blob: d36b2af4d1bf4b6621974823f36c52dda405f181 [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>
Ingo Molnarf7ccbae2017-02-08 18:51:30 +010012#include <linux/sched/coredump.h>
Ingo Molnar6a3827d2017-02-08 18:51:31 +010013#include <linux/sched/numa_balancing.h>
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080014#include <linux/highmem.h>
15#include <linux/hugetlb.h>
16#include <linux/mmu_notifier.h>
17#include <linux/rmap.h>
18#include <linux/swap.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080019#include <linux/shrinker.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080020#include <linux/mm_inline.h>
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -080021#include <linux/swapops.h>
Matthew Wilcox4897c762015-09-08 14:58:45 -070022#include <linux/dax.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080023#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080024#include <linux/freezer.h>
Dan Williamsf25748e32016-01-15 16:56:43 -080025#include <linux/pfn_t.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080026#include <linux/mman.h>
Dan Williams3565fce2016-01-15 16:56:55 -080027#include <linux/memremap.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020028#include <linux/pagemap.h>
Kirill A. Shutemov49071d42016-01-15 16:54:40 -080029#include <linux/debugfs.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000030#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080031#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070032#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070033#include <linux/page_idle.h>
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -070034#include <linux/shmem_fs.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080035
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080036#include <asm/tlb.h>
37#include <asm/pgalloc.h>
38#include "internal.h"
39
Andrea Arcangeliba761492011-01-13 15:46:58 -080040/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080041 * By default transparent hugepage support is disabled in order that avoid
42 * to risk increase the memory footprint of applications without a guaranteed
43 * benefit. When transparent hugepage support is enabled, is for all mappings,
44 * and khugepaged scans all mappings.
45 * Defrag is invoked by khugepaged hugepage allocations and by page faults
46 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080047 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080048unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080049#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080050 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080051#endif
52#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
53 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
54#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070055 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080056 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080058
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080059static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080060
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080061static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080062struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080063
Aaron Lu6fcb52a2016-10-07 17:00:08 -070064static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080065{
66 struct page *zero_page;
67retry:
68 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070069 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080070
71 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
72 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080073 if (!zero_page) {
74 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070075 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080076 }
77 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080078 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070079 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080080 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070081 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080082 goto retry;
83 }
84
85 /* We take additional reference here. It will be put back by shrinker */
86 atomic_set(&huge_zero_refcount, 2);
87 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070088 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080089}
90
Aaron Lu6fcb52a2016-10-07 17:00:08 -070091static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080092{
93 /*
94 * Counter should never go to zero here. Only shrinker can put
95 * last reference.
96 */
97 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
98}
99
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700100struct page *mm_get_huge_zero_page(struct mm_struct *mm)
101{
102 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
103 return READ_ONCE(huge_zero_page);
104
105 if (!get_huge_zero_page())
106 return NULL;
107
108 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
109 put_huge_zero_page();
110
111 return READ_ONCE(huge_zero_page);
112}
113
114void mm_put_huge_zero_page(struct mm_struct *mm)
115{
116 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
117 put_huge_zero_page();
118}
119
Glauber Costa48896462013-08-28 10:18:15 +1000120static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
121 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800122{
Glauber Costa48896462013-08-28 10:18:15 +1000123 /* we can free zero page only if last reference remains */
124 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
125}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800126
Glauber Costa48896462013-08-28 10:18:15 +1000127static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
128 struct shrink_control *sc)
129{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800130 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700131 struct page *zero_page = xchg(&huge_zero_page, NULL);
132 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700133 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000134 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800135 }
136
137 return 0;
138}
139
140static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000141 .count_objects = shrink_huge_zero_page_count,
142 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800143 .seeks = DEFAULT_SEEKS,
144};
145
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800146#ifdef CONFIG_SYSFS
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800147static ssize_t enabled_show(struct kobject *kobj,
148 struct kobj_attribute *attr, char *buf)
149{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700150 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
151 return sprintf(buf, "[always] madvise never\n");
152 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
153 return sprintf(buf, "always [madvise] never\n");
154 else
155 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800156}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700157
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800158static ssize_t enabled_store(struct kobject *kobj,
159 struct kobj_attribute *attr,
160 const char *buf, size_t count)
161{
David Rientjes21440d72017-02-22 15:45:49 -0800162 ssize_t ret = count;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800163
David Rientjes21440d72017-02-22 15:45:49 -0800164 if (!memcmp("always", buf,
165 min(sizeof("always")-1, count))) {
166 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
167 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
168 } else if (!memcmp("madvise", buf,
169 min(sizeof("madvise")-1, count))) {
170 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
171 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
172 } else if (!memcmp("never", buf,
173 min(sizeof("never")-1, count))) {
174 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
175 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
176 } else
177 ret = -EINVAL;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800178
179 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700180 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800181 if (err)
182 ret = err;
183 }
Andrea Arcangeliba761492011-01-13 15:46:58 -0800184 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800185}
186static struct kobj_attribute enabled_attr =
187 __ATTR(enabled, 0644, enabled_show, enabled_store);
188
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700189ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800190 struct kobj_attribute *attr, char *buf,
191 enum transparent_hugepage_flag flag)
192{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700193 return sprintf(buf, "%d\n",
194 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800195}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700196
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700197ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800198 struct kobj_attribute *attr,
199 const char *buf, size_t count,
200 enum transparent_hugepage_flag flag)
201{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700202 unsigned long value;
203 int ret;
204
205 ret = kstrtoul(buf, 10, &value);
206 if (ret < 0)
207 return ret;
208 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800209 return -EINVAL;
210
Ben Hutchingse27e6152011-04-14 15:22:21 -0700211 if (value)
212 set_bit(flag, &transparent_hugepage_flags);
213 else
214 clear_bit(flag, &transparent_hugepage_flags);
215
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800216 return count;
217}
218
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800219static ssize_t defrag_show(struct kobject *kobj,
220 struct kobj_attribute *attr, char *buf)
221{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700222 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800223 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
Mel Gorman444eb2a42016-03-17 14:19:23 -0700224 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800225 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
226 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
227 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
228 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
229 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
230 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800231}
David Rientjes21440d72017-02-22 15:45:49 -0800232
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800233static ssize_t defrag_store(struct kobject *kobj,
234 struct kobj_attribute *attr,
235 const char *buf, size_t count)
236{
David Rientjes21440d72017-02-22 15:45:49 -0800237 if (!memcmp("always", buf,
238 min(sizeof("always")-1, count))) {
239 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
242 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
243 } else if (!memcmp("defer", buf,
244 min(sizeof("defer")-1, count))) {
245 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
248 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
249 } else if (!memcmp("defer+madvise", buf,
250 min(sizeof("defer+madvise")-1, count))) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
255 } else if (!memcmp("madvise", buf,
256 min(sizeof("madvise")-1, count))) {
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
260 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
261 } else if (!memcmp("never", buf,
262 min(sizeof("never")-1, count))) {
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
267 } else
268 return -EINVAL;
269
270 return count;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800271}
272static struct kobj_attribute defrag_attr =
273 __ATTR(defrag, 0644, defrag_show, defrag_store);
274
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800275static ssize_t use_zero_page_show(struct kobject *kobj,
276 struct kobj_attribute *attr, char *buf)
277{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700278 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800279 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
280}
281static ssize_t use_zero_page_store(struct kobject *kobj,
282 struct kobj_attribute *attr, const char *buf, size_t count)
283{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700284 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800285 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
286}
287static struct kobj_attribute use_zero_page_attr =
288 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Hugh Dickins49920d22016-12-12 16:44:50 -0800289
290static ssize_t hpage_pmd_size_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
292{
293 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
294}
295static struct kobj_attribute hpage_pmd_size_attr =
296 __ATTR_RO(hpage_pmd_size);
297
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800298#ifdef CONFIG_DEBUG_VM
299static ssize_t debug_cow_show(struct kobject *kobj,
300 struct kobj_attribute *attr, char *buf)
301{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700302 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800303 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
304}
305static ssize_t debug_cow_store(struct kobject *kobj,
306 struct kobj_attribute *attr,
307 const char *buf, size_t count)
308{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700309 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800310 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
311}
312static struct kobj_attribute debug_cow_attr =
313 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
314#endif /* CONFIG_DEBUG_VM */
315
316static struct attribute *hugepage_attr[] = {
317 &enabled_attr.attr,
318 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800319 &use_zero_page_attr.attr,
Hugh Dickins49920d22016-12-12 16:44:50 -0800320 &hpage_pmd_size_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700321#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700322 &shmem_enabled_attr.attr,
323#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800324#ifdef CONFIG_DEBUG_VM
325 &debug_cow_attr.attr,
326#endif
327 NULL,
328};
329
330static struct attribute_group hugepage_attr_group = {
331 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800332};
333
Shaohua Li569e5592012-01-12 17:19:11 -0800334static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
335{
336 int err;
337
338 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
339 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700340 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800341 return -ENOMEM;
342 }
343
344 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
345 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700346 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800347 goto delete_obj;
348 }
349
350 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
351 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700352 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800353 goto remove_hp_group;
354 }
355
356 return 0;
357
358remove_hp_group:
359 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
360delete_obj:
361 kobject_put(*hugepage_kobj);
362 return err;
363}
364
365static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
366{
367 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
368 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
369 kobject_put(hugepage_kobj);
370}
371#else
372static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
373{
374 return 0;
375}
376
377static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
378{
379}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800380#endif /* CONFIG_SYSFS */
381
382static int __init hugepage_init(void)
383{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800384 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800385 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800386
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800387 if (!has_transparent_hugepage()) {
388 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800389 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800390 }
391
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530392 /*
393 * hugepages can't be allocated by the buddy allocator
394 */
395 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
396 /*
397 * we use page->mapping and page->index in second tail page
398 * as list_head: assuming THP order >= 2
399 */
400 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
401
Shaohua Li569e5592012-01-12 17:19:11 -0800402 err = hugepage_init_sysfs(&hugepage_kobj);
403 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700404 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800405
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700406 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800407 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700408 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800409
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700410 err = register_shrinker(&huge_zero_page_shrinker);
411 if (err)
412 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800413 err = register_shrinker(&deferred_split_shrinker);
414 if (err)
415 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800416
Rik van Riel97562cd2011-01-13 15:47:12 -0800417 /*
418 * By default disable transparent hugepages on smaller systems,
419 * where the extra memory used could hurt more than TLB overhead
420 * is likely to save. The admin can still enable it through /sys.
421 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700422 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800423 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700424 return 0;
425 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800426
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700427 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700428 if (err)
429 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800430
Shaohua Li569e5592012-01-12 17:19:11 -0800431 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700432err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800433 unregister_shrinker(&deferred_split_shrinker);
434err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700435 unregister_shrinker(&huge_zero_page_shrinker);
436err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700437 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700438err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800439 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700440err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800441 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800442}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800443subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800444
445static int __init setup_transparent_hugepage(char *str)
446{
447 int ret = 0;
448 if (!str)
449 goto out;
450 if (!strcmp(str, "always")) {
451 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
452 &transparent_hugepage_flags);
453 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
454 &transparent_hugepage_flags);
455 ret = 1;
456 } else if (!strcmp(str, "madvise")) {
457 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
458 &transparent_hugepage_flags);
459 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
460 &transparent_hugepage_flags);
461 ret = 1;
462 } else if (!strcmp(str, "never")) {
463 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
464 &transparent_hugepage_flags);
465 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
466 &transparent_hugepage_flags);
467 ret = 1;
468 }
469out:
470 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700471 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800472 return ret;
473}
474__setup("transparent_hugepage=", setup_transparent_hugepage);
475
Mel Gormanb32967f2012-11-19 12:35:47 +0000476pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800477{
478 if (likely(vma->vm_flags & VM_WRITE))
479 pmd = pmd_mkwrite(pmd);
480 return pmd;
481}
482
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800483static inline struct list_head *page_deferred_list(struct page *page)
484{
485 /*
486 * ->lru in the tail pages is occupied by compound_head.
487 * Let's use ->mapping + ->index in the second tail page as list_head.
488 */
489 return (struct list_head *)&page[2].mapping;
490}
491
492void prep_transhuge_page(struct page *page)
493{
494 /*
495 * we use page->mapping and page->indexlru in second tail page
496 * as list_head: assuming THP order >= 2
497 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800498
499 INIT_LIST_HEAD(page_deferred_list(page));
500 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
501}
502
Toshi Kani74d2fad2016-10-07 16:59:56 -0700503unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
504 loff_t off, unsigned long flags, unsigned long size)
505{
506 unsigned long addr;
507 loff_t off_end = off + len;
508 loff_t off_align = round_up(off, size);
509 unsigned long len_pad;
510
511 if (off_end <= off_align || (off_end - off_align) < size)
512 return 0;
513
514 len_pad = len + size;
515 if (len_pad < len || (off + len_pad) < off)
516 return 0;
517
518 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
519 off >> PAGE_SHIFT, flags);
520 if (IS_ERR_VALUE(addr))
521 return 0;
522
523 addr += (off - addr) & (size - 1);
524 return addr;
525}
526
527unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
528 unsigned long len, unsigned long pgoff, unsigned long flags)
529{
530 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
531
532 if (addr)
533 goto out;
534 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
535 goto out;
536
537 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
538 if (addr)
539 return addr;
540
541 out:
542 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
543}
544EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
545
Jan Kara82b0f8c2016-12-14 15:06:58 -0800546static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700547 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800548{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800549 struct vm_area_struct *vma = vmf->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700550 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800551 pgtable_t pgtable;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800552 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800553
Sasha Levin309381fea2014-01-23 15:52:54 -0800554 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700555
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700556 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700557 put_page(page);
558 count_vm_event(THP_FAULT_FALLBACK);
559 return VM_FAULT_FALLBACK;
560 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800561
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700562 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700563 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800564 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700565 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700566 return VM_FAULT_OOM;
567 }
568
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800569 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700570 /*
571 * The memory barrier inside __SetPageUptodate makes sure that
572 * clear_huge_page writes become visible before the set_pmd_at()
573 * write.
574 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800575 __SetPageUptodate(page);
576
Jan Kara82b0f8c2016-12-14 15:06:58 -0800577 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
578 if (unlikely(!pmd_none(*vmf->pmd))) {
579 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800580 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800581 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700582 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800583 } else {
584 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700585
586 /* Deliver the page fault to userland */
587 if (userfaultfd_missing(vma)) {
588 int ret;
589
Jan Kara82b0f8c2016-12-14 15:06:58 -0800590 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800591 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700592 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700593 pte_free(vma->vm_mm, pgtable);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800594 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700595 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
596 return ret;
597 }
598
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700599 entry = mk_huge_pmd(page, vma->vm_page_prot);
600 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800601 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800602 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700603 lru_cache_add_active_or_unevictable(page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800604 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
605 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700606 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
607 atomic_long_inc(&vma->vm_mm->nr_ptes);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800608 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700609 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800610 }
611
David Rientjesaa2e8782012-05-29 15:06:17 -0700612 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800613}
614
Mel Gorman444eb2a42016-03-17 14:19:23 -0700615/*
David Rientjes21440d72017-02-22 15:45:49 -0800616 * always: directly stall for all thp allocations
617 * defer: wake kswapd and fail if not immediately available
618 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
619 * fail if not immediately available
620 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
621 * available
622 * never: never stall for any thp allocation
Mel Gorman444eb2a42016-03-17 14:19:23 -0700623 */
624static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800625{
David Rientjes21440d72017-02-22 15:45:49 -0800626 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700627
David Rientjes21440d72017-02-22 15:45:49 -0800628 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
Vlastimil Babka25160352016-07-28 15:49:25 -0700629 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
David Rientjes21440d72017-02-22 15:45:49 -0800630 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
631 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
632 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
633 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
634 __GFP_KSWAPD_RECLAIM);
635 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
636 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
637 0);
Vlastimil Babka25160352016-07-28 15:49:25 -0700638 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700639}
640
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800641/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700642static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800643 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700644 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800645{
646 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700647 if (!pmd_none(*pmd))
648 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700649 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800650 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800651 if (pgtable)
652 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800653 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800654 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700655 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800656}
657
Jan Kara82b0f8c2016-12-14 15:06:58 -0800658int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800659{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800660 struct vm_area_struct *vma = vmf->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800661 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800662 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800663 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800664
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700665 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700666 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700667 if (unlikely(anon_vma_prepare(vma)))
668 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700669 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700670 return VM_FAULT_OOM;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800671 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700672 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700673 transparent_hugepage_use_zero_page()) {
674 pgtable_t pgtable;
675 struct page *zero_page;
676 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700677 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700678 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700679 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800680 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700681 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700682 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700683 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700684 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700685 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700686 }
Jan Kara82b0f8c2016-12-14 15:06:58 -0800687 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700688 ret = 0;
689 set = false;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800690 if (pmd_none(*vmf->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700691 if (userfaultfd_missing(vma)) {
Jan Kara82b0f8c2016-12-14 15:06:58 -0800692 spin_unlock(vmf->ptl);
693 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700694 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
695 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700696 set_huge_zero_page(pgtable, vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -0800697 haddr, vmf->pmd, zero_page);
698 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700699 set = true;
700 }
701 } else
Jan Kara82b0f8c2016-12-14 15:06:58 -0800702 spin_unlock(vmf->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700703 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700704 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700705 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800706 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700707 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800708 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700709 if (unlikely(!page)) {
710 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700711 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700712 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800713 prep_transhuge_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800714 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800715}
716
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700717static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800718 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700719{
720 struct mm_struct *mm = vma->vm_mm;
721 pmd_t entry;
722 spinlock_t *ptl;
723
724 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800725 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
726 if (pfn_t_devmap(pfn))
727 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800728 if (write) {
729 entry = pmd_mkyoung(pmd_mkdirty(entry));
730 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700731 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800732 set_pmd_at(mm, addr, pmd, entry);
733 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700734 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700735}
736
737int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800738 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700739{
740 pgprot_t pgprot = vma->vm_page_prot;
741 /*
742 * If we had pmd_special, we could avoid all these restrictions,
743 * but we need to be consistent with PTEs and architectures that
744 * can't support a 'special' bit.
745 */
746 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
747 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
748 (VM_PFNMAP|VM_MIXEDMAP));
749 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800750 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700751
752 if (addr < vma->vm_start || addr >= vma->vm_end)
753 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200754
755 track_pfn_insert(vma, &pgprot, pfn);
756
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700757 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
758 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700759}
Dan Williamsdee41072016-05-14 12:20:44 -0700760EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700761
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800762#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
763static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
764{
765 if (likely(vma->vm_flags & VM_WRITE))
766 pud = pud_mkwrite(pud);
767 return pud;
768}
769
770static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
771 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
772{
773 struct mm_struct *mm = vma->vm_mm;
774 pud_t entry;
775 spinlock_t *ptl;
776
777 ptl = pud_lock(mm, pud);
778 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
779 if (pfn_t_devmap(pfn))
780 entry = pud_mkdevmap(entry);
781 if (write) {
782 entry = pud_mkyoung(pud_mkdirty(entry));
783 entry = maybe_pud_mkwrite(entry, vma);
784 }
785 set_pud_at(mm, addr, pud, entry);
786 update_mmu_cache_pud(vma, addr, pud);
787 spin_unlock(ptl);
788}
789
790int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
791 pud_t *pud, pfn_t pfn, bool write)
792{
793 pgprot_t pgprot = vma->vm_page_prot;
794 /*
795 * If we had pud_special, we could avoid all these restrictions,
796 * but we need to be consistent with PTEs and architectures that
797 * can't support a 'special' bit.
798 */
799 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
800 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
801 (VM_PFNMAP|VM_MIXEDMAP));
802 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
803 BUG_ON(!pfn_t_devmap(pfn));
804
805 if (addr < vma->vm_start || addr >= vma->vm_end)
806 return VM_FAULT_SIGBUS;
807
808 track_pfn_insert(vma, &pgprot, pfn);
809
810 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
811 return VM_FAULT_NOPAGE;
812}
813EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
814#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
815
Dan Williams3565fce2016-01-15 16:56:55 -0800816static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
817 pmd_t *pmd)
818{
819 pmd_t _pmd;
820
821 /*
822 * We should set the dirty bit only for FOLL_WRITE but for now
823 * the dirty bit in the pmd is meaningless. And if the dirty
824 * bit will become meaningful and we'll only set it with
825 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
826 * set the young bit, instead of the current set_pmd_at.
827 */
828 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
829 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
830 pmd, _pmd, 1))
831 update_mmu_cache_pmd(vma, addr, pmd);
832}
833
834struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
835 pmd_t *pmd, int flags)
836{
837 unsigned long pfn = pmd_pfn(*pmd);
838 struct mm_struct *mm = vma->vm_mm;
839 struct dev_pagemap *pgmap;
840 struct page *page;
841
842 assert_spin_locked(pmd_lockptr(mm, pmd));
843
Keno Fischer8310d482017-01-24 15:17:48 -0800844 /*
845 * When we COW a devmap PMD entry, we split it into PTEs, so we should
846 * not be in this function with `flags & FOLL_COW` set.
847 */
848 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
849
Dan Williams3565fce2016-01-15 16:56:55 -0800850 if (flags & FOLL_WRITE && !pmd_write(*pmd))
851 return NULL;
852
853 if (pmd_present(*pmd) && pmd_devmap(*pmd))
854 /* pass */;
855 else
856 return NULL;
857
858 if (flags & FOLL_TOUCH)
859 touch_pmd(vma, addr, pmd);
860
861 /*
862 * device mapped pages can only be returned if the
863 * caller will manage the page reference count.
864 */
865 if (!(flags & FOLL_GET))
866 return ERR_PTR(-EEXIST);
867
868 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
869 pgmap = get_dev_pagemap(pfn, NULL);
870 if (!pgmap)
871 return ERR_PTR(-EFAULT);
872 page = pfn_to_page(pfn);
873 get_page(page);
874 put_dev_pagemap(pgmap);
875
876 return page;
877}
878
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800879int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
880 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
881 struct vm_area_struct *vma)
882{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800883 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800884 struct page *src_page;
885 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800886 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700887 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800888
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700889 /* Skip if can be re-fill on fault */
890 if (!vma_is_anonymous(vma))
891 return 0;
892
893 pgtable = pte_alloc_one(dst_mm, addr);
894 if (unlikely(!pgtable))
895 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800897 dst_ptl = pmd_lock(dst_mm, dst_pmd);
898 src_ptl = pmd_lockptr(src_mm, src_pmd);
899 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800900
901 ret = -EAGAIN;
902 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700903 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800904 pte_free(dst_mm, pgtable);
905 goto out_unlock;
906 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800907 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800908 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800909 * under splitting since we don't split the page itself, only pmd to
910 * a page table.
911 */
912 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700913 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800914 /*
915 * get_huge_zero_page() will never allocate a new page here,
916 * since we already have a zero page to copy. It just takes a
917 * reference.
918 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700919 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700920 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700921 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800922 ret = 0;
923 goto out_unlock;
924 }
Mel Gormande466bd2013-12-18 17:08:42 -0800925
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700926 src_page = pmd_page(pmd);
927 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
928 get_page(src_page);
929 page_dup_rmap(src_page, true);
930 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
931 atomic_long_inc(&dst_mm->nr_ptes);
932 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800933
934 pmdp_set_wrprotect(src_mm, addr, src_pmd);
935 pmd = pmd_mkold(pmd_wrprotect(pmd));
936 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800937
938 ret = 0;
939out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800940 spin_unlock(src_ptl);
941 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800942out:
943 return ret;
944}
945
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800946#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
947static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
948 pud_t *pud)
949{
950 pud_t _pud;
951
952 /*
953 * We should set the dirty bit only for FOLL_WRITE but for now
954 * the dirty bit in the pud is meaningless. And if the dirty
955 * bit will become meaningful and we'll only set it with
956 * FOLL_WRITE, an atomic set_bit will be required on the pud to
957 * set the young bit, instead of the current set_pud_at.
958 */
959 _pud = pud_mkyoung(pud_mkdirty(*pud));
960 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
961 pud, _pud, 1))
962 update_mmu_cache_pud(vma, addr, pud);
963}
964
965struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
966 pud_t *pud, int flags)
967{
968 unsigned long pfn = pud_pfn(*pud);
969 struct mm_struct *mm = vma->vm_mm;
970 struct dev_pagemap *pgmap;
971 struct page *page;
972
973 assert_spin_locked(pud_lockptr(mm, pud));
974
975 if (flags & FOLL_WRITE && !pud_write(*pud))
976 return NULL;
977
978 if (pud_present(*pud) && pud_devmap(*pud))
979 /* pass */;
980 else
981 return NULL;
982
983 if (flags & FOLL_TOUCH)
984 touch_pud(vma, addr, pud);
985
986 /*
987 * device mapped pages can only be returned if the
988 * caller will manage the page reference count.
989 */
990 if (!(flags & FOLL_GET))
991 return ERR_PTR(-EEXIST);
992
993 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
994 pgmap = get_dev_pagemap(pfn, NULL);
995 if (!pgmap)
996 return ERR_PTR(-EFAULT);
997 page = pfn_to_page(pfn);
998 get_page(page);
999 put_dev_pagemap(pgmap);
1000
1001 return page;
1002}
1003
1004int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1005 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1006 struct vm_area_struct *vma)
1007{
1008 spinlock_t *dst_ptl, *src_ptl;
1009 pud_t pud;
1010 int ret;
1011
1012 dst_ptl = pud_lock(dst_mm, dst_pud);
1013 src_ptl = pud_lockptr(src_mm, src_pud);
1014 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1015
1016 ret = -EAGAIN;
1017 pud = *src_pud;
1018 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1019 goto out_unlock;
1020
1021 /*
1022 * When page table lock is held, the huge zero pud should not be
1023 * under splitting since we don't split the page itself, only pud to
1024 * a page table.
1025 */
1026 if (is_huge_zero_pud(pud)) {
1027 /* No huge zero pud yet */
1028 }
1029
1030 pudp_set_wrprotect(src_mm, addr, src_pud);
1031 pud = pud_mkold(pud_wrprotect(pud));
1032 set_pud_at(dst_mm, addr, dst_pud, pud);
1033
1034 ret = 0;
1035out_unlock:
1036 spin_unlock(src_ptl);
1037 spin_unlock(dst_ptl);
1038 return ret;
1039}
1040
1041void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1042{
1043 pud_t entry;
1044 unsigned long haddr;
1045 bool write = vmf->flags & FAULT_FLAG_WRITE;
1046
1047 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1048 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1049 goto unlock;
1050
1051 entry = pud_mkyoung(orig_pud);
1052 if (write)
1053 entry = pud_mkdirty(entry);
1054 haddr = vmf->address & HPAGE_PUD_MASK;
1055 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1056 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1057
1058unlock:
1059 spin_unlock(vmf->ptl);
1060}
1061#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1062
Jan Kara82b0f8c2016-12-14 15:06:58 -08001063void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001064{
1065 pmd_t entry;
1066 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001067 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001068
Jan Kara82b0f8c2016-12-14 15:06:58 -08001069 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1070 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001071 goto unlock;
1072
1073 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001074 if (write)
1075 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001076 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001077 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001078 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001079
1080unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001081 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001082}
1083
Jan Kara82b0f8c2016-12-14 15:06:58 -08001084static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001085 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001086{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001087 struct vm_area_struct *vma = vmf->vma;
1088 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001089 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001090 pgtable_t pgtable;
1091 pmd_t _pmd;
1092 int ret = 0, i;
1093 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001094 unsigned long mmun_start; /* For mmu_notifiers */
1095 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001096
1097 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1098 GFP_KERNEL);
1099 if (unlikely(!pages)) {
1100 ret |= VM_FAULT_OOM;
1101 goto out;
1102 }
1103
1104 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001105 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001106 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001107 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001108 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1109 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001110 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001111 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001112 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001113 memcg = (void *)page_private(pages[i]);
1114 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001115 mem_cgroup_cancel_charge(pages[i], memcg,
1116 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001117 put_page(pages[i]);
1118 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001119 kfree(pages);
1120 ret |= VM_FAULT_OOM;
1121 goto out;
1122 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001123 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001124 }
1125
1126 for (i = 0; i < HPAGE_PMD_NR; i++) {
1127 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001128 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001129 __SetPageUptodate(pages[i]);
1130 cond_resched();
1131 }
1132
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001133 mmun_start = haddr;
1134 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001135 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001136
Jan Kara82b0f8c2016-12-14 15:06:58 -08001137 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1138 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001139 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001140 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001141
Jan Kara82b0f8c2016-12-14 15:06:58 -08001142 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001143 /* leave pmd empty until pte is filled */
1144
Jan Kara82b0f8c2016-12-14 15:06:58 -08001145 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001146 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001147
1148 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001149 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001150 entry = mk_pte(pages[i], vma->vm_page_prot);
1151 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001152 memcg = (void *)page_private(pages[i]);
1153 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001154 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001155 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001156 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001157 vmf->pte = pte_offset_map(&_pmd, haddr);
1158 VM_BUG_ON(!pte_none(*vmf->pte));
1159 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1160 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001161 }
1162 kfree(pages);
1163
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001165 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001166 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001167 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001168
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001169 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001170
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001171 ret |= VM_FAULT_WRITE;
1172 put_page(page);
1173
1174out:
1175 return ret;
1176
1177out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001178 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001179 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001180 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001181 memcg = (void *)page_private(pages[i]);
1182 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001183 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001184 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001185 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001186 kfree(pages);
1187 goto out;
1188}
1189
Jan Kara82b0f8c2016-12-14 15:06:58 -08001190int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001192 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001193 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001194 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001195 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001196 unsigned long mmun_start; /* For mmu_notifiers */
1197 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001198 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001199 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001200
Jan Kara82b0f8c2016-12-14 15:06:58 -08001201 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001202 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001203 if (is_huge_zero_pmd(orig_pmd))
1204 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001205 spin_lock(vmf->ptl);
1206 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001207 goto out_unlock;
1208
1209 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001210 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001211 /*
1212 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001213 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001214 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001215 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001216 pmd_t entry;
1217 entry = pmd_mkyoung(orig_pmd);
1218 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001219 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1220 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001221 ret |= VM_FAULT_WRITE;
1222 goto out_unlock;
1223 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001224 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001225 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001226alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001227 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001228 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001229 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001230 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001231 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001232 new_page = NULL;
1233
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001234 if (likely(new_page)) {
1235 prep_transhuge_page(new_page);
1236 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001237 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001238 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001239 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001240 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001241 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001242 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001243 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001244 ret |= VM_FAULT_FALLBACK;
1245 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001246 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001247 }
David Rientjes17766dd2013-09-12 15:14:06 -07001248 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001249 goto out;
1250 }
1251
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001252 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1253 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001254 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001255 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001256 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001257 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001258 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001259 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001260 goto out;
1261 }
1262
David Rientjes17766dd2013-09-12 15:14:06 -07001263 count_vm_event(THP_FAULT_ALLOC);
1264
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001265 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001266 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1267 else
1268 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001269 __SetPageUptodate(new_page);
1270
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001271 mmun_start = haddr;
1272 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001273 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001274
Jan Kara82b0f8c2016-12-14 15:06:58 -08001275 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001276 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001277 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001278 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1279 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001280 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001281 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001282 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001283 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001284 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001285 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1286 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001287 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001288 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001289 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001290 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001291 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1292 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001293 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001294 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001295 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001296 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001297 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001298 put_page(page);
1299 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001300 ret |= VM_FAULT_WRITE;
1301 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001302 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001303out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001304 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001305out:
1306 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001307out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001308 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001309 return ret;
1310}
1311
Keno Fischer8310d482017-01-24 15:17:48 -08001312/*
1313 * FOLL_FORCE can write to even unwritable pmd's, but only
1314 * after we've gone through a COW cycle and they are dirty.
1315 */
1316static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1317{
1318 return pmd_write(pmd) ||
1319 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1320}
1321
David Rientjesb676b292012-10-08 16:34:03 -07001322struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001323 unsigned long addr,
1324 pmd_t *pmd,
1325 unsigned int flags)
1326{
David Rientjesb676b292012-10-08 16:34:03 -07001327 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001328 struct page *page = NULL;
1329
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001330 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001331
Keno Fischer8310d482017-01-24 15:17:48 -08001332 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001333 goto out;
1334
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001335 /* Avoid dumping huge zero page */
1336 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1337 return ERR_PTR(-EFAULT);
1338
Mel Gorman2b4847e2013-12-18 17:08:32 -08001339 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001340 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001341 goto out;
1342
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001343 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001344 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001345 if (flags & FOLL_TOUCH)
1346 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001347 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001348 /*
1349 * We don't mlock() pte-mapped THPs. This way we can avoid
1350 * leaking mlocked pages into non-VM_LOCKED VMAs.
1351 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001352 * For anon THP:
1353 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001354 * In most cases the pmd is the only mapping of the page as we
1355 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1356 * writable private mappings in populate_vma_page_range().
1357 *
1358 * The only scenario when we have the page shared here is if we
1359 * mlocking read-only mapping shared over fork(). We skip
1360 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001361 *
1362 * For file THP:
1363 *
1364 * We can expect PageDoubleMap() to be stable under page lock:
1365 * for file pages we set it in page_add_file_rmap(), which
1366 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001367 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001368
1369 if (PageAnon(page) && compound_mapcount(page) != 1)
1370 goto skip_mlock;
1371 if (PageDoubleMap(page) || !page->mapping)
1372 goto skip_mlock;
1373 if (!trylock_page(page))
1374 goto skip_mlock;
1375 lru_add_drain();
1376 if (page->mapping && !PageDoubleMap(page))
1377 mlock_vma_page(page);
1378 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001379 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001380skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001381 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001382 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001383 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001384 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001385
1386out:
1387 return page;
1388}
1389
Mel Gormand10e63f2012-10-25 14:16:31 +02001390/* NUMA hinting page fault entry point for trans huge pmds */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001391int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001392{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001393 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001394 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001395 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001396 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001397 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001398 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001399 bool page_locked;
1400 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001401 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001402 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001403
Jan Kara82b0f8c2016-12-14 15:06:58 -08001404 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1405 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001406 goto out_unlock;
1407
Mel Gormande466bd2013-12-18 17:08:42 -08001408 /*
1409 * If there are potential migrations, wait for completion and retry
1410 * without disrupting NUMA hinting information. Do not relock and
1411 * check_same as the page may no longer be mapped.
1412 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001413 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1414 page = pmd_page(*vmf->pmd);
1415 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001416 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001417 goto out;
1418 }
1419
Mel Gormand10e63f2012-10-25 14:16:31 +02001420 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001421 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001422 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001423 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001424 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001425 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001426 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001427 flags |= TNF_FAULT_LOCAL;
1428 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001429
Mel Gormanbea66fb2015-03-25 15:55:37 -07001430 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001431 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001432 flags |= TNF_NO_GROUP;
1433
1434 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001435 * Acquire the page lock to serialise THP migrations but avoid dropping
1436 * page_table_lock if at all possible
1437 */
Mel Gormanb8916632013-10-07 11:28:44 +01001438 page_locked = trylock_page(page);
1439 target_nid = mpol_misplaced(page, vma, haddr);
1440 if (target_nid == -1) {
1441 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001442 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001443 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001444 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001445
Mel Gormande466bd2013-12-18 17:08:42 -08001446 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001447 if (!page_locked) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001448 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001449 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001450 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001451 goto out;
1452 }
1453
Mel Gorman2b4847e2013-12-18 17:08:32 -08001454 /*
1455 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1456 * to serialises splits
1457 */
Mel Gormanb8916632013-10-07 11:28:44 +01001458 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001459 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001460 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001461
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001462 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001463 spin_lock(vmf->ptl);
1464 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001465 unlock_page(page);
1466 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001467 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001468 goto out_unlock;
1469 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001470
Mel Gormanc3a489c2013-12-18 17:08:38 -08001471 /* Bail if we fail to protect against THP splits for any reason */
1472 if (unlikely(!anon_vma)) {
1473 put_page(page);
1474 page_nid = -1;
1475 goto clear_pmdnuma;
1476 }
1477
Mel Gormana54a4072013-10-07 11:28:46 +01001478 /*
1479 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001480 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001481 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001482 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001483 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001484 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001485 if (migrated) {
1486 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001487 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001488 } else
1489 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001490
Mel Gorman8191acb2013-10-07 11:28:45 +01001491 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001492clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001493 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001494 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001495 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001496 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001497 if (was_writable)
1498 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001499 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1500 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001501 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001502out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001503 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001504
1505out:
1506 if (anon_vma)
1507 page_unlock_anon_vma_read(anon_vma);
1508
Mel Gorman8191acb2013-10-07 11:28:45 +01001509 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001510 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001511 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001512
Mel Gormand10e63f2012-10-25 14:16:31 +02001513 return 0;
1514}
1515
Huang Ying319904a2016-07-28 15:48:03 -07001516/*
1517 * Return true if we do MADV_FREE successfully on entire pmd page.
1518 * Otherwise, return false.
1519 */
1520bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001521 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001522{
1523 spinlock_t *ptl;
1524 pmd_t orig_pmd;
1525 struct page *page;
1526 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001527 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001528
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001529 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1530
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001531 ptl = pmd_trans_huge_lock(pmd, vma);
1532 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001533 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001534
1535 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001536 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001537 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001538
1539 page = pmd_page(orig_pmd);
1540 /*
1541 * If other processes are mapping this page, we couldn't discard
1542 * the page unless they all do MADV_FREE so let's skip the page.
1543 */
1544 if (page_mapcount(page) != 1)
1545 goto out;
1546
1547 if (!trylock_page(page))
1548 goto out;
1549
1550 /*
1551 * If user want to discard part-pages of THP, split it so MADV_FREE
1552 * will deactivate only them.
1553 */
1554 if (next - addr != HPAGE_PMD_SIZE) {
1555 get_page(page);
1556 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001557 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001558 put_page(page);
1559 unlock_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001560 goto out_unlocked;
1561 }
1562
1563 if (PageDirty(page))
1564 ClearPageDirty(page);
1565 unlock_page(page);
1566
1567 if (PageActive(page))
1568 deactivate_page(page);
1569
1570 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1571 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1572 tlb->fullmm);
1573 orig_pmd = pmd_mkold(orig_pmd);
1574 orig_pmd = pmd_mkclean(orig_pmd);
1575
1576 set_pmd_at(mm, addr, pmd, orig_pmd);
1577 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1578 }
Huang Ying319904a2016-07-28 15:48:03 -07001579 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001580out:
1581 spin_unlock(ptl);
1582out_unlocked:
1583 return ret;
1584}
1585
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001586static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1587{
1588 pgtable_t pgtable;
1589
1590 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1591 pte_free(mm, pgtable);
1592 atomic_long_dec(&mm->nr_ptes);
1593}
1594
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001595int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001596 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001597{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001598 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001599 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001600
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001601 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1602
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001603 ptl = __pmd_trans_huge_lock(pmd, vma);
1604 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001605 return 0;
1606 /*
1607 * For architectures like ppc64 we look at deposited pgtable
1608 * when calling pmdp_huge_get_and_clear. So do the
1609 * pgtable_trans_huge_withdraw after finishing pmdp related
1610 * operations.
1611 */
1612 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1613 tlb->fullmm);
1614 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1615 if (vma_is_dax(vma)) {
1616 spin_unlock(ptl);
1617 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001618 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001619 } else if (is_huge_zero_pmd(orig_pmd)) {
1620 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1621 atomic_long_dec(&tlb->mm->nr_ptes);
1622 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001623 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001624 } else {
1625 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001626 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001627 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001628 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001629 if (PageAnon(page)) {
1630 pgtable_t pgtable;
1631 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1632 pte_free(tlb->mm, pgtable);
1633 atomic_long_dec(&tlb->mm->nr_ptes);
1634 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1635 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001636 if (arch_needs_pgtable_deposit())
1637 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001638 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1639 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001640 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001641 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001642 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001643 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001644}
1645
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001646#ifndef pmd_move_must_withdraw
1647static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1648 spinlock_t *old_pmd_ptl,
1649 struct vm_area_struct *vma)
1650{
1651 /*
1652 * With split pmd lock we also need to move preallocated
1653 * PTE page table if new_pmd is on different PMD page table.
1654 *
1655 * We also don't deposit and withdraw tables for file pages.
1656 */
1657 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1658}
1659#endif
1660
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001661bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001662 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001663 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001664{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001665 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001666 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001667 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001668 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001669
1670 if ((old_addr & ~HPAGE_PMD_MASK) ||
1671 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001672 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001673 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001674
1675 /*
1676 * The destination pmd shouldn't be established, free_pgtables()
1677 * should have release it.
1678 */
1679 if (WARN_ON(!pmd_none(*new_pmd))) {
1680 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001681 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001682 }
1683
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001684 /*
1685 * We don't have to worry about the ordering of src and dst
1686 * ptlocks because exclusive mmap_sem prevents deadlock.
1687 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001688 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1689 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001690 new_ptl = pmd_lockptr(mm, new_pmd);
1691 if (new_ptl != old_ptl)
1692 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001693 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001694 if (pmd_present(pmd) && pmd_dirty(pmd))
1695 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001696 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001697
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001698 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301699 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001700 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1701 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001702 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301703 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1704 if (new_ptl != old_ptl)
1705 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001706 if (force_flush)
1707 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1708 else
1709 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001710 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001711 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001712 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001713 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001714}
1715
Mel Gormanf123d742013-10-07 11:28:49 +01001716/*
1717 * Returns
1718 * - 0 if PMD could not be locked
1719 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1720 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1721 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001722int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001723 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001724{
1725 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001726 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001727 int ret = 0;
1728
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001729 ptl = __pmd_trans_huge_lock(pmd, vma);
1730 if (ptl) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001731 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001732 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001733 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001734
1735 /*
1736 * Avoid trapping faults against the zero page. The read-only
1737 * data is likely to be read-cached on the local CPU and
1738 * local/remote hits to the zero page are not interesting.
1739 */
1740 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1741 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001742 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001743 }
1744
Mel Gorman10c10452015-02-12 14:58:44 -08001745 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001746 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001747 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001748 if (preserve_write)
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001749 entry = pmd_mk_savedwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001750 ret = HPAGE_PMD_NR;
1751 set_pmd_at(mm, addr, pmd, entry);
Kirill A. Shutemovb237ade2016-07-26 15:25:45 -07001752 BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1753 pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001754 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001755 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001756 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001757
1758 return ret;
1759}
1760
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001761/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001762 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001763 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001764 * Note that if it returns page table lock pointer, this routine returns without
1765 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001766 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001767spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001768{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001769 spinlock_t *ptl;
1770 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001771 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001772 return ptl;
1773 spin_unlock(ptl);
1774 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001775}
1776
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001777/*
1778 * Returns true if a given pud maps a thp, false otherwise.
1779 *
1780 * Note that if it returns true, this routine returns without unlocking page
1781 * table lock. So callers must unlock it.
1782 */
1783spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1784{
1785 spinlock_t *ptl;
1786
1787 ptl = pud_lock(vma->vm_mm, pud);
1788 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1789 return ptl;
1790 spin_unlock(ptl);
1791 return NULL;
1792}
1793
1794#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1795int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1796 pud_t *pud, unsigned long addr)
1797{
1798 pud_t orig_pud;
1799 spinlock_t *ptl;
1800
1801 ptl = __pud_trans_huge_lock(pud, vma);
1802 if (!ptl)
1803 return 0;
1804 /*
1805 * For architectures like ppc64 we look at deposited pgtable
1806 * when calling pudp_huge_get_and_clear. So do the
1807 * pgtable_trans_huge_withdraw after finishing pudp related
1808 * operations.
1809 */
1810 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1811 tlb->fullmm);
1812 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1813 if (vma_is_dax(vma)) {
1814 spin_unlock(ptl);
1815 /* No zero page support yet */
1816 } else {
1817 /* No support for anonymous PUD pages yet */
1818 BUG();
1819 }
1820 return 1;
1821}
1822
1823static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1824 unsigned long haddr)
1825{
1826 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1827 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1828 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1829 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1830
1831 count_vm_event(THP_SPLIT_PMD);
1832
1833 pudp_huge_clear_flush_notify(vma, haddr, pud);
1834}
1835
1836void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1837 unsigned long address)
1838{
1839 spinlock_t *ptl;
1840 struct mm_struct *mm = vma->vm_mm;
1841 unsigned long haddr = address & HPAGE_PUD_MASK;
1842
1843 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1844 ptl = pud_lock(mm, pud);
1845 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1846 goto out;
1847 __split_huge_pud_locked(vma, pud, haddr);
1848
1849out:
1850 spin_unlock(ptl);
1851 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1852}
1853#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1854
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001855static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1856 unsigned long haddr, pmd_t *pmd)
1857{
1858 struct mm_struct *mm = vma->vm_mm;
1859 pgtable_t pgtable;
1860 pmd_t _pmd;
1861 int i;
1862
1863 /* leave pmd empty until pte is filled */
1864 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1865
1866 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1867 pmd_populate(mm, &_pmd, pgtable);
1868
1869 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1870 pte_t *pte, entry;
1871 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1872 entry = pte_mkspecial(entry);
1873 pte = pte_offset_map(&_pmd, haddr);
1874 VM_BUG_ON(!pte_none(*pte));
1875 set_pte_at(mm, haddr, pte, entry);
1876 pte_unmap(pte);
1877 }
1878 smp_wmb(); /* make pte visible before pmd */
1879 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001880}
1881
1882static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001883 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001884{
1885 struct mm_struct *mm = vma->vm_mm;
1886 struct page *page;
1887 pgtable_t pgtable;
1888 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001889 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001890 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001891 int i;
1892
1893 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1894 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1895 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001896 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001897
1898 count_vm_event(THP_SPLIT_PMD);
1899
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001900 if (!vma_is_anonymous(vma)) {
1901 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001902 /*
1903 * We are going to unmap this huge page. So
1904 * just go ahead and zap it
1905 */
1906 if (arch_needs_pgtable_deposit())
1907 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001908 if (vma_is_dax(vma))
1909 return;
1910 page = pmd_page(_pmd);
1911 if (!PageReferenced(page) && pmd_young(_pmd))
1912 SetPageReferenced(page);
1913 page_remove_rmap(page, true);
1914 put_page(page);
1915 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001916 return;
1917 } else if (is_huge_zero_pmd(*pmd)) {
1918 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1919 }
1920
1921 page = pmd_page(*pmd);
1922 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001923 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001924 write = pmd_write(*pmd);
1925 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001926 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001927 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001928
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301929 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001930 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1931 pmd_populate(mm, &_pmd, pgtable);
1932
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001933 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001934 pte_t entry, *pte;
1935 /*
1936 * Note that NUMA hinting access restrictions are not
1937 * transferred to avoid any possibility of altering
1938 * permissions across VMAs.
1939 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001940 if (freeze) {
1941 swp_entry_t swp_entry;
1942 swp_entry = make_migration_entry(page + i, write);
1943 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001944 if (soft_dirty)
1945 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001946 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001947 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001948 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001949 if (!write)
1950 entry = pte_wrprotect(entry);
1951 if (!young)
1952 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001953 if (soft_dirty)
1954 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001955 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001956 if (dirty)
1957 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001958 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001959 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001960 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001961 atomic_inc(&page[i]._mapcount);
1962 pte_unmap(pte);
1963 }
1964
1965 /*
1966 * Set PG_double_map before dropping compound_mapcount to avoid
1967 * false-negative page_mapped().
1968 */
1969 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1970 for (i = 0; i < HPAGE_PMD_NR; i++)
1971 atomic_inc(&page[i]._mapcount);
1972 }
1973
1974 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1975 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001976 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001977 if (TestClearPageDoubleMap(page)) {
1978 /* No need in mapcount reference anymore */
1979 for (i = 0; i < HPAGE_PMD_NR; i++)
1980 atomic_dec(&page[i]._mapcount);
1981 }
1982 }
1983
1984 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001985 /*
1986 * Up to this point the pmd is present and huge and userland has the
1987 * whole access to the hugepage during the split (which happens in
1988 * place). If we overwrite the pmd with the not-huge version pointing
1989 * to the pte here (which of course we could if all CPUs were bug
1990 * free), userland could trigger a small page size TLB miss on the
1991 * small sized TLB while the hugepage TLB entry is still established in
1992 * the huge TLB. Some CPU doesn't like that.
1993 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1994 * 383 on page 93. Intel should be safe but is also warns that it's
1995 * only safe if the permission and cache attributes of the two entries
1996 * loaded in the two TLB is identical (which should be the case here).
1997 * But it is generally safer to never allow small and huge TLB entries
1998 * for the same virtual address to be loaded simultaneously. So instead
1999 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2000 * current pmd notpresent (atomically because here the pmd_trans_huge
2001 * and pmd_trans_splitting must remain set at all times on the pmd
2002 * until the split is complete for this pmd), then we flush the SMP TLB
2003 * and finally we write the non-huge version of the pmd entry with
2004 * pmd_populate.
2005 */
2006 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002007 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002008
2009 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002010 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002011 page_remove_rmap(page + i, false);
2012 put_page(page + i);
2013 }
2014 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002015}
2016
2017void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002018 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002019{
2020 spinlock_t *ptl;
2021 struct mm_struct *mm = vma->vm_mm;
2022 unsigned long haddr = address & HPAGE_PMD_MASK;
2023
2024 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2025 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002026
2027 /*
2028 * If caller asks to setup a migration entries, we need a page to check
2029 * pmd against. Otherwise we can end up replacing wrong page.
2030 */
2031 VM_BUG_ON(freeze && !page);
2032 if (page && page != pmd_page(*pmd))
2033 goto out;
2034
Dan Williams5c7fb562016-01-15 16:56:52 -08002035 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002036 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002037 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002038 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08002039 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002040 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002041 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002042out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002043 spin_unlock(ptl);
2044 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2045}
2046
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002047void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2048 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002049{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002050 pgd_t *pgd;
2051 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002052 pmd_t *pmd;
2053
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002054 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002055 if (!pgd_present(*pgd))
2056 return;
2057
2058 pud = pud_offset(pgd, address);
2059 if (!pud_present(*pud))
2060 return;
2061
2062 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002063
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002064 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002065}
2066
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002067void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002068 unsigned long start,
2069 unsigned long end,
2070 long adjust_next)
2071{
2072 /*
2073 * If the new start address isn't hpage aligned and it could
2074 * previously contain an hugepage: check if we need to split
2075 * an huge pmd.
2076 */
2077 if (start & ~HPAGE_PMD_MASK &&
2078 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2079 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002080 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002081
2082 /*
2083 * If the new end address isn't hpage aligned and it could
2084 * previously contain an hugepage: check if we need to split
2085 * an huge pmd.
2086 */
2087 if (end & ~HPAGE_PMD_MASK &&
2088 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2089 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002090 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002091
2092 /*
2093 * If we're also updating the vma->vm_next->vm_start, if the new
2094 * vm_next->vm_start isn't page aligned and it could previously
2095 * contain an hugepage: check if we need to split an huge pmd.
2096 */
2097 if (adjust_next > 0) {
2098 struct vm_area_struct *next = vma->vm_next;
2099 unsigned long nstart = next->vm_start;
2100 nstart += adjust_next << PAGE_SHIFT;
2101 if (nstart & ~HPAGE_PMD_MASK &&
2102 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2103 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002104 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002105 }
2106}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002107
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002108static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002109{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002110 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002111 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2112 int ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002113
2114 VM_BUG_ON_PAGE(!PageHead(page), page);
2115
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002116 if (PageAnon(page))
2117 ttu_flags |= TTU_MIGRATION;
2118
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002119 ret = try_to_unmap(page, ttu_flags);
2120 VM_BUG_ON_PAGE(ret, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002121}
2122
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002123static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002124{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002125 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002126 if (PageTransHuge(page)) {
2127 remove_migration_ptes(page, page, true);
2128 } else {
2129 for (i = 0; i < HPAGE_PMD_NR; i++)
2130 remove_migration_ptes(page + i, page + i, true);
2131 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002132}
2133
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002134static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002135 struct lruvec *lruvec, struct list_head *list)
2136{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002137 struct page *page_tail = head + tail;
2138
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002139 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002140 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002141
2142 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002143 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002144 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002145 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2146 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002147 * get_page_unless_zero(), and atomic_set() is implemented in C not
2148 * using locked ops. spin_unlock on x86 sometime uses locked ops
2149 * because of PPro errata 66, 92, so unless somebody can guarantee
2150 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002151 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002152 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002153 if (PageAnon(head)) {
2154 page_ref_inc(page_tail);
2155 } else {
2156 /* Additional pin to radix tree */
2157 page_ref_add(page_tail, 2);
2158 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002159
2160 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2161 page_tail->flags |= (head->flags &
2162 ((1L << PG_referenced) |
2163 (1L << PG_swapbacked) |
2164 (1L << PG_mlocked) |
2165 (1L << PG_uptodate) |
2166 (1L << PG_active) |
2167 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002168 (1L << PG_unevictable) |
2169 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002170
2171 /*
2172 * After clearing PageTail the gup refcount can be released.
2173 * Page flags also must be visible before we make the page non-compound.
2174 */
2175 smp_wmb();
2176
2177 clear_compound_head(page_tail);
2178
2179 if (page_is_young(head))
2180 set_page_young(page_tail);
2181 if (page_is_idle(head))
2182 set_page_idle(page_tail);
2183
2184 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002185 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002186 page_tail);
2187 page_tail->mapping = head->mapping;
2188
2189 page_tail->index = head->index + tail;
2190 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2191 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002192}
2193
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002194static void __split_huge_page(struct page *page, struct list_head *list,
2195 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002196{
2197 struct page *head = compound_head(page);
2198 struct zone *zone = page_zone(head);
2199 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002200 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002201 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002202
Mel Gorman599d0c92016-07-28 15:45:31 -07002203 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002204
2205 /* complete memcg works before add pages to LRU */
2206 mem_cgroup_split_huge_fixup(head);
2207
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002208 if (!PageAnon(page))
2209 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2210
2211 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002212 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002213 /* Some pages can be beyond i_size: drop them from page cache */
2214 if (head[i].index >= end) {
2215 __ClearPageDirty(head + i);
2216 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002217 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2218 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002219 put_page(head + i);
2220 }
2221 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002222
2223 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002224 /* See comment in __split_huge_page_tail() */
2225 if (PageAnon(head)) {
2226 page_ref_inc(head);
2227 } else {
2228 /* Additional pin to radix tree */
2229 page_ref_add(head, 2);
2230 spin_unlock(&head->mapping->tree_lock);
2231 }
2232
Mel Gormana52633d2016-07-28 15:45:28 -07002233 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002234
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002235 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002236
2237 for (i = 0; i < HPAGE_PMD_NR; i++) {
2238 struct page *subpage = head + i;
2239 if (subpage == page)
2240 continue;
2241 unlock_page(subpage);
2242
2243 /*
2244 * Subpages may be freed if there wasn't any mapping
2245 * like if add_to_swap() is running on a lru page that
2246 * had its mapping zapped. And freeing these pages
2247 * requires taking the lru_lock so we do the put_page
2248 * of the tail pages after the split is complete.
2249 */
2250 put_page(subpage);
2251 }
2252}
2253
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002254int total_mapcount(struct page *page)
2255{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002256 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002257
2258 VM_BUG_ON_PAGE(PageTail(page), page);
2259
2260 if (likely(!PageCompound(page)))
2261 return atomic_read(&page->_mapcount) + 1;
2262
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002263 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002264 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002265 return compound;
2266 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002267 for (i = 0; i < HPAGE_PMD_NR; i++)
2268 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002269 /* File pages has compound_mapcount included in _mapcount */
2270 if (!PageAnon(page))
2271 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002272 if (PageDoubleMap(page))
2273 ret -= HPAGE_PMD_NR;
2274 return ret;
2275}
2276
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002277/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002278 * This calculates accurately how many mappings a transparent hugepage
2279 * has (unlike page_mapcount() which isn't fully accurate). This full
2280 * accuracy is primarily needed to know if copy-on-write faults can
2281 * reuse the page and change the mapping to read-write instead of
2282 * copying them. At the same time this returns the total_mapcount too.
2283 *
2284 * The function returns the highest mapcount any one of the subpages
2285 * has. If the return value is one, even if different processes are
2286 * mapping different subpages of the transparent hugepage, they can
2287 * all reuse it, because each process is reusing a different subpage.
2288 *
2289 * The total_mapcount is instead counting all virtual mappings of the
2290 * subpages. If the total_mapcount is equal to "one", it tells the
2291 * caller all mappings belong to the same "mm" and in turn the
2292 * anon_vma of the transparent hugepage can become the vma->anon_vma
2293 * local one as no other process may be mapping any of the subpages.
2294 *
2295 * It would be more accurate to replace page_mapcount() with
2296 * page_trans_huge_mapcount(), however we only use
2297 * page_trans_huge_mapcount() in the copy-on-write faults where we
2298 * need full accuracy to avoid breaking page pinning, because
2299 * page_trans_huge_mapcount() is slower than page_mapcount().
2300 */
2301int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2302{
2303 int i, ret, _total_mapcount, mapcount;
2304
2305 /* hugetlbfs shouldn't call it */
2306 VM_BUG_ON_PAGE(PageHuge(page), page);
2307
2308 if (likely(!PageTransCompound(page))) {
2309 mapcount = atomic_read(&page->_mapcount) + 1;
2310 if (total_mapcount)
2311 *total_mapcount = mapcount;
2312 return mapcount;
2313 }
2314
2315 page = compound_head(page);
2316
2317 _total_mapcount = ret = 0;
2318 for (i = 0; i < HPAGE_PMD_NR; i++) {
2319 mapcount = atomic_read(&page[i]._mapcount) + 1;
2320 ret = max(ret, mapcount);
2321 _total_mapcount += mapcount;
2322 }
2323 if (PageDoubleMap(page)) {
2324 ret -= 1;
2325 _total_mapcount -= HPAGE_PMD_NR;
2326 }
2327 mapcount = compound_mapcount(page);
2328 ret += mapcount;
2329 _total_mapcount += mapcount;
2330 if (total_mapcount)
2331 *total_mapcount = _total_mapcount;
2332 return ret;
2333}
2334
2335/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002336 * This function splits huge page into normal pages. @page can point to any
2337 * subpage of huge page to split. Split doesn't change the position of @page.
2338 *
2339 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2340 * The huge page must be locked.
2341 *
2342 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2343 *
2344 * Both head page and tail pages will inherit mapping, flags, and so on from
2345 * the hugepage.
2346 *
2347 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2348 * they are not mapped.
2349 *
2350 * Returns 0 if the hugepage is split successfully.
2351 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2352 * us.
2353 */
2354int split_huge_page_to_list(struct page *page, struct list_head *list)
2355{
2356 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002357 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002358 struct anon_vma *anon_vma = NULL;
2359 struct address_space *mapping = NULL;
2360 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002361 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002362 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002363
2364 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002365 VM_BUG_ON_PAGE(!PageLocked(page), page);
2366 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2367 VM_BUG_ON_PAGE(!PageCompound(page), page);
2368
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002369 if (PageAnon(head)) {
2370 /*
2371 * The caller does not necessarily hold an mmap_sem that would
2372 * prevent the anon_vma disappearing so we first we take a
2373 * reference to it and then lock the anon_vma for write. This
2374 * is similar to page_lock_anon_vma_read except the write lock
2375 * is taken to serialise against parallel split or collapse
2376 * operations.
2377 */
2378 anon_vma = page_get_anon_vma(head);
2379 if (!anon_vma) {
2380 ret = -EBUSY;
2381 goto out;
2382 }
2383 extra_pins = 0;
2384 mapping = NULL;
2385 anon_vma_lock_write(anon_vma);
2386 } else {
2387 mapping = head->mapping;
2388
2389 /* Truncated ? */
2390 if (!mapping) {
2391 ret = -EBUSY;
2392 goto out;
2393 }
2394
2395 /* Addidional pins from radix tree */
2396 extra_pins = HPAGE_PMD_NR;
2397 anon_vma = NULL;
2398 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002399 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002400
2401 /*
2402 * Racy check if we can split the page, before freeze_page() will
2403 * split PMDs
2404 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002405 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002406 ret = -EBUSY;
2407 goto out_unlock;
2408 }
2409
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002410 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002411 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002412 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2413
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002414 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2415 if (mlocked)
2416 lru_add_drain();
2417
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002418 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002419 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002420
2421 if (mapping) {
2422 void **pslot;
2423
2424 spin_lock(&mapping->tree_lock);
2425 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2426 page_index(head));
2427 /*
2428 * Check if the head page is present in radix tree.
2429 * We assume all tail are present too, if head is there.
2430 */
2431 if (radix_tree_deref_slot_protected(pslot,
2432 &mapping->tree_lock) != head)
2433 goto fail;
2434 }
2435
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002436 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002437 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002438 count = page_count(head);
2439 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002440 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002441 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002442 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002443 list_del(page_deferred_list(head));
2444 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002445 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002446 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002447 spin_unlock(&pgdata->split_queue_lock);
2448 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002449 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002450 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002451 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2452 pr_alert("total_mapcount: %u, page_count(): %u\n",
2453 mapcount, count);
2454 if (PageTail(page))
2455 dump_page(head, NULL);
2456 dump_page(page, "total_mapcount(head) > 0");
2457 BUG();
2458 }
2459 spin_unlock(&pgdata->split_queue_lock);
2460fail: if (mapping)
2461 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002462 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002463 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002464 ret = -EBUSY;
2465 }
2466
2467out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002468 if (anon_vma) {
2469 anon_vma_unlock_write(anon_vma);
2470 put_anon_vma(anon_vma);
2471 }
2472 if (mapping)
2473 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002474out:
2475 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2476 return ret;
2477}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002478
2479void free_transhuge_page(struct page *page)
2480{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002481 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002482 unsigned long flags;
2483
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002484 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002485 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002486 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002487 list_del(page_deferred_list(page));
2488 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002489 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002490 free_compound_page(page);
2491}
2492
2493void deferred_split_huge_page(struct page *page)
2494{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002495 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002496 unsigned long flags;
2497
2498 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2499
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002500 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002501 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002502 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002503 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2504 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002505 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002506 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002507}
2508
2509static unsigned long deferred_split_count(struct shrinker *shrink,
2510 struct shrink_control *sc)
2511{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002512 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002513 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002514}
2515
2516static unsigned long deferred_split_scan(struct shrinker *shrink,
2517 struct shrink_control *sc)
2518{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002519 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002520 unsigned long flags;
2521 LIST_HEAD(list), *pos, *next;
2522 struct page *page;
2523 int split = 0;
2524
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002525 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002526 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002527 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002528 page = list_entry((void *)pos, struct page, mapping);
2529 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002530 if (get_page_unless_zero(page)) {
2531 list_move(page_deferred_list(page), &list);
2532 } else {
2533 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002534 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002535 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002536 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002537 if (!--sc->nr_to_scan)
2538 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002539 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002540 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002541
2542 list_for_each_safe(pos, next, &list) {
2543 page = list_entry((void *)pos, struct page, mapping);
2544 lock_page(page);
2545 /* split_huge_page() removes page from list on success */
2546 if (!split_huge_page(page))
2547 split++;
2548 unlock_page(page);
2549 put_page(page);
2550 }
2551
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002552 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2553 list_splice_tail(&list, &pgdata->split_queue);
2554 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002555
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002556 /*
2557 * Stop shrinker if we didn't split any page, but the queue is empty.
2558 * This can happen if pages were freed under us.
2559 */
2560 if (!split && list_empty(&pgdata->split_queue))
2561 return SHRINK_STOP;
2562 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002563}
2564
2565static struct shrinker deferred_split_shrinker = {
2566 .count_objects = deferred_split_count,
2567 .scan_objects = deferred_split_scan,
2568 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002569 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002570};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002571
2572#ifdef CONFIG_DEBUG_FS
2573static int split_huge_pages_set(void *data, u64 val)
2574{
2575 struct zone *zone;
2576 struct page *page;
2577 unsigned long pfn, max_zone_pfn;
2578 unsigned long total = 0, split = 0;
2579
2580 if (val != 1)
2581 return -EINVAL;
2582
2583 for_each_populated_zone(zone) {
2584 max_zone_pfn = zone_end_pfn(zone);
2585 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2586 if (!pfn_valid(pfn))
2587 continue;
2588
2589 page = pfn_to_page(pfn);
2590 if (!get_page_unless_zero(page))
2591 continue;
2592
2593 if (zone != page_zone(page))
2594 goto next;
2595
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002596 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002597 goto next;
2598
2599 total++;
2600 lock_page(page);
2601 if (!split_huge_page(page))
2602 split++;
2603 unlock_page(page);
2604next:
2605 put_page(page);
2606 }
2607 }
2608
Yang Shi145bdaa2016-05-05 16:22:00 -07002609 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002610
2611 return 0;
2612}
2613DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2614 "%llu\n");
2615
2616static int __init split_huge_pages_debugfs(void)
2617{
2618 void *ret;
2619
Yang Shi145bdaa2016-05-05 16:22:00 -07002620 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002621 &split_huge_pages_fops);
2622 if (!ret)
2623 pr_warn("Failed to create split_huge_pages in debugfs");
2624 return 0;
2625}
2626late_initcall(split_huge_pages_debugfs);
2627#endif