blob: c107094f79bae9ee895bd6bf30976d900f16c141 [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);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -070070static void khugepaged_slab_exit(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080071
Sasha Levin43b5fbb2013-02-22 16:32:27 -080072#define MM_SLOTS_HASH_BITS 10
73static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
74
Andrea Arcangeliba761492011-01-13 15:46:58 -080075static struct kmem_cache *mm_slot_cache __read_mostly;
76
77/**
78 * struct mm_slot - hash lookup from mm to mm_slot
79 * @hash: hash collision list
80 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
81 * @mm: the mm that this information is valid for
82 */
83struct mm_slot {
84 struct hlist_node hash;
85 struct list_head mm_node;
86 struct mm_struct *mm;
87};
88
89/**
90 * struct khugepaged_scan - cursor for scanning
91 * @mm_head: the head of the mm list to scan
92 * @mm_slot: the current mm_slot we are scanning
93 * @address: the next address inside that to be scanned
94 *
95 * There is only the one khugepaged_scan instance of this cursor structure.
96 */
97struct khugepaged_scan {
98 struct list_head mm_head;
99 struct mm_slot *mm_slot;
100 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -0700101};
102static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800103 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104};
105
Andrea Arcangelif0005652011-01-13 15:47:04 -0800106
107static int set_recommended_min_free_kbytes(void)
108{
109 struct zone *zone;
110 int nr_zones = 0;
111 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800112
Andrea Arcangelif0005652011-01-13 15:47:04 -0800113 for_each_populated_zone(zone)
114 nr_zones++;
115
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119 /*
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
124 */
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
132
Han Pingtian42aa83c2014-01-23 15:53:28 -0800133 if (recommended_min > min_free_kbytes) {
134 if (user_min_free_kbytes >= 0)
135 pr_info("raising min_free_kbytes from %d to %lu "
136 "to help transparent hugepage allocations\n",
137 min_free_kbytes, recommended_min);
138
Andrea Arcangelif0005652011-01-13 15:47:04 -0800139 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800140 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800141 setup_per_zone_wmarks();
142 return 0;
143}
Andrea Arcangelif0005652011-01-13 15:47:04 -0800144
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700145static int start_stop_khugepaged(void)
Andrea Arcangeliba761492011-01-13 15:46:58 -0800146{
147 int err = 0;
148 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800149 if (!khugepaged_thread)
150 khugepaged_thread = kthread_run(khugepaged, NULL,
151 "khugepaged");
152 if (unlikely(IS_ERR(khugepaged_thread))) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700153 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
Andrea Arcangeliba761492011-01-13 15:46:58 -0800154 err = PTR_ERR(khugepaged_thread);
155 khugepaged_thread = NULL;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700156 goto fail;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800157 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700158
159 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800160 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800161
162 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700163 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700164 kthread_stop(khugepaged_thread);
165 khugepaged_thread = NULL;
166 }
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700167fail:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800168 return err;
169}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800170
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800171static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -0800172struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800173
174static inline bool is_huge_zero_pmd(pmd_t pmd)
175{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700176 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800177}
178
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700179static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800180{
181 struct page *zero_page;
182retry:
183 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -0700184 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800185
186 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
187 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800188 if (!zero_page) {
189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700190 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800191 }
192 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800193 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700194 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800195 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700196 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800197 goto retry;
198 }
199
200 /* We take additional reference here. It will be put back by shrinker */
201 atomic_set(&huge_zero_refcount, 2);
202 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -0700203 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800204}
205
206static void put_huge_zero_page(void)
207{
208 /*
209 * Counter should never go to zero here. Only shrinker can put
210 * last reference.
211 */
212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
213}
214
Glauber Costa48896462013-08-28 10:18:15 +1000215static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
216 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800217{
Glauber Costa48896462013-08-28 10:18:15 +1000218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
220}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800221
Glauber Costa48896462013-08-28 10:18:15 +1000222static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
223 struct shrink_control *sc)
224{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800225 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700226 struct page *zero_page = xchg(&huge_zero_page, NULL);
227 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700228 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000229 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800230 }
231
232 return 0;
233}
234
235static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000236 .count_objects = shrink_huge_zero_page_count,
237 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800238 .seeks = DEFAULT_SEEKS,
239};
240
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800241#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800242
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800243static ssize_t double_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag enabled,
246 enum transparent_hugepage_flag req_madv)
247{
248 if (test_bit(enabled, &transparent_hugepage_flags)) {
249 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
250 return sprintf(buf, "[always] madvise never\n");
251 } else if (test_bit(req_madv, &transparent_hugepage_flags))
252 return sprintf(buf, "always [madvise] never\n");
253 else
254 return sprintf(buf, "always madvise [never]\n");
255}
256static ssize_t double_flag_store(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 const char *buf, size_t count,
259 enum transparent_hugepage_flag enabled,
260 enum transparent_hugepage_flag req_madv)
261{
262 if (!memcmp("always", buf,
263 min(sizeof("always")-1, count))) {
264 set_bit(enabled, &transparent_hugepage_flags);
265 clear_bit(req_madv, &transparent_hugepage_flags);
266 } else if (!memcmp("madvise", buf,
267 min(sizeof("madvise")-1, count))) {
268 clear_bit(enabled, &transparent_hugepage_flags);
269 set_bit(req_madv, &transparent_hugepage_flags);
270 } else if (!memcmp("never", buf,
271 min(sizeof("never")-1, count))) {
272 clear_bit(enabled, &transparent_hugepage_flags);
273 clear_bit(req_madv, &transparent_hugepage_flags);
274 } else
275 return -EINVAL;
276
277 return count;
278}
279
280static ssize_t enabled_show(struct kobject *kobj,
281 struct kobj_attribute *attr, char *buf)
282{
283 return double_flag_show(kobj, attr, buf,
284 TRANSPARENT_HUGEPAGE_FLAG,
285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
286}
287static ssize_t enabled_store(struct kobject *kobj,
288 struct kobj_attribute *attr,
289 const char *buf, size_t count)
290{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800291 ssize_t ret;
292
293 ret = double_flag_store(kobj, attr, buf, count,
294 TRANSPARENT_HUGEPAGE_FLAG,
295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
296
297 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700298 int err;
299
300 mutex_lock(&khugepaged_mutex);
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700301 err = start_stop_khugepaged();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700302 mutex_unlock(&khugepaged_mutex);
303
Andrea Arcangeliba761492011-01-13 15:46:58 -0800304 if (err)
305 ret = err;
306 }
307
308 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800309}
310static struct kobj_attribute enabled_attr =
311 __ATTR(enabled, 0644, enabled_show, enabled_store);
312
313static ssize_t single_flag_show(struct kobject *kobj,
314 struct kobj_attribute *attr, char *buf,
315 enum transparent_hugepage_flag flag)
316{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700317 return sprintf(buf, "%d\n",
318 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800319}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700320
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800321static ssize_t single_flag_store(struct kobject *kobj,
322 struct kobj_attribute *attr,
323 const char *buf, size_t count,
324 enum transparent_hugepage_flag flag)
325{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700326 unsigned long value;
327 int ret;
328
329 ret = kstrtoul(buf, 10, &value);
330 if (ret < 0)
331 return ret;
332 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800333 return -EINVAL;
334
Ben Hutchingse27e6152011-04-14 15:22:21 -0700335 if (value)
336 set_bit(flag, &transparent_hugepage_flags);
337 else
338 clear_bit(flag, &transparent_hugepage_flags);
339
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800340 return count;
341}
342
343/*
344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
346 * memory just to allocate one more hugepage.
347 */
348static ssize_t defrag_show(struct kobject *kobj,
349 struct kobj_attribute *attr, char *buf)
350{
351 return double_flag_show(kobj, attr, buf,
352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
354}
355static ssize_t defrag_store(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 const char *buf, size_t count)
358{
359 return double_flag_store(kobj, attr, buf, count,
360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
362}
363static struct kobj_attribute defrag_attr =
364 __ATTR(defrag, 0644, defrag_show, defrag_store);
365
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800366static ssize_t use_zero_page_show(struct kobject *kobj,
367 struct kobj_attribute *attr, char *buf)
368{
369 return single_flag_show(kobj, attr, buf,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
371}
372static ssize_t use_zero_page_store(struct kobject *kobj,
373 struct kobj_attribute *attr, const char *buf, size_t count)
374{
375 return single_flag_store(kobj, attr, buf, count,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static struct kobj_attribute use_zero_page_attr =
379 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800380#ifdef CONFIG_DEBUG_VM
381static ssize_t debug_cow_show(struct kobject *kobj,
382 struct kobj_attribute *attr, char *buf)
383{
384 return single_flag_show(kobj, attr, buf,
385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
386}
387static ssize_t debug_cow_store(struct kobject *kobj,
388 struct kobj_attribute *attr,
389 const char *buf, size_t count)
390{
391 return single_flag_store(kobj, attr, buf, count,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
393}
394static struct kobj_attribute debug_cow_attr =
395 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
396#endif /* CONFIG_DEBUG_VM */
397
398static struct attribute *hugepage_attr[] = {
399 &enabled_attr.attr,
400 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800401 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800402#ifdef CONFIG_DEBUG_VM
403 &debug_cow_attr.attr,
404#endif
405 NULL,
406};
407
408static struct attribute_group hugepage_attr_group = {
409 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800410};
411
412static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
413 struct kobj_attribute *attr,
414 char *buf)
415{
416 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
417}
418
419static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
420 struct kobj_attribute *attr,
421 const char *buf, size_t count)
422{
423 unsigned long msecs;
424 int err;
425
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700426 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800427 if (err || msecs > UINT_MAX)
428 return -EINVAL;
429
430 khugepaged_scan_sleep_millisecs = msecs;
431 wake_up_interruptible(&khugepaged_wait);
432
433 return count;
434}
435static struct kobj_attribute scan_sleep_millisecs_attr =
436 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
437 scan_sleep_millisecs_store);
438
439static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
440 struct kobj_attribute *attr,
441 char *buf)
442{
443 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
444}
445
446static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
447 struct kobj_attribute *attr,
448 const char *buf, size_t count)
449{
450 unsigned long msecs;
451 int err;
452
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700453 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800454 if (err || msecs > UINT_MAX)
455 return -EINVAL;
456
457 khugepaged_alloc_sleep_millisecs = msecs;
458 wake_up_interruptible(&khugepaged_wait);
459
460 return count;
461}
462static struct kobj_attribute alloc_sleep_millisecs_attr =
463 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
464 alloc_sleep_millisecs_store);
465
466static ssize_t pages_to_scan_show(struct kobject *kobj,
467 struct kobj_attribute *attr,
468 char *buf)
469{
470 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
471}
472static ssize_t pages_to_scan_store(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 const char *buf, size_t count)
475{
476 int err;
477 unsigned long pages;
478
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700479 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800480 if (err || !pages || pages > UINT_MAX)
481 return -EINVAL;
482
483 khugepaged_pages_to_scan = pages;
484
485 return count;
486}
487static struct kobj_attribute pages_to_scan_attr =
488 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
489 pages_to_scan_store);
490
491static ssize_t pages_collapsed_show(struct kobject *kobj,
492 struct kobj_attribute *attr,
493 char *buf)
494{
495 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
496}
497static struct kobj_attribute pages_collapsed_attr =
498 __ATTR_RO(pages_collapsed);
499
500static ssize_t full_scans_show(struct kobject *kobj,
501 struct kobj_attribute *attr,
502 char *buf)
503{
504 return sprintf(buf, "%u\n", khugepaged_full_scans);
505}
506static struct kobj_attribute full_scans_attr =
507 __ATTR_RO(full_scans);
508
509static ssize_t khugepaged_defrag_show(struct kobject *kobj,
510 struct kobj_attribute *attr, char *buf)
511{
512 return single_flag_show(kobj, attr, buf,
513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
514}
515static ssize_t khugepaged_defrag_store(struct kobject *kobj,
516 struct kobj_attribute *attr,
517 const char *buf, size_t count)
518{
519 return single_flag_store(kobj, attr, buf, count,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
521}
522static struct kobj_attribute khugepaged_defrag_attr =
523 __ATTR(defrag, 0644, khugepaged_defrag_show,
524 khugepaged_defrag_store);
525
526/*
527 * max_ptes_none controls if khugepaged should collapse hugepages over
528 * any unmapped ptes in turn potentially increasing the memory
529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
530 * reduce the available free memory in the system as it
531 * runs. Increasing max_ptes_none will instead potentially reduce the
532 * free memory in the system during the khugepaged scan.
533 */
534static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
535 struct kobj_attribute *attr,
536 char *buf)
537{
538 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
539}
540static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 const char *buf, size_t count)
543{
544 int err;
545 unsigned long max_ptes_none;
546
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700547 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800548 if (err || max_ptes_none > HPAGE_PMD_NR-1)
549 return -EINVAL;
550
551 khugepaged_max_ptes_none = max_ptes_none;
552
553 return count;
554}
555static struct kobj_attribute khugepaged_max_ptes_none_attr =
556 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
557 khugepaged_max_ptes_none_store);
558
559static struct attribute *khugepaged_attr[] = {
560 &khugepaged_defrag_attr.attr,
561 &khugepaged_max_ptes_none_attr.attr,
562 &pages_to_scan_attr.attr,
563 &pages_collapsed_attr.attr,
564 &full_scans_attr.attr,
565 &scan_sleep_millisecs_attr.attr,
566 &alloc_sleep_millisecs_attr.attr,
567 NULL,
568};
569
570static struct attribute_group khugepaged_attr_group = {
571 .attrs = khugepaged_attr,
572 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800573};
Shaohua Li569e5592012-01-12 17:19:11 -0800574
575static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
576{
577 int err;
578
579 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
580 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700581 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800582 return -ENOMEM;
583 }
584
585 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
586 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700587 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800588 goto delete_obj;
589 }
590
591 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
592 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700593 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800594 goto remove_hp_group;
595 }
596
597 return 0;
598
599remove_hp_group:
600 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
601delete_obj:
602 kobject_put(*hugepage_kobj);
603 return err;
604}
605
606static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
607{
608 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
609 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
610 kobject_put(hugepage_kobj);
611}
612#else
613static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
614{
615 return 0;
616}
617
618static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
619{
620}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800621#endif /* CONFIG_SYSFS */
622
623static int __init hugepage_init(void)
624{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800625 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800626 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800627
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800628 if (!has_transparent_hugepage()) {
629 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800630 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800631 }
632
Shaohua Li569e5592012-01-12 17:19:11 -0800633 err = hugepage_init_sysfs(&hugepage_kobj);
634 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700635 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800636
637 err = khugepaged_slab_init();
638 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700639 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800640
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700641 err = register_shrinker(&huge_zero_page_shrinker);
642 if (err)
643 goto err_hzp_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800644
Rik van Riel97562cd2011-01-13 15:47:12 -0800645 /*
646 * By default disable transparent hugepages on smaller systems,
647 * where the extra memory used could hurt more than TLB overhead
648 * is likely to save. The admin can still enable it through /sys.
649 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700650 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800651 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700652 return 0;
653 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800654
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700655 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700656 if (err)
657 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800658
Shaohua Li569e5592012-01-12 17:19:11 -0800659 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700660err_khugepaged:
661 unregister_shrinker(&huge_zero_page_shrinker);
662err_hzp_shrinker:
663 khugepaged_slab_exit();
664err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800665 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700666err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800667 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800668}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800669subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800670
671static int __init setup_transparent_hugepage(char *str)
672{
673 int ret = 0;
674 if (!str)
675 goto out;
676 if (!strcmp(str, "always")) {
677 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
678 &transparent_hugepage_flags);
679 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
680 &transparent_hugepage_flags);
681 ret = 1;
682 } else if (!strcmp(str, "madvise")) {
683 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
684 &transparent_hugepage_flags);
685 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
686 &transparent_hugepage_flags);
687 ret = 1;
688 } else if (!strcmp(str, "never")) {
689 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
690 &transparent_hugepage_flags);
691 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
692 &transparent_hugepage_flags);
693 ret = 1;
694 }
695out:
696 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700697 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800698 return ret;
699}
700__setup("transparent_hugepage=", setup_transparent_hugepage);
701
Mel Gormanb32967f2012-11-19 12:35:47 +0000702pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800703{
704 if (likely(vma->vm_flags & VM_WRITE))
705 pmd = pmd_mkwrite(pmd);
706 return pmd;
707}
708
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700709static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800710{
711 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700712 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800713 entry = pmd_mkhuge(entry);
714 return entry;
715}
716
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800717static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
718 struct vm_area_struct *vma,
719 unsigned long haddr, pmd_t *pmd,
Michal Hocko3b363692015-04-15 16:13:29 -0700720 struct page *page, gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800721{
Johannes Weiner00501b52014-08-08 14:19:20 -0700722 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800723 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800724 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725
Sasha Levin309381fea2014-01-23 15:52:54 -0800726 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700727
Michal Hocko3b363692015-04-15 16:13:29 -0700728 if (mem_cgroup_try_charge(page, mm, gfp, &memcg))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800729 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800730
Johannes Weiner00501b52014-08-08 14:19:20 -0700731 pgtable = pte_alloc_one(mm, haddr);
732 if (unlikely(!pgtable)) {
733 mem_cgroup_cancel_charge(page, memcg);
734 return VM_FAULT_OOM;
735 }
736
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800737 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700738 /*
739 * The memory barrier inside __SetPageUptodate makes sure that
740 * clear_huge_page writes become visible before the set_pmd_at()
741 * write.
742 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800743 __SetPageUptodate(page);
744
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800745 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800746 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800747 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -0700748 mem_cgroup_cancel_charge(page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800749 put_page(page);
750 pte_free(mm, pgtable);
751 } else {
752 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700753 entry = mk_huge_pmd(page, vma->vm_page_prot);
754 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800755 page_add_new_anon_rmap(page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700756 mem_cgroup_commit_charge(page, memcg, false);
757 lru_cache_add_active_or_unevictable(page, vma);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700758 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800759 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800760 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800761 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800762 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800763 }
764
David Rientjesaa2e8782012-05-29 15:06:17 -0700765 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800766}
767
Andi Kleencc5d4622011-03-22 16:33:13 -0700768static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800769{
Andi Kleencc5d4622011-03-22 16:33:13 -0700770 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800771}
772
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800773/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800774static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800775 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700776 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800777{
778 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800779 if (!pmd_none(*pmd))
780 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700781 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800782 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700783 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800784 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800785 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800786 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800787}
788
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800789int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
790 unsigned long address, pmd_t *pmd,
791 unsigned int flags)
792{
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800793 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800794 struct page *page;
795 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800796
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700797 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700798 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700799 if (unlikely(anon_vma_prepare(vma)))
800 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700801 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700802 return VM_FAULT_OOM;
Dominik Dingel593befa2014-10-23 12:07:44 +0200803 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700804 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800805 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700806 pgtable_t pgtable;
807 struct page *zero_page;
808 bool set;
809 pgtable = pte_alloc_one(mm, haddr);
810 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800811 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700812 zero_page = get_huge_zero_page();
813 if (unlikely(!zero_page)) {
814 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700815 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700816 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700817 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800818 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700819 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
820 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800821 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700822 if (!set) {
823 pte_free(mm, pgtable);
824 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800825 }
David Rientjesedad9d22012-05-29 15:06:17 -0700826 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800827 }
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800828 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
829 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700830 if (unlikely(!page)) {
831 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700832 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700833 }
Michal Hocko3b363692015-04-15 16:13:29 -0700834 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page, gfp))) {
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700835 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700836 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700837 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700838 }
839
David Rientjes17766dd2013-09-12 15:14:06 -0700840 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700841 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800842}
843
844int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
845 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
846 struct vm_area_struct *vma)
847{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800848 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800849 struct page *src_page;
850 pmd_t pmd;
851 pgtable_t pgtable;
852 int ret;
853
854 ret = -ENOMEM;
855 pgtable = pte_alloc_one(dst_mm, addr);
856 if (unlikely(!pgtable))
857 goto out;
858
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800859 dst_ptl = pmd_lock(dst_mm, dst_pmd);
860 src_ptl = pmd_lockptr(src_mm, src_pmd);
861 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800862
863 ret = -EAGAIN;
864 pmd = *src_pmd;
865 if (unlikely(!pmd_trans_huge(pmd))) {
866 pte_free(dst_mm, pgtable);
867 goto out_unlock;
868 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800869 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800870 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800871 * under splitting since we don't split the page itself, only pmd to
872 * a page table.
873 */
874 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700875 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800876 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800877 /*
878 * get_huge_zero_page() will never allocate a new page here,
879 * since we already have a zero page to copy. It just takes a
880 * reference.
881 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700882 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800883 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700884 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800885 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800886 ret = 0;
887 goto out_unlock;
888 }
Mel Gormande466bd2013-12-18 17:08:42 -0800889
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800890 if (unlikely(pmd_trans_splitting(pmd))) {
891 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800892 spin_unlock(src_ptl);
893 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800894 pte_free(dst_mm, pgtable);
895
896 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
897 goto out;
898 }
899 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800900 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800901 get_page(src_page);
902 page_dup_rmap(src_page);
903 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
904
905 pmdp_set_wrprotect(src_mm, addr, src_pmd);
906 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700907 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800908 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800909 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910
911 ret = 0;
912out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800913 spin_unlock(src_ptl);
914 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800915out:
916 return ret;
917}
918
Will Deacona1dd4502012-12-11 16:01:27 -0800919void huge_pmd_set_accessed(struct mm_struct *mm,
920 struct vm_area_struct *vma,
921 unsigned long address,
922 pmd_t *pmd, pmd_t orig_pmd,
923 int dirty)
924{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800925 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800926 pmd_t entry;
927 unsigned long haddr;
928
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800929 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800930 if (unlikely(!pmd_same(*pmd, orig_pmd)))
931 goto unlock;
932
933 entry = pmd_mkyoung(orig_pmd);
934 haddr = address & HPAGE_PMD_MASK;
935 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
936 update_mmu_cache_pmd(vma, address, pmd);
937
938unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800939 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800940}
941
Hugh Dickins5338a932014-06-23 13:22:05 -0700942/*
943 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
944 * during copy_user_huge_page()'s copy_page_rep(): in the case when
945 * the source page gets split and a tail freed before copy completes.
946 * Called under pmd_lock of checked pmd, so safe from splitting itself.
947 */
948static void get_user_huge_page(struct page *page)
949{
950 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
951 struct page *endpage = page + HPAGE_PMD_NR;
952
953 atomic_add(HPAGE_PMD_NR, &page->_count);
954 while (++page < endpage)
955 get_huge_page_tail(page);
956 } else {
957 get_page(page);
958 }
959}
960
961static void put_user_huge_page(struct page *page)
962{
963 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
964 struct page *endpage = page + HPAGE_PMD_NR;
965
966 while (page < endpage)
967 put_page(page++);
968 } else {
969 put_page(page);
970 }
971}
972
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800973static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
974 struct vm_area_struct *vma,
975 unsigned long address,
976 pmd_t *pmd, pmd_t orig_pmd,
977 struct page *page,
978 unsigned long haddr)
979{
Johannes Weiner00501b52014-08-08 14:19:20 -0700980 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800981 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800982 pgtable_t pgtable;
983 pmd_t _pmd;
984 int ret = 0, i;
985 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700986 unsigned long mmun_start; /* For mmu_notifiers */
987 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800988
989 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
990 GFP_KERNEL);
991 if (unlikely(!pages)) {
992 ret |= VM_FAULT_OOM;
993 goto out;
994 }
995
996 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700997 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
998 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -0800999 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001000 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -07001001 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1002 &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001003 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001004 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001005 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001006 memcg = (void *)page_private(pages[i]);
1007 set_page_private(pages[i], 0);
1008 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001009 put_page(pages[i]);
1010 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001011 kfree(pages);
1012 ret |= VM_FAULT_OOM;
1013 goto out;
1014 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001015 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001016 }
1017
1018 for (i = 0; i < HPAGE_PMD_NR; i++) {
1019 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001020 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001021 __SetPageUptodate(pages[i]);
1022 cond_resched();
1023 }
1024
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001025 mmun_start = haddr;
1026 mmun_end = haddr + HPAGE_PMD_SIZE;
1027 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1028
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001029 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001030 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1031 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001032 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001033
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001034 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001035 /* leave pmd empty until pte is filled */
1036
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001037 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001038 pmd_populate(mm, &_pmd, pgtable);
1039
1040 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1041 pte_t *pte, entry;
1042 entry = mk_pte(pages[i], vma->vm_page_prot);
1043 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001044 memcg = (void *)page_private(pages[i]);
1045 set_page_private(pages[i], 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001046 page_add_new_anon_rmap(pages[i], vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001047 mem_cgroup_commit_charge(pages[i], memcg, false);
1048 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001049 pte = pte_offset_map(&_pmd, haddr);
1050 VM_BUG_ON(!pte_none(*pte));
1051 set_pte_at(mm, haddr, pte, entry);
1052 pte_unmap(pte);
1053 }
1054 kfree(pages);
1055
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056 smp_wmb(); /* make pte visible before pmd */
1057 pmd_populate(mm, pmd, pgtable);
1058 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001059 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001060
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001061 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1062
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001063 ret |= VM_FAULT_WRITE;
1064 put_page(page);
1065
1066out:
1067 return ret;
1068
1069out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001070 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001071 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001072 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001073 memcg = (void *)page_private(pages[i]);
1074 set_page_private(pages[i], 0);
1075 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001076 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001077 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001078 kfree(pages);
1079 goto out;
1080}
1081
1082int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1083 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1084{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001085 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001086 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001087 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001088 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001089 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001090 unsigned long mmun_start; /* For mmu_notifiers */
1091 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001092 gfp_t huge_gfp; /* for allocation and charge */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001093
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001094 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001095 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001096 haddr = address & HPAGE_PMD_MASK;
1097 if (is_huge_zero_pmd(orig_pmd))
1098 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001099 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001100 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1101 goto out_unlock;
1102
1103 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001104 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001105 if (page_mapcount(page) == 1) {
1106 pmd_t entry;
1107 entry = pmd_mkyoung(orig_pmd);
1108 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1109 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001110 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001111 ret |= VM_FAULT_WRITE;
1112 goto out_unlock;
1113 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001114 get_user_huge_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001115 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001116alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001117 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001118 !transparent_hugepage_debug_cow()) {
Michal Hocko3b363692015-04-15 16:13:29 -07001119 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1120 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001121 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001122 new_page = NULL;
1123
1124 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001125 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001126 split_huge_page_pmd(vma, address, pmd);
1127 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001128 } else {
1129 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1130 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001131 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001132 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001133 ret |= VM_FAULT_FALLBACK;
1134 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001135 put_user_huge_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001136 }
David Rientjes17766dd2013-09-12 15:14:06 -07001137 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001138 goto out;
1139 }
1140
Michal Hocko3b363692015-04-15 16:13:29 -07001141 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001142 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001143 if (page) {
1144 split_huge_page(page);
Hugh Dickins5338a932014-06-23 13:22:05 -07001145 put_user_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001146 } else
1147 split_huge_page_pmd(vma, address, pmd);
1148 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001149 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001150 goto out;
1151 }
1152
David Rientjes17766dd2013-09-12 15:14:06 -07001153 count_vm_event(THP_FAULT_ALLOC);
1154
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001155 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001156 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1157 else
1158 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159 __SetPageUptodate(new_page);
1160
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001161 mmun_start = haddr;
1162 mmun_end = haddr + HPAGE_PMD_SIZE;
1163 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1164
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001165 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001166 if (page)
Hugh Dickins5338a932014-06-23 13:22:05 -07001167 put_user_huge_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001168 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001169 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -07001170 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001171 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001172 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001173 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001174 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001175 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1176 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001177 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001178 page_add_new_anon_rmap(new_page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001179 mem_cgroup_commit_charge(new_page, memcg, false);
1180 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001181 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001182 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001183 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001184 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001185 put_huge_zero_page();
1186 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001187 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001188 page_remove_rmap(page);
1189 put_page(page);
1190 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191 ret |= VM_FAULT_WRITE;
1192 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001193 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001194out_mn:
1195 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1196out:
1197 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001198out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001199 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001200 return ret;
1201}
1202
David Rientjesb676b292012-10-08 16:34:03 -07001203struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001204 unsigned long addr,
1205 pmd_t *pmd,
1206 unsigned int flags)
1207{
David Rientjesb676b292012-10-08 16:34:03 -07001208 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209 struct page *page = NULL;
1210
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001211 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001212
1213 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1214 goto out;
1215
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001216 /* Avoid dumping huge zero page */
1217 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1218 return ERR_PTR(-EFAULT);
1219
Mel Gorman2b4847e2013-12-18 17:08:32 -08001220 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001221 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001222 goto out;
1223
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001224 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001225 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001226 if (flags & FOLL_TOUCH) {
1227 pmd_t _pmd;
1228 /*
1229 * We should set the dirty bit only for FOLL_WRITE but
1230 * for now the dirty bit in the pmd is meaningless.
1231 * And if the dirty bit will become meaningful and
1232 * we'll only set it with FOLL_WRITE, an atomic
1233 * set_bit will be required on the pmd to set the
1234 * young bit, instead of the current set_pmd_at.
1235 */
1236 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001237 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1238 pmd, _pmd, 1))
1239 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001240 }
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -07001241 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
David Rientjesb676b292012-10-08 16:34:03 -07001242 if (page->mapping && trylock_page(page)) {
1243 lru_add_drain();
1244 if (page->mapping)
1245 mlock_vma_page(page);
1246 unlock_page(page);
1247 }
1248 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001249 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001250 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001251 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001252 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001253
1254out:
1255 return page;
1256}
1257
Mel Gormand10e63f2012-10-25 14:16:31 +02001258/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001259int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1260 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001261{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001262 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001263 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001264 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001265 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001266 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001267 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001268 bool page_locked;
1269 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001270 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001271 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001272
Mel Gormanc0e7cad2015-02-12 14:58:41 -08001273 /* A PROT_NONE fault should not end up here */
1274 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1275
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))) {
Mel Gorman5d833062015-02-12 14:58:16 -08001286 page = pmd_page(*pmdp);
Mel Gormande466bd2013-12-18 17:08:42 -08001287 spin_unlock(ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001288 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001289 goto out;
1290 }
1291
Mel Gormand10e63f2012-10-25 14:16:31 +02001292 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001293 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001294 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001295 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001296 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001297 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001298 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001299 flags |= TNF_FAULT_LOCAL;
1300 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001301
Mel Gormanbea66fb2015-03-25 15:55:37 -07001302 /* See similar comment in do_numa_page for explanation */
1303 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001304 flags |= TNF_NO_GROUP;
1305
1306 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001307 * Acquire the page lock to serialise THP migrations but avoid dropping
1308 * page_table_lock if at all possible
1309 */
Mel Gormanb8916632013-10-07 11:28:44 +01001310 page_locked = trylock_page(page);
1311 target_nid = mpol_misplaced(page, vma, haddr);
1312 if (target_nid == -1) {
1313 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001314 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001315 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001316 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001317
Mel Gormande466bd2013-12-18 17:08:42 -08001318 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001319 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001320 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001321 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001322 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001323 goto out;
1324 }
1325
Mel Gorman2b4847e2013-12-18 17:08:32 -08001326 /*
1327 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1328 * to serialises splits
1329 */
Mel Gormanb8916632013-10-07 11:28:44 +01001330 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001331 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001332 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001333
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001334 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001335 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001336 if (unlikely(!pmd_same(pmd, *pmdp))) {
1337 unlock_page(page);
1338 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001339 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001340 goto out_unlock;
1341 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001342
Mel Gormanc3a489c2013-12-18 17:08:38 -08001343 /* Bail if we fail to protect against THP splits for any reason */
1344 if (unlikely(!anon_vma)) {
1345 put_page(page);
1346 page_nid = -1;
1347 goto clear_pmdnuma;
1348 }
1349
Mel Gormana54a4072013-10-07 11:28:46 +01001350 /*
1351 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001352 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001353 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001354 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001355 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001356 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001357 if (migrated) {
1358 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001359 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001360 } else
1361 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001362
Mel Gorman8191acb2013-10-07 11:28:45 +01001363 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001364clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001365 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001366 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001367 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001368 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001369 if (was_writable)
1370 pmd = pmd_mkwrite(pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001371 set_pmd_at(mm, haddr, pmdp, pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001372 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001373 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001374out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001375 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001376
1377out:
1378 if (anon_vma)
1379 page_unlock_anon_vma_read(anon_vma);
1380
Mel Gorman8191acb2013-10-07 11:28:45 +01001381 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001382 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001383
Mel Gormand10e63f2012-10-25 14:16:31 +02001384 return 0;
1385}
1386
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001387int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001388 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001389{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001390 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001391 int ret = 0;
1392
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001393 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001394 struct page *page;
1395 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001396 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001397 /*
1398 * For architectures like ppc64 we look at deposited pgtable
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001399 * when calling pmdp_huge_get_and_clear. So do the
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001400 * pgtable_trans_huge_withdraw after finishing pmdp related
1401 * operations.
1402 */
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001403 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1404 tlb->fullmm);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001405 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001406 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001407 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001408 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001409 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001410 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001411 } else {
1412 page = pmd_page(orig_pmd);
1413 page_remove_rmap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001414 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001415 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
Sasha Levin309381fea2014-01-23 15:52:54 -08001416 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001417 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001418 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001419 tlb_remove_page(tlb, page);
1420 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001421 pte_free(tlb->mm, pgtable);
1422 ret = 1;
1423 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001424 return ret;
1425}
1426
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001427int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1428 unsigned long old_addr,
1429 unsigned long new_addr, unsigned long old_end,
1430 pmd_t *old_pmd, pmd_t *new_pmd)
1431{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001432 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001433 int ret = 0;
1434 pmd_t pmd;
1435
1436 struct mm_struct *mm = vma->vm_mm;
1437
1438 if ((old_addr & ~HPAGE_PMD_MASK) ||
1439 (new_addr & ~HPAGE_PMD_MASK) ||
1440 old_end - old_addr < HPAGE_PMD_SIZE ||
1441 (new_vma->vm_flags & VM_NOHUGEPAGE))
1442 goto out;
1443
1444 /*
1445 * The destination pmd shouldn't be established, free_pgtables()
1446 * should have release it.
1447 */
1448 if (WARN_ON(!pmd_none(*new_pmd))) {
1449 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1450 goto out;
1451 }
1452
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001453 /*
1454 * We don't have to worry about the ordering of src and dst
1455 * ptlocks because exclusive mmap_sem prevents deadlock.
1456 */
1457 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001458 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001459 new_ptl = pmd_lockptr(mm, new_pmd);
1460 if (new_ptl != old_ptl)
1461 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001462 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001463 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001464
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301465 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1466 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001467 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1468 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001469 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301470 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1471 if (new_ptl != old_ptl)
1472 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001473 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001474 }
1475out:
1476 return ret;
1477}
1478
Mel Gormanf123d742013-10-07 11:28:49 +01001479/*
1480 * Returns
1481 * - 0 if PMD could not be locked
1482 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1483 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1484 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001485int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001486 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001487{
1488 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001489 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001490 int ret = 0;
1491
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001492 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001493 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001494 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001495 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001496
1497 /*
1498 * Avoid trapping faults against the zero page. The read-only
1499 * data is likely to be read-cached on the local CPU and
1500 * local/remote hits to the zero page are not interesting.
1501 */
1502 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1503 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001504 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001505 }
1506
Mel Gorman10c10452015-02-12 14:58:44 -08001507 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001508 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001509 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001510 if (preserve_write)
1511 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001512 ret = HPAGE_PMD_NR;
1513 set_pmd_at(mm, addr, pmd, entry);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001514 BUG_ON(!preserve_write && pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001515 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001516 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001517 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001518
1519 return ret;
1520}
1521
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001522/*
1523 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1524 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1525 *
1526 * Note that if it returns 1, this routine returns without unlocking page
1527 * table locks. So callers must unlock them.
1528 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001529int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1530 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001531{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001532 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001533 if (likely(pmd_trans_huge(*pmd))) {
1534 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001535 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001536 wait_split_huge_page(vma->anon_vma, pmd);
1537 return -1;
1538 } else {
1539 /* Thp mapped by 'pmd' is stable, so we can
1540 * handle it as it is. */
1541 return 1;
1542 }
1543 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001544 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001545 return 0;
1546}
1547
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001548/*
1549 * This function returns whether a given @page is mapped onto the @address
1550 * in the virtual space of @mm.
1551 *
1552 * When it's true, this function returns *pmd with holding the page table lock
1553 * and passing it back to the caller via @ptl.
1554 * If it's false, returns NULL without holding the page table lock.
1555 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001556pmd_t *page_check_address_pmd(struct page *page,
1557 struct mm_struct *mm,
1558 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001559 enum page_check_address_pmd_flag flag,
1560 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001561{
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001562 pgd_t *pgd;
1563 pud_t *pud;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001564 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001565
1566 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001567 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001568
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001569 pgd = pgd_offset(mm, address);
1570 if (!pgd_present(*pgd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001571 return NULL;
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001572 pud = pud_offset(pgd, address);
1573 if (!pud_present(*pud))
1574 return NULL;
1575 pmd = pmd_offset(pud, address);
1576
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001577 *ptl = pmd_lock(mm, pmd);
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001578 if (!pmd_present(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001579 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001580 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001581 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001582 /*
1583 * split_vma() may create temporary aliased mappings. There is
1584 * no risk as long as all huge pmd are found and have their
1585 * splitting bit set before __split_huge_page_refcount
1586 * runs. Finding the same huge pmd more than once during the
1587 * same rmap walk is not a problem.
1588 */
1589 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1590 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001591 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001592 if (pmd_trans_huge(*pmd)) {
1593 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1594 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001595 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001596 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001597unlock:
1598 spin_unlock(*ptl);
1599 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001600}
1601
1602static int __split_huge_page_splitting(struct page *page,
1603 struct vm_area_struct *vma,
1604 unsigned long address)
1605{
1606 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001607 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001608 pmd_t *pmd;
1609 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001610 /* For mmu_notifiers */
1611 const unsigned long mmun_start = address;
1612 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001613
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001614 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001615 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001616 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001617 if (pmd) {
1618 /*
1619 * We can't temporarily set the pmd to null in order
1620 * to split it, the pmd must remain marked huge at all
1621 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001622 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001623 * serialize against split_huge_page*.
1624 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001625 pmdp_splitting_flush(vma, address, pmd);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001626
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001627 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001628 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001629 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001630 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001631
1632 return ret;
1633}
1634
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001635static void __split_huge_page_refcount(struct page *page,
1636 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001637{
1638 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001639 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001640 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001641 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001642
1643 /* prevent PageLRU to go away from under us, and freeze lru stats */
1644 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001645 lruvec = mem_cgroup_page_lruvec(page, zone);
1646
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001647 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001648 /* complete memcg works before add pages to LRU */
1649 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001650
Shaohua Li45676882012-01-12 17:19:18 -08001651 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001652 struct page *page_tail = page + i;
1653
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001654 /* tail_page->_mapcount cannot change */
1655 BUG_ON(page_mapcount(page_tail) < 0);
1656 tail_count += page_mapcount(page_tail);
1657 /* check for overflow */
1658 BUG_ON(tail_count < 0);
1659 BUG_ON(atomic_read(&page_tail->_count) != 0);
1660 /*
1661 * tail_page->_count is zero and not changing from
1662 * under us. But get_page_unless_zero() may be running
1663 * from under us on the tail_page. If we used
1664 * atomic_set() below instead of atomic_add(), we
1665 * would then run atomic_set() concurrently with
1666 * get_page_unless_zero(), and atomic_set() is
1667 * implemented in C not using locked ops. spin_unlock
1668 * on x86 sometime uses locked ops because of PPro
1669 * errata 66, 92, so unless somebody can guarantee
1670 * atomic_set() here would be safe on all archs (and
1671 * not only on x86), it's safer to use atomic_add().
1672 */
1673 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1674 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001675
1676 /* after clearing PageTail the gup refcount can be released */
Waiman Long3a79d522014-08-06 16:05:38 -07001677 smp_mb__after_atomic();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001678
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001679 /*
1680 * retain hwpoison flag of the poisoned tail page:
1681 * fix for the unsuitable process killed on Guest Machine(KVM)
1682 * by the memory-failure.
1683 */
1684 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001685 page_tail->flags |= (page->flags &
1686 ((1L << PG_referenced) |
1687 (1L << PG_swapbacked) |
1688 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001689 (1L << PG_uptodate) |
1690 (1L << PG_active) |
1691 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001692 page_tail->flags |= (1L << PG_dirty);
1693
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001694 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001695 smp_wmb();
1696
1697 /*
1698 * __split_huge_page_splitting() already set the
1699 * splitting bit in all pmd that could map this
1700 * hugepage, that will ensure no CPU can alter the
1701 * mapcount on the head page. The mapcount is only
1702 * accounted in the head page and it has to be
1703 * transferred to all tail pages in the below code. So
1704 * for this code to be safe, the split the mapcount
1705 * can't change. But that doesn't mean userland can't
1706 * keep changing and reading the page contents while
1707 * we transfer the mapcount, so the pmd splitting
1708 * status is achieved setting a reserved bit in the
1709 * pmd, not by clearing the present bit.
1710 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001711 page_tail->_mapcount = page->_mapcount;
1712
1713 BUG_ON(page_tail->mapping);
1714 page_tail->mapping = page->mapping;
1715
Shaohua Li45676882012-01-12 17:19:18 -08001716 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001717 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001718
1719 BUG_ON(!PageAnon(page_tail));
1720 BUG_ON(!PageUptodate(page_tail));
1721 BUG_ON(!PageDirty(page_tail));
1722 BUG_ON(!PageSwapBacked(page_tail));
1723
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001724 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001725 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001726 atomic_sub(tail_count, &page->_count);
1727 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001728
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001729 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001730
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001731 ClearPageCompound(page);
1732 compound_unlock(page);
1733 spin_unlock_irq(&zone->lru_lock);
1734
1735 for (i = 1; i < HPAGE_PMD_NR; i++) {
1736 struct page *page_tail = page + i;
1737 BUG_ON(page_count(page_tail) <= 0);
1738 /*
1739 * Tail pages may be freed if there wasn't any mapping
1740 * like if add_to_swap() is running on a lru page that
1741 * had its mapping zapped. And freeing these pages
1742 * requires taking the lru_lock so we do the put_page
1743 * of the tail pages after the split is complete.
1744 */
1745 put_page(page_tail);
1746 }
1747
1748 /*
1749 * Only the head page (now become a regular page) is required
1750 * to be pinned by the caller.
1751 */
1752 BUG_ON(page_count(page) <= 0);
1753}
1754
1755static int __split_huge_page_map(struct page *page,
1756 struct vm_area_struct *vma,
1757 unsigned long address)
1758{
1759 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001760 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001761 pmd_t *pmd, _pmd;
1762 int ret = 0, i;
1763 pgtable_t pgtable;
1764 unsigned long haddr;
1765
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001766 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001767 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001768 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001769 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001770 pmd_populate(mm, &_pmd, pgtable);
Waiman Longf8303c22014-08-06 16:05:36 -07001771 if (pmd_write(*pmd))
1772 BUG_ON(page_mapcount(page) != 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001773
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001774 haddr = address;
1775 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001776 pte_t *pte, entry;
1777 BUG_ON(PageCompound(page+i));
Mel Gormanabc40bd2014-10-02 19:47:42 +01001778 /*
Mel Gorman8a0516e2015-02-12 14:58:22 -08001779 * Note that NUMA hinting access restrictions are not
1780 * transferred to avoid any possibility of altering
1781 * permissions across VMAs.
Mel Gormanabc40bd2014-10-02 19:47:42 +01001782 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001783 entry = mk_pte(page + i, vma->vm_page_prot);
1784 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1785 if (!pmd_write(*pmd))
1786 entry = pte_wrprotect(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001787 if (!pmd_young(*pmd))
1788 entry = pte_mkold(entry);
1789 pte = pte_offset_map(&_pmd, haddr);
1790 BUG_ON(!pte_none(*pte));
1791 set_pte_at(mm, haddr, pte, entry);
1792 pte_unmap(pte);
1793 }
1794
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001795 smp_wmb(); /* make pte visible before pmd */
1796 /*
1797 * Up to this point the pmd is present and huge and
1798 * userland has the whole access to the hugepage
1799 * during the split (which happens in place). If we
1800 * overwrite the pmd with the not-huge version
1801 * pointing to the pte here (which of course we could
1802 * if all CPUs were bug free), userland could trigger
1803 * a small page size TLB miss on the small sized TLB
1804 * while the hugepage TLB entry is still established
1805 * in the huge TLB. Some CPU doesn't like that. See
1806 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1807 * Erratum 383 on page 93. Intel should be safe but is
1808 * also warns that it's only safe if the permission
1809 * and cache attributes of the two entries loaded in
1810 * the two TLB is identical (which should be the case
1811 * here). But it is generally safer to never allow
1812 * small and huge TLB entries for the same virtual
1813 * address to be loaded simultaneously. So instead of
1814 * doing "pmd_populate(); flush_tlb_range();" we first
1815 * mark the current pmd notpresent (atomically because
1816 * here the pmd_trans_huge and pmd_trans_splitting
1817 * must remain set at all times on the pmd until the
1818 * split is complete for this pmd), then we flush the
1819 * SMP TLB and finally we write the non-huge version
1820 * of the pmd entry with pmd_populate.
1821 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001822 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001823 pmd_populate(mm, pmd, pgtable);
1824 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001825 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001826 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001827
1828 return ret;
1829}
1830
Ingo Molnar5a505082012-12-02 19:56:46 +00001831/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001832static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001833 struct anon_vma *anon_vma,
1834 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001835{
1836 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001837 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001838 struct anon_vma_chain *avc;
1839
1840 BUG_ON(!PageHead(page));
1841 BUG_ON(PageTail(page));
1842
1843 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001844 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001845 struct vm_area_struct *vma = avc->vma;
1846 unsigned long addr = vma_address(page, vma);
1847 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001848 mapcount += __split_huge_page_splitting(page, vma, addr);
1849 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001850 /*
1851 * It is critical that new vmas are added to the tail of the
1852 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1853 * and establishes a child pmd before
1854 * __split_huge_page_splitting() freezes the parent pmd (so if
1855 * we fail to prevent copy_huge_pmd() from running until the
1856 * whole __split_huge_page() is complete), we will still see
1857 * the newly established pmd of the child later during the
1858 * walk, to be able to set it as pmd_trans_splitting too.
1859 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001860 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001861 pr_err("mapcount %d page_mapcount %d\n",
1862 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001863 BUG();
1864 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001865
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001866 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001867
1868 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001869 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001870 struct vm_area_struct *vma = avc->vma;
1871 unsigned long addr = vma_address(page, vma);
1872 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001873 mapcount2 += __split_huge_page_map(page, vma, addr);
1874 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001875 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001876 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1877 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001878 BUG();
1879 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001880}
1881
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001882/*
1883 * Split a hugepage into normal pages. This doesn't change the position of head
1884 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1885 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1886 * from the hugepage.
1887 * Return 0 if the hugepage is split successfully otherwise return 1.
1888 */
1889int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001890{
1891 struct anon_vma *anon_vma;
1892 int ret = 1;
1893
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001894 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001895 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001896
1897 /*
1898 * The caller does not necessarily hold an mmap_sem that would prevent
1899 * the anon_vma disappearing so we first we take a reference to it
1900 * and then lock the anon_vma for write. This is similar to
1901 * page_lock_anon_vma_read except the write lock is taken to serialise
1902 * against parallel split or collapse operations.
1903 */
1904 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001905 if (!anon_vma)
1906 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001907 anon_vma_lock_write(anon_vma);
1908
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001909 ret = 0;
1910 if (!PageCompound(page))
1911 goto out_unlock;
1912
1913 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001914 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001915 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001916
1917 BUG_ON(PageCompound(page));
1918out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001919 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001920 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001921out:
1922 return ret;
1923}
1924
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001925#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001926
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001927int hugepage_madvise(struct vm_area_struct *vma,
1928 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001929{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001930 switch (advice) {
1931 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001932#ifdef CONFIG_S390
1933 /*
1934 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1935 * can't handle this properly after s390_enable_sie, so we simply
1936 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1937 */
1938 if (mm_has_pgste(vma->vm_mm))
1939 return 0;
1940#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001941 /*
1942 * Be somewhat over-protective like KSM for now!
1943 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001944 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001945 return -EINVAL;
1946 *vm_flags &= ~VM_NOHUGEPAGE;
1947 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001948 /*
1949 * If the vma become good for khugepaged to scan,
1950 * register it here without waiting a page fault that
1951 * may not happen any time soon.
1952 */
David Rientjes6d50e602014-10-29 14:50:31 -07001953 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001954 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001955 break;
1956 case MADV_NOHUGEPAGE:
1957 /*
1958 * Be somewhat over-protective like KSM for now!
1959 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001960 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001961 return -EINVAL;
1962 *vm_flags &= ~VM_HUGEPAGE;
1963 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001964 /*
1965 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1966 * this vma even if we leave the mm registered in khugepaged if
1967 * it got registered before VM_NOHUGEPAGE was set.
1968 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001969 break;
1970 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001971
1972 return 0;
1973}
1974
Andrea Arcangeliba761492011-01-13 15:46:58 -08001975static int __init khugepaged_slab_init(void)
1976{
1977 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1978 sizeof(struct mm_slot),
1979 __alignof__(struct mm_slot), 0, NULL);
1980 if (!mm_slot_cache)
1981 return -ENOMEM;
1982
1983 return 0;
1984}
1985
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -07001986static void __init khugepaged_slab_exit(void)
1987{
1988 kmem_cache_destroy(mm_slot_cache);
1989}
1990
Andrea Arcangeliba761492011-01-13 15:46:58 -08001991static inline struct mm_slot *alloc_mm_slot(void)
1992{
1993 if (!mm_slot_cache) /* initialization failed */
1994 return NULL;
1995 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1996}
1997
1998static inline void free_mm_slot(struct mm_slot *mm_slot)
1999{
2000 kmem_cache_free(mm_slot_cache, mm_slot);
2001}
2002
Andrea Arcangeliba761492011-01-13 15:46:58 -08002003static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2004{
2005 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002006
Sasha Levinb67bfe02013-02-27 17:06:00 -08002007 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002008 if (mm == mm_slot->mm)
2009 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002010
Andrea Arcangeliba761492011-01-13 15:46:58 -08002011 return NULL;
2012}
2013
2014static void insert_to_mm_slots_hash(struct mm_struct *mm,
2015 struct mm_slot *mm_slot)
2016{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002017 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002018 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002019}
2020
2021static inline int khugepaged_test_exit(struct mm_struct *mm)
2022{
2023 return atomic_read(&mm->mm_users) == 0;
2024}
2025
2026int __khugepaged_enter(struct mm_struct *mm)
2027{
2028 struct mm_slot *mm_slot;
2029 int wakeup;
2030
2031 mm_slot = alloc_mm_slot();
2032 if (!mm_slot)
2033 return -ENOMEM;
2034
2035 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07002036 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002037 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2038 free_mm_slot(mm_slot);
2039 return 0;
2040 }
2041
2042 spin_lock(&khugepaged_mm_lock);
2043 insert_to_mm_slots_hash(mm, mm_slot);
2044 /*
2045 * Insert just behind the scanning cursor, to let the area settle
2046 * down a little.
2047 */
2048 wakeup = list_empty(&khugepaged_scan.mm_head);
2049 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2050 spin_unlock(&khugepaged_mm_lock);
2051
2052 atomic_inc(&mm->mm_count);
2053 if (wakeup)
2054 wake_up_interruptible(&khugepaged_wait);
2055
2056 return 0;
2057}
2058
David Rientjes6d50e602014-10-29 14:50:31 -07002059int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2060 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002061{
2062 unsigned long hstart, hend;
2063 if (!vma->anon_vma)
2064 /*
2065 * Not yet faulted in so we will register later in the
2066 * page fault if needed.
2067 */
2068 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002069 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002070 /* khugepaged not yet working on file or special mappings */
2071 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07002072 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002073 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2074 hend = vma->vm_end & HPAGE_PMD_MASK;
2075 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07002076 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077 return 0;
2078}
2079
2080void __khugepaged_exit(struct mm_struct *mm)
2081{
2082 struct mm_slot *mm_slot;
2083 int free = 0;
2084
2085 spin_lock(&khugepaged_mm_lock);
2086 mm_slot = get_mm_slot(mm);
2087 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002088 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002089 list_del(&mm_slot->mm_node);
2090 free = 1;
2091 }
Chris Wrightd788e802011-07-25 17:12:14 -07002092 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002093
2094 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002095 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2096 free_mm_slot(mm_slot);
2097 mmdrop(mm);
2098 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002099 /*
2100 * This is required to serialize against
2101 * khugepaged_test_exit() (which is guaranteed to run
2102 * under mmap sem read mode). Stop here (after we
2103 * return all pagetables will be destroyed) until
2104 * khugepaged has finished working on the pagetables
2105 * under the mmap_sem.
2106 */
2107 down_write(&mm->mmap_sem);
2108 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002109 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002110}
2111
2112static void release_pte_page(struct page *page)
2113{
2114 /* 0 stands for page_is_file_cache(page) == false */
2115 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2116 unlock_page(page);
2117 putback_lru_page(page);
2118}
2119
2120static void release_pte_pages(pte_t *pte, pte_t *_pte)
2121{
2122 while (--_pte >= pte) {
2123 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002124 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002125 release_pte_page(pte_page(pteval));
2126 }
2127}
2128
Andrea Arcangeliba761492011-01-13 15:46:58 -08002129static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2130 unsigned long address,
2131 pte_t *pte)
2132{
2133 struct page *page;
2134 pte_t *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002135 int none_or_zero = 0;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002136 bool referenced = false, writable = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002137 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2138 _pte++, address += PAGE_SIZE) {
2139 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002140 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2141 if (++none_or_zero <= khugepaged_max_ptes_none)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002142 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002143 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002144 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002145 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002146 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002147 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002148 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002149 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002150 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002151
Sasha Levin309381fea2014-01-23 15:52:54 -08002152 VM_BUG_ON_PAGE(PageCompound(page), page);
2153 VM_BUG_ON_PAGE(!PageAnon(page), page);
2154 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002155
Andrea Arcangeliba761492011-01-13 15:46:58 -08002156 /*
2157 * We can do it before isolate_lru_page because the
2158 * page can't be freed from under us. NOTE: PG_lock
2159 * is needed to serialize against split_huge_page
2160 * when invoked from the VM.
2161 */
Bob Liu344aa352012-12-11 16:00:34 -08002162 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002163 goto out;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002164
2165 /*
2166 * cannot use mapcount: can't collapse if there's a gup pin.
2167 * The page must only be referenced by the scanned process
2168 * and page swap cache.
2169 */
2170 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2171 unlock_page(page);
2172 goto out;
2173 }
2174 if (pte_write(pteval)) {
2175 writable = true;
2176 } else {
2177 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2178 unlock_page(page);
2179 goto out;
2180 }
2181 /*
2182 * Page is not in the swap cache. It can be collapsed
2183 * into a THP.
2184 */
2185 }
2186
Andrea Arcangeliba761492011-01-13 15:46:58 -08002187 /*
2188 * Isolate the page to avoid collapsing an hugepage
2189 * currently in use by the VM.
2190 */
2191 if (isolate_lru_page(page)) {
2192 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002193 goto out;
2194 }
2195 /* 0 stands for page_is_file_cache(page) == false */
2196 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002197 VM_BUG_ON_PAGE(!PageLocked(page), page);
2198 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002199
2200 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002201 if (pte_young(pteval) || PageReferenced(page) ||
2202 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002203 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002204 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002205 if (likely(referenced && writable))
Bob Liu344aa352012-12-11 16:00:34 -08002206 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002207out:
Bob Liu344aa352012-12-11 16:00:34 -08002208 release_pte_pages(pte, _pte);
2209 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002210}
2211
2212static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2213 struct vm_area_struct *vma,
2214 unsigned long address,
2215 spinlock_t *ptl)
2216{
2217 pte_t *_pte;
2218 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2219 pte_t pteval = *_pte;
2220 struct page *src_page;
2221
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002222 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002223 clear_user_highpage(page, address);
2224 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002225 if (is_zero_pfn(pte_pfn(pteval))) {
2226 /*
2227 * ptl mostly unnecessary.
2228 */
2229 spin_lock(ptl);
2230 /*
2231 * paravirt calls inside pte_clear here are
2232 * superfluous.
2233 */
2234 pte_clear(vma->vm_mm, address, _pte);
2235 spin_unlock(ptl);
2236 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002237 } else {
2238 src_page = pte_page(pteval);
2239 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002240 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002241 release_pte_page(src_page);
2242 /*
2243 * ptl mostly unnecessary, but preempt has to
2244 * be disabled to update the per-cpu stats
2245 * inside page_remove_rmap().
2246 */
2247 spin_lock(ptl);
2248 /*
2249 * paravirt calls inside pte_clear here are
2250 * superfluous.
2251 */
2252 pte_clear(vma->vm_mm, address, _pte);
2253 page_remove_rmap(src_page);
2254 spin_unlock(ptl);
2255 free_page_and_swap_cache(src_page);
2256 }
2257
2258 address += PAGE_SIZE;
2259 page++;
2260 }
2261}
2262
Xiao Guangrong26234f32012-10-08 16:29:51 -07002263static void khugepaged_alloc_sleep(void)
2264{
2265 wait_event_freezable_timeout(khugepaged_wait, false,
2266 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2267}
2268
Bob Liu9f1b8682013-11-12 15:07:37 -08002269static int khugepaged_node_load[MAX_NUMNODES];
2270
David Rientjes14a4e212014-08-06 16:07:29 -07002271static bool khugepaged_scan_abort(int nid)
2272{
2273 int i;
2274
2275 /*
2276 * If zone_reclaim_mode is disabled, then no extra effort is made to
2277 * allocate memory locally.
2278 */
2279 if (!zone_reclaim_mode)
2280 return false;
2281
2282 /* If there is a count for this node already, it must be acceptable */
2283 if (khugepaged_node_load[nid])
2284 return false;
2285
2286 for (i = 0; i < MAX_NUMNODES; i++) {
2287 if (!khugepaged_node_load[i])
2288 continue;
2289 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2290 return true;
2291 }
2292 return false;
2293}
2294
Xiao Guangrong26234f32012-10-08 16:29:51 -07002295#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002296static int khugepaged_find_target_node(void)
2297{
2298 static int last_khugepaged_target_node = NUMA_NO_NODE;
2299 int nid, target_node = 0, max_value = 0;
2300
2301 /* find first node with max normal pages hit */
2302 for (nid = 0; nid < MAX_NUMNODES; nid++)
2303 if (khugepaged_node_load[nid] > max_value) {
2304 max_value = khugepaged_node_load[nid];
2305 target_node = nid;
2306 }
2307
2308 /* do some balance if several nodes have the same hit record */
2309 if (target_node <= last_khugepaged_target_node)
2310 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2311 nid++)
2312 if (max_value == khugepaged_node_load[nid]) {
2313 target_node = nid;
2314 break;
2315 }
2316
2317 last_khugepaged_target_node = target_node;
2318 return target_node;
2319}
2320
Xiao Guangrong26234f32012-10-08 16:29:51 -07002321static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2322{
2323 if (IS_ERR(*hpage)) {
2324 if (!*wait)
2325 return false;
2326
2327 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002328 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002329 khugepaged_alloc_sleep();
2330 } else if (*hpage) {
2331 put_page(*hpage);
2332 *hpage = NULL;
2333 }
2334
2335 return true;
2336}
2337
Michal Hocko3b363692015-04-15 16:13:29 -07002338static struct page *
2339khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002340 struct vm_area_struct *vma, unsigned long address,
2341 int node)
2342{
Sasha Levin309381fea2014-01-23 15:52:54 -08002343 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002344
Xiao Guangrong26234f32012-10-08 16:29:51 -07002345 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002346 * Before allocating the hugepage, release the mmap_sem read lock.
2347 * The allocation can take potentially a long time if it involves
2348 * sync compaction, and we do not need to hold the mmap_sem during
2349 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002350 */
2351 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002352
Michal Hocko3b363692015-04-15 16:13:29 -07002353 *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002354 if (unlikely(!*hpage)) {
2355 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2356 *hpage = ERR_PTR(-ENOMEM);
2357 return NULL;
2358 }
2359
2360 count_vm_event(THP_COLLAPSE_ALLOC);
2361 return *hpage;
2362}
2363#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002364static int khugepaged_find_target_node(void)
2365{
2366 return 0;
2367}
2368
Bob Liu10dc4152013-11-12 15:07:35 -08002369static inline struct page *alloc_hugepage(int defrag)
2370{
2371 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2372 HPAGE_PMD_ORDER);
2373}
2374
Xiao Guangrong26234f32012-10-08 16:29:51 -07002375static struct page *khugepaged_alloc_hugepage(bool *wait)
2376{
2377 struct page *hpage;
2378
2379 do {
2380 hpage = alloc_hugepage(khugepaged_defrag());
2381 if (!hpage) {
2382 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2383 if (!*wait)
2384 return NULL;
2385
2386 *wait = false;
2387 khugepaged_alloc_sleep();
2388 } else
2389 count_vm_event(THP_COLLAPSE_ALLOC);
2390 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2391
2392 return hpage;
2393}
2394
2395static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2396{
2397 if (!*hpage)
2398 *hpage = khugepaged_alloc_hugepage(wait);
2399
2400 if (unlikely(!*hpage))
2401 return false;
2402
2403 return true;
2404}
2405
Michal Hocko3b363692015-04-15 16:13:29 -07002406static struct page *
2407khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002408 struct vm_area_struct *vma, unsigned long address,
2409 int node)
2410{
2411 up_read(&mm->mmap_sem);
2412 VM_BUG_ON(!*hpage);
Michal Hocko3b363692015-04-15 16:13:29 -07002413
Xiao Guangrong26234f32012-10-08 16:29:51 -07002414 return *hpage;
2415}
2416#endif
2417
Bob Liufa475e52012-12-11 16:00:39 -08002418static bool hugepage_vma_check(struct vm_area_struct *vma)
2419{
2420 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2421 (vma->vm_flags & VM_NOHUGEPAGE))
2422 return false;
2423
2424 if (!vma->anon_vma || vma->vm_ops)
2425 return false;
2426 if (is_vma_temporary_stack(vma))
2427 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002428 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002429 return true;
2430}
2431
Andrea Arcangeliba761492011-01-13 15:46:58 -08002432static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002433 unsigned long address,
2434 struct page **hpage,
2435 struct vm_area_struct *vma,
2436 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002437{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002438 pmd_t *pmd, _pmd;
2439 pte_t *pte;
2440 pgtable_t pgtable;
2441 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002442 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002443 int isolated;
2444 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002445 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002446 unsigned long mmun_start; /* For mmu_notifiers */
2447 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07002448 gfp_t gfp;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002449
2450 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002451
Michal Hocko3b363692015-04-15 16:13:29 -07002452 /* Only allocate from the target node */
2453 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2454 __GFP_THISNODE;
2455
Xiao Guangrong26234f32012-10-08 16:29:51 -07002456 /* release the mmap_sem read lock. */
Michal Hocko3b363692015-04-15 16:13:29 -07002457 new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002458 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002459 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002460
Johannes Weiner00501b52014-08-08 14:19:20 -07002461 if (unlikely(mem_cgroup_try_charge(new_page, mm,
Michal Hocko3b363692015-04-15 16:13:29 -07002462 gfp, &memcg)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002463 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002464
2465 /*
2466 * Prevent all access to pagetables with the exception of
2467 * gup_fast later hanlded by the ptep_clear_flush and the VM
2468 * handled by the anon_vma lock + PG_lock.
2469 */
2470 down_write(&mm->mmap_sem);
2471 if (unlikely(khugepaged_test_exit(mm)))
2472 goto out;
2473
2474 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002475 if (!vma)
2476 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002477 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2478 hend = vma->vm_end & HPAGE_PMD_MASK;
2479 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2480 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002481 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002482 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002483 pmd = mm_find_pmd(mm, address);
2484 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002485 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002486
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002487 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002488
2489 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002490 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002491
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002492 mmun_start = address;
2493 mmun_end = address + HPAGE_PMD_SIZE;
2494 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002495 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002496 /*
2497 * After this gup_fast can't run anymore. This also removes
2498 * any huge TLB entry from the CPU so we won't allow
2499 * huge and small TLB entries for the same virtual address
2500 * to avoid the risk of CPU bugs in that area.
2501 */
Aneesh Kumar K.V15a25b22015-06-24 16:57:39 -07002502 _pmd = pmdp_collapse_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002503 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002504 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002505
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002506 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002507 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002508 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002509
2510 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002511 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002512 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002513 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002514 /*
2515 * We can only use set_pmd_at when establishing
2516 * hugepmds and never for establishing regular pmds that
2517 * points to regular pagetables. Use pmd_populate for that
2518 */
2519 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002520 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002521 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002522 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002523 }
2524
2525 /*
2526 * All pages are isolated and locked so anon_vma rmap
2527 * can't run anymore.
2528 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002529 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002530
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002531 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002532 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002533 __SetPageUptodate(new_page);
2534 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002535
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002536 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2537 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002538
2539 /*
2540 * spin_lock() below is not the equivalent of smp_wmb(), so
2541 * this is needed to avoid the copy_huge_page writes to become
2542 * visible after the set_pmd_at() write.
2543 */
2544 smp_wmb();
2545
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002546 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002547 BUG_ON(!pmd_none(*pmd));
2548 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002549 mem_cgroup_commit_charge(new_page, memcg, false);
2550 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002551 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002552 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002553 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002554 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002555
2556 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002557
Andrea Arcangeliba761492011-01-13 15:46:58 -08002558 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002559out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002560 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002561 return;
2562
Andrea Arcangelice83d212011-01-13 15:47:06 -08002563out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002564 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002565 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002566}
2567
2568static int khugepaged_scan_pmd(struct mm_struct *mm,
2569 struct vm_area_struct *vma,
2570 unsigned long address,
2571 struct page **hpage)
2572{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002573 pmd_t *pmd;
2574 pte_t *pte, *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002575 int ret = 0, none_or_zero = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002576 struct page *page;
2577 unsigned long _address;
2578 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002579 int node = NUMA_NO_NODE;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002580 bool writable = false, referenced = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002581
2582 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2583
Bob Liu62190492012-12-11 16:00:37 -08002584 pmd = mm_find_pmd(mm, address);
2585 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002586 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002587
Bob Liu9f1b8682013-11-12 15:07:37 -08002588 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002589 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2590 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2591 _pte++, _address += PAGE_SIZE) {
2592 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002593 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2594 if (++none_or_zero <= khugepaged_max_ptes_none)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002595 continue;
2596 else
2597 goto out_unmap;
2598 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002599 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002600 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002601 if (pte_write(pteval))
2602 writable = true;
2603
Andrea Arcangeliba761492011-01-13 15:46:58 -08002604 page = vm_normal_page(vma, _address, pteval);
2605 if (unlikely(!page))
2606 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002607 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002608 * Record which node the original page is from and save this
2609 * information to khugepaged_node_load[].
2610 * Khupaged will allocate hugepage from the node has the max
2611 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002612 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002613 node = page_to_nid(page);
David Rientjes14a4e212014-08-06 16:07:29 -07002614 if (khugepaged_scan_abort(node))
2615 goto out_unmap;
Bob Liu9f1b8682013-11-12 15:07:37 -08002616 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002617 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002618 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2619 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002620 /*
2621 * cannot use mapcount: can't collapse if there's a gup pin.
2622 * The page must only be referenced by the scanned process
2623 * and page swap cache.
2624 */
2625 if (page_count(page) != 1 + !!PageSwapCache(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002626 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002627 if (pte_young(pteval) || PageReferenced(page) ||
2628 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002629 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002630 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002631 if (referenced && writable)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002632 ret = 1;
2633out_unmap:
2634 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002635 if (ret) {
2636 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002637 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002638 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002639 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002640out:
2641 return ret;
2642}
2643
2644static void collect_mm_slot(struct mm_slot *mm_slot)
2645{
2646 struct mm_struct *mm = mm_slot->mm;
2647
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_test_exit(mm)) {
2651 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002652 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002653 list_del(&mm_slot->mm_node);
2654
2655 /*
2656 * Not strictly needed because the mm exited already.
2657 *
2658 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2659 */
2660
2661 /* khugepaged_mm_lock actually not necessary for the below */
2662 free_mm_slot(mm_slot);
2663 mmdrop(mm);
2664 }
2665}
2666
2667static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2668 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002669 __releases(&khugepaged_mm_lock)
2670 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002671{
2672 struct mm_slot *mm_slot;
2673 struct mm_struct *mm;
2674 struct vm_area_struct *vma;
2675 int progress = 0;
2676
2677 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002678 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002679
2680 if (khugepaged_scan.mm_slot)
2681 mm_slot = khugepaged_scan.mm_slot;
2682 else {
2683 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2684 struct mm_slot, mm_node);
2685 khugepaged_scan.address = 0;
2686 khugepaged_scan.mm_slot = mm_slot;
2687 }
2688 spin_unlock(&khugepaged_mm_lock);
2689
2690 mm = mm_slot->mm;
2691 down_read(&mm->mmap_sem);
2692 if (unlikely(khugepaged_test_exit(mm)))
2693 vma = NULL;
2694 else
2695 vma = find_vma(mm, khugepaged_scan.address);
2696
2697 progress++;
2698 for (; vma; vma = vma->vm_next) {
2699 unsigned long hstart, hend;
2700
2701 cond_resched();
2702 if (unlikely(khugepaged_test_exit(mm))) {
2703 progress++;
2704 break;
2705 }
Bob Liufa475e52012-12-11 16:00:39 -08002706 if (!hugepage_vma_check(vma)) {
2707skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002708 progress++;
2709 continue;
2710 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002711 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2712 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002713 if (hstart >= hend)
2714 goto skip;
2715 if (khugepaged_scan.address > hend)
2716 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002717 if (khugepaged_scan.address < hstart)
2718 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002719 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002720
2721 while (khugepaged_scan.address < hend) {
2722 int ret;
2723 cond_resched();
2724 if (unlikely(khugepaged_test_exit(mm)))
2725 goto breakouterloop;
2726
2727 VM_BUG_ON(khugepaged_scan.address < hstart ||
2728 khugepaged_scan.address + HPAGE_PMD_SIZE >
2729 hend);
2730 ret = khugepaged_scan_pmd(mm, vma,
2731 khugepaged_scan.address,
2732 hpage);
2733 /* move to next address */
2734 khugepaged_scan.address += HPAGE_PMD_SIZE;
2735 progress += HPAGE_PMD_NR;
2736 if (ret)
2737 /* we released mmap_sem so break loop */
2738 goto breakouterloop_mmap_sem;
2739 if (progress >= pages)
2740 goto breakouterloop;
2741 }
2742 }
2743breakouterloop:
2744 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2745breakouterloop_mmap_sem:
2746
2747 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002748 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002749 /*
2750 * Release the current mm_slot if this mm is about to die, or
2751 * if we scanned all vmas of this mm.
2752 */
2753 if (khugepaged_test_exit(mm) || !vma) {
2754 /*
2755 * Make sure that if mm_users is reaching zero while
2756 * khugepaged runs here, khugepaged_exit will find
2757 * mm_slot not pointing to the exiting mm.
2758 */
2759 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2760 khugepaged_scan.mm_slot = list_entry(
2761 mm_slot->mm_node.next,
2762 struct mm_slot, mm_node);
2763 khugepaged_scan.address = 0;
2764 } else {
2765 khugepaged_scan.mm_slot = NULL;
2766 khugepaged_full_scans++;
2767 }
2768
2769 collect_mm_slot(mm_slot);
2770 }
2771
2772 return progress;
2773}
2774
2775static int khugepaged_has_work(void)
2776{
2777 return !list_empty(&khugepaged_scan.mm_head) &&
2778 khugepaged_enabled();
2779}
2780
2781static int khugepaged_wait_event(void)
2782{
2783 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002784 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002785}
2786
Xiao Guangrongd5169042012-10-08 16:29:48 -07002787static void khugepaged_do_scan(void)
2788{
2789 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002790 unsigned int progress = 0, pass_through_head = 0;
2791 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002792 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002793
2794 barrier(); /* write khugepaged_pages_to_scan to local stack */
2795
2796 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002797 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002798 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002799
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002800 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002801
Jiri Kosinacd092412015-06-24 16:56:07 -07002802 if (unlikely(kthread_should_stop() || try_to_freeze()))
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002803 break;
2804
Andrea Arcangeliba761492011-01-13 15:46:58 -08002805 spin_lock(&khugepaged_mm_lock);
2806 if (!khugepaged_scan.mm_slot)
2807 pass_through_head++;
2808 if (khugepaged_has_work() &&
2809 pass_through_head < 2)
2810 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002811 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002812 else
2813 progress = pages;
2814 spin_unlock(&khugepaged_mm_lock);
2815 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002816
Xiao Guangrongd5169042012-10-08 16:29:48 -07002817 if (!IS_ERR_OR_NULL(hpage))
2818 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002819}
2820
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002821static void khugepaged_wait_work(void)
2822{
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002823 if (khugepaged_has_work()) {
2824 if (!khugepaged_scan_sleep_millisecs)
2825 return;
2826
2827 wait_event_freezable_timeout(khugepaged_wait,
2828 kthread_should_stop(),
2829 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2830 return;
2831 }
2832
2833 if (khugepaged_enabled())
2834 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2835}
2836
Andrea Arcangeliba761492011-01-13 15:46:58 -08002837static int khugepaged(void *none)
2838{
2839 struct mm_slot *mm_slot;
2840
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002841 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002842 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002843
Xiao Guangrongb7231782012-10-08 16:29:54 -07002844 while (!kthread_should_stop()) {
2845 khugepaged_do_scan();
2846 khugepaged_wait_work();
2847 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002848
2849 spin_lock(&khugepaged_mm_lock);
2850 mm_slot = khugepaged_scan.mm_slot;
2851 khugepaged_scan.mm_slot = NULL;
2852 if (mm_slot)
2853 collect_mm_slot(mm_slot);
2854 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002855 return 0;
2856}
2857
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002858static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2859 unsigned long haddr, pmd_t *pmd)
2860{
2861 struct mm_struct *mm = vma->vm_mm;
2862 pgtable_t pgtable;
2863 pmd_t _pmd;
2864 int i;
2865
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07002866 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002867 /* leave pmd empty until pte is filled */
2868
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002869 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002870 pmd_populate(mm, &_pmd, pgtable);
2871
2872 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2873 pte_t *pte, entry;
2874 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2875 entry = pte_mkspecial(entry);
2876 pte = pte_offset_map(&_pmd, haddr);
2877 VM_BUG_ON(!pte_none(*pte));
2878 set_pte_at(mm, haddr, pte, entry);
2879 pte_unmap(pte);
2880 }
2881 smp_wmb(); /* make pte visible before pmd */
2882 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002883 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002884}
2885
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002886void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2887 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002888{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002889 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002890 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002891 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002892 unsigned long haddr = address & HPAGE_PMD_MASK;
2893 unsigned long mmun_start; /* For mmu_notifiers */
2894 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002895
2896 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002897
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002898 mmun_start = haddr;
2899 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002900again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002901 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002902 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002903 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002904 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002905 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2906 return;
2907 }
2908 if (is_huge_zero_pmd(*pmd)) {
2909 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002910 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002911 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002912 return;
2913 }
2914 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002915 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002916 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002917 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002918 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002919
2920 split_huge_page(page);
2921
2922 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002923
2924 /*
2925 * We don't always have down_write of mmap_sem here: a racing
2926 * do_huge_pmd_wp_page() might have copied-on-write to another
2927 * huge page before our split_huge_page() got the anon_vma lock.
2928 */
2929 if (unlikely(pmd_trans_huge(*pmd)))
2930 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002931}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002932
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002933void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2934 pmd_t *pmd)
2935{
2936 struct vm_area_struct *vma;
2937
2938 vma = find_vma(mm, address);
2939 BUG_ON(vma == NULL);
2940 split_huge_page_pmd(vma, address, pmd);
2941}
2942
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002943static void split_huge_page_address(struct mm_struct *mm,
2944 unsigned long address)
2945{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002946 pgd_t *pgd;
2947 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002948 pmd_t *pmd;
2949
2950 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2951
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002952 pgd = pgd_offset(mm, address);
2953 if (!pgd_present(*pgd))
2954 return;
2955
2956 pud = pud_offset(pgd, address);
2957 if (!pud_present(*pud))
2958 return;
2959
2960 pmd = pmd_offset(pud, address);
2961 if (!pmd_present(*pmd))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002962 return;
2963 /*
2964 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2965 * materialize from under us.
2966 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002967 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002968}
2969
2970void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2971 unsigned long start,
2972 unsigned long end,
2973 long adjust_next)
2974{
2975 /*
2976 * If the new start address isn't hpage aligned and it could
2977 * previously contain an hugepage: check if we need to split
2978 * an huge pmd.
2979 */
2980 if (start & ~HPAGE_PMD_MASK &&
2981 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2982 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2983 split_huge_page_address(vma->vm_mm, start);
2984
2985 /*
2986 * If the new end address isn't hpage aligned and it could
2987 * previously contain an hugepage: check if we need to split
2988 * an huge pmd.
2989 */
2990 if (end & ~HPAGE_PMD_MASK &&
2991 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2992 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2993 split_huge_page_address(vma->vm_mm, end);
2994
2995 /*
2996 * If we're also updating the vma->vm_next->vm_start, if the new
2997 * vm_next->vm_start isn't page aligned and it could previously
2998 * contain an hugepage: check if we need to split an huge pmd.
2999 */
3000 if (adjust_next > 0) {
3001 struct vm_area_struct *next = vma->vm_next;
3002 unsigned long nstart = next->vm_start;
3003 nstart += adjust_next << PAGE_SHIFT;
3004 if (nstart & ~HPAGE_PMD_MASK &&
3005 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3006 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3007 split_huge_page_address(next->vm_mm, nstart);
3008 }
3009}