blob: 817a875f2b8c0338dc9e56511b92d99a7ff7b992 [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_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700788 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800789 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800790 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800791 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800792}
793
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800794int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
795 unsigned long address, pmd_t *pmd,
796 unsigned int flags)
797{
798 struct page *page;
799 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800800
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700801 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700802 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700803 if (unlikely(anon_vma_prepare(vma)))
804 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700805 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700806 return VM_FAULT_OOM;
Dominik Dingel593befa2014-10-23 12:07:44 +0200807 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700808 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800809 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700810 pgtable_t pgtable;
811 struct page *zero_page;
812 bool set;
813 pgtable = pte_alloc_one(mm, haddr);
814 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800815 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700816 zero_page = get_huge_zero_page();
817 if (unlikely(!zero_page)) {
818 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700819 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700820 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700821 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800822 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700823 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
824 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800825 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700826 if (!set) {
827 pte_free(mm, pgtable);
828 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800829 }
David Rientjesedad9d22012-05-29 15:06:17 -0700830 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800831 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700832 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
833 vma, haddr, numa_node_id(), 0);
834 if (unlikely(!page)) {
835 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700836 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700837 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700838 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700839 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700840 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700841 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700842 }
843
David Rientjes17766dd2013-09-12 15:14:06 -0700844 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700845 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800846}
847
848int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
849 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
850 struct vm_area_struct *vma)
851{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800852 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800853 struct page *src_page;
854 pmd_t pmd;
855 pgtable_t pgtable;
856 int ret;
857
858 ret = -ENOMEM;
859 pgtable = pte_alloc_one(dst_mm, addr);
860 if (unlikely(!pgtable))
861 goto out;
862
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800863 dst_ptl = pmd_lock(dst_mm, dst_pmd);
864 src_ptl = pmd_lockptr(src_mm, src_pmd);
865 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800866
867 ret = -EAGAIN;
868 pmd = *src_pmd;
869 if (unlikely(!pmd_trans_huge(pmd))) {
870 pte_free(dst_mm, pgtable);
871 goto out_unlock;
872 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800873 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800874 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800875 * under splitting since we don't split the page itself, only pmd to
876 * a page table.
877 */
878 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700879 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800880 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800881 /*
882 * get_huge_zero_page() will never allocate a new page here,
883 * since we already have a zero page to copy. It just takes a
884 * reference.
885 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700886 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800887 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700888 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800889 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800890 ret = 0;
891 goto out_unlock;
892 }
Mel Gormande466bd2013-12-18 17:08:42 -0800893
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800894 if (unlikely(pmd_trans_splitting(pmd))) {
895 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800896 spin_unlock(src_ptl);
897 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800898 pte_free(dst_mm, pgtable);
899
900 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
901 goto out;
902 }
903 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800904 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800905 get_page(src_page);
906 page_dup_rmap(src_page);
907 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
908
909 pmdp_set_wrprotect(src_mm, addr, src_pmd);
910 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700911 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800912 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800913 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800914
915 ret = 0;
916out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800917 spin_unlock(src_ptl);
918 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800919out:
920 return ret;
921}
922
Will Deacona1dd4502012-12-11 16:01:27 -0800923void huge_pmd_set_accessed(struct mm_struct *mm,
924 struct vm_area_struct *vma,
925 unsigned long address,
926 pmd_t *pmd, pmd_t orig_pmd,
927 int dirty)
928{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800929 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800930 pmd_t entry;
931 unsigned long haddr;
932
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800933 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800934 if (unlikely(!pmd_same(*pmd, orig_pmd)))
935 goto unlock;
936
937 entry = pmd_mkyoung(orig_pmd);
938 haddr = address & HPAGE_PMD_MASK;
939 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
940 update_mmu_cache_pmd(vma, address, pmd);
941
942unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800943 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800944}
945
Hugh Dickins5338a932014-06-23 13:22:05 -0700946/*
947 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
948 * during copy_user_huge_page()'s copy_page_rep(): in the case when
949 * the source page gets split and a tail freed before copy completes.
950 * Called under pmd_lock of checked pmd, so safe from splitting itself.
951 */
952static void get_user_huge_page(struct page *page)
953{
954 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
955 struct page *endpage = page + HPAGE_PMD_NR;
956
957 atomic_add(HPAGE_PMD_NR, &page->_count);
958 while (++page < endpage)
959 get_huge_page_tail(page);
960 } else {
961 get_page(page);
962 }
963}
964
965static void put_user_huge_page(struct page *page)
966{
967 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
968 struct page *endpage = page + HPAGE_PMD_NR;
969
970 while (page < endpage)
971 put_page(page++);
972 } else {
973 put_page(page);
974 }
975}
976
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800977static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
978 struct vm_area_struct *vma,
979 unsigned long address,
980 pmd_t *pmd, pmd_t orig_pmd,
981 struct page *page,
982 unsigned long haddr)
983{
Johannes Weiner00501b52014-08-08 14:19:20 -0700984 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800985 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800986 pgtable_t pgtable;
987 pmd_t _pmd;
988 int ret = 0, i;
989 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700990 unsigned long mmun_start; /* For mmu_notifiers */
991 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800992
993 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
994 GFP_KERNEL);
995 if (unlikely(!pages)) {
996 ret |= VM_FAULT_OOM;
997 goto out;
998 }
999
1000 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001001 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1002 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001003 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001004 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -07001005 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1006 &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001007 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001008 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001009 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001010 memcg = (void *)page_private(pages[i]);
1011 set_page_private(pages[i], 0);
1012 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001013 put_page(pages[i]);
1014 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001015 kfree(pages);
1016 ret |= VM_FAULT_OOM;
1017 goto out;
1018 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001019 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001020 }
1021
1022 for (i = 0; i < HPAGE_PMD_NR; i++) {
1023 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001024 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001025 __SetPageUptodate(pages[i]);
1026 cond_resched();
1027 }
1028
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001029 mmun_start = haddr;
1030 mmun_end = haddr + HPAGE_PMD_SIZE;
1031 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1032
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001033 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001034 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1035 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001036 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001037
Joerg Roedel34ee6452014-11-13 13:46:09 +11001038 pmdp_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001039 /* leave pmd empty until pte is filled */
1040
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001041 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001042 pmd_populate(mm, &_pmd, pgtable);
1043
1044 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1045 pte_t *pte, entry;
1046 entry = mk_pte(pages[i], vma->vm_page_prot);
1047 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001048 memcg = (void *)page_private(pages[i]);
1049 set_page_private(pages[i], 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001050 page_add_new_anon_rmap(pages[i], vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001051 mem_cgroup_commit_charge(pages[i], memcg, false);
1052 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001053 pte = pte_offset_map(&_pmd, haddr);
1054 VM_BUG_ON(!pte_none(*pte));
1055 set_pte_at(mm, haddr, pte, entry);
1056 pte_unmap(pte);
1057 }
1058 kfree(pages);
1059
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001060 smp_wmb(); /* make pte visible before pmd */
1061 pmd_populate(mm, pmd, pgtable);
1062 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001063 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001064
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001065 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1066
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001067 ret |= VM_FAULT_WRITE;
1068 put_page(page);
1069
1070out:
1071 return ret;
1072
1073out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001074 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001075 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001076 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001077 memcg = (void *)page_private(pages[i]);
1078 set_page_private(pages[i], 0);
1079 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001081 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001082 kfree(pages);
1083 goto out;
1084}
1085
1086int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1087 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1088{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001089 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001090 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001091 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001092 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001093 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001094 unsigned long mmun_start; /* For mmu_notifiers */
1095 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001096
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001097 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001098 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001099 haddr = address & HPAGE_PMD_MASK;
1100 if (is_huge_zero_pmd(orig_pmd))
1101 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001102 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001103 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1104 goto out_unlock;
1105
1106 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001107 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001108 if (page_mapcount(page) == 1) {
1109 pmd_t entry;
1110 entry = pmd_mkyoung(orig_pmd);
1111 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001113 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001114 ret |= VM_FAULT_WRITE;
1115 goto out_unlock;
1116 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001117 get_user_huge_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001118 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001119alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001120 if (transparent_hugepage_enabled(vma) &&
1121 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001122 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001123 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001124 else
1125 new_page = NULL;
1126
1127 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001128 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001129 split_huge_page_pmd(vma, address, pmd);
1130 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001131 } else {
1132 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1133 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001134 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001135 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001136 ret |= VM_FAULT_FALLBACK;
1137 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001138 put_user_huge_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001139 }
David Rientjes17766dd2013-09-12 15:14:06 -07001140 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001141 goto out;
1142 }
1143
Johannes Weiner00501b52014-08-08 14:19:20 -07001144 if (unlikely(mem_cgroup_try_charge(new_page, mm,
1145 GFP_TRANSHUGE, &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001146 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001147 if (page) {
1148 split_huge_page(page);
Hugh Dickins5338a932014-06-23 13:22:05 -07001149 put_user_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001150 } else
1151 split_huge_page_pmd(vma, address, pmd);
1152 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001153 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001154 goto out;
1155 }
1156
David Rientjes17766dd2013-09-12 15:14:06 -07001157 count_vm_event(THP_FAULT_ALLOC);
1158
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001159 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001160 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1161 else
1162 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001163 __SetPageUptodate(new_page);
1164
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001165 mmun_start = haddr;
1166 mmun_end = haddr + HPAGE_PMD_SIZE;
1167 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1168
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001169 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001170 if (page)
Hugh Dickins5338a932014-06-23 13:22:05 -07001171 put_user_huge_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001172 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001173 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -07001174 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001175 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001176 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001177 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001178 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001179 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1180 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001181 pmdp_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001182 page_add_new_anon_rmap(new_page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001183 mem_cgroup_commit_charge(new_page, memcg, false);
1184 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001185 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001186 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001187 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001188 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001189 put_huge_zero_page();
1190 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001191 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001192 page_remove_rmap(page);
1193 put_page(page);
1194 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001195 ret |= VM_FAULT_WRITE;
1196 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001197 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001198out_mn:
1199 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1200out:
1201 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001202out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001203 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001204 return ret;
1205}
1206
David Rientjesb676b292012-10-08 16:34:03 -07001207struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001208 unsigned long addr,
1209 pmd_t *pmd,
1210 unsigned int flags)
1211{
David Rientjesb676b292012-10-08 16:34:03 -07001212 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001213 struct page *page = NULL;
1214
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001215 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001216
1217 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1218 goto out;
1219
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001220 /* Avoid dumping huge zero page */
1221 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1222 return ERR_PTR(-EFAULT);
1223
Mel Gorman2b4847e2013-12-18 17:08:32 -08001224 /* Full NUMA hinting faults to serialise migration in fault paths */
1225 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1226 goto out;
1227
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001228 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001229 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 if (flags & FOLL_TOUCH) {
1231 pmd_t _pmd;
1232 /*
1233 * We should set the dirty bit only for FOLL_WRITE but
1234 * for now the dirty bit in the pmd is meaningless.
1235 * And if the dirty bit will become meaningful and
1236 * we'll only set it with FOLL_WRITE, an atomic
1237 * set_bit will be required on the pmd to set the
1238 * young bit, instead of the current set_pmd_at.
1239 */
1240 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001241 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1242 pmd, _pmd, 1))
1243 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001244 }
David Rientjesb676b292012-10-08 16:34:03 -07001245 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1246 if (page->mapping && trylock_page(page)) {
1247 lru_add_drain();
1248 if (page->mapping)
1249 mlock_vma_page(page);
1250 unlock_page(page);
1251 }
1252 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001253 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001254 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001255 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001256 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001257
1258out:
1259 return page;
1260}
1261
Mel Gormand10e63f2012-10-25 14:16:31 +02001262/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001263int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1264 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001265{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001266 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001267 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001268 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001269 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001270 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001271 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001272 bool page_locked;
1273 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001274 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001275
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001276 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001277 if (unlikely(!pmd_same(pmd, *pmdp)))
1278 goto out_unlock;
1279
Mel Gormande466bd2013-12-18 17:08:42 -08001280 /*
1281 * If there are potential migrations, wait for completion and retry
1282 * without disrupting NUMA hinting information. Do not relock and
1283 * check_same as the page may no longer be mapped.
1284 */
1285 if (unlikely(pmd_trans_migrating(*pmdp))) {
1286 spin_unlock(ptl);
1287 wait_migrate_huge_page(vma->anon_vma, pmdp);
1288 goto out;
1289 }
1290
Mel Gormand10e63f2012-10-25 14:16:31 +02001291 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001292 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001293 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001294 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001295 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001296 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001297 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001298 flags |= TNF_FAULT_LOCAL;
1299 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001300
Mel Gormanff9042b2013-10-07 11:28:43 +01001301 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001302 * Avoid grouping on DSO/COW pages in specific and RO pages
1303 * in general, RO pages shouldn't hurt as much anyway since
1304 * they can be in shared cache state.
1305 */
1306 if (!pmd_write(pmd))
1307 flags |= TNF_NO_GROUP;
1308
1309 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001310 * Acquire the page lock to serialise THP migrations but avoid dropping
1311 * page_table_lock if at all possible
1312 */
Mel Gormanb8916632013-10-07 11:28:44 +01001313 page_locked = trylock_page(page);
1314 target_nid = mpol_misplaced(page, vma, haddr);
1315 if (target_nid == -1) {
1316 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001317 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001318 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001319 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001320
Mel Gormande466bd2013-12-18 17:08:42 -08001321 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001322 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001323 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001324 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001325 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001326 goto out;
1327 }
1328
Mel Gorman2b4847e2013-12-18 17:08:32 -08001329 /*
1330 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1331 * to serialises splits
1332 */
Mel Gormanb8916632013-10-07 11:28:44 +01001333 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001334 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001335 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001336
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001337 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001338 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001339 if (unlikely(!pmd_same(pmd, *pmdp))) {
1340 unlock_page(page);
1341 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001342 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001343 goto out_unlock;
1344 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001345
Mel Gormanc3a489c2013-12-18 17:08:38 -08001346 /* Bail if we fail to protect against THP splits for any reason */
1347 if (unlikely(!anon_vma)) {
1348 put_page(page);
1349 page_nid = -1;
1350 goto clear_pmdnuma;
1351 }
1352
Mel Gormana54a4072013-10-07 11:28:46 +01001353 /*
1354 * Migrate the THP to the requested node, returns with page unlocked
1355 * and pmd_numa cleared.
1356 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001357 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001358 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001359 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001360 if (migrated) {
1361 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001362 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001363 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001364
Mel Gorman8191acb2013-10-07 11:28:45 +01001365 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001366clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001367 BUG_ON(!PageLocked(page));
Mel Gormand10e63f2012-10-25 14:16:31 +02001368 pmd = pmd_mknonnuma(pmd);
1369 set_pmd_at(mm, haddr, pmdp, pmd);
1370 VM_BUG_ON(pmd_numa(*pmdp));
1371 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001372 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001373out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001374 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001375
1376out:
1377 if (anon_vma)
1378 page_unlock_anon_vma_read(anon_vma);
1379
Mel Gorman8191acb2013-10-07 11:28:45 +01001380 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001381 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001382
Mel Gormand10e63f2012-10-25 14:16:31 +02001383 return 0;
1384}
1385
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001386int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001387 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001388{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001389 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001390 int ret = 0;
1391
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001392 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001393 struct page *page;
1394 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001395 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001396 /*
1397 * For architectures like ppc64 we look at deposited pgtable
1398 * when calling pmdp_get_and_clear. So do the
1399 * pgtable_trans_huge_withdraw after finishing pmdp related
1400 * operations.
1401 */
Martin Schwidefskyfcbe08d62014-10-24 10:52:29 +02001402 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1403 tlb->fullmm);
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) {
Joerg Roedel34ee6452014-11-13 13:46:09 +11001515 entry = pmdp_get_and_clear_notify(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);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001647
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001648 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001649 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001650 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001651 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001652
1653 return ret;
1654}
1655
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001656static void __split_huge_page_refcount(struct page *page,
1657 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001658{
1659 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001660 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001661 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001662 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001663
1664 /* prevent PageLRU to go away from under us, and freeze lru stats */
1665 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001666 lruvec = mem_cgroup_page_lruvec(page, zone);
1667
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001668 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001669 /* complete memcg works before add pages to LRU */
1670 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001671
Shaohua Li45676882012-01-12 17:19:18 -08001672 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001673 struct page *page_tail = page + i;
1674
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001675 /* tail_page->_mapcount cannot change */
1676 BUG_ON(page_mapcount(page_tail) < 0);
1677 tail_count += page_mapcount(page_tail);
1678 /* check for overflow */
1679 BUG_ON(tail_count < 0);
1680 BUG_ON(atomic_read(&page_tail->_count) != 0);
1681 /*
1682 * tail_page->_count is zero and not changing from
1683 * under us. But get_page_unless_zero() may be running
1684 * from under us on the tail_page. If we used
1685 * atomic_set() below instead of atomic_add(), we
1686 * would then run atomic_set() concurrently with
1687 * get_page_unless_zero(), and atomic_set() is
1688 * implemented in C not using locked ops. spin_unlock
1689 * on x86 sometime uses locked ops because of PPro
1690 * errata 66, 92, so unless somebody can guarantee
1691 * atomic_set() here would be safe on all archs (and
1692 * not only on x86), it's safer to use atomic_add().
1693 */
1694 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1695 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001696
1697 /* after clearing PageTail the gup refcount can be released */
Waiman Long3a79d522014-08-06 16:05:38 -07001698 smp_mb__after_atomic();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001699
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001700 /*
1701 * retain hwpoison flag of the poisoned tail page:
1702 * fix for the unsuitable process killed on Guest Machine(KVM)
1703 * by the memory-failure.
1704 */
1705 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001706 page_tail->flags |= (page->flags &
1707 ((1L << PG_referenced) |
1708 (1L << PG_swapbacked) |
1709 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001710 (1L << PG_uptodate) |
1711 (1L << PG_active) |
1712 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001713 page_tail->flags |= (1L << PG_dirty);
1714
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001715 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001716 smp_wmb();
1717
1718 /*
1719 * __split_huge_page_splitting() already set the
1720 * splitting bit in all pmd that could map this
1721 * hugepage, that will ensure no CPU can alter the
1722 * mapcount on the head page. The mapcount is only
1723 * accounted in the head page and it has to be
1724 * transferred to all tail pages in the below code. So
1725 * for this code to be safe, the split the mapcount
1726 * can't change. But that doesn't mean userland can't
1727 * keep changing and reading the page contents while
1728 * we transfer the mapcount, so the pmd splitting
1729 * status is achieved setting a reserved bit in the
1730 * pmd, not by clearing the present bit.
1731 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001732 page_tail->_mapcount = page->_mapcount;
1733
1734 BUG_ON(page_tail->mapping);
1735 page_tail->mapping = page->mapping;
1736
Shaohua Li45676882012-01-12 17:19:18 -08001737 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001738 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001739
1740 BUG_ON(!PageAnon(page_tail));
1741 BUG_ON(!PageUptodate(page_tail));
1742 BUG_ON(!PageDirty(page_tail));
1743 BUG_ON(!PageSwapBacked(page_tail));
1744
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001745 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001746 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001747 atomic_sub(tail_count, &page->_count);
1748 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001749
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001750 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001751
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001752 ClearPageCompound(page);
1753 compound_unlock(page);
1754 spin_unlock_irq(&zone->lru_lock);
1755
1756 for (i = 1; i < HPAGE_PMD_NR; i++) {
1757 struct page *page_tail = page + i;
1758 BUG_ON(page_count(page_tail) <= 0);
1759 /*
1760 * Tail pages may be freed if there wasn't any mapping
1761 * like if add_to_swap() is running on a lru page that
1762 * had its mapping zapped. And freeing these pages
1763 * requires taking the lru_lock so we do the put_page
1764 * of the tail pages after the split is complete.
1765 */
1766 put_page(page_tail);
1767 }
1768
1769 /*
1770 * Only the head page (now become a regular page) is required
1771 * to be pinned by the caller.
1772 */
1773 BUG_ON(page_count(page) <= 0);
1774}
1775
1776static int __split_huge_page_map(struct page *page,
1777 struct vm_area_struct *vma,
1778 unsigned long address)
1779{
1780 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001781 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001782 pmd_t *pmd, _pmd;
1783 int ret = 0, i;
1784 pgtable_t pgtable;
1785 unsigned long haddr;
1786
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001787 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001788 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001789 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001790 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001791 pmd_populate(mm, &_pmd, pgtable);
Waiman Longf8303c22014-08-06 16:05:36 -07001792 if (pmd_write(*pmd))
1793 BUG_ON(page_mapcount(page) != 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001794
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001795 haddr = address;
1796 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001797 pte_t *pte, entry;
1798 BUG_ON(PageCompound(page+i));
Mel Gormanabc40bd2014-10-02 19:47:42 +01001799 /*
1800 * Note that pmd_numa is not transferred deliberately
1801 * to avoid any possibility that pte_numa leaks to
1802 * a PROT_NONE VMA by accident.
1803 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001804 entry = mk_pte(page + i, vma->vm_page_prot);
1805 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1806 if (!pmd_write(*pmd))
1807 entry = pte_wrprotect(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001808 if (!pmd_young(*pmd))
1809 entry = pte_mkold(entry);
1810 pte = pte_offset_map(&_pmd, haddr);
1811 BUG_ON(!pte_none(*pte));
1812 set_pte_at(mm, haddr, pte, entry);
1813 pte_unmap(pte);
1814 }
1815
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001816 smp_wmb(); /* make pte visible before pmd */
1817 /*
1818 * Up to this point the pmd is present and huge and
1819 * userland has the whole access to the hugepage
1820 * during the split (which happens in place). If we
1821 * overwrite the pmd with the not-huge version
1822 * pointing to the pte here (which of course we could
1823 * if all CPUs were bug free), userland could trigger
1824 * a small page size TLB miss on the small sized TLB
1825 * while the hugepage TLB entry is still established
1826 * in the huge TLB. Some CPU doesn't like that. See
1827 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1828 * Erratum 383 on page 93. Intel should be safe but is
1829 * also warns that it's only safe if the permission
1830 * and cache attributes of the two entries loaded in
1831 * the two TLB is identical (which should be the case
1832 * here). But it is generally safer to never allow
1833 * small and huge TLB entries for the same virtual
1834 * address to be loaded simultaneously. So instead of
1835 * doing "pmd_populate(); flush_tlb_range();" we first
1836 * mark the current pmd notpresent (atomically because
1837 * here the pmd_trans_huge and pmd_trans_splitting
1838 * must remain set at all times on the pmd until the
1839 * split is complete for this pmd), then we flush the
1840 * SMP TLB and finally we write the non-huge version
1841 * of the pmd entry with pmd_populate.
1842 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001843 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001844 pmd_populate(mm, pmd, pgtable);
1845 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001846 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001847 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001848
1849 return ret;
1850}
1851
Ingo Molnar5a505082012-12-02 19:56:46 +00001852/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001853static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001854 struct anon_vma *anon_vma,
1855 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001856{
1857 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001858 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001859 struct anon_vma_chain *avc;
1860
1861 BUG_ON(!PageHead(page));
1862 BUG_ON(PageTail(page));
1863
1864 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001865 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001866 struct vm_area_struct *vma = avc->vma;
1867 unsigned long addr = vma_address(page, vma);
1868 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001869 mapcount += __split_huge_page_splitting(page, vma, addr);
1870 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001871 /*
1872 * It is critical that new vmas are added to the tail of the
1873 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1874 * and establishes a child pmd before
1875 * __split_huge_page_splitting() freezes the parent pmd (so if
1876 * we fail to prevent copy_huge_pmd() from running until the
1877 * whole __split_huge_page() is complete), we will still see
1878 * the newly established pmd of the child later during the
1879 * walk, to be able to set it as pmd_trans_splitting too.
1880 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001881 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001882 pr_err("mapcount %d page_mapcount %d\n",
1883 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001884 BUG();
1885 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001886
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001887 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001888
1889 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001890 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001891 struct vm_area_struct *vma = avc->vma;
1892 unsigned long addr = vma_address(page, vma);
1893 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001894 mapcount2 += __split_huge_page_map(page, vma, addr);
1895 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001896 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001897 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1898 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001899 BUG();
1900 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001901}
1902
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001903/*
1904 * Split a hugepage into normal pages. This doesn't change the position of head
1905 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1906 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1907 * from the hugepage.
1908 * Return 0 if the hugepage is split successfully otherwise return 1.
1909 */
1910int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001911{
1912 struct anon_vma *anon_vma;
1913 int ret = 1;
1914
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001915 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001916 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001917
1918 /*
1919 * The caller does not necessarily hold an mmap_sem that would prevent
1920 * the anon_vma disappearing so we first we take a reference to it
1921 * and then lock the anon_vma for write. This is similar to
1922 * page_lock_anon_vma_read except the write lock is taken to serialise
1923 * against parallel split or collapse operations.
1924 */
1925 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001926 if (!anon_vma)
1927 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001928 anon_vma_lock_write(anon_vma);
1929
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001930 ret = 0;
1931 if (!PageCompound(page))
1932 goto out_unlock;
1933
1934 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001935 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001936 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001937
1938 BUG_ON(PageCompound(page));
1939out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001940 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001941 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001942out:
1943 return ret;
1944}
1945
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001946#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001947
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001948int hugepage_madvise(struct vm_area_struct *vma,
1949 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001950{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001951 switch (advice) {
1952 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001953#ifdef CONFIG_S390
1954 /*
1955 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1956 * can't handle this properly after s390_enable_sie, so we simply
1957 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1958 */
1959 if (mm_has_pgste(vma->vm_mm))
1960 return 0;
1961#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001962 /*
1963 * Be somewhat over-protective like KSM for now!
1964 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001965 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001966 return -EINVAL;
1967 *vm_flags &= ~VM_NOHUGEPAGE;
1968 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001969 /*
1970 * If the vma become good for khugepaged to scan,
1971 * register it here without waiting a page fault that
1972 * may not happen any time soon.
1973 */
David Rientjes6d50e602014-10-29 14:50:31 -07001974 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001975 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001976 break;
1977 case MADV_NOHUGEPAGE:
1978 /*
1979 * Be somewhat over-protective like KSM for now!
1980 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001981 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001982 return -EINVAL;
1983 *vm_flags &= ~VM_HUGEPAGE;
1984 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001985 /*
1986 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1987 * this vma even if we leave the mm registered in khugepaged if
1988 * it got registered before VM_NOHUGEPAGE was set.
1989 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001990 break;
1991 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001992
1993 return 0;
1994}
1995
Andrea Arcangeliba761492011-01-13 15:46:58 -08001996static int __init khugepaged_slab_init(void)
1997{
1998 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1999 sizeof(struct mm_slot),
2000 __alignof__(struct mm_slot), 0, NULL);
2001 if (!mm_slot_cache)
2002 return -ENOMEM;
2003
2004 return 0;
2005}
2006
Andrea Arcangeliba761492011-01-13 15:46:58 -08002007static inline struct mm_slot *alloc_mm_slot(void)
2008{
2009 if (!mm_slot_cache) /* initialization failed */
2010 return NULL;
2011 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2012}
2013
2014static inline void free_mm_slot(struct mm_slot *mm_slot)
2015{
2016 kmem_cache_free(mm_slot_cache, mm_slot);
2017}
2018
Andrea Arcangeliba761492011-01-13 15:46:58 -08002019static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2020{
2021 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002022
Sasha Levinb67bfe02013-02-27 17:06:00 -08002023 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002024 if (mm == mm_slot->mm)
2025 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002026
Andrea Arcangeliba761492011-01-13 15:46:58 -08002027 return NULL;
2028}
2029
2030static void insert_to_mm_slots_hash(struct mm_struct *mm,
2031 struct mm_slot *mm_slot)
2032{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002033 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002034 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002035}
2036
2037static inline int khugepaged_test_exit(struct mm_struct *mm)
2038{
2039 return atomic_read(&mm->mm_users) == 0;
2040}
2041
2042int __khugepaged_enter(struct mm_struct *mm)
2043{
2044 struct mm_slot *mm_slot;
2045 int wakeup;
2046
2047 mm_slot = alloc_mm_slot();
2048 if (!mm_slot)
2049 return -ENOMEM;
2050
2051 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07002052 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002053 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2054 free_mm_slot(mm_slot);
2055 return 0;
2056 }
2057
2058 spin_lock(&khugepaged_mm_lock);
2059 insert_to_mm_slots_hash(mm, mm_slot);
2060 /*
2061 * Insert just behind the scanning cursor, to let the area settle
2062 * down a little.
2063 */
2064 wakeup = list_empty(&khugepaged_scan.mm_head);
2065 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2066 spin_unlock(&khugepaged_mm_lock);
2067
2068 atomic_inc(&mm->mm_count);
2069 if (wakeup)
2070 wake_up_interruptible(&khugepaged_wait);
2071
2072 return 0;
2073}
2074
David Rientjes6d50e602014-10-29 14:50:31 -07002075int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2076 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077{
2078 unsigned long hstart, hend;
2079 if (!vma->anon_vma)
2080 /*
2081 * Not yet faulted in so we will register later in the
2082 * page fault if needed.
2083 */
2084 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002085 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002086 /* khugepaged not yet working on file or special mappings */
2087 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07002088 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002089 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2090 hend = vma->vm_end & HPAGE_PMD_MASK;
2091 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07002092 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002093 return 0;
2094}
2095
2096void __khugepaged_exit(struct mm_struct *mm)
2097{
2098 struct mm_slot *mm_slot;
2099 int free = 0;
2100
2101 spin_lock(&khugepaged_mm_lock);
2102 mm_slot = get_mm_slot(mm);
2103 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002104 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002105 list_del(&mm_slot->mm_node);
2106 free = 1;
2107 }
Chris Wrightd788e802011-07-25 17:12:14 -07002108 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002109
2110 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002111 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2112 free_mm_slot(mm_slot);
2113 mmdrop(mm);
2114 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002115 /*
2116 * This is required to serialize against
2117 * khugepaged_test_exit() (which is guaranteed to run
2118 * under mmap sem read mode). Stop here (after we
2119 * return all pagetables will be destroyed) until
2120 * khugepaged has finished working on the pagetables
2121 * under the mmap_sem.
2122 */
2123 down_write(&mm->mmap_sem);
2124 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002125 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002126}
2127
2128static void release_pte_page(struct page *page)
2129{
2130 /* 0 stands for page_is_file_cache(page) == false */
2131 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2132 unlock_page(page);
2133 putback_lru_page(page);
2134}
2135
2136static void release_pte_pages(pte_t *pte, pte_t *_pte)
2137{
2138 while (--_pte >= pte) {
2139 pte_t pteval = *_pte;
2140 if (!pte_none(pteval))
2141 release_pte_page(pte_page(pteval));
2142 }
2143}
2144
Andrea Arcangeliba761492011-01-13 15:46:58 -08002145static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2146 unsigned long address,
2147 pte_t *pte)
2148{
2149 struct page *page;
2150 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002151 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002152 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2153 _pte++, address += PAGE_SIZE) {
2154 pte_t pteval = *_pte;
2155 if (pte_none(pteval)) {
2156 if (++none <= khugepaged_max_ptes_none)
2157 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002158 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002159 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002160 }
Bob Liu344aa352012-12-11 16:00:34 -08002161 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002162 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002163 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002164 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002165 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002166
Sasha Levin309381fea2014-01-23 15:52:54 -08002167 VM_BUG_ON_PAGE(PageCompound(page), page);
2168 VM_BUG_ON_PAGE(!PageAnon(page), page);
2169 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002170
2171 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002172 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002173 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002174 /*
2175 * We can do it before isolate_lru_page because the
2176 * page can't be freed from under us. NOTE: PG_lock
2177 * is needed to serialize against split_huge_page
2178 * when invoked from the VM.
2179 */
Bob Liu344aa352012-12-11 16:00:34 -08002180 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002181 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002182 /*
2183 * Isolate the page to avoid collapsing an hugepage
2184 * currently in use by the VM.
2185 */
2186 if (isolate_lru_page(page)) {
2187 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002188 goto out;
2189 }
2190 /* 0 stands for page_is_file_cache(page) == false */
2191 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002192 VM_BUG_ON_PAGE(!PageLocked(page), page);
2193 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002194
2195 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002196 if (pte_young(pteval) || PageReferenced(page) ||
2197 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002198 referenced = 1;
2199 }
Bob Liu344aa352012-12-11 16:00:34 -08002200 if (likely(referenced))
2201 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002202out:
Bob Liu344aa352012-12-11 16:00:34 -08002203 release_pte_pages(pte, _pte);
2204 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002205}
2206
2207static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2208 struct vm_area_struct *vma,
2209 unsigned long address,
2210 spinlock_t *ptl)
2211{
2212 pte_t *_pte;
2213 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2214 pte_t pteval = *_pte;
2215 struct page *src_page;
2216
2217 if (pte_none(pteval)) {
2218 clear_user_highpage(page, address);
2219 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2220 } else {
2221 src_page = pte_page(pteval);
2222 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002223 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002224 release_pte_page(src_page);
2225 /*
2226 * ptl mostly unnecessary, but preempt has to
2227 * be disabled to update the per-cpu stats
2228 * inside page_remove_rmap().
2229 */
2230 spin_lock(ptl);
2231 /*
2232 * paravirt calls inside pte_clear here are
2233 * superfluous.
2234 */
2235 pte_clear(vma->vm_mm, address, _pte);
2236 page_remove_rmap(src_page);
2237 spin_unlock(ptl);
2238 free_page_and_swap_cache(src_page);
2239 }
2240
2241 address += PAGE_SIZE;
2242 page++;
2243 }
2244}
2245
Xiao Guangrong26234f32012-10-08 16:29:51 -07002246static void khugepaged_alloc_sleep(void)
2247{
2248 wait_event_freezable_timeout(khugepaged_wait, false,
2249 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2250}
2251
Bob Liu9f1b8682013-11-12 15:07:37 -08002252static int khugepaged_node_load[MAX_NUMNODES];
2253
David Rientjes14a4e212014-08-06 16:07:29 -07002254static bool khugepaged_scan_abort(int nid)
2255{
2256 int i;
2257
2258 /*
2259 * If zone_reclaim_mode is disabled, then no extra effort is made to
2260 * allocate memory locally.
2261 */
2262 if (!zone_reclaim_mode)
2263 return false;
2264
2265 /* If there is a count for this node already, it must be acceptable */
2266 if (khugepaged_node_load[nid])
2267 return false;
2268
2269 for (i = 0; i < MAX_NUMNODES; i++) {
2270 if (!khugepaged_node_load[i])
2271 continue;
2272 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2273 return true;
2274 }
2275 return false;
2276}
2277
Xiao Guangrong26234f32012-10-08 16:29:51 -07002278#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002279static int khugepaged_find_target_node(void)
2280{
2281 static int last_khugepaged_target_node = NUMA_NO_NODE;
2282 int nid, target_node = 0, max_value = 0;
2283
2284 /* find first node with max normal pages hit */
2285 for (nid = 0; nid < MAX_NUMNODES; nid++)
2286 if (khugepaged_node_load[nid] > max_value) {
2287 max_value = khugepaged_node_load[nid];
2288 target_node = nid;
2289 }
2290
2291 /* do some balance if several nodes have the same hit record */
2292 if (target_node <= last_khugepaged_target_node)
2293 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2294 nid++)
2295 if (max_value == khugepaged_node_load[nid]) {
2296 target_node = nid;
2297 break;
2298 }
2299
2300 last_khugepaged_target_node = target_node;
2301 return target_node;
2302}
2303
Xiao Guangrong26234f32012-10-08 16:29:51 -07002304static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2305{
2306 if (IS_ERR(*hpage)) {
2307 if (!*wait)
2308 return false;
2309
2310 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002311 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002312 khugepaged_alloc_sleep();
2313 } else if (*hpage) {
2314 put_page(*hpage);
2315 *hpage = NULL;
2316 }
2317
2318 return true;
2319}
2320
2321static struct page
2322*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2323 struct vm_area_struct *vma, unsigned long address,
2324 int node)
2325{
Sasha Levin309381fea2014-01-23 15:52:54 -08002326 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002327
Xiao Guangrong26234f32012-10-08 16:29:51 -07002328 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002329 * Before allocating the hugepage, release the mmap_sem read lock.
2330 * The allocation can take potentially a long time if it involves
2331 * sync compaction, and we do not need to hold the mmap_sem during
2332 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002333 */
2334 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002335
2336 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2337 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002338 if (unlikely(!*hpage)) {
2339 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2340 *hpage = ERR_PTR(-ENOMEM);
2341 return NULL;
2342 }
2343
2344 count_vm_event(THP_COLLAPSE_ALLOC);
2345 return *hpage;
2346}
2347#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002348static int khugepaged_find_target_node(void)
2349{
2350 return 0;
2351}
2352
Bob Liu10dc4152013-11-12 15:07:35 -08002353static inline struct page *alloc_hugepage(int defrag)
2354{
2355 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2356 HPAGE_PMD_ORDER);
2357}
2358
Xiao Guangrong26234f32012-10-08 16:29:51 -07002359static struct page *khugepaged_alloc_hugepage(bool *wait)
2360{
2361 struct page *hpage;
2362
2363 do {
2364 hpage = alloc_hugepage(khugepaged_defrag());
2365 if (!hpage) {
2366 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2367 if (!*wait)
2368 return NULL;
2369
2370 *wait = false;
2371 khugepaged_alloc_sleep();
2372 } else
2373 count_vm_event(THP_COLLAPSE_ALLOC);
2374 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2375
2376 return hpage;
2377}
2378
2379static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2380{
2381 if (!*hpage)
2382 *hpage = khugepaged_alloc_hugepage(wait);
2383
2384 if (unlikely(!*hpage))
2385 return false;
2386
2387 return true;
2388}
2389
2390static struct page
2391*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2392 struct vm_area_struct *vma, unsigned long address,
2393 int node)
2394{
2395 up_read(&mm->mmap_sem);
2396 VM_BUG_ON(!*hpage);
2397 return *hpage;
2398}
2399#endif
2400
Bob Liufa475e52012-12-11 16:00:39 -08002401static bool hugepage_vma_check(struct vm_area_struct *vma)
2402{
2403 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2404 (vma->vm_flags & VM_NOHUGEPAGE))
2405 return false;
2406
2407 if (!vma->anon_vma || vma->vm_ops)
2408 return false;
2409 if (is_vma_temporary_stack(vma))
2410 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002411 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002412 return true;
2413}
2414
Andrea Arcangeliba761492011-01-13 15:46:58 -08002415static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002416 unsigned long address,
2417 struct page **hpage,
2418 struct vm_area_struct *vma,
2419 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002420{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002421 pmd_t *pmd, _pmd;
2422 pte_t *pte;
2423 pgtable_t pgtable;
2424 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002425 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002426 int isolated;
2427 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002428 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002429 unsigned long mmun_start; /* For mmu_notifiers */
2430 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002431
2432 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002433
Xiao Guangrong26234f32012-10-08 16:29:51 -07002434 /* release the mmap_sem read lock. */
2435 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2436 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002437 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002438
Johannes Weiner00501b52014-08-08 14:19:20 -07002439 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2440 GFP_TRANSHUGE, &memcg)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002441 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002442
2443 /*
2444 * Prevent all access to pagetables with the exception of
2445 * gup_fast later hanlded by the ptep_clear_flush and the VM
2446 * handled by the anon_vma lock + PG_lock.
2447 */
2448 down_write(&mm->mmap_sem);
2449 if (unlikely(khugepaged_test_exit(mm)))
2450 goto out;
2451
2452 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002453 if (!vma)
2454 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002455 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2456 hend = vma->vm_end & HPAGE_PMD_MASK;
2457 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2458 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002459 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002460 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002461 pmd = mm_find_pmd(mm, address);
2462 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002463 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002464
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002465 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002466
2467 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002468 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002469
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002470 mmun_start = address;
2471 mmun_end = address + HPAGE_PMD_SIZE;
2472 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002473 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002474 /*
2475 * After this gup_fast can't run anymore. This also removes
2476 * any huge TLB entry from the CPU so we won't allow
2477 * huge and small TLB entries for the same virtual address
2478 * to avoid the risk of CPU bugs in that area.
2479 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002480 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002481 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002482 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002483
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002484 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002485 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002486 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002487
2488 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002489 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002490 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002491 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002492 /*
2493 * We can only use set_pmd_at when establishing
2494 * hugepmds and never for establishing regular pmds that
2495 * points to regular pagetables. Use pmd_populate for that
2496 */
2497 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002498 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002499 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002500 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002501 }
2502
2503 /*
2504 * All pages are isolated and locked so anon_vma rmap
2505 * can't run anymore.
2506 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002507 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002508
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002509 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002510 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002511 __SetPageUptodate(new_page);
2512 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002513
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002514 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2515 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002516
2517 /*
2518 * spin_lock() below is not the equivalent of smp_wmb(), so
2519 * this is needed to avoid the copy_huge_page writes to become
2520 * visible after the set_pmd_at() write.
2521 */
2522 smp_wmb();
2523
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002524 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002525 BUG_ON(!pmd_none(*pmd));
2526 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002527 mem_cgroup_commit_charge(new_page, memcg, false);
2528 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002529 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002530 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002531 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002532 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002533
2534 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002535
Andrea Arcangeliba761492011-01-13 15:46:58 -08002536 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002537out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002538 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002539 return;
2540
Andrea Arcangelice83d212011-01-13 15:47:06 -08002541out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002542 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002543 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002544}
2545
2546static int khugepaged_scan_pmd(struct mm_struct *mm,
2547 struct vm_area_struct *vma,
2548 unsigned long address,
2549 struct page **hpage)
2550{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002551 pmd_t *pmd;
2552 pte_t *pte, *_pte;
2553 int ret = 0, referenced = 0, none = 0;
2554 struct page *page;
2555 unsigned long _address;
2556 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002557 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002558
2559 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2560
Bob Liu62190492012-12-11 16:00:37 -08002561 pmd = mm_find_pmd(mm, address);
2562 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002563 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002564
Bob Liu9f1b8682013-11-12 15:07:37 -08002565 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002566 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2567 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2568 _pte++, _address += PAGE_SIZE) {
2569 pte_t pteval = *_pte;
2570 if (pte_none(pteval)) {
2571 if (++none <= khugepaged_max_ptes_none)
2572 continue;
2573 else
2574 goto out_unmap;
2575 }
2576 if (!pte_present(pteval) || !pte_write(pteval))
2577 goto out_unmap;
2578 page = vm_normal_page(vma, _address, pteval);
2579 if (unlikely(!page))
2580 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002581 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002582 * Record which node the original page is from and save this
2583 * information to khugepaged_node_load[].
2584 * Khupaged will allocate hugepage from the node has the max
2585 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002586 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002587 node = page_to_nid(page);
David Rientjes14a4e212014-08-06 16:07:29 -07002588 if (khugepaged_scan_abort(node))
2589 goto out_unmap;
Bob Liu9f1b8682013-11-12 15:07:37 -08002590 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002591 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002592 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2593 goto out_unmap;
2594 /* cannot use mapcount: can't collapse if there's a gup pin */
2595 if (page_count(page) != 1)
2596 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002597 if (pte_young(pteval) || PageReferenced(page) ||
2598 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002599 referenced = 1;
2600 }
2601 if (referenced)
2602 ret = 1;
2603out_unmap:
2604 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002605 if (ret) {
2606 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002607 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002608 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002609 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002610out:
2611 return ret;
2612}
2613
2614static void collect_mm_slot(struct mm_slot *mm_slot)
2615{
2616 struct mm_struct *mm = mm_slot->mm;
2617
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002618 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002619
2620 if (khugepaged_test_exit(mm)) {
2621 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002622 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002623 list_del(&mm_slot->mm_node);
2624
2625 /*
2626 * Not strictly needed because the mm exited already.
2627 *
2628 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2629 */
2630
2631 /* khugepaged_mm_lock actually not necessary for the below */
2632 free_mm_slot(mm_slot);
2633 mmdrop(mm);
2634 }
2635}
2636
2637static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2638 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002639 __releases(&khugepaged_mm_lock)
2640 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002641{
2642 struct mm_slot *mm_slot;
2643 struct mm_struct *mm;
2644 struct vm_area_struct *vma;
2645 int progress = 0;
2646
2647 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002648 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002649
2650 if (khugepaged_scan.mm_slot)
2651 mm_slot = khugepaged_scan.mm_slot;
2652 else {
2653 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2654 struct mm_slot, mm_node);
2655 khugepaged_scan.address = 0;
2656 khugepaged_scan.mm_slot = mm_slot;
2657 }
2658 spin_unlock(&khugepaged_mm_lock);
2659
2660 mm = mm_slot->mm;
2661 down_read(&mm->mmap_sem);
2662 if (unlikely(khugepaged_test_exit(mm)))
2663 vma = NULL;
2664 else
2665 vma = find_vma(mm, khugepaged_scan.address);
2666
2667 progress++;
2668 for (; vma; vma = vma->vm_next) {
2669 unsigned long hstart, hend;
2670
2671 cond_resched();
2672 if (unlikely(khugepaged_test_exit(mm))) {
2673 progress++;
2674 break;
2675 }
Bob Liufa475e52012-12-11 16:00:39 -08002676 if (!hugepage_vma_check(vma)) {
2677skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002678 progress++;
2679 continue;
2680 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002681 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2682 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002683 if (hstart >= hend)
2684 goto skip;
2685 if (khugepaged_scan.address > hend)
2686 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002687 if (khugepaged_scan.address < hstart)
2688 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002689 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002690
2691 while (khugepaged_scan.address < hend) {
2692 int ret;
2693 cond_resched();
2694 if (unlikely(khugepaged_test_exit(mm)))
2695 goto breakouterloop;
2696
2697 VM_BUG_ON(khugepaged_scan.address < hstart ||
2698 khugepaged_scan.address + HPAGE_PMD_SIZE >
2699 hend);
2700 ret = khugepaged_scan_pmd(mm, vma,
2701 khugepaged_scan.address,
2702 hpage);
2703 /* move to next address */
2704 khugepaged_scan.address += HPAGE_PMD_SIZE;
2705 progress += HPAGE_PMD_NR;
2706 if (ret)
2707 /* we released mmap_sem so break loop */
2708 goto breakouterloop_mmap_sem;
2709 if (progress >= pages)
2710 goto breakouterloop;
2711 }
2712 }
2713breakouterloop:
2714 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2715breakouterloop_mmap_sem:
2716
2717 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002718 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002719 /*
2720 * Release the current mm_slot if this mm is about to die, or
2721 * if we scanned all vmas of this mm.
2722 */
2723 if (khugepaged_test_exit(mm) || !vma) {
2724 /*
2725 * Make sure that if mm_users is reaching zero while
2726 * khugepaged runs here, khugepaged_exit will find
2727 * mm_slot not pointing to the exiting mm.
2728 */
2729 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2730 khugepaged_scan.mm_slot = list_entry(
2731 mm_slot->mm_node.next,
2732 struct mm_slot, mm_node);
2733 khugepaged_scan.address = 0;
2734 } else {
2735 khugepaged_scan.mm_slot = NULL;
2736 khugepaged_full_scans++;
2737 }
2738
2739 collect_mm_slot(mm_slot);
2740 }
2741
2742 return progress;
2743}
2744
2745static int khugepaged_has_work(void)
2746{
2747 return !list_empty(&khugepaged_scan.mm_head) &&
2748 khugepaged_enabled();
2749}
2750
2751static int khugepaged_wait_event(void)
2752{
2753 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002754 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002755}
2756
Xiao Guangrongd5169042012-10-08 16:29:48 -07002757static void khugepaged_do_scan(void)
2758{
2759 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002760 unsigned int progress = 0, pass_through_head = 0;
2761 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002762 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002763
2764 barrier(); /* write khugepaged_pages_to_scan to local stack */
2765
2766 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002767 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002768 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002769
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002770 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002771
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002772 if (unlikely(kthread_should_stop() || freezing(current)))
2773 break;
2774
Andrea Arcangeliba761492011-01-13 15:46:58 -08002775 spin_lock(&khugepaged_mm_lock);
2776 if (!khugepaged_scan.mm_slot)
2777 pass_through_head++;
2778 if (khugepaged_has_work() &&
2779 pass_through_head < 2)
2780 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002781 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002782 else
2783 progress = pages;
2784 spin_unlock(&khugepaged_mm_lock);
2785 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002786
Xiao Guangrongd5169042012-10-08 16:29:48 -07002787 if (!IS_ERR_OR_NULL(hpage))
2788 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002789}
2790
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002791static void khugepaged_wait_work(void)
2792{
2793 try_to_freeze();
2794
2795 if (khugepaged_has_work()) {
2796 if (!khugepaged_scan_sleep_millisecs)
2797 return;
2798
2799 wait_event_freezable_timeout(khugepaged_wait,
2800 kthread_should_stop(),
2801 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2802 return;
2803 }
2804
2805 if (khugepaged_enabled())
2806 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2807}
2808
Andrea Arcangeliba761492011-01-13 15:46:58 -08002809static int khugepaged(void *none)
2810{
2811 struct mm_slot *mm_slot;
2812
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002813 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002814 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002815
Xiao Guangrongb7231782012-10-08 16:29:54 -07002816 while (!kthread_should_stop()) {
2817 khugepaged_do_scan();
2818 khugepaged_wait_work();
2819 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002820
2821 spin_lock(&khugepaged_mm_lock);
2822 mm_slot = khugepaged_scan.mm_slot;
2823 khugepaged_scan.mm_slot = NULL;
2824 if (mm_slot)
2825 collect_mm_slot(mm_slot);
2826 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002827 return 0;
2828}
2829
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002830static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2831 unsigned long haddr, pmd_t *pmd)
2832{
2833 struct mm_struct *mm = vma->vm_mm;
2834 pgtable_t pgtable;
2835 pmd_t _pmd;
2836 int i;
2837
Joerg Roedel34ee6452014-11-13 13:46:09 +11002838 pmdp_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002839 /* leave pmd empty until pte is filled */
2840
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002841 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002842 pmd_populate(mm, &_pmd, pgtable);
2843
2844 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2845 pte_t *pte, entry;
2846 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2847 entry = pte_mkspecial(entry);
2848 pte = pte_offset_map(&_pmd, haddr);
2849 VM_BUG_ON(!pte_none(*pte));
2850 set_pte_at(mm, haddr, pte, entry);
2851 pte_unmap(pte);
2852 }
2853 smp_wmb(); /* make pte visible before pmd */
2854 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002855 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002856}
2857
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002858void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2859 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002860{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002861 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002862 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002863 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002864 unsigned long haddr = address & HPAGE_PMD_MASK;
2865 unsigned long mmun_start; /* For mmu_notifiers */
2866 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002867
2868 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002869
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002870 mmun_start = haddr;
2871 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002872again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002873 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002874 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002875 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002876 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002877 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2878 return;
2879 }
2880 if (is_huge_zero_pmd(*pmd)) {
2881 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002882 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002883 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002884 return;
2885 }
2886 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002887 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002888 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002889 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002890 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002891
2892 split_huge_page(page);
2893
2894 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002895
2896 /*
2897 * We don't always have down_write of mmap_sem here: a racing
2898 * do_huge_pmd_wp_page() might have copied-on-write to another
2899 * huge page before our split_huge_page() got the anon_vma lock.
2900 */
2901 if (unlikely(pmd_trans_huge(*pmd)))
2902 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002903}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002904
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002905void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2906 pmd_t *pmd)
2907{
2908 struct vm_area_struct *vma;
2909
2910 vma = find_vma(mm, address);
2911 BUG_ON(vma == NULL);
2912 split_huge_page_pmd(vma, address, pmd);
2913}
2914
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002915static void split_huge_page_address(struct mm_struct *mm,
2916 unsigned long address)
2917{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002918 pgd_t *pgd;
2919 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002920 pmd_t *pmd;
2921
2922 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2923
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002924 pgd = pgd_offset(mm, address);
2925 if (!pgd_present(*pgd))
2926 return;
2927
2928 pud = pud_offset(pgd, address);
2929 if (!pud_present(*pud))
2930 return;
2931
2932 pmd = pmd_offset(pud, address);
2933 if (!pmd_present(*pmd))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002934 return;
2935 /*
2936 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2937 * materialize from under us.
2938 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002939 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002940}
2941
2942void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2943 unsigned long start,
2944 unsigned long end,
2945 long adjust_next)
2946{
2947 /*
2948 * If the new start address isn't hpage aligned and it could
2949 * previously contain an hugepage: check if we need to split
2950 * an huge pmd.
2951 */
2952 if (start & ~HPAGE_PMD_MASK &&
2953 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2954 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2955 split_huge_page_address(vma->vm_mm, start);
2956
2957 /*
2958 * If the new end address isn't hpage aligned and it could
2959 * previously contain an hugepage: check if we need to split
2960 * an huge pmd.
2961 */
2962 if (end & ~HPAGE_PMD_MASK &&
2963 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2964 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2965 split_huge_page_address(vma->vm_mm, end);
2966
2967 /*
2968 * If we're also updating the vma->vm_next->vm_start, if the new
2969 * vm_next->vm_start isn't page aligned and it could previously
2970 * contain an hugepage: check if we need to split an huge pmd.
2971 */
2972 if (adjust_next > 0) {
2973 struct vm_area_struct *next = vma->vm_next;
2974 unsigned long nstart = next->vm_start;
2975 nstart += adjust_next << PAGE_SHIFT;
2976 if (nstart & ~HPAGE_PMD_MASK &&
2977 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2978 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2979 split_huge_page_address(next->vm_mm, nstart);
2980 }
2981}