blob: 3f3a86cc62b68339a322973dc67b9dbc1085c501 [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>
Vlastimil Babkaba37a942019-08-24 17:54:59 -070033#include <linux/page_owner.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080034
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080035#include <asm/tlb.h>
36#include <asm/pgalloc.h>
37#include "internal.h"
38
Andrea Arcangeliba761492011-01-13 15:46:58 -080039/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080040 * By default transparent hugepage support is disabled in order that avoid
41 * to risk increase the memory footprint of applications without a guaranteed
42 * benefit. When transparent hugepage support is enabled, is for all mappings,
43 * and khugepaged scans all mappings.
44 * Defrag is invoked by khugepaged hugepage allocations and by page faults
45 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080046 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080047unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080048#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080049 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080050#endif
51#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
52 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
53#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070054 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080055 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
56 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080057
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080058static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080059
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080060static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080061struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080062
Aaron Lu6fcb52a2016-10-07 17:00:08 -070063static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080064{
65 struct page *zero_page;
66retry:
67 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070068 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080069
70 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
71 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080072 if (!zero_page) {
73 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070074 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080075 }
76 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080077 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070078 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080079 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070080 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080081 goto retry;
82 }
83
84 /* We take additional reference here. It will be put back by shrinker */
85 atomic_set(&huge_zero_refcount, 2);
86 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070087 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080088}
89
Aaron Lu6fcb52a2016-10-07 17:00:08 -070090static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080091{
92 /*
93 * Counter should never go to zero here. Only shrinker can put
94 * last reference.
95 */
96 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
97}
98
Aaron Lu6fcb52a2016-10-07 17:00:08 -070099struct page *mm_get_huge_zero_page(struct mm_struct *mm)
100{
101 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
102 return READ_ONCE(huge_zero_page);
103
104 if (!get_huge_zero_page())
105 return NULL;
106
107 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
108 put_huge_zero_page();
109
110 return READ_ONCE(huge_zero_page);
111}
112
113void mm_put_huge_zero_page(struct mm_struct *mm)
114{
115 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
116 put_huge_zero_page();
117}
118
Glauber Costa48896462013-08-28 10:18:15 +1000119static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
120 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800121{
Glauber Costa48896462013-08-28 10:18:15 +1000122 /* we can free zero page only if last reference remains */
123 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
124}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800125
Glauber Costa48896462013-08-28 10:18:15 +1000126static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
127 struct shrink_control *sc)
128{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800129 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700130 struct page *zero_page = xchg(&huge_zero_page, NULL);
131 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700132 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000133 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800134 }
135
136 return 0;
137}
138
139static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000140 .count_objects = shrink_huge_zero_page_count,
141 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800142 .seeks = DEFAULT_SEEKS,
143};
144
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800145#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800146
Mel Gorman444eb2a42016-03-17 14:19:23 -0700147static ssize_t triple_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800148 struct kobj_attribute *attr,
149 const char *buf, size_t count,
150 enum transparent_hugepage_flag enabled,
Mel Gorman444eb2a42016-03-17 14:19:23 -0700151 enum transparent_hugepage_flag deferred,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800152 enum transparent_hugepage_flag req_madv)
153{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700154 if (!memcmp("defer", buf,
155 min(sizeof("defer")-1, count))) {
156 if (enabled == deferred)
157 return -EINVAL;
158 clear_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800159 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700160 set_bit(deferred, &transparent_hugepage_flags);
161 } else if (!memcmp("always", buf,
162 min(sizeof("always")-1, count))) {
163 clear_bit(deferred, &transparent_hugepage_flags);
164 clear_bit(req_madv, &transparent_hugepage_flags);
165 set_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800166 } else if (!memcmp("madvise", buf,
167 min(sizeof("madvise")-1, count))) {
168 clear_bit(enabled, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700169 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800170 set_bit(req_madv, &transparent_hugepage_flags);
171 } else if (!memcmp("never", buf,
172 min(sizeof("never")-1, count))) {
173 clear_bit(enabled, &transparent_hugepage_flags);
174 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700175 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800176 } else
177 return -EINVAL;
178
179 return count;
180}
181
182static ssize_t enabled_show(struct kobject *kobj,
183 struct kobj_attribute *attr, char *buf)
184{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700185 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
186 return sprintf(buf, "[always] madvise never\n");
187 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
188 return sprintf(buf, "always [madvise] never\n");
189 else
190 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800191}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700192
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800193static ssize_t enabled_store(struct kobject *kobj,
194 struct kobj_attribute *attr,
195 const char *buf, size_t count)
196{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800197 ssize_t ret;
198
Mel Gorman444eb2a42016-03-17 14:19:23 -0700199 ret = triple_flag_store(kobj, attr, buf, count,
200 TRANSPARENT_HUGEPAGE_FLAG,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800201 TRANSPARENT_HUGEPAGE_FLAG,
202 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
203
204 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700205 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800206 if (err)
207 ret = err;
208 }
209
210 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800211}
212static struct kobj_attribute enabled_attr =
213 __ATTR(enabled, 0644, enabled_show, enabled_store);
214
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700215ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800216 struct kobj_attribute *attr, char *buf,
217 enum transparent_hugepage_flag flag)
218{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700219 return sprintf(buf, "%d\n",
220 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800221}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700222
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700223ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800224 struct kobj_attribute *attr,
225 const char *buf, size_t count,
226 enum transparent_hugepage_flag flag)
227{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700228 unsigned long value;
229 int ret;
230
231 ret = kstrtoul(buf, 10, &value);
232 if (ret < 0)
233 return ret;
234 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800235 return -EINVAL;
236
Ben Hutchingse27e6152011-04-14 15:22:21 -0700237 if (value)
238 set_bit(flag, &transparent_hugepage_flags);
239 else
240 clear_bit(flag, &transparent_hugepage_flags);
241
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800242 return count;
243}
244
245/*
246 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
247 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
248 * memory just to allocate one more hugepage.
249 */
250static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
252{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700253 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
254 return sprintf(buf, "[always] defer madvise never\n");
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
256 return sprintf(buf, "always [defer] madvise never\n");
257 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
258 return sprintf(buf, "always defer [madvise] never\n");
259 else
260 return sprintf(buf, "always defer madvise [never]\n");
261
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800262}
263static ssize_t defrag_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
266{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700267 return triple_flag_store(kobj, attr, buf, count,
268 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
269 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800270 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
271}
272static struct kobj_attribute defrag_attr =
273 __ATTR(defrag, 0644, defrag_show, defrag_store);
274
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800275static ssize_t use_zero_page_show(struct kobject *kobj,
276 struct kobj_attribute *attr, char *buf)
277{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700278 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800279 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
280}
281static ssize_t use_zero_page_store(struct kobject *kobj,
282 struct kobj_attribute *attr, const char *buf, size_t count)
283{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700284 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800285 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
286}
287static struct kobj_attribute use_zero_page_attr =
288 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800289#ifdef CONFIG_DEBUG_VM
290static ssize_t debug_cow_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
292{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700293 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800294 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
295}
296static ssize_t debug_cow_store(struct kobject *kobj,
297 struct kobj_attribute *attr,
298 const char *buf, size_t count)
299{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700300 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
302}
303static struct kobj_attribute debug_cow_attr =
304 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
305#endif /* CONFIG_DEBUG_VM */
306
307static struct attribute *hugepage_attr[] = {
308 &enabled_attr.attr,
309 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800310 &use_zero_page_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700311#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700312 &shmem_enabled_attr.attr,
313#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800314#ifdef CONFIG_DEBUG_VM
315 &debug_cow_attr.attr,
316#endif
317 NULL,
318};
319
320static struct attribute_group hugepage_attr_group = {
321 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800322};
323
Shaohua Li569e5592012-01-12 17:19:11 -0800324static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
325{
326 int err;
327
328 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
329 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700330 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800331 return -ENOMEM;
332 }
333
334 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
335 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700336 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800337 goto delete_obj;
338 }
339
340 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
341 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700342 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800343 goto remove_hp_group;
344 }
345
346 return 0;
347
348remove_hp_group:
349 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
350delete_obj:
351 kobject_put(*hugepage_kobj);
352 return err;
353}
354
355static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
356{
357 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
358 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
359 kobject_put(hugepage_kobj);
360}
361#else
362static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
363{
364 return 0;
365}
366
367static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
368{
369}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800370#endif /* CONFIG_SYSFS */
371
372static int __init hugepage_init(void)
373{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800374 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800375 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800376
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800377 if (!has_transparent_hugepage()) {
378 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800379 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800380 }
381
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530382 /*
383 * hugepages can't be allocated by the buddy allocator
384 */
385 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
386 /*
387 * we use page->mapping and page->index in second tail page
388 * as list_head: assuming THP order >= 2
389 */
390 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
391
Shaohua Li569e5592012-01-12 17:19:11 -0800392 err = hugepage_init_sysfs(&hugepage_kobj);
393 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700394 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800395
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700396 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800397 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700398 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800399
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700400 err = register_shrinker(&huge_zero_page_shrinker);
401 if (err)
402 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800403 err = register_shrinker(&deferred_split_shrinker);
404 if (err)
405 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800406
Rik van Riel97562cd2011-01-13 15:47:12 -0800407 /*
408 * By default disable transparent hugepages on smaller systems,
409 * where the extra memory used could hurt more than TLB overhead
410 * is likely to save. The admin can still enable it through /sys.
411 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700412 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800413 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700414 return 0;
415 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800416
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700417 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700418 if (err)
419 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800420
Shaohua Li569e5592012-01-12 17:19:11 -0800421 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700422err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800423 unregister_shrinker(&deferred_split_shrinker);
424err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700425 unregister_shrinker(&huge_zero_page_shrinker);
426err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700427 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700428err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800429 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700430err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800431 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800432}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800433subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800434
435static int __init setup_transparent_hugepage(char *str)
436{
437 int ret = 0;
438 if (!str)
439 goto out;
440 if (!strcmp(str, "always")) {
441 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
442 &transparent_hugepage_flags);
443 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
444 &transparent_hugepage_flags);
445 ret = 1;
446 } else if (!strcmp(str, "madvise")) {
447 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
448 &transparent_hugepage_flags);
449 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
450 &transparent_hugepage_flags);
451 ret = 1;
452 } else if (!strcmp(str, "never")) {
453 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
454 &transparent_hugepage_flags);
455 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
456 &transparent_hugepage_flags);
457 ret = 1;
458 }
459out:
460 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700461 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800462 return ret;
463}
464__setup("transparent_hugepage=", setup_transparent_hugepage);
465
Mel Gormanb32967f2012-11-19 12:35:47 +0000466pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800467{
468 if (likely(vma->vm_flags & VM_WRITE))
469 pmd = pmd_mkwrite(pmd);
470 return pmd;
471}
472
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800473static inline struct list_head *page_deferred_list(struct page *page)
474{
475 /*
476 * ->lru in the tail pages is occupied by compound_head.
477 * Let's use ->mapping + ->index in the second tail page as list_head.
478 */
479 return (struct list_head *)&page[2].mapping;
480}
481
482void prep_transhuge_page(struct page *page)
483{
484 /*
485 * we use page->mapping and page->indexlru in second tail page
486 * as list_head: assuming THP order >= 2
487 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800488
489 INIT_LIST_HEAD(page_deferred_list(page));
490 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
491}
492
Toshi Kani74d2fad2016-10-07 16:59:56 -0700493unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
494 loff_t off, unsigned long flags, unsigned long size)
495{
496 unsigned long addr;
497 loff_t off_end = off + len;
498 loff_t off_align = round_up(off, size);
499 unsigned long len_pad;
500
501 if (off_end <= off_align || (off_end - off_align) < size)
502 return 0;
503
504 len_pad = len + size;
505 if (len_pad < len || (off + len_pad) < off)
506 return 0;
507
508 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
509 off >> PAGE_SHIFT, flags);
510 if (IS_ERR_VALUE(addr))
511 return 0;
512
513 addr += (off - addr) & (size - 1);
514 return addr;
515}
516
517unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
518 unsigned long len, unsigned long pgoff, unsigned long flags)
519{
520 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
521
522 if (addr)
523 goto out;
524 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
525 goto out;
526
527 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
528 if (addr)
529 return addr;
530
531 out:
532 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
533}
534EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
535
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700536static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
537 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800538{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700539 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700540 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800541 pgtable_t pgtable;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700542 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800543
Sasha Levin309381fea2014-01-23 15:52:54 -0800544 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700545
David Rientjes6b7ff8e2018-03-22 16:17:45 -0700546 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
547 true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700548 put_page(page);
549 count_vm_event(THP_FAULT_FALLBACK);
550 return VM_FAULT_FALLBACK;
551 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800552
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700553 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700554 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800555 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700556 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700557 return VM_FAULT_OOM;
558 }
559
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800560 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700561 /*
562 * The memory barrier inside __SetPageUptodate makes sure that
563 * clear_huge_page writes become visible before the set_pmd_at()
564 * write.
565 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800566 __SetPageUptodate(page);
567
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700568 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
569 if (unlikely(!pmd_none(*fe->pmd))) {
570 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800571 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800572 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700573 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800574 } else {
575 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700576
577 /* Deliver the page fault to userland */
578 if (userfaultfd_missing(vma)) {
579 int ret;
580
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700581 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800582 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700583 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700584 pte_free(vma->vm_mm, pgtable);
585 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700586 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
587 return ret;
588 }
589
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700590 entry = mk_huge_pmd(page, vma->vm_page_prot);
591 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800592 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800593 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700594 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700595 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
596 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
597 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
598 atomic_long_inc(&vma->vm_mm->nr_ptes);
599 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700600 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800601 }
602
David Rientjesaa2e8782012-05-29 15:06:17 -0700603 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800604}
605
Mel Gorman444eb2a42016-03-17 14:19:23 -0700606/*
Vlastimil Babka25160352016-07-28 15:49:25 -0700607 * If THP defrag is set to always then directly reclaim/compact as necessary
608 * If set to defer then do only background reclaim/compact and defer to khugepaged
Mel Gorman444eb2a42016-03-17 14:19:23 -0700609 * If set to madvise and the VMA is flagged then directly reclaim/compact
Vlastimil Babka25160352016-07-28 15:49:25 -0700610 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
Mel Gorman444eb2a42016-03-17 14:19:23 -0700611 */
612static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800613{
Vlastimil Babka25160352016-07-28 15:49:25 -0700614 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700615
Vlastimil Babka25160352016-07-28 15:49:25 -0700616 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
617 &transparent_hugepage_flags) && vma_madvised)
618 return GFP_TRANSHUGE;
619 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
620 &transparent_hugepage_flags))
621 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
622 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
623 &transparent_hugepage_flags))
624 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700625
Vlastimil Babka25160352016-07-28 15:49:25 -0700626 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700627}
628
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800629/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700630static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800631 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700632 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800633{
634 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700635 if (!pmd_none(*pmd))
636 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700637 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800638 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800639 if (pgtable)
640 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800641 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800642 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700643 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800644}
645
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700646int do_huge_pmd_anonymous_page(struct fault_env *fe)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800647{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700648 struct vm_area_struct *vma = fe->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800649 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800650 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700651 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800652
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700653 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700654 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700655 if (unlikely(anon_vma_prepare(vma)))
656 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700657 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700658 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700659 if (!(fe->flags & FAULT_FLAG_WRITE) &&
660 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700661 transparent_hugepage_use_zero_page()) {
662 pgtable_t pgtable;
663 struct page *zero_page;
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;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700676 if (pmd_none(*fe->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700677 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700678 spin_unlock(fe->ptl);
Gerald Schaefer003be902020-11-23 18:05:38 +0100679 pte_free(vma->vm_mm, pgtable);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700680 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 }
Gerald Schaefer003be902020-11-23 18:05:38 +0100687 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700688 spin_unlock(fe->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700689 pte_free(vma->vm_mm, pgtable);
Gerald Schaefer003be902020-11-23 18:05:38 +0100690 }
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700691 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800692 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700693 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800694 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700695 if (unlikely(!page)) {
696 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700697 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700698 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800699 prep_transhuge_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700700 return __do_huge_pmd_anonymous_page(fe, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800701}
702
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700703static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800704 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700705{
706 struct mm_struct *mm = vma->vm_mm;
707 pmd_t entry;
708 spinlock_t *ptl;
709
710 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800711 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
712 if (pfn_t_devmap(pfn))
713 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800714 if (write) {
715 entry = pmd_mkyoung(pmd_mkdirty(entry));
716 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700717 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800718 set_pmd_at(mm, addr, pmd, entry);
719 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700720 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700721}
722
723int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800724 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700725{
726 pgprot_t pgprot = vma->vm_page_prot;
727 /*
728 * If we had pmd_special, we could avoid all these restrictions,
729 * but we need to be consistent with PTEs and architectures that
730 * can't support a 'special' bit.
731 */
732 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
733 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
734 (VM_PFNMAP|VM_MIXEDMAP));
735 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800736 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700737
738 if (addr < vma->vm_start || addr >= vma->vm_end)
739 return VM_FAULT_SIGBUS;
740 if (track_pfn_insert(vma, &pgprot, pfn))
741 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700742 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
743 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700744}
Dan Williamsdee41072016-05-14 12:20:44 -0700745EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700746
Dan Williams3565fce2016-01-15 16:56:55 -0800747static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300748 pmd_t *pmd, int flags)
Dan Williams3565fce2016-01-15 16:56:55 -0800749{
750 pmd_t _pmd;
751
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300752 _pmd = pmd_mkyoung(*pmd);
753 if (flags & FOLL_WRITE)
754 _pmd = pmd_mkdirty(_pmd);
Dan Williams3565fce2016-01-15 16:56:55 -0800755 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300756 pmd, _pmd, flags & FOLL_WRITE))
Dan Williams3565fce2016-01-15 16:56:55 -0800757 update_mmu_cache_pmd(vma, addr, pmd);
758}
759
760struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
761 pmd_t *pmd, int flags)
762{
763 unsigned long pfn = pmd_pfn(*pmd);
764 struct mm_struct *mm = vma->vm_mm;
765 struct dev_pagemap *pgmap;
766 struct page *page;
767
768 assert_spin_locked(pmd_lockptr(mm, pmd));
769
Keno Fischer6676aa62017-01-24 15:17:48 -0800770 /*
771 * When we COW a devmap PMD entry, we split it into PTEs, so we should
772 * not be in this function with `flags & FOLL_COW` set.
773 */
774 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
775
Dan Williams3565fce2016-01-15 16:56:55 -0800776 if (flags & FOLL_WRITE && !pmd_write(*pmd))
777 return NULL;
778
779 if (pmd_present(*pmd) && pmd_devmap(*pmd))
780 /* pass */;
781 else
782 return NULL;
783
784 if (flags & FOLL_TOUCH)
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +0300785 touch_pmd(vma, addr, pmd, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800786
787 /*
788 * device mapped pages can only be returned if the
789 * caller will manage the page reference count.
790 */
791 if (!(flags & FOLL_GET))
792 return ERR_PTR(-EEXIST);
793
794 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
795 pgmap = get_dev_pagemap(pfn, NULL);
796 if (!pgmap)
797 return ERR_PTR(-EFAULT);
798 page = pfn_to_page(pfn);
799 get_page(page);
800 put_dev_pagemap(pgmap);
801
802 return page;
803}
804
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800805int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
806 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
807 struct vm_area_struct *vma)
808{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800809 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800810 struct page *src_page;
811 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800812 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700813 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800814
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700815 /* Skip if can be re-fill on fault */
816 if (!vma_is_anonymous(vma))
817 return 0;
818
819 pgtable = pte_alloc_one(dst_mm, addr);
820 if (unlikely(!pgtable))
821 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800822
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800823 dst_ptl = pmd_lock(dst_mm, dst_pmd);
824 src_ptl = pmd_lockptr(src_mm, src_pmd);
825 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800826
827 ret = -EAGAIN;
828 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700829 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800830 pte_free(dst_mm, pgtable);
831 goto out_unlock;
832 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800833 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800834 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800835 * under splitting since we don't split the page itself, only pmd to
836 * a page table.
837 */
838 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700839 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800840 /*
841 * get_huge_zero_page() will never allocate a new page here,
842 * since we already have a zero page to copy. It just takes a
843 * reference.
844 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700845 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700846 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700847 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800848 ret = 0;
849 goto out_unlock;
850 }
Mel Gormande466bd2013-12-18 17:08:42 -0800851
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700852 src_page = pmd_page(pmd);
853 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
854 get_page(src_page);
855 page_dup_rmap(src_page, true);
856 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
857 atomic_long_inc(&dst_mm->nr_ptes);
858 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800859
860 pmdp_set_wrprotect(src_mm, addr, src_pmd);
861 pmd = pmd_mkold(pmd_wrprotect(pmd));
862 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800863
864 ret = 0;
865out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800866 spin_unlock(src_ptl);
867 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800868out:
869 return ret;
870}
871
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700872void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -0800873{
874 pmd_t entry;
875 unsigned long haddr;
Minchan Kim8edd3652017-01-10 16:57:51 -0800876 bool write = fe->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -0800877
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700878 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
879 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -0800880 goto unlock;
881
882 entry = pmd_mkyoung(orig_pmd);
Minchan Kim8edd3652017-01-10 16:57:51 -0800883 if (write)
884 entry = pmd_mkdirty(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700885 haddr = fe->address & HPAGE_PMD_MASK;
Minchan Kim8edd3652017-01-10 16:57:51 -0800886 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700887 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800888
889unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700890 spin_unlock(fe->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800891}
892
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700893static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
894 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800895{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700896 struct vm_area_struct *vma = fe->vma;
897 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -0700898 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800899 pgtable_t pgtable;
900 pmd_t _pmd;
901 int ret = 0, i;
902 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700903 unsigned long mmun_start; /* For mmu_notifiers */
904 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800905
906 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
907 GFP_KERNEL);
908 if (unlikely(!pages)) {
909 ret |= VM_FAULT_OOM;
910 goto out;
911 }
912
913 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700914 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700915 __GFP_OTHER_NODE, vma,
916 fe->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800917 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700918 mem_cgroup_try_charge(pages[i], vma->vm_mm,
919 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800920 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800921 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800922 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700923 memcg = (void *)page_private(pages[i]);
924 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800925 mem_cgroup_cancel_charge(pages[i], memcg,
926 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800927 put_page(pages[i]);
928 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800929 kfree(pages);
930 ret |= VM_FAULT_OOM;
931 goto out;
932 }
Johannes Weiner00501b52014-08-08 14:19:20 -0700933 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800934 }
935
936 for (i = 0; i < HPAGE_PMD_NR; i++) {
937 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700938 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800939 __SetPageUptodate(pages[i]);
940 cond_resched();
941 }
942
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700943 mmun_start = haddr;
944 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700945 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700946
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700947 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
948 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800949 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -0800950 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800951
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700952 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800953 /* leave pmd empty until pte is filled */
954
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700955 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
956 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800957
958 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700959 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800960 entry = mk_pte(pages[i], vma->vm_page_prot);
961 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -0700962 memcg = (void *)page_private(pages[i]);
963 set_page_private(pages[i], 0);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700964 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800965 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -0700966 lru_cache_add_active_or_unevictable(pages[i], vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700967 fe->pte = pte_offset_map(&_pmd, haddr);
968 VM_BUG_ON(!pte_none(*fe->pte));
969 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
970 pte_unmap(fe->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800971 }
972 kfree(pages);
973
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800974 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700975 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800976 page_remove_rmap(page, true);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700977 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800978
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700979 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700980
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800981 ret |= VM_FAULT_WRITE;
982 put_page(page);
983
984out:
985 return ret;
986
987out_free_pages:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700988 spin_unlock(fe->ptl);
989 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800990 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700991 memcg = (void *)page_private(pages[i]);
992 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800993 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800994 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800995 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800996 kfree(pages);
997 goto out;
998}
999
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001000int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001001{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001002 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001003 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001004 struct mem_cgroup *memcg;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001005 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001006 unsigned long mmun_start; /* For mmu_notifiers */
1007 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001008 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001009 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001010
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001011 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001012 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001013 if (is_huge_zero_pmd(orig_pmd))
1014 goto alloc;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001015 spin_lock(fe->ptl);
1016 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001017 goto out_unlock;
1018
1019 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001020 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001021 /*
1022 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001023 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001024 */
Vlastimil Babkacc916622021-02-26 17:22:00 +01001025 if (!trylock_page(page)) {
1026 get_page(page);
1027 spin_unlock(fe->ptl);
1028 lock_page(page);
1029 spin_lock(fe->ptl);
1030 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1031 unlock_page(page);
1032 put_page(page);
1033 goto out_unlock;
1034 }
1035 put_page(page);
1036 }
1037
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001038 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001039 pmd_t entry;
1040 entry = pmd_mkyoung(orig_pmd);
1041 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001042 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1043 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001044 ret |= VM_FAULT_WRITE;
Vlastimil Babkacc916622021-02-26 17:22:00 +01001045 unlock_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001046 goto out_unlock;
1047 }
Vlastimil Babkacc916622021-02-26 17:22:00 +01001048 unlock_page(page);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001049 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001050 spin_unlock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001051alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001052 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001053 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001054 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001055 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001056 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001057 new_page = NULL;
1058
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001059 if (likely(new_page)) {
1060 prep_transhuge_page(new_page);
1061 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001062 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001063 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001064 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001065 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001066 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001067 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001068 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001069 ret |= VM_FAULT_FALLBACK;
1070 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001071 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001072 }
David Rientjes17766dd2013-09-12 15:14:06 -07001073 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001074 goto out;
1075 }
1076
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001077 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
David Rientjes6b7ff8e2018-03-22 16:17:45 -07001078 huge_gfp | __GFP_NORETRY, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001079 put_page(new_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001080 split_huge_pmd(vma, fe->pmd, fe->address);
1081 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001082 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001083 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001084 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001085 goto out;
1086 }
1087
David Rientjes17766dd2013-09-12 15:14:06 -07001088 count_vm_event(THP_FAULT_ALLOC);
1089
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001090 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001091 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1092 else
1093 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001094 __SetPageUptodate(new_page);
1095
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001096 mmun_start = haddr;
1097 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001098 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001099
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001100 spin_lock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001101 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001102 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001103 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1104 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001105 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001106 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001107 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001108 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001109 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001110 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1111 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001112 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001113 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001114 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001115 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001116 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1117 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001118 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001119 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001120 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001121 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001122 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001123 put_page(page);
1124 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 ret |= VM_FAULT_WRITE;
1126 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001127 spin_unlock(fe->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001128out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001129 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001130out:
1131 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001132out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001133 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134 return ret;
1135}
1136
Keno Fischer6676aa62017-01-24 15:17:48 -08001137/*
Linus Torvalds0c296402022-01-24 14:41:50 +01001138 * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1139 * but only after we've gone through a COW cycle and they are dirty.
Keno Fischer6676aa62017-01-24 15:17:48 -08001140 */
1141static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1142{
Linus Torvalds0c296402022-01-24 14:41:50 +01001143 return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
Keno Fischer6676aa62017-01-24 15:17:48 -08001144}
1145
David Rientjesb676b292012-10-08 16:34:03 -07001146struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001147 unsigned long addr,
1148 pmd_t *pmd,
1149 unsigned int flags)
1150{
David Rientjesb676b292012-10-08 16:34:03 -07001151 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001152 struct page *page = NULL;
1153
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001154 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001155
Keno Fischer6676aa62017-01-24 15:17:48 -08001156 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001157 goto out;
1158
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001159 /* Avoid dumping huge zero page */
1160 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1161 return ERR_PTR(-EFAULT);
1162
Mel Gorman2b4847e2013-12-18 17:08:32 -08001163 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001164 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001165 goto out;
1166
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001167 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001168 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001169 if (flags & FOLL_TOUCH)
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +03001170 touch_pmd(vma, addr, pmd, flags);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001171 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001172 /*
1173 * We don't mlock() pte-mapped THPs. This way we can avoid
1174 * leaking mlocked pages into non-VM_LOCKED VMAs.
1175 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001176 * For anon THP:
1177 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001178 * In most cases the pmd is the only mapping of the page as we
1179 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1180 * writable private mappings in populate_vma_page_range().
1181 *
1182 * The only scenario when we have the page shared here is if we
1183 * mlocking read-only mapping shared over fork(). We skip
1184 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001185 *
1186 * For file THP:
1187 *
1188 * We can expect PageDoubleMap() to be stable under page lock:
1189 * for file pages we set it in page_add_file_rmap(), which
1190 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001191 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001192
1193 if (PageAnon(page) && compound_mapcount(page) != 1)
1194 goto skip_mlock;
1195 if (PageDoubleMap(page) || !page->mapping)
1196 goto skip_mlock;
1197 if (!trylock_page(page))
1198 goto skip_mlock;
1199 lru_add_drain();
1200 if (page->mapping && !PageDoubleMap(page))
1201 mlock_vma_page(page);
1202 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001203 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001204skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001206 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001207 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001208 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209
1210out:
1211 return page;
1212}
1213
Mel Gormand10e63f2012-10-25 14:16:31 +02001214/* NUMA hinting page fault entry point for trans huge pmds */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001215int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001216{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001217 struct vm_area_struct *vma = fe->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001218 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001219 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001220 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001221 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001222 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001223 bool page_locked;
1224 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001225 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001226 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001227
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001228 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1229 if (unlikely(!pmd_same(pmd, *fe->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001230 goto out_unlock;
1231
Mel Gormande466bd2013-12-18 17:08:42 -08001232 /*
1233 * If there are potential migrations, wait for completion and retry
1234 * without disrupting NUMA hinting information. Do not relock and
1235 * check_same as the page may no longer be mapped.
1236 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001237 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1238 page = pmd_page(*fe->pmd);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001239 if (!get_page_unless_zero(page))
1240 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001241 spin_unlock(fe->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001242 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001243 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001244 goto out;
1245 }
1246
Mel Gormand10e63f2012-10-25 14:16:31 +02001247 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001248 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001249 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001250 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001251 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001252 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001253 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001254 flags |= TNF_FAULT_LOCAL;
1255 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001256
Mel Gormanbea66fb2015-03-25 15:55:37 -07001257 /* See similar comment in do_numa_page for explanation */
Rik van Rield59dc7b2016-09-08 21:30:53 -04001258 if (!pmd_write(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001259 flags |= TNF_NO_GROUP;
1260
1261 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001262 * Acquire the page lock to serialise THP migrations but avoid dropping
1263 * page_table_lock if at all possible
1264 */
Mel Gormanb8916632013-10-07 11:28:44 +01001265 page_locked = trylock_page(page);
1266 target_nid = mpol_misplaced(page, vma, haddr);
1267 if (target_nid == -1) {
1268 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001269 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001270 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001271 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001272
Mel Gormande466bd2013-12-18 17:08:42 -08001273 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001274 if (!page_locked) {
Chas Williams77d1a432018-09-06 11:09:10 -04001275 page_nid = -1;
Mark Rutland2aa6d032017-06-16 14:02:34 -07001276 if (!get_page_unless_zero(page))
1277 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001278 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001279 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001280 put_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001281 goto out;
1282 }
1283
Mel Gorman2b4847e2013-12-18 17:08:32 -08001284 /*
1285 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1286 * to serialises splits
1287 */
Mel Gormanb8916632013-10-07 11:28:44 +01001288 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001289 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001290 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001291
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001292 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001293 spin_lock(fe->ptl);
1294 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001295 unlock_page(page);
1296 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001297 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001298 goto out_unlock;
1299 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001300
Mel Gormanc3a489c2013-12-18 17:08:38 -08001301 /* Bail if we fail to protect against THP splits for any reason */
1302 if (unlikely(!anon_vma)) {
1303 put_page(page);
1304 page_nid = -1;
1305 goto clear_pmdnuma;
1306 }
1307
Mel Gormana54a4072013-10-07 11:28:46 +01001308 /*
1309 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001310 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001311 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001312 spin_unlock(fe->ptl);
1313 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1314 fe->pmd, pmd, fe->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001315 if (migrated) {
1316 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001317 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001318 } else
1319 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001320
Mel Gorman8191acb2013-10-07 11:28:45 +01001321 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001322clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001323 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001324 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001325 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001326 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001327 if (was_writable)
1328 pmd = pmd_mkwrite(pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001329 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1330 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001331 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001332out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001333 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001334
1335out:
1336 if (anon_vma)
1337 page_unlock_anon_vma_read(anon_vma);
1338
Mel Gorman8191acb2013-10-07 11:28:45 +01001339 if (page_nid != -1)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001340 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001341
Mel Gormand10e63f2012-10-25 14:16:31 +02001342 return 0;
1343}
1344
Huang Ying319904a2016-07-28 15:48:03 -07001345/*
1346 * Return true if we do MADV_FREE successfully on entire pmd page.
1347 * Otherwise, return false.
1348 */
1349bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001350 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001351{
1352 spinlock_t *ptl;
1353 pmd_t orig_pmd;
1354 struct page *page;
1355 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001356 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001357
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001358 ptl = pmd_trans_huge_lock(pmd, vma);
1359 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001360 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001361
1362 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001363 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001364 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001365
1366 page = pmd_page(orig_pmd);
1367 /*
1368 * If other processes are mapping this page, we couldn't discard
1369 * the page unless they all do MADV_FREE so let's skip the page.
1370 */
Miaohe Lin535f6102021-06-30 18:47:57 -07001371 if (total_mapcount(page) != 1)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001372 goto out;
1373
1374 if (!trylock_page(page))
1375 goto out;
1376
1377 /*
1378 * If user want to discard part-pages of THP, split it so MADV_FREE
1379 * will deactivate only them.
1380 */
1381 if (next - addr != HPAGE_PMD_SIZE) {
1382 get_page(page);
1383 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001384 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001385 unlock_page(page);
Kirill A. Shutemovd2b64682017-07-06 15:35:28 -07001386 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001387 goto out_unlocked;
1388 }
1389
1390 if (PageDirty(page))
1391 ClearPageDirty(page);
1392 unlock_page(page);
1393
1394 if (PageActive(page))
1395 deactivate_page(page);
1396
1397 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemovf5848032017-04-13 14:56:26 -07001398 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001399 orig_pmd = pmd_mkold(orig_pmd);
1400 orig_pmd = pmd_mkclean(orig_pmd);
1401
1402 set_pmd_at(mm, addr, pmd, orig_pmd);
1403 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1404 }
Huang Ying319904a2016-07-28 15:48:03 -07001405 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001406out:
1407 spin_unlock(ptl);
1408out_unlocked:
1409 return ret;
1410}
1411
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001412int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001413 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001414{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001415 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001416 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001417
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001418 ptl = __pmd_trans_huge_lock(pmd, vma);
1419 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001420 return 0;
1421 /*
1422 * For architectures like ppc64 we look at deposited pgtable
1423 * when calling pmdp_huge_get_and_clear. So do the
1424 * pgtable_trans_huge_withdraw after finishing pmdp related
1425 * operations.
1426 */
1427 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1428 tlb->fullmm);
1429 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1430 if (vma_is_dax(vma)) {
1431 spin_unlock(ptl);
1432 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001433 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001434 } else if (is_huge_zero_pmd(orig_pmd)) {
1435 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1436 atomic_long_dec(&tlb->mm->nr_ptes);
1437 spin_unlock(ptl);
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001438 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001439 } else {
1440 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001441 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001442 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001443 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001444 if (PageAnon(page)) {
1445 pgtable_t pgtable;
1446 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1447 pte_free(tlb->mm, pgtable);
1448 atomic_long_dec(&tlb->mm->nr_ptes);
1449 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1450 } else {
1451 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1452 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001453 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001454 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001455 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001456 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001457}
1458
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001459bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001460 unsigned long new_addr, unsigned long old_end,
Linus Torvaldse34bd9a2018-10-12 15:22:59 -07001461 pmd_t *old_pmd, pmd_t *new_pmd)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001462{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001463 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001464 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001465 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001466 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001467
1468 if ((old_addr & ~HPAGE_PMD_MASK) ||
1469 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001470 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001471 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001472
1473 /*
1474 * The destination pmd shouldn't be established, free_pgtables()
1475 * should have release it.
1476 */
1477 if (WARN_ON(!pmd_none(*new_pmd))) {
1478 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001479 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001480 }
1481
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001482 /*
1483 * We don't have to worry about the ordering of src and dst
1484 * ptlocks because exclusive mmap_sem prevents deadlock.
1485 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001486 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1487 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001488 new_ptl = pmd_lockptr(mm, new_pmd);
1489 if (new_ptl != old_ptl)
1490 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001491 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Linus Torvaldse34bd9a2018-10-12 15:22:59 -07001492 if (pmd_present(pmd))
Aaron Lua2ce2662016-11-29 13:27:31 +08001493 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001494 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001495
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001496 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1497 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301498 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001499 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1500 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001501 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301502 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
Aaron Lu5d190422016-11-10 17:16:33 +08001503 if (force_flush)
1504 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
Linus Torvaldse34bd9a2018-10-12 15:22:59 -07001505 if (new_ptl != old_ptl)
1506 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001507 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001508 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001509 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001510 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001511}
1512
Mel Gormanf123d742013-10-07 11:28:49 +01001513/*
1514 * Returns
1515 * - 0 if PMD could not be locked
1516 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1517 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1518 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001519int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001520 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001521{
1522 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001523 spinlock_t *ptl;
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001524 pmd_t entry;
1525 bool preserve_write;
1526 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001527
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001528 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001529 if (!ptl)
1530 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001531
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001532 preserve_write = prot_numa && pmd_write(*pmd);
1533 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001534
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001535 /*
1536 * Avoid trapping faults against the zero page. The read-only
1537 * data is likely to be read-cached on the local CPU and
1538 * local/remote hits to the zero page are not interesting.
1539 */
1540 if (prot_numa && is_huge_zero_pmd(*pmd))
1541 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001542
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001543 if (prot_numa && pmd_protnone(*pmd))
1544 goto unlock;
1545
Kirill A. Shutemovc2edc332017-04-13 14:56:20 -07001546 /*
1547 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1548 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1549 * which is also under down_read(mmap_sem):
1550 *
1551 * CPU0: CPU1:
1552 * change_huge_pmd(prot_numa=1)
1553 * pmdp_huge_get_and_clear_notify()
1554 * madvise_dontneed()
1555 * zap_pmd_range()
1556 * pmd_trans_huge(*pmd) == 0 (without ptl)
1557 * // skip the pmd
1558 * set_pmd_at();
1559 * // pmd is re-established
1560 *
1561 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1562 * which may break userspace.
1563 *
1564 * pmdp_invalidate() is required to make sure we don't miss
1565 * dirty/young flags set by hardware.
1566 */
1567 entry = *pmd;
1568 pmdp_invalidate(vma, addr, pmd);
1569
1570 /*
1571 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1572 * corrupt them.
1573 */
1574 if (pmd_dirty(*pmd))
1575 entry = pmd_mkdirty(entry);
1576 if (pmd_young(*pmd))
1577 entry = pmd_mkyoung(entry);
1578
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001579 entry = pmd_modify(entry, newprot);
1580 if (preserve_write)
1581 entry = pmd_mkwrite(entry);
1582 ret = HPAGE_PMD_NR;
1583 set_pmd_at(mm, addr, pmd, entry);
1584 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1585unlock:
1586 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001587 return ret;
1588}
1589
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001590/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001591 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001592 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001593 * Note that if it returns page table lock pointer, this routine returns without
1594 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001595 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001596spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001597{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001598 spinlock_t *ptl;
1599 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001600 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001601 return ptl;
1602 spin_unlock(ptl);
1603 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001604}
1605
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001606static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1607 unsigned long haddr, pmd_t *pmd)
1608{
1609 struct mm_struct *mm = vma->vm_mm;
1610 pgtable_t pgtable;
1611 pmd_t _pmd;
1612 int i;
1613
1614 /* leave pmd empty until pte is filled */
1615 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1616
1617 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1618 pmd_populate(mm, &_pmd, pgtable);
1619
1620 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1621 pte_t *pte, entry;
1622 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1623 entry = pte_mkspecial(entry);
1624 pte = pte_offset_map(&_pmd, haddr);
1625 VM_BUG_ON(!pte_none(*pte));
1626 set_pte_at(mm, haddr, pte, entry);
1627 pte_unmap(pte);
1628 }
1629 smp_wmb(); /* make pte visible before pmd */
1630 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001631}
1632
1633static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001634 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001635{
1636 struct mm_struct *mm = vma->vm_mm;
1637 struct page *page;
1638 pgtable_t pgtable;
1639 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001640 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001641 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001642 int i;
1643
1644 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1645 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1646 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001647 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001648
1649 count_vm_event(THP_SPLIT_PMD);
1650
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001651 if (!vma_is_anonymous(vma)) {
1652 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001653 if (vma_is_dax(vma))
1654 return;
1655 page = pmd_page(_pmd);
Hugh Dickins3472e372018-07-20 17:53:45 -07001656 if (!PageDirty(page) && pmd_dirty(_pmd))
1657 set_page_dirty(page);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001658 if (!PageReferenced(page) && pmd_young(_pmd))
1659 SetPageReferenced(page);
1660 page_remove_rmap(page, true);
1661 put_page(page);
1662 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001663 return;
1664 } else if (is_huge_zero_pmd(*pmd)) {
1665 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1666 }
1667
1668 page = pmd_page(*pmd);
1669 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001670 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001671 write = pmd_write(*pmd);
1672 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001673 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001674 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001675
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301676 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001677 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678 pmd_populate(mm, &_pmd, pgtable);
1679
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001680 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001681 pte_t entry, *pte;
1682 /*
1683 * Note that NUMA hinting access restrictions are not
1684 * transferred to avoid any possibility of altering
1685 * permissions across VMAs.
1686 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001687 if (freeze) {
1688 swp_entry_t swp_entry;
1689 swp_entry = make_migration_entry(page + i, write);
1690 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001691 if (soft_dirty)
1692 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001693 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001694 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001695 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001696 if (!write)
1697 entry = pte_wrprotect(entry);
1698 if (!young)
1699 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001700 if (soft_dirty)
1701 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001702 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001703 if (dirty)
1704 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001705 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001706 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001707 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001708 atomic_inc(&page[i]._mapcount);
1709 pte_unmap(pte);
1710 }
1711
1712 /*
1713 * Set PG_double_map before dropping compound_mapcount to avoid
1714 * false-negative page_mapped().
1715 */
1716 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1717 for (i = 0; i < HPAGE_PMD_NR; i++)
1718 atomic_inc(&page[i]._mapcount);
1719 }
1720
1721 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1722 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001723 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001724 if (TestClearPageDoubleMap(page)) {
1725 /* No need in mapcount reference anymore */
1726 for (i = 0; i < HPAGE_PMD_NR; i++)
1727 atomic_dec(&page[i]._mapcount);
1728 }
1729 }
1730
1731 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001732 /*
1733 * Up to this point the pmd is present and huge and userland has the
1734 * whole access to the hugepage during the split (which happens in
1735 * place). If we overwrite the pmd with the not-huge version pointing
1736 * to the pte here (which of course we could if all CPUs were bug
1737 * free), userland could trigger a small page size TLB miss on the
1738 * small sized TLB while the hugepage TLB entry is still established in
1739 * the huge TLB. Some CPU doesn't like that.
1740 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1741 * 383 on page 93. Intel should be safe but is also warns that it's
1742 * only safe if the permission and cache attributes of the two entries
1743 * loaded in the two TLB is identical (which should be the case here).
1744 * But it is generally safer to never allow small and huge TLB entries
1745 * for the same virtual address to be loaded simultaneously. So instead
1746 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1747 * current pmd notpresent (atomically because here the pmd_trans_huge
1748 * and pmd_trans_splitting must remain set at all times on the pmd
1749 * until the split is complete for this pmd), then we flush the SMP TLB
1750 * and finally we write the non-huge version of the pmd entry with
1751 * pmd_populate.
1752 */
1753 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001754 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001755
1756 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001757 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001758 page_remove_rmap(page + i, false);
1759 put_page(page + i);
1760 }
1761 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001762}
1763
1764void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001765 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001766{
1767 spinlock_t *ptl;
1768 struct mm_struct *mm = vma->vm_mm;
1769 unsigned long haddr = address & HPAGE_PMD_MASK;
Hugh Dickins332293a2021-02-04 18:32:31 -08001770 bool do_unlock_page = false;
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001771 pmd_t _pmd;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001772
1773 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1774 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001775
1776 /*
1777 * If caller asks to setup a migration entries, we need a page to check
1778 * pmd against. Otherwise we can end up replacing wrong page.
1779 */
1780 VM_BUG_ON(freeze && !page);
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001781 if (page) {
1782 VM_WARN_ON_ONCE(!PageLocked(page));
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001783 if (page != pmd_page(*pmd))
1784 goto out;
1785 }
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001786
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001787repeat:
Dan Williams5c7fb562016-01-15 16:56:52 -08001788 if (pmd_trans_huge(*pmd)) {
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001789 if (!page) {
1790 page = pmd_page(*pmd);
Hugh Dickins332293a2021-02-04 18:32:31 -08001791 /*
1792 * An anonymous page must be locked, to ensure that a
1793 * concurrent reuse_swap_page() sees stable mapcount;
1794 * but reuse_swap_page() is not used on shmem or file,
1795 * and page lock must not be taken when zap_pmd_range()
1796 * calls __split_huge_pmd() while i_mmap_lock is held.
1797 */
1798 if (PageAnon(page)) {
1799 if (unlikely(!trylock_page(page))) {
1800 get_page(page);
1801 _pmd = *pmd;
1802 spin_unlock(ptl);
1803 lock_page(page);
1804 spin_lock(ptl);
1805 if (unlikely(!pmd_same(*pmd, _pmd))) {
1806 unlock_page(page);
1807 put_page(page);
1808 page = NULL;
1809 goto repeat;
1810 }
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001811 put_page(page);
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001812 }
Hugh Dickins332293a2021-02-04 18:32:31 -08001813 do_unlock_page = true;
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001814 }
1815 }
Dan Williams5c7fb562016-01-15 16:56:52 -08001816 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07001817 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08001818 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001819 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001820 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001821out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001822 spin_unlock(ptl);
Hugh Dickins332293a2021-02-04 18:32:31 -08001823 if (do_unlock_page)
Andrea Arcangelid93d7bd2020-05-27 19:06:24 -04001824 unlock_page(page);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001825 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1826}
1827
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001828void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1829 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001830{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001831 pgd_t *pgd;
1832 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001833 pmd_t *pmd;
1834
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001835 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001836 if (!pgd_present(*pgd))
1837 return;
1838
1839 pud = pud_offset(pgd, address);
1840 if (!pud_present(*pud))
1841 return;
1842
1843 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001844
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001845 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001846}
1847
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07001848void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001849 unsigned long start,
1850 unsigned long end,
1851 long adjust_next)
1852{
1853 /*
1854 * If the new start address isn't hpage aligned and it could
1855 * previously contain an hugepage: check if we need to split
1856 * an huge pmd.
1857 */
1858 if (start & ~HPAGE_PMD_MASK &&
1859 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1860 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001861 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001862
1863 /*
1864 * If the new end address isn't hpage aligned and it could
1865 * previously contain an hugepage: check if we need to split
1866 * an huge pmd.
1867 */
1868 if (end & ~HPAGE_PMD_MASK &&
1869 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1870 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001871 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001872
1873 /*
1874 * If we're also updating the vma->vm_next->vm_start, if the new
1875 * vm_next->vm_start isn't page aligned and it could previously
1876 * contain an hugepage: check if we need to split an huge pmd.
1877 */
1878 if (adjust_next > 0) {
1879 struct vm_area_struct *next = vma->vm_next;
1880 unsigned long nstart = next->vm_start;
1881 nstart += adjust_next << PAGE_SHIFT;
1882 if (nstart & ~HPAGE_PMD_MASK &&
1883 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1884 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001885 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001886 }
1887}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001888
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08001889static void unmap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001890{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001891 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1892 TTU_RMAP_LOCKED;
Yang Shi07cc82b2021-06-15 18:24:07 -07001893 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001894
1895 VM_BUG_ON_PAGE(!PageHead(page), page);
1896
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001897 if (PageAnon(page))
1898 ttu_flags |= TTU_MIGRATION;
1899
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001900 /* We only need TTU_SPLIT_HUGE_PMD once */
Yang Shi07cc82b2021-06-15 18:24:07 -07001901 try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1902 for (i = 1; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001903 /* Cut short if the page is unmapped */
1904 if (page_count(page) == 1)
1905 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001906
Yang Shi07cc82b2021-06-15 18:24:07 -07001907 try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001908 }
Yang Shi07cc82b2021-06-15 18:24:07 -07001909
1910 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001911}
1912
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08001913static void remap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001914{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001915 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001916
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001917 for (i = 0; i < HPAGE_PMD_NR; i++)
1918 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001919}
1920
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001921static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001922 struct lruvec *lruvec, struct list_head *list)
1923{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001924 struct page *page_tail = head + tail;
1925
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001926 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001927
1928 /*
Konstantin Khlebnikovfb732e62018-04-05 16:23:28 -07001929 * Clone page flags before unfreezing refcount.
1930 *
1931 * After successful get_page_unless_zero() might follow flags change,
1932 * for exmaple lock_page() which set PG_waiters.
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001933 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001934 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1935 page_tail->flags |= (head->flags &
1936 ((1L << PG_referenced) |
1937 (1L << PG_swapbacked) |
1938 (1L << PG_mlocked) |
1939 (1L << PG_uptodate) |
1940 (1L << PG_active) |
1941 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001942 (1L << PG_unevictable) |
1943 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001944
Hugh Dickinsffdad592018-11-30 14:10:16 -08001945 /* ->mapping in first tail page is compound_mapcount */
1946 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1947 page_tail);
1948 page_tail->mapping = head->mapping;
1949 page_tail->index = head->index + tail;
1950
Konstantin Khlebnikovfb732e62018-04-05 16:23:28 -07001951 /* Page flags must be visible before we make the page non-compound. */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001952 smp_wmb();
1953
Konstantin Khlebnikovfb732e62018-04-05 16:23:28 -07001954 /*
1955 * Clear PageTail before unfreezing page refcount.
1956 *
1957 * After successful get_page_unless_zero() might follow put_page()
1958 * which needs correct compound_head().
1959 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001960 clear_compound_head(page_tail);
1961
Konstantin Khlebnikovfb732e62018-04-05 16:23:28 -07001962 /* Finally unfreeze refcount. Additional reference from page cache. */
1963 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
1964 PageSwapCache(head)));
1965
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001966 if (page_is_young(head))
1967 set_page_young(page_tail);
1968 if (page_is_idle(head))
1969 set_page_idle(page_tail);
1970
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001971 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1972 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001973}
1974
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001975static void __split_huge_page(struct page *page, struct list_head *list,
Hugh Dickinsb59b24f2018-11-30 14:10:21 -08001976 pgoff_t end, unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001977{
1978 struct page *head = compound_head(page);
1979 struct zone *zone = page_zone(head);
1980 struct lruvec *lruvec;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001981 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001982
Mel Gorman599d0c92016-07-28 15:45:31 -07001983 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001984
1985 /* complete memcg works before add pages to LRU */
1986 mem_cgroup_split_huge_fixup(head);
1987
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001988 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001989 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001990 /* Some pages can be beyond i_size: drop them from page cache */
1991 if (head[i].index >= end) {
1992 __ClearPageDirty(head + i);
1993 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001994 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1995 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001996 put_page(head + i);
1997 }
1998 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001999
2000 ClearPageCompound(head);
Vlastimil Babkaba37a942019-08-24 17:54:59 -07002001
2002 split_page_owner(head, HPAGE_PMD_ORDER);
2003
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002004 /* See comment in __split_huge_page_tail() */
2005 if (PageAnon(head)) {
2006 page_ref_inc(head);
2007 } else {
2008 /* Additional pin to radix tree */
2009 page_ref_add(head, 2);
2010 spin_unlock(&head->mapping->tree_lock);
2011 }
2012
Mel Gormana52633d2016-07-28 15:45:28 -07002013 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002014
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08002015 remap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002016
2017 for (i = 0; i < HPAGE_PMD_NR; i++) {
2018 struct page *subpage = head + i;
2019 if (subpage == page)
2020 continue;
2021 unlock_page(subpage);
2022
2023 /*
2024 * Subpages may be freed if there wasn't any mapping
2025 * like if add_to_swap() is running on a lru page that
2026 * had its mapping zapped. And freeing these pages
2027 * requires taking the lru_lock so we do the put_page
2028 * of the tail pages after the split is complete.
2029 */
2030 put_page(subpage);
2031 }
2032}
2033
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002034int total_mapcount(struct page *page)
2035{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002036 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002037
2038 VM_BUG_ON_PAGE(PageTail(page), page);
2039
2040 if (likely(!PageCompound(page)))
2041 return atomic_read(&page->_mapcount) + 1;
2042
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002043 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002044 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002045 return compound;
2046 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002047 for (i = 0; i < HPAGE_PMD_NR; i++)
2048 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002049 /* File pages has compound_mapcount included in _mapcount */
2050 if (!PageAnon(page))
2051 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002052 if (PageDoubleMap(page))
2053 ret -= HPAGE_PMD_NR;
2054 return ret;
2055}
2056
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002057/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002058 * This calculates accurately how many mappings a transparent hugepage
2059 * has (unlike page_mapcount() which isn't fully accurate). This full
2060 * accuracy is primarily needed to know if copy-on-write faults can
2061 * reuse the page and change the mapping to read-write instead of
2062 * copying them. At the same time this returns the total_mapcount too.
2063 *
2064 * The function returns the highest mapcount any one of the subpages
2065 * has. If the return value is one, even if different processes are
2066 * mapping different subpages of the transparent hugepage, they can
2067 * all reuse it, because each process is reusing a different subpage.
2068 *
2069 * The total_mapcount is instead counting all virtual mappings of the
2070 * subpages. If the total_mapcount is equal to "one", it tells the
2071 * caller all mappings belong to the same "mm" and in turn the
2072 * anon_vma of the transparent hugepage can become the vma->anon_vma
2073 * local one as no other process may be mapping any of the subpages.
2074 *
2075 * It would be more accurate to replace page_mapcount() with
2076 * page_trans_huge_mapcount(), however we only use
2077 * page_trans_huge_mapcount() in the copy-on-write faults where we
2078 * need full accuracy to avoid breaking page pinning, because
2079 * page_trans_huge_mapcount() is slower than page_mapcount().
2080 */
2081int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2082{
2083 int i, ret, _total_mapcount, mapcount;
2084
2085 /* hugetlbfs shouldn't call it */
2086 VM_BUG_ON_PAGE(PageHuge(page), page);
2087
2088 if (likely(!PageTransCompound(page))) {
2089 mapcount = atomic_read(&page->_mapcount) + 1;
2090 if (total_mapcount)
2091 *total_mapcount = mapcount;
2092 return mapcount;
2093 }
2094
2095 page = compound_head(page);
2096
2097 _total_mapcount = ret = 0;
2098 for (i = 0; i < HPAGE_PMD_NR; i++) {
2099 mapcount = atomic_read(&page[i]._mapcount) + 1;
2100 ret = max(ret, mapcount);
2101 _total_mapcount += mapcount;
2102 }
2103 if (PageDoubleMap(page)) {
2104 ret -= 1;
2105 _total_mapcount -= HPAGE_PMD_NR;
2106 }
2107 mapcount = compound_mapcount(page);
2108 ret += mapcount;
2109 _total_mapcount += mapcount;
2110 if (total_mapcount)
2111 *total_mapcount = _total_mapcount;
2112 return ret;
2113}
2114
2115/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002116 * This function splits huge page into normal pages. @page can point to any
2117 * subpage of huge page to split. Split doesn't change the position of @page.
2118 *
2119 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2120 * The huge page must be locked.
2121 *
2122 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2123 *
2124 * Both head page and tail pages will inherit mapping, flags, and so on from
2125 * the hugepage.
2126 *
2127 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2128 * they are not mapped.
2129 *
2130 * Returns 0 if the hugepage is split successfully.
2131 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2132 * us.
2133 */
2134int split_huge_page_to_list(struct page *page, struct list_head *list)
2135{
2136 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002137 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002138 struct anon_vma *anon_vma = NULL;
2139 struct address_space *mapping = NULL;
Yang Shi07cc82b2021-06-15 18:24:07 -07002140 int extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002141 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002142 unsigned long flags;
Hugh Dickinsb59b24f2018-11-30 14:10:21 -08002143 pgoff_t end;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002144
Wei Yangb2596722020-01-30 22:14:29 -08002145 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002146 VM_BUG_ON_PAGE(!PageLocked(page), page);
2147 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2148 VM_BUG_ON_PAGE(!PageCompound(page), page);
2149
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002150 if (PageAnon(head)) {
2151 /*
2152 * The caller does not necessarily hold an mmap_sem that would
2153 * prevent the anon_vma disappearing so we first we take a
2154 * reference to it and then lock the anon_vma for write. This
2155 * is similar to page_lock_anon_vma_read except the write lock
2156 * is taken to serialise against parallel split or collapse
2157 * operations.
2158 */
2159 anon_vma = page_get_anon_vma(head);
2160 if (!anon_vma) {
2161 ret = -EBUSY;
2162 goto out;
2163 }
2164 extra_pins = 0;
Hugh Dickinsb59b24f2018-11-30 14:10:21 -08002165 end = -1;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002166 mapping = NULL;
2167 anon_vma_lock_write(anon_vma);
2168 } else {
2169 mapping = head->mapping;
2170
2171 /* Truncated ? */
2172 if (!mapping) {
2173 ret = -EBUSY;
2174 goto out;
2175 }
2176
2177 /* Addidional pins from radix tree */
2178 extra_pins = HPAGE_PMD_NR;
2179 anon_vma = NULL;
2180 i_mmap_lock_read(mapping);
Hugh Dickinsb59b24f2018-11-30 14:10:21 -08002181
2182 /*
2183 *__split_huge_page() may need to trim off pages beyond EOF:
2184 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2185 * which cannot be nested inside the page tree lock. So note
2186 * end now: i_size itself may be changed at any moment, but
2187 * head page lock is good enough to serialize the trimming.
2188 */
2189 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002190 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002191
2192 /*
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08002193 * Racy check if we can split the page, before unmap_page() will
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002194 * split PMDs
2195 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002196 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002197 ret = -EBUSY;
2198 goto out_unlock;
2199 }
2200
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002201 mlocked = PageMlocked(page);
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08002202 unmap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002203
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002204 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2205 if (mlocked)
2206 lru_add_drain();
2207
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002208 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002209 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002210
2211 if (mapping) {
2212 void **pslot;
2213
2214 spin_lock(&mapping->tree_lock);
2215 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2216 page_index(head));
2217 /*
2218 * Check if the head page is present in radix tree.
2219 * We assume all tail are present too, if head is there.
2220 */
2221 if (radix_tree_deref_slot_protected(pslot,
2222 &mapping->tree_lock) != head)
2223 goto fail;
2224 }
2225
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002226 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002227 spin_lock(&pgdata->split_queue_lock);
Yang Shi07cc82b2021-06-15 18:24:07 -07002228 if (page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002229 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002230 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002231 list_del(page_deferred_list(head));
2232 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002233 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002234 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002235 spin_unlock(&pgdata->split_queue_lock);
Hugh Dickinsb59b24f2018-11-30 14:10:21 -08002236 __split_huge_page(page, list, end, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002237 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002238 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002239 spin_unlock(&pgdata->split_queue_lock);
Yang Shi07cc82b2021-06-15 18:24:07 -07002240fail:
2241 if (mapping)
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002242 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002243 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Hugh Dickinsb48c29b2018-11-30 14:10:13 -08002244 remap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002245 ret = -EBUSY;
2246 }
2247
2248out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002249 if (anon_vma) {
2250 anon_vma_unlock_write(anon_vma);
2251 put_anon_vma(anon_vma);
2252 }
2253 if (mapping)
2254 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002255out:
2256 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2257 return ret;
2258}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002259
2260void free_transhuge_page(struct page *page)
2261{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002262 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002263 unsigned long flags;
2264
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002265 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002266 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002267 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002268 list_del(page_deferred_list(page));
2269 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002270 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002271 free_compound_page(page);
2272}
2273
2274void deferred_split_huge_page(struct page *page)
2275{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002276 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002277 unsigned long flags;
2278
2279 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2280
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002281 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002282 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002283 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002284 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2285 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002286 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002287 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002288}
2289
2290static unsigned long deferred_split_count(struct shrinker *shrink,
2291 struct shrink_control *sc)
2292{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002293 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002294 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002295}
2296
2297static unsigned long deferred_split_scan(struct shrinker *shrink,
2298 struct shrink_control *sc)
2299{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002300 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002301 unsigned long flags;
2302 LIST_HEAD(list), *pos, *next;
2303 struct page *page;
2304 int split = 0;
2305
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002306 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002307 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002308 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002309 page = list_entry((void *)pos, struct page, mapping);
2310 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002311 if (get_page_unless_zero(page)) {
2312 list_move(page_deferred_list(page), &list);
2313 } else {
2314 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002315 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002316 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002317 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002318 if (!--sc->nr_to_scan)
2319 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002320 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002321 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002322
2323 list_for_each_safe(pos, next, &list) {
2324 page = list_entry((void *)pos, struct page, mapping);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002325 if (!trylock_page(page))
2326 goto next;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002327 /* split_huge_page() removes page from list on success */
2328 if (!split_huge_page(page))
2329 split++;
2330 unlock_page(page);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002331next:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002332 put_page(page);
2333 }
2334
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002335 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2336 list_splice_tail(&list, &pgdata->split_queue);
2337 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002338
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002339 /*
2340 * Stop shrinker if we didn't split any page, but the queue is empty.
2341 * This can happen if pages were freed under us.
2342 */
2343 if (!split && list_empty(&pgdata->split_queue))
2344 return SHRINK_STOP;
2345 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002346}
2347
2348static struct shrinker deferred_split_shrinker = {
2349 .count_objects = deferred_split_count,
2350 .scan_objects = deferred_split_scan,
2351 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002352 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002353};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002354
2355#ifdef CONFIG_DEBUG_FS
2356static int split_huge_pages_set(void *data, u64 val)
2357{
2358 struct zone *zone;
2359 struct page *page;
2360 unsigned long pfn, max_zone_pfn;
2361 unsigned long total = 0, split = 0;
2362
2363 if (val != 1)
2364 return -EINVAL;
2365
2366 for_each_populated_zone(zone) {
2367 max_zone_pfn = zone_end_pfn(zone);
2368 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2369 if (!pfn_valid(pfn))
2370 continue;
2371
2372 page = pfn_to_page(pfn);
2373 if (!get_page_unless_zero(page))
2374 continue;
2375
2376 if (zone != page_zone(page))
2377 goto next;
2378
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002379 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002380 goto next;
2381
2382 total++;
2383 lock_page(page);
2384 if (!split_huge_page(page))
2385 split++;
2386 unlock_page(page);
2387next:
2388 put_page(page);
2389 }
2390 }
2391
Yang Shi145bdaa2016-05-05 16:22:00 -07002392 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002393
2394 return 0;
2395}
2396DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2397 "%llu\n");
2398
2399static int __init split_huge_pages_debugfs(void)
2400{
2401 void *ret;
2402
Yang Shi145bdaa2016-05-05 16:22:00 -07002403 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002404 &split_huge_pages_fops);
2405 if (!ret)
2406 pr_warn("Failed to create split_huge_pages in debugfs");
2407 return 0;
2408}
2409late_initcall(split_huge_pages_debugfs);
2410#endif