blob: d4a6e40015128c626f606e339474d8b23d90607e [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 Arcangeliba761492011-01-13 15:46:58 -0800145
Mel Gorman444eb2a42016-03-17 14:19:23 -0700146static ssize_t triple_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800147 struct kobj_attribute *attr,
148 const char *buf, size_t count,
149 enum transparent_hugepage_flag enabled,
Mel Gorman444eb2a42016-03-17 14:19:23 -0700150 enum transparent_hugepage_flag deferred,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800151 enum transparent_hugepage_flag req_madv)
152{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700153 if (!memcmp("defer", buf,
154 min(sizeof("defer")-1, count))) {
155 if (enabled == deferred)
156 return -EINVAL;
157 clear_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800158 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700159 set_bit(deferred, &transparent_hugepage_flags);
160 } else if (!memcmp("always", buf,
161 min(sizeof("always")-1, count))) {
162 clear_bit(deferred, &transparent_hugepage_flags);
163 clear_bit(req_madv, &transparent_hugepage_flags);
164 set_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800165 } else if (!memcmp("madvise", buf,
166 min(sizeof("madvise")-1, count))) {
167 clear_bit(enabled, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700168 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800169 set_bit(req_madv, &transparent_hugepage_flags);
170 } else if (!memcmp("never", buf,
171 min(sizeof("never")-1, count))) {
172 clear_bit(enabled, &transparent_hugepage_flags);
173 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700174 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800175 } else
176 return -EINVAL;
177
178 return count;
179}
180
181static ssize_t enabled_show(struct kobject *kobj,
182 struct kobj_attribute *attr, char *buf)
183{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700184 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
185 return sprintf(buf, "[always] madvise never\n");
186 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
187 return sprintf(buf, "always [madvise] never\n");
188 else
189 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800190}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700191
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800192static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
195{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800196 ssize_t ret;
197
Mel Gorman444eb2a42016-03-17 14:19:23 -0700198 ret = triple_flag_store(kobj, attr, buf, count,
199 TRANSPARENT_HUGEPAGE_FLAG,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800200 TRANSPARENT_HUGEPAGE_FLAG,
201 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
202
203 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700204 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800205 if (err)
206 ret = err;
207 }
208
209 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800210}
211static struct kobj_attribute enabled_attr =
212 __ATTR(enabled, 0644, enabled_show, enabled_store);
213
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700214ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800215 struct kobj_attribute *attr, char *buf,
216 enum transparent_hugepage_flag flag)
217{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700218 return sprintf(buf, "%d\n",
219 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800220}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700221
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700222ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800223 struct kobj_attribute *attr,
224 const char *buf, size_t count,
225 enum transparent_hugepage_flag flag)
226{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700227 unsigned long value;
228 int ret;
229
230 ret = kstrtoul(buf, 10, &value);
231 if (ret < 0)
232 return ret;
233 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800234 return -EINVAL;
235
Ben Hutchingse27e6152011-04-14 15:22:21 -0700236 if (value)
237 set_bit(flag, &transparent_hugepage_flags);
238 else
239 clear_bit(flag, &transparent_hugepage_flags);
240
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800241 return count;
242}
243
244/*
245 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
246 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
247 * memory just to allocate one more hugepage.
248 */
249static ssize_t defrag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf)
251{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700252 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
253 return sprintf(buf, "[always] defer madvise never\n");
254 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
255 return sprintf(buf, "always [defer] madvise never\n");
256 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
257 return sprintf(buf, "always defer [madvise] never\n");
258 else
259 return sprintf(buf, "always defer madvise [never]\n");
260
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800261}
262static ssize_t defrag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count)
265{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700266 return triple_flag_store(kobj, attr, buf, count,
267 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
268 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800269 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
270}
271static struct kobj_attribute defrag_attr =
272 __ATTR(defrag, 0644, defrag_show, defrag_store);
273
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800274static ssize_t use_zero_page_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
276{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700277 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800278 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
279}
280static ssize_t use_zero_page_store(struct kobject *kobj,
281 struct kobj_attribute *attr, const char *buf, size_t count)
282{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700283 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800284 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
285}
286static struct kobj_attribute use_zero_page_attr =
287 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800288#ifdef CONFIG_DEBUG_VM
289static ssize_t debug_cow_show(struct kobject *kobj,
290 struct kobj_attribute *attr, char *buf)
291{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700292 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800293 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
294}
295static ssize_t debug_cow_store(struct kobject *kobj,
296 struct kobj_attribute *attr,
297 const char *buf, size_t count)
298{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700299 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800300 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
301}
302static struct kobj_attribute debug_cow_attr =
303 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
304#endif /* CONFIG_DEBUG_VM */
305
306static struct attribute *hugepage_attr[] = {
307 &enabled_attr.attr,
308 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800309 &use_zero_page_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700310#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700311 &shmem_enabled_attr.attr,
312#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800313#ifdef CONFIG_DEBUG_VM
314 &debug_cow_attr.attr,
315#endif
316 NULL,
317};
318
319static struct attribute_group hugepage_attr_group = {
320 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800321};
322
Shaohua Li569e5592012-01-12 17:19:11 -0800323static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
324{
325 int err;
326
327 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
328 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700329 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800330 return -ENOMEM;
331 }
332
333 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
334 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700335 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800336 goto delete_obj;
337 }
338
339 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
340 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700341 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800342 goto remove_hp_group;
343 }
344
345 return 0;
346
347remove_hp_group:
348 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
349delete_obj:
350 kobject_put(*hugepage_kobj);
351 return err;
352}
353
354static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
355{
356 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
357 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
358 kobject_put(hugepage_kobj);
359}
360#else
361static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
362{
363 return 0;
364}
365
366static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800369#endif /* CONFIG_SYSFS */
370
371static int __init hugepage_init(void)
372{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800373 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800374 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800375
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800376 if (!has_transparent_hugepage()) {
377 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800378 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800379 }
380
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530381 /*
382 * hugepages can't be allocated by the buddy allocator
383 */
384 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
385 /*
386 * we use page->mapping and page->index in second tail page
387 * as list_head: assuming THP order >= 2
388 */
389 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
390
Shaohua Li569e5592012-01-12 17:19:11 -0800391 err = hugepage_init_sysfs(&hugepage_kobj);
392 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700393 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800394
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700395 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800396 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700397 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800398
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700399 err = register_shrinker(&huge_zero_page_shrinker);
400 if (err)
401 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800402 err = register_shrinker(&deferred_split_shrinker);
403 if (err)
404 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800405
Rik van Riel97562cd2011-01-13 15:47:12 -0800406 /*
407 * By default disable transparent hugepages on smaller systems,
408 * where the extra memory used could hurt more than TLB overhead
409 * is likely to save. The admin can still enable it through /sys.
410 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700411 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800412 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700413 return 0;
414 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800415
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700416 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700417 if (err)
418 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800419
Shaohua Li569e5592012-01-12 17:19:11 -0800420 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700421err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800422 unregister_shrinker(&deferred_split_shrinker);
423err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700424 unregister_shrinker(&huge_zero_page_shrinker);
425err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700426 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700427err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800428 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700429err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800430 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800431}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800432subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800433
434static int __init setup_transparent_hugepage(char *str)
435{
436 int ret = 0;
437 if (!str)
438 goto out;
439 if (!strcmp(str, "always")) {
440 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
441 &transparent_hugepage_flags);
442 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
443 &transparent_hugepage_flags);
444 ret = 1;
445 } else if (!strcmp(str, "madvise")) {
446 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
447 &transparent_hugepage_flags);
448 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
449 &transparent_hugepage_flags);
450 ret = 1;
451 } else if (!strcmp(str, "never")) {
452 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 }
458out:
459 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700460 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800461 return ret;
462}
463__setup("transparent_hugepage=", setup_transparent_hugepage);
464
Mel Gormanb32967f2012-11-19 12:35:47 +0000465pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800466{
467 if (likely(vma->vm_flags & VM_WRITE))
468 pmd = pmd_mkwrite(pmd);
469 return pmd;
470}
471
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800472static inline struct list_head *page_deferred_list(struct page *page)
473{
474 /*
475 * ->lru in the tail pages is occupied by compound_head.
476 * Let's use ->mapping + ->index in the second tail page as list_head.
477 */
478 return (struct list_head *)&page[2].mapping;
479}
480
481void prep_transhuge_page(struct page *page)
482{
483 /*
484 * we use page->mapping and page->indexlru in second tail page
485 * as list_head: assuming THP order >= 2
486 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800487
488 INIT_LIST_HEAD(page_deferred_list(page));
489 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
490}
491
Toshi Kani74d2fad2016-10-07 16:59:56 -0700492unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
493 loff_t off, unsigned long flags, unsigned long size)
494{
495 unsigned long addr;
496 loff_t off_end = off + len;
497 loff_t off_align = round_up(off, size);
498 unsigned long len_pad;
499
500 if (off_end <= off_align || (off_end - off_align) < size)
501 return 0;
502
503 len_pad = len + size;
504 if (len_pad < len || (off + len_pad) < off)
505 return 0;
506
507 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
508 off >> PAGE_SHIFT, flags);
509 if (IS_ERR_VALUE(addr))
510 return 0;
511
512 addr += (off - addr) & (size - 1);
513 return addr;
514}
515
516unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
517 unsigned long len, unsigned long pgoff, unsigned long flags)
518{
519 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
520
521 if (addr)
522 goto out;
523 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
524 goto out;
525
526 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
527 if (addr)
528 return addr;
529
530 out:
531 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
532}
533EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
534
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700535static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
536 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800537{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700538 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700539 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800540 pgtable_t pgtable;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700541 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800542
Sasha Levin309381fea2014-01-23 15:52:54 -0800543 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700544
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700545 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700546 put_page(page);
547 count_vm_event(THP_FAULT_FALLBACK);
548 return VM_FAULT_FALLBACK;
549 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800550
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700551 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700552 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800553 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700554 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700555 return VM_FAULT_OOM;
556 }
557
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800558 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700559 /*
560 * The memory barrier inside __SetPageUptodate makes sure that
561 * clear_huge_page writes become visible before the set_pmd_at()
562 * write.
563 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800564 __SetPageUptodate(page);
565
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700566 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
567 if (unlikely(!pmd_none(*fe->pmd))) {
568 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800569 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800570 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700571 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800572 } else {
573 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700574
575 /* Deliver the page fault to userland */
576 if (userfaultfd_missing(vma)) {
577 int ret;
578
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700579 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800580 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700581 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700582 pte_free(vma->vm_mm, pgtable);
583 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700584 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
585 return ret;
586 }
587
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700588 entry = mk_huge_pmd(page, vma->vm_page_prot);
589 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800590 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800591 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700592 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700593 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
594 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
595 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
596 atomic_long_inc(&vma->vm_mm->nr_ptes);
597 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700598 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800599 }
600
David Rientjesaa2e8782012-05-29 15:06:17 -0700601 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800602}
603
Mel Gorman444eb2a42016-03-17 14:19:23 -0700604/*
Vlastimil Babka25160352016-07-28 15:49:25 -0700605 * If THP defrag is set to always then directly reclaim/compact as necessary
606 * If set to defer then do only background reclaim/compact and defer to khugepaged
Mel Gorman444eb2a42016-03-17 14:19:23 -0700607 * If set to madvise and the VMA is flagged then directly reclaim/compact
Vlastimil Babka25160352016-07-28 15:49:25 -0700608 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
Mel Gorman444eb2a42016-03-17 14:19:23 -0700609 */
610static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800611{
Vlastimil Babka25160352016-07-28 15:49:25 -0700612 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700613
Vlastimil Babka25160352016-07-28 15:49:25 -0700614 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
615 &transparent_hugepage_flags) && vma_madvised)
616 return GFP_TRANSHUGE;
617 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
618 &transparent_hugepage_flags))
619 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
620 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
621 &transparent_hugepage_flags))
622 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700623
Vlastimil Babka25160352016-07-28 15:49:25 -0700624 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700625}
626
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800627/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700628static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800629 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700630 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800631{
632 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700633 if (!pmd_none(*pmd))
634 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700635 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800636 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800637 if (pgtable)
638 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800639 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800640 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700641 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800642}
643
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700644int do_huge_pmd_anonymous_page(struct fault_env *fe)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800645{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700646 struct vm_area_struct *vma = fe->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800647 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800648 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700649 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800650
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700651 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700652 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700653 if (unlikely(anon_vma_prepare(vma)))
654 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700655 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700656 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700657 if (!(fe->flags & FAULT_FLAG_WRITE) &&
658 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700659 transparent_hugepage_use_zero_page()) {
660 pgtable_t pgtable;
661 struct page *zero_page;
662 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700663 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700664 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700665 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800666 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700667 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700668 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700669 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700670 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700671 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700672 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700673 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700674 ret = 0;
675 set = false;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700676 if (pmd_none(*fe->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700677 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700678 spin_unlock(fe->ptl);
679 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700680 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
681 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700682 set_huge_zero_page(pgtable, vma->vm_mm, vma,
683 haddr, fe->pmd, zero_page);
684 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700685 set = true;
686 }
687 } else
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700688 spin_unlock(fe->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700689 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700690 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700691 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800692 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700693 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800694 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700695 if (unlikely(!page)) {
696 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700697 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700698 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800699 prep_transhuge_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700700 return __do_huge_pmd_anonymous_page(fe, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800701}
702
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700703static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800704 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700705{
706 struct mm_struct *mm = vma->vm_mm;
707 pmd_t entry;
708 spinlock_t *ptl;
709
710 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800711 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
712 if (pfn_t_devmap(pfn))
713 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800714 if (write) {
715 entry = pmd_mkyoung(pmd_mkdirty(entry));
716 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700717 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800718 set_pmd_at(mm, addr, pmd, entry);
719 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700720 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700721}
722
723int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800724 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700725{
726 pgprot_t pgprot = vma->vm_page_prot;
727 /*
728 * If we had pmd_special, we could avoid all these restrictions,
729 * but we need to be consistent with PTEs and architectures that
730 * can't support a 'special' bit.
731 */
732 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
733 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
734 (VM_PFNMAP|VM_MIXEDMAP));
735 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800736 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700737
738 if (addr < vma->vm_start || addr >= vma->vm_end)
739 return VM_FAULT_SIGBUS;
740 if (track_pfn_insert(vma, &pgprot, pfn))
741 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700742 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
743 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700744}
Dan Williamsdee41072016-05-14 12:20:44 -0700745EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700746
Dan Williams3565fce2016-01-15 16:56:55 -0800747static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
748 pmd_t *pmd)
749{
750 pmd_t _pmd;
751
752 /*
753 * We should set the dirty bit only for FOLL_WRITE but for now
754 * the dirty bit in the pmd is meaningless. And if the dirty
755 * bit will become meaningful and we'll only set it with
756 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
757 * set the young bit, instead of the current set_pmd_at.
758 */
759 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
760 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
761 pmd, _pmd, 1))
762 update_mmu_cache_pmd(vma, addr, pmd);
763}
764
765struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
766 pmd_t *pmd, int flags)
767{
768 unsigned long pfn = pmd_pfn(*pmd);
769 struct mm_struct *mm = vma->vm_mm;
770 struct dev_pagemap *pgmap;
771 struct page *page;
772
773 assert_spin_locked(pmd_lockptr(mm, pmd));
774
775 if (flags & FOLL_WRITE && !pmd_write(*pmd))
776 return NULL;
777
778 if (pmd_present(*pmd) && pmd_devmap(*pmd))
779 /* pass */;
780 else
781 return NULL;
782
783 if (flags & FOLL_TOUCH)
784 touch_pmd(vma, addr, pmd);
785
786 /*
787 * device mapped pages can only be returned if the
788 * caller will manage the page reference count.
789 */
790 if (!(flags & FOLL_GET))
791 return ERR_PTR(-EEXIST);
792
793 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
794 pgmap = get_dev_pagemap(pfn, NULL);
795 if (!pgmap)
796 return ERR_PTR(-EFAULT);
797 page = pfn_to_page(pfn);
798 get_page(page);
799 put_dev_pagemap(pgmap);
800
801 return page;
802}
803
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800804int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
805 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
806 struct vm_area_struct *vma)
807{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800808 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800809 struct page *src_page;
810 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800811 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700812 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800813
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700814 /* Skip if can be re-fill on fault */
815 if (!vma_is_anonymous(vma))
816 return 0;
817
818 pgtable = pte_alloc_one(dst_mm, addr);
819 if (unlikely(!pgtable))
820 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800821
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800822 dst_ptl = pmd_lock(dst_mm, dst_pmd);
823 src_ptl = pmd_lockptr(src_mm, src_pmd);
824 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800825
826 ret = -EAGAIN;
827 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700828 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800829 pte_free(dst_mm, pgtable);
830 goto out_unlock;
831 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800832 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800833 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800834 * under splitting since we don't split the page itself, only pmd to
835 * a page table.
836 */
837 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700838 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800839 /*
840 * get_huge_zero_page() will never allocate a new page here,
841 * since we already have a zero page to copy. It just takes a
842 * reference.
843 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700844 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700845 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700846 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800847 ret = 0;
848 goto out_unlock;
849 }
Mel Gormande466bd2013-12-18 17:08:42 -0800850
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700851 src_page = pmd_page(pmd);
852 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
853 get_page(src_page);
854 page_dup_rmap(src_page, true);
855 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
856 atomic_long_inc(&dst_mm->nr_ptes);
857 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800858
859 pmdp_set_wrprotect(src_mm, addr, src_pmd);
860 pmd = pmd_mkold(pmd_wrprotect(pmd));
861 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800862
863 ret = 0;
864out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800865 spin_unlock(src_ptl);
866 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800867out:
868 return ret;
869}
870
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700871void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -0800872{
873 pmd_t entry;
874 unsigned long haddr;
875
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700876 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
877 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -0800878 goto unlock;
879
880 entry = pmd_mkyoung(orig_pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700881 haddr = fe->address & HPAGE_PMD_MASK;
882 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
883 fe->flags & FAULT_FLAG_WRITE))
884 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800885
886unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700887 spin_unlock(fe->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800888}
889
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700890static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
891 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800892{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700893 struct vm_area_struct *vma = fe->vma;
894 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -0700895 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896 pgtable_t pgtable;
897 pmd_t _pmd;
898 int ret = 0, i;
899 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700900 unsigned long mmun_start; /* For mmu_notifiers */
901 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800902
903 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
904 GFP_KERNEL);
905 if (unlikely(!pages)) {
906 ret |= VM_FAULT_OOM;
907 goto out;
908 }
909
910 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700911 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700912 __GFP_OTHER_NODE, vma,
913 fe->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800914 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700915 mem_cgroup_try_charge(pages[i], vma->vm_mm,
916 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800917 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800918 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800919 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700920 memcg = (void *)page_private(pages[i]);
921 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800922 mem_cgroup_cancel_charge(pages[i], memcg,
923 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800924 put_page(pages[i]);
925 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800926 kfree(pages);
927 ret |= VM_FAULT_OOM;
928 goto out;
929 }
Johannes Weiner00501b52014-08-08 14:19:20 -0700930 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800931 }
932
933 for (i = 0; i < HPAGE_PMD_NR; i++) {
934 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700935 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800936 __SetPageUptodate(pages[i]);
937 cond_resched();
938 }
939
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700940 mmun_start = haddr;
941 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700942 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700943
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700944 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
945 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800946 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -0800947 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800948
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700949 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800950 /* leave pmd empty until pte is filled */
951
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700952 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
953 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800954
955 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700956 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800957 entry = mk_pte(pages[i], vma->vm_page_prot);
958 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -0700959 memcg = (void *)page_private(pages[i]);
960 set_page_private(pages[i], 0);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700961 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800962 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -0700963 lru_cache_add_active_or_unevictable(pages[i], vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700964 fe->pte = pte_offset_map(&_pmd, haddr);
965 VM_BUG_ON(!pte_none(*fe->pte));
966 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
967 pte_unmap(fe->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800968 }
969 kfree(pages);
970
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800971 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700972 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800973 page_remove_rmap(page, true);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700974 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800975
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700976 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700977
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800978 ret |= VM_FAULT_WRITE;
979 put_page(page);
980
981out:
982 return ret;
983
984out_free_pages:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700985 spin_unlock(fe->ptl);
986 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800987 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700988 memcg = (void *)page_private(pages[i]);
989 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800990 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800991 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800992 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800993 kfree(pages);
994 goto out;
995}
996
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700997int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800998{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700999 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001000 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001001 struct mem_cgroup *memcg;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001002 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001003 unsigned long mmun_start; /* For mmu_notifiers */
1004 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001005 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001006 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001007
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001008 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001009 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001010 if (is_huge_zero_pmd(orig_pmd))
1011 goto alloc;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001012 spin_lock(fe->ptl);
1013 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001014 goto out_unlock;
1015
1016 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001017 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001018 /*
1019 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001020 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001021 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001022 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001023 pmd_t entry;
1024 entry = pmd_mkyoung(orig_pmd);
1025 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001026 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1027 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001028 ret |= VM_FAULT_WRITE;
1029 goto out_unlock;
1030 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001031 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001032 spin_unlock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001033alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001034 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001035 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001036 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001037 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001038 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001039 new_page = NULL;
1040
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001041 if (likely(new_page)) {
1042 prep_transhuge_page(new_page);
1043 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001044 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001045 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001046 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001047 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001048 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001049 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001050 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001051 ret |= VM_FAULT_FALLBACK;
1052 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001053 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001054 }
David Rientjes17766dd2013-09-12 15:14:06 -07001055 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056 goto out;
1057 }
1058
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001059 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1060 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001061 put_page(new_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001062 split_huge_pmd(vma, fe->pmd, fe->address);
1063 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001064 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001065 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001066 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001067 goto out;
1068 }
1069
David Rientjes17766dd2013-09-12 15:14:06 -07001070 count_vm_event(THP_FAULT_ALLOC);
1071
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001072 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001073 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1074 else
1075 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001076 __SetPageUptodate(new_page);
1077
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001078 mmun_start = haddr;
1079 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001080 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001081
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001082 spin_lock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001083 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001084 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001085 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1086 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001087 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001088 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001089 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001090 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001091 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001092 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1093 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001094 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001095 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001096 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001097 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001098 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1099 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001100 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001101 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001102 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001103 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001104 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001105 put_page(page);
1106 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001107 ret |= VM_FAULT_WRITE;
1108 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001109 spin_unlock(fe->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001110out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001111 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001112out:
1113 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001114out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001115 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001116 return ret;
1117}
1118
David Rientjesb676b292012-10-08 16:34:03 -07001119struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001120 unsigned long addr,
1121 pmd_t *pmd,
1122 unsigned int flags)
1123{
David Rientjesb676b292012-10-08 16:34:03 -07001124 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 struct page *page = NULL;
1126
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001127 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001128
1129 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1130 goto out;
1131
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001132 /* Avoid dumping huge zero page */
1133 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1134 return ERR_PTR(-EFAULT);
1135
Mel Gorman2b4847e2013-12-18 17:08:32 -08001136 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001137 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001138 goto out;
1139
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001140 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001141 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001142 if (flags & FOLL_TOUCH)
1143 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001144 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001145 /*
1146 * We don't mlock() pte-mapped THPs. This way we can avoid
1147 * leaking mlocked pages into non-VM_LOCKED VMAs.
1148 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001149 * For anon THP:
1150 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001151 * In most cases the pmd is the only mapping of the page as we
1152 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1153 * writable private mappings in populate_vma_page_range().
1154 *
1155 * The only scenario when we have the page shared here is if we
1156 * mlocking read-only mapping shared over fork(). We skip
1157 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001158 *
1159 * For file THP:
1160 *
1161 * We can expect PageDoubleMap() to be stable under page lock:
1162 * for file pages we set it in page_add_file_rmap(), which
1163 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001164 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001165
1166 if (PageAnon(page) && compound_mapcount(page) != 1)
1167 goto skip_mlock;
1168 if (PageDoubleMap(page) || !page->mapping)
1169 goto skip_mlock;
1170 if (!trylock_page(page))
1171 goto skip_mlock;
1172 lru_add_drain();
1173 if (page->mapping && !PageDoubleMap(page))
1174 mlock_vma_page(page);
1175 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001176 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001177skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001178 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001179 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001180 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001181 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001182
1183out:
1184 return page;
1185}
1186
Mel Gormand10e63f2012-10-25 14:16:31 +02001187/* NUMA hinting page fault entry point for trans huge pmds */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001188int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001189{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001190 struct vm_area_struct *vma = fe->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001191 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001192 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001193 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001194 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001195 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001196 bool page_locked;
1197 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001198 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001199 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001200
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001201 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1202 if (unlikely(!pmd_same(pmd, *fe->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001203 goto out_unlock;
1204
Mel Gormande466bd2013-12-18 17:08:42 -08001205 /*
1206 * If there are potential migrations, wait for completion and retry
1207 * without disrupting NUMA hinting information. Do not relock and
1208 * check_same as the page may no longer be mapped.
1209 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001210 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1211 page = pmd_page(*fe->pmd);
1212 spin_unlock(fe->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001213 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001214 goto out;
1215 }
1216
Mel Gormand10e63f2012-10-25 14:16:31 +02001217 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001218 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001219 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001220 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001221 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001222 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001223 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001224 flags |= TNF_FAULT_LOCAL;
1225 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001226
Mel Gormanbea66fb2015-03-25 15:55:37 -07001227 /* See similar comment in do_numa_page for explanation */
Rik van Rield59dc7b2016-09-08 21:30:53 -04001228 if (!pmd_write(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001229 flags |= TNF_NO_GROUP;
1230
1231 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001232 * Acquire the page lock to serialise THP migrations but avoid dropping
1233 * page_table_lock if at all possible
1234 */
Mel Gormanb8916632013-10-07 11:28:44 +01001235 page_locked = trylock_page(page);
1236 target_nid = mpol_misplaced(page, vma, haddr);
1237 if (target_nid == -1) {
1238 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001239 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001240 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001241 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001242
Mel Gormande466bd2013-12-18 17:08:42 -08001243 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001244 if (!page_locked) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001245 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001246 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001247 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001248 goto out;
1249 }
1250
Mel Gorman2b4847e2013-12-18 17:08:32 -08001251 /*
1252 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1253 * to serialises splits
1254 */
Mel Gormanb8916632013-10-07 11:28:44 +01001255 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001256 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001257 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001258
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001259 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001260 spin_lock(fe->ptl);
1261 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001262 unlock_page(page);
1263 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001264 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001265 goto out_unlock;
1266 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001267
Mel Gormanc3a489c2013-12-18 17:08:38 -08001268 /* Bail if we fail to protect against THP splits for any reason */
1269 if (unlikely(!anon_vma)) {
1270 put_page(page);
1271 page_nid = -1;
1272 goto clear_pmdnuma;
1273 }
1274
Mel Gormana54a4072013-10-07 11:28:46 +01001275 /*
1276 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001277 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001278 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001279 spin_unlock(fe->ptl);
1280 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1281 fe->pmd, pmd, fe->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001282 if (migrated) {
1283 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001284 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001285 } else
1286 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001287
Mel Gorman8191acb2013-10-07 11:28:45 +01001288 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001289clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001290 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001291 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001292 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001293 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001294 if (was_writable)
1295 pmd = pmd_mkwrite(pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001296 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1297 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001298 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001299out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001300 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001301
1302out:
1303 if (anon_vma)
1304 page_unlock_anon_vma_read(anon_vma);
1305
Mel Gorman8191acb2013-10-07 11:28:45 +01001306 if (page_nid != -1)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001307 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001308
Mel Gormand10e63f2012-10-25 14:16:31 +02001309 return 0;
1310}
1311
Huang Ying319904a2016-07-28 15:48:03 -07001312/*
1313 * Return true if we do MADV_FREE successfully on entire pmd page.
1314 * Otherwise, return false.
1315 */
1316bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001317 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001318{
1319 spinlock_t *ptl;
1320 pmd_t orig_pmd;
1321 struct page *page;
1322 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001323 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001324
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001325 ptl = pmd_trans_huge_lock(pmd, vma);
1326 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001327 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001328
1329 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001330 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001331 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001332
1333 page = pmd_page(orig_pmd);
1334 /*
1335 * If other processes are mapping this page, we couldn't discard
1336 * the page unless they all do MADV_FREE so let's skip the page.
1337 */
1338 if (page_mapcount(page) != 1)
1339 goto out;
1340
1341 if (!trylock_page(page))
1342 goto out;
1343
1344 /*
1345 * If user want to discard part-pages of THP, split it so MADV_FREE
1346 * will deactivate only them.
1347 */
1348 if (next - addr != HPAGE_PMD_SIZE) {
1349 get_page(page);
1350 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001351 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001352 put_page(page);
1353 unlock_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001354 goto out_unlocked;
1355 }
1356
1357 if (PageDirty(page))
1358 ClearPageDirty(page);
1359 unlock_page(page);
1360
1361 if (PageActive(page))
1362 deactivate_page(page);
1363
1364 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1365 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1366 tlb->fullmm);
1367 orig_pmd = pmd_mkold(orig_pmd);
1368 orig_pmd = pmd_mkclean(orig_pmd);
1369
1370 set_pmd_at(mm, addr, pmd, orig_pmd);
1371 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1372 }
Huang Ying319904a2016-07-28 15:48:03 -07001373 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001374out:
1375 spin_unlock(ptl);
1376out_unlocked:
1377 return ret;
1378}
1379
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001380int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001381 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001382{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001383 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001384 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001385
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001386 ptl = __pmd_trans_huge_lock(pmd, vma);
1387 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001388 return 0;
1389 /*
1390 * For architectures like ppc64 we look at deposited pgtable
1391 * when calling pmdp_huge_get_and_clear. So do the
1392 * pgtable_trans_huge_withdraw after finishing pmdp related
1393 * operations.
1394 */
1395 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1396 tlb->fullmm);
1397 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1398 if (vma_is_dax(vma)) {
1399 spin_unlock(ptl);
1400 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001401 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001402 } else if (is_huge_zero_pmd(orig_pmd)) {
1403 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1404 atomic_long_dec(&tlb->mm->nr_ptes);
1405 spin_unlock(ptl);
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001406 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001407 } else {
1408 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001409 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001410 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001411 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001412 if (PageAnon(page)) {
1413 pgtable_t pgtable;
1414 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1415 pte_free(tlb->mm, pgtable);
1416 atomic_long_dec(&tlb->mm->nr_ptes);
1417 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1418 } else {
1419 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1420 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001421 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001422 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001423 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001424 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001425}
1426
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001427bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001428 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001429 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001430{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001431 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001432 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001433 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001434 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001435
1436 if ((old_addr & ~HPAGE_PMD_MASK) ||
1437 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001438 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001439 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001440
1441 /*
1442 * The destination pmd shouldn't be established, free_pgtables()
1443 * should have release it.
1444 */
1445 if (WARN_ON(!pmd_none(*new_pmd))) {
1446 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001447 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001448 }
1449
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001450 /*
1451 * We don't have to worry about the ordering of src and dst
1452 * ptlocks because exclusive mmap_sem prevents deadlock.
1453 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001454 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1455 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001456 new_ptl = pmd_lockptr(mm, new_pmd);
1457 if (new_ptl != old_ptl)
1458 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001459 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001460 if (pmd_present(pmd) && pmd_dirty(pmd))
1461 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001462 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001463
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001464 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1465 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301466 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001467 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1468 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001469 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301470 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1471 if (new_ptl != old_ptl)
1472 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001473 if (force_flush)
1474 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1475 else
1476 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001477 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001478 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001479 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001480 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001481}
1482
Mel Gormanf123d742013-10-07 11:28:49 +01001483/*
1484 * Returns
1485 * - 0 if PMD could not be locked
1486 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1487 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1488 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001489int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001490 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001491{
1492 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001493 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001494 int ret = 0;
1495
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001496 ptl = __pmd_trans_huge_lock(pmd, vma);
1497 if (ptl) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001498 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001499 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001500 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001501
1502 /*
1503 * Avoid trapping faults against the zero page. The read-only
1504 * data is likely to be read-cached on the local CPU and
1505 * local/remote hits to the zero page are not interesting.
1506 */
1507 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1508 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001509 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001510 }
1511
Mel Gorman10c10452015-02-12 14:58:44 -08001512 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001513 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001514 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001515 if (preserve_write)
1516 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001517 ret = HPAGE_PMD_NR;
1518 set_pmd_at(mm, addr, pmd, entry);
Kirill A. Shutemovb237ade2016-07-26 15:25:45 -07001519 BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1520 pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001521 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001522 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001523 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001524
1525 return ret;
1526}
1527
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001528/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001529 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001530 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001531 * Note that if it returns page table lock pointer, this routine returns without
1532 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001533 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001534spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001535{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001536 spinlock_t *ptl;
1537 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001538 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001539 return ptl;
1540 spin_unlock(ptl);
1541 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001542}
1543
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001544static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1545 unsigned long haddr, pmd_t *pmd)
1546{
1547 struct mm_struct *mm = vma->vm_mm;
1548 pgtable_t pgtable;
1549 pmd_t _pmd;
1550 int i;
1551
1552 /* leave pmd empty until pte is filled */
1553 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1554
1555 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1556 pmd_populate(mm, &_pmd, pgtable);
1557
1558 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1559 pte_t *pte, entry;
1560 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1561 entry = pte_mkspecial(entry);
1562 pte = pte_offset_map(&_pmd, haddr);
1563 VM_BUG_ON(!pte_none(*pte));
1564 set_pte_at(mm, haddr, pte, entry);
1565 pte_unmap(pte);
1566 }
1567 smp_wmb(); /* make pte visible before pmd */
1568 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001569}
1570
1571static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001572 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001573{
1574 struct mm_struct *mm = vma->vm_mm;
1575 struct page *page;
1576 pgtable_t pgtable;
1577 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001578 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001579 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001580 int i;
1581
1582 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1583 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1584 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001585 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001586
1587 count_vm_event(THP_SPLIT_PMD);
1588
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001589 if (!vma_is_anonymous(vma)) {
1590 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001591 if (vma_is_dax(vma))
1592 return;
1593 page = pmd_page(_pmd);
1594 if (!PageReferenced(page) && pmd_young(_pmd))
1595 SetPageReferenced(page);
1596 page_remove_rmap(page, true);
1597 put_page(page);
1598 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001599 return;
1600 } else if (is_huge_zero_pmd(*pmd)) {
1601 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1602 }
1603
1604 page = pmd_page(*pmd);
1605 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001606 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001607 write = pmd_write(*pmd);
1608 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001609 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001610 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001611
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301612 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001613 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1614 pmd_populate(mm, &_pmd, pgtable);
1615
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001616 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001617 pte_t entry, *pte;
1618 /*
1619 * Note that NUMA hinting access restrictions are not
1620 * transferred to avoid any possibility of altering
1621 * permissions across VMAs.
1622 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001623 if (freeze) {
1624 swp_entry_t swp_entry;
1625 swp_entry = make_migration_entry(page + i, write);
1626 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001627 if (soft_dirty)
1628 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001629 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001630 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001631 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001632 if (!write)
1633 entry = pte_wrprotect(entry);
1634 if (!young)
1635 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001636 if (soft_dirty)
1637 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001638 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001639 if (dirty)
1640 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001641 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001642 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001643 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001644 atomic_inc(&page[i]._mapcount);
1645 pte_unmap(pte);
1646 }
1647
1648 /*
1649 * Set PG_double_map before dropping compound_mapcount to avoid
1650 * false-negative page_mapped().
1651 */
1652 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1653 for (i = 0; i < HPAGE_PMD_NR; i++)
1654 atomic_inc(&page[i]._mapcount);
1655 }
1656
1657 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1658 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001659 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001660 if (TestClearPageDoubleMap(page)) {
1661 /* No need in mapcount reference anymore */
1662 for (i = 0; i < HPAGE_PMD_NR; i++)
1663 atomic_dec(&page[i]._mapcount);
1664 }
1665 }
1666
1667 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001668 /*
1669 * Up to this point the pmd is present and huge and userland has the
1670 * whole access to the hugepage during the split (which happens in
1671 * place). If we overwrite the pmd with the not-huge version pointing
1672 * to the pte here (which of course we could if all CPUs were bug
1673 * free), userland could trigger a small page size TLB miss on the
1674 * small sized TLB while the hugepage TLB entry is still established in
1675 * the huge TLB. Some CPU doesn't like that.
1676 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1677 * 383 on page 93. Intel should be safe but is also warns that it's
1678 * only safe if the permission and cache attributes of the two entries
1679 * loaded in the two TLB is identical (which should be the case here).
1680 * But it is generally safer to never allow small and huge TLB entries
1681 * for the same virtual address to be loaded simultaneously. So instead
1682 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1683 * current pmd notpresent (atomically because here the pmd_trans_huge
1684 * and pmd_trans_splitting must remain set at all times on the pmd
1685 * until the split is complete for this pmd), then we flush the SMP TLB
1686 * and finally we write the non-huge version of the pmd entry with
1687 * pmd_populate.
1688 */
1689 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001690 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001691
1692 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001693 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001694 page_remove_rmap(page + i, false);
1695 put_page(page + i);
1696 }
1697 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001698}
1699
1700void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001701 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001702{
1703 spinlock_t *ptl;
1704 struct mm_struct *mm = vma->vm_mm;
1705 unsigned long haddr = address & HPAGE_PMD_MASK;
1706
1707 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1708 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001709
1710 /*
1711 * If caller asks to setup a migration entries, we need a page to check
1712 * pmd against. Otherwise we can end up replacing wrong page.
1713 */
1714 VM_BUG_ON(freeze && !page);
1715 if (page && page != pmd_page(*pmd))
1716 goto out;
1717
Dan Williams5c7fb562016-01-15 16:56:52 -08001718 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001719 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001720 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07001721 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08001722 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001723 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001724 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001725out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001726 spin_unlock(ptl);
1727 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1728}
1729
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001730void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1731 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001732{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001733 pgd_t *pgd;
1734 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001735 pmd_t *pmd;
1736
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001737 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001738 if (!pgd_present(*pgd))
1739 return;
1740
1741 pud = pud_offset(pgd, address);
1742 if (!pud_present(*pud))
1743 return;
1744
1745 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001746
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001747 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001748}
1749
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07001750void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001751 unsigned long start,
1752 unsigned long end,
1753 long adjust_next)
1754{
1755 /*
1756 * If the new start address isn't hpage aligned and it could
1757 * previously contain an hugepage: check if we need to split
1758 * an huge pmd.
1759 */
1760 if (start & ~HPAGE_PMD_MASK &&
1761 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1762 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001763 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001764
1765 /*
1766 * If the new end address isn't hpage aligned and it could
1767 * previously contain an hugepage: check if we need to split
1768 * an huge pmd.
1769 */
1770 if (end & ~HPAGE_PMD_MASK &&
1771 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1772 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001773 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001774
1775 /*
1776 * If we're also updating the vma->vm_next->vm_start, if the new
1777 * vm_next->vm_start isn't page aligned and it could previously
1778 * contain an hugepage: check if we need to split an huge pmd.
1779 */
1780 if (adjust_next > 0) {
1781 struct vm_area_struct *next = vma->vm_next;
1782 unsigned long nstart = next->vm_start;
1783 nstart += adjust_next << PAGE_SHIFT;
1784 if (nstart & ~HPAGE_PMD_MASK &&
1785 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1786 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001787 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001788 }
1789}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001790
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001791static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001792{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001793 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1794 TTU_RMAP_LOCKED;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001795 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001796
1797 VM_BUG_ON_PAGE(!PageHead(page), page);
1798
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001799 if (PageAnon(page))
1800 ttu_flags |= TTU_MIGRATION;
1801
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001802 /* We only need TTU_SPLIT_HUGE_PMD once */
1803 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1804 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1805 /* Cut short if the page is unmapped */
1806 if (page_count(page) == 1)
1807 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001808
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001809 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001810 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001811 VM_BUG_ON_PAGE(ret, page + i - 1);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001812}
1813
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001814static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001815{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001816 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001817
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001818 for (i = 0; i < HPAGE_PMD_NR; i++)
1819 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001820}
1821
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001822static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001823 struct lruvec *lruvec, struct list_head *list)
1824{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001825 struct page *page_tail = head + tail;
1826
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001827 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001828 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001829
1830 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001831 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001832 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001833 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1834 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001835 * get_page_unless_zero(), and atomic_set() is implemented in C not
1836 * using locked ops. spin_unlock on x86 sometime uses locked ops
1837 * because of PPro errata 66, 92, so unless somebody can guarantee
1838 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001839 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001840 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001841 if (PageAnon(head)) {
1842 page_ref_inc(page_tail);
1843 } else {
1844 /* Additional pin to radix tree */
1845 page_ref_add(page_tail, 2);
1846 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001847
1848 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1849 page_tail->flags |= (head->flags &
1850 ((1L << PG_referenced) |
1851 (1L << PG_swapbacked) |
1852 (1L << PG_mlocked) |
1853 (1L << PG_uptodate) |
1854 (1L << PG_active) |
1855 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001856 (1L << PG_unevictable) |
1857 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001858
1859 /*
1860 * After clearing PageTail the gup refcount can be released.
1861 * Page flags also must be visible before we make the page non-compound.
1862 */
1863 smp_wmb();
1864
1865 clear_compound_head(page_tail);
1866
1867 if (page_is_young(head))
1868 set_page_young(page_tail);
1869 if (page_is_idle(head))
1870 set_page_idle(page_tail);
1871
1872 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001873 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001874 page_tail);
1875 page_tail->mapping = head->mapping;
1876
1877 page_tail->index = head->index + tail;
1878 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1879 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001880}
1881
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001882static void __split_huge_page(struct page *page, struct list_head *list,
1883 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001884{
1885 struct page *head = compound_head(page);
1886 struct zone *zone = page_zone(head);
1887 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001888 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001889 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001890
Mel Gorman599d0c92016-07-28 15:45:31 -07001891 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001892
1893 /* complete memcg works before add pages to LRU */
1894 mem_cgroup_split_huge_fixup(head);
1895
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001896 if (!PageAnon(page))
1897 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1898
1899 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001900 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001901 /* Some pages can be beyond i_size: drop them from page cache */
1902 if (head[i].index >= end) {
1903 __ClearPageDirty(head + i);
1904 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001905 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1906 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001907 put_page(head + i);
1908 }
1909 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001910
1911 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001912 /* See comment in __split_huge_page_tail() */
1913 if (PageAnon(head)) {
1914 page_ref_inc(head);
1915 } else {
1916 /* Additional pin to radix tree */
1917 page_ref_add(head, 2);
1918 spin_unlock(&head->mapping->tree_lock);
1919 }
1920
Mel Gormana52633d2016-07-28 15:45:28 -07001921 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001922
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001923 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001924
1925 for (i = 0; i < HPAGE_PMD_NR; i++) {
1926 struct page *subpage = head + i;
1927 if (subpage == page)
1928 continue;
1929 unlock_page(subpage);
1930
1931 /*
1932 * Subpages may be freed if there wasn't any mapping
1933 * like if add_to_swap() is running on a lru page that
1934 * had its mapping zapped. And freeing these pages
1935 * requires taking the lru_lock so we do the put_page
1936 * of the tail pages after the split is complete.
1937 */
1938 put_page(subpage);
1939 }
1940}
1941
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001942int total_mapcount(struct page *page)
1943{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001944 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001945
1946 VM_BUG_ON_PAGE(PageTail(page), page);
1947
1948 if (likely(!PageCompound(page)))
1949 return atomic_read(&page->_mapcount) + 1;
1950
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001951 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001952 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001953 return compound;
1954 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001955 for (i = 0; i < HPAGE_PMD_NR; i++)
1956 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001957 /* File pages has compound_mapcount included in _mapcount */
1958 if (!PageAnon(page))
1959 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001960 if (PageDoubleMap(page))
1961 ret -= HPAGE_PMD_NR;
1962 return ret;
1963}
1964
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001965/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001966 * This calculates accurately how many mappings a transparent hugepage
1967 * has (unlike page_mapcount() which isn't fully accurate). This full
1968 * accuracy is primarily needed to know if copy-on-write faults can
1969 * reuse the page and change the mapping to read-write instead of
1970 * copying them. At the same time this returns the total_mapcount too.
1971 *
1972 * The function returns the highest mapcount any one of the subpages
1973 * has. If the return value is one, even if different processes are
1974 * mapping different subpages of the transparent hugepage, they can
1975 * all reuse it, because each process is reusing a different subpage.
1976 *
1977 * The total_mapcount is instead counting all virtual mappings of the
1978 * subpages. If the total_mapcount is equal to "one", it tells the
1979 * caller all mappings belong to the same "mm" and in turn the
1980 * anon_vma of the transparent hugepage can become the vma->anon_vma
1981 * local one as no other process may be mapping any of the subpages.
1982 *
1983 * It would be more accurate to replace page_mapcount() with
1984 * page_trans_huge_mapcount(), however we only use
1985 * page_trans_huge_mapcount() in the copy-on-write faults where we
1986 * need full accuracy to avoid breaking page pinning, because
1987 * page_trans_huge_mapcount() is slower than page_mapcount().
1988 */
1989int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
1990{
1991 int i, ret, _total_mapcount, mapcount;
1992
1993 /* hugetlbfs shouldn't call it */
1994 VM_BUG_ON_PAGE(PageHuge(page), page);
1995
1996 if (likely(!PageTransCompound(page))) {
1997 mapcount = atomic_read(&page->_mapcount) + 1;
1998 if (total_mapcount)
1999 *total_mapcount = mapcount;
2000 return mapcount;
2001 }
2002
2003 page = compound_head(page);
2004
2005 _total_mapcount = ret = 0;
2006 for (i = 0; i < HPAGE_PMD_NR; i++) {
2007 mapcount = atomic_read(&page[i]._mapcount) + 1;
2008 ret = max(ret, mapcount);
2009 _total_mapcount += mapcount;
2010 }
2011 if (PageDoubleMap(page)) {
2012 ret -= 1;
2013 _total_mapcount -= HPAGE_PMD_NR;
2014 }
2015 mapcount = compound_mapcount(page);
2016 ret += mapcount;
2017 _total_mapcount += mapcount;
2018 if (total_mapcount)
2019 *total_mapcount = _total_mapcount;
2020 return ret;
2021}
2022
2023/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002024 * This function splits huge page into normal pages. @page can point to any
2025 * subpage of huge page to split. Split doesn't change the position of @page.
2026 *
2027 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2028 * The huge page must be locked.
2029 *
2030 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2031 *
2032 * Both head page and tail pages will inherit mapping, flags, and so on from
2033 * the hugepage.
2034 *
2035 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2036 * they are not mapped.
2037 *
2038 * Returns 0 if the hugepage is split successfully.
2039 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2040 * us.
2041 */
2042int split_huge_page_to_list(struct page *page, struct list_head *list)
2043{
2044 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002045 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002046 struct anon_vma *anon_vma = NULL;
2047 struct address_space *mapping = NULL;
2048 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002049 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002050 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002051
2052 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002053 VM_BUG_ON_PAGE(!PageLocked(page), page);
2054 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2055 VM_BUG_ON_PAGE(!PageCompound(page), page);
2056
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002057 if (PageAnon(head)) {
2058 /*
2059 * The caller does not necessarily hold an mmap_sem that would
2060 * prevent the anon_vma disappearing so we first we take a
2061 * reference to it and then lock the anon_vma for write. This
2062 * is similar to page_lock_anon_vma_read except the write lock
2063 * is taken to serialise against parallel split or collapse
2064 * operations.
2065 */
2066 anon_vma = page_get_anon_vma(head);
2067 if (!anon_vma) {
2068 ret = -EBUSY;
2069 goto out;
2070 }
2071 extra_pins = 0;
2072 mapping = NULL;
2073 anon_vma_lock_write(anon_vma);
2074 } else {
2075 mapping = head->mapping;
2076
2077 /* Truncated ? */
2078 if (!mapping) {
2079 ret = -EBUSY;
2080 goto out;
2081 }
2082
2083 /* Addidional pins from radix tree */
2084 extra_pins = HPAGE_PMD_NR;
2085 anon_vma = NULL;
2086 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002087 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002088
2089 /*
2090 * Racy check if we can split the page, before freeze_page() will
2091 * split PMDs
2092 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002093 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002094 ret = -EBUSY;
2095 goto out_unlock;
2096 }
2097
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002098 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002099 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002100 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2101
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002102 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2103 if (mlocked)
2104 lru_add_drain();
2105
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002106 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002107 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002108
2109 if (mapping) {
2110 void **pslot;
2111
2112 spin_lock(&mapping->tree_lock);
2113 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2114 page_index(head));
2115 /*
2116 * Check if the head page is present in radix tree.
2117 * We assume all tail are present too, if head is there.
2118 */
2119 if (radix_tree_deref_slot_protected(pslot,
2120 &mapping->tree_lock) != head)
2121 goto fail;
2122 }
2123
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002124 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002125 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002126 count = page_count(head);
2127 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002128 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002129 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002130 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002131 list_del(page_deferred_list(head));
2132 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002133 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002134 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002135 spin_unlock(&pgdata->split_queue_lock);
2136 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002137 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002138 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002139 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2140 pr_alert("total_mapcount: %u, page_count(): %u\n",
2141 mapcount, count);
2142 if (PageTail(page))
2143 dump_page(head, NULL);
2144 dump_page(page, "total_mapcount(head) > 0");
2145 BUG();
2146 }
2147 spin_unlock(&pgdata->split_queue_lock);
2148fail: if (mapping)
2149 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002150 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002151 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002152 ret = -EBUSY;
2153 }
2154
2155out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002156 if (anon_vma) {
2157 anon_vma_unlock_write(anon_vma);
2158 put_anon_vma(anon_vma);
2159 }
2160 if (mapping)
2161 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002162out:
2163 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2164 return ret;
2165}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002166
2167void free_transhuge_page(struct page *page)
2168{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002169 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002170 unsigned long flags;
2171
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002172 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002173 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002174 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002175 list_del(page_deferred_list(page));
2176 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002177 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002178 free_compound_page(page);
2179}
2180
2181void deferred_split_huge_page(struct page *page)
2182{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002183 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002184 unsigned long flags;
2185
2186 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2187
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002188 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002189 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002190 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002191 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2192 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002193 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002194 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002195}
2196
2197static unsigned long deferred_split_count(struct shrinker *shrink,
2198 struct shrink_control *sc)
2199{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002200 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002201 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002202}
2203
2204static unsigned long deferred_split_scan(struct shrinker *shrink,
2205 struct shrink_control *sc)
2206{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002207 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002208 unsigned long flags;
2209 LIST_HEAD(list), *pos, *next;
2210 struct page *page;
2211 int split = 0;
2212
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002213 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002214 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002215 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002216 page = list_entry((void *)pos, struct page, mapping);
2217 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002218 if (get_page_unless_zero(page)) {
2219 list_move(page_deferred_list(page), &list);
2220 } else {
2221 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002222 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002223 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002224 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002225 if (!--sc->nr_to_scan)
2226 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002227 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002228 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002229
2230 list_for_each_safe(pos, next, &list) {
2231 page = list_entry((void *)pos, struct page, mapping);
2232 lock_page(page);
2233 /* split_huge_page() removes page from list on success */
2234 if (!split_huge_page(page))
2235 split++;
2236 unlock_page(page);
2237 put_page(page);
2238 }
2239
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002240 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2241 list_splice_tail(&list, &pgdata->split_queue);
2242 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002243
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002244 /*
2245 * Stop shrinker if we didn't split any page, but the queue is empty.
2246 * This can happen if pages were freed under us.
2247 */
2248 if (!split && list_empty(&pgdata->split_queue))
2249 return SHRINK_STOP;
2250 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002251}
2252
2253static struct shrinker deferred_split_shrinker = {
2254 .count_objects = deferred_split_count,
2255 .scan_objects = deferred_split_scan,
2256 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002257 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002258};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002259
2260#ifdef CONFIG_DEBUG_FS
2261static int split_huge_pages_set(void *data, u64 val)
2262{
2263 struct zone *zone;
2264 struct page *page;
2265 unsigned long pfn, max_zone_pfn;
2266 unsigned long total = 0, split = 0;
2267
2268 if (val != 1)
2269 return -EINVAL;
2270
2271 for_each_populated_zone(zone) {
2272 max_zone_pfn = zone_end_pfn(zone);
2273 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2274 if (!pfn_valid(pfn))
2275 continue;
2276
2277 page = pfn_to_page(pfn);
2278 if (!get_page_unless_zero(page))
2279 continue;
2280
2281 if (zone != page_zone(page))
2282 goto next;
2283
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002284 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002285 goto next;
2286
2287 total++;
2288 lock_page(page);
2289 if (!split_huge_page(page))
2290 split++;
2291 unlock_page(page);
2292next:
2293 put_page(page);
2294 }
2295 }
2296
Yang Shi145bdaa2016-05-05 16:22:00 -07002297 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002298
2299 return 0;
2300}
2301DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2302 "%llu\n");
2303
2304static int __init split_huge_pages_debugfs(void)
2305{
2306 void *ret;
2307
Yang Shi145bdaa2016-05-05 16:22:00 -07002308 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002309 &split_huge_pages_fops);
2310 if (!ret)
2311 pr_warn("Failed to create split_huge_pages in debugfs");
2312 return 0;
2313}
2314late_initcall(split_huge_pages_debugfs);
2315#endif