blob: e2982ea26090e8f21782b6cb5b984a89d93dba96 [file] [log] [blame]
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
Andrew Mortonae3a8c12014-06-04 16:06:58 -07008#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080010#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/highmem.h>
13#include <linux/hugetlb.h>
14#include <linux/mmu_notifier.h>
15#include <linux/rmap.h>
16#include <linux/swap.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080017#include <linux/shrinker.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080018#include <linux/mm_inline.h>
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -080019#include <linux/swapops.h>
Matthew Wilcox4897c762015-09-08 14:58:45 -070020#include <linux/dax.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080021#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080022#include <linux/freezer.h>
Dan Williamsf25748e32016-01-15 16:56:43 -080023#include <linux/pfn_t.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080024#include <linux/mman.h>
Dan Williams3565fce2016-01-15 16:56:55 -080025#include <linux/memremap.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020026#include <linux/pagemap.h>
Kirill A. Shutemov49071d42016-01-15 16:54:40 -080027#include <linux/debugfs.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000028#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080029#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070030#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070031#include <linux/page_idle.h>
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -070032#include <linux/shmem_fs.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080033
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080034#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
Andrea Arcangeliba761492011-01-13 15:46:58 -080038/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080039 * By default transparent hugepage support is disabled in order that avoid
40 * to risk increase the memory footprint of applications without a guaranteed
41 * benefit. When transparent hugepage support is enabled, is for all mappings,
42 * and khugepaged scans all mappings.
43 * Defrag is invoked by khugepaged hugepage allocations and by page faults
44 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080045 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080046unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080047#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080048 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080049#endif
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070053 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080054 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080056
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080057static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080058
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080059static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080060struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080061
Aaron Lu6fcb52a2016-10-07 17:00:08 -070062static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080063{
64 struct page *zero_page;
65retry:
66 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070067 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080068
69 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080071 if (!zero_page) {
72 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070073 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080074 }
75 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080076 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070077 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080078 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070079 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080080 goto retry;
81 }
82
83 /* We take additional reference here. It will be put back by shrinker */
84 atomic_set(&huge_zero_refcount, 2);
85 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070086 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080087}
88
Aaron Lu6fcb52a2016-10-07 17:00:08 -070089static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080090{
91 /*
92 * Counter should never go to zero here. Only shrinker can put
93 * last reference.
94 */
95 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96}
97
Aaron Lu6fcb52a2016-10-07 17:00:08 -070098struct page *mm_get_huge_zero_page(struct mm_struct *mm)
99{
100 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
101 return READ_ONCE(huge_zero_page);
102
103 if (!get_huge_zero_page())
104 return NULL;
105
106 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
107 put_huge_zero_page();
108
109 return READ_ONCE(huge_zero_page);
110}
111
112void mm_put_huge_zero_page(struct mm_struct *mm)
113{
114 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 put_huge_zero_page();
116}
117
Glauber Costa48896462013-08-28 10:18:15 +1000118static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
119 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800120{
Glauber Costa48896462013-08-28 10:18:15 +1000121 /* we can free zero page only if last reference remains */
122 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
123}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800124
Glauber Costa48896462013-08-28 10:18:15 +1000125static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
126 struct shrink_control *sc)
127{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800128 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700129 struct page *zero_page = xchg(&huge_zero_page, NULL);
130 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700131 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000132 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800133 }
134
135 return 0;
136}
137
138static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000139 .count_objects = shrink_huge_zero_page_count,
140 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800141 .seeks = DEFAULT_SEEKS,
142};
143
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800144#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800145
Mel Gorman444eb2a42016-03-17 14:19:23 -0700146static ssize_t triple_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800147 struct kobj_attribute *attr,
148 const char *buf, size_t count,
149 enum transparent_hugepage_flag enabled,
Mel Gorman444eb2a42016-03-17 14:19:23 -0700150 enum transparent_hugepage_flag deferred,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800151 enum transparent_hugepage_flag req_madv)
152{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700153 if (!memcmp("defer", buf,
154 min(sizeof("defer")-1, count))) {
155 if (enabled == deferred)
156 return -EINVAL;
157 clear_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800158 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700159 set_bit(deferred, &transparent_hugepage_flags);
160 } else if (!memcmp("always", buf,
161 min(sizeof("always")-1, count))) {
162 clear_bit(deferred, &transparent_hugepage_flags);
163 clear_bit(req_madv, &transparent_hugepage_flags);
164 set_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800165 } else if (!memcmp("madvise", buf,
166 min(sizeof("madvise")-1, count))) {
167 clear_bit(enabled, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700168 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800169 set_bit(req_madv, &transparent_hugepage_flags);
170 } else if (!memcmp("never", buf,
171 min(sizeof("never")-1, count))) {
172 clear_bit(enabled, &transparent_hugepage_flags);
173 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700174 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800175 } else
176 return -EINVAL;
177
178 return count;
179}
180
181static ssize_t enabled_show(struct kobject *kobj,
182 struct kobj_attribute *attr, char *buf)
183{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700184 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
185 return sprintf(buf, "[always] madvise never\n");
186 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
187 return sprintf(buf, "always [madvise] never\n");
188 else
189 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800190}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700191
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800192static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
195{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800196 ssize_t ret;
197
Mel Gorman444eb2a42016-03-17 14:19:23 -0700198 ret = triple_flag_store(kobj, attr, buf, count,
199 TRANSPARENT_HUGEPAGE_FLAG,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800200 TRANSPARENT_HUGEPAGE_FLAG,
201 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
202
203 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700204 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800205 if (err)
206 ret = err;
207 }
208
209 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800210}
211static struct kobj_attribute enabled_attr =
212 __ATTR(enabled, 0644, enabled_show, enabled_store);
213
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700214ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800215 struct kobj_attribute *attr, char *buf,
216 enum transparent_hugepage_flag flag)
217{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700218 return sprintf(buf, "%d\n",
219 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800220}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700221
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700222ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800223 struct kobj_attribute *attr,
224 const char *buf, size_t count,
225 enum transparent_hugepage_flag flag)
226{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700227 unsigned long value;
228 int ret;
229
230 ret = kstrtoul(buf, 10, &value);
231 if (ret < 0)
232 return ret;
233 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800234 return -EINVAL;
235
Ben Hutchingse27e6152011-04-14 15:22:21 -0700236 if (value)
237 set_bit(flag, &transparent_hugepage_flags);
238 else
239 clear_bit(flag, &transparent_hugepage_flags);
240
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800241 return count;
242}
243
244/*
245 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
246 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
247 * memory just to allocate one more hugepage.
248 */
249static ssize_t defrag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf)
251{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700252 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
253 return sprintf(buf, "[always] defer madvise never\n");
254 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
255 return sprintf(buf, "always [defer] madvise never\n");
256 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
257 return sprintf(buf, "always defer [madvise] never\n");
258 else
259 return sprintf(buf, "always defer madvise [never]\n");
260
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800261}
262static ssize_t defrag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count)
265{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700266 return triple_flag_store(kobj, attr, buf, count,
267 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
268 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800269 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
270}
271static struct kobj_attribute defrag_attr =
272 __ATTR(defrag, 0644, defrag_show, defrag_store);
273
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800274static ssize_t use_zero_page_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
276{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700277 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800278 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
279}
280static ssize_t use_zero_page_store(struct kobject *kobj,
281 struct kobj_attribute *attr, const char *buf, size_t count)
282{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700283 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800284 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
285}
286static struct kobj_attribute use_zero_page_attr =
287 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800288#ifdef CONFIG_DEBUG_VM
289static ssize_t debug_cow_show(struct kobject *kobj,
290 struct kobj_attribute *attr, char *buf)
291{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700292 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800293 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
294}
295static ssize_t debug_cow_store(struct kobject *kobj,
296 struct kobj_attribute *attr,
297 const char *buf, size_t count)
298{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700299 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800300 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
301}
302static struct kobj_attribute debug_cow_attr =
303 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
304#endif /* CONFIG_DEBUG_VM */
305
306static struct attribute *hugepage_attr[] = {
307 &enabled_attr.attr,
308 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800309 &use_zero_page_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700310#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700311 &shmem_enabled_attr.attr,
312#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800313#ifdef CONFIG_DEBUG_VM
314 &debug_cow_attr.attr,
315#endif
316 NULL,
317};
318
319static struct attribute_group hugepage_attr_group = {
320 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800321};
322
Shaohua Li569e5592012-01-12 17:19:11 -0800323static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
324{
325 int err;
326
327 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
328 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700329 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800330 return -ENOMEM;
331 }
332
333 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
334 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700335 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800336 goto delete_obj;
337 }
338
339 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
340 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700341 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800342 goto remove_hp_group;
343 }
344
345 return 0;
346
347remove_hp_group:
348 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
349delete_obj:
350 kobject_put(*hugepage_kobj);
351 return err;
352}
353
354static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
355{
356 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
357 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
358 kobject_put(hugepage_kobj);
359}
360#else
361static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
362{
363 return 0;
364}
365
366static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800369#endif /* CONFIG_SYSFS */
370
371static int __init hugepage_init(void)
372{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800373 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800374 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800375
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800376 if (!has_transparent_hugepage()) {
377 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800378 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800379 }
380
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530381 /*
382 * hugepages can't be allocated by the buddy allocator
383 */
384 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
385 /*
386 * we use page->mapping and page->index in second tail page
387 * as list_head: assuming THP order >= 2
388 */
389 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
390
Shaohua Li569e5592012-01-12 17:19:11 -0800391 err = hugepage_init_sysfs(&hugepage_kobj);
392 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700393 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800394
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700395 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800396 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700397 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800398
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700399 err = register_shrinker(&huge_zero_page_shrinker);
400 if (err)
401 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800402 err = register_shrinker(&deferred_split_shrinker);
403 if (err)
404 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800405
Rik van Riel97562cd2011-01-13 15:47:12 -0800406 /*
407 * By default disable transparent hugepages on smaller systems,
408 * where the extra memory used could hurt more than TLB overhead
409 * is likely to save. The admin can still enable it through /sys.
410 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700411 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800412 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700413 return 0;
414 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800415
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700416 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700417 if (err)
418 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800419
Shaohua Li569e5592012-01-12 17:19:11 -0800420 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700421err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800422 unregister_shrinker(&deferred_split_shrinker);
423err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700424 unregister_shrinker(&huge_zero_page_shrinker);
425err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700426 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700427err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800428 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700429err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800430 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800431}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800432subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800433
434static int __init setup_transparent_hugepage(char *str)
435{
436 int ret = 0;
437 if (!str)
438 goto out;
439 if (!strcmp(str, "always")) {
440 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
441 &transparent_hugepage_flags);
442 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
443 &transparent_hugepage_flags);
444 ret = 1;
445 } else if (!strcmp(str, "madvise")) {
446 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
447 &transparent_hugepage_flags);
448 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
449 &transparent_hugepage_flags);
450 ret = 1;
451 } else if (!strcmp(str, "never")) {
452 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 }
458out:
459 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700460 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800461 return ret;
462}
463__setup("transparent_hugepage=", setup_transparent_hugepage);
464
Mel Gormanb32967f2012-11-19 12:35:47 +0000465pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800466{
467 if (likely(vma->vm_flags & VM_WRITE))
468 pmd = pmd_mkwrite(pmd);
469 return pmd;
470}
471
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800472static inline struct list_head *page_deferred_list(struct page *page)
473{
474 /*
475 * ->lru in the tail pages is occupied by compound_head.
476 * Let's use ->mapping + ->index in the second tail page as list_head.
477 */
478 return (struct list_head *)&page[2].mapping;
479}
480
481void prep_transhuge_page(struct page *page)
482{
483 /*
484 * we use page->mapping and page->indexlru in second tail page
485 * as list_head: assuming THP order >= 2
486 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800487
488 INIT_LIST_HEAD(page_deferred_list(page));
489 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
490}
491
Toshi Kani74d2fad2016-10-07 16:59:56 -0700492unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
493 loff_t off, unsigned long flags, unsigned long size)
494{
495 unsigned long addr;
496 loff_t off_end = off + len;
497 loff_t off_align = round_up(off, size);
498 unsigned long len_pad;
499
500 if (off_end <= off_align || (off_end - off_align) < size)
501 return 0;
502
503 len_pad = len + size;
504 if (len_pad < len || (off + len_pad) < off)
505 return 0;
506
507 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
508 off >> PAGE_SHIFT, flags);
509 if (IS_ERR_VALUE(addr))
510 return 0;
511
512 addr += (off - addr) & (size - 1);
513 return addr;
514}
515
516unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
517 unsigned long len, unsigned long pgoff, unsigned long flags)
518{
519 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
520
521 if (addr)
522 goto out;
523 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
524 goto out;
525
526 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
527 if (addr)
528 return addr;
529
530 out:
531 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
532}
533EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
534
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700535static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
536 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800537{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700538 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700539 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800540 pgtable_t pgtable;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700541 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800542
Sasha Levin309381fea2014-01-23 15:52:54 -0800543 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700544
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700545 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700546 put_page(page);
547 count_vm_event(THP_FAULT_FALLBACK);
548 return VM_FAULT_FALLBACK;
549 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800550
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700551 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700552 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800553 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700554 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700555 return VM_FAULT_OOM;
556 }
557
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800558 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700559 /*
560 * The memory barrier inside __SetPageUptodate makes sure that
561 * clear_huge_page writes become visible before the set_pmd_at()
562 * write.
563 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800564 __SetPageUptodate(page);
565
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700566 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
567 if (unlikely(!pmd_none(*fe->pmd))) {
568 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800569 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800570 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700571 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800572 } else {
573 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700574
575 /* Deliver the page fault to userland */
576 if (userfaultfd_missing(vma)) {
577 int ret;
578
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700579 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800580 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700581 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700582 pte_free(vma->vm_mm, pgtable);
583 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700584 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
585 return ret;
586 }
587
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700588 entry = mk_huge_pmd(page, vma->vm_page_prot);
589 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800590 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800591 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700592 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700593 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
594 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
595 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
596 atomic_long_inc(&vma->vm_mm->nr_ptes);
597 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700598 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800599 }
600
David Rientjesaa2e8782012-05-29 15:06:17 -0700601 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800602}
603
Mel Gorman444eb2a42016-03-17 14:19:23 -0700604/*
Vlastimil Babka25160352016-07-28 15:49:25 -0700605 * If THP defrag is set to always then directly reclaim/compact as necessary
606 * If set to defer then do only background reclaim/compact and defer to khugepaged
Mel Gorman444eb2a42016-03-17 14:19:23 -0700607 * If set to madvise and the VMA is flagged then directly reclaim/compact
Vlastimil Babka25160352016-07-28 15:49:25 -0700608 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
Mel Gorman444eb2a42016-03-17 14:19:23 -0700609 */
610static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800611{
Vlastimil Babka25160352016-07-28 15:49:25 -0700612 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700613
Vlastimil Babka25160352016-07-28 15:49:25 -0700614 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
615 &transparent_hugepage_flags) && vma_madvised)
616 return GFP_TRANSHUGE;
617 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
618 &transparent_hugepage_flags))
619 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
620 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
621 &transparent_hugepage_flags))
622 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700623
Vlastimil Babka25160352016-07-28 15:49:25 -0700624 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700625}
626
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800627/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700628static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800629 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700630 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800631{
632 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700633 if (!pmd_none(*pmd))
634 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700635 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800636 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800637 if (pgtable)
638 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800639 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800640 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700641 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800642}
643
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700644int do_huge_pmd_anonymous_page(struct fault_env *fe)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800645{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700646 struct vm_area_struct *vma = fe->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800647 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800648 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700649 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800650
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700651 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700652 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700653 if (unlikely(anon_vma_prepare(vma)))
654 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700655 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700656 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700657 if (!(fe->flags & FAULT_FLAG_WRITE) &&
658 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700659 transparent_hugepage_use_zero_page()) {
660 pgtable_t pgtable;
661 struct page *zero_page;
662 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700663 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700664 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700665 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800666 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700667 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700668 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700669 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700670 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700671 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700672 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700673 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700674 ret = 0;
675 set = false;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700676 if (pmd_none(*fe->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700677 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700678 spin_unlock(fe->ptl);
679 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700680 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
681 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700682 set_huge_zero_page(pgtable, vma->vm_mm, vma,
683 haddr, fe->pmd, zero_page);
684 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700685 set = true;
686 }
687 } else
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700688 spin_unlock(fe->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700689 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700690 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700691 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800692 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700693 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800694 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700695 if (unlikely(!page)) {
696 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700697 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700698 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800699 prep_transhuge_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700700 return __do_huge_pmd_anonymous_page(fe, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800701}
702
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700703static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800704 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700705{
706 struct mm_struct *mm = vma->vm_mm;
707 pmd_t entry;
708 spinlock_t *ptl;
709
710 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800711 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
712 if (pfn_t_devmap(pfn))
713 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800714 if (write) {
715 entry = pmd_mkyoung(pmd_mkdirty(entry));
716 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700717 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800718 set_pmd_at(mm, addr, pmd, entry);
719 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700720 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700721}
722
723int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800724 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700725{
726 pgprot_t pgprot = vma->vm_page_prot;
727 /*
728 * If we had pmd_special, we could avoid all these restrictions,
729 * but we need to be consistent with PTEs and architectures that
730 * can't support a 'special' bit.
731 */
732 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
733 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
734 (VM_PFNMAP|VM_MIXEDMAP));
735 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800736 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700737
738 if (addr < vma->vm_start || addr >= vma->vm_end)
739 return VM_FAULT_SIGBUS;
740 if (track_pfn_insert(vma, &pgprot, pfn))
741 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700742 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
743 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700744}
Dan Williamsdee41072016-05-14 12:20:44 -0700745EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700746
Dan Williams3565fce2016-01-15 16:56:55 -0800747static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
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 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001025 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001026 pmd_t entry;
1027 entry = pmd_mkyoung(orig_pmd);
1028 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001029 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1030 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001031 ret |= VM_FAULT_WRITE;
1032 goto out_unlock;
1033 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001034 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001035 spin_unlock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001036alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001037 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001038 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001039 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001040 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001041 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001042 new_page = NULL;
1043
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001044 if (likely(new_page)) {
1045 prep_transhuge_page(new_page);
1046 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001047 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001048 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001049 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001050 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001051 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001052 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001053 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001054 ret |= VM_FAULT_FALLBACK;
1055 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001056 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001057 }
David Rientjes17766dd2013-09-12 15:14:06 -07001058 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001059 goto out;
1060 }
1061
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001062 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1063 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001064 put_page(new_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001065 split_huge_pmd(vma, fe->pmd, fe->address);
1066 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001067 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001068 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001069 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001070 goto out;
1071 }
1072
David Rientjes17766dd2013-09-12 15:14:06 -07001073 count_vm_event(THP_FAULT_ALLOC);
1074
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001075 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001076 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1077 else
1078 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001079 __SetPageUptodate(new_page);
1080
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001081 mmun_start = haddr;
1082 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001083 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001084
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001085 spin_lock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001086 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001087 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001088 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1089 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001090 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001091 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001092 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001093 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001094 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001095 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1096 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001097 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001098 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001099 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001100 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001101 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1102 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001103 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001104 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001105 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001106 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001107 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001108 put_page(page);
1109 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001110 ret |= VM_FAULT_WRITE;
1111 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001112 spin_unlock(fe->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001113out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001114 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001115out:
1116 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001117out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001118 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001119 return ret;
1120}
1121
Keno Fischer6676aa62017-01-24 15:17:48 -08001122/*
1123 * FOLL_FORCE can write to even unwritable pmd's, but only
1124 * after we've gone through a COW cycle and they are dirty.
1125 */
1126static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1127{
1128 return pmd_write(pmd) ||
1129 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1130}
1131
David Rientjesb676b292012-10-08 16:34:03 -07001132struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001133 unsigned long addr,
1134 pmd_t *pmd,
1135 unsigned int flags)
1136{
David Rientjesb676b292012-10-08 16:34:03 -07001137 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001138 struct page *page = NULL;
1139
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001140 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001141
Keno Fischer6676aa62017-01-24 15:17:48 -08001142 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001143 goto out;
1144
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001145 /* Avoid dumping huge zero page */
1146 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1147 return ERR_PTR(-EFAULT);
1148
Mel Gorman2b4847e2013-12-18 17:08:32 -08001149 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001150 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001151 goto out;
1152
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001153 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001154 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001155 if (flags & FOLL_TOUCH)
Kirill A. Shutemov7031ae22017-11-27 06:21:25 +03001156 touch_pmd(vma, addr, pmd, flags);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001157 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001158 /*
1159 * We don't mlock() pte-mapped THPs. This way we can avoid
1160 * leaking mlocked pages into non-VM_LOCKED VMAs.
1161 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001162 * For anon THP:
1163 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001164 * In most cases the pmd is the only mapping of the page as we
1165 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1166 * writable private mappings in populate_vma_page_range().
1167 *
1168 * The only scenario when we have the page shared here is if we
1169 * mlocking read-only mapping shared over fork(). We skip
1170 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001171 *
1172 * For file THP:
1173 *
1174 * We can expect PageDoubleMap() to be stable under page lock:
1175 * for file pages we set it in page_add_file_rmap(), which
1176 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001177 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001178
1179 if (PageAnon(page) && compound_mapcount(page) != 1)
1180 goto skip_mlock;
1181 if (PageDoubleMap(page) || !page->mapping)
1182 goto skip_mlock;
1183 if (!trylock_page(page))
1184 goto skip_mlock;
1185 lru_add_drain();
1186 if (page->mapping && !PageDoubleMap(page))
1187 mlock_vma_page(page);
1188 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001189 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001190skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001192 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001193 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001194 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001195
1196out:
1197 return page;
1198}
1199
Mel Gormand10e63f2012-10-25 14:16:31 +02001200/* NUMA hinting page fault entry point for trans huge pmds */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001201int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001202{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001203 struct vm_area_struct *vma = fe->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001204 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001205 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001206 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001207 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001208 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001209 bool page_locked;
1210 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001211 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001212 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001213
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001214 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1215 if (unlikely(!pmd_same(pmd, *fe->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001216 goto out_unlock;
1217
Mel Gormande466bd2013-12-18 17:08:42 -08001218 /*
1219 * If there are potential migrations, wait for completion and retry
1220 * without disrupting NUMA hinting information. Do not relock and
1221 * check_same as the page may no longer be mapped.
1222 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001223 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1224 page = pmd_page(*fe->pmd);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001225 if (!get_page_unless_zero(page))
1226 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001227 spin_unlock(fe->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001228 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001229 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001230 goto out;
1231 }
1232
Mel Gormand10e63f2012-10-25 14:16:31 +02001233 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001234 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001235 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001236 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001237 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001238 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001239 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001240 flags |= TNF_FAULT_LOCAL;
1241 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001242
Mel Gormanbea66fb2015-03-25 15:55:37 -07001243 /* See similar comment in do_numa_page for explanation */
Rik van Rield59dc7b2016-09-08 21:30:53 -04001244 if (!pmd_write(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001245 flags |= TNF_NO_GROUP;
1246
1247 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001248 * Acquire the page lock to serialise THP migrations but avoid dropping
1249 * page_table_lock if at all possible
1250 */
Mel Gormanb8916632013-10-07 11:28:44 +01001251 page_locked = trylock_page(page);
1252 target_nid = mpol_misplaced(page, vma, haddr);
1253 if (target_nid == -1) {
1254 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001255 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001256 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001257 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001258
Mel Gormande466bd2013-12-18 17:08:42 -08001259 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001260 if (!page_locked) {
Mark Rutland2aa6d032017-06-16 14:02:34 -07001261 if (!get_page_unless_zero(page))
1262 goto out_unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001263 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001264 wait_on_page_locked(page);
Mark Rutland2aa6d032017-06-16 14:02:34 -07001265 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001266 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001267 goto out;
1268 }
1269
Mel Gorman2b4847e2013-12-18 17:08:32 -08001270 /*
1271 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1272 * to serialises splits
1273 */
Mel Gormanb8916632013-10-07 11:28:44 +01001274 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001275 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001276 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001277
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001278 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001279 spin_lock(fe->ptl);
1280 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001281 unlock_page(page);
1282 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001283 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001284 goto out_unlock;
1285 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001286
Mel Gormanc3a489c2013-12-18 17:08:38 -08001287 /* Bail if we fail to protect against THP splits for any reason */
1288 if (unlikely(!anon_vma)) {
1289 put_page(page);
1290 page_nid = -1;
1291 goto clear_pmdnuma;
1292 }
1293
Mel Gormana54a4072013-10-07 11:28:46 +01001294 /*
1295 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001296 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001297 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001298 spin_unlock(fe->ptl);
1299 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1300 fe->pmd, pmd, fe->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001301 if (migrated) {
1302 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001303 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001304 } else
1305 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001306
Mel Gorman8191acb2013-10-07 11:28:45 +01001307 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001308clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001309 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001310 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001311 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001312 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001313 if (was_writable)
1314 pmd = pmd_mkwrite(pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001315 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1316 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001317 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001318out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001319 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001320
1321out:
1322 if (anon_vma)
1323 page_unlock_anon_vma_read(anon_vma);
1324
Mel Gorman8191acb2013-10-07 11:28:45 +01001325 if (page_nid != -1)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001326 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001327
Mel Gormand10e63f2012-10-25 14:16:31 +02001328 return 0;
1329}
1330
Huang Ying319904a2016-07-28 15:48:03 -07001331/*
1332 * Return true if we do MADV_FREE successfully on entire pmd page.
1333 * Otherwise, return false.
1334 */
1335bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001336 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001337{
1338 spinlock_t *ptl;
1339 pmd_t orig_pmd;
1340 struct page *page;
1341 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001342 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001343
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001344 ptl = pmd_trans_huge_lock(pmd, vma);
1345 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001346 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001347
1348 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001349 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001350 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001351
1352 page = pmd_page(orig_pmd);
1353 /*
1354 * If other processes are mapping this page, we couldn't discard
1355 * the page unless they all do MADV_FREE so let's skip the page.
1356 */
1357 if (page_mapcount(page) != 1)
1358 goto out;
1359
1360 if (!trylock_page(page))
1361 goto out;
1362
1363 /*
1364 * If user want to discard part-pages of THP, split it so MADV_FREE
1365 * will deactivate only them.
1366 */
1367 if (next - addr != HPAGE_PMD_SIZE) {
1368 get_page(page);
1369 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001370 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001371 unlock_page(page);
Kirill A. Shutemovd2b64682017-07-06 15:35:28 -07001372 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001373 goto out_unlocked;
1374 }
1375
1376 if (PageDirty(page))
1377 ClearPageDirty(page);
1378 unlock_page(page);
1379
1380 if (PageActive(page))
1381 deactivate_page(page);
1382
1383 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemovf5848032017-04-13 14:56:26 -07001384 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001385 orig_pmd = pmd_mkold(orig_pmd);
1386 orig_pmd = pmd_mkclean(orig_pmd);
1387
1388 set_pmd_at(mm, addr, pmd, orig_pmd);
1389 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1390 }
Huang Ying319904a2016-07-28 15:48:03 -07001391 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001392out:
1393 spin_unlock(ptl);
1394out_unlocked:
1395 return ret;
1396}
1397
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001398int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001399 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001400{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001401 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001402 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001403
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001404 ptl = __pmd_trans_huge_lock(pmd, vma);
1405 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001406 return 0;
1407 /*
1408 * For architectures like ppc64 we look at deposited pgtable
1409 * when calling pmdp_huge_get_and_clear. So do the
1410 * pgtable_trans_huge_withdraw after finishing pmdp related
1411 * operations.
1412 */
1413 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1414 tlb->fullmm);
1415 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1416 if (vma_is_dax(vma)) {
1417 spin_unlock(ptl);
1418 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001419 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001420 } else if (is_huge_zero_pmd(orig_pmd)) {
1421 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1422 atomic_long_dec(&tlb->mm->nr_ptes);
1423 spin_unlock(ptl);
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001424 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001425 } else {
1426 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001427 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001428 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001429 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001430 if (PageAnon(page)) {
1431 pgtable_t pgtable;
1432 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1433 pte_free(tlb->mm, pgtable);
1434 atomic_long_dec(&tlb->mm->nr_ptes);
1435 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1436 } else {
1437 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1438 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001439 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001440 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001441 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001442 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001443}
1444
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001445bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001446 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001447 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001448{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001449 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001450 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001451 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001452 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001453
1454 if ((old_addr & ~HPAGE_PMD_MASK) ||
1455 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001456 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001457 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001458
1459 /*
1460 * The destination pmd shouldn't be established, free_pgtables()
1461 * should have release it.
1462 */
1463 if (WARN_ON(!pmd_none(*new_pmd))) {
1464 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001465 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001466 }
1467
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001468 /*
1469 * We don't have to worry about the ordering of src and dst
1470 * ptlocks because exclusive mmap_sem prevents deadlock.
1471 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001472 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1473 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001474 new_ptl = pmd_lockptr(mm, new_pmd);
1475 if (new_ptl != old_ptl)
1476 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001477 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001478 if (pmd_present(pmd) && pmd_dirty(pmd))
1479 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001480 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001481
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001482 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1483 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301484 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001485 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1486 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001487 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301488 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1489 if (new_ptl != old_ptl)
1490 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001491 if (force_flush)
1492 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1493 else
1494 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001495 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001496 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001497 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001498 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001499}
1500
Mel Gormanf123d742013-10-07 11:28:49 +01001501/*
1502 * Returns
1503 * - 0 if PMD could not be locked
1504 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1505 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1506 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001507int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001508 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001509{
1510 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001511 spinlock_t *ptl;
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001512 pmd_t entry;
1513 bool preserve_write;
1514 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001515
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001516 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001517 if (!ptl)
1518 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001519
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001520 preserve_write = prot_numa && pmd_write(*pmd);
1521 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001522
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001523 /*
1524 * Avoid trapping faults against the zero page. The read-only
1525 * data is likely to be read-cached on the local CPU and
1526 * local/remote hits to the zero page are not interesting.
1527 */
1528 if (prot_numa && is_huge_zero_pmd(*pmd))
1529 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001530
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001531 if (prot_numa && pmd_protnone(*pmd))
1532 goto unlock;
1533
Kirill A. Shutemovc2edc332017-04-13 14:56:20 -07001534 /*
1535 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1536 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1537 * which is also under down_read(mmap_sem):
1538 *
1539 * CPU0: CPU1:
1540 * change_huge_pmd(prot_numa=1)
1541 * pmdp_huge_get_and_clear_notify()
1542 * madvise_dontneed()
1543 * zap_pmd_range()
1544 * pmd_trans_huge(*pmd) == 0 (without ptl)
1545 * // skip the pmd
1546 * set_pmd_at();
1547 * // pmd is re-established
1548 *
1549 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1550 * which may break userspace.
1551 *
1552 * pmdp_invalidate() is required to make sure we don't miss
1553 * dirty/young flags set by hardware.
1554 */
1555 entry = *pmd;
1556 pmdp_invalidate(vma, addr, pmd);
1557
1558 /*
1559 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1560 * corrupt them.
1561 */
1562 if (pmd_dirty(*pmd))
1563 entry = pmd_mkdirty(entry);
1564 if (pmd_young(*pmd))
1565 entry = pmd_mkyoung(entry);
1566
Kirill A. Shutemov7bdd6852017-04-13 14:56:17 -07001567 entry = pmd_modify(entry, newprot);
1568 if (preserve_write)
1569 entry = pmd_mkwrite(entry);
1570 ret = HPAGE_PMD_NR;
1571 set_pmd_at(mm, addr, pmd, entry);
1572 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1573unlock:
1574 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001575 return ret;
1576}
1577
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001578/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001579 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001580 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001581 * Note that if it returns page table lock pointer, this routine returns without
1582 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001583 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001584spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001585{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001586 spinlock_t *ptl;
1587 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001588 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001589 return ptl;
1590 spin_unlock(ptl);
1591 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001592}
1593
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001594static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1595 unsigned long haddr, pmd_t *pmd)
1596{
1597 struct mm_struct *mm = vma->vm_mm;
1598 pgtable_t pgtable;
1599 pmd_t _pmd;
1600 int i;
1601
1602 /* leave pmd empty until pte is filled */
1603 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1604
1605 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1606 pmd_populate(mm, &_pmd, pgtable);
1607
1608 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1609 pte_t *pte, entry;
1610 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1611 entry = pte_mkspecial(entry);
1612 pte = pte_offset_map(&_pmd, haddr);
1613 VM_BUG_ON(!pte_none(*pte));
1614 set_pte_at(mm, haddr, pte, entry);
1615 pte_unmap(pte);
1616 }
1617 smp_wmb(); /* make pte visible before pmd */
1618 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001619}
1620
1621static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001622 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001623{
1624 struct mm_struct *mm = vma->vm_mm;
1625 struct page *page;
1626 pgtable_t pgtable;
1627 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001628 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001629 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001630 int i;
1631
1632 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1633 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1634 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001635 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001636
1637 count_vm_event(THP_SPLIT_PMD);
1638
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001639 if (!vma_is_anonymous(vma)) {
1640 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001641 if (vma_is_dax(vma))
1642 return;
1643 page = pmd_page(_pmd);
1644 if (!PageReferenced(page) && pmd_young(_pmd))
1645 SetPageReferenced(page);
1646 page_remove_rmap(page, true);
1647 put_page(page);
1648 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001649 return;
1650 } else if (is_huge_zero_pmd(*pmd)) {
1651 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1652 }
1653
1654 page = pmd_page(*pmd);
1655 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001656 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001657 write = pmd_write(*pmd);
1658 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001659 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001660 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001661
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301662 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001663 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1664 pmd_populate(mm, &_pmd, pgtable);
1665
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001666 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001667 pte_t entry, *pte;
1668 /*
1669 * Note that NUMA hinting access restrictions are not
1670 * transferred to avoid any possibility of altering
1671 * permissions across VMAs.
1672 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001673 if (freeze) {
1674 swp_entry_t swp_entry;
1675 swp_entry = make_migration_entry(page + i, write);
1676 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001677 if (soft_dirty)
1678 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001679 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001680 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001681 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001682 if (!write)
1683 entry = pte_wrprotect(entry);
1684 if (!young)
1685 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001686 if (soft_dirty)
1687 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001688 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001689 if (dirty)
1690 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001691 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001692 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001693 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001694 atomic_inc(&page[i]._mapcount);
1695 pte_unmap(pte);
1696 }
1697
1698 /*
1699 * Set PG_double_map before dropping compound_mapcount to avoid
1700 * false-negative page_mapped().
1701 */
1702 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1703 for (i = 0; i < HPAGE_PMD_NR; i++)
1704 atomic_inc(&page[i]._mapcount);
1705 }
1706
1707 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1708 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001709 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001710 if (TestClearPageDoubleMap(page)) {
1711 /* No need in mapcount reference anymore */
1712 for (i = 0; i < HPAGE_PMD_NR; i++)
1713 atomic_dec(&page[i]._mapcount);
1714 }
1715 }
1716
1717 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001718 /*
1719 * Up to this point the pmd is present and huge and userland has the
1720 * whole access to the hugepage during the split (which happens in
1721 * place). If we overwrite the pmd with the not-huge version pointing
1722 * to the pte here (which of course we could if all CPUs were bug
1723 * free), userland could trigger a small page size TLB miss on the
1724 * small sized TLB while the hugepage TLB entry is still established in
1725 * the huge TLB. Some CPU doesn't like that.
1726 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1727 * 383 on page 93. Intel should be safe but is also warns that it's
1728 * only safe if the permission and cache attributes of the two entries
1729 * loaded in the two TLB is identical (which should be the case here).
1730 * But it is generally safer to never allow small and huge TLB entries
1731 * for the same virtual address to be loaded simultaneously. So instead
1732 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1733 * current pmd notpresent (atomically because here the pmd_trans_huge
1734 * and pmd_trans_splitting must remain set at all times on the pmd
1735 * until the split is complete for this pmd), then we flush the SMP TLB
1736 * and finally we write the non-huge version of the pmd entry with
1737 * pmd_populate.
1738 */
1739 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001740 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001741
1742 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001743 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001744 page_remove_rmap(page + i, false);
1745 put_page(page + i);
1746 }
1747 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001748}
1749
1750void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001751 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001752{
1753 spinlock_t *ptl;
1754 struct mm_struct *mm = vma->vm_mm;
1755 unsigned long haddr = address & HPAGE_PMD_MASK;
1756
1757 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1758 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001759
1760 /*
1761 * If caller asks to setup a migration entries, we need a page to check
1762 * pmd against. Otherwise we can end up replacing wrong page.
1763 */
1764 VM_BUG_ON(freeze && !page);
1765 if (page && page != pmd_page(*pmd))
1766 goto out;
1767
Dan Williams5c7fb562016-01-15 16:56:52 -08001768 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001769 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001770 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07001771 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08001772 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001773 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001774 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001775out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001776 spin_unlock(ptl);
1777 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1778}
1779
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001780void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1781 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001782{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001783 pgd_t *pgd;
1784 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001785 pmd_t *pmd;
1786
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001787 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001788 if (!pgd_present(*pgd))
1789 return;
1790
1791 pud = pud_offset(pgd, address);
1792 if (!pud_present(*pud))
1793 return;
1794
1795 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001796
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001797 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001798}
1799
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07001800void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001801 unsigned long start,
1802 unsigned long end,
1803 long adjust_next)
1804{
1805 /*
1806 * If the new start address isn't hpage aligned and it could
1807 * previously contain an hugepage: check if we need to split
1808 * an huge pmd.
1809 */
1810 if (start & ~HPAGE_PMD_MASK &&
1811 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1812 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001813 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001814
1815 /*
1816 * If the new end address isn't hpage aligned and it could
1817 * previously contain an hugepage: check if we need to split
1818 * an huge pmd.
1819 */
1820 if (end & ~HPAGE_PMD_MASK &&
1821 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1822 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001823 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001824
1825 /*
1826 * If we're also updating the vma->vm_next->vm_start, if the new
1827 * vm_next->vm_start isn't page aligned and it could previously
1828 * contain an hugepage: check if we need to split an huge pmd.
1829 */
1830 if (adjust_next > 0) {
1831 struct vm_area_struct *next = vma->vm_next;
1832 unsigned long nstart = next->vm_start;
1833 nstart += adjust_next << PAGE_SHIFT;
1834 if (nstart & ~HPAGE_PMD_MASK &&
1835 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1836 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001837 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001838 }
1839}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001840
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001841static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001842{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001843 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1844 TTU_RMAP_LOCKED;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001845 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001846
1847 VM_BUG_ON_PAGE(!PageHead(page), page);
1848
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001849 if (PageAnon(page))
1850 ttu_flags |= TTU_MIGRATION;
1851
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001852 /* We only need TTU_SPLIT_HUGE_PMD once */
1853 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1854 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1855 /* Cut short if the page is unmapped */
1856 if (page_count(page) == 1)
1857 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001858
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001859 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001860 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001861 VM_BUG_ON_PAGE(ret, page + i - 1);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001862}
1863
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001864static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001865{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001866 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001867
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001868 for (i = 0; i < HPAGE_PMD_NR; i++)
1869 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001870}
1871
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001872static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001873 struct lruvec *lruvec, struct list_head *list)
1874{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001875 struct page *page_tail = head + tail;
1876
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001877 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001878 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001879
1880 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001881 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001882 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001883 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1884 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001885 * get_page_unless_zero(), and atomic_set() is implemented in C not
1886 * using locked ops. spin_unlock on x86 sometime uses locked ops
1887 * because of PPro errata 66, 92, so unless somebody can guarantee
1888 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001889 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001890 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001891 if (PageAnon(head)) {
1892 page_ref_inc(page_tail);
1893 } else {
1894 /* Additional pin to radix tree */
1895 page_ref_add(page_tail, 2);
1896 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001897
1898 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1899 page_tail->flags |= (head->flags &
1900 ((1L << PG_referenced) |
1901 (1L << PG_swapbacked) |
1902 (1L << PG_mlocked) |
1903 (1L << PG_uptodate) |
1904 (1L << PG_active) |
1905 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001906 (1L << PG_unevictable) |
1907 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001908
1909 /*
1910 * After clearing PageTail the gup refcount can be released.
1911 * Page flags also must be visible before we make the page non-compound.
1912 */
1913 smp_wmb();
1914
1915 clear_compound_head(page_tail);
1916
1917 if (page_is_young(head))
1918 set_page_young(page_tail);
1919 if (page_is_idle(head))
1920 set_page_idle(page_tail);
1921
1922 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001923 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001924 page_tail);
1925 page_tail->mapping = head->mapping;
1926
1927 page_tail->index = head->index + tail;
1928 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1929 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001930}
1931
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001932static void __split_huge_page(struct page *page, struct list_head *list,
1933 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001934{
1935 struct page *head = compound_head(page);
1936 struct zone *zone = page_zone(head);
1937 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001938 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001939 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001940
Mel Gorman599d0c92016-07-28 15:45:31 -07001941 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001942
1943 /* complete memcg works before add pages to LRU */
1944 mem_cgroup_split_huge_fixup(head);
1945
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001946 if (!PageAnon(page))
1947 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1948
1949 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001950 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001951 /* Some pages can be beyond i_size: drop them from page cache */
1952 if (head[i].index >= end) {
1953 __ClearPageDirty(head + i);
1954 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001955 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1956 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001957 put_page(head + i);
1958 }
1959 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001960
1961 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001962 /* See comment in __split_huge_page_tail() */
1963 if (PageAnon(head)) {
1964 page_ref_inc(head);
1965 } else {
1966 /* Additional pin to radix tree */
1967 page_ref_add(head, 2);
1968 spin_unlock(&head->mapping->tree_lock);
1969 }
1970
Mel Gormana52633d2016-07-28 15:45:28 -07001971 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001972
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001973 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001974
1975 for (i = 0; i < HPAGE_PMD_NR; i++) {
1976 struct page *subpage = head + i;
1977 if (subpage == page)
1978 continue;
1979 unlock_page(subpage);
1980
1981 /*
1982 * Subpages may be freed if there wasn't any mapping
1983 * like if add_to_swap() is running on a lru page that
1984 * had its mapping zapped. And freeing these pages
1985 * requires taking the lru_lock so we do the put_page
1986 * of the tail pages after the split is complete.
1987 */
1988 put_page(subpage);
1989 }
1990}
1991
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001992int total_mapcount(struct page *page)
1993{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001994 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001995
1996 VM_BUG_ON_PAGE(PageTail(page), page);
1997
1998 if (likely(!PageCompound(page)))
1999 return atomic_read(&page->_mapcount) + 1;
2000
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002001 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002002 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002003 return compound;
2004 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002005 for (i = 0; i < HPAGE_PMD_NR; i++)
2006 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002007 /* File pages has compound_mapcount included in _mapcount */
2008 if (!PageAnon(page))
2009 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002010 if (PageDoubleMap(page))
2011 ret -= HPAGE_PMD_NR;
2012 return ret;
2013}
2014
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002015/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002016 * This calculates accurately how many mappings a transparent hugepage
2017 * has (unlike page_mapcount() which isn't fully accurate). This full
2018 * accuracy is primarily needed to know if copy-on-write faults can
2019 * reuse the page and change the mapping to read-write instead of
2020 * copying them. At the same time this returns the total_mapcount too.
2021 *
2022 * The function returns the highest mapcount any one of the subpages
2023 * has. If the return value is one, even if different processes are
2024 * mapping different subpages of the transparent hugepage, they can
2025 * all reuse it, because each process is reusing a different subpage.
2026 *
2027 * The total_mapcount is instead counting all virtual mappings of the
2028 * subpages. If the total_mapcount is equal to "one", it tells the
2029 * caller all mappings belong to the same "mm" and in turn the
2030 * anon_vma of the transparent hugepage can become the vma->anon_vma
2031 * local one as no other process may be mapping any of the subpages.
2032 *
2033 * It would be more accurate to replace page_mapcount() with
2034 * page_trans_huge_mapcount(), however we only use
2035 * page_trans_huge_mapcount() in the copy-on-write faults where we
2036 * need full accuracy to avoid breaking page pinning, because
2037 * page_trans_huge_mapcount() is slower than page_mapcount().
2038 */
2039int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2040{
2041 int i, ret, _total_mapcount, mapcount;
2042
2043 /* hugetlbfs shouldn't call it */
2044 VM_BUG_ON_PAGE(PageHuge(page), page);
2045
2046 if (likely(!PageTransCompound(page))) {
2047 mapcount = atomic_read(&page->_mapcount) + 1;
2048 if (total_mapcount)
2049 *total_mapcount = mapcount;
2050 return mapcount;
2051 }
2052
2053 page = compound_head(page);
2054
2055 _total_mapcount = ret = 0;
2056 for (i = 0; i < HPAGE_PMD_NR; i++) {
2057 mapcount = atomic_read(&page[i]._mapcount) + 1;
2058 ret = max(ret, mapcount);
2059 _total_mapcount += mapcount;
2060 }
2061 if (PageDoubleMap(page)) {
2062 ret -= 1;
2063 _total_mapcount -= HPAGE_PMD_NR;
2064 }
2065 mapcount = compound_mapcount(page);
2066 ret += mapcount;
2067 _total_mapcount += mapcount;
2068 if (total_mapcount)
2069 *total_mapcount = _total_mapcount;
2070 return ret;
2071}
2072
2073/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002074 * This function splits huge page into normal pages. @page can point to any
2075 * subpage of huge page to split. Split doesn't change the position of @page.
2076 *
2077 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2078 * The huge page must be locked.
2079 *
2080 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2081 *
2082 * Both head page and tail pages will inherit mapping, flags, and so on from
2083 * the hugepage.
2084 *
2085 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2086 * they are not mapped.
2087 *
2088 * Returns 0 if the hugepage is split successfully.
2089 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2090 * us.
2091 */
2092int split_huge_page_to_list(struct page *page, struct list_head *list)
2093{
2094 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002095 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002096 struct anon_vma *anon_vma = NULL;
2097 struct address_space *mapping = NULL;
2098 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002099 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002100 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002101
2102 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002103 VM_BUG_ON_PAGE(!PageLocked(page), page);
2104 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2105 VM_BUG_ON_PAGE(!PageCompound(page), page);
2106
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002107 if (PageAnon(head)) {
2108 /*
2109 * The caller does not necessarily hold an mmap_sem that would
2110 * prevent the anon_vma disappearing so we first we take a
2111 * reference to it and then lock the anon_vma for write. This
2112 * is similar to page_lock_anon_vma_read except the write lock
2113 * is taken to serialise against parallel split or collapse
2114 * operations.
2115 */
2116 anon_vma = page_get_anon_vma(head);
2117 if (!anon_vma) {
2118 ret = -EBUSY;
2119 goto out;
2120 }
2121 extra_pins = 0;
2122 mapping = NULL;
2123 anon_vma_lock_write(anon_vma);
2124 } else {
2125 mapping = head->mapping;
2126
2127 /* Truncated ? */
2128 if (!mapping) {
2129 ret = -EBUSY;
2130 goto out;
2131 }
2132
2133 /* Addidional pins from radix tree */
2134 extra_pins = HPAGE_PMD_NR;
2135 anon_vma = NULL;
2136 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002137 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002138
2139 /*
2140 * Racy check if we can split the page, before freeze_page() will
2141 * split PMDs
2142 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002143 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002144 ret = -EBUSY;
2145 goto out_unlock;
2146 }
2147
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002148 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002149 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002150 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2151
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002152 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2153 if (mlocked)
2154 lru_add_drain();
2155
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002156 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002157 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002158
2159 if (mapping) {
2160 void **pslot;
2161
2162 spin_lock(&mapping->tree_lock);
2163 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2164 page_index(head));
2165 /*
2166 * Check if the head page is present in radix tree.
2167 * We assume all tail are present too, if head is there.
2168 */
2169 if (radix_tree_deref_slot_protected(pslot,
2170 &mapping->tree_lock) != head)
2171 goto fail;
2172 }
2173
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002174 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002175 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002176 count = page_count(head);
2177 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002178 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002179 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002180 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002181 list_del(page_deferred_list(head));
2182 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002183 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002184 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002185 spin_unlock(&pgdata->split_queue_lock);
2186 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002187 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002188 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002189 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2190 pr_alert("total_mapcount: %u, page_count(): %u\n",
2191 mapcount, count);
2192 if (PageTail(page))
2193 dump_page(head, NULL);
2194 dump_page(page, "total_mapcount(head) > 0");
2195 BUG();
2196 }
2197 spin_unlock(&pgdata->split_queue_lock);
2198fail: if (mapping)
2199 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002200 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002201 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002202 ret = -EBUSY;
2203 }
2204
2205out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002206 if (anon_vma) {
2207 anon_vma_unlock_write(anon_vma);
2208 put_anon_vma(anon_vma);
2209 }
2210 if (mapping)
2211 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002212out:
2213 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2214 return ret;
2215}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002216
2217void free_transhuge_page(struct page *page)
2218{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002219 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002220 unsigned long flags;
2221
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002222 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002223 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002224 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002225 list_del(page_deferred_list(page));
2226 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002227 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002228 free_compound_page(page);
2229}
2230
2231void deferred_split_huge_page(struct page *page)
2232{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002233 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002234 unsigned long flags;
2235
2236 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2237
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002238 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002239 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002240 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002241 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2242 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002243 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002244 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002245}
2246
2247static unsigned long deferred_split_count(struct shrinker *shrink,
2248 struct shrink_control *sc)
2249{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002250 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002251 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002252}
2253
2254static unsigned long deferred_split_scan(struct shrinker *shrink,
2255 struct shrink_control *sc)
2256{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002257 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002258 unsigned long flags;
2259 LIST_HEAD(list), *pos, *next;
2260 struct page *page;
2261 int split = 0;
2262
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002263 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002264 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002265 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002266 page = list_entry((void *)pos, struct page, mapping);
2267 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002268 if (get_page_unless_zero(page)) {
2269 list_move(page_deferred_list(page), &list);
2270 } else {
2271 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002272 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002273 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002274 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002275 if (!--sc->nr_to_scan)
2276 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002277 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002278 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002279
2280 list_for_each_safe(pos, next, &list) {
2281 page = list_entry((void *)pos, struct page, mapping);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002282 if (!trylock_page(page))
2283 goto next;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002284 /* split_huge_page() removes page from list on success */
2285 if (!split_huge_page(page))
2286 split++;
2287 unlock_page(page);
Kirill A. Shutemov142d9dd2018-03-22 16:17:31 -07002288next:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002289 put_page(page);
2290 }
2291
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002292 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2293 list_splice_tail(&list, &pgdata->split_queue);
2294 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002295
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002296 /*
2297 * Stop shrinker if we didn't split any page, but the queue is empty.
2298 * This can happen if pages were freed under us.
2299 */
2300 if (!split && list_empty(&pgdata->split_queue))
2301 return SHRINK_STOP;
2302 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002303}
2304
2305static struct shrinker deferred_split_shrinker = {
2306 .count_objects = deferred_split_count,
2307 .scan_objects = deferred_split_scan,
2308 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002309 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002310};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002311
2312#ifdef CONFIG_DEBUG_FS
2313static int split_huge_pages_set(void *data, u64 val)
2314{
2315 struct zone *zone;
2316 struct page *page;
2317 unsigned long pfn, max_zone_pfn;
2318 unsigned long total = 0, split = 0;
2319
2320 if (val != 1)
2321 return -EINVAL;
2322
2323 for_each_populated_zone(zone) {
2324 max_zone_pfn = zone_end_pfn(zone);
2325 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2326 if (!pfn_valid(pfn))
2327 continue;
2328
2329 page = pfn_to_page(pfn);
2330 if (!get_page_unless_zero(page))
2331 continue;
2332
2333 if (zone != page_zone(page))
2334 goto next;
2335
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002336 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002337 goto next;
2338
2339 total++;
2340 lock_page(page);
2341 if (!split_huge_page(page))
2342 split++;
2343 unlock_page(page);
2344next:
2345 put_page(page);
2346 }
2347 }
2348
Yang Shi145bdaa2016-05-05 16:22:00 -07002349 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002350
2351 return 0;
2352}
2353DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2354 "%llu\n");
2355
2356static int __init split_huge_pages_debugfs(void)
2357{
2358 void *ret;
2359
Yang Shi145bdaa2016-05-05 16:22:00 -07002360 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002361 &split_huge_pages_fops);
2362 if (!ret)
2363 pr_warn("Failed to create split_huge_pages in debugfs");
2364 return 0;
2365}
2366late_initcall(split_huge_pages_debugfs);
2367#endif