blob: 269b5df58543e44d6a283c6268036f010aad37c2 [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>
Michal Hocko6b31d592017-08-18 15:16:15 -070035#include <linux/oom.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080036
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080037#include <asm/tlb.h>
38#include <asm/pgalloc.h>
39#include "internal.h"
40
Andrea Arcangeliba761492011-01-13 15:46:58 -080041/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080042 * By default transparent hugepage support is disabled in order that avoid
43 * to risk increase the memory footprint of applications without a guaranteed
44 * benefit. When transparent hugepage support is enabled, is for all mappings,
45 * and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080048 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080049unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080050#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080051 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080052#endif
53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070056 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080057 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080059
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080060static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080061
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080062static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080063struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080064
Aaron Lu6fcb52a2016-10-07 17:00:08 -070065static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080066{
67 struct page *zero_page;
68retry:
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070070 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080071
72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080074 if (!zero_page) {
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070076 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080077 }
78 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080079 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070080 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080081 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070082 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080083 goto retry;
84 }
85
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount, 2);
88 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070089 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080090}
91
Aaron Lu6fcb52a2016-10-07 17:00:08 -070092static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080093{
94 /*
95 * Counter should never go to zero here. Only shrinker can put
96 * last reference.
97 */
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99}
100
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700101struct page *mm_get_huge_zero_page(struct mm_struct *mm)
102{
103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104 return READ_ONCE(huge_zero_page);
105
106 if (!get_huge_zero_page())
107 return NULL;
108
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110 put_huge_zero_page();
111
112 return READ_ONCE(huge_zero_page);
113}
114
115void mm_put_huge_zero_page(struct mm_struct *mm)
116{
117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118 put_huge_zero_page();
119}
120
Glauber Costa48896462013-08-28 10:18:15 +1000121static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800123{
Glauber Costa48896462013-08-28 10:18:15 +1000124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800127
Glauber Costa48896462013-08-28 10:18:15 +1000128static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129 struct shrink_control *sc)
130{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700132 struct page *zero_page = xchg(&huge_zero_page, NULL);
133 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700134 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000135 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800136 }
137
138 return 0;
139}
140
141static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000142 .count_objects = shrink_huge_zero_page_count,
143 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800144 .seeks = DEFAULT_SEEKS,
145};
146
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800147#ifdef CONFIG_SYSFS
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800148static ssize_t enabled_show(struct kobject *kobj,
149 struct kobj_attribute *attr, char *buf)
150{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152 return sprintf(buf, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154 return sprintf(buf, "always [madvise] never\n");
155 else
156 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800157}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700158
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800159static ssize_t enabled_store(struct kobject *kobj,
160 struct kobj_attribute *attr,
161 const char *buf, size_t count)
162{
David Rientjes21440d72017-02-22 15:45:49 -0800163 ssize_t ret = count;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800164
David Rientjes21440d72017-02-22 15:45:49 -0800165 if (!memcmp("always", buf,
166 min(sizeof("always")-1, count))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 } else if (!memcmp("madvise", buf,
170 min(sizeof("madvise")-1, count))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173 } else if (!memcmp("never", buf,
174 min(sizeof("never")-1, count))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177 } else
178 ret = -EINVAL;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800179
180 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700181 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800182 if (err)
183 ret = err;
184 }
Andrea Arcangeliba761492011-01-13 15:46:58 -0800185 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800186}
187static struct kobj_attribute enabled_attr =
188 __ATTR(enabled, 0644, enabled_show, enabled_store);
189
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700190ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag flag)
193{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700194 return sprintf(buf, "%d\n",
195 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800196}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700197
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700198ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800199 struct kobj_attribute *attr,
200 const char *buf, size_t count,
201 enum transparent_hugepage_flag flag)
202{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700203 unsigned long value;
204 int ret;
205
206 ret = kstrtoul(buf, 10, &value);
207 if (ret < 0)
208 return ret;
209 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800210 return -EINVAL;
211
Ben Hutchingse27e6152011-04-14 15:22:21 -0700212 if (value)
213 set_bit(flag, &transparent_hugepage_flags);
214 else
215 clear_bit(flag, &transparent_hugepage_flags);
216
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800217 return count;
218}
219
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800220static ssize_t defrag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
222{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800224 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
Mel Gorman444eb2a42016-03-17 14:19:23 -0700225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800226 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800232}
David Rientjes21440d72017-02-22 15:45:49 -0800233
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800234static ssize_t defrag_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
237{
David Rientjes21440d72017-02-22 15:45:49 -0800238 if (!memcmp("always", buf,
239 min(sizeof("always")-1, count))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
David Rientjes21440d72017-02-22 15:45:49 -0800244 } else if (!memcmp("defer+madvise", buf,
245 min(sizeof("defer+madvise")-1, count))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
David Rientjes4fad7fb2017-04-07 16:04:54 -0700250 } else if (!memcmp("defer", buf,
251 min(sizeof("defer")-1, count))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
David Rientjes21440d72017-02-22 15:45:49 -0800256 } else if (!memcmp("madvise", buf,
257 min(sizeof("madvise")-1, count))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 } else if (!memcmp("never", buf,
263 min(sizeof("never")-1, count))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
270
271 return count;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800272}
273static struct kobj_attribute defrag_attr =
274 __ATTR(defrag, 0644, defrag_show, defrag_store);
275
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800276static ssize_t use_zero_page_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
278{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700279 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281}
282static ssize_t use_zero_page_store(struct kobject *kobj,
283 struct kobj_attribute *attr, const char *buf, size_t count)
284{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700285 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287}
288static struct kobj_attribute use_zero_page_attr =
289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Hugh Dickins49920d22016-12-12 16:44:50 -0800290
291static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
293{
294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
295}
296static struct kobj_attribute hpage_pmd_size_attr =
297 __ATTR_RO(hpage_pmd_size);
298
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800299#ifdef CONFIG_DEBUG_VM
300static ssize_t debug_cow_show(struct kobject *kobj,
301 struct kobj_attribute *attr, char *buf)
302{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700303 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305}
306static ssize_t debug_cow_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
309{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700310 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312}
313static struct kobj_attribute debug_cow_attr =
314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315#endif /* CONFIG_DEBUG_VM */
316
317static struct attribute *hugepage_attr[] = {
318 &enabled_attr.attr,
319 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800320 &use_zero_page_attr.attr,
Hugh Dickins49920d22016-12-12 16:44:50 -0800321 &hpage_pmd_size_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700322#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700323 &shmem_enabled_attr.attr,
324#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800325#ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr.attr,
327#endif
328 NULL,
329};
330
Arvind Yadav8aa95a22017-09-06 16:22:03 -0700331static const struct attribute_group hugepage_attr_group = {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800332 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800333};
334
Shaohua Li569e5592012-01-12 17:19:11 -0800335static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
336{
337 int err;
338
339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700341 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800342 return -ENOMEM;
343 }
344
345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
346 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700347 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800348 goto delete_obj;
349 }
350
351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
352 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700353 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800354 goto remove_hp_group;
355 }
356
357 return 0;
358
359remove_hp_group:
360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
361delete_obj:
362 kobject_put(*hugepage_kobj);
363 return err;
364}
365
366static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370 kobject_put(hugepage_kobj);
371}
372#else
373static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
374{
375 return 0;
376}
377
378static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379{
380}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800381#endif /* CONFIG_SYSFS */
382
383static int __init hugepage_init(void)
384{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800385 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800386 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800387
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800390 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800391 }
392
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530393 /*
394 * hugepages can't be allocated by the buddy allocator
395 */
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
397 /*
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
400 */
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
402
Shaohua Li569e5592012-01-12 17:19:11 -0800403 err = hugepage_init_sysfs(&hugepage_kobj);
404 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700405 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800406
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700407 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800408 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700409 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800410
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700411 err = register_shrinker(&huge_zero_page_shrinker);
412 if (err)
413 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800414 err = register_shrinker(&deferred_split_shrinker);
415 if (err)
416 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800417
Rik van Riel97562cd2011-01-13 15:47:12 -0800418 /*
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
422 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800424 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700425 return 0;
426 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800427
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700428 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700429 if (err)
430 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800431
Shaohua Li569e5592012-01-12 17:19:11 -0800432 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700433err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800434 unregister_shrinker(&deferred_split_shrinker);
435err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700436 unregister_shrinker(&huge_zero_page_shrinker);
437err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700438 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700439err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800440 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700441err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800442 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800443}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800444subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800445
446static int __init setup_transparent_hugepage(char *str)
447{
448 int ret = 0;
449 if (!str)
450 goto out;
451 if (!strcmp(str, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 } else if (!strcmp(str, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459 &transparent_hugepage_flags);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461 &transparent_hugepage_flags);
462 ret = 1;
463 } else if (!strcmp(str, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465 &transparent_hugepage_flags);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467 &transparent_hugepage_flags);
468 ret = 1;
469 }
470out:
471 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800473 return ret;
474}
475__setup("transparent_hugepage=", setup_transparent_hugepage);
476
Mel Gormanb32967f2012-11-19 12:35:47 +0000477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800478{
479 if (likely(vma->vm_flags & VM_WRITE))
480 pmd = pmd_mkwrite(pmd);
481 return pmd;
482}
483
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800484static inline struct list_head *page_deferred_list(struct page *page)
485{
486 /*
487 * ->lru in the tail pages is occupied by compound_head.
488 * Let's use ->mapping + ->index in the second tail page as list_head.
489 */
490 return (struct list_head *)&page[2].mapping;
491}
492
493void prep_transhuge_page(struct page *page)
494{
495 /*
496 * we use page->mapping and page->indexlru in second tail page
497 * as list_head: assuming THP order >= 2
498 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800499
500 INIT_LIST_HEAD(page_deferred_list(page));
501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
502}
503
Toshi Kani74d2fad2016-10-07 16:59:56 -0700504unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
505 loff_t off, unsigned long flags, unsigned long size)
506{
507 unsigned long addr;
508 loff_t off_end = off + len;
509 loff_t off_align = round_up(off, size);
510 unsigned long len_pad;
511
512 if (off_end <= off_align || (off_end - off_align) < size)
513 return 0;
514
515 len_pad = len + size;
516 if (len_pad < len || (off + len_pad) < off)
517 return 0;
518
519 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
520 off >> PAGE_SHIFT, flags);
521 if (IS_ERR_VALUE(addr))
522 return 0;
523
524 addr += (off - addr) & (size - 1);
525 return addr;
526}
527
528unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
529 unsigned long len, unsigned long pgoff, unsigned long flags)
530{
531 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
532
533 if (addr)
534 goto out;
535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
536 goto out;
537
538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
539 if (addr)
540 return addr;
541
542 out:
543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
544}
545EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
546
Jan Kara82b0f8c2016-12-14 15:06:58 -0800547static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700548 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800549{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800550 struct vm_area_struct *vma = vmf->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700551 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800552 pgtable_t pgtable;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Michal Hocko6b31d592017-08-18 15:16:15 -0700554 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800555
Sasha Levin309381fea2014-01-23 15:52:54 -0800556 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700557
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700559 put_page(page);
560 count_vm_event(THP_FAULT_FALLBACK);
561 return VM_FAULT_FALLBACK;
562 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800563
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700564 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700565 if (unlikely(!pgtable)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700566 ret = VM_FAULT_OOM;
567 goto release;
Johannes Weiner00501b52014-08-08 14:19:20 -0700568 }
569
Huang Yingc79b57e2017-09-06 16:25:04 -0700570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700571 /*
572 * The memory barrier inside __SetPageUptodate makes sure that
573 * clear_huge_page writes become visible before the set_pmd_at()
574 * write.
575 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800576 __SetPageUptodate(page);
577
Jan Kara82b0f8c2016-12-14 15:06:58 -0800578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
579 if (unlikely(!pmd_none(*vmf->pmd))) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700580 goto unlock_release;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800581 } else {
582 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700583
Michal Hocko6b31d592017-08-18 15:16:15 -0700584 ret = check_stable_address_space(vma->vm_mm);
585 if (ret)
586 goto unlock_release;
587
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700588 /* Deliver the page fault to userland */
589 if (userfaultfd_missing(vma)) {
590 int ret;
591
Jan Kara82b0f8c2016-12-14 15:06:58 -0800592 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800593 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700594 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700595 pte_free(vma->vm_mm, pgtable);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800596 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700597 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
598 return ret;
599 }
600
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700601 entry = mk_huge_pmd(page, vma->vm_page_prot);
602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800603 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800604 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700605 lru_cache_add_active_or_unevictable(page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
609 atomic_long_inc(&vma->vm_mm->nr_ptes);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800610 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700611 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800612 }
613
David Rientjesaa2e8782012-05-29 15:06:17 -0700614 return 0;
Michal Hocko6b31d592017-08-18 15:16:15 -0700615unlock_release:
616 spin_unlock(vmf->ptl);
617release:
618 if (pgtable)
619 pte_free(vma->vm_mm, pgtable);
620 mem_cgroup_cancel_charge(page, memcg, true);
621 put_page(page);
622 return ret;
623
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800624}
625
Mel Gorman444eb2a42016-03-17 14:19:23 -0700626/*
David Rientjes21440d72017-02-22 15:45:49 -0800627 * always: directly stall for all thp allocations
628 * defer: wake kswapd and fail if not immediately available
629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
630 * fail if not immediately available
631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
632 * available
633 * never: never stall for any thp allocation
Mel Gorman444eb2a42016-03-17 14:19:23 -0700634 */
635static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800636{
David Rientjes21440d72017-02-22 15:45:49 -0800637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700638
David Rientjes21440d72017-02-22 15:45:49 -0800639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
Vlastimil Babka25160352016-07-28 15:49:25 -0700640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
David Rientjes21440d72017-02-22 15:45:49 -0800641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645 __GFP_KSWAPD_RECLAIM);
646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
648 0);
Vlastimil Babka25160352016-07-28 15:49:25 -0700649 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700650}
651
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800652/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700653static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700655 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800656{
657 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700658 if (!pmd_none(*pmd))
659 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700660 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800661 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800662 if (pgtable)
663 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800664 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800665 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700666 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800667}
668
Jan Kara82b0f8c2016-12-14 15:06:58 -0800669int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800670{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800671 struct vm_area_struct *vma = vmf->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800672 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800673 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800675
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700677 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700678 if (unlikely(anon_vma_prepare(vma)))
679 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700680 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700681 return VM_FAULT_OOM;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800682 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700683 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700684 transparent_hugepage_use_zero_page()) {
685 pgtable_t pgtable;
686 struct page *zero_page;
687 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700688 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700689 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700690 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800691 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700692 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700693 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700694 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700695 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700696 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700697 }
Jan Kara82b0f8c2016-12-14 15:06:58 -0800698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700699 ret = 0;
700 set = false;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800701 if (pmd_none(*vmf->pmd)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700702 ret = check_stable_address_space(vma->vm_mm);
703 if (ret) {
704 spin_unlock(vmf->ptl);
705 } else if (userfaultfd_missing(vma)) {
Jan Kara82b0f8c2016-12-14 15:06:58 -0800706 spin_unlock(vmf->ptl);
707 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700708 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
709 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700710 set_huge_zero_page(pgtable, vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -0800711 haddr, vmf->pmd, zero_page);
712 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700713 set = true;
714 }
715 } else
Jan Kara82b0f8c2016-12-14 15:06:58 -0800716 spin_unlock(vmf->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700717 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700718 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700719 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700721 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700723 if (unlikely(!page)) {
724 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700725 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700726 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800727 prep_transhuge_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800728 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800729}
730
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700731static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
733 pgtable_t pgtable)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700734{
735 struct mm_struct *mm = vma->vm_mm;
736 pmd_t entry;
737 spinlock_t *ptl;
738
739 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
741 if (pfn_t_devmap(pfn))
742 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800743 if (write) {
744 entry = pmd_mkyoung(pmd_mkdirty(entry));
745 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700746 }
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700747
748 if (pgtable) {
749 pgtable_trans_huge_deposit(mm, pmd, pgtable);
750 atomic_long_inc(&mm->nr_ptes);
751 }
752
Ross Zwisler01871e52016-01-15 16:56:02 -0800753 set_pmd_at(mm, addr, pmd, entry);
754 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700755 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700756}
757
758int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800759 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700760{
761 pgprot_t pgprot = vma->vm_page_prot;
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700762 pgtable_t pgtable = NULL;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700763 /*
764 * If we had pmd_special, we could avoid all these restrictions,
765 * but we need to be consistent with PTEs and architectures that
766 * can't support a 'special' bit.
767 */
768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
770 (VM_PFNMAP|VM_MIXEDMAP));
771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800772 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700773
774 if (addr < vma->vm_start || addr >= vma->vm_end)
775 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200776
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700777 if (arch_needs_pgtable_deposit()) {
778 pgtable = pte_alloc_one(vma->vm_mm, addr);
779 if (!pgtable)
780 return VM_FAULT_OOM;
781 }
782
Borislav Petkov308a0472016-10-26 19:43:43 +0200783 track_pfn_insert(vma, &pgprot, pfn);
784
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700786 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700787}
Dan Williamsdee41072016-05-14 12:20:44 -0700788EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700789
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800790#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
792{
793 if (likely(vma->vm_flags & VM_WRITE))
794 pud = pud_mkwrite(pud);
795 return pud;
796}
797
798static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
800{
801 struct mm_struct *mm = vma->vm_mm;
802 pud_t entry;
803 spinlock_t *ptl;
804
805 ptl = pud_lock(mm, pud);
806 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
807 if (pfn_t_devmap(pfn))
808 entry = pud_mkdevmap(entry);
809 if (write) {
810 entry = pud_mkyoung(pud_mkdirty(entry));
811 entry = maybe_pud_mkwrite(entry, vma);
812 }
813 set_pud_at(mm, addr, pud, entry);
814 update_mmu_cache_pud(vma, addr, pud);
815 spin_unlock(ptl);
816}
817
818int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
819 pud_t *pud, pfn_t pfn, bool write)
820{
821 pgprot_t pgprot = vma->vm_page_prot;
822 /*
823 * If we had pud_special, we could avoid all these restrictions,
824 * but we need to be consistent with PTEs and architectures that
825 * can't support a 'special' bit.
826 */
827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
829 (VM_PFNMAP|VM_MIXEDMAP));
830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
831 BUG_ON(!pfn_t_devmap(pfn));
832
833 if (addr < vma->vm_start || addr >= vma->vm_end)
834 return VM_FAULT_SIGBUS;
835
836 track_pfn_insert(vma, &pgprot, pfn);
837
838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
839 return VM_FAULT_NOPAGE;
840}
841EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
842#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
843
Dan Williams3565fce2016-01-15 16:56:55 -0800844static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
845 pmd_t *pmd)
846{
847 pmd_t _pmd;
848
849 /*
850 * We should set the dirty bit only for FOLL_WRITE but for now
851 * the dirty bit in the pmd is meaningless. And if the dirty
852 * bit will become meaningful and we'll only set it with
853 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
854 * set the young bit, instead of the current set_pmd_at.
855 */
856 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
857 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
858 pmd, _pmd, 1))
859 update_mmu_cache_pmd(vma, addr, pmd);
860}
861
862struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
863 pmd_t *pmd, int flags)
864{
865 unsigned long pfn = pmd_pfn(*pmd);
866 struct mm_struct *mm = vma->vm_mm;
867 struct dev_pagemap *pgmap;
868 struct page *page;
869
870 assert_spin_locked(pmd_lockptr(mm, pmd));
871
Keno Fischer8310d482017-01-24 15:17:48 -0800872 /*
873 * When we COW a devmap PMD entry, we split it into PTEs, so we should
874 * not be in this function with `flags & FOLL_COW` set.
875 */
876 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
877
Dan Williams3565fce2016-01-15 16:56:55 -0800878 if (flags & FOLL_WRITE && !pmd_write(*pmd))
879 return NULL;
880
881 if (pmd_present(*pmd) && pmd_devmap(*pmd))
882 /* pass */;
883 else
884 return NULL;
885
886 if (flags & FOLL_TOUCH)
887 touch_pmd(vma, addr, pmd);
888
889 /*
890 * device mapped pages can only be returned if the
891 * caller will manage the page reference count.
892 */
893 if (!(flags & FOLL_GET))
894 return ERR_PTR(-EEXIST);
895
896 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
897 pgmap = get_dev_pagemap(pfn, NULL);
898 if (!pgmap)
899 return ERR_PTR(-EFAULT);
900 page = pfn_to_page(pfn);
901 get_page(page);
902 put_dev_pagemap(pgmap);
903
904 return page;
905}
906
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800907int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
908 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
909 struct vm_area_struct *vma)
910{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800911 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800912 struct page *src_page;
913 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800914 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700915 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800916
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700917 /* Skip if can be re-fill on fault */
918 if (!vma_is_anonymous(vma))
919 return 0;
920
921 pgtable = pte_alloc_one(dst_mm, addr);
922 if (unlikely(!pgtable))
923 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800924
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800925 dst_ptl = pmd_lock(dst_mm, dst_pmd);
926 src_ptl = pmd_lockptr(src_mm, src_pmd);
927 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800928
929 ret = -EAGAIN;
930 pmd = *src_pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -0700931
932#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
933 if (unlikely(is_swap_pmd(pmd))) {
934 swp_entry_t entry = pmd_to_swp_entry(pmd);
935
936 VM_BUG_ON(!is_pmd_migration_entry(pmd));
937 if (is_write_migration_entry(entry)) {
938 make_migration_entry_read(&entry);
939 pmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -0700940 if (pmd_swp_soft_dirty(*src_pmd))
941 pmd = pmd_swp_mksoft_dirty(pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -0700942 set_pmd_at(src_mm, addr, src_pmd, pmd);
943 }
944 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
945 ret = 0;
946 goto out_unlock;
947 }
948#endif
949
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700950 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800951 pte_free(dst_mm, pgtable);
952 goto out_unlock;
953 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800954 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800955 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800956 * under splitting since we don't split the page itself, only pmd to
957 * a page table.
958 */
959 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700960 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800961 /*
962 * get_huge_zero_page() will never allocate a new page here,
963 * since we already have a zero page to copy. It just takes a
964 * reference.
965 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700966 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700967 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700968 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800969 ret = 0;
970 goto out_unlock;
971 }
Mel Gormande466bd2013-12-18 17:08:42 -0800972
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700973 src_page = pmd_page(pmd);
974 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
975 get_page(src_page);
976 page_dup_rmap(src_page, true);
977 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
978 atomic_long_inc(&dst_mm->nr_ptes);
979 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800980
981 pmdp_set_wrprotect(src_mm, addr, src_pmd);
982 pmd = pmd_mkold(pmd_wrprotect(pmd));
983 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800984
985 ret = 0;
986out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800987 spin_unlock(src_ptl);
988 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800989out:
990 return ret;
991}
992
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800993#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
994static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
995 pud_t *pud)
996{
997 pud_t _pud;
998
999 /*
1000 * We should set the dirty bit only for FOLL_WRITE but for now
1001 * the dirty bit in the pud is meaningless. And if the dirty
1002 * bit will become meaningful and we'll only set it with
1003 * FOLL_WRITE, an atomic set_bit will be required on the pud to
1004 * set the young bit, instead of the current set_pud_at.
1005 */
1006 _pud = pud_mkyoung(pud_mkdirty(*pud));
1007 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1008 pud, _pud, 1))
1009 update_mmu_cache_pud(vma, addr, pud);
1010}
1011
1012struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1013 pud_t *pud, int flags)
1014{
1015 unsigned long pfn = pud_pfn(*pud);
1016 struct mm_struct *mm = vma->vm_mm;
1017 struct dev_pagemap *pgmap;
1018 struct page *page;
1019
1020 assert_spin_locked(pud_lockptr(mm, pud));
1021
1022 if (flags & FOLL_WRITE && !pud_write(*pud))
1023 return NULL;
1024
1025 if (pud_present(*pud) && pud_devmap(*pud))
1026 /* pass */;
1027 else
1028 return NULL;
1029
1030 if (flags & FOLL_TOUCH)
1031 touch_pud(vma, addr, pud);
1032
1033 /*
1034 * device mapped pages can only be returned if the
1035 * caller will manage the page reference count.
1036 */
1037 if (!(flags & FOLL_GET))
1038 return ERR_PTR(-EEXIST);
1039
1040 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1041 pgmap = get_dev_pagemap(pfn, NULL);
1042 if (!pgmap)
1043 return ERR_PTR(-EFAULT);
1044 page = pfn_to_page(pfn);
1045 get_page(page);
1046 put_dev_pagemap(pgmap);
1047
1048 return page;
1049}
1050
1051int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1052 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1053 struct vm_area_struct *vma)
1054{
1055 spinlock_t *dst_ptl, *src_ptl;
1056 pud_t pud;
1057 int ret;
1058
1059 dst_ptl = pud_lock(dst_mm, dst_pud);
1060 src_ptl = pud_lockptr(src_mm, src_pud);
1061 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1062
1063 ret = -EAGAIN;
1064 pud = *src_pud;
1065 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1066 goto out_unlock;
1067
1068 /*
1069 * When page table lock is held, the huge zero pud should not be
1070 * under splitting since we don't split the page itself, only pud to
1071 * a page table.
1072 */
1073 if (is_huge_zero_pud(pud)) {
1074 /* No huge zero pud yet */
1075 }
1076
1077 pudp_set_wrprotect(src_mm, addr, src_pud);
1078 pud = pud_mkold(pud_wrprotect(pud));
1079 set_pud_at(dst_mm, addr, dst_pud, pud);
1080
1081 ret = 0;
1082out_unlock:
1083 spin_unlock(src_ptl);
1084 spin_unlock(dst_ptl);
1085 return ret;
1086}
1087
1088void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1089{
1090 pud_t entry;
1091 unsigned long haddr;
1092 bool write = vmf->flags & FAULT_FLAG_WRITE;
1093
1094 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1095 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1096 goto unlock;
1097
1098 entry = pud_mkyoung(orig_pud);
1099 if (write)
1100 entry = pud_mkdirty(entry);
1101 haddr = vmf->address & HPAGE_PUD_MASK;
1102 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1103 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1104
1105unlock:
1106 spin_unlock(vmf->ptl);
1107}
1108#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1109
Jan Kara82b0f8c2016-12-14 15:06:58 -08001110void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001111{
1112 pmd_t entry;
1113 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001114 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001115
Jan Kara82b0f8c2016-12-14 15:06:58 -08001116 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1117 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001118 goto unlock;
1119
1120 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001121 if (write)
1122 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001123 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001124 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001125 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001126
1127unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001128 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001129}
1130
Jan Kara82b0f8c2016-12-14 15:06:58 -08001131static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001132 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001133{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001134 struct vm_area_struct *vma = vmf->vma;
1135 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001136 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001137 pgtable_t pgtable;
1138 pmd_t _pmd;
1139 int ret = 0, i;
1140 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001141 unsigned long mmun_start; /* For mmu_notifiers */
1142 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001143
1144 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1145 GFP_KERNEL);
1146 if (unlikely(!pages)) {
1147 ret |= VM_FAULT_OOM;
1148 goto out;
1149 }
1150
1151 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001152 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001153 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001154 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001155 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1156 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001157 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001158 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001159 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001160 memcg = (void *)page_private(pages[i]);
1161 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001162 mem_cgroup_cancel_charge(pages[i], memcg,
1163 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001164 put_page(pages[i]);
1165 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001166 kfree(pages);
1167 ret |= VM_FAULT_OOM;
1168 goto out;
1169 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001170 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001171 }
1172
1173 for (i = 0; i < HPAGE_PMD_NR; i++) {
1174 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001175 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001176 __SetPageUptodate(pages[i]);
1177 cond_resched();
1178 }
1179
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001180 mmun_start = haddr;
1181 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001182 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001183
Jan Kara82b0f8c2016-12-14 15:06:58 -08001184 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1185 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001186 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001187 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001188
Jan Kara82b0f8c2016-12-14 15:06:58 -08001189 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001190 /* leave pmd empty until pte is filled */
1191
Jan Kara82b0f8c2016-12-14 15:06:58 -08001192 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001193 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001194
1195 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001196 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001197 entry = mk_pte(pages[i], vma->vm_page_prot);
1198 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001201 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001202 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001203 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001204 vmf->pte = pte_offset_map(&_pmd, haddr);
1205 VM_BUG_ON(!pte_none(*vmf->pte));
1206 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001208 }
1209 kfree(pages);
1210
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001211 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001212 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001213 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001214 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001215
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001216 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001217
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001218 ret |= VM_FAULT_WRITE;
1219 put_page(page);
1220
1221out:
1222 return ret;
1223
1224out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001225 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001226 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001227 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001228 memcg = (void *)page_private(pages[i]);
1229 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001230 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001231 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001232 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 kfree(pages);
1234 goto out;
1235}
1236
Jan Kara82b0f8c2016-12-14 15:06:58 -08001237int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001238{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001239 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001240 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001241 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001242 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001243 unsigned long mmun_start; /* For mmu_notifiers */
1244 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001245 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001246 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001247
Jan Kara82b0f8c2016-12-14 15:06:58 -08001248 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001249 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001250 if (is_huge_zero_pmd(orig_pmd))
1251 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001252 spin_lock(vmf->ptl);
1253 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001254 goto out_unlock;
1255
1256 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001257 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001258 /*
1259 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001260 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001261 */
Huang Yingba3c4ce2017-09-06 16:22:19 -07001262 if (!trylock_page(page)) {
1263 get_page(page);
1264 spin_unlock(vmf->ptl);
1265 lock_page(page);
1266 spin_lock(vmf->ptl);
1267 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1268 unlock_page(page);
1269 put_page(page);
1270 goto out_unlock;
1271 }
1272 put_page(page);
1273 }
1274 if (reuse_swap_page(page, NULL)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001275 pmd_t entry;
1276 entry = pmd_mkyoung(orig_pmd);
1277 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001278 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1279 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001280 ret |= VM_FAULT_WRITE;
Huang Yingba3c4ce2017-09-06 16:22:19 -07001281 unlock_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001282 goto out_unlock;
1283 }
Huang Yingba3c4ce2017-09-06 16:22:19 -07001284 unlock_page(page);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001285 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001286 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001287alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001288 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001289 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001290 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001291 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001292 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001293 new_page = NULL;
1294
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001295 if (likely(new_page)) {
1296 prep_transhuge_page(new_page);
1297 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001298 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001299 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001300 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001301 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001302 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001303 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001304 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001305 ret |= VM_FAULT_FALLBACK;
1306 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001307 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001308 }
David Rientjes17766dd2013-09-12 15:14:06 -07001309 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001310 goto out;
1311 }
1312
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001313 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1314 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001315 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001316 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001317 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001318 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001319 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001320 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001321 goto out;
1322 }
1323
David Rientjes17766dd2013-09-12 15:14:06 -07001324 count_vm_event(THP_FAULT_ALLOC);
1325
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001326 if (!page)
Huang Yingc79b57e2017-09-06 16:25:04 -07001327 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001328 else
1329 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001330 __SetPageUptodate(new_page);
1331
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001332 mmun_start = haddr;
1333 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001334 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001335
Jan Kara82b0f8c2016-12-14 15:06:58 -08001336 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001337 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001338 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001339 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1340 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001341 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001342 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001343 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001344 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001345 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001346 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1347 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001348 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001349 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001350 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001351 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001352 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1353 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001354 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001355 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001356 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001357 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001358 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001359 put_page(page);
1360 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001361 ret |= VM_FAULT_WRITE;
1362 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001363 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001364out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001365 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001366out:
1367 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001368out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001369 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001370 return ret;
1371}
1372
Keno Fischer8310d482017-01-24 15:17:48 -08001373/*
1374 * FOLL_FORCE can write to even unwritable pmd's, but only
1375 * after we've gone through a COW cycle and they are dirty.
1376 */
1377static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1378{
1379 return pmd_write(pmd) ||
1380 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1381}
1382
David Rientjesb676b292012-10-08 16:34:03 -07001383struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001384 unsigned long addr,
1385 pmd_t *pmd,
1386 unsigned int flags)
1387{
David Rientjesb676b292012-10-08 16:34:03 -07001388 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001389 struct page *page = NULL;
1390
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001391 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001392
Keno Fischer8310d482017-01-24 15:17:48 -08001393 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001394 goto out;
1395
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001396 /* Avoid dumping huge zero page */
1397 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1398 return ERR_PTR(-EFAULT);
1399
Mel Gorman2b4847e2013-12-18 17:08:32 -08001400 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001401 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001402 goto out;
1403
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001404 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001405 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001406 if (flags & FOLL_TOUCH)
1407 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001408 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001409 /*
1410 * We don't mlock() pte-mapped THPs. This way we can avoid
1411 * leaking mlocked pages into non-VM_LOCKED VMAs.
1412 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001413 * For anon THP:
1414 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001415 * In most cases the pmd is the only mapping of the page as we
1416 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1417 * writable private mappings in populate_vma_page_range().
1418 *
1419 * The only scenario when we have the page shared here is if we
1420 * mlocking read-only mapping shared over fork(). We skip
1421 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001422 *
1423 * For file THP:
1424 *
1425 * We can expect PageDoubleMap() to be stable under page lock:
1426 * for file pages we set it in page_add_file_rmap(), which
1427 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001428 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001429
1430 if (PageAnon(page) && compound_mapcount(page) != 1)
1431 goto skip_mlock;
1432 if (PageDoubleMap(page) || !page->mapping)
1433 goto skip_mlock;
1434 if (!trylock_page(page))
1435 goto skip_mlock;
1436 lru_add_drain();
1437 if (page->mapping && !PageDoubleMap(page))
1438 mlock_vma_page(page);
1439 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001440 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001441skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001442 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001443 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001444 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001445 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001446
1447out:
1448 return page;
1449}
1450
Mel Gormand10e63f2012-10-25 14:16:31 +02001451/* NUMA hinting page fault entry point for trans huge pmds */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001452int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001453{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001454 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001455 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001456 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001457 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001458 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001459 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001460 bool page_locked;
1461 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001462 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001463 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001464
Jan Kara82b0f8c2016-12-14 15:06:58 -08001465 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1466 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001467 goto out_unlock;
1468
Mel Gormande466bd2013-12-18 17:08:42 -08001469 /*
1470 * If there are potential migrations, wait for completion and retry
1471 * without disrupting NUMA hinting information. Do not relock and
1472 * check_same as the page may no longer be mapped.
1473 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001474 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1475 page = pmd_page(*vmf->pmd);
Mark Rutland3c226c62017-06-16 14:02:34 -07001476 if (!get_page_unless_zero(page))
1477 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001478 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001479 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001480 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001481 goto out;
1482 }
1483
Mel Gormand10e63f2012-10-25 14:16:31 +02001484 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001485 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001486 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001487 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001488 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001489 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001490 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001491 flags |= TNF_FAULT_LOCAL;
1492 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001493
Mel Gormanbea66fb2015-03-25 15:55:37 -07001494 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001495 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001496 flags |= TNF_NO_GROUP;
1497
1498 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001499 * Acquire the page lock to serialise THP migrations but avoid dropping
1500 * page_table_lock if at all possible
1501 */
Mel Gormanb8916632013-10-07 11:28:44 +01001502 page_locked = trylock_page(page);
1503 target_nid = mpol_misplaced(page, vma, haddr);
1504 if (target_nid == -1) {
1505 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001506 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001507 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001508 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001509
Mel Gormande466bd2013-12-18 17:08:42 -08001510 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001511 if (!page_locked) {
Mark Rutland3c226c62017-06-16 14:02:34 -07001512 page_nid = -1;
1513 if (!get_page_unless_zero(page))
1514 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001515 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001516 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001517 put_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001518 goto out;
1519 }
1520
Mel Gorman2b4847e2013-12-18 17:08:32 -08001521 /*
1522 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1523 * to serialises splits
1524 */
Mel Gormanb8916632013-10-07 11:28:44 +01001525 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001526 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001527 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001528
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001529 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001530 spin_lock(vmf->ptl);
1531 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001532 unlock_page(page);
1533 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001534 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001535 goto out_unlock;
1536 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001537
Mel Gormanc3a489c2013-12-18 17:08:38 -08001538 /* Bail if we fail to protect against THP splits for any reason */
1539 if (unlikely(!anon_vma)) {
1540 put_page(page);
1541 page_nid = -1;
1542 goto clear_pmdnuma;
1543 }
1544
Mel Gormana54a4072013-10-07 11:28:46 +01001545 /*
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001546 * Since we took the NUMA fault, we must have observed the !accessible
1547 * bit. Make sure all other CPUs agree with that, to avoid them
1548 * modifying the page we're about to migrate.
1549 *
1550 * Must be done under PTL such that we'll observe the relevant
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001551 * inc_tlb_flush_pending().
1552 *
1553 * We are not sure a pending tlb flush here is for a huge page
1554 * mapping or not. Hence use the tlb range variant
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001555 */
1556 if (mm_tlb_flush_pending(vma->vm_mm))
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001557 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001558
1559 /*
Mel Gormana54a4072013-10-07 11:28:46 +01001560 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001561 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001562 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001563 spin_unlock(vmf->ptl);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001564
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001565 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001566 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001567 if (migrated) {
1568 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001569 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001570 } else
1571 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001572
Mel Gorman8191acb2013-10-07 11:28:45 +01001573 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001574clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001575 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001576 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001577 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001578 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001579 if (was_writable)
1580 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001581 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1582 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001583 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001584out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001585 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001586
1587out:
1588 if (anon_vma)
1589 page_unlock_anon_vma_read(anon_vma);
1590
Mel Gorman8191acb2013-10-07 11:28:45 +01001591 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001592 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001593 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001594
Mel Gormand10e63f2012-10-25 14:16:31 +02001595 return 0;
1596}
1597
Huang Ying319904a2016-07-28 15:48:03 -07001598/*
1599 * Return true if we do MADV_FREE successfully on entire pmd page.
1600 * Otherwise, return false.
1601 */
1602bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001603 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001604{
1605 spinlock_t *ptl;
1606 pmd_t orig_pmd;
1607 struct page *page;
1608 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001609 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001610
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001611 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1612
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001613 ptl = pmd_trans_huge_lock(pmd, vma);
1614 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001615 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001616
1617 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001618 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001619 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001620
Zi Yan84c3fc42017-09-08 16:11:01 -07001621 if (unlikely(!pmd_present(orig_pmd))) {
1622 VM_BUG_ON(thp_migration_supported() &&
1623 !is_pmd_migration_entry(orig_pmd));
1624 goto out;
1625 }
1626
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001627 page = pmd_page(orig_pmd);
1628 /*
1629 * If other processes are mapping this page, we couldn't discard
1630 * the page unless they all do MADV_FREE so let's skip the page.
1631 */
1632 if (page_mapcount(page) != 1)
1633 goto out;
1634
1635 if (!trylock_page(page))
1636 goto out;
1637
1638 /*
1639 * If user want to discard part-pages of THP, split it so MADV_FREE
1640 * will deactivate only them.
1641 */
1642 if (next - addr != HPAGE_PMD_SIZE) {
1643 get_page(page);
1644 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001645 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001646 unlock_page(page);
Kirill A. Shutemovbbf29ff2017-07-06 15:35:28 -07001647 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001648 goto out_unlocked;
1649 }
1650
1651 if (PageDirty(page))
1652 ClearPageDirty(page);
1653 unlock_page(page);
1654
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001655 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemov58ceeb62017-04-13 14:56:26 -07001656 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001657 orig_pmd = pmd_mkold(orig_pmd);
1658 orig_pmd = pmd_mkclean(orig_pmd);
1659
1660 set_pmd_at(mm, addr, pmd, orig_pmd);
1661 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1662 }
Shaohua Li802a3a92017-05-03 14:52:32 -07001663
1664 mark_page_lazyfree(page);
Huang Ying319904a2016-07-28 15:48:03 -07001665 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001666out:
1667 spin_unlock(ptl);
1668out_unlocked:
1669 return ret;
1670}
1671
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001672static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1673{
1674 pgtable_t pgtable;
1675
1676 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1677 pte_free(mm, pgtable);
1678 atomic_long_dec(&mm->nr_ptes);
1679}
1680
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001681int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001682 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001683{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001684 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001685 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001686
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001687 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1688
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001689 ptl = __pmd_trans_huge_lock(pmd, vma);
1690 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001691 return 0;
1692 /*
1693 * For architectures like ppc64 we look at deposited pgtable
1694 * when calling pmdp_huge_get_and_clear. So do the
1695 * pgtable_trans_huge_withdraw after finishing pmdp related
1696 * operations.
1697 */
1698 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1699 tlb->fullmm);
1700 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1701 if (vma_is_dax(vma)) {
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -07001702 if (arch_needs_pgtable_deposit())
1703 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001704 spin_unlock(ptl);
1705 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001706 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001707 } else if (is_huge_zero_pmd(orig_pmd)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001708 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001709 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001710 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001711 } else {
Zi Yan616b8372017-09-08 16:10:57 -07001712 struct page *page = NULL;
1713 int flush_needed = 1;
1714
1715 if (pmd_present(orig_pmd)) {
1716 page = pmd_page(orig_pmd);
1717 page_remove_rmap(page, true);
1718 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1719 VM_BUG_ON_PAGE(!PageHead(page), page);
1720 } else if (thp_migration_supported()) {
1721 swp_entry_t entry;
1722
1723 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1724 entry = pmd_to_swp_entry(orig_pmd);
1725 page = pfn_to_page(swp_offset(entry));
1726 flush_needed = 0;
1727 } else
1728 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1729
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001730 if (PageAnon(page)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001731 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001732 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1733 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001734 if (arch_needs_pgtable_deposit())
1735 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001736 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1737 }
Zi Yan616b8372017-09-08 16:10:57 -07001738
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001739 spin_unlock(ptl);
Zi Yan616b8372017-09-08 16:10:57 -07001740 if (flush_needed)
1741 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001742 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001743 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001744}
1745
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001746#ifndef pmd_move_must_withdraw
1747static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1748 spinlock_t *old_pmd_ptl,
1749 struct vm_area_struct *vma)
1750{
1751 /*
1752 * With split pmd lock we also need to move preallocated
1753 * PTE page table if new_pmd is on different PMD page table.
1754 *
1755 * We also don't deposit and withdraw tables for file pages.
1756 */
1757 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1758}
1759#endif
1760
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001761static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1762{
1763#ifdef CONFIG_MEM_SOFT_DIRTY
1764 if (unlikely(is_pmd_migration_entry(pmd)))
1765 pmd = pmd_swp_mksoft_dirty(pmd);
1766 else if (pmd_present(pmd))
1767 pmd = pmd_mksoft_dirty(pmd);
1768#endif
1769 return pmd;
1770}
1771
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001772bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001773 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001774 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001775{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001776 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001777 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001778 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001779 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001780
1781 if ((old_addr & ~HPAGE_PMD_MASK) ||
1782 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001783 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001784 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001785
1786 /*
1787 * The destination pmd shouldn't be established, free_pgtables()
1788 * should have release it.
1789 */
1790 if (WARN_ON(!pmd_none(*new_pmd))) {
1791 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001792 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001793 }
1794
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001795 /*
1796 * We don't have to worry about the ordering of src and dst
1797 * ptlocks because exclusive mmap_sem prevents deadlock.
1798 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001799 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1800 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001801 new_ptl = pmd_lockptr(mm, new_pmd);
1802 if (new_ptl != old_ptl)
1803 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001804 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001805 if (pmd_present(pmd) && pmd_dirty(pmd))
1806 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001807 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001808
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001809 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301810 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001811 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1812 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001813 }
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001814 pmd = move_soft_dirty_pmd(pmd);
1815 set_pmd_at(mm, new_addr, new_pmd, pmd);
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301816 if (new_ptl != old_ptl)
1817 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001818 if (force_flush)
1819 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1820 else
1821 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001822 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001823 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001824 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001825 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001826}
1827
Mel Gormanf123d742013-10-07 11:28:49 +01001828/*
1829 * Returns
1830 * - 0 if PMD could not be locked
1831 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1832 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1833 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001834int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001835 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001836{
1837 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001838 spinlock_t *ptl;
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001839 pmd_t entry;
1840 bool preserve_write;
1841 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001842
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001843 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001844 if (!ptl)
1845 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001846
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001847 preserve_write = prot_numa && pmd_write(*pmd);
1848 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001849
Zi Yan84c3fc42017-09-08 16:11:01 -07001850#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1851 if (is_swap_pmd(*pmd)) {
1852 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1853
1854 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1855 if (is_write_migration_entry(entry)) {
1856 pmd_t newpmd;
1857 /*
1858 * A protection check is difficult so
1859 * just be safe and disable write
1860 */
1861 make_migration_entry_read(&entry);
1862 newpmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001863 if (pmd_swp_soft_dirty(*pmd))
1864 newpmd = pmd_swp_mksoft_dirty(newpmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001865 set_pmd_at(mm, addr, pmd, newpmd);
1866 }
1867 goto unlock;
1868 }
1869#endif
1870
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001871 /*
1872 * Avoid trapping faults against the zero page. The read-only
1873 * data is likely to be read-cached on the local CPU and
1874 * local/remote hits to the zero page are not interesting.
1875 */
1876 if (prot_numa && is_huge_zero_pmd(*pmd))
1877 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001878
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001879 if (prot_numa && pmd_protnone(*pmd))
1880 goto unlock;
1881
Kirill A. Shutemovced10802017-04-13 14:56:20 -07001882 /*
1883 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1884 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1885 * which is also under down_read(mmap_sem):
1886 *
1887 * CPU0: CPU1:
1888 * change_huge_pmd(prot_numa=1)
1889 * pmdp_huge_get_and_clear_notify()
1890 * madvise_dontneed()
1891 * zap_pmd_range()
1892 * pmd_trans_huge(*pmd) == 0 (without ptl)
1893 * // skip the pmd
1894 * set_pmd_at();
1895 * // pmd is re-established
1896 *
1897 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1898 * which may break userspace.
1899 *
1900 * pmdp_invalidate() is required to make sure we don't miss
1901 * dirty/young flags set by hardware.
1902 */
1903 entry = *pmd;
1904 pmdp_invalidate(vma, addr, pmd);
1905
1906 /*
1907 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1908 * corrupt them.
1909 */
1910 if (pmd_dirty(*pmd))
1911 entry = pmd_mkdirty(entry);
1912 if (pmd_young(*pmd))
1913 entry = pmd_mkyoung(entry);
1914
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001915 entry = pmd_modify(entry, newprot);
1916 if (preserve_write)
1917 entry = pmd_mk_savedwrite(entry);
1918 ret = HPAGE_PMD_NR;
1919 set_pmd_at(mm, addr, pmd, entry);
1920 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921unlock:
1922 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001923 return ret;
1924}
1925
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001926/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001927 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001928 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001929 * Note that if it returns page table lock pointer, this routine returns without
1930 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001931 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001932spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001933{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001934 spinlock_t *ptl;
1935 ptl = pmd_lock(vma->vm_mm, pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001936 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937 pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001938 return ptl;
1939 spin_unlock(ptl);
1940 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001941}
1942
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001943/*
1944 * Returns true if a given pud maps a thp, false otherwise.
1945 *
1946 * Note that if it returns true, this routine returns without unlocking page
1947 * table lock. So callers must unlock it.
1948 */
1949spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950{
1951 spinlock_t *ptl;
1952
1953 ptl = pud_lock(vma->vm_mm, pud);
1954 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955 return ptl;
1956 spin_unlock(ptl);
1957 return NULL;
1958}
1959
1960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962 pud_t *pud, unsigned long addr)
1963{
1964 pud_t orig_pud;
1965 spinlock_t *ptl;
1966
1967 ptl = __pud_trans_huge_lock(pud, vma);
1968 if (!ptl)
1969 return 0;
1970 /*
1971 * For architectures like ppc64 we look at deposited pgtable
1972 * when calling pudp_huge_get_and_clear. So do the
1973 * pgtable_trans_huge_withdraw after finishing pudp related
1974 * operations.
1975 */
1976 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977 tlb->fullmm);
1978 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979 if (vma_is_dax(vma)) {
1980 spin_unlock(ptl);
1981 /* No zero page support yet */
1982 } else {
1983 /* No support for anonymous PUD pages yet */
1984 BUG();
1985 }
1986 return 1;
1987}
1988
1989static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990 unsigned long haddr)
1991{
1992 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996
Yisheng Xiece9311c2017-03-09 16:17:00 -08001997 count_vm_event(THP_SPLIT_PUD);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001998
1999 pudp_huge_clear_flush_notify(vma, haddr, pud);
2000}
2001
2002void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003 unsigned long address)
2004{
2005 spinlock_t *ptl;
2006 struct mm_struct *mm = vma->vm_mm;
2007 unsigned long haddr = address & HPAGE_PUD_MASK;
2008
2009 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010 ptl = pud_lock(mm, pud);
2011 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012 goto out;
2013 __split_huge_pud_locked(vma, pud, haddr);
2014
2015out:
2016 spin_unlock(ptl);
2017 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
2018}
2019#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2020
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002021static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2022 unsigned long haddr, pmd_t *pmd)
2023{
2024 struct mm_struct *mm = vma->vm_mm;
2025 pgtable_t pgtable;
2026 pmd_t _pmd;
2027 int i;
2028
2029 /* leave pmd empty until pte is filled */
2030 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2031
2032 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2033 pmd_populate(mm, &_pmd, pgtable);
2034
2035 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2036 pte_t *pte, entry;
2037 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2038 entry = pte_mkspecial(entry);
2039 pte = pte_offset_map(&_pmd, haddr);
2040 VM_BUG_ON(!pte_none(*pte));
2041 set_pte_at(mm, haddr, pte, entry);
2042 pte_unmap(pte);
2043 }
2044 smp_wmb(); /* make pte visible before pmd */
2045 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002046}
2047
2048static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002049 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002050{
2051 struct mm_struct *mm = vma->vm_mm;
2052 struct page *page;
2053 pgtable_t pgtable;
2054 pmd_t _pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -07002055 bool young, write, dirty, soft_dirty, pmd_migration = false;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002056 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002057 int i;
2058
2059 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2060 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2061 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Zi Yan84c3fc42017-09-08 16:11:01 -07002062 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2063 && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002064
2065 count_vm_event(THP_SPLIT_PMD);
2066
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002067 if (!vma_is_anonymous(vma)) {
2068 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08002069 /*
2070 * We are going to unmap this huge page. So
2071 * just go ahead and zap it
2072 */
2073 if (arch_needs_pgtable_deposit())
2074 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002075 if (vma_is_dax(vma))
2076 return;
2077 page = pmd_page(_pmd);
2078 if (!PageReferenced(page) && pmd_young(_pmd))
2079 SetPageReferenced(page);
2080 page_remove_rmap(page, true);
2081 put_page(page);
2082 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002083 return;
2084 } else if (is_huge_zero_pmd(*pmd)) {
2085 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2086 }
2087
Zi Yan84c3fc42017-09-08 16:11:01 -07002088#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2089 pmd_migration = is_pmd_migration_entry(*pmd);
2090 if (pmd_migration) {
2091 swp_entry_t entry;
2092
2093 entry = pmd_to_swp_entry(*pmd);
2094 page = pfn_to_page(swp_offset(entry));
2095 } else
2096#endif
2097 page = pmd_page(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002098 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002099 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002100 write = pmd_write(*pmd);
2101 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002102 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002103 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002104
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05302105 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002106 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2107 pmd_populate(mm, &_pmd, pgtable);
2108
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002109 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002110 pte_t entry, *pte;
2111 /*
2112 * Note that NUMA hinting access restrictions are not
2113 * transferred to avoid any possibility of altering
2114 * permissions across VMAs.
2115 */
Zi Yan84c3fc42017-09-08 16:11:01 -07002116 if (freeze || pmd_migration) {
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002117 swp_entry_t swp_entry;
2118 swp_entry = make_migration_entry(page + i, write);
2119 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002120 if (soft_dirty)
2121 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002122 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07002123 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002124 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002125 if (!write)
2126 entry = pte_wrprotect(entry);
2127 if (!young)
2128 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002129 if (soft_dirty)
2130 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002131 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002132 if (dirty)
2133 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002134 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002135 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002136 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002137 atomic_inc(&page[i]._mapcount);
2138 pte_unmap(pte);
2139 }
2140
2141 /*
2142 * Set PG_double_map before dropping compound_mapcount to avoid
2143 * false-negative page_mapped().
2144 */
2145 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2146 for (i = 0; i < HPAGE_PMD_NR; i++)
2147 atomic_inc(&page[i]._mapcount);
2148 }
2149
2150 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2151 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07002152 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002153 if (TestClearPageDoubleMap(page)) {
2154 /* No need in mapcount reference anymore */
2155 for (i = 0; i < HPAGE_PMD_NR; i++)
2156 atomic_dec(&page[i]._mapcount);
2157 }
2158 }
2159
2160 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002161 /*
2162 * Up to this point the pmd is present and huge and userland has the
2163 * whole access to the hugepage during the split (which happens in
2164 * place). If we overwrite the pmd with the not-huge version pointing
2165 * to the pte here (which of course we could if all CPUs were bug
2166 * free), userland could trigger a small page size TLB miss on the
2167 * small sized TLB while the hugepage TLB entry is still established in
2168 * the huge TLB. Some CPU doesn't like that.
2169 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2170 * 383 on page 93. Intel should be safe but is also warns that it's
2171 * only safe if the permission and cache attributes of the two entries
2172 * loaded in the two TLB is identical (which should be the case here).
2173 * But it is generally safer to never allow small and huge TLB entries
2174 * for the same virtual address to be loaded simultaneously. So instead
2175 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2176 * current pmd notpresent (atomically because here the pmd_trans_huge
2177 * and pmd_trans_splitting must remain set at all times on the pmd
2178 * until the split is complete for this pmd), then we flush the SMP TLB
2179 * and finally we write the non-huge version of the pmd entry with
2180 * pmd_populate.
2181 */
2182 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002183 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002184
2185 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002186 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002187 page_remove_rmap(page + i, false);
2188 put_page(page + i);
2189 }
2190 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002191}
2192
2193void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002194 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002195{
2196 spinlock_t *ptl;
2197 struct mm_struct *mm = vma->vm_mm;
2198 unsigned long haddr = address & HPAGE_PMD_MASK;
2199
2200 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2201 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002202
2203 /*
2204 * If caller asks to setup a migration entries, we need a page to check
2205 * pmd against. Otherwise we can end up replacing wrong page.
2206 */
2207 VM_BUG_ON(freeze && !page);
2208 if (page && page != pmd_page(*pmd))
2209 goto out;
2210
Dan Williams5c7fb562016-01-15 16:56:52 -08002211 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002212 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002213 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002214 clear_page_mlock(page);
Zi Yan84c3fc42017-09-08 16:11:01 -07002215 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002216 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002217 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002218out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002219 spin_unlock(ptl);
2220 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2221}
2222
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002223void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2224 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002225{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002226 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002227 p4d_t *p4d;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002228 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002229 pmd_t *pmd;
2230
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002231 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002232 if (!pgd_present(*pgd))
2233 return;
2234
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002235 p4d = p4d_offset(pgd, address);
2236 if (!p4d_present(*p4d))
2237 return;
2238
2239 pud = pud_offset(p4d, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002240 if (!pud_present(*pud))
2241 return;
2242
2243 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002244
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002245 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002246}
2247
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002248void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002249 unsigned long start,
2250 unsigned long end,
2251 long adjust_next)
2252{
2253 /*
2254 * If the new start address isn't hpage aligned and it could
2255 * previously contain an hugepage: check if we need to split
2256 * an huge pmd.
2257 */
2258 if (start & ~HPAGE_PMD_MASK &&
2259 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2260 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002261 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002262
2263 /*
2264 * If the new end address isn't hpage aligned and it could
2265 * previously contain an hugepage: check if we need to split
2266 * an huge pmd.
2267 */
2268 if (end & ~HPAGE_PMD_MASK &&
2269 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2270 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002271 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002272
2273 /*
2274 * If we're also updating the vma->vm_next->vm_start, if the new
2275 * vm_next->vm_start isn't page aligned and it could previously
2276 * contain an hugepage: check if we need to split an huge pmd.
2277 */
2278 if (adjust_next > 0) {
2279 struct vm_area_struct *next = vma->vm_next;
2280 unsigned long nstart = next->vm_start;
2281 nstart += adjust_next << PAGE_SHIFT;
2282 if (nstart & ~HPAGE_PMD_MASK &&
2283 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2284 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002285 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002286 }
2287}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002288
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002289static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002290{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002291 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002292 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
Minchan Kim666e5a42017-05-03 14:54:20 -07002293 bool unmap_success;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002294
2295 VM_BUG_ON_PAGE(!PageHead(page), page);
2296
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002297 if (PageAnon(page))
Naoya Horiguchib5ff8162017-09-08 16:10:49 -07002298 ttu_flags |= TTU_SPLIT_FREEZE;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002299
Minchan Kim666e5a42017-05-03 14:54:20 -07002300 unmap_success = try_to_unmap(page, ttu_flags);
2301 VM_BUG_ON_PAGE(!unmap_success, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002302}
2303
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002304static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002305{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002306 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002307 if (PageTransHuge(page)) {
2308 remove_migration_ptes(page, page, true);
2309 } else {
2310 for (i = 0; i < HPAGE_PMD_NR; i++)
2311 remove_migration_ptes(page + i, page + i, true);
2312 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002313}
2314
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002315static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002316 struct lruvec *lruvec, struct list_head *list)
2317{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002318 struct page *page_tail = head + tail;
2319
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002320 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002321 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002322
2323 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002324 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002325 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002326 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2327 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002328 * get_page_unless_zero(), and atomic_set() is implemented in C not
2329 * using locked ops. spin_unlock on x86 sometime uses locked ops
2330 * because of PPro errata 66, 92, so unless somebody can guarantee
2331 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002332 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002333 */
Huang Ying38d8b4e2017-07-06 15:37:18 -07002334 if (PageAnon(head) && !PageSwapCache(head)) {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002335 page_ref_inc(page_tail);
2336 } else {
2337 /* Additional pin to radix tree */
2338 page_ref_add(page_tail, 2);
2339 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002340
2341 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2342 page_tail->flags |= (head->flags &
2343 ((1L << PG_referenced) |
2344 (1L << PG_swapbacked) |
Huang Ying38d8b4e2017-07-06 15:37:18 -07002345 (1L << PG_swapcache) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002346 (1L << PG_mlocked) |
2347 (1L << PG_uptodate) |
2348 (1L << PG_active) |
2349 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002350 (1L << PG_unevictable) |
2351 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002352
2353 /*
2354 * After clearing PageTail the gup refcount can be released.
2355 * Page flags also must be visible before we make the page non-compound.
2356 */
2357 smp_wmb();
2358
2359 clear_compound_head(page_tail);
2360
2361 if (page_is_young(head))
2362 set_page_young(page_tail);
2363 if (page_is_idle(head))
2364 set_page_idle(page_tail);
2365
2366 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002367 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002368 page_tail);
2369 page_tail->mapping = head->mapping;
2370
2371 page_tail->index = head->index + tail;
2372 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2373 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002374}
2375
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002376static void __split_huge_page(struct page *page, struct list_head *list,
2377 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002378{
2379 struct page *head = compound_head(page);
2380 struct zone *zone = page_zone(head);
2381 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002382 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002383 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002384
Mel Gorman599d0c92016-07-28 15:45:31 -07002385 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002386
2387 /* complete memcg works before add pages to LRU */
2388 mem_cgroup_split_huge_fixup(head);
2389
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002390 if (!PageAnon(page))
2391 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2392
2393 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002394 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002395 /* Some pages can be beyond i_size: drop them from page cache */
2396 if (head[i].index >= end) {
2397 __ClearPageDirty(head + i);
2398 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002399 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2400 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002401 put_page(head + i);
2402 }
2403 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002404
2405 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002406 /* See comment in __split_huge_page_tail() */
2407 if (PageAnon(head)) {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002408 /* Additional pin to radix tree of swap cache */
2409 if (PageSwapCache(head))
2410 page_ref_add(head, 2);
2411 else
2412 page_ref_inc(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002413 } else {
2414 /* Additional pin to radix tree */
2415 page_ref_add(head, 2);
2416 spin_unlock(&head->mapping->tree_lock);
2417 }
2418
Mel Gormana52633d2016-07-28 15:45:28 -07002419 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002420
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002421 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002422
2423 for (i = 0; i < HPAGE_PMD_NR; i++) {
2424 struct page *subpage = head + i;
2425 if (subpage == page)
2426 continue;
2427 unlock_page(subpage);
2428
2429 /*
2430 * Subpages may be freed if there wasn't any mapping
2431 * like if add_to_swap() is running on a lru page that
2432 * had its mapping zapped. And freeing these pages
2433 * requires taking the lru_lock so we do the put_page
2434 * of the tail pages after the split is complete.
2435 */
2436 put_page(subpage);
2437 }
2438}
2439
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002440int total_mapcount(struct page *page)
2441{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002442 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002443
2444 VM_BUG_ON_PAGE(PageTail(page), page);
2445
2446 if (likely(!PageCompound(page)))
2447 return atomic_read(&page->_mapcount) + 1;
2448
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002449 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002450 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002451 return compound;
2452 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002453 for (i = 0; i < HPAGE_PMD_NR; i++)
2454 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002455 /* File pages has compound_mapcount included in _mapcount */
2456 if (!PageAnon(page))
2457 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002458 if (PageDoubleMap(page))
2459 ret -= HPAGE_PMD_NR;
2460 return ret;
2461}
2462
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002463/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002464 * This calculates accurately how many mappings a transparent hugepage
2465 * has (unlike page_mapcount() which isn't fully accurate). This full
2466 * accuracy is primarily needed to know if copy-on-write faults can
2467 * reuse the page and change the mapping to read-write instead of
2468 * copying them. At the same time this returns the total_mapcount too.
2469 *
2470 * The function returns the highest mapcount any one of the subpages
2471 * has. If the return value is one, even if different processes are
2472 * mapping different subpages of the transparent hugepage, they can
2473 * all reuse it, because each process is reusing a different subpage.
2474 *
2475 * The total_mapcount is instead counting all virtual mappings of the
2476 * subpages. If the total_mapcount is equal to "one", it tells the
2477 * caller all mappings belong to the same "mm" and in turn the
2478 * anon_vma of the transparent hugepage can become the vma->anon_vma
2479 * local one as no other process may be mapping any of the subpages.
2480 *
2481 * It would be more accurate to replace page_mapcount() with
2482 * page_trans_huge_mapcount(), however we only use
2483 * page_trans_huge_mapcount() in the copy-on-write faults where we
2484 * need full accuracy to avoid breaking page pinning, because
2485 * page_trans_huge_mapcount() is slower than page_mapcount().
2486 */
2487int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2488{
2489 int i, ret, _total_mapcount, mapcount;
2490
2491 /* hugetlbfs shouldn't call it */
2492 VM_BUG_ON_PAGE(PageHuge(page), page);
2493
2494 if (likely(!PageTransCompound(page))) {
2495 mapcount = atomic_read(&page->_mapcount) + 1;
2496 if (total_mapcount)
2497 *total_mapcount = mapcount;
2498 return mapcount;
2499 }
2500
2501 page = compound_head(page);
2502
2503 _total_mapcount = ret = 0;
2504 for (i = 0; i < HPAGE_PMD_NR; i++) {
2505 mapcount = atomic_read(&page[i]._mapcount) + 1;
2506 ret = max(ret, mapcount);
2507 _total_mapcount += mapcount;
2508 }
2509 if (PageDoubleMap(page)) {
2510 ret -= 1;
2511 _total_mapcount -= HPAGE_PMD_NR;
2512 }
2513 mapcount = compound_mapcount(page);
2514 ret += mapcount;
2515 _total_mapcount += mapcount;
2516 if (total_mapcount)
2517 *total_mapcount = _total_mapcount;
2518 return ret;
2519}
2520
Huang Yingb8f593c2017-07-06 15:37:28 -07002521/* Racy check whether the huge page can be split */
2522bool can_split_huge_page(struct page *page, int *pextra_pins)
2523{
2524 int extra_pins;
2525
2526 /* Additional pins from radix tree */
2527 if (PageAnon(page))
2528 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2529 else
2530 extra_pins = HPAGE_PMD_NR;
2531 if (pextra_pins)
2532 *pextra_pins = extra_pins;
2533 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2534}
2535
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002536/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002537 * This function splits huge page into normal pages. @page can point to any
2538 * subpage of huge page to split. Split doesn't change the position of @page.
2539 *
2540 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2541 * The huge page must be locked.
2542 *
2543 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2544 *
2545 * Both head page and tail pages will inherit mapping, flags, and so on from
2546 * the hugepage.
2547 *
2548 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2549 * they are not mapped.
2550 *
2551 * Returns 0 if the hugepage is split successfully.
2552 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2553 * us.
2554 */
2555int split_huge_page_to_list(struct page *page, struct list_head *list)
2556{
2557 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002558 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002559 struct anon_vma *anon_vma = NULL;
2560 struct address_space *mapping = NULL;
2561 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002562 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002563 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002564
2565 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002566 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002567 VM_BUG_ON_PAGE(!PageCompound(page), page);
2568
Huang Ying59807682017-09-06 16:22:34 -07002569 if (PageWriteback(page))
2570 return -EBUSY;
2571
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002572 if (PageAnon(head)) {
2573 /*
2574 * The caller does not necessarily hold an mmap_sem that would
2575 * prevent the anon_vma disappearing so we first we take a
2576 * reference to it and then lock the anon_vma for write. This
2577 * is similar to page_lock_anon_vma_read except the write lock
2578 * is taken to serialise against parallel split or collapse
2579 * operations.
2580 */
2581 anon_vma = page_get_anon_vma(head);
2582 if (!anon_vma) {
2583 ret = -EBUSY;
2584 goto out;
2585 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002586 mapping = NULL;
2587 anon_vma_lock_write(anon_vma);
2588 } else {
2589 mapping = head->mapping;
2590
2591 /* Truncated ? */
2592 if (!mapping) {
2593 ret = -EBUSY;
2594 goto out;
2595 }
2596
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002597 anon_vma = NULL;
2598 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002599 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002600
2601 /*
2602 * Racy check if we can split the page, before freeze_page() will
2603 * split PMDs
2604 */
Huang Yingb8f593c2017-07-06 15:37:28 -07002605 if (!can_split_huge_page(head, &extra_pins)) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002606 ret = -EBUSY;
2607 goto out_unlock;
2608 }
2609
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002610 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002611 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002612 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2613
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002614 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2615 if (mlocked)
2616 lru_add_drain();
2617
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002618 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002619 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002620
2621 if (mapping) {
2622 void **pslot;
2623
2624 spin_lock(&mapping->tree_lock);
2625 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2626 page_index(head));
2627 /*
2628 * Check if the head page is present in radix tree.
2629 * We assume all tail are present too, if head is there.
2630 */
2631 if (radix_tree_deref_slot_protected(pslot,
2632 &mapping->tree_lock) != head)
2633 goto fail;
2634 }
2635
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002636 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002637 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002638 count = page_count(head);
2639 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002640 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002641 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002642 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002643 list_del(page_deferred_list(head));
2644 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002645 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002646 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002647 spin_unlock(&pgdata->split_queue_lock);
2648 __split_huge_page(page, list, flags);
Huang Ying59807682017-09-06 16:22:34 -07002649 if (PageSwapCache(head)) {
2650 swp_entry_t entry = { .val = page_private(head) };
2651
2652 ret = split_swap_cluster(entry);
2653 } else
2654 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002655 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002656 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2657 pr_alert("total_mapcount: %u, page_count(): %u\n",
2658 mapcount, count);
2659 if (PageTail(page))
2660 dump_page(head, NULL);
2661 dump_page(page, "total_mapcount(head) > 0");
2662 BUG();
2663 }
2664 spin_unlock(&pgdata->split_queue_lock);
2665fail: if (mapping)
2666 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002667 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002668 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002669 ret = -EBUSY;
2670 }
2671
2672out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002673 if (anon_vma) {
2674 anon_vma_unlock_write(anon_vma);
2675 put_anon_vma(anon_vma);
2676 }
2677 if (mapping)
2678 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002679out:
2680 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2681 return ret;
2682}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002683
2684void free_transhuge_page(struct page *page)
2685{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002686 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002687 unsigned long flags;
2688
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002689 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002690 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002691 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002692 list_del(page_deferred_list(page));
2693 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002694 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002695 free_compound_page(page);
2696}
2697
2698void deferred_split_huge_page(struct page *page)
2699{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002700 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002701 unsigned long flags;
2702
2703 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2704
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002705 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002706 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002707 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002708 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2709 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002710 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002711 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002712}
2713
2714static unsigned long deferred_split_count(struct shrinker *shrink,
2715 struct shrink_control *sc)
2716{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002717 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002718 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002719}
2720
2721static unsigned long deferred_split_scan(struct shrinker *shrink,
2722 struct shrink_control *sc)
2723{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002724 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002725 unsigned long flags;
2726 LIST_HEAD(list), *pos, *next;
2727 struct page *page;
2728 int split = 0;
2729
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002730 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002731 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002732 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002733 page = list_entry((void *)pos, struct page, mapping);
2734 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002735 if (get_page_unless_zero(page)) {
2736 list_move(page_deferred_list(page), &list);
2737 } else {
2738 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002739 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002740 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002741 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002742 if (!--sc->nr_to_scan)
2743 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002744 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002745 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002746
2747 list_for_each_safe(pos, next, &list) {
2748 page = list_entry((void *)pos, struct page, mapping);
2749 lock_page(page);
2750 /* split_huge_page() removes page from list on success */
2751 if (!split_huge_page(page))
2752 split++;
2753 unlock_page(page);
2754 put_page(page);
2755 }
2756
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002757 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2758 list_splice_tail(&list, &pgdata->split_queue);
2759 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002760
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002761 /*
2762 * Stop shrinker if we didn't split any page, but the queue is empty.
2763 * This can happen if pages were freed under us.
2764 */
2765 if (!split && list_empty(&pgdata->split_queue))
2766 return SHRINK_STOP;
2767 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002768}
2769
2770static struct shrinker deferred_split_shrinker = {
2771 .count_objects = deferred_split_count,
2772 .scan_objects = deferred_split_scan,
2773 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002774 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002775};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002776
2777#ifdef CONFIG_DEBUG_FS
2778static int split_huge_pages_set(void *data, u64 val)
2779{
2780 struct zone *zone;
2781 struct page *page;
2782 unsigned long pfn, max_zone_pfn;
2783 unsigned long total = 0, split = 0;
2784
2785 if (val != 1)
2786 return -EINVAL;
2787
2788 for_each_populated_zone(zone) {
2789 max_zone_pfn = zone_end_pfn(zone);
2790 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2791 if (!pfn_valid(pfn))
2792 continue;
2793
2794 page = pfn_to_page(pfn);
2795 if (!get_page_unless_zero(page))
2796 continue;
2797
2798 if (zone != page_zone(page))
2799 goto next;
2800
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002801 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002802 goto next;
2803
2804 total++;
2805 lock_page(page);
2806 if (!split_huge_page(page))
2807 split++;
2808 unlock_page(page);
2809next:
2810 put_page(page);
2811 }
2812 }
2813
Yang Shi145bdaa2016-05-05 16:22:00 -07002814 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002815
2816 return 0;
2817}
2818DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2819 "%llu\n");
2820
2821static int __init split_huge_pages_debugfs(void)
2822{
2823 void *ret;
2824
Yang Shi145bdaa2016-05-05 16:22:00 -07002825 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002826 &split_huge_pages_fops);
2827 if (!ret)
2828 pr_warn("Failed to create split_huge_pages in debugfs");
2829 return 0;
2830}
2831late_initcall(split_huge_pages_debugfs);
2832#endif
Zi Yan616b8372017-09-08 16:10:57 -07002833
2834#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2835void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2836 struct page *page)
2837{
2838 struct vm_area_struct *vma = pvmw->vma;
2839 struct mm_struct *mm = vma->vm_mm;
2840 unsigned long address = pvmw->address;
2841 pmd_t pmdval;
2842 swp_entry_t entry;
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002843 pmd_t pmdswp;
Zi Yan616b8372017-09-08 16:10:57 -07002844
2845 if (!(pvmw->pmd && !pvmw->pte))
2846 return;
2847
2848 mmu_notifier_invalidate_range_start(mm, address,
2849 address + HPAGE_PMD_SIZE);
2850
2851 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2852 pmdval = *pvmw->pmd;
2853 pmdp_invalidate(vma, address, pvmw->pmd);
2854 if (pmd_dirty(pmdval))
2855 set_page_dirty(page);
2856 entry = make_migration_entry(page, pmd_write(pmdval));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002857 pmdswp = swp_entry_to_pmd(entry);
2858 if (pmd_soft_dirty(pmdval))
2859 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2860 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
Zi Yan616b8372017-09-08 16:10:57 -07002861 page_remove_rmap(page, true);
2862 put_page(page);
2863
2864 mmu_notifier_invalidate_range_end(mm, address,
2865 address + HPAGE_PMD_SIZE);
2866}
2867
2868void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2869{
2870 struct vm_area_struct *vma = pvmw->vma;
2871 struct mm_struct *mm = vma->vm_mm;
2872 unsigned long address = pvmw->address;
2873 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2874 pmd_t pmde;
2875 swp_entry_t entry;
2876
2877 if (!(pvmw->pmd && !pvmw->pte))
2878 return;
2879
2880 entry = pmd_to_swp_entry(*pvmw->pmd);
2881 get_page(new);
2882 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002883 if (pmd_swp_soft_dirty(*pvmw->pmd))
2884 pmde = pmd_mksoft_dirty(pmde);
Zi Yan616b8372017-09-08 16:10:57 -07002885 if (is_write_migration_entry(entry))
2886 pmde = maybe_pmd_mkwrite(pmde, vma);
2887
2888 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2889 page_add_anon_rmap(new, vma, mmun_start, true);
2890 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2891 if (vma->vm_flags & VM_LOCKED)
2892 mlock_vma_page(new);
2893 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2894}
2895#endif