blob: 724372866e676b3916ccdbf0b3ae66dd7ae7033a [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
David Rientjes6b7ff8e2018-03-22 16:17:45 -0700545 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
546 true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700547 put_page(page);
548 count_vm_event(THP_FAULT_FALLBACK);
549 return VM_FAULT_FALLBACK;
550 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800551
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700552 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700553 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800554 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700555 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700556 return VM_FAULT_OOM;
557 }
558
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800559 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700560 /*
561 * The memory barrier inside __SetPageUptodate makes sure that
562 * clear_huge_page writes become visible before the set_pmd_at()
563 * write.
564 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800565 __SetPageUptodate(page);
566
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700567 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
568 if (unlikely(!pmd_none(*fe->pmd))) {
569 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800570 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800571 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700572 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800573 } else {
574 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700575
576 /* Deliver the page fault to userland */
577 if (userfaultfd_missing(vma)) {
578 int ret;
579
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700580 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800581 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700582 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700583 pte_free(vma->vm_mm, pgtable);
584 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700585 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
586 return ret;
587 }
588
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700589 entry = mk_huge_pmd(page, vma->vm_page_prot);
590 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800591 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800592 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700593 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700594 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
595 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
596 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
597 atomic_long_inc(&vma->vm_mm->nr_ptes);
598 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700599 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800600 }
601
David Rientjesaa2e8782012-05-29 15:06:17 -0700602 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800603}
604
Mel Gorman444eb2a42016-03-17 14:19:23 -0700605/*
Vlastimil Babka25160352016-07-28 15:49:25 -0700606 * If THP defrag is set to always then directly reclaim/compact as necessary
607 * If set to defer then do only background reclaim/compact and defer to khugepaged
Mel Gorman444eb2a42016-03-17 14:19:23 -0700608 * If set to madvise and the VMA is flagged then directly reclaim/compact
Vlastimil Babka25160352016-07-28 15:49:25 -0700609 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
Mel Gorman444eb2a42016-03-17 14:19:23 -0700610 */
611static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800612{
Vlastimil Babka25160352016-07-28 15:49:25 -0700613 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700614
Vlastimil Babka25160352016-07-28 15:49:25 -0700615 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
616 &transparent_hugepage_flags) && vma_madvised)
617 return GFP_TRANSHUGE;
618 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
619 &transparent_hugepage_flags))
620 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
621 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
622 &transparent_hugepage_flags))
623 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700624
Vlastimil Babka25160352016-07-28 15:49:25 -0700625 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700626}
627
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800628/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700629static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800630 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700631 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800632{
633 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700634 if (!pmd_none(*pmd))
635 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700636 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800637 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800638 if (pgtable)
639 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800640 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800641 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700642 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800643}
644
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700645int do_huge_pmd_anonymous_page(struct fault_env *fe)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800646{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700647 struct vm_area_struct *vma = fe->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800648 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800649 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700650 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800651
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700652 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700653 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700654 if (unlikely(anon_vma_prepare(vma)))
655 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700656 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700657 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700658 if (!(fe->flags & FAULT_FLAG_WRITE) &&
659 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700660 transparent_hugepage_use_zero_page()) {
661 pgtable_t pgtable;
662 struct page *zero_page;
663 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700664 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700665 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700666 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800667 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700668 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700669 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700670 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700671 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700672 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700673 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700674 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700675 ret = 0;
676 set = false;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700677 if (pmd_none(*fe->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700678 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700679 spin_unlock(fe->ptl);
680 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700681 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
682 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700683 set_huge_zero_page(pgtable, vma->vm_mm, vma,
684 haddr, fe->pmd, zero_page);
685 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700686 set = true;
687 }
688 } else
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700689 spin_unlock(fe->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700690 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700691 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700692 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800693 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700694 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800695 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700696 if (unlikely(!page)) {
697 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700698 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700699 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800700 prep_transhuge_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700701 return __do_huge_pmd_anonymous_page(fe, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800702}
703
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700704static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800705 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700706{
707 struct mm_struct *mm = vma->vm_mm;
708 pmd_t entry;
709 spinlock_t *ptl;
710
711 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800712 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
713 if (pfn_t_devmap(pfn))
714 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800715 if (write) {
716 entry = pmd_mkyoung(pmd_mkdirty(entry));
717 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700718 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800719 set_pmd_at(mm, addr, pmd, entry);
720 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700721 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700722}
723
724int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800725 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700726{
727 pgprot_t pgprot = vma->vm_page_prot;
728 /*
729 * If we had pmd_special, we could avoid all these restrictions,
730 * but we need to be consistent with PTEs and architectures that
731 * can't support a 'special' bit.
732 */
733 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
734 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
735 (VM_PFNMAP|VM_MIXEDMAP));
736 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800737 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700738
739 if (addr < vma->vm_start || addr >= vma->vm_end)
740 return VM_FAULT_SIGBUS;
741 if (track_pfn_insert(vma, &pgprot, pfn))
742 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700743 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
744 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700745}
Dan Williamsdee41072016-05-14 12:20:44 -0700746EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700747
Dan Williams3565fce2016-01-15 16:56:55 -0800748static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300749 pmd_t *pmd, int flags)
Dan Williams3565fce2016-01-15 16:56:55 -0800750{
751 pmd_t _pmd;
752
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300753 _pmd = pmd_mkyoung(*pmd);
754 if (flags & FOLL_WRITE)
755 _pmd = pmd_mkdirty(_pmd);
Dan Williams3565fce2016-01-15 16:56:55 -0800756 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300757 pmd, _pmd, flags & FOLL_WRITE))
Dan Williams3565fce2016-01-15 16:56:55 -0800758 update_mmu_cache_pmd(vma, addr, pmd);
759}
760
761struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
762 pmd_t *pmd, int flags)
763{
764 unsigned long pfn = pmd_pfn(*pmd);
765 struct mm_struct *mm = vma->vm_mm;
766 struct dev_pagemap *pgmap;
767 struct page *page;
768
769 assert_spin_locked(pmd_lockptr(mm, pmd));
770
Keno Fischer6676aa62017-01-24 15:17:48 -0800771 /*
772 * When we COW a devmap PMD entry, we split it into PTEs, so we should
773 * not be in this function with `flags & FOLL_COW` set.
774 */
775 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
776
Dan Williams3565fce2016-01-15 16:56:55 -0800777 if (flags & FOLL_WRITE && !pmd_write(*pmd))
778 return NULL;
779
780 if (pmd_present(*pmd) && pmd_devmap(*pmd))
781 /* pass */;
782 else
783 return NULL;
784
785 if (flags & FOLL_TOUCH)
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300786 touch_pmd(vma, addr, pmd, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800787
788 /*
789 * device mapped pages can only be returned if the
790 * caller will manage the page reference count.
791 */
792 if (!(flags & FOLL_GET))
793 return ERR_PTR(-EEXIST);
794
795 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
796 pgmap = get_dev_pagemap(pfn, NULL);
797 if (!pgmap)
798 return ERR_PTR(-EFAULT);
799 page = pfn_to_page(pfn);
800 get_page(page);
801 put_dev_pagemap(pgmap);
802
803 return page;
804}
805
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800806int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
807 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
808 struct vm_area_struct *vma)
809{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800810 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800811 struct page *src_page;
812 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800813 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700814 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800815
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700816 /* Skip if can be re-fill on fault */
817 if (!vma_is_anonymous(vma))
818 return 0;
819
820 pgtable = pte_alloc_one(dst_mm, addr);
821 if (unlikely(!pgtable))
822 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800823
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800824 dst_ptl = pmd_lock(dst_mm, dst_pmd);
825 src_ptl = pmd_lockptr(src_mm, src_pmd);
826 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800827
828 ret = -EAGAIN;
829 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700830 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800831 pte_free(dst_mm, pgtable);
832 goto out_unlock;
833 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800834 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800835 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800836 * under splitting since we don't split the page itself, only pmd to
837 * a page table.
838 */
839 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700840 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800841 /*
842 * get_huge_zero_page() will never allocate a new page here,
843 * since we already have a zero page to copy. It just takes a
844 * reference.
845 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700846 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700847 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700848 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800849 ret = 0;
850 goto out_unlock;
851 }
Mel Gormande466bd2013-12-18 17:08:42 -0800852
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700853 src_page = pmd_page(pmd);
854 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
855 get_page(src_page);
856 page_dup_rmap(src_page, true);
857 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
858 atomic_long_inc(&dst_mm->nr_ptes);
859 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800860
861 pmdp_set_wrprotect(src_mm, addr, src_pmd);
862 pmd = pmd_mkold(pmd_wrprotect(pmd));
863 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800864
865 ret = 0;
866out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800867 spin_unlock(src_ptl);
868 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800869out:
870 return ret;
871}
872
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700873void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -0800874{
875 pmd_t entry;
876 unsigned long haddr;
Minchan Kim8edd3652017-01-10 16:57:51 -0800877 bool write = fe->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -0800878
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700879 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
880 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -0800881 goto unlock;
882
883 entry = pmd_mkyoung(orig_pmd);
Minchan Kim8edd3652017-01-10 16:57:51 -0800884 if (write)
885 entry = pmd_mkdirty(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700886 haddr = fe->address & HPAGE_PMD_MASK;
Minchan Kim8edd3652017-01-10 16:57:51 -0800887 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700888 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800889
890unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700891 spin_unlock(fe->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800892}
893
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700894static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
895 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700897 struct vm_area_struct *vma = fe->vma;
898 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -0700899 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800900 pgtable_t pgtable;
901 pmd_t _pmd;
902 int ret = 0, i;
903 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700904 unsigned long mmun_start; /* For mmu_notifiers */
905 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800906
907 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
908 GFP_KERNEL);
909 if (unlikely(!pages)) {
910 ret |= VM_FAULT_OOM;
911 goto out;
912 }
913
914 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700915 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700916 __GFP_OTHER_NODE, vma,
917 fe->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800918 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700919 mem_cgroup_try_charge(pages[i], vma->vm_mm,
920 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800921 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800922 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800923 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700924 memcg = (void *)page_private(pages[i]);
925 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800926 mem_cgroup_cancel_charge(pages[i], memcg,
927 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800928 put_page(pages[i]);
929 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800930 kfree(pages);
931 ret |= VM_FAULT_OOM;
932 goto out;
933 }
Johannes Weiner00501b52014-08-08 14:19:20 -0700934 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800935 }
936
937 for (i = 0; i < HPAGE_PMD_NR; i++) {
938 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700939 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800940 __SetPageUptodate(pages[i]);
941 cond_resched();
942 }
943
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700944 mmun_start = haddr;
945 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700946 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700947
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700948 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
949 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800950 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -0800951 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800952
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700953 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800954 /* leave pmd empty until pte is filled */
955
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700956 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
957 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800958
959 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700960 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800961 entry = mk_pte(pages[i], vma->vm_page_prot);
962 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -0700963 memcg = (void *)page_private(pages[i]);
964 set_page_private(pages[i], 0);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700965 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800966 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -0700967 lru_cache_add_active_or_unevictable(pages[i], vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700968 fe->pte = pte_offset_map(&_pmd, haddr);
969 VM_BUG_ON(!pte_none(*fe->pte));
970 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
971 pte_unmap(fe->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800972 }
973 kfree(pages);
974
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800975 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700976 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800977 page_remove_rmap(page, true);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700978 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800979
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700980 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700981
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800982 ret |= VM_FAULT_WRITE;
983 put_page(page);
984
985out:
986 return ret;
987
988out_free_pages:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700989 spin_unlock(fe->ptl);
990 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800991 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700992 memcg = (void *)page_private(pages[i]);
993 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800994 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800995 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800996 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800997 kfree(pages);
998 goto out;
999}
1000
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001001int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001002{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001003 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001004 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001005 struct mem_cgroup *memcg;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001006 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001007 unsigned long mmun_start; /* For mmu_notifiers */
1008 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001009 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001010 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001011
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001012 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001013 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001014 if (is_huge_zero_pmd(orig_pmd))
1015 goto alloc;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001016 spin_lock(fe->ptl);
1017 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001018 goto out_unlock;
1019
1020 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001021 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001022 /*
1023 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001024 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001025 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001026 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001027 pmd_t entry;
1028 entry = pmd_mkyoung(orig_pmd);
1029 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001030 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1031 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001032 ret |= VM_FAULT_WRITE;
1033 goto out_unlock;
1034 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001035 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001036 spin_unlock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001037alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001038 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001039 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001040 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001041 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001042 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001043 new_page = NULL;
1044
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001045 if (likely(new_page)) {
1046 prep_transhuge_page(new_page);
1047 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001048 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001049 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001050 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001051 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001052 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001053 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001054 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001055 ret |= VM_FAULT_FALLBACK;
1056 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001057 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001058 }
David Rientjes17766dd2013-09-12 15:14:06 -07001059 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001060 goto out;
1061 }
1062
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001063 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
David Rientjes6b7ff8e2018-03-22 16:17:45 -07001064 huge_gfp | __GFP_NORETRY, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001065 put_page(new_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001066 split_huge_pmd(vma, fe->pmd, fe->address);
1067 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001068 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001069 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001070 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001071 goto out;
1072 }
1073
David Rientjes17766dd2013-09-12 15:14:06 -07001074 count_vm_event(THP_FAULT_ALLOC);
1075
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001076 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001077 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1078 else
1079 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 __SetPageUptodate(new_page);
1081
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001082 mmun_start = haddr;
1083 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001084 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001085
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001086 spin_lock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001087 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001088 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001089 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1090 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001091 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001092 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001093 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001094 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001095 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001096 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1097 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001098 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001099 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001100 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001101 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001102 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1103 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001104 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001105 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001106 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001107 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001108 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001109 put_page(page);
1110 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001111 ret |= VM_FAULT_WRITE;
1112 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001113 spin_unlock(fe->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001114out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001115 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001116out:
1117 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001118out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001119 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001120 return ret;
1121}
1122
Keno Fischer6676aa62017-01-24 15:17:48 -08001123/*
1124 * FOLL_FORCE can write to even unwritable pmd's, but only
1125 * after we've gone through a COW cycle and they are dirty.
1126 */
1127static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1128{
1129 return pmd_write(pmd) ||
1130 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1131}
1132
David Rientjesb676b292012-10-08 16:34:03 -07001133struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134 unsigned long addr,
1135 pmd_t *pmd,
1136 unsigned int flags)
1137{
David Rientjesb676b292012-10-08 16:34:03 -07001138 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001139 struct page *page = NULL;
1140
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001141 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001142
Keno Fischer6676aa62017-01-24 15:17:48 -08001143 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001144 goto out;
1145
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001146 /* Avoid dumping huge zero page */
1147 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1148 return ERR_PTR(-EFAULT);
1149
Mel Gorman2b4847e2013-12-18 17:08:32 -08001150 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001151 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001152 goto out;
1153
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001154 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001155 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001156 if (flags & FOLL_TOUCH)
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +03001157 touch_pmd(vma, addr, pmd, flags);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001158 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001159 /*
1160 * We don't mlock() pte-mapped THPs. This way we can avoid
1161 * leaking mlocked pages into non-VM_LOCKED VMAs.
1162 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001163 * For anon THP:
1164 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001165 * In most cases the pmd is the only mapping of the page as we
1166 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1167 * writable private mappings in populate_vma_page_range().
1168 *
1169 * The only scenario when we have the page shared here is if we
1170 * mlocking read-only mapping shared over fork(). We skip
1171 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001172 *
1173 * For file THP:
1174 *
1175 * We can expect PageDoubleMap() to be stable under page lock:
1176 * for file pages we set it in page_add_file_rmap(), which
1177 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001178 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001179
1180 if (PageAnon(page) && compound_mapcount(page) != 1)
1181 goto skip_mlock;
1182 if (PageDoubleMap(page) || !page->mapping)
1183 goto skip_mlock;
1184 if (!trylock_page(page))
1185 goto skip_mlock;
1186 lru_add_drain();
1187 if (page->mapping && !PageDoubleMap(page))
1188 mlock_vma_page(page);
1189 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001190 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001191skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001192 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001193 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001194 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001195 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001196
1197out:
1198 return page;
1199}
1200
Mel Gormand10e63f2012-10-25 14:16:31 +02001201/* NUMA hinting page fault entry point for trans huge pmds */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001202int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001203{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001204 struct vm_area_struct *vma = fe->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001205 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001206 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001207 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001208 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001209 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001210 bool page_locked;
1211 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001212 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001213 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001214
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001215 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1216 if (unlikely(!pmd_same(pmd, *fe->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001217 goto out_unlock;
1218
Mel Gormande466bd2013-12-18 17:08:42 -08001219 /*
1220 * If there are potential migrations, wait for completion and retry
1221 * without disrupting NUMA hinting information. Do not relock and
1222 * check_same as the page may no longer be mapped.
1223 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001224 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1225 page = pmd_page(*fe->pmd);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001226 if (!get_page_unless_zero(page))
1227 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001228 spin_unlock(fe->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001229 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001230 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001231 goto out;
1232 }
1233
Mel Gormand10e63f2012-10-25 14:16:31 +02001234 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001235 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001236 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001237 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001238 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001239 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001240 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001241 flags |= TNF_FAULT_LOCAL;
1242 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001243
Mel Gormanbea66fb2015-03-25 15:55:37 -07001244 /* See similar comment in do_numa_page for explanation */
Rik van Rield59dc7b2016-09-08 21:30:53 -04001245 if (!pmd_write(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001246 flags |= TNF_NO_GROUP;
1247
1248 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001249 * Acquire the page lock to serialise THP migrations but avoid dropping
1250 * page_table_lock if at all possible
1251 */
Mel Gormanb8916632013-10-07 11:28:44 +01001252 page_locked = trylock_page(page);
1253 target_nid = mpol_misplaced(page, vma, haddr);
1254 if (target_nid == -1) {
1255 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001256 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001257 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001258 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001259
Mel Gormande466bd2013-12-18 17:08:42 -08001260 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001261 if (!page_locked) {
Mark Rutland2aa6d032017-06-16 14:02:34 -07001262 if (!get_page_unless_zero(page))
1263 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001264 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001265 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001266 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001267 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001268 goto out;
1269 }
1270
Mel Gorman2b4847e2013-12-18 17:08:32 -08001271 /*
1272 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1273 * to serialises splits
1274 */
Mel Gormanb8916632013-10-07 11:28:44 +01001275 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001276 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001277 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001278
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001279 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001280 spin_lock(fe->ptl);
1281 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001282 unlock_page(page);
1283 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001284 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001285 goto out_unlock;
1286 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001287
Mel Gormanc3a489c2013-12-18 17:08:38 -08001288 /* Bail if we fail to protect against THP splits for any reason */
1289 if (unlikely(!anon_vma)) {
1290 put_page(page);
1291 page_nid = -1;
1292 goto clear_pmdnuma;
1293 }
1294
Mel Gormana54a4072013-10-07 11:28:46 +01001295 /*
1296 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001297 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001298 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001299 spin_unlock(fe->ptl);
1300 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1301 fe->pmd, pmd, fe->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001302 if (migrated) {
1303 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001304 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001305 } else
1306 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001307
Mel Gorman8191acb2013-10-07 11:28:45 +01001308 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001309clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001310 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001311 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001312 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001313 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001314 if (was_writable)
1315 pmd = pmd_mkwrite(pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001316 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1317 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001318 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001319out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001320 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001321
1322out:
1323 if (anon_vma)
1324 page_unlock_anon_vma_read(anon_vma);
1325
Mel Gorman8191acb2013-10-07 11:28:45 +01001326 if (page_nid != -1)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001327 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001328
Mel Gormand10e63f2012-10-25 14:16:31 +02001329 return 0;
1330}
1331
Huang Ying319904a2016-07-28 15:48:03 -07001332/*
1333 * Return true if we do MADV_FREE successfully on entire pmd page.
1334 * Otherwise, return false.
1335 */
1336bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001337 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001338{
1339 spinlock_t *ptl;
1340 pmd_t orig_pmd;
1341 struct page *page;
1342 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001343 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001344
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001345 ptl = pmd_trans_huge_lock(pmd, vma);
1346 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001347 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001348
1349 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001350 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001351 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001352
1353 page = pmd_page(orig_pmd);
1354 /*
1355 * If other processes are mapping this page, we couldn't discard
1356 * the page unless they all do MADV_FREE so let's skip the page.
1357 */
1358 if (page_mapcount(page) != 1)
1359 goto out;
1360
1361 if (!trylock_page(page))
1362 goto out;
1363
1364 /*
1365 * If user want to discard part-pages of THP, split it so MADV_FREE
1366 * will deactivate only them.
1367 */
1368 if (next - addr != HPAGE_PMD_SIZE) {
1369 get_page(page);
1370 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001371 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001372 unlock_page(page);
Kirill A. Shutemovd2b64682017-07-06 15:35:28 -07001373 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001374 goto out_unlocked;
1375 }
1376
1377 if (PageDirty(page))
1378 ClearPageDirty(page);
1379 unlock_page(page);
1380
1381 if (PageActive(page))
1382 deactivate_page(page);
1383
1384 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemovf5848032017-04-13 14:56:26 -07001385 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001386 orig_pmd = pmd_mkold(orig_pmd);
1387 orig_pmd = pmd_mkclean(orig_pmd);
1388
1389 set_pmd_at(mm, addr, pmd, orig_pmd);
1390 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1391 }
Huang Ying319904a2016-07-28 15:48:03 -07001392 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001393out:
1394 spin_unlock(ptl);
1395out_unlocked:
1396 return ret;
1397}
1398
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001399int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001400 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001401{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001402 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001403 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001404
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001405 ptl = __pmd_trans_huge_lock(pmd, vma);
1406 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001407 return 0;
1408 /*
1409 * For architectures like ppc64 we look at deposited pgtable
1410 * when calling pmdp_huge_get_and_clear. So do the
1411 * pgtable_trans_huge_withdraw after finishing pmdp related
1412 * operations.
1413 */
1414 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1415 tlb->fullmm);
1416 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1417 if (vma_is_dax(vma)) {
1418 spin_unlock(ptl);
1419 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001420 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001421 } else if (is_huge_zero_pmd(orig_pmd)) {
1422 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1423 atomic_long_dec(&tlb->mm->nr_ptes);
1424 spin_unlock(ptl);
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001425 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001426 } else {
1427 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001428 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001429 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001430 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001431 if (PageAnon(page)) {
1432 pgtable_t pgtable;
1433 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1434 pte_free(tlb->mm, pgtable);
1435 atomic_long_dec(&tlb->mm->nr_ptes);
1436 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1437 } else {
1438 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1439 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001440 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001441 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001442 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001443 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001444}
1445
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001446bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001447 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001448 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001449{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001450 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001451 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001452 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001453 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001454
1455 if ((old_addr & ~HPAGE_PMD_MASK) ||
1456 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001457 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001458 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001459
1460 /*
1461 * The destination pmd shouldn't be established, free_pgtables()
1462 * should have release it.
1463 */
1464 if (WARN_ON(!pmd_none(*new_pmd))) {
1465 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001466 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001467 }
1468
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001469 /*
1470 * We don't have to worry about the ordering of src and dst
1471 * ptlocks because exclusive mmap_sem prevents deadlock.
1472 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001473 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1474 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001475 new_ptl = pmd_lockptr(mm, new_pmd);
1476 if (new_ptl != old_ptl)
1477 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001478 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001479 if (pmd_present(pmd) && pmd_dirty(pmd))
1480 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001481 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001482
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001483 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1484 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301485 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001486 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1487 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001488 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301489 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1490 if (new_ptl != old_ptl)
1491 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001492 if (force_flush)
1493 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1494 else
1495 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001496 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001497 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001498 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001499 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001500}
1501
Mel Gormanf123d742013-10-07 11:28:49 +01001502/*
1503 * Returns
1504 * - 0 if PMD could not be locked
1505 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1506 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1507 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001508int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001509 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001510{
1511 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001512 spinlock_t *ptl;
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001513 pmd_t entry;
1514 bool preserve_write;
1515 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001516
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001517 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001518 if (!ptl)
1519 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001520
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001521 preserve_write = prot_numa && pmd_write(*pmd);
1522 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001523
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001524 /*
1525 * Avoid trapping faults against the zero page. The read-only
1526 * data is likely to be read-cached on the local CPU and
1527 * local/remote hits to the zero page are not interesting.
1528 */
1529 if (prot_numa && is_huge_zero_pmd(*pmd))
1530 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001531
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001532 if (prot_numa && pmd_protnone(*pmd))
1533 goto unlock;
1534
Kirill A. Shutemovc2edc332017-04-13 14:56:20 -07001535 /*
1536 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1537 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1538 * which is also under down_read(mmap_sem):
1539 *
1540 * CPU0: CPU1:
1541 * change_huge_pmd(prot_numa=1)
1542 * pmdp_huge_get_and_clear_notify()
1543 * madvise_dontneed()
1544 * zap_pmd_range()
1545 * pmd_trans_huge(*pmd) == 0 (without ptl)
1546 * // skip the pmd
1547 * set_pmd_at();
1548 * // pmd is re-established
1549 *
1550 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1551 * which may break userspace.
1552 *
1553 * pmdp_invalidate() is required to make sure we don't miss
1554 * dirty/young flags set by hardware.
1555 */
1556 entry = *pmd;
1557 pmdp_invalidate(vma, addr, pmd);
1558
1559 /*
1560 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1561 * corrupt them.
1562 */
1563 if (pmd_dirty(*pmd))
1564 entry = pmd_mkdirty(entry);
1565 if (pmd_young(*pmd))
1566 entry = pmd_mkyoung(entry);
1567
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001568 entry = pmd_modify(entry, newprot);
1569 if (preserve_write)
1570 entry = pmd_mkwrite(entry);
1571 ret = HPAGE_PMD_NR;
1572 set_pmd_at(mm, addr, pmd, entry);
1573 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1574unlock:
1575 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001576 return ret;
1577}
1578
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001579/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001580 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001581 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001582 * Note that if it returns page table lock pointer, this routine returns without
1583 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001584 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001585spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001586{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001587 spinlock_t *ptl;
1588 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001589 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001590 return ptl;
1591 spin_unlock(ptl);
1592 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001593}
1594
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001595static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1596 unsigned long haddr, pmd_t *pmd)
1597{
1598 struct mm_struct *mm = vma->vm_mm;
1599 pgtable_t pgtable;
1600 pmd_t _pmd;
1601 int i;
1602
1603 /* leave pmd empty until pte is filled */
1604 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1605
1606 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1607 pmd_populate(mm, &_pmd, pgtable);
1608
1609 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1610 pte_t *pte, entry;
1611 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1612 entry = pte_mkspecial(entry);
1613 pte = pte_offset_map(&_pmd, haddr);
1614 VM_BUG_ON(!pte_none(*pte));
1615 set_pte_at(mm, haddr, pte, entry);
1616 pte_unmap(pte);
1617 }
1618 smp_wmb(); /* make pte visible before pmd */
1619 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001620}
1621
1622static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001623 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001624{
1625 struct mm_struct *mm = vma->vm_mm;
1626 struct page *page;
1627 pgtable_t pgtable;
1628 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001629 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001630 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001631 int i;
1632
1633 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1634 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1635 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001636 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001637
1638 count_vm_event(THP_SPLIT_PMD);
1639
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001640 if (!vma_is_anonymous(vma)) {
1641 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001642 if (vma_is_dax(vma))
1643 return;
1644 page = pmd_page(_pmd);
1645 if (!PageReferenced(page) && pmd_young(_pmd))
1646 SetPageReferenced(page);
1647 page_remove_rmap(page, true);
1648 put_page(page);
1649 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001650 return;
1651 } else if (is_huge_zero_pmd(*pmd)) {
1652 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1653 }
1654
1655 page = pmd_page(*pmd);
1656 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001657 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001658 write = pmd_write(*pmd);
1659 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001660 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001661 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001662
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301663 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001664 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1665 pmd_populate(mm, &_pmd, pgtable);
1666
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001667 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001668 pte_t entry, *pte;
1669 /*
1670 * Note that NUMA hinting access restrictions are not
1671 * transferred to avoid any possibility of altering
1672 * permissions across VMAs.
1673 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001674 if (freeze) {
1675 swp_entry_t swp_entry;
1676 swp_entry = make_migration_entry(page + i, write);
1677 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001678 if (soft_dirty)
1679 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001680 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001681 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001682 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001683 if (!write)
1684 entry = pte_wrprotect(entry);
1685 if (!young)
1686 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001687 if (soft_dirty)
1688 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001689 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001690 if (dirty)
1691 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001692 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001693 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001694 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001695 atomic_inc(&page[i]._mapcount);
1696 pte_unmap(pte);
1697 }
1698
1699 /*
1700 * Set PG_double_map before dropping compound_mapcount to avoid
1701 * false-negative page_mapped().
1702 */
1703 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1704 for (i = 0; i < HPAGE_PMD_NR; i++)
1705 atomic_inc(&page[i]._mapcount);
1706 }
1707
1708 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1709 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001710 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001711 if (TestClearPageDoubleMap(page)) {
1712 /* No need in mapcount reference anymore */
1713 for (i = 0; i < HPAGE_PMD_NR; i++)
1714 atomic_dec(&page[i]._mapcount);
1715 }
1716 }
1717
1718 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001719 /*
1720 * Up to this point the pmd is present and huge and userland has the
1721 * whole access to the hugepage during the split (which happens in
1722 * place). If we overwrite the pmd with the not-huge version pointing
1723 * to the pte here (which of course we could if all CPUs were bug
1724 * free), userland could trigger a small page size TLB miss on the
1725 * small sized TLB while the hugepage TLB entry is still established in
1726 * the huge TLB. Some CPU doesn't like that.
1727 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1728 * 383 on page 93. Intel should be safe but is also warns that it's
1729 * only safe if the permission and cache attributes of the two entries
1730 * loaded in the two TLB is identical (which should be the case here).
1731 * But it is generally safer to never allow small and huge TLB entries
1732 * for the same virtual address to be loaded simultaneously. So instead
1733 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1734 * current pmd notpresent (atomically because here the pmd_trans_huge
1735 * and pmd_trans_splitting must remain set at all times on the pmd
1736 * until the split is complete for this pmd), then we flush the SMP TLB
1737 * and finally we write the non-huge version of the pmd entry with
1738 * pmd_populate.
1739 */
1740 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001741 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001742
1743 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001744 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001745 page_remove_rmap(page + i, false);
1746 put_page(page + i);
1747 }
1748 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001749}
1750
1751void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001752 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001753{
1754 spinlock_t *ptl;
1755 struct mm_struct *mm = vma->vm_mm;
1756 unsigned long haddr = address & HPAGE_PMD_MASK;
1757
1758 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1759 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001760
1761 /*
1762 * If caller asks to setup a migration entries, we need a page to check
1763 * pmd against. Otherwise we can end up replacing wrong page.
1764 */
1765 VM_BUG_ON(freeze && !page);
1766 if (page && page != pmd_page(*pmd))
1767 goto out;
1768
Dan Williams5c7fb562016-01-15 16:56:52 -08001769 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001770 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001771 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07001772 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08001773 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001774 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001775 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001776out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001777 spin_unlock(ptl);
1778 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1779}
1780
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001781void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1782 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001783{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001784 pgd_t *pgd;
1785 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001786 pmd_t *pmd;
1787
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001788 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001789 if (!pgd_present(*pgd))
1790 return;
1791
1792 pud = pud_offset(pgd, address);
1793 if (!pud_present(*pud))
1794 return;
1795
1796 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001797
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001798 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001799}
1800
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07001801void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001802 unsigned long start,
1803 unsigned long end,
1804 long adjust_next)
1805{
1806 /*
1807 * If the new start address isn't hpage aligned and it could
1808 * previously contain an hugepage: check if we need to split
1809 * an huge pmd.
1810 */
1811 if (start & ~HPAGE_PMD_MASK &&
1812 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1813 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001814 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001815
1816 /*
1817 * If the new end address isn't hpage aligned and it could
1818 * previously contain an hugepage: check if we need to split
1819 * an huge pmd.
1820 */
1821 if (end & ~HPAGE_PMD_MASK &&
1822 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1823 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001824 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001825
1826 /*
1827 * If we're also updating the vma->vm_next->vm_start, if the new
1828 * vm_next->vm_start isn't page aligned and it could previously
1829 * contain an hugepage: check if we need to split an huge pmd.
1830 */
1831 if (adjust_next > 0) {
1832 struct vm_area_struct *next = vma->vm_next;
1833 unsigned long nstart = next->vm_start;
1834 nstart += adjust_next << PAGE_SHIFT;
1835 if (nstart & ~HPAGE_PMD_MASK &&
1836 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1837 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001838 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001839 }
1840}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001841
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001842static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001843{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001844 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1845 TTU_RMAP_LOCKED;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001846 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001847
1848 VM_BUG_ON_PAGE(!PageHead(page), page);
1849
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001850 if (PageAnon(page))
1851 ttu_flags |= TTU_MIGRATION;
1852
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001853 /* We only need TTU_SPLIT_HUGE_PMD once */
1854 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1855 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1856 /* Cut short if the page is unmapped */
1857 if (page_count(page) == 1)
1858 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001859
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001860 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001861 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001862 VM_BUG_ON_PAGE(ret, page + i - 1);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001863}
1864
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001865static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001866{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001867 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001868
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001869 for (i = 0; i < HPAGE_PMD_NR; i++)
1870 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001871}
1872
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001873static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001874 struct lruvec *lruvec, struct list_head *list)
1875{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001876 struct page *page_tail = head + tail;
1877
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001878 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001879 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001880
1881 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001882 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001883 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001884 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1885 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001886 * get_page_unless_zero(), and atomic_set() is implemented in C not
1887 * using locked ops. spin_unlock on x86 sometime uses locked ops
1888 * because of PPro errata 66, 92, so unless somebody can guarantee
1889 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001890 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001891 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001892 if (PageAnon(head)) {
1893 page_ref_inc(page_tail);
1894 } else {
1895 /* Additional pin to radix tree */
1896 page_ref_add(page_tail, 2);
1897 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001898
1899 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1900 page_tail->flags |= (head->flags &
1901 ((1L << PG_referenced) |
1902 (1L << PG_swapbacked) |
1903 (1L << PG_mlocked) |
1904 (1L << PG_uptodate) |
1905 (1L << PG_active) |
1906 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001907 (1L << PG_unevictable) |
1908 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001909
1910 /*
1911 * After clearing PageTail the gup refcount can be released.
1912 * Page flags also must be visible before we make the page non-compound.
1913 */
1914 smp_wmb();
1915
1916 clear_compound_head(page_tail);
1917
1918 if (page_is_young(head))
1919 set_page_young(page_tail);
1920 if (page_is_idle(head))
1921 set_page_idle(page_tail);
1922
1923 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001924 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001925 page_tail);
1926 page_tail->mapping = head->mapping;
1927
1928 page_tail->index = head->index + tail;
1929 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1930 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001931}
1932
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001933static void __split_huge_page(struct page *page, struct list_head *list,
1934 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001935{
1936 struct page *head = compound_head(page);
1937 struct zone *zone = page_zone(head);
1938 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001939 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001940 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001941
Mel Gorman599d0c92016-07-28 15:45:31 -07001942 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001943
1944 /* complete memcg works before add pages to LRU */
1945 mem_cgroup_split_huge_fixup(head);
1946
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001947 if (!PageAnon(page))
1948 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1949
1950 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001951 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001952 /* Some pages can be beyond i_size: drop them from page cache */
1953 if (head[i].index >= end) {
1954 __ClearPageDirty(head + i);
1955 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001956 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1957 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001958 put_page(head + i);
1959 }
1960 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001961
1962 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001963 /* See comment in __split_huge_page_tail() */
1964 if (PageAnon(head)) {
1965 page_ref_inc(head);
1966 } else {
1967 /* Additional pin to radix tree */
1968 page_ref_add(head, 2);
1969 spin_unlock(&head->mapping->tree_lock);
1970 }
1971
Mel Gormana52633d2016-07-28 15:45:28 -07001972 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001973
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001974 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001975
1976 for (i = 0; i < HPAGE_PMD_NR; i++) {
1977 struct page *subpage = head + i;
1978 if (subpage == page)
1979 continue;
1980 unlock_page(subpage);
1981
1982 /*
1983 * Subpages may be freed if there wasn't any mapping
1984 * like if add_to_swap() is running on a lru page that
1985 * had its mapping zapped. And freeing these pages
1986 * requires taking the lru_lock so we do the put_page
1987 * of the tail pages after the split is complete.
1988 */
1989 put_page(subpage);
1990 }
1991}
1992
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001993int total_mapcount(struct page *page)
1994{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001995 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001996
1997 VM_BUG_ON_PAGE(PageTail(page), page);
1998
1999 if (likely(!PageCompound(page)))
2000 return atomic_read(&page->_mapcount) + 1;
2001
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002002 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002003 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002004 return compound;
2005 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002006 for (i = 0; i < HPAGE_PMD_NR; i++)
2007 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002008 /* File pages has compound_mapcount included in _mapcount */
2009 if (!PageAnon(page))
2010 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002011 if (PageDoubleMap(page))
2012 ret -= HPAGE_PMD_NR;
2013 return ret;
2014}
2015
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002016/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002017 * This calculates accurately how many mappings a transparent hugepage
2018 * has (unlike page_mapcount() which isn't fully accurate). This full
2019 * accuracy is primarily needed to know if copy-on-write faults can
2020 * reuse the page and change the mapping to read-write instead of
2021 * copying them. At the same time this returns the total_mapcount too.
2022 *
2023 * The function returns the highest mapcount any one of the subpages
2024 * has. If the return value is one, even if different processes are
2025 * mapping different subpages of the transparent hugepage, they can
2026 * all reuse it, because each process is reusing a different subpage.
2027 *
2028 * The total_mapcount is instead counting all virtual mappings of the
2029 * subpages. If the total_mapcount is equal to "one", it tells the
2030 * caller all mappings belong to the same "mm" and in turn the
2031 * anon_vma of the transparent hugepage can become the vma->anon_vma
2032 * local one as no other process may be mapping any of the subpages.
2033 *
2034 * It would be more accurate to replace page_mapcount() with
2035 * page_trans_huge_mapcount(), however we only use
2036 * page_trans_huge_mapcount() in the copy-on-write faults where we
2037 * need full accuracy to avoid breaking page pinning, because
2038 * page_trans_huge_mapcount() is slower than page_mapcount().
2039 */
2040int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2041{
2042 int i, ret, _total_mapcount, mapcount;
2043
2044 /* hugetlbfs shouldn't call it */
2045 VM_BUG_ON_PAGE(PageHuge(page), page);
2046
2047 if (likely(!PageTransCompound(page))) {
2048 mapcount = atomic_read(&page->_mapcount) + 1;
2049 if (total_mapcount)
2050 *total_mapcount = mapcount;
2051 return mapcount;
2052 }
2053
2054 page = compound_head(page);
2055
2056 _total_mapcount = ret = 0;
2057 for (i = 0; i < HPAGE_PMD_NR; i++) {
2058 mapcount = atomic_read(&page[i]._mapcount) + 1;
2059 ret = max(ret, mapcount);
2060 _total_mapcount += mapcount;
2061 }
2062 if (PageDoubleMap(page)) {
2063 ret -= 1;
2064 _total_mapcount -= HPAGE_PMD_NR;
2065 }
2066 mapcount = compound_mapcount(page);
2067 ret += mapcount;
2068 _total_mapcount += mapcount;
2069 if (total_mapcount)
2070 *total_mapcount = _total_mapcount;
2071 return ret;
2072}
2073
2074/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002075 * This function splits huge page into normal pages. @page can point to any
2076 * subpage of huge page to split. Split doesn't change the position of @page.
2077 *
2078 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2079 * The huge page must be locked.
2080 *
2081 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2082 *
2083 * Both head page and tail pages will inherit mapping, flags, and so on from
2084 * the hugepage.
2085 *
2086 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2087 * they are not mapped.
2088 *
2089 * Returns 0 if the hugepage is split successfully.
2090 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2091 * us.
2092 */
2093int split_huge_page_to_list(struct page *page, struct list_head *list)
2094{
2095 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002096 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002097 struct anon_vma *anon_vma = NULL;
2098 struct address_space *mapping = NULL;
2099 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002100 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002101 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002102
2103 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002104 VM_BUG_ON_PAGE(!PageLocked(page), page);
2105 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2106 VM_BUG_ON_PAGE(!PageCompound(page), page);
2107
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002108 if (PageAnon(head)) {
2109 /*
2110 * The caller does not necessarily hold an mmap_sem that would
2111 * prevent the anon_vma disappearing so we first we take a
2112 * reference to it and then lock the anon_vma for write. This
2113 * is similar to page_lock_anon_vma_read except the write lock
2114 * is taken to serialise against parallel split or collapse
2115 * operations.
2116 */
2117 anon_vma = page_get_anon_vma(head);
2118 if (!anon_vma) {
2119 ret = -EBUSY;
2120 goto out;
2121 }
2122 extra_pins = 0;
2123 mapping = NULL;
2124 anon_vma_lock_write(anon_vma);
2125 } else {
2126 mapping = head->mapping;
2127
2128 /* Truncated ? */
2129 if (!mapping) {
2130 ret = -EBUSY;
2131 goto out;
2132 }
2133
2134 /* Addidional pins from radix tree */
2135 extra_pins = HPAGE_PMD_NR;
2136 anon_vma = NULL;
2137 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002138 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002139
2140 /*
2141 * Racy check if we can split the page, before freeze_page() will
2142 * split PMDs
2143 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002144 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002145 ret = -EBUSY;
2146 goto out_unlock;
2147 }
2148
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002149 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002150 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002151 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2152
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002153 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2154 if (mlocked)
2155 lru_add_drain();
2156
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002157 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002158 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002159
2160 if (mapping) {
2161 void **pslot;
2162
2163 spin_lock(&mapping->tree_lock);
2164 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2165 page_index(head));
2166 /*
2167 * Check if the head page is present in radix tree.
2168 * We assume all tail are present too, if head is there.
2169 */
2170 if (radix_tree_deref_slot_protected(pslot,
2171 &mapping->tree_lock) != head)
2172 goto fail;
2173 }
2174
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002175 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002176 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002177 count = page_count(head);
2178 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002179 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002180 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002181 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002182 list_del(page_deferred_list(head));
2183 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002184 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002185 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002186 spin_unlock(&pgdata->split_queue_lock);
2187 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002188 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002189 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002190 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2191 pr_alert("total_mapcount: %u, page_count(): %u\n",
2192 mapcount, count);
2193 if (PageTail(page))
2194 dump_page(head, NULL);
2195 dump_page(page, "total_mapcount(head) > 0");
2196 BUG();
2197 }
2198 spin_unlock(&pgdata->split_queue_lock);
2199fail: if (mapping)
2200 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002201 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002202 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002203 ret = -EBUSY;
2204 }
2205
2206out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002207 if (anon_vma) {
2208 anon_vma_unlock_write(anon_vma);
2209 put_anon_vma(anon_vma);
2210 }
2211 if (mapping)
2212 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002213out:
2214 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2215 return ret;
2216}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002217
2218void free_transhuge_page(struct page *page)
2219{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002220 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002221 unsigned long flags;
2222
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002223 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002224 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002225 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002226 list_del(page_deferred_list(page));
2227 }
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 free_compound_page(page);
2230}
2231
2232void deferred_split_huge_page(struct page *page)
2233{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002234 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002235 unsigned long flags;
2236
2237 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2238
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002239 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002240 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002241 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002242 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2243 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002244 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002245 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002246}
2247
2248static unsigned long deferred_split_count(struct shrinker *shrink,
2249 struct shrink_control *sc)
2250{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002251 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002252 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002253}
2254
2255static unsigned long deferred_split_scan(struct shrinker *shrink,
2256 struct shrink_control *sc)
2257{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002258 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002259 unsigned long flags;
2260 LIST_HEAD(list), *pos, *next;
2261 struct page *page;
2262 int split = 0;
2263
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002264 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002265 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002266 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002267 page = list_entry((void *)pos, struct page, mapping);
2268 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002269 if (get_page_unless_zero(page)) {
2270 list_move(page_deferred_list(page), &list);
2271 } else {
2272 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002273 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002274 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002275 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002276 if (!--sc->nr_to_scan)
2277 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002278 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002279 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002280
2281 list_for_each_safe(pos, next, &list) {
2282 page = list_entry((void *)pos, struct page, mapping);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002283 if (!trylock_page(page))
2284 goto next;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002285 /* split_huge_page() removes page from list on success */
2286 if (!split_huge_page(page))
2287 split++;
2288 unlock_page(page);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002289next:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002290 put_page(page);
2291 }
2292
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002293 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2294 list_splice_tail(&list, &pgdata->split_queue);
2295 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002296
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002297 /*
2298 * Stop shrinker if we didn't split any page, but the queue is empty.
2299 * This can happen if pages were freed under us.
2300 */
2301 if (!split && list_empty(&pgdata->split_queue))
2302 return SHRINK_STOP;
2303 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002304}
2305
2306static struct shrinker deferred_split_shrinker = {
2307 .count_objects = deferred_split_count,
2308 .scan_objects = deferred_split_scan,
2309 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002310 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002311};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002312
2313#ifdef CONFIG_DEBUG_FS
2314static int split_huge_pages_set(void *data, u64 val)
2315{
2316 struct zone *zone;
2317 struct page *page;
2318 unsigned long pfn, max_zone_pfn;
2319 unsigned long total = 0, split = 0;
2320
2321 if (val != 1)
2322 return -EINVAL;
2323
2324 for_each_populated_zone(zone) {
2325 max_zone_pfn = zone_end_pfn(zone);
2326 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2327 if (!pfn_valid(pfn))
2328 continue;
2329
2330 page = pfn_to_page(pfn);
2331 if (!get_page_unless_zero(page))
2332 continue;
2333
2334 if (zone != page_zone(page))
2335 goto next;
2336
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002337 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002338 goto next;
2339
2340 total++;
2341 lock_page(page);
2342 if (!split_huge_page(page))
2343 split++;
2344 unlock_page(page);
2345next:
2346 put_page(page);
2347 }
2348 }
2349
Yang Shi145bdaa2016-05-05 16:22:00 -07002350 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002351
2352 return 0;
2353}
2354DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2355 "%llu\n");
2356
2357static int __init split_huge_pages_debugfs(void)
2358{
2359 void *ret;
2360
Yang Shi145bdaa2016-05-05 16:22:00 -07002361 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002362 &split_huge_pages_fops);
2363 if (!ret)
2364 pr_warn("Failed to create split_huge_pages in debugfs");
2365 return 0;
2366}
2367late_initcall(split_huge_pages_debugfs);
2368#endif