blob: de984159cf0b8a0be6a73b9f421e9a0b4f75c599 [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();
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700203 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800204 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);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700235 __free_pages(zero_page, compound_order(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{
Johannes Weiner00501b52014-08-08 14:19:20 -0700718 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800719 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800720 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800721
Sasha Levin309381fea2014-01-23 15:52:54 -0800722 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700723
724 if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800726
Johannes Weiner00501b52014-08-08 14:19:20 -0700727 pgtable = pte_alloc_one(mm, haddr);
728 if (unlikely(!pgtable)) {
729 mem_cgroup_cancel_charge(page, memcg);
730 return VM_FAULT_OOM;
731 }
732
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800733 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700734 /*
735 * The memory barrier inside __SetPageUptodate makes sure that
736 * clear_huge_page writes become visible before the set_pmd_at()
737 * write.
738 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800739 __SetPageUptodate(page);
740
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800741 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800742 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800743 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -0700744 mem_cgroup_cancel_charge(page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800745 put_page(page);
746 pte_free(mm, pgtable);
747 } else {
748 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700749 entry = mk_huge_pmd(page, vma->vm_page_prot);
750 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800751 page_add_new_anon_rmap(page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700752 mem_cgroup_commit_charge(page, memcg, false);
753 lru_cache_add_active_or_unevictable(page, vma);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700754 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800755 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800756 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800757 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800758 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800759 }
760
David Rientjesaa2e8782012-05-29 15:06:17 -0700761 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800762}
763
Andi Kleencc5d4622011-03-22 16:33:13 -0700764static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800765{
Andi Kleencc5d4622011-03-22 16:33:13 -0700766 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800767}
768
769static inline struct page *alloc_hugepage_vma(int defrag,
770 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700771 unsigned long haddr, int nd,
772 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800773{
Andi Kleencc5d4622011-03-22 16:33:13 -0700774 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800775 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800776}
777
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800778/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800779static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800780 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700781 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800782{
783 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800784 if (!pmd_none(*pmd))
785 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700786 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800787 entry = pmd_wrprotect(entry);
788 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700789 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800790 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800791 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800792 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800793}
794
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800795int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
796 unsigned long address, pmd_t *pmd,
797 unsigned int flags)
798{
799 struct page *page;
800 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800801
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700802 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700803 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700804 if (unlikely(anon_vma_prepare(vma)))
805 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700806 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700807 return VM_FAULT_OOM;
808 if (!(flags & FAULT_FLAG_WRITE) &&
809 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800810 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700811 pgtable_t pgtable;
812 struct page *zero_page;
813 bool set;
814 pgtable = pte_alloc_one(mm, haddr);
815 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800816 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700817 zero_page = get_huge_zero_page();
818 if (unlikely(!zero_page)) {
819 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700820 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700821 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700822 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800823 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700824 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
825 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800826 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700827 if (!set) {
828 pte_free(mm, pgtable);
829 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800830 }
David Rientjesedad9d22012-05-29 15:06:17 -0700831 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800832 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700833 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
834 vma, haddr, numa_node_id(), 0);
835 if (unlikely(!page)) {
836 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700837 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700838 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700839 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700840 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700841 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700842 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700843 }
844
David Rientjes17766dd2013-09-12 15:14:06 -0700845 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700846 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800847}
848
849int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
850 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
851 struct vm_area_struct *vma)
852{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800853 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800854 struct page *src_page;
855 pmd_t pmd;
856 pgtable_t pgtable;
857 int ret;
858
859 ret = -ENOMEM;
860 pgtable = pte_alloc_one(dst_mm, addr);
861 if (unlikely(!pgtable))
862 goto out;
863
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800864 dst_ptl = pmd_lock(dst_mm, dst_pmd);
865 src_ptl = pmd_lockptr(src_mm, src_pmd);
866 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800867
868 ret = -EAGAIN;
869 pmd = *src_pmd;
870 if (unlikely(!pmd_trans_huge(pmd))) {
871 pte_free(dst_mm, pgtable);
872 goto out_unlock;
873 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800874 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800875 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800876 * under splitting since we don't split the page itself, only pmd to
877 * a page table.
878 */
879 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700880 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800881 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800882 /*
883 * get_huge_zero_page() will never allocate a new page here,
884 * since we already have a zero page to copy. It just takes a
885 * reference.
886 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700887 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800888 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700889 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800890 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800891 ret = 0;
892 goto out_unlock;
893 }
Mel Gormande466bd2013-12-18 17:08:42 -0800894
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800895 if (unlikely(pmd_trans_splitting(pmd))) {
896 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800897 spin_unlock(src_ptl);
898 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800899 pte_free(dst_mm, pgtable);
900
901 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
902 goto out;
903 }
904 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800905 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800906 get_page(src_page);
907 page_dup_rmap(src_page);
908 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
909
910 pmdp_set_wrprotect(src_mm, addr, src_pmd);
911 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700912 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800913 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800914 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800915
916 ret = 0;
917out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800918 spin_unlock(src_ptl);
919 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800920out:
921 return ret;
922}
923
Will Deacona1dd4502012-12-11 16:01:27 -0800924void huge_pmd_set_accessed(struct mm_struct *mm,
925 struct vm_area_struct *vma,
926 unsigned long address,
927 pmd_t *pmd, pmd_t orig_pmd,
928 int dirty)
929{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800930 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800931 pmd_t entry;
932 unsigned long haddr;
933
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800934 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800935 if (unlikely(!pmd_same(*pmd, orig_pmd)))
936 goto unlock;
937
938 entry = pmd_mkyoung(orig_pmd);
939 haddr = address & HPAGE_PMD_MASK;
940 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
941 update_mmu_cache_pmd(vma, address, pmd);
942
943unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800944 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800945}
946
Hugh Dickins5338a932014-06-23 13:22:05 -0700947/*
948 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
949 * during copy_user_huge_page()'s copy_page_rep(): in the case when
950 * the source page gets split and a tail freed before copy completes.
951 * Called under pmd_lock of checked pmd, so safe from splitting itself.
952 */
953static void get_user_huge_page(struct page *page)
954{
955 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
956 struct page *endpage = page + HPAGE_PMD_NR;
957
958 atomic_add(HPAGE_PMD_NR, &page->_count);
959 while (++page < endpage)
960 get_huge_page_tail(page);
961 } else {
962 get_page(page);
963 }
964}
965
966static void put_user_huge_page(struct page *page)
967{
968 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
969 struct page *endpage = page + HPAGE_PMD_NR;
970
971 while (page < endpage)
972 put_page(page++);
973 } else {
974 put_page(page);
975 }
976}
977
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800978static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
979 struct vm_area_struct *vma,
980 unsigned long address,
981 pmd_t *pmd, pmd_t orig_pmd,
982 struct page *page,
983 unsigned long haddr)
984{
Johannes Weiner00501b52014-08-08 14:19:20 -0700985 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800986 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800987 pgtable_t pgtable;
988 pmd_t _pmd;
989 int ret = 0, i;
990 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700991 unsigned long mmun_start; /* For mmu_notifiers */
992 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800993
994 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
995 GFP_KERNEL);
996 if (unlikely(!pages)) {
997 ret |= VM_FAULT_OOM;
998 goto out;
999 }
1000
1001 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001002 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1003 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001004 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001005 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -07001006 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1007 &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001008 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001009 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001010 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001011 memcg = (void *)page_private(pages[i]);
1012 set_page_private(pages[i], 0);
1013 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001014 put_page(pages[i]);
1015 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001016 kfree(pages);
1017 ret |= VM_FAULT_OOM;
1018 goto out;
1019 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001020 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001021 }
1022
1023 for (i = 0; i < HPAGE_PMD_NR; i++) {
1024 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001025 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001026 __SetPageUptodate(pages[i]);
1027 cond_resched();
1028 }
1029
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001030 mmun_start = haddr;
1031 mmun_end = haddr + HPAGE_PMD_SIZE;
1032 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1033
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001034 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001035 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1036 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001037 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001038
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001039 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001040 /* leave pmd empty until pte is filled */
1041
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001042 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001043 pmd_populate(mm, &_pmd, pgtable);
1044
1045 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1046 pte_t *pte, entry;
1047 entry = mk_pte(pages[i], vma->vm_page_prot);
1048 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001049 memcg = (void *)page_private(pages[i]);
1050 set_page_private(pages[i], 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001051 page_add_new_anon_rmap(pages[i], vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001052 mem_cgroup_commit_charge(pages[i], memcg, false);
1053 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001054 pte = pte_offset_map(&_pmd, haddr);
1055 VM_BUG_ON(!pte_none(*pte));
1056 set_pte_at(mm, haddr, pte, entry);
1057 pte_unmap(pte);
1058 }
1059 kfree(pages);
1060
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001061 smp_wmb(); /* make pte visible before pmd */
1062 pmd_populate(mm, pmd, pgtable);
1063 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001064 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001065
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001066 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1067
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001068 ret |= VM_FAULT_WRITE;
1069 put_page(page);
1070
1071out:
1072 return ret;
1073
1074out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001075 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001076 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001077 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001078 memcg = (void *)page_private(pages[i]);
1079 set_page_private(pages[i], 0);
1080 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001081 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001082 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001083 kfree(pages);
1084 goto out;
1085}
1086
1087int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1088 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1089{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001090 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001091 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001092 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001093 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001094 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001095 unsigned long mmun_start; /* For mmu_notifiers */
1096 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001097
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001098 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001099 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001100 haddr = address & HPAGE_PMD_MASK;
1101 if (is_huge_zero_pmd(orig_pmd))
1102 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001103 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001104 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1105 goto out_unlock;
1106
1107 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001108 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001109 if (page_mapcount(page) == 1) {
1110 pmd_t entry;
1111 entry = pmd_mkyoung(orig_pmd);
1112 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1113 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001114 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001115 ret |= VM_FAULT_WRITE;
1116 goto out_unlock;
1117 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001118 get_user_huge_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001119 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001120alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001121 if (transparent_hugepage_enabled(vma) &&
1122 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001123 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001124 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 else
1126 new_page = NULL;
1127
1128 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001129 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001130 split_huge_page_pmd(vma, address, pmd);
1131 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001132 } else {
1133 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1134 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001135 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001136 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001137 ret |= VM_FAULT_FALLBACK;
1138 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001139 put_user_huge_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001140 }
David Rientjes17766dd2013-09-12 15:14:06 -07001141 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001142 goto out;
1143 }
1144
Johannes Weiner00501b52014-08-08 14:19:20 -07001145 if (unlikely(mem_cgroup_try_charge(new_page, mm,
1146 GFP_TRANSHUGE, &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001147 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001148 if (page) {
1149 split_huge_page(page);
Hugh Dickins5338a932014-06-23 13:22:05 -07001150 put_user_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001151 } else
1152 split_huge_page_pmd(vma, address, pmd);
1153 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001154 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001155 goto out;
1156 }
1157
David Rientjes17766dd2013-09-12 15:14:06 -07001158 count_vm_event(THP_FAULT_ALLOC);
1159
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001160 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001161 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1162 else
1163 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 __SetPageUptodate(new_page);
1165
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001166 mmun_start = haddr;
1167 mmun_end = haddr + HPAGE_PMD_SIZE;
1168 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1169
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001170 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001171 if (page)
Hugh Dickins5338a932014-06-23 13:22:05 -07001172 put_user_huge_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001173 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001174 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -07001175 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001176 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001177 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001178 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001179 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001180 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1181 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001182 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001183 page_add_new_anon_rmap(new_page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001184 mem_cgroup_commit_charge(new_page, memcg, false);
1185 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001186 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001187 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001188 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001189 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001190 put_huge_zero_page();
1191 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001192 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001193 page_remove_rmap(page);
1194 put_page(page);
1195 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001196 ret |= VM_FAULT_WRITE;
1197 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001198 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001199out_mn:
1200 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1201out:
1202 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001203out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001204 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205 return ret;
1206}
1207
David Rientjesb676b292012-10-08 16:34:03 -07001208struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209 unsigned long addr,
1210 pmd_t *pmd,
1211 unsigned int flags)
1212{
David Rientjesb676b292012-10-08 16:34:03 -07001213 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001214 struct page *page = NULL;
1215
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001216 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001217
1218 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1219 goto out;
1220
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001221 /* Avoid dumping huge zero page */
1222 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1223 return ERR_PTR(-EFAULT);
1224
Mel Gorman2b4847e2013-12-18 17:08:32 -08001225 /* Full NUMA hinting faults to serialise migration in fault paths */
1226 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1227 goto out;
1228
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001229 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001230 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001231 if (flags & FOLL_TOUCH) {
1232 pmd_t _pmd;
1233 /*
1234 * We should set the dirty bit only for FOLL_WRITE but
1235 * for now the dirty bit in the pmd is meaningless.
1236 * And if the dirty bit will become meaningful and
1237 * we'll only set it with FOLL_WRITE, an atomic
1238 * set_bit will be required on the pmd to set the
1239 * young bit, instead of the current set_pmd_at.
1240 */
1241 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001242 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1243 pmd, _pmd, 1))
1244 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001245 }
David Rientjesb676b292012-10-08 16:34:03 -07001246 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1247 if (page->mapping && trylock_page(page)) {
1248 lru_add_drain();
1249 if (page->mapping)
1250 mlock_vma_page(page);
1251 unlock_page(page);
1252 }
1253 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001254 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001255 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001256 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001257 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001258
1259out:
1260 return page;
1261}
1262
Mel Gormand10e63f2012-10-25 14:16:31 +02001263/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001264int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1265 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001266{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001267 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001268 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001269 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001270 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001271 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001272 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001273 bool page_locked;
1274 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001275 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001276
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001277 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001278 if (unlikely(!pmd_same(pmd, *pmdp)))
1279 goto out_unlock;
1280
Mel Gormande466bd2013-12-18 17:08:42 -08001281 /*
1282 * If there are potential migrations, wait for completion and retry
1283 * without disrupting NUMA hinting information. Do not relock and
1284 * check_same as the page may no longer be mapped.
1285 */
1286 if (unlikely(pmd_trans_migrating(*pmdp))) {
1287 spin_unlock(ptl);
1288 wait_migrate_huge_page(vma->anon_vma, pmdp);
1289 goto out;
1290 }
1291
Mel Gormand10e63f2012-10-25 14:16:31 +02001292 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001293 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001294 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001295 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001296 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001297 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001298 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001299 flags |= TNF_FAULT_LOCAL;
1300 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001301
Mel Gormanff9042b2013-10-07 11:28:43 +01001302 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001303 * Avoid grouping on DSO/COW pages in specific and RO pages
1304 * in general, RO pages shouldn't hurt as much anyway since
1305 * they can be in shared cache state.
1306 */
1307 if (!pmd_write(pmd))
1308 flags |= TNF_NO_GROUP;
1309
1310 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001311 * Acquire the page lock to serialise THP migrations but avoid dropping
1312 * page_table_lock if at all possible
1313 */
Mel Gormanb8916632013-10-07 11:28:44 +01001314 page_locked = trylock_page(page);
1315 target_nid = mpol_misplaced(page, vma, haddr);
1316 if (target_nid == -1) {
1317 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001318 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001319 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001320 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001321
Mel Gormande466bd2013-12-18 17:08:42 -08001322 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001323 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001324 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001325 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001326 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001327 goto out;
1328 }
1329
Mel Gorman2b4847e2013-12-18 17:08:32 -08001330 /*
1331 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1332 * to serialises splits
1333 */
Mel Gormanb8916632013-10-07 11:28:44 +01001334 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001335 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001336 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001337
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001338 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001339 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001340 if (unlikely(!pmd_same(pmd, *pmdp))) {
1341 unlock_page(page);
1342 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001343 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001344 goto out_unlock;
1345 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001346
Mel Gormanc3a489c2013-12-18 17:08:38 -08001347 /* Bail if we fail to protect against THP splits for any reason */
1348 if (unlikely(!anon_vma)) {
1349 put_page(page);
1350 page_nid = -1;
1351 goto clear_pmdnuma;
1352 }
1353
Mel Gormana54a4072013-10-07 11:28:46 +01001354 /*
1355 * Migrate the THP to the requested node, returns with page unlocked
1356 * and pmd_numa cleared.
1357 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001358 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001359 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001360 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001361 if (migrated) {
1362 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001363 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001364 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001365
Mel Gorman8191acb2013-10-07 11:28:45 +01001366 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001367clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001368 BUG_ON(!PageLocked(page));
Mel Gormand10e63f2012-10-25 14:16:31 +02001369 pmd = pmd_mknonnuma(pmd);
1370 set_pmd_at(mm, haddr, pmdp, pmd);
1371 VM_BUG_ON(pmd_numa(*pmdp));
1372 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001373 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001374out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001375 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001376
1377out:
1378 if (anon_vma)
1379 page_unlock_anon_vma_read(anon_vma);
1380
Mel Gorman8191acb2013-10-07 11:28:45 +01001381 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001382 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001383
Mel Gormand10e63f2012-10-25 14:16:31 +02001384 return 0;
1385}
1386
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001387int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001388 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001389{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001390 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001391 int ret = 0;
1392
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001393 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001394 struct page *page;
1395 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001396 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001397 /*
1398 * For architectures like ppc64 we look at deposited pgtable
1399 * when calling pmdp_get_and_clear. So do the
1400 * pgtable_trans_huge_withdraw after finishing pmdp related
1401 * operations.
1402 */
David Millerf5c8ad42012-10-08 16:34:26 -07001403 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001404 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001405 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001406 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001407 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001408 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001409 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001410 } else {
1411 page = pmd_page(orig_pmd);
1412 page_remove_rmap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001413 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001414 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
Sasha Levin309381fea2014-01-23 15:52:54 -08001415 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001416 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001417 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001418 tlb_remove_page(tlb, page);
1419 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001420 pte_free(tlb->mm, pgtable);
1421 ret = 1;
1422 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001423 return ret;
1424}
1425
Johannes Weiner0ca16342011-01-13 15:47:02 -08001426int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427 unsigned long addr, unsigned long end,
1428 unsigned char *vec)
1429{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001430 spinlock_t *ptl;
Johannes Weiner0ca16342011-01-13 15:47:02 -08001431 int ret = 0;
1432
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001433 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001434 /*
1435 * All logical pages in the range are present
1436 * if backed by a huge page.
1437 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001438 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001439 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440 ret = 1;
1441 }
Johannes Weiner0ca16342011-01-13 15:47:02 -08001442
1443 return ret;
1444}
1445
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001446int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447 unsigned long old_addr,
1448 unsigned long new_addr, unsigned long old_end,
1449 pmd_t *old_pmd, pmd_t *new_pmd)
1450{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001451 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001452 int ret = 0;
1453 pmd_t pmd;
1454
1455 struct mm_struct *mm = vma->vm_mm;
1456
1457 if ((old_addr & ~HPAGE_PMD_MASK) ||
1458 (new_addr & ~HPAGE_PMD_MASK) ||
1459 old_end - old_addr < HPAGE_PMD_SIZE ||
1460 (new_vma->vm_flags & VM_NOHUGEPAGE))
1461 goto out;
1462
1463 /*
1464 * The destination pmd shouldn't be established, free_pgtables()
1465 * should have release it.
1466 */
1467 if (WARN_ON(!pmd_none(*new_pmd))) {
1468 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469 goto out;
1470 }
1471
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001472 /*
1473 * We don't have to worry about the ordering of src and dst
1474 * ptlocks because exclusive mmap_sem prevents deadlock.
1475 */
1476 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001477 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001478 new_ptl = pmd_lockptr(mm, new_pmd);
1479 if (new_ptl != old_ptl)
1480 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001481 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001483
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301484 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1485 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001486 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1487 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001488 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301489 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1490 if (new_ptl != old_ptl)
1491 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001492 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001493 }
1494out:
1495 return ret;
1496}
1497
Mel Gormanf123d742013-10-07 11:28:49 +01001498/*
1499 * Returns
1500 * - 0 if PMD could not be locked
1501 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1502 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1503 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001504int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001505 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001506{
1507 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001508 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001509 int ret = 0;
1510
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001511 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001512 pmd_t entry;
Mel Gormanf123d742013-10-07 11:28:49 +01001513 ret = 1;
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001514 if (!prot_numa) {
Mel Gormanf123d742013-10-07 11:28:49 +01001515 entry = pmdp_get_and_clear(mm, addr, pmd);
Mel Gorman16679182013-12-18 17:08:41 -08001516 if (pmd_numa(entry))
1517 entry = pmd_mknonnuma(entry);
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001518 entry = pmd_modify(entry, newprot);
Mel Gormanf123d742013-10-07 11:28:49 +01001519 ret = HPAGE_PMD_NR;
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301520 set_pmd_at(mm, addr, pmd, entry);
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001521 BUG_ON(pmd_write(entry));
1522 } else {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001523 struct page *page = pmd_page(*pmd);
1524
Mel Gormana1a46182013-10-07 11:28:50 +01001525 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001526 * Do not trap faults against the zero page. The
1527 * read-only data is likely to be read-cached on the
1528 * local CPU cache and it is less useful to know about
1529 * local vs remote hits on the zero page.
Mel Gormana1a46182013-10-07 11:28:50 +01001530 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001531 if (!is_huge_zero_page(page) &&
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001532 !pmd_numa(*pmd)) {
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301533 pmdp_set_numa(mm, addr, pmd);
Mel Gormanf123d742013-10-07 11:28:49 +01001534 ret = HPAGE_PMD_NR;
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001535 }
1536 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001537 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001538 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001539
1540 return ret;
1541}
1542
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001543/*
1544 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1545 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1546 *
1547 * Note that if it returns 1, this routine returns without unlocking page
1548 * table locks. So callers must unlock them.
1549 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001550int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1551 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001552{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001553 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001554 if (likely(pmd_trans_huge(*pmd))) {
1555 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001556 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001557 wait_split_huge_page(vma->anon_vma, pmd);
1558 return -1;
1559 } else {
1560 /* Thp mapped by 'pmd' is stable, so we can
1561 * handle it as it is. */
1562 return 1;
1563 }
1564 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001565 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001566 return 0;
1567}
1568
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001569/*
1570 * This function returns whether a given @page is mapped onto the @address
1571 * in the virtual space of @mm.
1572 *
1573 * When it's true, this function returns *pmd with holding the page table lock
1574 * and passing it back to the caller via @ptl.
1575 * If it's false, returns NULL without holding the page table lock.
1576 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001577pmd_t *page_check_address_pmd(struct page *page,
1578 struct mm_struct *mm,
1579 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001580 enum page_check_address_pmd_flag flag,
1581 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001582{
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001583 pgd_t *pgd;
1584 pud_t *pud;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001585 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001586
1587 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001588 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001589
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001590 pgd = pgd_offset(mm, address);
1591 if (!pgd_present(*pgd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001592 return NULL;
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001593 pud = pud_offset(pgd, address);
1594 if (!pud_present(*pud))
1595 return NULL;
1596 pmd = pmd_offset(pud, address);
1597
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001598 *ptl = pmd_lock(mm, pmd);
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001599 if (!pmd_present(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001600 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001601 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001602 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001603 /*
1604 * split_vma() may create temporary aliased mappings. There is
1605 * no risk as long as all huge pmd are found and have their
1606 * splitting bit set before __split_huge_page_refcount
1607 * runs. Finding the same huge pmd more than once during the
1608 * same rmap walk is not a problem.
1609 */
1610 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1611 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001612 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001613 if (pmd_trans_huge(*pmd)) {
1614 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1615 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001616 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001617 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001618unlock:
1619 spin_unlock(*ptl);
1620 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001621}
1622
1623static int __split_huge_page_splitting(struct page *page,
1624 struct vm_area_struct *vma,
1625 unsigned long address)
1626{
1627 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001628 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001629 pmd_t *pmd;
1630 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001631 /* For mmu_notifiers */
1632 const unsigned long mmun_start = address;
1633 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001634
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001635 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001636 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001637 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001638 if (pmd) {
1639 /*
1640 * We can't temporarily set the pmd to null in order
1641 * to split it, the pmd must remain marked huge at all
1642 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001643 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001644 * serialize against split_huge_page*.
1645 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001646 pmdp_splitting_flush(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001647 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001648 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001649 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001650 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001651
1652 return ret;
1653}
1654
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001655static void __split_huge_page_refcount(struct page *page,
1656 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001657{
1658 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001659 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001660 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001661 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001662
1663 /* prevent PageLRU to go away from under us, and freeze lru stats */
1664 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001665 lruvec = mem_cgroup_page_lruvec(page, zone);
1666
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001667 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001668 /* complete memcg works before add pages to LRU */
1669 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001670
Shaohua Li45676882012-01-12 17:19:18 -08001671 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001672 struct page *page_tail = page + i;
1673
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001674 /* tail_page->_mapcount cannot change */
1675 BUG_ON(page_mapcount(page_tail) < 0);
1676 tail_count += page_mapcount(page_tail);
1677 /* check for overflow */
1678 BUG_ON(tail_count < 0);
1679 BUG_ON(atomic_read(&page_tail->_count) != 0);
1680 /*
1681 * tail_page->_count is zero and not changing from
1682 * under us. But get_page_unless_zero() may be running
1683 * from under us on the tail_page. If we used
1684 * atomic_set() below instead of atomic_add(), we
1685 * would then run atomic_set() concurrently with
1686 * get_page_unless_zero(), and atomic_set() is
1687 * implemented in C not using locked ops. spin_unlock
1688 * on x86 sometime uses locked ops because of PPro
1689 * errata 66, 92, so unless somebody can guarantee
1690 * atomic_set() here would be safe on all archs (and
1691 * not only on x86), it's safer to use atomic_add().
1692 */
1693 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1694 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001695
1696 /* after clearing PageTail the gup refcount can be released */
Waiman Long3a79d522014-08-06 16:05:38 -07001697 smp_mb__after_atomic();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001698
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001699 /*
1700 * retain hwpoison flag of the poisoned tail page:
1701 * fix for the unsuitable process killed on Guest Machine(KVM)
1702 * by the memory-failure.
1703 */
1704 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001705 page_tail->flags |= (page->flags &
1706 ((1L << PG_referenced) |
1707 (1L << PG_swapbacked) |
1708 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001709 (1L << PG_uptodate) |
1710 (1L << PG_active) |
1711 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001712 page_tail->flags |= (1L << PG_dirty);
1713
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001714 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001715 smp_wmb();
1716
1717 /*
1718 * __split_huge_page_splitting() already set the
1719 * splitting bit in all pmd that could map this
1720 * hugepage, that will ensure no CPU can alter the
1721 * mapcount on the head page. The mapcount is only
1722 * accounted in the head page and it has to be
1723 * transferred to all tail pages in the below code. So
1724 * for this code to be safe, the split the mapcount
1725 * can't change. But that doesn't mean userland can't
1726 * keep changing and reading the page contents while
1727 * we transfer the mapcount, so the pmd splitting
1728 * status is achieved setting a reserved bit in the
1729 * pmd, not by clearing the present bit.
1730 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001731 page_tail->_mapcount = page->_mapcount;
1732
1733 BUG_ON(page_tail->mapping);
1734 page_tail->mapping = page->mapping;
1735
Shaohua Li45676882012-01-12 17:19:18 -08001736 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001737 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001738
1739 BUG_ON(!PageAnon(page_tail));
1740 BUG_ON(!PageUptodate(page_tail));
1741 BUG_ON(!PageDirty(page_tail));
1742 BUG_ON(!PageSwapBacked(page_tail));
1743
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001744 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001745 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001746 atomic_sub(tail_count, &page->_count);
1747 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001748
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001749 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001750
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001751 ClearPageCompound(page);
1752 compound_unlock(page);
1753 spin_unlock_irq(&zone->lru_lock);
1754
1755 for (i = 1; i < HPAGE_PMD_NR; i++) {
1756 struct page *page_tail = page + i;
1757 BUG_ON(page_count(page_tail) <= 0);
1758 /*
1759 * Tail pages may be freed if there wasn't any mapping
1760 * like if add_to_swap() is running on a lru page that
1761 * had its mapping zapped. And freeing these pages
1762 * requires taking the lru_lock so we do the put_page
1763 * of the tail pages after the split is complete.
1764 */
1765 put_page(page_tail);
1766 }
1767
1768 /*
1769 * Only the head page (now become a regular page) is required
1770 * to be pinned by the caller.
1771 */
1772 BUG_ON(page_count(page) <= 0);
1773}
1774
1775static int __split_huge_page_map(struct page *page,
1776 struct vm_area_struct *vma,
1777 unsigned long address)
1778{
1779 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001780 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001781 pmd_t *pmd, _pmd;
1782 int ret = 0, i;
1783 pgtable_t pgtable;
1784 unsigned long haddr;
1785
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001786 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001787 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001788 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001789 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001790 pmd_populate(mm, &_pmd, pgtable);
Waiman Longf8303c22014-08-06 16:05:36 -07001791 if (pmd_write(*pmd))
1792 BUG_ON(page_mapcount(page) != 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001793
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001794 haddr = address;
1795 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001796 pte_t *pte, entry;
1797 BUG_ON(PageCompound(page+i));
Mel Gormanabc40bd2014-10-02 19:47:42 +01001798 /*
1799 * Note that pmd_numa is not transferred deliberately
1800 * to avoid any possibility that pte_numa leaks to
1801 * a PROT_NONE VMA by accident.
1802 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001803 entry = mk_pte(page + i, vma->vm_page_prot);
1804 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1805 if (!pmd_write(*pmd))
1806 entry = pte_wrprotect(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001807 if (!pmd_young(*pmd))
1808 entry = pte_mkold(entry);
1809 pte = pte_offset_map(&_pmd, haddr);
1810 BUG_ON(!pte_none(*pte));
1811 set_pte_at(mm, haddr, pte, entry);
1812 pte_unmap(pte);
1813 }
1814
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001815 smp_wmb(); /* make pte visible before pmd */
1816 /*
1817 * Up to this point the pmd is present and huge and
1818 * userland has the whole access to the hugepage
1819 * during the split (which happens in place). If we
1820 * overwrite the pmd with the not-huge version
1821 * pointing to the pte here (which of course we could
1822 * if all CPUs were bug free), userland could trigger
1823 * a small page size TLB miss on the small sized TLB
1824 * while the hugepage TLB entry is still established
1825 * in the huge TLB. Some CPU doesn't like that. See
1826 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1827 * Erratum 383 on page 93. Intel should be safe but is
1828 * also warns that it's only safe if the permission
1829 * and cache attributes of the two entries loaded in
1830 * the two TLB is identical (which should be the case
1831 * here). But it is generally safer to never allow
1832 * small and huge TLB entries for the same virtual
1833 * address to be loaded simultaneously. So instead of
1834 * doing "pmd_populate(); flush_tlb_range();" we first
1835 * mark the current pmd notpresent (atomically because
1836 * here the pmd_trans_huge and pmd_trans_splitting
1837 * must remain set at all times on the pmd until the
1838 * split is complete for this pmd), then we flush the
1839 * SMP TLB and finally we write the non-huge version
1840 * of the pmd entry with pmd_populate.
1841 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001842 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001843 pmd_populate(mm, pmd, pgtable);
1844 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001845 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001846 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001847
1848 return ret;
1849}
1850
Ingo Molnar5a505082012-12-02 19:56:46 +00001851/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001852static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001853 struct anon_vma *anon_vma,
1854 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001855{
1856 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001857 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001858 struct anon_vma_chain *avc;
1859
1860 BUG_ON(!PageHead(page));
1861 BUG_ON(PageTail(page));
1862
1863 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001864 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001865 struct vm_area_struct *vma = avc->vma;
1866 unsigned long addr = vma_address(page, vma);
1867 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001868 mapcount += __split_huge_page_splitting(page, vma, addr);
1869 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001870 /*
1871 * It is critical that new vmas are added to the tail of the
1872 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1873 * and establishes a child pmd before
1874 * __split_huge_page_splitting() freezes the parent pmd (so if
1875 * we fail to prevent copy_huge_pmd() from running until the
1876 * whole __split_huge_page() is complete), we will still see
1877 * the newly established pmd of the child later during the
1878 * walk, to be able to set it as pmd_trans_splitting too.
1879 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001880 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001881 pr_err("mapcount %d page_mapcount %d\n",
1882 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001883 BUG();
1884 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001885
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001886 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001887
1888 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001889 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001890 struct vm_area_struct *vma = avc->vma;
1891 unsigned long addr = vma_address(page, vma);
1892 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001893 mapcount2 += __split_huge_page_map(page, vma, addr);
1894 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001895 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001896 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1897 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001898 BUG();
1899 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001900}
1901
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001902/*
1903 * Split a hugepage into normal pages. This doesn't change the position of head
1904 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1905 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1906 * from the hugepage.
1907 * Return 0 if the hugepage is split successfully otherwise return 1.
1908 */
1909int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001910{
1911 struct anon_vma *anon_vma;
1912 int ret = 1;
1913
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001914 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001915 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001916
1917 /*
1918 * The caller does not necessarily hold an mmap_sem that would prevent
1919 * the anon_vma disappearing so we first we take a reference to it
1920 * and then lock the anon_vma for write. This is similar to
1921 * page_lock_anon_vma_read except the write lock is taken to serialise
1922 * against parallel split or collapse operations.
1923 */
1924 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001925 if (!anon_vma)
1926 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001927 anon_vma_lock_write(anon_vma);
1928
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001929 ret = 0;
1930 if (!PageCompound(page))
1931 goto out_unlock;
1932
1933 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001934 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001935 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001936
1937 BUG_ON(PageCompound(page));
1938out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001939 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001940 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001941out:
1942 return ret;
1943}
1944
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001945#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001946
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001947int hugepage_madvise(struct vm_area_struct *vma,
1948 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001949{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001950 switch (advice) {
1951 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001952#ifdef CONFIG_S390
1953 /*
1954 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1955 * can't handle this properly after s390_enable_sie, so we simply
1956 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1957 */
1958 if (mm_has_pgste(vma->vm_mm))
1959 return 0;
1960#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001961 /*
1962 * Be somewhat over-protective like KSM for now!
1963 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001964 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001965 return -EINVAL;
1966 *vm_flags &= ~VM_NOHUGEPAGE;
1967 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001968 /*
1969 * If the vma become good for khugepaged to scan,
1970 * register it here without waiting a page fault that
1971 * may not happen any time soon.
1972 */
David Rientjes6d50e602014-10-29 14:50:31 -07001973 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001974 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001975 break;
1976 case MADV_NOHUGEPAGE:
1977 /*
1978 * Be somewhat over-protective like KSM for now!
1979 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001980 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001981 return -EINVAL;
1982 *vm_flags &= ~VM_HUGEPAGE;
1983 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001984 /*
1985 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1986 * this vma even if we leave the mm registered in khugepaged if
1987 * it got registered before VM_NOHUGEPAGE was set.
1988 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001989 break;
1990 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001991
1992 return 0;
1993}
1994
Andrea Arcangeliba761492011-01-13 15:46:58 -08001995static int __init khugepaged_slab_init(void)
1996{
1997 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1998 sizeof(struct mm_slot),
1999 __alignof__(struct mm_slot), 0, NULL);
2000 if (!mm_slot_cache)
2001 return -ENOMEM;
2002
2003 return 0;
2004}
2005
Andrea Arcangeliba761492011-01-13 15:46:58 -08002006static inline struct mm_slot *alloc_mm_slot(void)
2007{
2008 if (!mm_slot_cache) /* initialization failed */
2009 return NULL;
2010 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2011}
2012
2013static inline void free_mm_slot(struct mm_slot *mm_slot)
2014{
2015 kmem_cache_free(mm_slot_cache, mm_slot);
2016}
2017
Andrea Arcangeliba761492011-01-13 15:46:58 -08002018static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2019{
2020 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002021
Sasha Levinb67bfe02013-02-27 17:06:00 -08002022 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002023 if (mm == mm_slot->mm)
2024 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002025
Andrea Arcangeliba761492011-01-13 15:46:58 -08002026 return NULL;
2027}
2028
2029static void insert_to_mm_slots_hash(struct mm_struct *mm,
2030 struct mm_slot *mm_slot)
2031{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002032 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002033 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002034}
2035
2036static inline int khugepaged_test_exit(struct mm_struct *mm)
2037{
2038 return atomic_read(&mm->mm_users) == 0;
2039}
2040
2041int __khugepaged_enter(struct mm_struct *mm)
2042{
2043 struct mm_slot *mm_slot;
2044 int wakeup;
2045
2046 mm_slot = alloc_mm_slot();
2047 if (!mm_slot)
2048 return -ENOMEM;
2049
2050 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07002051 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002052 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2053 free_mm_slot(mm_slot);
2054 return 0;
2055 }
2056
2057 spin_lock(&khugepaged_mm_lock);
2058 insert_to_mm_slots_hash(mm, mm_slot);
2059 /*
2060 * Insert just behind the scanning cursor, to let the area settle
2061 * down a little.
2062 */
2063 wakeup = list_empty(&khugepaged_scan.mm_head);
2064 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2065 spin_unlock(&khugepaged_mm_lock);
2066
2067 atomic_inc(&mm->mm_count);
2068 if (wakeup)
2069 wake_up_interruptible(&khugepaged_wait);
2070
2071 return 0;
2072}
2073
David Rientjes6d50e602014-10-29 14:50:31 -07002074int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2075 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002076{
2077 unsigned long hstart, hend;
2078 if (!vma->anon_vma)
2079 /*
2080 * Not yet faulted in so we will register later in the
2081 * page fault if needed.
2082 */
2083 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002084 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002085 /* khugepaged not yet working on file or special mappings */
2086 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07002087 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002088 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2089 hend = vma->vm_end & HPAGE_PMD_MASK;
2090 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07002091 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002092 return 0;
2093}
2094
2095void __khugepaged_exit(struct mm_struct *mm)
2096{
2097 struct mm_slot *mm_slot;
2098 int free = 0;
2099
2100 spin_lock(&khugepaged_mm_lock);
2101 mm_slot = get_mm_slot(mm);
2102 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002103 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002104 list_del(&mm_slot->mm_node);
2105 free = 1;
2106 }
Chris Wrightd788e802011-07-25 17:12:14 -07002107 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002108
2109 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002110 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2111 free_mm_slot(mm_slot);
2112 mmdrop(mm);
2113 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002114 /*
2115 * This is required to serialize against
2116 * khugepaged_test_exit() (which is guaranteed to run
2117 * under mmap sem read mode). Stop here (after we
2118 * return all pagetables will be destroyed) until
2119 * khugepaged has finished working on the pagetables
2120 * under the mmap_sem.
2121 */
2122 down_write(&mm->mmap_sem);
2123 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002124 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002125}
2126
2127static void release_pte_page(struct page *page)
2128{
2129 /* 0 stands for page_is_file_cache(page) == false */
2130 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2131 unlock_page(page);
2132 putback_lru_page(page);
2133}
2134
2135static void release_pte_pages(pte_t *pte, pte_t *_pte)
2136{
2137 while (--_pte >= pte) {
2138 pte_t pteval = *_pte;
2139 if (!pte_none(pteval))
2140 release_pte_page(pte_page(pteval));
2141 }
2142}
2143
Andrea Arcangeliba761492011-01-13 15:46:58 -08002144static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2145 unsigned long address,
2146 pte_t *pte)
2147{
2148 struct page *page;
2149 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002150 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002151 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2152 _pte++, address += PAGE_SIZE) {
2153 pte_t pteval = *_pte;
2154 if (pte_none(pteval)) {
2155 if (++none <= khugepaged_max_ptes_none)
2156 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002157 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002158 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002159 }
Bob Liu344aa352012-12-11 16:00:34 -08002160 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002161 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002162 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002163 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002164 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002165
Sasha Levin309381fea2014-01-23 15:52:54 -08002166 VM_BUG_ON_PAGE(PageCompound(page), page);
2167 VM_BUG_ON_PAGE(!PageAnon(page), page);
2168 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002169
2170 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002171 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002172 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002173 /*
2174 * We can do it before isolate_lru_page because the
2175 * page can't be freed from under us. NOTE: PG_lock
2176 * is needed to serialize against split_huge_page
2177 * when invoked from the VM.
2178 */
Bob Liu344aa352012-12-11 16:00:34 -08002179 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002180 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002181 /*
2182 * Isolate the page to avoid collapsing an hugepage
2183 * currently in use by the VM.
2184 */
2185 if (isolate_lru_page(page)) {
2186 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002187 goto out;
2188 }
2189 /* 0 stands for page_is_file_cache(page) == false */
2190 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002191 VM_BUG_ON_PAGE(!PageLocked(page), page);
2192 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002193
2194 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002195 if (pte_young(pteval) || PageReferenced(page) ||
2196 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002197 referenced = 1;
2198 }
Bob Liu344aa352012-12-11 16:00:34 -08002199 if (likely(referenced))
2200 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002201out:
Bob Liu344aa352012-12-11 16:00:34 -08002202 release_pte_pages(pte, _pte);
2203 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002204}
2205
2206static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2207 struct vm_area_struct *vma,
2208 unsigned long address,
2209 spinlock_t *ptl)
2210{
2211 pte_t *_pte;
2212 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2213 pte_t pteval = *_pte;
2214 struct page *src_page;
2215
2216 if (pte_none(pteval)) {
2217 clear_user_highpage(page, address);
2218 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2219 } else {
2220 src_page = pte_page(pteval);
2221 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002222 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002223 release_pte_page(src_page);
2224 /*
2225 * ptl mostly unnecessary, but preempt has to
2226 * be disabled to update the per-cpu stats
2227 * inside page_remove_rmap().
2228 */
2229 spin_lock(ptl);
2230 /*
2231 * paravirt calls inside pte_clear here are
2232 * superfluous.
2233 */
2234 pte_clear(vma->vm_mm, address, _pte);
2235 page_remove_rmap(src_page);
2236 spin_unlock(ptl);
2237 free_page_and_swap_cache(src_page);
2238 }
2239
2240 address += PAGE_SIZE;
2241 page++;
2242 }
2243}
2244
Xiao Guangrong26234f32012-10-08 16:29:51 -07002245static void khugepaged_alloc_sleep(void)
2246{
2247 wait_event_freezable_timeout(khugepaged_wait, false,
2248 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2249}
2250
Bob Liu9f1b8682013-11-12 15:07:37 -08002251static int khugepaged_node_load[MAX_NUMNODES];
2252
David Rientjes14a4e212014-08-06 16:07:29 -07002253static bool khugepaged_scan_abort(int nid)
2254{
2255 int i;
2256
2257 /*
2258 * If zone_reclaim_mode is disabled, then no extra effort is made to
2259 * allocate memory locally.
2260 */
2261 if (!zone_reclaim_mode)
2262 return false;
2263
2264 /* If there is a count for this node already, it must be acceptable */
2265 if (khugepaged_node_load[nid])
2266 return false;
2267
2268 for (i = 0; i < MAX_NUMNODES; i++) {
2269 if (!khugepaged_node_load[i])
2270 continue;
2271 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2272 return true;
2273 }
2274 return false;
2275}
2276
Xiao Guangrong26234f32012-10-08 16:29:51 -07002277#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002278static int khugepaged_find_target_node(void)
2279{
2280 static int last_khugepaged_target_node = NUMA_NO_NODE;
2281 int nid, target_node = 0, max_value = 0;
2282
2283 /* find first node with max normal pages hit */
2284 for (nid = 0; nid < MAX_NUMNODES; nid++)
2285 if (khugepaged_node_load[nid] > max_value) {
2286 max_value = khugepaged_node_load[nid];
2287 target_node = nid;
2288 }
2289
2290 /* do some balance if several nodes have the same hit record */
2291 if (target_node <= last_khugepaged_target_node)
2292 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2293 nid++)
2294 if (max_value == khugepaged_node_load[nid]) {
2295 target_node = nid;
2296 break;
2297 }
2298
2299 last_khugepaged_target_node = target_node;
2300 return target_node;
2301}
2302
Xiao Guangrong26234f32012-10-08 16:29:51 -07002303static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2304{
2305 if (IS_ERR(*hpage)) {
2306 if (!*wait)
2307 return false;
2308
2309 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002310 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002311 khugepaged_alloc_sleep();
2312 } else if (*hpage) {
2313 put_page(*hpage);
2314 *hpage = NULL;
2315 }
2316
2317 return true;
2318}
2319
2320static struct page
2321*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2322 struct vm_area_struct *vma, unsigned long address,
2323 int node)
2324{
Sasha Levin309381fea2014-01-23 15:52:54 -08002325 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002326
Xiao Guangrong26234f32012-10-08 16:29:51 -07002327 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002328 * Before allocating the hugepage, release the mmap_sem read lock.
2329 * The allocation can take potentially a long time if it involves
2330 * sync compaction, and we do not need to hold the mmap_sem during
2331 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002332 */
2333 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002334
2335 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2336 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002337 if (unlikely(!*hpage)) {
2338 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2339 *hpage = ERR_PTR(-ENOMEM);
2340 return NULL;
2341 }
2342
2343 count_vm_event(THP_COLLAPSE_ALLOC);
2344 return *hpage;
2345}
2346#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002347static int khugepaged_find_target_node(void)
2348{
2349 return 0;
2350}
2351
Bob Liu10dc4152013-11-12 15:07:35 -08002352static inline struct page *alloc_hugepage(int defrag)
2353{
2354 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2355 HPAGE_PMD_ORDER);
2356}
2357
Xiao Guangrong26234f32012-10-08 16:29:51 -07002358static struct page *khugepaged_alloc_hugepage(bool *wait)
2359{
2360 struct page *hpage;
2361
2362 do {
2363 hpage = alloc_hugepage(khugepaged_defrag());
2364 if (!hpage) {
2365 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2366 if (!*wait)
2367 return NULL;
2368
2369 *wait = false;
2370 khugepaged_alloc_sleep();
2371 } else
2372 count_vm_event(THP_COLLAPSE_ALLOC);
2373 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2374
2375 return hpage;
2376}
2377
2378static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2379{
2380 if (!*hpage)
2381 *hpage = khugepaged_alloc_hugepage(wait);
2382
2383 if (unlikely(!*hpage))
2384 return false;
2385
2386 return true;
2387}
2388
2389static struct page
2390*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2391 struct vm_area_struct *vma, unsigned long address,
2392 int node)
2393{
2394 up_read(&mm->mmap_sem);
2395 VM_BUG_ON(!*hpage);
2396 return *hpage;
2397}
2398#endif
2399
Bob Liufa475e52012-12-11 16:00:39 -08002400static bool hugepage_vma_check(struct vm_area_struct *vma)
2401{
2402 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2403 (vma->vm_flags & VM_NOHUGEPAGE))
2404 return false;
2405
2406 if (!vma->anon_vma || vma->vm_ops)
2407 return false;
2408 if (is_vma_temporary_stack(vma))
2409 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002410 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002411 return true;
2412}
2413
Andrea Arcangeliba761492011-01-13 15:46:58 -08002414static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002415 unsigned long address,
2416 struct page **hpage,
2417 struct vm_area_struct *vma,
2418 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002419{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002420 pmd_t *pmd, _pmd;
2421 pte_t *pte;
2422 pgtable_t pgtable;
2423 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002424 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002425 int isolated;
2426 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002427 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002428 unsigned long mmun_start; /* For mmu_notifiers */
2429 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002430
2431 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002432
Xiao Guangrong26234f32012-10-08 16:29:51 -07002433 /* release the mmap_sem read lock. */
2434 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2435 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002436 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002437
Johannes Weiner00501b52014-08-08 14:19:20 -07002438 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2439 GFP_TRANSHUGE, &memcg)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002440 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002441
2442 /*
2443 * Prevent all access to pagetables with the exception of
2444 * gup_fast later hanlded by the ptep_clear_flush and the VM
2445 * handled by the anon_vma lock + PG_lock.
2446 */
2447 down_write(&mm->mmap_sem);
2448 if (unlikely(khugepaged_test_exit(mm)))
2449 goto out;
2450
2451 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002452 if (!vma)
2453 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002454 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2455 hend = vma->vm_end & HPAGE_PMD_MASK;
2456 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2457 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002458 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002459 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002460 pmd = mm_find_pmd(mm, address);
2461 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002462 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002463
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002464 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002465
2466 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002467 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002468
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002469 mmun_start = address;
2470 mmun_end = address + HPAGE_PMD_SIZE;
2471 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002472 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002473 /*
2474 * After this gup_fast can't run anymore. This also removes
2475 * any huge TLB entry from the CPU so we won't allow
2476 * huge and small TLB entries for the same virtual address
2477 * to avoid the risk of CPU bugs in that area.
2478 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002479 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002480 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002481 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002482
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002483 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002484 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002485 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002486
2487 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002488 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002489 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002490 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002491 /*
2492 * We can only use set_pmd_at when establishing
2493 * hugepmds and never for establishing regular pmds that
2494 * points to regular pagetables. Use pmd_populate for that
2495 */
2496 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002497 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002498 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002499 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002500 }
2501
2502 /*
2503 * All pages are isolated and locked so anon_vma rmap
2504 * can't run anymore.
2505 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002506 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002507
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002508 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002509 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002510 __SetPageUptodate(new_page);
2511 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002512
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002513 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2514 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002515
2516 /*
2517 * spin_lock() below is not the equivalent of smp_wmb(), so
2518 * this is needed to avoid the copy_huge_page writes to become
2519 * visible after the set_pmd_at() write.
2520 */
2521 smp_wmb();
2522
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002523 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002524 BUG_ON(!pmd_none(*pmd));
2525 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002526 mem_cgroup_commit_charge(new_page, memcg, false);
2527 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002528 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002529 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002530 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002531 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002532
2533 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002534
Andrea Arcangeliba761492011-01-13 15:46:58 -08002535 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002536out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002537 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002538 return;
2539
Andrea Arcangelice83d212011-01-13 15:47:06 -08002540out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002541 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002542 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002543}
2544
2545static int khugepaged_scan_pmd(struct mm_struct *mm,
2546 struct vm_area_struct *vma,
2547 unsigned long address,
2548 struct page **hpage)
2549{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002550 pmd_t *pmd;
2551 pte_t *pte, *_pte;
2552 int ret = 0, referenced = 0, none = 0;
2553 struct page *page;
2554 unsigned long _address;
2555 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002556 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002557
2558 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2559
Bob Liu62190492012-12-11 16:00:37 -08002560 pmd = mm_find_pmd(mm, address);
2561 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002562 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002563
Bob Liu9f1b8682013-11-12 15:07:37 -08002564 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002565 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2566 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2567 _pte++, _address += PAGE_SIZE) {
2568 pte_t pteval = *_pte;
2569 if (pte_none(pteval)) {
2570 if (++none <= khugepaged_max_ptes_none)
2571 continue;
2572 else
2573 goto out_unmap;
2574 }
2575 if (!pte_present(pteval) || !pte_write(pteval))
2576 goto out_unmap;
2577 page = vm_normal_page(vma, _address, pteval);
2578 if (unlikely(!page))
2579 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002580 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002581 * Record which node the original page is from and save this
2582 * information to khugepaged_node_load[].
2583 * Khupaged will allocate hugepage from the node has the max
2584 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002585 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002586 node = page_to_nid(page);
David Rientjes14a4e212014-08-06 16:07:29 -07002587 if (khugepaged_scan_abort(node))
2588 goto out_unmap;
Bob Liu9f1b8682013-11-12 15:07:37 -08002589 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002590 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002591 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2592 goto out_unmap;
2593 /* cannot use mapcount: can't collapse if there's a gup pin */
2594 if (page_count(page) != 1)
2595 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002596 if (pte_young(pteval) || PageReferenced(page) ||
2597 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002598 referenced = 1;
2599 }
2600 if (referenced)
2601 ret = 1;
2602out_unmap:
2603 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002604 if (ret) {
2605 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002606 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002607 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002608 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002609out:
2610 return ret;
2611}
2612
2613static void collect_mm_slot(struct mm_slot *mm_slot)
2614{
2615 struct mm_struct *mm = mm_slot->mm;
2616
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002617 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002618
2619 if (khugepaged_test_exit(mm)) {
2620 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002621 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002622 list_del(&mm_slot->mm_node);
2623
2624 /*
2625 * Not strictly needed because the mm exited already.
2626 *
2627 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2628 */
2629
2630 /* khugepaged_mm_lock actually not necessary for the below */
2631 free_mm_slot(mm_slot);
2632 mmdrop(mm);
2633 }
2634}
2635
2636static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2637 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002638 __releases(&khugepaged_mm_lock)
2639 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002640{
2641 struct mm_slot *mm_slot;
2642 struct mm_struct *mm;
2643 struct vm_area_struct *vma;
2644 int progress = 0;
2645
2646 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002647 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002648
2649 if (khugepaged_scan.mm_slot)
2650 mm_slot = khugepaged_scan.mm_slot;
2651 else {
2652 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2653 struct mm_slot, mm_node);
2654 khugepaged_scan.address = 0;
2655 khugepaged_scan.mm_slot = mm_slot;
2656 }
2657 spin_unlock(&khugepaged_mm_lock);
2658
2659 mm = mm_slot->mm;
2660 down_read(&mm->mmap_sem);
2661 if (unlikely(khugepaged_test_exit(mm)))
2662 vma = NULL;
2663 else
2664 vma = find_vma(mm, khugepaged_scan.address);
2665
2666 progress++;
2667 for (; vma; vma = vma->vm_next) {
2668 unsigned long hstart, hend;
2669
2670 cond_resched();
2671 if (unlikely(khugepaged_test_exit(mm))) {
2672 progress++;
2673 break;
2674 }
Bob Liufa475e52012-12-11 16:00:39 -08002675 if (!hugepage_vma_check(vma)) {
2676skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002677 progress++;
2678 continue;
2679 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002680 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2681 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002682 if (hstart >= hend)
2683 goto skip;
2684 if (khugepaged_scan.address > hend)
2685 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002686 if (khugepaged_scan.address < hstart)
2687 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002688 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002689
2690 while (khugepaged_scan.address < hend) {
2691 int ret;
2692 cond_resched();
2693 if (unlikely(khugepaged_test_exit(mm)))
2694 goto breakouterloop;
2695
2696 VM_BUG_ON(khugepaged_scan.address < hstart ||
2697 khugepaged_scan.address + HPAGE_PMD_SIZE >
2698 hend);
2699 ret = khugepaged_scan_pmd(mm, vma,
2700 khugepaged_scan.address,
2701 hpage);
2702 /* move to next address */
2703 khugepaged_scan.address += HPAGE_PMD_SIZE;
2704 progress += HPAGE_PMD_NR;
2705 if (ret)
2706 /* we released mmap_sem so break loop */
2707 goto breakouterloop_mmap_sem;
2708 if (progress >= pages)
2709 goto breakouterloop;
2710 }
2711 }
2712breakouterloop:
2713 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2714breakouterloop_mmap_sem:
2715
2716 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002717 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002718 /*
2719 * Release the current mm_slot if this mm is about to die, or
2720 * if we scanned all vmas of this mm.
2721 */
2722 if (khugepaged_test_exit(mm) || !vma) {
2723 /*
2724 * Make sure that if mm_users is reaching zero while
2725 * khugepaged runs here, khugepaged_exit will find
2726 * mm_slot not pointing to the exiting mm.
2727 */
2728 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2729 khugepaged_scan.mm_slot = list_entry(
2730 mm_slot->mm_node.next,
2731 struct mm_slot, mm_node);
2732 khugepaged_scan.address = 0;
2733 } else {
2734 khugepaged_scan.mm_slot = NULL;
2735 khugepaged_full_scans++;
2736 }
2737
2738 collect_mm_slot(mm_slot);
2739 }
2740
2741 return progress;
2742}
2743
2744static int khugepaged_has_work(void)
2745{
2746 return !list_empty(&khugepaged_scan.mm_head) &&
2747 khugepaged_enabled();
2748}
2749
2750static int khugepaged_wait_event(void)
2751{
2752 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002753 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002754}
2755
Xiao Guangrongd5169042012-10-08 16:29:48 -07002756static void khugepaged_do_scan(void)
2757{
2758 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002759 unsigned int progress = 0, pass_through_head = 0;
2760 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002761 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002762
2763 barrier(); /* write khugepaged_pages_to_scan to local stack */
2764
2765 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002766 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002767 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002768
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002769 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002770
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002771 if (unlikely(kthread_should_stop() || freezing(current)))
2772 break;
2773
Andrea Arcangeliba761492011-01-13 15:46:58 -08002774 spin_lock(&khugepaged_mm_lock);
2775 if (!khugepaged_scan.mm_slot)
2776 pass_through_head++;
2777 if (khugepaged_has_work() &&
2778 pass_through_head < 2)
2779 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002780 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002781 else
2782 progress = pages;
2783 spin_unlock(&khugepaged_mm_lock);
2784 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002785
Xiao Guangrongd5169042012-10-08 16:29:48 -07002786 if (!IS_ERR_OR_NULL(hpage))
2787 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002788}
2789
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002790static void khugepaged_wait_work(void)
2791{
2792 try_to_freeze();
2793
2794 if (khugepaged_has_work()) {
2795 if (!khugepaged_scan_sleep_millisecs)
2796 return;
2797
2798 wait_event_freezable_timeout(khugepaged_wait,
2799 kthread_should_stop(),
2800 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2801 return;
2802 }
2803
2804 if (khugepaged_enabled())
2805 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2806}
2807
Andrea Arcangeliba761492011-01-13 15:46:58 -08002808static int khugepaged(void *none)
2809{
2810 struct mm_slot *mm_slot;
2811
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002812 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002813 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002814
Xiao Guangrongb7231782012-10-08 16:29:54 -07002815 while (!kthread_should_stop()) {
2816 khugepaged_do_scan();
2817 khugepaged_wait_work();
2818 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002819
2820 spin_lock(&khugepaged_mm_lock);
2821 mm_slot = khugepaged_scan.mm_slot;
2822 khugepaged_scan.mm_slot = NULL;
2823 if (mm_slot)
2824 collect_mm_slot(mm_slot);
2825 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002826 return 0;
2827}
2828
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002829static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2830 unsigned long haddr, pmd_t *pmd)
2831{
2832 struct mm_struct *mm = vma->vm_mm;
2833 pgtable_t pgtable;
2834 pmd_t _pmd;
2835 int i;
2836
2837 pmdp_clear_flush(vma, haddr, pmd);
2838 /* leave pmd empty until pte is filled */
2839
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002840 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002841 pmd_populate(mm, &_pmd, pgtable);
2842
2843 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2844 pte_t *pte, entry;
2845 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2846 entry = pte_mkspecial(entry);
2847 pte = pte_offset_map(&_pmd, haddr);
2848 VM_BUG_ON(!pte_none(*pte));
2849 set_pte_at(mm, haddr, pte, entry);
2850 pte_unmap(pte);
2851 }
2852 smp_wmb(); /* make pte visible before pmd */
2853 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002854 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002855}
2856
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002857void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2858 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002859{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002860 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002861 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002862 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002863 unsigned long haddr = address & HPAGE_PMD_MASK;
2864 unsigned long mmun_start; /* For mmu_notifiers */
2865 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002866
2867 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002868
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002869 mmun_start = haddr;
2870 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002871again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002872 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002873 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002874 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002875 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002876 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2877 return;
2878 }
2879 if (is_huge_zero_pmd(*pmd)) {
2880 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002881 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002882 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002883 return;
2884 }
2885 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002886 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002887 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002888 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002889 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002890
2891 split_huge_page(page);
2892
2893 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002894
2895 /*
2896 * We don't always have down_write of mmap_sem here: a racing
2897 * do_huge_pmd_wp_page() might have copied-on-write to another
2898 * huge page before our split_huge_page() got the anon_vma lock.
2899 */
2900 if (unlikely(pmd_trans_huge(*pmd)))
2901 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002902}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002903
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002904void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2905 pmd_t *pmd)
2906{
2907 struct vm_area_struct *vma;
2908
2909 vma = find_vma(mm, address);
2910 BUG_ON(vma == NULL);
2911 split_huge_page_pmd(vma, address, pmd);
2912}
2913
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002914static void split_huge_page_address(struct mm_struct *mm,
2915 unsigned long address)
2916{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002917 pgd_t *pgd;
2918 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002919 pmd_t *pmd;
2920
2921 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2922
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002923 pgd = pgd_offset(mm, address);
2924 if (!pgd_present(*pgd))
2925 return;
2926
2927 pud = pud_offset(pgd, address);
2928 if (!pud_present(*pud))
2929 return;
2930
2931 pmd = pmd_offset(pud, address);
2932 if (!pmd_present(*pmd))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002933 return;
2934 /*
2935 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2936 * materialize from under us.
2937 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002938 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002939}
2940
2941void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2942 unsigned long start,
2943 unsigned long end,
2944 long adjust_next)
2945{
2946 /*
2947 * If the new start address isn't hpage aligned and it could
2948 * previously contain an hugepage: check if we need to split
2949 * an huge pmd.
2950 */
2951 if (start & ~HPAGE_PMD_MASK &&
2952 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2953 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2954 split_huge_page_address(vma->vm_mm, start);
2955
2956 /*
2957 * If the new end address isn't hpage aligned and it could
2958 * previously contain an hugepage: check if we need to split
2959 * an huge pmd.
2960 */
2961 if (end & ~HPAGE_PMD_MASK &&
2962 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2963 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2964 split_huge_page_address(vma->vm_mm, end);
2965
2966 /*
2967 * If we're also updating the vma->vm_next->vm_start, if the new
2968 * vm_next->vm_start isn't page aligned and it could previously
2969 * contain an hugepage: check if we need to split an huge pmd.
2970 */
2971 if (adjust_next > 0) {
2972 struct vm_area_struct *next = vma->vm_next;
2973 unsigned long nstart = next->vm_start;
2974 nstart += adjust_next << PAGE_SHIFT;
2975 if (nstart & ~HPAGE_PMD_MASK &&
2976 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2977 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2978 split_huge_page_address(next->vm_mm, nstart);
2979 }
2980}