blob: a7bac4f2b78aa62e750f6282dd6215b54a983dd3 [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>
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -080019#include <linux/swapops.h>
Matthew Wilcox4897c762015-09-08 14:58:45 -070020#include <linux/dax.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080021#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080022#include <linux/freezer.h>
Dan Williamsf25748e32016-01-15 16:56:43 -080023#include <linux/pfn_t.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080024#include <linux/mman.h>
Dan Williams3565fce2016-01-15 16:56:55 -080025#include <linux/memremap.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020026#include <linux/pagemap.h>
Kirill A. Shutemov49071d42016-01-15 16:54:40 -080027#include <linux/debugfs.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000028#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080029#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070030#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070031#include <linux/page_idle.h>
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -070032#include <linux/shmem_fs.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080033
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080034#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
Andrea Arcangeliba761492011-01-13 15:46:58 -080038/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080039 * By default transparent hugepage support is disabled in order that avoid
40 * to risk increase the memory footprint of applications without a guaranteed
41 * benefit. When transparent hugepage support is enabled, is for all mappings,
42 * and khugepaged scans all mappings.
43 * Defrag is invoked by khugepaged hugepage allocations and by page faults
44 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080045 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080046unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080047#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080048 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080049#endif
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070053 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080054 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080056
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080057static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080058
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080059static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080060struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080061
Aaron Lu6fcb52a2016-10-07 17:00:08 -070062static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080063{
64 struct page *zero_page;
65retry:
66 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070067 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080068
69 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080071 if (!zero_page) {
72 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070073 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080074 }
75 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080076 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070077 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080078 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070079 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080080 goto retry;
81 }
82
83 /* We take additional reference here. It will be put back by shrinker */
84 atomic_set(&huge_zero_refcount, 2);
85 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070086 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080087}
88
Aaron Lu6fcb52a2016-10-07 17:00:08 -070089static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080090{
91 /*
92 * Counter should never go to zero here. Only shrinker can put
93 * last reference.
94 */
95 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96}
97
Aaron Lu6fcb52a2016-10-07 17:00:08 -070098struct page *mm_get_huge_zero_page(struct mm_struct *mm)
99{
100 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
101 return READ_ONCE(huge_zero_page);
102
103 if (!get_huge_zero_page())
104 return NULL;
105
106 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
107 put_huge_zero_page();
108
109 return READ_ONCE(huge_zero_page);
110}
111
112void mm_put_huge_zero_page(struct mm_struct *mm)
113{
114 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 put_huge_zero_page();
116}
117
Glauber Costa48896462013-08-28 10:18:15 +1000118static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
119 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800120{
Glauber Costa48896462013-08-28 10:18:15 +1000121 /* we can free zero page only if last reference remains */
122 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
123}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800124
Glauber Costa48896462013-08-28 10:18:15 +1000125static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
126 struct shrink_control *sc)
127{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800128 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700129 struct page *zero_page = xchg(&huge_zero_page, NULL);
130 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700131 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000132 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800133 }
134
135 return 0;
136}
137
138static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000139 .count_objects = shrink_huge_zero_page_count,
140 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800141 .seeks = DEFAULT_SEEKS,
142};
143
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800144#ifdef CONFIG_SYSFS
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800145static ssize_t enabled_show(struct kobject *kobj,
146 struct kobj_attribute *attr, char *buf)
147{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700148 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
149 return sprintf(buf, "[always] madvise never\n");
150 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
151 return sprintf(buf, "always [madvise] never\n");
152 else
153 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800154}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700155
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800156static ssize_t enabled_store(struct kobject *kobj,
157 struct kobj_attribute *attr,
158 const char *buf, size_t count)
159{
David Rientjes21440d72017-02-22 15:45:49 -0800160 ssize_t ret = count;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800161
David Rientjes21440d72017-02-22 15:45:49 -0800162 if (!memcmp("always", buf,
163 min(sizeof("always")-1, count))) {
164 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
165 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
166 } else if (!memcmp("madvise", buf,
167 min(sizeof("madvise")-1, count))) {
168 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
170 } else if (!memcmp("never", buf,
171 min(sizeof("never")-1, count))) {
172 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
173 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
174 } else
175 ret = -EINVAL;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800176
177 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700178 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800179 if (err)
180 ret = err;
181 }
Andrea Arcangeliba761492011-01-13 15:46:58 -0800182 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800183}
184static struct kobj_attribute enabled_attr =
185 __ATTR(enabled, 0644, enabled_show, enabled_store);
186
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700187ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800188 struct kobj_attribute *attr, char *buf,
189 enum transparent_hugepage_flag flag)
190{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700191 return sprintf(buf, "%d\n",
192 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800193}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700194
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700195ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800196 struct kobj_attribute *attr,
197 const char *buf, size_t count,
198 enum transparent_hugepage_flag flag)
199{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700200 unsigned long value;
201 int ret;
202
203 ret = kstrtoul(buf, 10, &value);
204 if (ret < 0)
205 return ret;
206 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800207 return -EINVAL;
208
Ben Hutchingse27e6152011-04-14 15:22:21 -0700209 if (value)
210 set_bit(flag, &transparent_hugepage_flags);
211 else
212 clear_bit(flag, &transparent_hugepage_flags);
213
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800214 return count;
215}
216
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800217static ssize_t defrag_show(struct kobject *kobj,
218 struct kobj_attribute *attr, char *buf)
219{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700220 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800221 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
Mel Gorman444eb2a42016-03-17 14:19:23 -0700222 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800223 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
224 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
225 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
226 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
227 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
228 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800229}
David Rientjes21440d72017-02-22 15:45:49 -0800230
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800231static ssize_t defrag_store(struct kobject *kobj,
232 struct kobj_attribute *attr,
233 const char *buf, size_t count)
234{
David Rientjes21440d72017-02-22 15:45:49 -0800235 if (!memcmp("always", buf,
236 min(sizeof("always")-1, count))) {
237 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
238 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
239 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
240 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
241 } else if (!memcmp("defer", buf,
242 min(sizeof("defer")-1, count))) {
243 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
245 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
246 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
247 } else if (!memcmp("defer+madvise", buf,
248 min(sizeof("defer+madvise")-1, count))) {
249 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
250 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
252 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 } else if (!memcmp("madvise", buf,
254 min(sizeof("madvise")-1, count))) {
255 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
258 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
259 } else if (!memcmp("never", buf,
260 min(sizeof("never")-1, count))) {
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
265 } else
266 return -EINVAL;
267
268 return count;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800269}
270static struct kobj_attribute defrag_attr =
271 __ATTR(defrag, 0644, defrag_show, defrag_store);
272
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800273static ssize_t use_zero_page_show(struct kobject *kobj,
274 struct kobj_attribute *attr, char *buf)
275{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700276 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800277 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
278}
279static ssize_t use_zero_page_store(struct kobject *kobj,
280 struct kobj_attribute *attr, const char *buf, size_t count)
281{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700282 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800283 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
284}
285static struct kobj_attribute use_zero_page_attr =
286 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Hugh Dickins49920d22016-12-12 16:44:50 -0800287
288static ssize_t hpage_pmd_size_show(struct kobject *kobj,
289 struct kobj_attribute *attr, char *buf)
290{
291 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
292}
293static struct kobj_attribute hpage_pmd_size_attr =
294 __ATTR_RO(hpage_pmd_size);
295
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800296#ifdef CONFIG_DEBUG_VM
297static ssize_t debug_cow_show(struct kobject *kobj,
298 struct kobj_attribute *attr, char *buf)
299{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700300 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
302}
303static ssize_t debug_cow_store(struct kobject *kobj,
304 struct kobj_attribute *attr,
305 const char *buf, size_t count)
306{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700307 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800308 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
309}
310static struct kobj_attribute debug_cow_attr =
311 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
312#endif /* CONFIG_DEBUG_VM */
313
314static struct attribute *hugepage_attr[] = {
315 &enabled_attr.attr,
316 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800317 &use_zero_page_attr.attr,
Hugh Dickins49920d22016-12-12 16:44:50 -0800318 &hpage_pmd_size_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700319#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700320 &shmem_enabled_attr.attr,
321#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800322#ifdef CONFIG_DEBUG_VM
323 &debug_cow_attr.attr,
324#endif
325 NULL,
326};
327
328static struct attribute_group hugepage_attr_group = {
329 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800330};
331
Shaohua Li569e5592012-01-12 17:19:11 -0800332static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
333{
334 int err;
335
336 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
337 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700338 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800339 return -ENOMEM;
340 }
341
342 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
343 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700344 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800345 goto delete_obj;
346 }
347
348 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
349 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700350 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800351 goto remove_hp_group;
352 }
353
354 return 0;
355
356remove_hp_group:
357 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
358delete_obj:
359 kobject_put(*hugepage_kobj);
360 return err;
361}
362
363static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
364{
365 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
366 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
367 kobject_put(hugepage_kobj);
368}
369#else
370static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
371{
372 return 0;
373}
374
375static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
376{
377}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800378#endif /* CONFIG_SYSFS */
379
380static int __init hugepage_init(void)
381{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800382 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800383 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800384
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800385 if (!has_transparent_hugepage()) {
386 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800387 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800388 }
389
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530390 /*
391 * hugepages can't be allocated by the buddy allocator
392 */
393 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
394 /*
395 * we use page->mapping and page->index in second tail page
396 * as list_head: assuming THP order >= 2
397 */
398 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
399
Shaohua Li569e5592012-01-12 17:19:11 -0800400 err = hugepage_init_sysfs(&hugepage_kobj);
401 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700402 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800403
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700404 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800405 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700406 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800407
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700408 err = register_shrinker(&huge_zero_page_shrinker);
409 if (err)
410 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800411 err = register_shrinker(&deferred_split_shrinker);
412 if (err)
413 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800414
Rik van Riel97562cd2011-01-13 15:47:12 -0800415 /*
416 * By default disable transparent hugepages on smaller systems,
417 * where the extra memory used could hurt more than TLB overhead
418 * is likely to save. The admin can still enable it through /sys.
419 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700420 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800421 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700422 return 0;
423 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800424
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700425 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700426 if (err)
427 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800428
Shaohua Li569e5592012-01-12 17:19:11 -0800429 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700430err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800431 unregister_shrinker(&deferred_split_shrinker);
432err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700433 unregister_shrinker(&huge_zero_page_shrinker);
434err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700435 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700436err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800437 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700438err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800439 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800440}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800441subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800442
443static int __init setup_transparent_hugepage(char *str)
444{
445 int ret = 0;
446 if (!str)
447 goto out;
448 if (!strcmp(str, "always")) {
449 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
450 &transparent_hugepage_flags);
451 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
452 &transparent_hugepage_flags);
453 ret = 1;
454 } else if (!strcmp(str, "madvise")) {
455 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
456 &transparent_hugepage_flags);
457 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
458 &transparent_hugepage_flags);
459 ret = 1;
460 } else if (!strcmp(str, "never")) {
461 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
462 &transparent_hugepage_flags);
463 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
464 &transparent_hugepage_flags);
465 ret = 1;
466 }
467out:
468 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700469 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800470 return ret;
471}
472__setup("transparent_hugepage=", setup_transparent_hugepage);
473
Mel Gormanb32967f2012-11-19 12:35:47 +0000474pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800475{
476 if (likely(vma->vm_flags & VM_WRITE))
477 pmd = pmd_mkwrite(pmd);
478 return pmd;
479}
480
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800481static inline struct list_head *page_deferred_list(struct page *page)
482{
483 /*
484 * ->lru in the tail pages is occupied by compound_head.
485 * Let's use ->mapping + ->index in the second tail page as list_head.
486 */
487 return (struct list_head *)&page[2].mapping;
488}
489
490void prep_transhuge_page(struct page *page)
491{
492 /*
493 * we use page->mapping and page->indexlru in second tail page
494 * as list_head: assuming THP order >= 2
495 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800496
497 INIT_LIST_HEAD(page_deferred_list(page));
498 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
499}
500
Toshi Kani74d2fad2016-10-07 16:59:56 -0700501unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
502 loff_t off, unsigned long flags, unsigned long size)
503{
504 unsigned long addr;
505 loff_t off_end = off + len;
506 loff_t off_align = round_up(off, size);
507 unsigned long len_pad;
508
509 if (off_end <= off_align || (off_end - off_align) < size)
510 return 0;
511
512 len_pad = len + size;
513 if (len_pad < len || (off + len_pad) < off)
514 return 0;
515
516 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
517 off >> PAGE_SHIFT, flags);
518 if (IS_ERR_VALUE(addr))
519 return 0;
520
521 addr += (off - addr) & (size - 1);
522 return addr;
523}
524
525unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
526 unsigned long len, unsigned long pgoff, unsigned long flags)
527{
528 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
529
530 if (addr)
531 goto out;
532 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
533 goto out;
534
535 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
536 if (addr)
537 return addr;
538
539 out:
540 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
541}
542EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
543
Jan Kara82b0f8c2016-12-14 15:06:58 -0800544static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700545 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800546{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800547 struct vm_area_struct *vma = vmf->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700548 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800549 pgtable_t pgtable;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800550 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800551
Sasha Levin309381fea2014-01-23 15:52:54 -0800552 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700553
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700554 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700555 put_page(page);
556 count_vm_event(THP_FAULT_FALLBACK);
557 return VM_FAULT_FALLBACK;
558 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800559
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700560 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700561 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800562 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700563 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700564 return VM_FAULT_OOM;
565 }
566
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800567 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700568 /*
569 * The memory barrier inside __SetPageUptodate makes sure that
570 * clear_huge_page writes become visible before the set_pmd_at()
571 * write.
572 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800573 __SetPageUptodate(page);
574
Jan Kara82b0f8c2016-12-14 15:06:58 -0800575 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
576 if (unlikely(!pmd_none(*vmf->pmd))) {
577 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800578 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800579 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700580 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800581 } else {
582 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700583
584 /* Deliver the page fault to userland */
585 if (userfaultfd_missing(vma)) {
586 int ret;
587
Jan Kara82b0f8c2016-12-14 15:06:58 -0800588 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800589 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700590 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700591 pte_free(vma->vm_mm, pgtable);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800592 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700593 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
594 return ret;
595 }
596
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700597 entry = mk_huge_pmd(page, vma->vm_page_prot);
598 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800599 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800600 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700601 lru_cache_add_active_or_unevictable(page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800602 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
603 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700604 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
605 atomic_long_inc(&vma->vm_mm->nr_ptes);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800606 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700607 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800608 }
609
David Rientjesaa2e8782012-05-29 15:06:17 -0700610 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800611}
612
Mel Gorman444eb2a42016-03-17 14:19:23 -0700613/*
David Rientjes21440d72017-02-22 15:45:49 -0800614 * always: directly stall for all thp allocations
615 * defer: wake kswapd and fail if not immediately available
616 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
617 * fail if not immediately available
618 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
619 * available
620 * never: never stall for any thp allocation
Mel Gorman444eb2a42016-03-17 14:19:23 -0700621 */
622static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800623{
David Rientjes21440d72017-02-22 15:45:49 -0800624 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700625
David Rientjes21440d72017-02-22 15:45:49 -0800626 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
Vlastimil Babka25160352016-07-28 15:49:25 -0700627 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
David Rientjes21440d72017-02-22 15:45:49 -0800628 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
629 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
630 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
631 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
632 __GFP_KSWAPD_RECLAIM);
633 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
634 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
635 0);
Vlastimil Babka25160352016-07-28 15:49:25 -0700636 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700637}
638
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800639/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700640static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800641 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700642 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800643{
644 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700645 if (!pmd_none(*pmd))
646 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700647 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800648 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800649 if (pgtable)
650 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800651 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800652 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700653 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800654}
655
Jan Kara82b0f8c2016-12-14 15:06:58 -0800656int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800657{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800658 struct vm_area_struct *vma = vmf->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800659 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800660 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800661 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800662
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700663 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700664 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700665 if (unlikely(anon_vma_prepare(vma)))
666 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700667 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700668 return VM_FAULT_OOM;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800669 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700670 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700671 transparent_hugepage_use_zero_page()) {
672 pgtable_t pgtable;
673 struct page *zero_page;
674 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700675 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700676 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700677 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800678 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700679 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700680 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700681 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700682 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700683 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700684 }
Jan Kara82b0f8c2016-12-14 15:06:58 -0800685 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700686 ret = 0;
687 set = false;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800688 if (pmd_none(*vmf->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700689 if (userfaultfd_missing(vma)) {
Jan Kara82b0f8c2016-12-14 15:06:58 -0800690 spin_unlock(vmf->ptl);
691 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700692 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
693 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700694 set_huge_zero_page(pgtable, vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -0800695 haddr, vmf->pmd, zero_page);
696 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700697 set = true;
698 }
699 } else
Jan Kara82b0f8c2016-12-14 15:06:58 -0800700 spin_unlock(vmf->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700701 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700702 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700703 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800704 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700705 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800706 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700707 if (unlikely(!page)) {
708 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700709 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700710 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800711 prep_transhuge_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800712 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800713}
714
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700715static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800716 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700717{
718 struct mm_struct *mm = vma->vm_mm;
719 pmd_t entry;
720 spinlock_t *ptl;
721
722 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800723 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
724 if (pfn_t_devmap(pfn))
725 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800726 if (write) {
727 entry = pmd_mkyoung(pmd_mkdirty(entry));
728 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700729 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800730 set_pmd_at(mm, addr, pmd, entry);
731 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700732 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700733}
734
735int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800736 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700737{
738 pgprot_t pgprot = vma->vm_page_prot;
739 /*
740 * If we had pmd_special, we could avoid all these restrictions,
741 * but we need to be consistent with PTEs and architectures that
742 * can't support a 'special' bit.
743 */
744 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
745 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
746 (VM_PFNMAP|VM_MIXEDMAP));
747 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800748 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700749
750 if (addr < vma->vm_start || addr >= vma->vm_end)
751 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200752
753 track_pfn_insert(vma, &pgprot, pfn);
754
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700755 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
756 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700757}
Dan Williamsdee41072016-05-14 12:20:44 -0700758EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700759
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800760#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
761static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
762{
763 if (likely(vma->vm_flags & VM_WRITE))
764 pud = pud_mkwrite(pud);
765 return pud;
766}
767
768static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
769 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
770{
771 struct mm_struct *mm = vma->vm_mm;
772 pud_t entry;
773 spinlock_t *ptl;
774
775 ptl = pud_lock(mm, pud);
776 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
777 if (pfn_t_devmap(pfn))
778 entry = pud_mkdevmap(entry);
779 if (write) {
780 entry = pud_mkyoung(pud_mkdirty(entry));
781 entry = maybe_pud_mkwrite(entry, vma);
782 }
783 set_pud_at(mm, addr, pud, entry);
784 update_mmu_cache_pud(vma, addr, pud);
785 spin_unlock(ptl);
786}
787
788int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
789 pud_t *pud, pfn_t pfn, bool write)
790{
791 pgprot_t pgprot = vma->vm_page_prot;
792 /*
793 * If we had pud_special, we could avoid all these restrictions,
794 * but we need to be consistent with PTEs and architectures that
795 * can't support a 'special' bit.
796 */
797 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
798 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
799 (VM_PFNMAP|VM_MIXEDMAP));
800 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
801 BUG_ON(!pfn_t_devmap(pfn));
802
803 if (addr < vma->vm_start || addr >= vma->vm_end)
804 return VM_FAULT_SIGBUS;
805
806 track_pfn_insert(vma, &pgprot, pfn);
807
808 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
809 return VM_FAULT_NOPAGE;
810}
811EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
812#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
813
Dan Williams3565fce2016-01-15 16:56:55 -0800814static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
815 pmd_t *pmd)
816{
817 pmd_t _pmd;
818
819 /*
820 * We should set the dirty bit only for FOLL_WRITE but for now
821 * the dirty bit in the pmd is meaningless. And if the dirty
822 * bit will become meaningful and we'll only set it with
823 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
824 * set the young bit, instead of the current set_pmd_at.
825 */
826 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
827 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
828 pmd, _pmd, 1))
829 update_mmu_cache_pmd(vma, addr, pmd);
830}
831
832struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
833 pmd_t *pmd, int flags)
834{
835 unsigned long pfn = pmd_pfn(*pmd);
836 struct mm_struct *mm = vma->vm_mm;
837 struct dev_pagemap *pgmap;
838 struct page *page;
839
840 assert_spin_locked(pmd_lockptr(mm, pmd));
841
Keno Fischer8310d482017-01-24 15:17:48 -0800842 /*
843 * When we COW a devmap PMD entry, we split it into PTEs, so we should
844 * not be in this function with `flags & FOLL_COW` set.
845 */
846 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
847
Dan Williams3565fce2016-01-15 16:56:55 -0800848 if (flags & FOLL_WRITE && !pmd_write(*pmd))
849 return NULL;
850
851 if (pmd_present(*pmd) && pmd_devmap(*pmd))
852 /* pass */;
853 else
854 return NULL;
855
856 if (flags & FOLL_TOUCH)
857 touch_pmd(vma, addr, pmd);
858
859 /*
860 * device mapped pages can only be returned if the
861 * caller will manage the page reference count.
862 */
863 if (!(flags & FOLL_GET))
864 return ERR_PTR(-EEXIST);
865
866 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
867 pgmap = get_dev_pagemap(pfn, NULL);
868 if (!pgmap)
869 return ERR_PTR(-EFAULT);
870 page = pfn_to_page(pfn);
871 get_page(page);
872 put_dev_pagemap(pgmap);
873
874 return page;
875}
876
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800877int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
878 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
879 struct vm_area_struct *vma)
880{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800881 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800882 struct page *src_page;
883 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800884 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700885 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800886
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700887 /* Skip if can be re-fill on fault */
888 if (!vma_is_anonymous(vma))
889 return 0;
890
891 pgtable = pte_alloc_one(dst_mm, addr);
892 if (unlikely(!pgtable))
893 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800894
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800895 dst_ptl = pmd_lock(dst_mm, dst_pmd);
896 src_ptl = pmd_lockptr(src_mm, src_pmd);
897 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800898
899 ret = -EAGAIN;
900 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700901 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800902 pte_free(dst_mm, pgtable);
903 goto out_unlock;
904 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800905 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800906 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800907 * under splitting since we don't split the page itself, only pmd to
908 * a page table.
909 */
910 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700911 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800912 /*
913 * get_huge_zero_page() will never allocate a new page here,
914 * since we already have a zero page to copy. It just takes a
915 * reference.
916 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700917 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700918 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700919 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800920 ret = 0;
921 goto out_unlock;
922 }
Mel Gormande466bd2013-12-18 17:08:42 -0800923
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700924 src_page = pmd_page(pmd);
925 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
926 get_page(src_page);
927 page_dup_rmap(src_page, true);
928 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
929 atomic_long_inc(&dst_mm->nr_ptes);
930 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800931
932 pmdp_set_wrprotect(src_mm, addr, src_pmd);
933 pmd = pmd_mkold(pmd_wrprotect(pmd));
934 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800935
936 ret = 0;
937out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800938 spin_unlock(src_ptl);
939 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800940out:
941 return ret;
942}
943
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800944#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
945static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
946 pud_t *pud)
947{
948 pud_t _pud;
949
950 /*
951 * We should set the dirty bit only for FOLL_WRITE but for now
952 * the dirty bit in the pud is meaningless. And if the dirty
953 * bit will become meaningful and we'll only set it with
954 * FOLL_WRITE, an atomic set_bit will be required on the pud to
955 * set the young bit, instead of the current set_pud_at.
956 */
957 _pud = pud_mkyoung(pud_mkdirty(*pud));
958 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
959 pud, _pud, 1))
960 update_mmu_cache_pud(vma, addr, pud);
961}
962
963struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
964 pud_t *pud, int flags)
965{
966 unsigned long pfn = pud_pfn(*pud);
967 struct mm_struct *mm = vma->vm_mm;
968 struct dev_pagemap *pgmap;
969 struct page *page;
970
971 assert_spin_locked(pud_lockptr(mm, pud));
972
973 if (flags & FOLL_WRITE && !pud_write(*pud))
974 return NULL;
975
976 if (pud_present(*pud) && pud_devmap(*pud))
977 /* pass */;
978 else
979 return NULL;
980
981 if (flags & FOLL_TOUCH)
982 touch_pud(vma, addr, pud);
983
984 /*
985 * device mapped pages can only be returned if the
986 * caller will manage the page reference count.
987 */
988 if (!(flags & FOLL_GET))
989 return ERR_PTR(-EEXIST);
990
991 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
992 pgmap = get_dev_pagemap(pfn, NULL);
993 if (!pgmap)
994 return ERR_PTR(-EFAULT);
995 page = pfn_to_page(pfn);
996 get_page(page);
997 put_dev_pagemap(pgmap);
998
999 return page;
1000}
1001
1002int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1003 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1004 struct vm_area_struct *vma)
1005{
1006 spinlock_t *dst_ptl, *src_ptl;
1007 pud_t pud;
1008 int ret;
1009
1010 dst_ptl = pud_lock(dst_mm, dst_pud);
1011 src_ptl = pud_lockptr(src_mm, src_pud);
1012 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1013
1014 ret = -EAGAIN;
1015 pud = *src_pud;
1016 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1017 goto out_unlock;
1018
1019 /*
1020 * When page table lock is held, the huge zero pud should not be
1021 * under splitting since we don't split the page itself, only pud to
1022 * a page table.
1023 */
1024 if (is_huge_zero_pud(pud)) {
1025 /* No huge zero pud yet */
1026 }
1027
1028 pudp_set_wrprotect(src_mm, addr, src_pud);
1029 pud = pud_mkold(pud_wrprotect(pud));
1030 set_pud_at(dst_mm, addr, dst_pud, pud);
1031
1032 ret = 0;
1033out_unlock:
1034 spin_unlock(src_ptl);
1035 spin_unlock(dst_ptl);
1036 return ret;
1037}
1038
1039void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1040{
1041 pud_t entry;
1042 unsigned long haddr;
1043 bool write = vmf->flags & FAULT_FLAG_WRITE;
1044
1045 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1046 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1047 goto unlock;
1048
1049 entry = pud_mkyoung(orig_pud);
1050 if (write)
1051 entry = pud_mkdirty(entry);
1052 haddr = vmf->address & HPAGE_PUD_MASK;
1053 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1054 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1055
1056unlock:
1057 spin_unlock(vmf->ptl);
1058}
1059#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1060
Jan Kara82b0f8c2016-12-14 15:06:58 -08001061void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001062{
1063 pmd_t entry;
1064 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001065 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001066
Jan Kara82b0f8c2016-12-14 15:06:58 -08001067 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1068 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001069 goto unlock;
1070
1071 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001072 if (write)
1073 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001074 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001075 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001076 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001077
1078unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001079 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001080}
1081
Jan Kara82b0f8c2016-12-14 15:06:58 -08001082static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001083 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001084{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001085 struct vm_area_struct *vma = vmf->vma;
1086 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001087 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001088 pgtable_t pgtable;
1089 pmd_t _pmd;
1090 int ret = 0, i;
1091 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001092 unsigned long mmun_start; /* For mmu_notifiers */
1093 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001094
1095 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1096 GFP_KERNEL);
1097 if (unlikely(!pages)) {
1098 ret |= VM_FAULT_OOM;
1099 goto out;
1100 }
1101
1102 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001103 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001104 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001105 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001106 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1107 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001108 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001109 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001110 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001111 memcg = (void *)page_private(pages[i]);
1112 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001113 mem_cgroup_cancel_charge(pages[i], memcg,
1114 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001115 put_page(pages[i]);
1116 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001117 kfree(pages);
1118 ret |= VM_FAULT_OOM;
1119 goto out;
1120 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001121 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001122 }
1123
1124 for (i = 0; i < HPAGE_PMD_NR; i++) {
1125 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001126 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001127 __SetPageUptodate(pages[i]);
1128 cond_resched();
1129 }
1130
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001131 mmun_start = haddr;
1132 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001133 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001134
Jan Kara82b0f8c2016-12-14 15:06:58 -08001135 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1136 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001137 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001138 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001139
Jan Kara82b0f8c2016-12-14 15:06:58 -08001140 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001141 /* leave pmd empty until pte is filled */
1142
Jan Kara82b0f8c2016-12-14 15:06:58 -08001143 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001144 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001145
1146 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001147 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001148 entry = mk_pte(pages[i], vma->vm_page_prot);
1149 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001150 memcg = (void *)page_private(pages[i]);
1151 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001152 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001153 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001154 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001155 vmf->pte = pte_offset_map(&_pmd, haddr);
1156 VM_BUG_ON(!pte_none(*vmf->pte));
1157 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1158 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159 }
1160 kfree(pages);
1161
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001162 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001163 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001164 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001165 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001166
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001167 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001168
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001169 ret |= VM_FAULT_WRITE;
1170 put_page(page);
1171
1172out:
1173 return ret;
1174
1175out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001176 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001177 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001178 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001179 memcg = (void *)page_private(pages[i]);
1180 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001181 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001182 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001183 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001184 kfree(pages);
1185 goto out;
1186}
1187
Jan Kara82b0f8c2016-12-14 15:06:58 -08001188int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001189{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001190 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001191 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001192 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001193 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001194 unsigned long mmun_start; /* For mmu_notifiers */
1195 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001196 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001197 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001198
Jan Kara82b0f8c2016-12-14 15:06:58 -08001199 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001200 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001201 if (is_huge_zero_pmd(orig_pmd))
1202 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001203 spin_lock(vmf->ptl);
1204 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205 goto out_unlock;
1206
1207 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001208 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001209 /*
1210 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001211 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001212 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001213 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001214 pmd_t entry;
1215 entry = pmd_mkyoung(orig_pmd);
1216 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001217 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1218 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001219 ret |= VM_FAULT_WRITE;
1220 goto out_unlock;
1221 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001222 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001223 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001224alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001225 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001226 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001227 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001228 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001229 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 new_page = NULL;
1231
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001232 if (likely(new_page)) {
1233 prep_transhuge_page(new_page);
1234 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001235 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001236 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001237 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001238 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001239 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001240 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001241 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001242 ret |= VM_FAULT_FALLBACK;
1243 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001244 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001245 }
David Rientjes17766dd2013-09-12 15:14:06 -07001246 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001247 goto out;
1248 }
1249
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001250 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1251 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001252 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001253 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001254 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001255 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001256 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001257 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001258 goto out;
1259 }
1260
David Rientjes17766dd2013-09-12 15:14:06 -07001261 count_vm_event(THP_FAULT_ALLOC);
1262
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001263 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001264 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1265 else
1266 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001267 __SetPageUptodate(new_page);
1268
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001269 mmun_start = haddr;
1270 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001271 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001272
Jan Kara82b0f8c2016-12-14 15:06:58 -08001273 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001274 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001275 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001276 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1277 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001278 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001279 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001280 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001281 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001282 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001283 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1284 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001285 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001286 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001287 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001288 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001289 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1290 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001291 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001292 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001293 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001294 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001295 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001296 put_page(page);
1297 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001298 ret |= VM_FAULT_WRITE;
1299 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001300 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001301out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001302 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001303out:
1304 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001305out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001306 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001307 return ret;
1308}
1309
Keno Fischer8310d482017-01-24 15:17:48 -08001310/*
1311 * FOLL_FORCE can write to even unwritable pmd's, but only
1312 * after we've gone through a COW cycle and they are dirty.
1313 */
1314static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1315{
1316 return pmd_write(pmd) ||
1317 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1318}
1319
David Rientjesb676b292012-10-08 16:34:03 -07001320struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001321 unsigned long addr,
1322 pmd_t *pmd,
1323 unsigned int flags)
1324{
David Rientjesb676b292012-10-08 16:34:03 -07001325 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001326 struct page *page = NULL;
1327
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001328 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001329
Keno Fischer8310d482017-01-24 15:17:48 -08001330 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001331 goto out;
1332
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001333 /* Avoid dumping huge zero page */
1334 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1335 return ERR_PTR(-EFAULT);
1336
Mel Gorman2b4847e2013-12-18 17:08:32 -08001337 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001338 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001339 goto out;
1340
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001341 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001342 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001343 if (flags & FOLL_TOUCH)
1344 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001345 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001346 /*
1347 * We don't mlock() pte-mapped THPs. This way we can avoid
1348 * leaking mlocked pages into non-VM_LOCKED VMAs.
1349 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001350 * For anon THP:
1351 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001352 * In most cases the pmd is the only mapping of the page as we
1353 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1354 * writable private mappings in populate_vma_page_range().
1355 *
1356 * The only scenario when we have the page shared here is if we
1357 * mlocking read-only mapping shared over fork(). We skip
1358 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001359 *
1360 * For file THP:
1361 *
1362 * We can expect PageDoubleMap() to be stable under page lock:
1363 * for file pages we set it in page_add_file_rmap(), which
1364 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001365 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001366
1367 if (PageAnon(page) && compound_mapcount(page) != 1)
1368 goto skip_mlock;
1369 if (PageDoubleMap(page) || !page->mapping)
1370 goto skip_mlock;
1371 if (!trylock_page(page))
1372 goto skip_mlock;
1373 lru_add_drain();
1374 if (page->mapping && !PageDoubleMap(page))
1375 mlock_vma_page(page);
1376 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001377 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001378skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001379 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001380 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001381 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001382 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001383
1384out:
1385 return page;
1386}
1387
Mel Gormand10e63f2012-10-25 14:16:31 +02001388/* NUMA hinting page fault entry point for trans huge pmds */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001389int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001390{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001391 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001392 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001393 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001394 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001395 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001396 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001397 bool page_locked;
1398 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001399 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001400 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001401
Jan Kara82b0f8c2016-12-14 15:06:58 -08001402 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1403 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001404 goto out_unlock;
1405
Mel Gormande466bd2013-12-18 17:08:42 -08001406 /*
1407 * If there are potential migrations, wait for completion and retry
1408 * without disrupting NUMA hinting information. Do not relock and
1409 * check_same as the page may no longer be mapped.
1410 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001411 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1412 page = pmd_page(*vmf->pmd);
1413 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001414 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001415 goto out;
1416 }
1417
Mel Gormand10e63f2012-10-25 14:16:31 +02001418 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001419 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001420 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001421 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001422 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001423 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001424 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001425 flags |= TNF_FAULT_LOCAL;
1426 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001427
Mel Gormanbea66fb2015-03-25 15:55:37 -07001428 /* See similar comment in do_numa_page for explanation */
Rik van Rield59dc7b2016-09-08 21:30:53 -04001429 if (!pmd_write(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001430 flags |= TNF_NO_GROUP;
1431
1432 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001433 * Acquire the page lock to serialise THP migrations but avoid dropping
1434 * page_table_lock if at all possible
1435 */
Mel Gormanb8916632013-10-07 11:28:44 +01001436 page_locked = trylock_page(page);
1437 target_nid = mpol_misplaced(page, vma, haddr);
1438 if (target_nid == -1) {
1439 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001440 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001441 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001442 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001443
Mel Gormande466bd2013-12-18 17:08:42 -08001444 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001445 if (!page_locked) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001446 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001447 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001448 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001449 goto out;
1450 }
1451
Mel Gorman2b4847e2013-12-18 17:08:32 -08001452 /*
1453 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1454 * to serialises splits
1455 */
Mel Gormanb8916632013-10-07 11:28:44 +01001456 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001457 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001458 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001459
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001460 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001461 spin_lock(vmf->ptl);
1462 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001463 unlock_page(page);
1464 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001465 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001466 goto out_unlock;
1467 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001468
Mel Gormanc3a489c2013-12-18 17:08:38 -08001469 /* Bail if we fail to protect against THP splits for any reason */
1470 if (unlikely(!anon_vma)) {
1471 put_page(page);
1472 page_nid = -1;
1473 goto clear_pmdnuma;
1474 }
1475
Mel Gormana54a4072013-10-07 11:28:46 +01001476 /*
1477 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001478 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001479 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001480 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001481 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001482 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001483 if (migrated) {
1484 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001485 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001486 } else
1487 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001488
Mel Gorman8191acb2013-10-07 11:28:45 +01001489 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001490clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001491 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001492 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001493 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001494 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001495 if (was_writable)
1496 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001497 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1498 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001499 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001500out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001501 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001502
1503out:
1504 if (anon_vma)
1505 page_unlock_anon_vma_read(anon_vma);
1506
Mel Gorman8191acb2013-10-07 11:28:45 +01001507 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001508 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1509 vmf->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001510
Mel Gormand10e63f2012-10-25 14:16:31 +02001511 return 0;
1512}
1513
Huang Ying319904a2016-07-28 15:48:03 -07001514/*
1515 * Return true if we do MADV_FREE successfully on entire pmd page.
1516 * Otherwise, return false.
1517 */
1518bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001519 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001520{
1521 spinlock_t *ptl;
1522 pmd_t orig_pmd;
1523 struct page *page;
1524 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001525 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001526
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001527 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1528
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001529 ptl = pmd_trans_huge_lock(pmd, vma);
1530 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001531 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001532
1533 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001534 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001535 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001536
1537 page = pmd_page(orig_pmd);
1538 /*
1539 * If other processes are mapping this page, we couldn't discard
1540 * the page unless they all do MADV_FREE so let's skip the page.
1541 */
1542 if (page_mapcount(page) != 1)
1543 goto out;
1544
1545 if (!trylock_page(page))
1546 goto out;
1547
1548 /*
1549 * If user want to discard part-pages of THP, split it so MADV_FREE
1550 * will deactivate only them.
1551 */
1552 if (next - addr != HPAGE_PMD_SIZE) {
1553 get_page(page);
1554 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001555 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001556 put_page(page);
1557 unlock_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001558 goto out_unlocked;
1559 }
1560
1561 if (PageDirty(page))
1562 ClearPageDirty(page);
1563 unlock_page(page);
1564
1565 if (PageActive(page))
1566 deactivate_page(page);
1567
1568 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1569 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1570 tlb->fullmm);
1571 orig_pmd = pmd_mkold(orig_pmd);
1572 orig_pmd = pmd_mkclean(orig_pmd);
1573
1574 set_pmd_at(mm, addr, pmd, orig_pmd);
1575 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1576 }
Huang Ying319904a2016-07-28 15:48:03 -07001577 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001578out:
1579 spin_unlock(ptl);
1580out_unlocked:
1581 return ret;
1582}
1583
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001584static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1585{
1586 pgtable_t pgtable;
1587
1588 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1589 pte_free(mm, pgtable);
1590 atomic_long_dec(&mm->nr_ptes);
1591}
1592
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001593int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001594 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001595{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001596 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001597 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001598
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001599 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1600
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001601 ptl = __pmd_trans_huge_lock(pmd, vma);
1602 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001603 return 0;
1604 /*
1605 * For architectures like ppc64 we look at deposited pgtable
1606 * when calling pmdp_huge_get_and_clear. So do the
1607 * pgtable_trans_huge_withdraw after finishing pmdp related
1608 * operations.
1609 */
1610 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1611 tlb->fullmm);
1612 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1613 if (vma_is_dax(vma)) {
1614 spin_unlock(ptl);
1615 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001616 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001617 } else if (is_huge_zero_pmd(orig_pmd)) {
1618 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1619 atomic_long_dec(&tlb->mm->nr_ptes);
1620 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001621 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001622 } else {
1623 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001624 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001625 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001626 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001627 if (PageAnon(page)) {
1628 pgtable_t pgtable;
1629 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1630 pte_free(tlb->mm, pgtable);
1631 atomic_long_dec(&tlb->mm->nr_ptes);
1632 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1633 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001634 if (arch_needs_pgtable_deposit())
1635 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001636 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1637 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001638 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001639 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001640 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001641 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001642}
1643
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001644#ifndef pmd_move_must_withdraw
1645static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1646 spinlock_t *old_pmd_ptl,
1647 struct vm_area_struct *vma)
1648{
1649 /*
1650 * With split pmd lock we also need to move preallocated
1651 * PTE page table if new_pmd is on different PMD page table.
1652 *
1653 * We also don't deposit and withdraw tables for file pages.
1654 */
1655 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1656}
1657#endif
1658
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001659bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001660 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001661 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001662{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001663 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001664 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001665 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001666 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001667
1668 if ((old_addr & ~HPAGE_PMD_MASK) ||
1669 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001670 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001671 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001672
1673 /*
1674 * The destination pmd shouldn't be established, free_pgtables()
1675 * should have release it.
1676 */
1677 if (WARN_ON(!pmd_none(*new_pmd))) {
1678 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001679 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001680 }
1681
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001682 /*
1683 * We don't have to worry about the ordering of src and dst
1684 * ptlocks because exclusive mmap_sem prevents deadlock.
1685 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001686 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1687 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001688 new_ptl = pmd_lockptr(mm, new_pmd);
1689 if (new_ptl != old_ptl)
1690 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001691 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001692 if (pmd_present(pmd) && pmd_dirty(pmd))
1693 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001694 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001695
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001696 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301697 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001698 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1699 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001700 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301701 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1702 if (new_ptl != old_ptl)
1703 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001704 if (force_flush)
1705 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1706 else
1707 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001708 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001709 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001710 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001711 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001712}
1713
Mel Gormanf123d742013-10-07 11:28:49 +01001714/*
1715 * Returns
1716 * - 0 if PMD could not be locked
1717 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1718 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1719 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001720int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001721 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001722{
1723 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001724 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001725 int ret = 0;
1726
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001727 ptl = __pmd_trans_huge_lock(pmd, vma);
1728 if (ptl) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001729 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001730 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001731 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001732
1733 /*
1734 * Avoid trapping faults against the zero page. The read-only
1735 * data is likely to be read-cached on the local CPU and
1736 * local/remote hits to the zero page are not interesting.
1737 */
1738 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1739 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001740 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001741 }
1742
Mel Gorman10c10452015-02-12 14:58:44 -08001743 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001744 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001745 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001746 if (preserve_write)
1747 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001748 ret = HPAGE_PMD_NR;
1749 set_pmd_at(mm, addr, pmd, entry);
Kirill A. Shutemovb237ade2016-07-26 15:25:45 -07001750 BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1751 pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001752 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001753 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001754 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001755
1756 return ret;
1757}
1758
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001759/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001760 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001761 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001762 * Note that if it returns page table lock pointer, this routine returns without
1763 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001764 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001765spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001766{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001767 spinlock_t *ptl;
1768 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001769 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001770 return ptl;
1771 spin_unlock(ptl);
1772 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001773}
1774
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001775/*
1776 * Returns true if a given pud maps a thp, false otherwise.
1777 *
1778 * Note that if it returns true, this routine returns without unlocking page
1779 * table lock. So callers must unlock it.
1780 */
1781spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1782{
1783 spinlock_t *ptl;
1784
1785 ptl = pud_lock(vma->vm_mm, pud);
1786 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1787 return ptl;
1788 spin_unlock(ptl);
1789 return NULL;
1790}
1791
1792#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1793int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1794 pud_t *pud, unsigned long addr)
1795{
1796 pud_t orig_pud;
1797 spinlock_t *ptl;
1798
1799 ptl = __pud_trans_huge_lock(pud, vma);
1800 if (!ptl)
1801 return 0;
1802 /*
1803 * For architectures like ppc64 we look at deposited pgtable
1804 * when calling pudp_huge_get_and_clear. So do the
1805 * pgtable_trans_huge_withdraw after finishing pudp related
1806 * operations.
1807 */
1808 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1809 tlb->fullmm);
1810 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1811 if (vma_is_dax(vma)) {
1812 spin_unlock(ptl);
1813 /* No zero page support yet */
1814 } else {
1815 /* No support for anonymous PUD pages yet */
1816 BUG();
1817 }
1818 return 1;
1819}
1820
1821static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1822 unsigned long haddr)
1823{
1824 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1825 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1826 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1827 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1828
1829 count_vm_event(THP_SPLIT_PMD);
1830
1831 pudp_huge_clear_flush_notify(vma, haddr, pud);
1832}
1833
1834void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1835 unsigned long address)
1836{
1837 spinlock_t *ptl;
1838 struct mm_struct *mm = vma->vm_mm;
1839 unsigned long haddr = address & HPAGE_PUD_MASK;
1840
1841 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1842 ptl = pud_lock(mm, pud);
1843 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1844 goto out;
1845 __split_huge_pud_locked(vma, pud, haddr);
1846
1847out:
1848 spin_unlock(ptl);
1849 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1850}
1851#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1852
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001853static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1854 unsigned long haddr, pmd_t *pmd)
1855{
1856 struct mm_struct *mm = vma->vm_mm;
1857 pgtable_t pgtable;
1858 pmd_t _pmd;
1859 int i;
1860
1861 /* leave pmd empty until pte is filled */
1862 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1863
1864 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1865 pmd_populate(mm, &_pmd, pgtable);
1866
1867 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1868 pte_t *pte, entry;
1869 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1870 entry = pte_mkspecial(entry);
1871 pte = pte_offset_map(&_pmd, haddr);
1872 VM_BUG_ON(!pte_none(*pte));
1873 set_pte_at(mm, haddr, pte, entry);
1874 pte_unmap(pte);
1875 }
1876 smp_wmb(); /* make pte visible before pmd */
1877 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001878}
1879
1880static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001881 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001882{
1883 struct mm_struct *mm = vma->vm_mm;
1884 struct page *page;
1885 pgtable_t pgtable;
1886 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001887 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001888 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001889 int i;
1890
1891 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1892 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1893 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001894 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001895
1896 count_vm_event(THP_SPLIT_PMD);
1897
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001898 if (!vma_is_anonymous(vma)) {
1899 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001900 /*
1901 * We are going to unmap this huge page. So
1902 * just go ahead and zap it
1903 */
1904 if (arch_needs_pgtable_deposit())
1905 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001906 if (vma_is_dax(vma))
1907 return;
1908 page = pmd_page(_pmd);
1909 if (!PageReferenced(page) && pmd_young(_pmd))
1910 SetPageReferenced(page);
1911 page_remove_rmap(page, true);
1912 put_page(page);
1913 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001914 return;
1915 } else if (is_huge_zero_pmd(*pmd)) {
1916 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1917 }
1918
1919 page = pmd_page(*pmd);
1920 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001921 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001922 write = pmd_write(*pmd);
1923 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001924 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001925 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001926
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301927 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001928 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1929 pmd_populate(mm, &_pmd, pgtable);
1930
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001931 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001932 pte_t entry, *pte;
1933 /*
1934 * Note that NUMA hinting access restrictions are not
1935 * transferred to avoid any possibility of altering
1936 * permissions across VMAs.
1937 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001938 if (freeze) {
1939 swp_entry_t swp_entry;
1940 swp_entry = make_migration_entry(page + i, write);
1941 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001942 if (soft_dirty)
1943 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001944 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001945 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001946 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001947 if (!write)
1948 entry = pte_wrprotect(entry);
1949 if (!young)
1950 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001951 if (soft_dirty)
1952 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001953 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001954 if (dirty)
1955 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001956 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001957 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001958 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001959 atomic_inc(&page[i]._mapcount);
1960 pte_unmap(pte);
1961 }
1962
1963 /*
1964 * Set PG_double_map before dropping compound_mapcount to avoid
1965 * false-negative page_mapped().
1966 */
1967 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1968 for (i = 0; i < HPAGE_PMD_NR; i++)
1969 atomic_inc(&page[i]._mapcount);
1970 }
1971
1972 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1973 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001974 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001975 if (TestClearPageDoubleMap(page)) {
1976 /* No need in mapcount reference anymore */
1977 for (i = 0; i < HPAGE_PMD_NR; i++)
1978 atomic_dec(&page[i]._mapcount);
1979 }
1980 }
1981
1982 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001983 /*
1984 * Up to this point the pmd is present and huge and userland has the
1985 * whole access to the hugepage during the split (which happens in
1986 * place). If we overwrite the pmd with the not-huge version pointing
1987 * to the pte here (which of course we could if all CPUs were bug
1988 * free), userland could trigger a small page size TLB miss on the
1989 * small sized TLB while the hugepage TLB entry is still established in
1990 * the huge TLB. Some CPU doesn't like that.
1991 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1992 * 383 on page 93. Intel should be safe but is also warns that it's
1993 * only safe if the permission and cache attributes of the two entries
1994 * loaded in the two TLB is identical (which should be the case here).
1995 * But it is generally safer to never allow small and huge TLB entries
1996 * for the same virtual address to be loaded simultaneously. So instead
1997 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1998 * current pmd notpresent (atomically because here the pmd_trans_huge
1999 * and pmd_trans_splitting must remain set at all times on the pmd
2000 * until the split is complete for this pmd), then we flush the SMP TLB
2001 * and finally we write the non-huge version of the pmd entry with
2002 * pmd_populate.
2003 */
2004 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002005 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002006
2007 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002008 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002009 page_remove_rmap(page + i, false);
2010 put_page(page + i);
2011 }
2012 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002013}
2014
2015void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002016 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002017{
2018 spinlock_t *ptl;
2019 struct mm_struct *mm = vma->vm_mm;
2020 unsigned long haddr = address & HPAGE_PMD_MASK;
2021
2022 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2023 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002024
2025 /*
2026 * If caller asks to setup a migration entries, we need a page to check
2027 * pmd against. Otherwise we can end up replacing wrong page.
2028 */
2029 VM_BUG_ON(freeze && !page);
2030 if (page && page != pmd_page(*pmd))
2031 goto out;
2032
Dan Williams5c7fb562016-01-15 16:56:52 -08002033 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002034 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002035 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002036 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08002037 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002038 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002039 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002040out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002041 spin_unlock(ptl);
2042 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2043}
2044
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002045void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2046 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002047{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002048 pgd_t *pgd;
2049 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002050 pmd_t *pmd;
2051
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002052 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002053 if (!pgd_present(*pgd))
2054 return;
2055
2056 pud = pud_offset(pgd, address);
2057 if (!pud_present(*pud))
2058 return;
2059
2060 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002061
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002062 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002063}
2064
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002065void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002066 unsigned long start,
2067 unsigned long end,
2068 long adjust_next)
2069{
2070 /*
2071 * If the new start address isn't hpage aligned and it could
2072 * previously contain an hugepage: check if we need to split
2073 * an huge pmd.
2074 */
2075 if (start & ~HPAGE_PMD_MASK &&
2076 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2077 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002078 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002079
2080 /*
2081 * If the new end address isn't hpage aligned and it could
2082 * previously contain an hugepage: check if we need to split
2083 * an huge pmd.
2084 */
2085 if (end & ~HPAGE_PMD_MASK &&
2086 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2087 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002088 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002089
2090 /*
2091 * If we're also updating the vma->vm_next->vm_start, if the new
2092 * vm_next->vm_start isn't page aligned and it could previously
2093 * contain an hugepage: check if we need to split an huge pmd.
2094 */
2095 if (adjust_next > 0) {
2096 struct vm_area_struct *next = vma->vm_next;
2097 unsigned long nstart = next->vm_start;
2098 nstart += adjust_next << PAGE_SHIFT;
2099 if (nstart & ~HPAGE_PMD_MASK &&
2100 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2101 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002102 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002103 }
2104}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002105
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002106static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002107{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002108 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2109 TTU_RMAP_LOCKED;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002110 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002111
2112 VM_BUG_ON_PAGE(!PageHead(page), page);
2113
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002114 if (PageAnon(page))
2115 ttu_flags |= TTU_MIGRATION;
2116
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002117 /* We only need TTU_SPLIT_HUGE_PMD once */
2118 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
2119 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
2120 /* Cut short if the page is unmapped */
2121 if (page_count(page) == 1)
2122 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002123
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002124 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002125 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002126 VM_BUG_ON_PAGE(ret, page + i - 1);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002127}
2128
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002129static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002130{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002131 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002132 if (PageTransHuge(page)) {
2133 remove_migration_ptes(page, page, true);
2134 } else {
2135 for (i = 0; i < HPAGE_PMD_NR; i++)
2136 remove_migration_ptes(page + i, page + i, true);
2137 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002138}
2139
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002140static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002141 struct lruvec *lruvec, struct list_head *list)
2142{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002143 struct page *page_tail = head + tail;
2144
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002145 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002146 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002147
2148 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002149 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002150 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002151 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2152 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002153 * get_page_unless_zero(), and atomic_set() is implemented in C not
2154 * using locked ops. spin_unlock on x86 sometime uses locked ops
2155 * because of PPro errata 66, 92, so unless somebody can guarantee
2156 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002157 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002158 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002159 if (PageAnon(head)) {
2160 page_ref_inc(page_tail);
2161 } else {
2162 /* Additional pin to radix tree */
2163 page_ref_add(page_tail, 2);
2164 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002165
2166 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2167 page_tail->flags |= (head->flags &
2168 ((1L << PG_referenced) |
2169 (1L << PG_swapbacked) |
2170 (1L << PG_mlocked) |
2171 (1L << PG_uptodate) |
2172 (1L << PG_active) |
2173 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002174 (1L << PG_unevictable) |
2175 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002176
2177 /*
2178 * After clearing PageTail the gup refcount can be released.
2179 * Page flags also must be visible before we make the page non-compound.
2180 */
2181 smp_wmb();
2182
2183 clear_compound_head(page_tail);
2184
2185 if (page_is_young(head))
2186 set_page_young(page_tail);
2187 if (page_is_idle(head))
2188 set_page_idle(page_tail);
2189
2190 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002191 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002192 page_tail);
2193 page_tail->mapping = head->mapping;
2194
2195 page_tail->index = head->index + tail;
2196 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2197 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002198}
2199
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002200static void __split_huge_page(struct page *page, struct list_head *list,
2201 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002202{
2203 struct page *head = compound_head(page);
2204 struct zone *zone = page_zone(head);
2205 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002206 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002207 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002208
Mel Gorman599d0c92016-07-28 15:45:31 -07002209 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002210
2211 /* complete memcg works before add pages to LRU */
2212 mem_cgroup_split_huge_fixup(head);
2213
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002214 if (!PageAnon(page))
2215 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2216
2217 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002218 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002219 /* Some pages can be beyond i_size: drop them from page cache */
2220 if (head[i].index >= end) {
2221 __ClearPageDirty(head + i);
2222 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002223 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2224 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002225 put_page(head + i);
2226 }
2227 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002228
2229 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002230 /* See comment in __split_huge_page_tail() */
2231 if (PageAnon(head)) {
2232 page_ref_inc(head);
2233 } else {
2234 /* Additional pin to radix tree */
2235 page_ref_add(head, 2);
2236 spin_unlock(&head->mapping->tree_lock);
2237 }
2238
Mel Gormana52633d2016-07-28 15:45:28 -07002239 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002240
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002241 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002242
2243 for (i = 0; i < HPAGE_PMD_NR; i++) {
2244 struct page *subpage = head + i;
2245 if (subpage == page)
2246 continue;
2247 unlock_page(subpage);
2248
2249 /*
2250 * Subpages may be freed if there wasn't any mapping
2251 * like if add_to_swap() is running on a lru page that
2252 * had its mapping zapped. And freeing these pages
2253 * requires taking the lru_lock so we do the put_page
2254 * of the tail pages after the split is complete.
2255 */
2256 put_page(subpage);
2257 }
2258}
2259
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002260int total_mapcount(struct page *page)
2261{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002262 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002263
2264 VM_BUG_ON_PAGE(PageTail(page), page);
2265
2266 if (likely(!PageCompound(page)))
2267 return atomic_read(&page->_mapcount) + 1;
2268
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002269 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002270 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002271 return compound;
2272 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002273 for (i = 0; i < HPAGE_PMD_NR; i++)
2274 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002275 /* File pages has compound_mapcount included in _mapcount */
2276 if (!PageAnon(page))
2277 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002278 if (PageDoubleMap(page))
2279 ret -= HPAGE_PMD_NR;
2280 return ret;
2281}
2282
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002283/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002284 * This calculates accurately how many mappings a transparent hugepage
2285 * has (unlike page_mapcount() which isn't fully accurate). This full
2286 * accuracy is primarily needed to know if copy-on-write faults can
2287 * reuse the page and change the mapping to read-write instead of
2288 * copying them. At the same time this returns the total_mapcount too.
2289 *
2290 * The function returns the highest mapcount any one of the subpages
2291 * has. If the return value is one, even if different processes are
2292 * mapping different subpages of the transparent hugepage, they can
2293 * all reuse it, because each process is reusing a different subpage.
2294 *
2295 * The total_mapcount is instead counting all virtual mappings of the
2296 * subpages. If the total_mapcount is equal to "one", it tells the
2297 * caller all mappings belong to the same "mm" and in turn the
2298 * anon_vma of the transparent hugepage can become the vma->anon_vma
2299 * local one as no other process may be mapping any of the subpages.
2300 *
2301 * It would be more accurate to replace page_mapcount() with
2302 * page_trans_huge_mapcount(), however we only use
2303 * page_trans_huge_mapcount() in the copy-on-write faults where we
2304 * need full accuracy to avoid breaking page pinning, because
2305 * page_trans_huge_mapcount() is slower than page_mapcount().
2306 */
2307int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2308{
2309 int i, ret, _total_mapcount, mapcount;
2310
2311 /* hugetlbfs shouldn't call it */
2312 VM_BUG_ON_PAGE(PageHuge(page), page);
2313
2314 if (likely(!PageTransCompound(page))) {
2315 mapcount = atomic_read(&page->_mapcount) + 1;
2316 if (total_mapcount)
2317 *total_mapcount = mapcount;
2318 return mapcount;
2319 }
2320
2321 page = compound_head(page);
2322
2323 _total_mapcount = ret = 0;
2324 for (i = 0; i < HPAGE_PMD_NR; i++) {
2325 mapcount = atomic_read(&page[i]._mapcount) + 1;
2326 ret = max(ret, mapcount);
2327 _total_mapcount += mapcount;
2328 }
2329 if (PageDoubleMap(page)) {
2330 ret -= 1;
2331 _total_mapcount -= HPAGE_PMD_NR;
2332 }
2333 mapcount = compound_mapcount(page);
2334 ret += mapcount;
2335 _total_mapcount += mapcount;
2336 if (total_mapcount)
2337 *total_mapcount = _total_mapcount;
2338 return ret;
2339}
2340
2341/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002342 * This function splits huge page into normal pages. @page can point to any
2343 * subpage of huge page to split. Split doesn't change the position of @page.
2344 *
2345 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2346 * The huge page must be locked.
2347 *
2348 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2349 *
2350 * Both head page and tail pages will inherit mapping, flags, and so on from
2351 * the hugepage.
2352 *
2353 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2354 * they are not mapped.
2355 *
2356 * Returns 0 if the hugepage is split successfully.
2357 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2358 * us.
2359 */
2360int split_huge_page_to_list(struct page *page, struct list_head *list)
2361{
2362 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002363 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002364 struct anon_vma *anon_vma = NULL;
2365 struct address_space *mapping = NULL;
2366 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002367 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002368 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002369
2370 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002371 VM_BUG_ON_PAGE(!PageLocked(page), page);
2372 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2373 VM_BUG_ON_PAGE(!PageCompound(page), page);
2374
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002375 if (PageAnon(head)) {
2376 /*
2377 * The caller does not necessarily hold an mmap_sem that would
2378 * prevent the anon_vma disappearing so we first we take a
2379 * reference to it and then lock the anon_vma for write. This
2380 * is similar to page_lock_anon_vma_read except the write lock
2381 * is taken to serialise against parallel split or collapse
2382 * operations.
2383 */
2384 anon_vma = page_get_anon_vma(head);
2385 if (!anon_vma) {
2386 ret = -EBUSY;
2387 goto out;
2388 }
2389 extra_pins = 0;
2390 mapping = NULL;
2391 anon_vma_lock_write(anon_vma);
2392 } else {
2393 mapping = head->mapping;
2394
2395 /* Truncated ? */
2396 if (!mapping) {
2397 ret = -EBUSY;
2398 goto out;
2399 }
2400
2401 /* Addidional pins from radix tree */
2402 extra_pins = HPAGE_PMD_NR;
2403 anon_vma = NULL;
2404 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002405 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002406
2407 /*
2408 * Racy check if we can split the page, before freeze_page() will
2409 * split PMDs
2410 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002411 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002412 ret = -EBUSY;
2413 goto out_unlock;
2414 }
2415
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002416 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002417 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002418 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2419
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002420 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2421 if (mlocked)
2422 lru_add_drain();
2423
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002424 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002425 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002426
2427 if (mapping) {
2428 void **pslot;
2429
2430 spin_lock(&mapping->tree_lock);
2431 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2432 page_index(head));
2433 /*
2434 * Check if the head page is present in radix tree.
2435 * We assume all tail are present too, if head is there.
2436 */
2437 if (radix_tree_deref_slot_protected(pslot,
2438 &mapping->tree_lock) != head)
2439 goto fail;
2440 }
2441
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002442 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002443 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002444 count = page_count(head);
2445 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002446 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002447 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002448 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002449 list_del(page_deferred_list(head));
2450 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002451 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002452 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002453 spin_unlock(&pgdata->split_queue_lock);
2454 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002455 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002456 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002457 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2458 pr_alert("total_mapcount: %u, page_count(): %u\n",
2459 mapcount, count);
2460 if (PageTail(page))
2461 dump_page(head, NULL);
2462 dump_page(page, "total_mapcount(head) > 0");
2463 BUG();
2464 }
2465 spin_unlock(&pgdata->split_queue_lock);
2466fail: if (mapping)
2467 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002468 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002469 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002470 ret = -EBUSY;
2471 }
2472
2473out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002474 if (anon_vma) {
2475 anon_vma_unlock_write(anon_vma);
2476 put_anon_vma(anon_vma);
2477 }
2478 if (mapping)
2479 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002480out:
2481 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2482 return ret;
2483}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002484
2485void free_transhuge_page(struct page *page)
2486{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002487 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002488 unsigned long flags;
2489
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002490 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002491 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002492 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002493 list_del(page_deferred_list(page));
2494 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002495 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002496 free_compound_page(page);
2497}
2498
2499void deferred_split_huge_page(struct page *page)
2500{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002501 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002502 unsigned long flags;
2503
2504 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2505
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002506 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002507 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002508 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002509 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2510 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002511 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002512 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002513}
2514
2515static unsigned long deferred_split_count(struct shrinker *shrink,
2516 struct shrink_control *sc)
2517{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002518 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002519 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002520}
2521
2522static unsigned long deferred_split_scan(struct shrinker *shrink,
2523 struct shrink_control *sc)
2524{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002525 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002526 unsigned long flags;
2527 LIST_HEAD(list), *pos, *next;
2528 struct page *page;
2529 int split = 0;
2530
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002531 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002532 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002533 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002534 page = list_entry((void *)pos, struct page, mapping);
2535 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002536 if (get_page_unless_zero(page)) {
2537 list_move(page_deferred_list(page), &list);
2538 } else {
2539 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002540 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002541 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002542 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002543 if (!--sc->nr_to_scan)
2544 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002545 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002546 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002547
2548 list_for_each_safe(pos, next, &list) {
2549 page = list_entry((void *)pos, struct page, mapping);
2550 lock_page(page);
2551 /* split_huge_page() removes page from list on success */
2552 if (!split_huge_page(page))
2553 split++;
2554 unlock_page(page);
2555 put_page(page);
2556 }
2557
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002558 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2559 list_splice_tail(&list, &pgdata->split_queue);
2560 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002561
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002562 /*
2563 * Stop shrinker if we didn't split any page, but the queue is empty.
2564 * This can happen if pages were freed under us.
2565 */
2566 if (!split && list_empty(&pgdata->split_queue))
2567 return SHRINK_STOP;
2568 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002569}
2570
2571static struct shrinker deferred_split_shrinker = {
2572 .count_objects = deferred_split_count,
2573 .scan_objects = deferred_split_scan,
2574 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002575 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002576};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002577
2578#ifdef CONFIG_DEBUG_FS
2579static int split_huge_pages_set(void *data, u64 val)
2580{
2581 struct zone *zone;
2582 struct page *page;
2583 unsigned long pfn, max_zone_pfn;
2584 unsigned long total = 0, split = 0;
2585
2586 if (val != 1)
2587 return -EINVAL;
2588
2589 for_each_populated_zone(zone) {
2590 max_zone_pfn = zone_end_pfn(zone);
2591 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2592 if (!pfn_valid(pfn))
2593 continue;
2594
2595 page = pfn_to_page(pfn);
2596 if (!get_page_unless_zero(page))
2597 continue;
2598
2599 if (zone != page_zone(page))
2600 goto next;
2601
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002602 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002603 goto next;
2604
2605 total++;
2606 lock_page(page);
2607 if (!split_huge_page(page))
2608 split++;
2609 unlock_page(page);
2610next:
2611 put_page(page);
2612 }
2613 }
2614
Yang Shi145bdaa2016-05-05 16:22:00 -07002615 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002616
2617 return 0;
2618}
2619DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2620 "%llu\n");
2621
2622static int __init split_huge_pages_debugfs(void)
2623{
2624 void *ret;
2625
Yang Shi145bdaa2016-05-05 16:22:00 -07002626 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002627 &split_huge_pages_fops);
2628 if (!ret)
2629 pr_warn("Failed to create split_huge_pages in debugfs");
2630 return 0;
2631}
2632late_initcall(split_huge_pages_debugfs);
2633#endif