blob: 86975dec0ba160feadfb8aa0d13b8f2be943638d [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);
David Rientjes21440d72017-02-22 15:45:49 -0800243 } else if (!memcmp("defer+madvise", buf,
244 min(sizeof("defer+madvise")-1, count))) {
245 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
248 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
David Rientjes4fad7fb2017-04-07 16:04:54 -0700249 } else if (!memcmp("defer", buf,
250 min(sizeof("defer")-1, count))) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
David Rientjes21440d72017-02-22 15:45:49 -0800255 } 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,
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700718 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
719 pgtable_t pgtable)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700720{
721 struct mm_struct *mm = vma->vm_mm;
722 pmd_t entry;
723 spinlock_t *ptl;
724
725 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800726 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
727 if (pfn_t_devmap(pfn))
728 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800729 if (write) {
730 entry = pmd_mkyoung(pmd_mkdirty(entry));
731 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700732 }
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700733
734 if (pgtable) {
735 pgtable_trans_huge_deposit(mm, pmd, pgtable);
736 atomic_long_inc(&mm->nr_ptes);
737 }
738
Ross Zwisler01871e52016-01-15 16:56:02 -0800739 set_pmd_at(mm, addr, pmd, entry);
740 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700741 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700742}
743
744int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800745 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700746{
747 pgprot_t pgprot = vma->vm_page_prot;
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700748 pgtable_t pgtable = NULL;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700749 /*
750 * If we had pmd_special, we could avoid all these restrictions,
751 * but we need to be consistent with PTEs and architectures that
752 * can't support a 'special' bit.
753 */
754 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
755 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
756 (VM_PFNMAP|VM_MIXEDMAP));
757 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800758 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700759
760 if (addr < vma->vm_start || addr >= vma->vm_end)
761 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200762
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700763 if (arch_needs_pgtable_deposit()) {
764 pgtable = pte_alloc_one(vma->vm_mm, addr);
765 if (!pgtable)
766 return VM_FAULT_OOM;
767 }
768
Borislav Petkov308a0472016-10-26 19:43:43 +0200769 track_pfn_insert(vma, &pgprot, pfn);
770
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700771 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700772 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700773}
Dan Williamsdee41072016-05-14 12:20:44 -0700774EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700775
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800776#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
777static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
778{
779 if (likely(vma->vm_flags & VM_WRITE))
780 pud = pud_mkwrite(pud);
781 return pud;
782}
783
784static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
785 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
786{
787 struct mm_struct *mm = vma->vm_mm;
788 pud_t entry;
789 spinlock_t *ptl;
790
791 ptl = pud_lock(mm, pud);
792 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
793 if (pfn_t_devmap(pfn))
794 entry = pud_mkdevmap(entry);
795 if (write) {
796 entry = pud_mkyoung(pud_mkdirty(entry));
797 entry = maybe_pud_mkwrite(entry, vma);
798 }
799 set_pud_at(mm, addr, pud, entry);
800 update_mmu_cache_pud(vma, addr, pud);
801 spin_unlock(ptl);
802}
803
804int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
805 pud_t *pud, pfn_t pfn, bool write)
806{
807 pgprot_t pgprot = vma->vm_page_prot;
808 /*
809 * If we had pud_special, we could avoid all these restrictions,
810 * but we need to be consistent with PTEs and architectures that
811 * can't support a 'special' bit.
812 */
813 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
814 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
815 (VM_PFNMAP|VM_MIXEDMAP));
816 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
817 BUG_ON(!pfn_t_devmap(pfn));
818
819 if (addr < vma->vm_start || addr >= vma->vm_end)
820 return VM_FAULT_SIGBUS;
821
822 track_pfn_insert(vma, &pgprot, pfn);
823
824 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
825 return VM_FAULT_NOPAGE;
826}
827EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
828#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
829
Dan Williams3565fce2016-01-15 16:56:55 -0800830static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
831 pmd_t *pmd)
832{
833 pmd_t _pmd;
834
835 /*
836 * We should set the dirty bit only for FOLL_WRITE but for now
837 * the dirty bit in the pmd is meaningless. And if the dirty
838 * bit will become meaningful and we'll only set it with
839 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
840 * set the young bit, instead of the current set_pmd_at.
841 */
842 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
843 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
844 pmd, _pmd, 1))
845 update_mmu_cache_pmd(vma, addr, pmd);
846}
847
848struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
849 pmd_t *pmd, int flags)
850{
851 unsigned long pfn = pmd_pfn(*pmd);
852 struct mm_struct *mm = vma->vm_mm;
853 struct dev_pagemap *pgmap;
854 struct page *page;
855
856 assert_spin_locked(pmd_lockptr(mm, pmd));
857
Keno Fischer8310d482017-01-24 15:17:48 -0800858 /*
859 * When we COW a devmap PMD entry, we split it into PTEs, so we should
860 * not be in this function with `flags & FOLL_COW` set.
861 */
862 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
863
Dan Williams3565fce2016-01-15 16:56:55 -0800864 if (flags & FOLL_WRITE && !pmd_write(*pmd))
865 return NULL;
866
867 if (pmd_present(*pmd) && pmd_devmap(*pmd))
868 /* pass */;
869 else
870 return NULL;
871
872 if (flags & FOLL_TOUCH)
873 touch_pmd(vma, addr, pmd);
874
875 /*
876 * device mapped pages can only be returned if the
877 * caller will manage the page reference count.
878 */
879 if (!(flags & FOLL_GET))
880 return ERR_PTR(-EEXIST);
881
882 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
883 pgmap = get_dev_pagemap(pfn, NULL);
884 if (!pgmap)
885 return ERR_PTR(-EFAULT);
886 page = pfn_to_page(pfn);
887 get_page(page);
888 put_dev_pagemap(pgmap);
889
890 return page;
891}
892
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800893int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
894 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
895 struct vm_area_struct *vma)
896{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800897 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800898 struct page *src_page;
899 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800900 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700901 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800902
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700903 /* Skip if can be re-fill on fault */
904 if (!vma_is_anonymous(vma))
905 return 0;
906
907 pgtable = pte_alloc_one(dst_mm, addr);
908 if (unlikely(!pgtable))
909 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800911 dst_ptl = pmd_lock(dst_mm, dst_pmd);
912 src_ptl = pmd_lockptr(src_mm, src_pmd);
913 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800914
915 ret = -EAGAIN;
916 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700917 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800918 pte_free(dst_mm, pgtable);
919 goto out_unlock;
920 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800921 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800922 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800923 * under splitting since we don't split the page itself, only pmd to
924 * a page table.
925 */
926 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700927 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800928 /*
929 * get_huge_zero_page() will never allocate a new page here,
930 * since we already have a zero page to copy. It just takes a
931 * reference.
932 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700933 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700934 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700935 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800936 ret = 0;
937 goto out_unlock;
938 }
Mel Gormande466bd2013-12-18 17:08:42 -0800939
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700940 src_page = pmd_page(pmd);
941 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
942 get_page(src_page);
943 page_dup_rmap(src_page, true);
944 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
945 atomic_long_inc(&dst_mm->nr_ptes);
946 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800947
948 pmdp_set_wrprotect(src_mm, addr, src_pmd);
949 pmd = pmd_mkold(pmd_wrprotect(pmd));
950 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800951
952 ret = 0;
953out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800954 spin_unlock(src_ptl);
955 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800956out:
957 return ret;
958}
959
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
961static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
962 pud_t *pud)
963{
964 pud_t _pud;
965
966 /*
967 * We should set the dirty bit only for FOLL_WRITE but for now
968 * the dirty bit in the pud is meaningless. And if the dirty
969 * bit will become meaningful and we'll only set it with
970 * FOLL_WRITE, an atomic set_bit will be required on the pud to
971 * set the young bit, instead of the current set_pud_at.
972 */
973 _pud = pud_mkyoung(pud_mkdirty(*pud));
974 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
975 pud, _pud, 1))
976 update_mmu_cache_pud(vma, addr, pud);
977}
978
979struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
980 pud_t *pud, int flags)
981{
982 unsigned long pfn = pud_pfn(*pud);
983 struct mm_struct *mm = vma->vm_mm;
984 struct dev_pagemap *pgmap;
985 struct page *page;
986
987 assert_spin_locked(pud_lockptr(mm, pud));
988
989 if (flags & FOLL_WRITE && !pud_write(*pud))
990 return NULL;
991
992 if (pud_present(*pud) && pud_devmap(*pud))
993 /* pass */;
994 else
995 return NULL;
996
997 if (flags & FOLL_TOUCH)
998 touch_pud(vma, addr, pud);
999
1000 /*
1001 * device mapped pages can only be returned if the
1002 * caller will manage the page reference count.
1003 */
1004 if (!(flags & FOLL_GET))
1005 return ERR_PTR(-EEXIST);
1006
1007 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1008 pgmap = get_dev_pagemap(pfn, NULL);
1009 if (!pgmap)
1010 return ERR_PTR(-EFAULT);
1011 page = pfn_to_page(pfn);
1012 get_page(page);
1013 put_dev_pagemap(pgmap);
1014
1015 return page;
1016}
1017
1018int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1019 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1020 struct vm_area_struct *vma)
1021{
1022 spinlock_t *dst_ptl, *src_ptl;
1023 pud_t pud;
1024 int ret;
1025
1026 dst_ptl = pud_lock(dst_mm, dst_pud);
1027 src_ptl = pud_lockptr(src_mm, src_pud);
1028 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1029
1030 ret = -EAGAIN;
1031 pud = *src_pud;
1032 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1033 goto out_unlock;
1034
1035 /*
1036 * When page table lock is held, the huge zero pud should not be
1037 * under splitting since we don't split the page itself, only pud to
1038 * a page table.
1039 */
1040 if (is_huge_zero_pud(pud)) {
1041 /* No huge zero pud yet */
1042 }
1043
1044 pudp_set_wrprotect(src_mm, addr, src_pud);
1045 pud = pud_mkold(pud_wrprotect(pud));
1046 set_pud_at(dst_mm, addr, dst_pud, pud);
1047
1048 ret = 0;
1049out_unlock:
1050 spin_unlock(src_ptl);
1051 spin_unlock(dst_ptl);
1052 return ret;
1053}
1054
1055void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1056{
1057 pud_t entry;
1058 unsigned long haddr;
1059 bool write = vmf->flags & FAULT_FLAG_WRITE;
1060
1061 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1062 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1063 goto unlock;
1064
1065 entry = pud_mkyoung(orig_pud);
1066 if (write)
1067 entry = pud_mkdirty(entry);
1068 haddr = vmf->address & HPAGE_PUD_MASK;
1069 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1070 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1071
1072unlock:
1073 spin_unlock(vmf->ptl);
1074}
1075#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1076
Jan Kara82b0f8c2016-12-14 15:06:58 -08001077void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001078{
1079 pmd_t entry;
1080 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001081 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001082
Jan Kara82b0f8c2016-12-14 15:06:58 -08001083 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1084 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001085 goto unlock;
1086
1087 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001088 if (write)
1089 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001090 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001091 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001092 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001093
1094unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001095 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001096}
1097
Jan Kara82b0f8c2016-12-14 15:06:58 -08001098static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001099 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001100{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001101 struct vm_area_struct *vma = vmf->vma;
1102 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001103 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001104 pgtable_t pgtable;
1105 pmd_t _pmd;
1106 int ret = 0, i;
1107 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001108 unsigned long mmun_start; /* For mmu_notifiers */
1109 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001110
1111 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1112 GFP_KERNEL);
1113 if (unlikely(!pages)) {
1114 ret |= VM_FAULT_OOM;
1115 goto out;
1116 }
1117
1118 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001119 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001120 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001121 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001122 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1123 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001124 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001126 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001127 memcg = (void *)page_private(pages[i]);
1128 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001129 mem_cgroup_cancel_charge(pages[i], memcg,
1130 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001131 put_page(pages[i]);
1132 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001133 kfree(pages);
1134 ret |= VM_FAULT_OOM;
1135 goto out;
1136 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001137 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001138 }
1139
1140 for (i = 0; i < HPAGE_PMD_NR; i++) {
1141 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001142 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001143 __SetPageUptodate(pages[i]);
1144 cond_resched();
1145 }
1146
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001147 mmun_start = haddr;
1148 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001149 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001150
Jan Kara82b0f8c2016-12-14 15:06:58 -08001151 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1152 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001153 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001154 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001155
Jan Kara82b0f8c2016-12-14 15:06:58 -08001156 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001157 /* leave pmd empty until pte is filled */
1158
Jan Kara82b0f8c2016-12-14 15:06:58 -08001159 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001160 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001161
1162 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001163 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 entry = mk_pte(pages[i], vma->vm_page_prot);
1165 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001166 memcg = (void *)page_private(pages[i]);
1167 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001168 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001169 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001170 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001171 vmf->pte = pte_offset_map(&_pmd, haddr);
1172 VM_BUG_ON(!pte_none(*vmf->pte));
1173 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1174 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001175 }
1176 kfree(pages);
1177
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001178 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001179 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001180 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001181 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001182
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001183 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001184
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001185 ret |= VM_FAULT_WRITE;
1186 put_page(page);
1187
1188out:
1189 return ret;
1190
1191out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001192 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001193 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001194 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001195 memcg = (void *)page_private(pages[i]);
1196 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001197 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001198 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001199 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001200 kfree(pages);
1201 goto out;
1202}
1203
Jan Kara82b0f8c2016-12-14 15:06:58 -08001204int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001206 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001207 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001208 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001209 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001210 unsigned long mmun_start; /* For mmu_notifiers */
1211 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001212 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001213 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001214
Jan Kara82b0f8c2016-12-14 15:06:58 -08001215 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001216 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001217 if (is_huge_zero_pmd(orig_pmd))
1218 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001219 spin_lock(vmf->ptl);
1220 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001221 goto out_unlock;
1222
1223 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001224 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001225 /*
1226 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001227 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001228 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001229 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 pmd_t entry;
1231 entry = pmd_mkyoung(orig_pmd);
1232 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001233 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1234 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235 ret |= VM_FAULT_WRITE;
1236 goto out_unlock;
1237 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001238 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001239 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001240alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001241 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001242 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001243 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001244 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001245 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246 new_page = NULL;
1247
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001248 if (likely(new_page)) {
1249 prep_transhuge_page(new_page);
1250 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001251 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001252 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001253 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001254 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001255 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001256 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001257 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001258 ret |= VM_FAULT_FALLBACK;
1259 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001260 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001261 }
David Rientjes17766dd2013-09-12 15:14:06 -07001262 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001263 goto out;
1264 }
1265
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001266 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1267 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001268 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001269 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001270 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001271 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001272 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001273 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001274 goto out;
1275 }
1276
David Rientjes17766dd2013-09-12 15:14:06 -07001277 count_vm_event(THP_FAULT_ALLOC);
1278
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001279 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001280 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1281 else
1282 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001283 __SetPageUptodate(new_page);
1284
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001285 mmun_start = haddr;
1286 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001287 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001288
Jan Kara82b0f8c2016-12-14 15:06:58 -08001289 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001290 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001291 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001292 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1293 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001294 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001295 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001296 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001297 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001298 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001299 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1300 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001301 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001302 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001303 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001304 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001305 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1306 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001307 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001308 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001309 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001310 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001311 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001312 put_page(page);
1313 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001314 ret |= VM_FAULT_WRITE;
1315 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001316 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001317out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001318 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001319out:
1320 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001321out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001322 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001323 return ret;
1324}
1325
Keno Fischer8310d482017-01-24 15:17:48 -08001326/*
1327 * FOLL_FORCE can write to even unwritable pmd's, but only
1328 * after we've gone through a COW cycle and they are dirty.
1329 */
1330static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1331{
1332 return pmd_write(pmd) ||
1333 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1334}
1335
David Rientjesb676b292012-10-08 16:34:03 -07001336struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001337 unsigned long addr,
1338 pmd_t *pmd,
1339 unsigned int flags)
1340{
David Rientjesb676b292012-10-08 16:34:03 -07001341 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001342 struct page *page = NULL;
1343
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001344 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001345
Keno Fischer8310d482017-01-24 15:17:48 -08001346 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001347 goto out;
1348
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001349 /* Avoid dumping huge zero page */
1350 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1351 return ERR_PTR(-EFAULT);
1352
Mel Gorman2b4847e2013-12-18 17:08:32 -08001353 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001354 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001355 goto out;
1356
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001357 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001358 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001359 if (flags & FOLL_TOUCH)
1360 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001361 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001362 /*
1363 * We don't mlock() pte-mapped THPs. This way we can avoid
1364 * leaking mlocked pages into non-VM_LOCKED VMAs.
1365 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001366 * For anon THP:
1367 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001368 * In most cases the pmd is the only mapping of the page as we
1369 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1370 * writable private mappings in populate_vma_page_range().
1371 *
1372 * The only scenario when we have the page shared here is if we
1373 * mlocking read-only mapping shared over fork(). We skip
1374 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001375 *
1376 * For file THP:
1377 *
1378 * We can expect PageDoubleMap() to be stable under page lock:
1379 * for file pages we set it in page_add_file_rmap(), which
1380 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001381 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001382
1383 if (PageAnon(page) && compound_mapcount(page) != 1)
1384 goto skip_mlock;
1385 if (PageDoubleMap(page) || !page->mapping)
1386 goto skip_mlock;
1387 if (!trylock_page(page))
1388 goto skip_mlock;
1389 lru_add_drain();
1390 if (page->mapping && !PageDoubleMap(page))
1391 mlock_vma_page(page);
1392 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001393 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001394skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001395 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001396 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001397 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001398 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001399
1400out:
1401 return page;
1402}
1403
Mel Gormand10e63f2012-10-25 14:16:31 +02001404/* NUMA hinting page fault entry point for trans huge pmds */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001405int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001406{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001407 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001408 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001409 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001410 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001411 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001412 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001413 bool page_locked;
1414 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001415 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001416 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001417
Jan Kara82b0f8c2016-12-14 15:06:58 -08001418 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1419 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001420 goto out_unlock;
1421
Mel Gormande466bd2013-12-18 17:08:42 -08001422 /*
1423 * If there are potential migrations, wait for completion and retry
1424 * without disrupting NUMA hinting information. Do not relock and
1425 * check_same as the page may no longer be mapped.
1426 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001427 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1428 page = pmd_page(*vmf->pmd);
Mark Rutland3c226c62017-06-16 14:02:34 -07001429 if (!get_page_unless_zero(page))
1430 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001431 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001432 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001433 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001434 goto out;
1435 }
1436
Mel Gormand10e63f2012-10-25 14:16:31 +02001437 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001438 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001439 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001440 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001441 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001442 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001443 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001444 flags |= TNF_FAULT_LOCAL;
1445 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001446
Mel Gormanbea66fb2015-03-25 15:55:37 -07001447 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001448 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001449 flags |= TNF_NO_GROUP;
1450
1451 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001452 * Acquire the page lock to serialise THP migrations but avoid dropping
1453 * page_table_lock if at all possible
1454 */
Mel Gormanb8916632013-10-07 11:28:44 +01001455 page_locked = trylock_page(page);
1456 target_nid = mpol_misplaced(page, vma, haddr);
1457 if (target_nid == -1) {
1458 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001459 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001460 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001461 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001462
Mel Gormande466bd2013-12-18 17:08:42 -08001463 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001464 if (!page_locked) {
Mark Rutland3c226c62017-06-16 14:02:34 -07001465 page_nid = -1;
1466 if (!get_page_unless_zero(page))
1467 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001468 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001469 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001470 put_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001471 goto out;
1472 }
1473
Mel Gorman2b4847e2013-12-18 17:08:32 -08001474 /*
1475 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1476 * to serialises splits
1477 */
Mel Gormanb8916632013-10-07 11:28:44 +01001478 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001479 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001480 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001481
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001482 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001483 spin_lock(vmf->ptl);
1484 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001485 unlock_page(page);
1486 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001487 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001488 goto out_unlock;
1489 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001490
Mel Gormanc3a489c2013-12-18 17:08:38 -08001491 /* Bail if we fail to protect against THP splits for any reason */
1492 if (unlikely(!anon_vma)) {
1493 put_page(page);
1494 page_nid = -1;
1495 goto clear_pmdnuma;
1496 }
1497
Mel Gormana54a4072013-10-07 11:28:46 +01001498 /*
1499 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001500 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001501 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001502 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001503 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001504 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001505 if (migrated) {
1506 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001507 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001508 } else
1509 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001510
Mel Gorman8191acb2013-10-07 11:28:45 +01001511 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001512clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001513 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001514 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001515 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001516 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001517 if (was_writable)
1518 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001519 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1520 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001521 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001522out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001523 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001524
1525out:
1526 if (anon_vma)
1527 page_unlock_anon_vma_read(anon_vma);
1528
Mel Gorman8191acb2013-10-07 11:28:45 +01001529 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001530 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001531 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001532
Mel Gormand10e63f2012-10-25 14:16:31 +02001533 return 0;
1534}
1535
Huang Ying319904a2016-07-28 15:48:03 -07001536/*
1537 * Return true if we do MADV_FREE successfully on entire pmd page.
1538 * Otherwise, return false.
1539 */
1540bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001541 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001542{
1543 spinlock_t *ptl;
1544 pmd_t orig_pmd;
1545 struct page *page;
1546 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001547 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001548
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001549 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1550
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001551 ptl = pmd_trans_huge_lock(pmd, vma);
1552 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001553 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001554
1555 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001556 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001557 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001558
1559 page = pmd_page(orig_pmd);
1560 /*
1561 * If other processes are mapping this page, we couldn't discard
1562 * the page unless they all do MADV_FREE so let's skip the page.
1563 */
1564 if (page_mapcount(page) != 1)
1565 goto out;
1566
1567 if (!trylock_page(page))
1568 goto out;
1569
1570 /*
1571 * If user want to discard part-pages of THP, split it so MADV_FREE
1572 * will deactivate only them.
1573 */
1574 if (next - addr != HPAGE_PMD_SIZE) {
1575 get_page(page);
1576 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001577 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001578 unlock_page(page);
Kirill A. Shutemovbbf29ff2017-07-06 15:35:28 -07001579 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001580 goto out_unlocked;
1581 }
1582
1583 if (PageDirty(page))
1584 ClearPageDirty(page);
1585 unlock_page(page);
1586
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001587 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemov58ceeb62017-04-13 14:56:26 -07001588 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001589 orig_pmd = pmd_mkold(orig_pmd);
1590 orig_pmd = pmd_mkclean(orig_pmd);
1591
1592 set_pmd_at(mm, addr, pmd, orig_pmd);
1593 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1594 }
Shaohua Li802a3a92017-05-03 14:52:32 -07001595
1596 mark_page_lazyfree(page);
Huang Ying319904a2016-07-28 15:48:03 -07001597 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001598out:
1599 spin_unlock(ptl);
1600out_unlocked:
1601 return ret;
1602}
1603
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001604static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1605{
1606 pgtable_t pgtable;
1607
1608 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1609 pte_free(mm, pgtable);
1610 atomic_long_dec(&mm->nr_ptes);
1611}
1612
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001613int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001614 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001615{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001616 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001617 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001618
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001619 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1620
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001621 ptl = __pmd_trans_huge_lock(pmd, vma);
1622 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001623 return 0;
1624 /*
1625 * For architectures like ppc64 we look at deposited pgtable
1626 * when calling pmdp_huge_get_and_clear. So do the
1627 * pgtable_trans_huge_withdraw after finishing pmdp related
1628 * operations.
1629 */
1630 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1631 tlb->fullmm);
1632 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1633 if (vma_is_dax(vma)) {
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -07001634 if (arch_needs_pgtable_deposit())
1635 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001636 spin_unlock(ptl);
1637 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001638 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001639 } else if (is_huge_zero_pmd(orig_pmd)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001640 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001641 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001642 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001643 } else {
1644 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001645 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001646 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001647 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001648 if (PageAnon(page)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001649 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001650 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1651 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001652 if (arch_needs_pgtable_deposit())
1653 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001654 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1655 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001656 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001657 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001658 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001659 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001660}
1661
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001662#ifndef pmd_move_must_withdraw
1663static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1664 spinlock_t *old_pmd_ptl,
1665 struct vm_area_struct *vma)
1666{
1667 /*
1668 * With split pmd lock we also need to move preallocated
1669 * PTE page table if new_pmd is on different PMD page table.
1670 *
1671 * We also don't deposit and withdraw tables for file pages.
1672 */
1673 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1674}
1675#endif
1676
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001677bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001678 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001679 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001680{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001681 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001682 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001683 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001684 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001685
1686 if ((old_addr & ~HPAGE_PMD_MASK) ||
1687 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001688 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001689 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001690
1691 /*
1692 * The destination pmd shouldn't be established, free_pgtables()
1693 * should have release it.
1694 */
1695 if (WARN_ON(!pmd_none(*new_pmd))) {
1696 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001697 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001698 }
1699
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001700 /*
1701 * We don't have to worry about the ordering of src and dst
1702 * ptlocks because exclusive mmap_sem prevents deadlock.
1703 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001704 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1705 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001706 new_ptl = pmd_lockptr(mm, new_pmd);
1707 if (new_ptl != old_ptl)
1708 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001709 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001710 if (pmd_present(pmd) && pmd_dirty(pmd))
1711 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001712 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001713
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001714 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301715 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001716 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1717 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001718 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301719 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1720 if (new_ptl != old_ptl)
1721 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001722 if (force_flush)
1723 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1724 else
1725 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001726 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001727 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001728 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001729 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001730}
1731
Mel Gormanf123d742013-10-07 11:28:49 +01001732/*
1733 * Returns
1734 * - 0 if PMD could not be locked
1735 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1736 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1737 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001738int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001739 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001740{
1741 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001742 spinlock_t *ptl;
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001743 pmd_t entry;
1744 bool preserve_write;
1745 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001746
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001747 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001748 if (!ptl)
1749 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001750
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001751 preserve_write = prot_numa && pmd_write(*pmd);
1752 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001753
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001754 /*
1755 * Avoid trapping faults against the zero page. The read-only
1756 * data is likely to be read-cached on the local CPU and
1757 * local/remote hits to the zero page are not interesting.
1758 */
1759 if (prot_numa && is_huge_zero_pmd(*pmd))
1760 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001761
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001762 if (prot_numa && pmd_protnone(*pmd))
1763 goto unlock;
1764
Kirill A. Shutemovced10802017-04-13 14:56:20 -07001765 /*
1766 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1767 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1768 * which is also under down_read(mmap_sem):
1769 *
1770 * CPU0: CPU1:
1771 * change_huge_pmd(prot_numa=1)
1772 * pmdp_huge_get_and_clear_notify()
1773 * madvise_dontneed()
1774 * zap_pmd_range()
1775 * pmd_trans_huge(*pmd) == 0 (without ptl)
1776 * // skip the pmd
1777 * set_pmd_at();
1778 * // pmd is re-established
1779 *
1780 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1781 * which may break userspace.
1782 *
1783 * pmdp_invalidate() is required to make sure we don't miss
1784 * dirty/young flags set by hardware.
1785 */
1786 entry = *pmd;
1787 pmdp_invalidate(vma, addr, pmd);
1788
1789 /*
1790 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1791 * corrupt them.
1792 */
1793 if (pmd_dirty(*pmd))
1794 entry = pmd_mkdirty(entry);
1795 if (pmd_young(*pmd))
1796 entry = pmd_mkyoung(entry);
1797
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001798 entry = pmd_modify(entry, newprot);
1799 if (preserve_write)
1800 entry = pmd_mk_savedwrite(entry);
1801 ret = HPAGE_PMD_NR;
1802 set_pmd_at(mm, addr, pmd, entry);
1803 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1804unlock:
1805 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001806 return ret;
1807}
1808
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001809/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001810 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001811 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001812 * Note that if it returns page table lock pointer, this routine returns without
1813 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001814 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001815spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001816{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001817 spinlock_t *ptl;
1818 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001819 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001820 return ptl;
1821 spin_unlock(ptl);
1822 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001823}
1824
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001825/*
1826 * Returns true if a given pud maps a thp, false otherwise.
1827 *
1828 * Note that if it returns true, this routine returns without unlocking page
1829 * table lock. So callers must unlock it.
1830 */
1831spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1832{
1833 spinlock_t *ptl;
1834
1835 ptl = pud_lock(vma->vm_mm, pud);
1836 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1837 return ptl;
1838 spin_unlock(ptl);
1839 return NULL;
1840}
1841
1842#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1843int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1844 pud_t *pud, unsigned long addr)
1845{
1846 pud_t orig_pud;
1847 spinlock_t *ptl;
1848
1849 ptl = __pud_trans_huge_lock(pud, vma);
1850 if (!ptl)
1851 return 0;
1852 /*
1853 * For architectures like ppc64 we look at deposited pgtable
1854 * when calling pudp_huge_get_and_clear. So do the
1855 * pgtable_trans_huge_withdraw after finishing pudp related
1856 * operations.
1857 */
1858 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1859 tlb->fullmm);
1860 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1861 if (vma_is_dax(vma)) {
1862 spin_unlock(ptl);
1863 /* No zero page support yet */
1864 } else {
1865 /* No support for anonymous PUD pages yet */
1866 BUG();
1867 }
1868 return 1;
1869}
1870
1871static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1872 unsigned long haddr)
1873{
1874 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1875 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1876 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1877 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1878
Yisheng Xiece9311c2017-03-09 16:17:00 -08001879 count_vm_event(THP_SPLIT_PUD);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001880
1881 pudp_huge_clear_flush_notify(vma, haddr, pud);
1882}
1883
1884void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1885 unsigned long address)
1886{
1887 spinlock_t *ptl;
1888 struct mm_struct *mm = vma->vm_mm;
1889 unsigned long haddr = address & HPAGE_PUD_MASK;
1890
1891 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
1892 ptl = pud_lock(mm, pud);
1893 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1894 goto out;
1895 __split_huge_pud_locked(vma, pud, haddr);
1896
1897out:
1898 spin_unlock(ptl);
1899 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
1900}
1901#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1902
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001903static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1904 unsigned long haddr, pmd_t *pmd)
1905{
1906 struct mm_struct *mm = vma->vm_mm;
1907 pgtable_t pgtable;
1908 pmd_t _pmd;
1909 int i;
1910
1911 /* leave pmd empty until pte is filled */
1912 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1913
1914 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1915 pmd_populate(mm, &_pmd, pgtable);
1916
1917 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1918 pte_t *pte, entry;
1919 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1920 entry = pte_mkspecial(entry);
1921 pte = pte_offset_map(&_pmd, haddr);
1922 VM_BUG_ON(!pte_none(*pte));
1923 set_pte_at(mm, haddr, pte, entry);
1924 pte_unmap(pte);
1925 }
1926 smp_wmb(); /* make pte visible before pmd */
1927 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001928}
1929
1930static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001931 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001932{
1933 struct mm_struct *mm = vma->vm_mm;
1934 struct page *page;
1935 pgtable_t pgtable;
1936 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001937 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001938 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001939 int i;
1940
1941 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1942 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1943 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001944 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001945
1946 count_vm_event(THP_SPLIT_PMD);
1947
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001948 if (!vma_is_anonymous(vma)) {
1949 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001950 /*
1951 * We are going to unmap this huge page. So
1952 * just go ahead and zap it
1953 */
1954 if (arch_needs_pgtable_deposit())
1955 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001956 if (vma_is_dax(vma))
1957 return;
1958 page = pmd_page(_pmd);
1959 if (!PageReferenced(page) && pmd_young(_pmd))
1960 SetPageReferenced(page);
1961 page_remove_rmap(page, true);
1962 put_page(page);
1963 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001964 return;
1965 } else if (is_huge_zero_pmd(*pmd)) {
1966 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1967 }
1968
1969 page = pmd_page(*pmd);
1970 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001971 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001972 write = pmd_write(*pmd);
1973 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001974 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001975 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001976
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301977 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001978 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1979 pmd_populate(mm, &_pmd, pgtable);
1980
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001981 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001982 pte_t entry, *pte;
1983 /*
1984 * Note that NUMA hinting access restrictions are not
1985 * transferred to avoid any possibility of altering
1986 * permissions across VMAs.
1987 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001988 if (freeze) {
1989 swp_entry_t swp_entry;
1990 swp_entry = make_migration_entry(page + i, write);
1991 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001992 if (soft_dirty)
1993 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001994 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07001995 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001996 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001997 if (!write)
1998 entry = pte_wrprotect(entry);
1999 if (!young)
2000 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002001 if (soft_dirty)
2002 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002003 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002004 if (dirty)
2005 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002006 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002007 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002008 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002009 atomic_inc(&page[i]._mapcount);
2010 pte_unmap(pte);
2011 }
2012
2013 /*
2014 * Set PG_double_map before dropping compound_mapcount to avoid
2015 * false-negative page_mapped().
2016 */
2017 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2018 for (i = 0; i < HPAGE_PMD_NR; i++)
2019 atomic_inc(&page[i]._mapcount);
2020 }
2021
2022 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2023 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07002024 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002025 if (TestClearPageDoubleMap(page)) {
2026 /* No need in mapcount reference anymore */
2027 for (i = 0; i < HPAGE_PMD_NR; i++)
2028 atomic_dec(&page[i]._mapcount);
2029 }
2030 }
2031
2032 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002033 /*
2034 * Up to this point the pmd is present and huge and userland has the
2035 * whole access to the hugepage during the split (which happens in
2036 * place). If we overwrite the pmd with the not-huge version pointing
2037 * to the pte here (which of course we could if all CPUs were bug
2038 * free), userland could trigger a small page size TLB miss on the
2039 * small sized TLB while the hugepage TLB entry is still established in
2040 * the huge TLB. Some CPU doesn't like that.
2041 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2042 * 383 on page 93. Intel should be safe but is also warns that it's
2043 * only safe if the permission and cache attributes of the two entries
2044 * loaded in the two TLB is identical (which should be the case here).
2045 * But it is generally safer to never allow small and huge TLB entries
2046 * for the same virtual address to be loaded simultaneously. So instead
2047 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2048 * current pmd notpresent (atomically because here the pmd_trans_huge
2049 * and pmd_trans_splitting must remain set at all times on the pmd
2050 * until the split is complete for this pmd), then we flush the SMP TLB
2051 * and finally we write the non-huge version of the pmd entry with
2052 * pmd_populate.
2053 */
2054 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002055 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002056
2057 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002058 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002059 page_remove_rmap(page + i, false);
2060 put_page(page + i);
2061 }
2062 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002063}
2064
2065void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002066 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002067{
2068 spinlock_t *ptl;
2069 struct mm_struct *mm = vma->vm_mm;
2070 unsigned long haddr = address & HPAGE_PMD_MASK;
2071
2072 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2073 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002074
2075 /*
2076 * If caller asks to setup a migration entries, we need a page to check
2077 * pmd against. Otherwise we can end up replacing wrong page.
2078 */
2079 VM_BUG_ON(freeze && !page);
2080 if (page && page != pmd_page(*pmd))
2081 goto out;
2082
Dan Williams5c7fb562016-01-15 16:56:52 -08002083 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002084 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002085 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002086 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08002087 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002088 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002089 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002090out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002091 spin_unlock(ptl);
2092 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2093}
2094
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002095void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2096 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002097{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002098 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002099 p4d_t *p4d;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002100 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002101 pmd_t *pmd;
2102
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002103 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002104 if (!pgd_present(*pgd))
2105 return;
2106
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002107 p4d = p4d_offset(pgd, address);
2108 if (!p4d_present(*p4d))
2109 return;
2110
2111 pud = pud_offset(p4d, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002112 if (!pud_present(*pud))
2113 return;
2114
2115 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002116
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002117 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002118}
2119
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002120void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002121 unsigned long start,
2122 unsigned long end,
2123 long adjust_next)
2124{
2125 /*
2126 * If the new start address isn't hpage aligned and it could
2127 * previously contain an hugepage: check if we need to split
2128 * an huge pmd.
2129 */
2130 if (start & ~HPAGE_PMD_MASK &&
2131 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2132 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002133 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002134
2135 /*
2136 * If the new end address isn't hpage aligned and it could
2137 * previously contain an hugepage: check if we need to split
2138 * an huge pmd.
2139 */
2140 if (end & ~HPAGE_PMD_MASK &&
2141 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2142 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002143 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002144
2145 /*
2146 * If we're also updating the vma->vm_next->vm_start, if the new
2147 * vm_next->vm_start isn't page aligned and it could previously
2148 * contain an hugepage: check if we need to split an huge pmd.
2149 */
2150 if (adjust_next > 0) {
2151 struct vm_area_struct *next = vma->vm_next;
2152 unsigned long nstart = next->vm_start;
2153 nstart += adjust_next << PAGE_SHIFT;
2154 if (nstart & ~HPAGE_PMD_MASK &&
2155 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2156 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002157 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002158 }
2159}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002160
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002161static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002162{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002163 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002164 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
Minchan Kim666e5a42017-05-03 14:54:20 -07002165 bool unmap_success;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002166
2167 VM_BUG_ON_PAGE(!PageHead(page), page);
2168
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002169 if (PageAnon(page))
2170 ttu_flags |= TTU_MIGRATION;
2171
Minchan Kim666e5a42017-05-03 14:54:20 -07002172 unmap_success = try_to_unmap(page, ttu_flags);
2173 VM_BUG_ON_PAGE(!unmap_success, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002174}
2175
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002176static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002177{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002178 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002179 if (PageTransHuge(page)) {
2180 remove_migration_ptes(page, page, true);
2181 } else {
2182 for (i = 0; i < HPAGE_PMD_NR; i++)
2183 remove_migration_ptes(page + i, page + i, true);
2184 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002185}
2186
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002187static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002188 struct lruvec *lruvec, struct list_head *list)
2189{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002190 struct page *page_tail = head + tail;
2191
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002192 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002193 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002194
2195 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002196 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002197 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002198 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2199 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002200 * get_page_unless_zero(), and atomic_set() is implemented in C not
2201 * using locked ops. spin_unlock on x86 sometime uses locked ops
2202 * because of PPro errata 66, 92, so unless somebody can guarantee
2203 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002204 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002205 */
Huang Ying38d8b4e2017-07-06 15:37:18 -07002206 if (PageAnon(head) && !PageSwapCache(head)) {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002207 page_ref_inc(page_tail);
2208 } else {
2209 /* Additional pin to radix tree */
2210 page_ref_add(page_tail, 2);
2211 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002212
2213 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2214 page_tail->flags |= (head->flags &
2215 ((1L << PG_referenced) |
2216 (1L << PG_swapbacked) |
Huang Ying38d8b4e2017-07-06 15:37:18 -07002217 (1L << PG_swapcache) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002218 (1L << PG_mlocked) |
2219 (1L << PG_uptodate) |
2220 (1L << PG_active) |
2221 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002222 (1L << PG_unevictable) |
2223 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002224
2225 /*
2226 * After clearing PageTail the gup refcount can be released.
2227 * Page flags also must be visible before we make the page non-compound.
2228 */
2229 smp_wmb();
2230
2231 clear_compound_head(page_tail);
2232
2233 if (page_is_young(head))
2234 set_page_young(page_tail);
2235 if (page_is_idle(head))
2236 set_page_idle(page_tail);
2237
2238 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002239 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002240 page_tail);
2241 page_tail->mapping = head->mapping;
2242
2243 page_tail->index = head->index + tail;
2244 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2245 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002246}
2247
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002248static void __split_huge_page(struct page *page, struct list_head *list,
2249 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002250{
2251 struct page *head = compound_head(page);
2252 struct zone *zone = page_zone(head);
2253 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002254 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002255 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002256
Mel Gorman599d0c92016-07-28 15:45:31 -07002257 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002258
2259 /* complete memcg works before add pages to LRU */
2260 mem_cgroup_split_huge_fixup(head);
2261
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002262 if (!PageAnon(page))
2263 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2264
2265 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002266 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002267 /* Some pages can be beyond i_size: drop them from page cache */
2268 if (head[i].index >= end) {
2269 __ClearPageDirty(head + i);
2270 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002271 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2272 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002273 put_page(head + i);
2274 }
2275 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002276
2277 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002278 /* See comment in __split_huge_page_tail() */
2279 if (PageAnon(head)) {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002280 /* Additional pin to radix tree of swap cache */
2281 if (PageSwapCache(head))
2282 page_ref_add(head, 2);
2283 else
2284 page_ref_inc(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002285 } else {
2286 /* Additional pin to radix tree */
2287 page_ref_add(head, 2);
2288 spin_unlock(&head->mapping->tree_lock);
2289 }
2290
Mel Gormana52633d2016-07-28 15:45:28 -07002291 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002292
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002293 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002294
2295 for (i = 0; i < HPAGE_PMD_NR; i++) {
2296 struct page *subpage = head + i;
2297 if (subpage == page)
2298 continue;
2299 unlock_page(subpage);
2300
2301 /*
2302 * Subpages may be freed if there wasn't any mapping
2303 * like if add_to_swap() is running on a lru page that
2304 * had its mapping zapped. And freeing these pages
2305 * requires taking the lru_lock so we do the put_page
2306 * of the tail pages after the split is complete.
2307 */
2308 put_page(subpage);
2309 }
2310}
2311
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002312int total_mapcount(struct page *page)
2313{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002314 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002315
2316 VM_BUG_ON_PAGE(PageTail(page), page);
2317
2318 if (likely(!PageCompound(page)))
2319 return atomic_read(&page->_mapcount) + 1;
2320
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002321 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002322 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002323 return compound;
2324 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002325 for (i = 0; i < HPAGE_PMD_NR; i++)
2326 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002327 /* File pages has compound_mapcount included in _mapcount */
2328 if (!PageAnon(page))
2329 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002330 if (PageDoubleMap(page))
2331 ret -= HPAGE_PMD_NR;
2332 return ret;
2333}
2334
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002335/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002336 * This calculates accurately how many mappings a transparent hugepage
2337 * has (unlike page_mapcount() which isn't fully accurate). This full
2338 * accuracy is primarily needed to know if copy-on-write faults can
2339 * reuse the page and change the mapping to read-write instead of
2340 * copying them. At the same time this returns the total_mapcount too.
2341 *
2342 * The function returns the highest mapcount any one of the subpages
2343 * has. If the return value is one, even if different processes are
2344 * mapping different subpages of the transparent hugepage, they can
2345 * all reuse it, because each process is reusing a different subpage.
2346 *
2347 * The total_mapcount is instead counting all virtual mappings of the
2348 * subpages. If the total_mapcount is equal to "one", it tells the
2349 * caller all mappings belong to the same "mm" and in turn the
2350 * anon_vma of the transparent hugepage can become the vma->anon_vma
2351 * local one as no other process may be mapping any of the subpages.
2352 *
2353 * It would be more accurate to replace page_mapcount() with
2354 * page_trans_huge_mapcount(), however we only use
2355 * page_trans_huge_mapcount() in the copy-on-write faults where we
2356 * need full accuracy to avoid breaking page pinning, because
2357 * page_trans_huge_mapcount() is slower than page_mapcount().
2358 */
2359int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2360{
2361 int i, ret, _total_mapcount, mapcount;
2362
2363 /* hugetlbfs shouldn't call it */
2364 VM_BUG_ON_PAGE(PageHuge(page), page);
2365
2366 if (likely(!PageTransCompound(page))) {
2367 mapcount = atomic_read(&page->_mapcount) + 1;
2368 if (total_mapcount)
2369 *total_mapcount = mapcount;
2370 return mapcount;
2371 }
2372
2373 page = compound_head(page);
2374
2375 _total_mapcount = ret = 0;
2376 for (i = 0; i < HPAGE_PMD_NR; i++) {
2377 mapcount = atomic_read(&page[i]._mapcount) + 1;
2378 ret = max(ret, mapcount);
2379 _total_mapcount += mapcount;
2380 }
2381 if (PageDoubleMap(page)) {
2382 ret -= 1;
2383 _total_mapcount -= HPAGE_PMD_NR;
2384 }
2385 mapcount = compound_mapcount(page);
2386 ret += mapcount;
2387 _total_mapcount += mapcount;
2388 if (total_mapcount)
2389 *total_mapcount = _total_mapcount;
2390 return ret;
2391}
2392
Huang Yingb8f593c2017-07-06 15:37:28 -07002393/* Racy check whether the huge page can be split */
2394bool can_split_huge_page(struct page *page, int *pextra_pins)
2395{
2396 int extra_pins;
2397
2398 /* Additional pins from radix tree */
2399 if (PageAnon(page))
2400 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2401 else
2402 extra_pins = HPAGE_PMD_NR;
2403 if (pextra_pins)
2404 *pextra_pins = extra_pins;
2405 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2406}
2407
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002408/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002409 * This function splits huge page into normal pages. @page can point to any
2410 * subpage of huge page to split. Split doesn't change the position of @page.
2411 *
2412 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2413 * The huge page must be locked.
2414 *
2415 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2416 *
2417 * Both head page and tail pages will inherit mapping, flags, and so on from
2418 * the hugepage.
2419 *
2420 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2421 * they are not mapped.
2422 *
2423 * Returns 0 if the hugepage is split successfully.
2424 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2425 * us.
2426 */
2427int split_huge_page_to_list(struct page *page, struct list_head *list)
2428{
2429 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002430 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002431 struct anon_vma *anon_vma = NULL;
2432 struct address_space *mapping = NULL;
2433 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002434 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002435 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002436
2437 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002438 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002439 VM_BUG_ON_PAGE(!PageCompound(page), page);
2440
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002441 if (PageAnon(head)) {
2442 /*
2443 * The caller does not necessarily hold an mmap_sem that would
2444 * prevent the anon_vma disappearing so we first we take a
2445 * reference to it and then lock the anon_vma for write. This
2446 * is similar to page_lock_anon_vma_read except the write lock
2447 * is taken to serialise against parallel split or collapse
2448 * operations.
2449 */
2450 anon_vma = page_get_anon_vma(head);
2451 if (!anon_vma) {
2452 ret = -EBUSY;
2453 goto out;
2454 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002455 mapping = NULL;
2456 anon_vma_lock_write(anon_vma);
2457 } else {
2458 mapping = head->mapping;
2459
2460 /* Truncated ? */
2461 if (!mapping) {
2462 ret = -EBUSY;
2463 goto out;
2464 }
2465
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002466 anon_vma = NULL;
2467 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002468 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002469
2470 /*
2471 * Racy check if we can split the page, before freeze_page() will
2472 * split PMDs
2473 */
Huang Yingb8f593c2017-07-06 15:37:28 -07002474 if (!can_split_huge_page(head, &extra_pins)) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002475 ret = -EBUSY;
2476 goto out_unlock;
2477 }
2478
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002479 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002480 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002481 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2482
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002483 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2484 if (mlocked)
2485 lru_add_drain();
2486
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002487 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002488 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002489
2490 if (mapping) {
2491 void **pslot;
2492
2493 spin_lock(&mapping->tree_lock);
2494 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2495 page_index(head));
2496 /*
2497 * Check if the head page is present in radix tree.
2498 * We assume all tail are present too, if head is there.
2499 */
2500 if (radix_tree_deref_slot_protected(pslot,
2501 &mapping->tree_lock) != head)
2502 goto fail;
2503 }
2504
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002505 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002506 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002507 count = page_count(head);
2508 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002509 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002510 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002511 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002512 list_del(page_deferred_list(head));
2513 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002514 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002515 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002516 spin_unlock(&pgdata->split_queue_lock);
2517 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002518 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002519 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002520 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2521 pr_alert("total_mapcount: %u, page_count(): %u\n",
2522 mapcount, count);
2523 if (PageTail(page))
2524 dump_page(head, NULL);
2525 dump_page(page, "total_mapcount(head) > 0");
2526 BUG();
2527 }
2528 spin_unlock(&pgdata->split_queue_lock);
2529fail: if (mapping)
2530 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002531 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002532 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002533 ret = -EBUSY;
2534 }
2535
2536out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002537 if (anon_vma) {
2538 anon_vma_unlock_write(anon_vma);
2539 put_anon_vma(anon_vma);
2540 }
2541 if (mapping)
2542 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002543out:
2544 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2545 return ret;
2546}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002547
2548void free_transhuge_page(struct page *page)
2549{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002550 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002551 unsigned long flags;
2552
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002553 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002554 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002555 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002556 list_del(page_deferred_list(page));
2557 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002558 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002559 free_compound_page(page);
2560}
2561
2562void deferred_split_huge_page(struct page *page)
2563{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002564 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002565 unsigned long flags;
2566
2567 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2568
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002569 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002570 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002571 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002572 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2573 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002574 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002575 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002576}
2577
2578static unsigned long deferred_split_count(struct shrinker *shrink,
2579 struct shrink_control *sc)
2580{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002581 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002582 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002583}
2584
2585static unsigned long deferred_split_scan(struct shrinker *shrink,
2586 struct shrink_control *sc)
2587{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002588 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002589 unsigned long flags;
2590 LIST_HEAD(list), *pos, *next;
2591 struct page *page;
2592 int split = 0;
2593
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002594 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002595 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002596 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002597 page = list_entry((void *)pos, struct page, mapping);
2598 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002599 if (get_page_unless_zero(page)) {
2600 list_move(page_deferred_list(page), &list);
2601 } else {
2602 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002603 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002604 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002605 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002606 if (!--sc->nr_to_scan)
2607 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002608 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002609 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002610
2611 list_for_each_safe(pos, next, &list) {
2612 page = list_entry((void *)pos, struct page, mapping);
2613 lock_page(page);
2614 /* split_huge_page() removes page from list on success */
2615 if (!split_huge_page(page))
2616 split++;
2617 unlock_page(page);
2618 put_page(page);
2619 }
2620
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002621 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2622 list_splice_tail(&list, &pgdata->split_queue);
2623 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002624
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002625 /*
2626 * Stop shrinker if we didn't split any page, but the queue is empty.
2627 * This can happen if pages were freed under us.
2628 */
2629 if (!split && list_empty(&pgdata->split_queue))
2630 return SHRINK_STOP;
2631 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002632}
2633
2634static struct shrinker deferred_split_shrinker = {
2635 .count_objects = deferred_split_count,
2636 .scan_objects = deferred_split_scan,
2637 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002638 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002639};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002640
2641#ifdef CONFIG_DEBUG_FS
2642static int split_huge_pages_set(void *data, u64 val)
2643{
2644 struct zone *zone;
2645 struct page *page;
2646 unsigned long pfn, max_zone_pfn;
2647 unsigned long total = 0, split = 0;
2648
2649 if (val != 1)
2650 return -EINVAL;
2651
2652 for_each_populated_zone(zone) {
2653 max_zone_pfn = zone_end_pfn(zone);
2654 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2655 if (!pfn_valid(pfn))
2656 continue;
2657
2658 page = pfn_to_page(pfn);
2659 if (!get_page_unless_zero(page))
2660 continue;
2661
2662 if (zone != page_zone(page))
2663 goto next;
2664
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002665 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002666 goto next;
2667
2668 total++;
2669 lock_page(page);
2670 if (!split_huge_page(page))
2671 split++;
2672 unlock_page(page);
2673next:
2674 put_page(page);
2675 }
2676 }
2677
Yang Shi145bdaa2016-05-05 16:22:00 -07002678 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002679
2680 return 0;
2681}
2682DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2683 "%llu\n");
2684
2685static int __init split_huge_pages_debugfs(void)
2686{
2687 void *ret;
2688
Yang Shi145bdaa2016-05-05 16:22:00 -07002689 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002690 &split_huge_pages_fops);
2691 if (!ret)
2692 pr_warn("Failed to create split_huge_pages in debugfs");
2693 return 0;
2694}
2695late_initcall(split_huge_pages_debugfs);
2696#endif