blob: 0e7ded98d114d184877d2fc9bd0f02c3187f2ed5 [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/*
Michael DeGuzisb14d5952017-05-17 15:19:21 -040042 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080046 * 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
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800478{
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800479 if (likely(vma->vm_flags & VM_WRITE))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800480 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);
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800602 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);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800609 mm_inc_nr_ptes(vma->vm_mm);
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. Shutemovc4812902017-11-15 17:35:37 -0800665 mm_inc_nr_ptes(mm);
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) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800744 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);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800750 mm_inc_nr_ptes(mm);
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700751 }
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
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800792{
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800793 if (likely(vma->vm_flags & VM_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800794 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) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800810 entry = pud_mkyoung(pud_mkdirty(entry));
811 entry = maybe_pud_mkwrite(entry, vma);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800812 }
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,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300845 pmd_t *pmd, int flags)
Dan Williams3565fce2016-01-15 16:56:55 -0800846{
847 pmd_t _pmd;
848
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300849 _pmd = pmd_mkyoung(*pmd);
850 if (flags & FOLL_WRITE)
851 _pmd = pmd_mkdirty(_pmd);
Dan Williams3565fce2016-01-15 16:56:55 -0800852 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300853 pmd, _pmd, flags & FOLL_WRITE))
Dan Williams3565fce2016-01-15 16:56:55 -0800854 update_mmu_cache_pmd(vma, addr, pmd);
855}
856
857struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
858 pmd_t *pmd, int flags)
859{
860 unsigned long pfn = pmd_pfn(*pmd);
861 struct mm_struct *mm = vma->vm_mm;
862 struct dev_pagemap *pgmap;
863 struct page *page;
864
865 assert_spin_locked(pmd_lockptr(mm, pmd));
866
Keno Fischer8310d482017-01-24 15:17:48 -0800867 /*
868 * When we COW a devmap PMD entry, we split it into PTEs, so we should
869 * not be in this function with `flags & FOLL_COW` set.
870 */
871 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
872
Dan Williams3565fce2016-01-15 16:56:55 -0800873 if (flags & FOLL_WRITE && !pmd_write(*pmd))
874 return NULL;
875
876 if (pmd_present(*pmd) && pmd_devmap(*pmd))
877 /* pass */;
878 else
879 return NULL;
880
881 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300882 touch_pmd(vma, addr, pmd, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800883
884 /*
885 * device mapped pages can only be returned if the
886 * caller will manage the page reference count.
887 */
888 if (!(flags & FOLL_GET))
889 return ERR_PTR(-EEXIST);
890
891 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
892 pgmap = get_dev_pagemap(pfn, NULL);
893 if (!pgmap)
894 return ERR_PTR(-EFAULT);
895 page = pfn_to_page(pfn);
896 get_page(page);
897 put_dev_pagemap(pgmap);
898
899 return page;
900}
901
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800902int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
903 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
904 struct vm_area_struct *vma)
905{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800906 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800907 struct page *src_page;
908 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800909 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700910 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800911
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700912 /* Skip if can be re-fill on fault */
913 if (!vma_is_anonymous(vma))
914 return 0;
915
916 pgtable = pte_alloc_one(dst_mm, addr);
917 if (unlikely(!pgtable))
918 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800919
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800920 dst_ptl = pmd_lock(dst_mm, dst_pmd);
921 src_ptl = pmd_lockptr(src_mm, src_pmd);
922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800923
924 ret = -EAGAIN;
925 pmd = *src_pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -0700926
927#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
928 if (unlikely(is_swap_pmd(pmd))) {
929 swp_entry_t entry = pmd_to_swp_entry(pmd);
930
931 VM_BUG_ON(!is_pmd_migration_entry(pmd));
932 if (is_write_migration_entry(entry)) {
933 make_migration_entry_read(&entry);
934 pmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -0700935 if (pmd_swp_soft_dirty(*src_pmd))
936 pmd = pmd_swp_mksoft_dirty(pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -0700937 set_pmd_at(src_mm, addr, src_pmd, pmd);
938 }
Zi Yandd8a67f2017-11-02 15:59:47 -0700939 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovaf5b0f62017-11-15 17:35:40 -0800940 mm_inc_nr_ptes(dst_mm);
Zi Yandd8a67f2017-11-02 15:59:47 -0700941 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Zi Yan84c3fc42017-09-08 16:11:01 -0700942 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
943 ret = 0;
944 goto out_unlock;
945 }
946#endif
947
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700948 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800949 pte_free(dst_mm, pgtable);
950 goto out_unlock;
951 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800952 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800953 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800954 * under splitting since we don't split the page itself, only pmd to
955 * a page table.
956 */
957 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700958 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800959 /*
960 * get_huge_zero_page() will never allocate a new page here,
961 * since we already have a zero page to copy. It just takes a
962 * reference.
963 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700964 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700965 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700966 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800967 ret = 0;
968 goto out_unlock;
969 }
Mel Gormande466bd2013-12-18 17:08:42 -0800970
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700971 src_page = pmd_page(pmd);
972 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
973 get_page(src_page);
974 page_dup_rmap(src_page, true);
975 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800976 mm_inc_nr_ptes(dst_mm);
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700977 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800978
979 pmdp_set_wrprotect(src_mm, addr, src_pmd);
980 pmd = pmd_mkold(pmd_wrprotect(pmd));
981 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800982
983 ret = 0;
984out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800985 spin_unlock(src_ptl);
986 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800987out:
988 return ret;
989}
990
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800991#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
992static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300993 pud_t *pud, int flags)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800994{
995 pud_t _pud;
996
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300997 _pud = pud_mkyoung(*pud);
998 if (flags & FOLL_WRITE)
999 _pud = pud_mkdirty(_pud);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001000 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001001 pud, _pud, flags & FOLL_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001002 update_mmu_cache_pud(vma, addr, pud);
1003}
1004
1005struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1006 pud_t *pud, int flags)
1007{
1008 unsigned long pfn = pud_pfn(*pud);
1009 struct mm_struct *mm = vma->vm_mm;
1010 struct dev_pagemap *pgmap;
1011 struct page *page;
1012
1013 assert_spin_locked(pud_lockptr(mm, pud));
1014
1015 if (flags & FOLL_WRITE && !pud_write(*pud))
1016 return NULL;
1017
1018 if (pud_present(*pud) && pud_devmap(*pud))
1019 /* pass */;
1020 else
1021 return NULL;
1022
1023 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001024 touch_pud(vma, addr, pud, flags);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001025
1026 /*
1027 * device mapped pages can only be returned if the
1028 * caller will manage the page reference count.
1029 */
1030 if (!(flags & FOLL_GET))
1031 return ERR_PTR(-EEXIST);
1032
1033 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1034 pgmap = get_dev_pagemap(pfn, NULL);
1035 if (!pgmap)
1036 return ERR_PTR(-EFAULT);
1037 page = pfn_to_page(pfn);
1038 get_page(page);
1039 put_dev_pagemap(pgmap);
1040
1041 return page;
1042}
1043
1044int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1046 struct vm_area_struct *vma)
1047{
1048 spinlock_t *dst_ptl, *src_ptl;
1049 pud_t pud;
1050 int ret;
1051
1052 dst_ptl = pud_lock(dst_mm, dst_pud);
1053 src_ptl = pud_lockptr(src_mm, src_pud);
1054 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1055
1056 ret = -EAGAIN;
1057 pud = *src_pud;
1058 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1059 goto out_unlock;
1060
1061 /*
1062 * When page table lock is held, the huge zero pud should not be
1063 * under splitting since we don't split the page itself, only pud to
1064 * a page table.
1065 */
1066 if (is_huge_zero_pud(pud)) {
1067 /* No huge zero pud yet */
1068 }
1069
1070 pudp_set_wrprotect(src_mm, addr, src_pud);
1071 pud = pud_mkold(pud_wrprotect(pud));
1072 set_pud_at(dst_mm, addr, dst_pud, pud);
1073
1074 ret = 0;
1075out_unlock:
1076 spin_unlock(src_ptl);
1077 spin_unlock(dst_ptl);
1078 return ret;
1079}
1080
1081void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1082{
1083 pud_t entry;
1084 unsigned long haddr;
1085 bool write = vmf->flags & FAULT_FLAG_WRITE;
1086
1087 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1088 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1089 goto unlock;
1090
1091 entry = pud_mkyoung(orig_pud);
1092 if (write)
1093 entry = pud_mkdirty(entry);
1094 haddr = vmf->address & HPAGE_PUD_MASK;
1095 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1096 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1097
1098unlock:
1099 spin_unlock(vmf->ptl);
1100}
1101#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1102
Jan Kara82b0f8c2016-12-14 15:06:58 -08001103void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001104{
1105 pmd_t entry;
1106 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001107 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001108
Jan Kara82b0f8c2016-12-14 15:06:58 -08001109 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1110 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001111 goto unlock;
1112
1113 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001114 if (write)
1115 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001116 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001117 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001118 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001119
1120unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001121 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001122}
1123
Jan Kara82b0f8c2016-12-14 15:06:58 -08001124static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001125 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001126{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001127 struct vm_area_struct *vma = vmf->vma;
1128 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001129 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001130 pgtable_t pgtable;
1131 pmd_t _pmd;
1132 int ret = 0, i;
1133 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001134 unsigned long mmun_start; /* For mmu_notifiers */
1135 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001136
1137 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1138 GFP_KERNEL);
1139 if (unlikely(!pages)) {
1140 ret |= VM_FAULT_OOM;
1141 goto out;
1142 }
1143
1144 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001145 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001146 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001147 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001148 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1149 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001150 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001151 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001152 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001153 memcg = (void *)page_private(pages[i]);
1154 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001155 mem_cgroup_cancel_charge(pages[i], memcg,
1156 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001157 put_page(pages[i]);
1158 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159 kfree(pages);
1160 ret |= VM_FAULT_OOM;
1161 goto out;
1162 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001163 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 }
1165
1166 for (i = 0; i < HPAGE_PMD_NR; i++) {
1167 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001168 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001169 __SetPageUptodate(pages[i]);
1170 cond_resched();
1171 }
1172
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001173 mmun_start = haddr;
1174 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001175 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001176
Jan Kara82b0f8c2016-12-14 15:06:58 -08001177 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1178 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001179 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001180 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001181
Jérôme Glisse0f108512017-11-15 17:34:07 -08001182 /*
1183 * Leave pmd empty until pte is filled note we must notify here as
1184 * concurrent CPU thread might write to new page before the call to
1185 * mmu_notifier_invalidate_range_end() happens which can lead to a
1186 * device seeing memory write in different order than CPU.
1187 *
1188 * See Documentation/vm/mmu_notifier.txt
1189 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001190 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191
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
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001216 /*
1217 * No need to double call mmu_notifier->invalidate_range() callback as
1218 * the above pmdp_huge_clear_flush_notify() did already call it.
1219 */
1220 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1221 mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001222
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001223 ret |= VM_FAULT_WRITE;
1224 put_page(page);
1225
1226out:
1227 return ret;
1228
1229out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001230 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001232 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001233 memcg = (void *)page_private(pages[i]);
1234 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001235 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001236 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001237 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001238 kfree(pages);
1239 goto out;
1240}
1241
Jan Kara82b0f8c2016-12-14 15:06:58 -08001242int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001243{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001244 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001245 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001246 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001247 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001248 unsigned long mmun_start; /* For mmu_notifiers */
1249 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001250 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001251 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001252
Jan Kara82b0f8c2016-12-14 15:06:58 -08001253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001254 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001255 if (is_huge_zero_pmd(orig_pmd))
1256 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001257 spin_lock(vmf->ptl);
1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001259 goto out_unlock;
1260
1261 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001263 /*
1264 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001265 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001266 */
Huang Yingba3c4ce2017-09-06 16:22:19 -07001267 if (!trylock_page(page)) {
1268 get_page(page);
1269 spin_unlock(vmf->ptl);
1270 lock_page(page);
1271 spin_lock(vmf->ptl);
1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273 unlock_page(page);
1274 put_page(page);
1275 goto out_unlock;
1276 }
1277 put_page(page);
1278 }
1279 if (reuse_swap_page(page, NULL)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001280 pmd_t entry;
1281 entry = pmd_mkyoung(orig_pmd);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001285 ret |= VM_FAULT_WRITE;
Huang Yingba3c4ce2017-09-06 16:22:19 -07001286 unlock_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001287 goto out_unlock;
1288 }
Huang Yingba3c4ce2017-09-06 16:22:19 -07001289 unlock_page(page);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001290 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001291 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001292alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001293 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001294 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001295 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001297 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001298 new_page = NULL;
1299
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001300 if (likely(new_page)) {
1301 prep_transhuge_page(new_page);
1302 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001303 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001304 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001305 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001306 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001308 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001309 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001310 ret |= VM_FAULT_FALLBACK;
1311 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001312 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001313 }
David Rientjes17766dd2013-09-12 15:14:06 -07001314 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001315 goto out;
1316 }
1317
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001318 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1319 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001320 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001321 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001322 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001323 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001324 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001325 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001326 goto out;
1327 }
1328
David Rientjes17766dd2013-09-12 15:14:06 -07001329 count_vm_event(THP_FAULT_ALLOC);
1330
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001331 if (!page)
Huang Yingc79b57e2017-09-06 16:25:04 -07001332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001333 else
1334 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001335 __SetPageUptodate(new_page);
1336
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001337 mmun_start = haddr;
1338 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001339 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001340
Jan Kara82b0f8c2016-12-14 15:06:58 -08001341 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001342 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001343 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001346 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001347 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001348 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001349 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001350 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001351 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001353 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001354 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001355 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001356 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001357 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1358 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001359 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001360 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001361 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001362 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001363 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001364 put_page(page);
1365 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001366 ret |= VM_FAULT_WRITE;
1367 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001368 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001369out_mn:
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001370 /*
1371 * No need to double call mmu_notifier->invalidate_range() callback as
1372 * the above pmdp_huge_clear_flush_notify() did already call it.
1373 */
1374 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1375 mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001376out:
1377 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001378out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001379 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001380 return ret;
1381}
1382
Keno Fischer8310d482017-01-24 15:17:48 -08001383/*
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1386 */
1387static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1388{
1389 return pmd_write(pmd) ||
1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1391}
1392
David Rientjesb676b292012-10-08 16:34:03 -07001393struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001394 unsigned long addr,
1395 pmd_t *pmd,
1396 unsigned int flags)
1397{
David Rientjesb676b292012-10-08 16:34:03 -07001398 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001399 struct page *page = NULL;
1400
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001401 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001402
Keno Fischer8310d482017-01-24 15:17:48 -08001403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001404 goto out;
1405
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001406 /* Avoid dumping huge zero page */
1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408 return ERR_PTR(-EFAULT);
1409
Mel Gorman2b4847e2013-12-18 17:08:32 -08001410 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001412 goto out;
1413
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001414 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001416 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001417 touch_pmd(vma, addr, pmd, flags);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001419 /*
1420 * We don't mlock() pte-mapped THPs. This way we can avoid
1421 * leaking mlocked pages into non-VM_LOCKED VMAs.
1422 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001423 * For anon THP:
1424 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001425 * In most cases the pmd is the only mapping of the page as we
1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427 * writable private mappings in populate_vma_page_range().
1428 *
1429 * The only scenario when we have the page shared here is if we
1430 * mlocking read-only mapping shared over fork(). We skip
1431 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001432 *
1433 * For file THP:
1434 *
1435 * We can expect PageDoubleMap() to be stable under page lock:
1436 * for file pages we set it in page_add_file_rmap(), which
1437 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001438 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001439
1440 if (PageAnon(page) && compound_mapcount(page) != 1)
1441 goto skip_mlock;
1442 if (PageDoubleMap(page) || !page->mapping)
1443 goto skip_mlock;
1444 if (!trylock_page(page))
1445 goto skip_mlock;
1446 lru_add_drain();
1447 if (page->mapping && !PageDoubleMap(page))
1448 mlock_vma_page(page);
1449 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001450 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001451skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001454 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001455 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001456
1457out:
1458 return page;
1459}
1460
Mel Gormand10e63f2012-10-25 14:16:31 +02001461/* NUMA hinting page fault entry point for trans huge pmds */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001462int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001463{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001464 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001465 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001466 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001468 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001469 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001470 bool page_locked;
1471 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001472 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001473 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001474
Jan Kara82b0f8c2016-12-14 15:06:58 -08001475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001477 goto out_unlock;
1478
Mel Gormande466bd2013-12-18 17:08:42 -08001479 /*
1480 * If there are potential migrations, wait for completion and retry
1481 * without disrupting NUMA hinting information. Do not relock and
1482 * check_same as the page may no longer be mapped.
1483 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485 page = pmd_page(*vmf->pmd);
Mark Rutland3c226c62017-06-16 14:02:34 -07001486 if (!get_page_unless_zero(page))
1487 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001488 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001489 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001490 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001491 goto out;
1492 }
1493
Mel Gormand10e63f2012-10-25 14:16:31 +02001494 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001495 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001496 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001497 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001498 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001499 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001500 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001501 flags |= TNF_FAULT_LOCAL;
1502 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001503
Mel Gormanbea66fb2015-03-25 15:55:37 -07001504 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001505 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001506 flags |= TNF_NO_GROUP;
1507
1508 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001509 * Acquire the page lock to serialise THP migrations but avoid dropping
1510 * page_table_lock if at all possible
1511 */
Mel Gormanb8916632013-10-07 11:28:44 +01001512 page_locked = trylock_page(page);
1513 target_nid = mpol_misplaced(page, vma, haddr);
1514 if (target_nid == -1) {
1515 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001516 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001517 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001518 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001519
Mel Gormande466bd2013-12-18 17:08:42 -08001520 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001521 if (!page_locked) {
Mark Rutland3c226c62017-06-16 14:02:34 -07001522 page_nid = -1;
1523 if (!get_page_unless_zero(page))
1524 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001525 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001526 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001527 put_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001528 goto out;
1529 }
1530
Mel Gorman2b4847e2013-12-18 17:08:32 -08001531 /*
1532 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1533 * to serialises splits
1534 */
Mel Gormanb8916632013-10-07 11:28:44 +01001535 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001536 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001537 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001538
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001539 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001540 spin_lock(vmf->ptl);
1541 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001542 unlock_page(page);
1543 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001544 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001545 goto out_unlock;
1546 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001547
Mel Gormanc3a489c2013-12-18 17:08:38 -08001548 /* Bail if we fail to protect against THP splits for any reason */
1549 if (unlikely(!anon_vma)) {
1550 put_page(page);
1551 page_nid = -1;
1552 goto clear_pmdnuma;
1553 }
1554
Mel Gormana54a4072013-10-07 11:28:46 +01001555 /*
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001556 * Since we took the NUMA fault, we must have observed the !accessible
1557 * bit. Make sure all other CPUs agree with that, to avoid them
1558 * modifying the page we're about to migrate.
1559 *
1560 * Must be done under PTL such that we'll observe the relevant
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001561 * inc_tlb_flush_pending().
1562 *
1563 * We are not sure a pending tlb flush here is for a huge page
1564 * mapping or not. Hence use the tlb range variant
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001565 */
1566 if (mm_tlb_flush_pending(vma->vm_mm))
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001567 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001568
1569 /*
Mel Gormana54a4072013-10-07 11:28:46 +01001570 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001571 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001572 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001573 spin_unlock(vmf->ptl);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001574
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001575 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001576 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001577 if (migrated) {
1578 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001579 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001580 } else
1581 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001582
Mel Gorman8191acb2013-10-07 11:28:45 +01001583 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001584clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001585 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001586 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001587 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001588 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001589 if (was_writable)
1590 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001593 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001594out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001595 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001596
1597out:
1598 if (anon_vma)
1599 page_unlock_anon_vma_read(anon_vma);
1600
Mel Gorman8191acb2013-10-07 11:28:45 +01001601 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001602 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001603 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001604
Mel Gormand10e63f2012-10-25 14:16:31 +02001605 return 0;
1606}
1607
Huang Ying319904a2016-07-28 15:48:03 -07001608/*
1609 * Return true if we do MADV_FREE successfully on entire pmd page.
1610 * Otherwise, return false.
1611 */
1612bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001613 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001614{
1615 spinlock_t *ptl;
1616 pmd_t orig_pmd;
1617 struct page *page;
1618 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001619 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001620
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001621 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1622
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001623 ptl = pmd_trans_huge_lock(pmd, vma);
1624 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001625 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001626
1627 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001628 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001629 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001630
Zi Yan84c3fc42017-09-08 16:11:01 -07001631 if (unlikely(!pmd_present(orig_pmd))) {
1632 VM_BUG_ON(thp_migration_supported() &&
1633 !is_pmd_migration_entry(orig_pmd));
1634 goto out;
1635 }
1636
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001637 page = pmd_page(orig_pmd);
1638 /*
1639 * If other processes are mapping this page, we couldn't discard
1640 * the page unless they all do MADV_FREE so let's skip the page.
1641 */
1642 if (page_mapcount(page) != 1)
1643 goto out;
1644
1645 if (!trylock_page(page))
1646 goto out;
1647
1648 /*
1649 * If user want to discard part-pages of THP, split it so MADV_FREE
1650 * will deactivate only them.
1651 */
1652 if (next - addr != HPAGE_PMD_SIZE) {
1653 get_page(page);
1654 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001655 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001656 unlock_page(page);
Kirill A. Shutemovbbf29ff2017-07-06 15:35:28 -07001657 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001658 goto out_unlocked;
1659 }
1660
1661 if (PageDirty(page))
1662 ClearPageDirty(page);
1663 unlock_page(page);
1664
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001665 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemov58ceeb62017-04-13 14:56:26 -07001666 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001667 orig_pmd = pmd_mkold(orig_pmd);
1668 orig_pmd = pmd_mkclean(orig_pmd);
1669
1670 set_pmd_at(mm, addr, pmd, orig_pmd);
1671 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1672 }
Shaohua Li802a3a92017-05-03 14:52:32 -07001673
1674 mark_page_lazyfree(page);
Huang Ying319904a2016-07-28 15:48:03 -07001675 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001676out:
1677 spin_unlock(ptl);
1678out_unlocked:
1679 return ret;
1680}
1681
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001682static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1683{
1684 pgtable_t pgtable;
1685
1686 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1687 pte_free(mm, pgtable);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -08001688 mm_dec_nr_ptes(mm);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001689}
1690
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001691int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001692 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001693{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001694 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001695 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001696
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001697 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1698
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001699 ptl = __pmd_trans_huge_lock(pmd, vma);
1700 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001701 return 0;
1702 /*
1703 * For architectures like ppc64 we look at deposited pgtable
1704 * when calling pmdp_huge_get_and_clear. So do the
1705 * pgtable_trans_huge_withdraw after finishing pmdp related
1706 * operations.
1707 */
1708 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1709 tlb->fullmm);
1710 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711 if (vma_is_dax(vma)) {
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -07001712 if (arch_needs_pgtable_deposit())
1713 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001714 spin_unlock(ptl);
1715 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001716 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001717 } else if (is_huge_zero_pmd(orig_pmd)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001718 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001719 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001720 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001721 } else {
Zi Yan616b8372017-09-08 16:10:57 -07001722 struct page *page = NULL;
1723 int flush_needed = 1;
1724
1725 if (pmd_present(orig_pmd)) {
1726 page = pmd_page(orig_pmd);
1727 page_remove_rmap(page, true);
1728 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1729 VM_BUG_ON_PAGE(!PageHead(page), page);
1730 } else if (thp_migration_supported()) {
1731 swp_entry_t entry;
1732
1733 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1734 entry = pmd_to_swp_entry(orig_pmd);
1735 page = pfn_to_page(swp_offset(entry));
1736 flush_needed = 0;
1737 } else
1738 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1739
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001740 if (PageAnon(page)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001741 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001742 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1743 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001744 if (arch_needs_pgtable_deposit())
1745 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001746 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1747 }
Zi Yan616b8372017-09-08 16:10:57 -07001748
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001749 spin_unlock(ptl);
Zi Yan616b8372017-09-08 16:10:57 -07001750 if (flush_needed)
1751 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001752 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001753 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001754}
1755
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001756#ifndef pmd_move_must_withdraw
1757static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1758 spinlock_t *old_pmd_ptl,
1759 struct vm_area_struct *vma)
1760{
1761 /*
1762 * With split pmd lock we also need to move preallocated
1763 * PTE page table if new_pmd is on different PMD page table.
1764 *
1765 * We also don't deposit and withdraw tables for file pages.
1766 */
1767 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1768}
1769#endif
1770
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001771static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1772{
1773#ifdef CONFIG_MEM_SOFT_DIRTY
1774 if (unlikely(is_pmd_migration_entry(pmd)))
1775 pmd = pmd_swp_mksoft_dirty(pmd);
1776 else if (pmd_present(pmd))
1777 pmd = pmd_mksoft_dirty(pmd);
1778#endif
1779 return pmd;
1780}
1781
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001782bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001783 unsigned long new_addr, unsigned long old_end,
Aaron Lu5d190422016-11-10 17:16:33 +08001784 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001785{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001786 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001787 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001788 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001789 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001790
1791 if ((old_addr & ~HPAGE_PMD_MASK) ||
1792 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001793 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001794 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001795
1796 /*
1797 * The destination pmd shouldn't be established, free_pgtables()
1798 * should have release it.
1799 */
1800 if (WARN_ON(!pmd_none(*new_pmd))) {
1801 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001802 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001803 }
1804
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001805 /*
1806 * We don't have to worry about the ordering of src and dst
1807 * ptlocks because exclusive mmap_sem prevents deadlock.
1808 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001809 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1810 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001811 new_ptl = pmd_lockptr(mm, new_pmd);
1812 if (new_ptl != old_ptl)
1813 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001814 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Aaron Lua2ce2662016-11-29 13:27:31 +08001815 if (pmd_present(pmd) && pmd_dirty(pmd))
1816 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001817 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001818
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001819 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301820 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001821 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1822 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001823 }
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001824 pmd = move_soft_dirty_pmd(pmd);
1825 set_pmd_at(mm, new_addr, new_pmd, pmd);
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301826 if (new_ptl != old_ptl)
1827 spin_unlock(new_ptl);
Aaron Lu5d190422016-11-10 17:16:33 +08001828 if (force_flush)
1829 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1830 else
1831 *need_flush = true;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001832 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001833 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001834 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001835 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001836}
1837
Mel Gormanf123d742013-10-07 11:28:49 +01001838/*
1839 * Returns
1840 * - 0 if PMD could not be locked
1841 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1842 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1843 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001844int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001845 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001846{
1847 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001848 spinlock_t *ptl;
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001849 pmd_t entry;
1850 bool preserve_write;
1851 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001852
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001853 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001854 if (!ptl)
1855 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001856
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001857 preserve_write = prot_numa && pmd_write(*pmd);
1858 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001859
Zi Yan84c3fc42017-09-08 16:11:01 -07001860#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1861 if (is_swap_pmd(*pmd)) {
1862 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1863
1864 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1865 if (is_write_migration_entry(entry)) {
1866 pmd_t newpmd;
1867 /*
1868 * A protection check is difficult so
1869 * just be safe and disable write
1870 */
1871 make_migration_entry_read(&entry);
1872 newpmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001873 if (pmd_swp_soft_dirty(*pmd))
1874 newpmd = pmd_swp_mksoft_dirty(newpmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001875 set_pmd_at(mm, addr, pmd, newpmd);
1876 }
1877 goto unlock;
1878 }
1879#endif
1880
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001881 /*
1882 * Avoid trapping faults against the zero page. The read-only
1883 * data is likely to be read-cached on the local CPU and
1884 * local/remote hits to the zero page are not interesting.
1885 */
1886 if (prot_numa && is_huge_zero_pmd(*pmd))
1887 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001888
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001889 if (prot_numa && pmd_protnone(*pmd))
1890 goto unlock;
1891
Kirill A. Shutemovced10802017-04-13 14:56:20 -07001892 /*
1893 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1894 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895 * which is also under down_read(mmap_sem):
1896 *
1897 * CPU0: CPU1:
1898 * change_huge_pmd(prot_numa=1)
1899 * pmdp_huge_get_and_clear_notify()
1900 * madvise_dontneed()
1901 * zap_pmd_range()
1902 * pmd_trans_huge(*pmd) == 0 (without ptl)
1903 * // skip the pmd
1904 * set_pmd_at();
1905 * // pmd is re-established
1906 *
1907 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908 * which may break userspace.
1909 *
1910 * pmdp_invalidate() is required to make sure we don't miss
1911 * dirty/young flags set by hardware.
1912 */
1913 entry = *pmd;
1914 pmdp_invalidate(vma, addr, pmd);
1915
1916 /*
1917 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1918 * corrupt them.
1919 */
1920 if (pmd_dirty(*pmd))
1921 entry = pmd_mkdirty(entry);
1922 if (pmd_young(*pmd))
1923 entry = pmd_mkyoung(entry);
1924
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001925 entry = pmd_modify(entry, newprot);
1926 if (preserve_write)
1927 entry = pmd_mk_savedwrite(entry);
1928 ret = HPAGE_PMD_NR;
1929 set_pmd_at(mm, addr, pmd, entry);
1930 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1931unlock:
1932 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001933 return ret;
1934}
1935
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001936/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001937 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001938 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001939 * Note that if it returns page table lock pointer, this routine returns without
1940 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001941 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001942spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001943{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001944 spinlock_t *ptl;
1945 ptl = pmd_lock(vma->vm_mm, pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001946 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1947 pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001948 return ptl;
1949 spin_unlock(ptl);
1950 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001951}
1952
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001953/*
1954 * Returns true if a given pud maps a thp, false otherwise.
1955 *
1956 * Note that if it returns true, this routine returns without unlocking page
1957 * table lock. So callers must unlock it.
1958 */
1959spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1960{
1961 spinlock_t *ptl;
1962
1963 ptl = pud_lock(vma->vm_mm, pud);
1964 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1965 return ptl;
1966 spin_unlock(ptl);
1967 return NULL;
1968}
1969
1970#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1971int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1972 pud_t *pud, unsigned long addr)
1973{
1974 pud_t orig_pud;
1975 spinlock_t *ptl;
1976
1977 ptl = __pud_trans_huge_lock(pud, vma);
1978 if (!ptl)
1979 return 0;
1980 /*
1981 * For architectures like ppc64 we look at deposited pgtable
1982 * when calling pudp_huge_get_and_clear. So do the
1983 * pgtable_trans_huge_withdraw after finishing pudp related
1984 * operations.
1985 */
1986 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1987 tlb->fullmm);
1988 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1989 if (vma_is_dax(vma)) {
1990 spin_unlock(ptl);
1991 /* No zero page support yet */
1992 } else {
1993 /* No support for anonymous PUD pages yet */
1994 BUG();
1995 }
1996 return 1;
1997}
1998
1999static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2000 unsigned long haddr)
2001{
2002 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2003 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2004 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2005 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2006
Yisheng Xiece9311c2017-03-09 16:17:00 -08002007 count_vm_event(THP_SPLIT_PUD);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002008
2009 pudp_huge_clear_flush_notify(vma, haddr, pud);
2010}
2011
2012void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2013 unsigned long address)
2014{
2015 spinlock_t *ptl;
2016 struct mm_struct *mm = vma->vm_mm;
2017 unsigned long haddr = address & HPAGE_PUD_MASK;
2018
2019 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2020 ptl = pud_lock(mm, pud);
2021 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2022 goto out;
2023 __split_huge_pud_locked(vma, pud, haddr);
2024
2025out:
2026 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002027 /*
2028 * No need to double call mmu_notifier->invalidate_range() callback as
2029 * the above pudp_huge_clear_flush_notify() did already call it.
2030 */
2031 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2032 HPAGE_PUD_SIZE);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002033}
2034#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2035
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002036static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2037 unsigned long haddr, pmd_t *pmd)
2038{
2039 struct mm_struct *mm = vma->vm_mm;
2040 pgtable_t pgtable;
2041 pmd_t _pmd;
2042 int i;
2043
Jérôme Glisse0f108512017-11-15 17:34:07 -08002044 /*
2045 * Leave pmd empty until pte is filled note that it is fine to delay
2046 * notification until mmu_notifier_invalidate_range_end() as we are
2047 * replacing a zero pmd write protected page with a zero pte write
2048 * protected page.
2049 *
2050 * See Documentation/vm/mmu_notifier.txt
2051 */
2052 pmdp_huge_clear_flush(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002053
2054 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2055 pmd_populate(mm, &_pmd, pgtable);
2056
2057 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2058 pte_t *pte, entry;
2059 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2060 entry = pte_mkspecial(entry);
2061 pte = pte_offset_map(&_pmd, haddr);
2062 VM_BUG_ON(!pte_none(*pte));
2063 set_pte_at(mm, haddr, pte, entry);
2064 pte_unmap(pte);
2065 }
2066 smp_wmb(); /* make pte visible before pmd */
2067 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002068}
2069
2070static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002071 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002072{
2073 struct mm_struct *mm = vma->vm_mm;
2074 struct page *page;
2075 pgtable_t pgtable;
2076 pmd_t _pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -07002077 bool young, write, dirty, soft_dirty, pmd_migration = false;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002078 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002079 int i;
2080
2081 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2082 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2083 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Zi Yan84c3fc42017-09-08 16:11:01 -07002084 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2085 && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002086
2087 count_vm_event(THP_SPLIT_PMD);
2088
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002089 if (!vma_is_anonymous(vma)) {
2090 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08002091 /*
2092 * We are going to unmap this huge page. So
2093 * just go ahead and zap it
2094 */
2095 if (arch_needs_pgtable_deposit())
2096 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002097 if (vma_is_dax(vma))
2098 return;
2099 page = pmd_page(_pmd);
2100 if (!PageReferenced(page) && pmd_young(_pmd))
2101 SetPageReferenced(page);
2102 page_remove_rmap(page, true);
2103 put_page(page);
2104 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002105 return;
2106 } else if (is_huge_zero_pmd(*pmd)) {
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002107 /*
2108 * FIXME: Do we want to invalidate secondary mmu by calling
2109 * mmu_notifier_invalidate_range() see comments below inside
2110 * __split_huge_pmd() ?
2111 *
2112 * We are going from a zero huge page write protected to zero
2113 * small page also write protected so it does not seems useful
2114 * to invalidate secondary mmu at this time.
2115 */
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002116 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2117 }
2118
Zi Yan84c3fc42017-09-08 16:11:01 -07002119#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2120 pmd_migration = is_pmd_migration_entry(*pmd);
2121 if (pmd_migration) {
2122 swp_entry_t entry;
2123
2124 entry = pmd_to_swp_entry(*pmd);
2125 page = pfn_to_page(swp_offset(entry));
2126 } else
2127#endif
2128 page = pmd_page(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002129 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002130 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002131 write = pmd_write(*pmd);
2132 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002133 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002134 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002135
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05302136 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002137 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2138 pmd_populate(mm, &_pmd, pgtable);
2139
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002140 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002141 pte_t entry, *pte;
2142 /*
2143 * Note that NUMA hinting access restrictions are not
2144 * transferred to avoid any possibility of altering
2145 * permissions across VMAs.
2146 */
Zi Yan84c3fc42017-09-08 16:11:01 -07002147 if (freeze || pmd_migration) {
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002148 swp_entry_t swp_entry;
2149 swp_entry = make_migration_entry(page + i, write);
2150 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002151 if (soft_dirty)
2152 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002153 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07002154 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002155 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002156 if (!write)
2157 entry = pte_wrprotect(entry);
2158 if (!young)
2159 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002160 if (soft_dirty)
2161 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002162 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002163 if (dirty)
2164 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002165 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002166 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002167 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002168 atomic_inc(&page[i]._mapcount);
2169 pte_unmap(pte);
2170 }
2171
2172 /*
2173 * Set PG_double_map before dropping compound_mapcount to avoid
2174 * false-negative page_mapped().
2175 */
2176 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2177 for (i = 0; i < HPAGE_PMD_NR; i++)
2178 atomic_inc(&page[i]._mapcount);
2179 }
2180
2181 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2182 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07002183 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002184 if (TestClearPageDoubleMap(page)) {
2185 /* No need in mapcount reference anymore */
2186 for (i = 0; i < HPAGE_PMD_NR; i++)
2187 atomic_dec(&page[i]._mapcount);
2188 }
2189 }
2190
2191 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002192 /*
2193 * Up to this point the pmd is present and huge and userland has the
2194 * whole access to the hugepage during the split (which happens in
2195 * place). If we overwrite the pmd with the not-huge version pointing
2196 * to the pte here (which of course we could if all CPUs were bug
2197 * free), userland could trigger a small page size TLB miss on the
2198 * small sized TLB while the hugepage TLB entry is still established in
2199 * the huge TLB. Some CPU doesn't like that.
2200 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2201 * 383 on page 93. Intel should be safe but is also warns that it's
2202 * only safe if the permission and cache attributes of the two entries
2203 * loaded in the two TLB is identical (which should be the case here).
2204 * But it is generally safer to never allow small and huge TLB entries
2205 * for the same virtual address to be loaded simultaneously. So instead
2206 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2207 * current pmd notpresent (atomically because here the pmd_trans_huge
2208 * and pmd_trans_splitting must remain set at all times on the pmd
2209 * until the split is complete for this pmd), then we flush the SMP TLB
2210 * and finally we write the non-huge version of the pmd entry with
2211 * pmd_populate.
2212 */
2213 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002214 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002215
2216 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002217 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002218 page_remove_rmap(page + i, false);
2219 put_page(page + i);
2220 }
2221 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002222}
2223
2224void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002225 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002226{
2227 spinlock_t *ptl;
2228 struct mm_struct *mm = vma->vm_mm;
2229 unsigned long haddr = address & HPAGE_PMD_MASK;
2230
2231 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2232 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002233
2234 /*
2235 * If caller asks to setup a migration entries, we need a page to check
2236 * pmd against. Otherwise we can end up replacing wrong page.
2237 */
2238 VM_BUG_ON(freeze && !page);
2239 if (page && page != pmd_page(*pmd))
2240 goto out;
2241
Dan Williams5c7fb562016-01-15 16:56:52 -08002242 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002243 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002244 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002245 clear_page_mlock(page);
Zi Yan84c3fc42017-09-08 16:11:01 -07002246 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002247 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002248 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002249out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002250 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002251 /*
2252 * No need to double call mmu_notifier->invalidate_range() callback.
2253 * They are 3 cases to consider inside __split_huge_pmd_locked():
2254 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2255 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2256 * fault will trigger a flush_notify before pointing to a new page
2257 * (it is fine if the secondary mmu keeps pointing to the old zero
2258 * page in the meantime)
2259 * 3) Split a huge pmd into pte pointing to the same page. No need
2260 * to invalidate secondary tlb entry they are all still valid.
2261 * any further changes to individual pte will notify. So no need
2262 * to call mmu_notifier->invalidate_range()
2263 */
2264 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2265 HPAGE_PMD_SIZE);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002266}
2267
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002268void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2269 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002270{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002271 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002272 p4d_t *p4d;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002273 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002274 pmd_t *pmd;
2275
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002276 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002277 if (!pgd_present(*pgd))
2278 return;
2279
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002280 p4d = p4d_offset(pgd, address);
2281 if (!p4d_present(*p4d))
2282 return;
2283
2284 pud = pud_offset(p4d, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002285 if (!pud_present(*pud))
2286 return;
2287
2288 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002289
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002290 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002291}
2292
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002293void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002294 unsigned long start,
2295 unsigned long end,
2296 long adjust_next)
2297{
2298 /*
2299 * If the new start address isn't hpage aligned and it could
2300 * previously contain an hugepage: check if we need to split
2301 * an huge pmd.
2302 */
2303 if (start & ~HPAGE_PMD_MASK &&
2304 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2305 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002306 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002307
2308 /*
2309 * If the new end address isn't hpage aligned and it could
2310 * previously contain an hugepage: check if we need to split
2311 * an huge pmd.
2312 */
2313 if (end & ~HPAGE_PMD_MASK &&
2314 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2315 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002316 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002317
2318 /*
2319 * If we're also updating the vma->vm_next->vm_start, if the new
2320 * vm_next->vm_start isn't page aligned and it could previously
2321 * contain an hugepage: check if we need to split an huge pmd.
2322 */
2323 if (adjust_next > 0) {
2324 struct vm_area_struct *next = vma->vm_next;
2325 unsigned long nstart = next->vm_start;
2326 nstart += adjust_next << PAGE_SHIFT;
2327 if (nstart & ~HPAGE_PMD_MASK &&
2328 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2329 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002330 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002331 }
2332}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002333
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002334static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002335{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002336 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002337 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
Minchan Kim666e5a42017-05-03 14:54:20 -07002338 bool unmap_success;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002339
2340 VM_BUG_ON_PAGE(!PageHead(page), page);
2341
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002342 if (PageAnon(page))
Naoya Horiguchib5ff8162017-09-08 16:10:49 -07002343 ttu_flags |= TTU_SPLIT_FREEZE;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002344
Minchan Kim666e5a42017-05-03 14:54:20 -07002345 unmap_success = try_to_unmap(page, ttu_flags);
2346 VM_BUG_ON_PAGE(!unmap_success, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002347}
2348
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002349static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002350{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002351 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002352 if (PageTransHuge(page)) {
2353 remove_migration_ptes(page, page, true);
2354 } else {
2355 for (i = 0; i < HPAGE_PMD_NR; i++)
2356 remove_migration_ptes(page + i, page + i, true);
2357 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002358}
2359
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002360static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002361 struct lruvec *lruvec, struct list_head *list)
2362{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002363 struct page *page_tail = head + tail;
2364
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002365 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002366 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002367
2368 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002369 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002370 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002371 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2372 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002373 * get_page_unless_zero(), and atomic_set() is implemented in C not
2374 * using locked ops. spin_unlock on x86 sometime uses locked ops
2375 * because of PPro errata 66, 92, so unless somebody can guarantee
2376 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002377 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002378 */
Huang Ying38d8b4e2017-07-06 15:37:18 -07002379 if (PageAnon(head) && !PageSwapCache(head)) {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002380 page_ref_inc(page_tail);
2381 } else {
2382 /* Additional pin to radix tree */
2383 page_ref_add(page_tail, 2);
2384 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002385
2386 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2387 page_tail->flags |= (head->flags &
2388 ((1L << PG_referenced) |
2389 (1L << PG_swapbacked) |
Huang Ying38d8b4e2017-07-06 15:37:18 -07002390 (1L << PG_swapcache) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002391 (1L << PG_mlocked) |
2392 (1L << PG_uptodate) |
2393 (1L << PG_active) |
2394 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002395 (1L << PG_unevictable) |
2396 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002397
2398 /*
2399 * After clearing PageTail the gup refcount can be released.
2400 * Page flags also must be visible before we make the page non-compound.
2401 */
2402 smp_wmb();
2403
2404 clear_compound_head(page_tail);
2405
2406 if (page_is_young(head))
2407 set_page_young(page_tail);
2408 if (page_is_idle(head))
2409 set_page_idle(page_tail);
2410
2411 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002412 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002413 page_tail);
2414 page_tail->mapping = head->mapping;
2415
2416 page_tail->index = head->index + tail;
2417 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2418 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002419}
2420
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002421static void __split_huge_page(struct page *page, struct list_head *list,
2422 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002423{
2424 struct page *head = compound_head(page);
2425 struct zone *zone = page_zone(head);
2426 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002427 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002428 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002429
Mel Gorman599d0c92016-07-28 15:45:31 -07002430 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002431
2432 /* complete memcg works before add pages to LRU */
2433 mem_cgroup_split_huge_fixup(head);
2434
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002435 if (!PageAnon(page))
2436 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2437
2438 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002439 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002440 /* Some pages can be beyond i_size: drop them from page cache */
2441 if (head[i].index >= end) {
2442 __ClearPageDirty(head + i);
2443 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002444 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2445 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002446 put_page(head + i);
2447 }
2448 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002449
2450 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002451 /* See comment in __split_huge_page_tail() */
2452 if (PageAnon(head)) {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002453 /* Additional pin to radix tree of swap cache */
2454 if (PageSwapCache(head))
2455 page_ref_add(head, 2);
2456 else
2457 page_ref_inc(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002458 } else {
2459 /* Additional pin to radix tree */
2460 page_ref_add(head, 2);
2461 spin_unlock(&head->mapping->tree_lock);
2462 }
2463
Mel Gormana52633d2016-07-28 15:45:28 -07002464 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002465
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002466 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002467
2468 for (i = 0; i < HPAGE_PMD_NR; i++) {
2469 struct page *subpage = head + i;
2470 if (subpage == page)
2471 continue;
2472 unlock_page(subpage);
2473
2474 /*
2475 * Subpages may be freed if there wasn't any mapping
2476 * like if add_to_swap() is running on a lru page that
2477 * had its mapping zapped. And freeing these pages
2478 * requires taking the lru_lock so we do the put_page
2479 * of the tail pages after the split is complete.
2480 */
2481 put_page(subpage);
2482 }
2483}
2484
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002485int total_mapcount(struct page *page)
2486{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002487 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002488
2489 VM_BUG_ON_PAGE(PageTail(page), page);
2490
2491 if (likely(!PageCompound(page)))
2492 return atomic_read(&page->_mapcount) + 1;
2493
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002494 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002495 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002496 return compound;
2497 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002498 for (i = 0; i < HPAGE_PMD_NR; i++)
2499 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002500 /* File pages has compound_mapcount included in _mapcount */
2501 if (!PageAnon(page))
2502 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002503 if (PageDoubleMap(page))
2504 ret -= HPAGE_PMD_NR;
2505 return ret;
2506}
2507
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002508/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002509 * This calculates accurately how many mappings a transparent hugepage
2510 * has (unlike page_mapcount() which isn't fully accurate). This full
2511 * accuracy is primarily needed to know if copy-on-write faults can
2512 * reuse the page and change the mapping to read-write instead of
2513 * copying them. At the same time this returns the total_mapcount too.
2514 *
2515 * The function returns the highest mapcount any one of the subpages
2516 * has. If the return value is one, even if different processes are
2517 * mapping different subpages of the transparent hugepage, they can
2518 * all reuse it, because each process is reusing a different subpage.
2519 *
2520 * The total_mapcount is instead counting all virtual mappings of the
2521 * subpages. If the total_mapcount is equal to "one", it tells the
2522 * caller all mappings belong to the same "mm" and in turn the
2523 * anon_vma of the transparent hugepage can become the vma->anon_vma
2524 * local one as no other process may be mapping any of the subpages.
2525 *
2526 * It would be more accurate to replace page_mapcount() with
2527 * page_trans_huge_mapcount(), however we only use
2528 * page_trans_huge_mapcount() in the copy-on-write faults where we
2529 * need full accuracy to avoid breaking page pinning, because
2530 * page_trans_huge_mapcount() is slower than page_mapcount().
2531 */
2532int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2533{
2534 int i, ret, _total_mapcount, mapcount;
2535
2536 /* hugetlbfs shouldn't call it */
2537 VM_BUG_ON_PAGE(PageHuge(page), page);
2538
2539 if (likely(!PageTransCompound(page))) {
2540 mapcount = atomic_read(&page->_mapcount) + 1;
2541 if (total_mapcount)
2542 *total_mapcount = mapcount;
2543 return mapcount;
2544 }
2545
2546 page = compound_head(page);
2547
2548 _total_mapcount = ret = 0;
2549 for (i = 0; i < HPAGE_PMD_NR; i++) {
2550 mapcount = atomic_read(&page[i]._mapcount) + 1;
2551 ret = max(ret, mapcount);
2552 _total_mapcount += mapcount;
2553 }
2554 if (PageDoubleMap(page)) {
2555 ret -= 1;
2556 _total_mapcount -= HPAGE_PMD_NR;
2557 }
2558 mapcount = compound_mapcount(page);
2559 ret += mapcount;
2560 _total_mapcount += mapcount;
2561 if (total_mapcount)
2562 *total_mapcount = _total_mapcount;
2563 return ret;
2564}
2565
Huang Yingb8f593c2017-07-06 15:37:28 -07002566/* Racy check whether the huge page can be split */
2567bool can_split_huge_page(struct page *page, int *pextra_pins)
2568{
2569 int extra_pins;
2570
2571 /* Additional pins from radix tree */
2572 if (PageAnon(page))
2573 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2574 else
2575 extra_pins = HPAGE_PMD_NR;
2576 if (pextra_pins)
2577 *pextra_pins = extra_pins;
2578 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2579}
2580
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002581/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002582 * This function splits huge page into normal pages. @page can point to any
2583 * subpage of huge page to split. Split doesn't change the position of @page.
2584 *
2585 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2586 * The huge page must be locked.
2587 *
2588 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2589 *
2590 * Both head page and tail pages will inherit mapping, flags, and so on from
2591 * the hugepage.
2592 *
2593 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2594 * they are not mapped.
2595 *
2596 * Returns 0 if the hugepage is split successfully.
2597 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2598 * us.
2599 */
2600int split_huge_page_to_list(struct page *page, struct list_head *list)
2601{
2602 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002603 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002604 struct anon_vma *anon_vma = NULL;
2605 struct address_space *mapping = NULL;
2606 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002607 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002608 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002609
2610 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002611 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002612 VM_BUG_ON_PAGE(!PageCompound(page), page);
2613
Huang Ying59807682017-09-06 16:22:34 -07002614 if (PageWriteback(page))
2615 return -EBUSY;
2616
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002617 if (PageAnon(head)) {
2618 /*
2619 * The caller does not necessarily hold an mmap_sem that would
2620 * prevent the anon_vma disappearing so we first we take a
2621 * reference to it and then lock the anon_vma for write. This
2622 * is similar to page_lock_anon_vma_read except the write lock
2623 * is taken to serialise against parallel split or collapse
2624 * operations.
2625 */
2626 anon_vma = page_get_anon_vma(head);
2627 if (!anon_vma) {
2628 ret = -EBUSY;
2629 goto out;
2630 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002631 mapping = NULL;
2632 anon_vma_lock_write(anon_vma);
2633 } else {
2634 mapping = head->mapping;
2635
2636 /* Truncated ? */
2637 if (!mapping) {
2638 ret = -EBUSY;
2639 goto out;
2640 }
2641
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002642 anon_vma = NULL;
2643 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002644 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002645
2646 /*
2647 * Racy check if we can split the page, before freeze_page() will
2648 * split PMDs
2649 */
Huang Yingb8f593c2017-07-06 15:37:28 -07002650 if (!can_split_huge_page(head, &extra_pins)) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002651 ret = -EBUSY;
2652 goto out_unlock;
2653 }
2654
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002655 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002656 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002657 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2658
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002659 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2660 if (mlocked)
2661 lru_add_drain();
2662
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002663 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002664 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002665
2666 if (mapping) {
2667 void **pslot;
2668
2669 spin_lock(&mapping->tree_lock);
2670 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2671 page_index(head));
2672 /*
2673 * Check if the head page is present in radix tree.
2674 * We assume all tail are present too, if head is there.
2675 */
2676 if (radix_tree_deref_slot_protected(pslot,
2677 &mapping->tree_lock) != head)
2678 goto fail;
2679 }
2680
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002681 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002682 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002683 count = page_count(head);
2684 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002685 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002686 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002687 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002688 list_del(page_deferred_list(head));
2689 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002690 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002691 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002692 spin_unlock(&pgdata->split_queue_lock);
2693 __split_huge_page(page, list, flags);
Huang Ying59807682017-09-06 16:22:34 -07002694 if (PageSwapCache(head)) {
2695 swp_entry_t entry = { .val = page_private(head) };
2696
2697 ret = split_swap_cluster(entry);
2698 } else
2699 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002700 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002701 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2702 pr_alert("total_mapcount: %u, page_count(): %u\n",
2703 mapcount, count);
2704 if (PageTail(page))
2705 dump_page(head, NULL);
2706 dump_page(page, "total_mapcount(head) > 0");
2707 BUG();
2708 }
2709 spin_unlock(&pgdata->split_queue_lock);
2710fail: if (mapping)
2711 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002712 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002713 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002714 ret = -EBUSY;
2715 }
2716
2717out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002718 if (anon_vma) {
2719 anon_vma_unlock_write(anon_vma);
2720 put_anon_vma(anon_vma);
2721 }
2722 if (mapping)
2723 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002724out:
2725 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2726 return ret;
2727}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002728
2729void free_transhuge_page(struct page *page)
2730{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002731 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002732 unsigned long flags;
2733
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002734 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002735 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002736 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002737 list_del(page_deferred_list(page));
2738 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002739 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002740 free_compound_page(page);
2741}
2742
2743void deferred_split_huge_page(struct page *page)
2744{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002745 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002746 unsigned long flags;
2747
2748 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2749
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002750 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002751 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002752 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002753 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2754 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002755 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002756 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002757}
2758
2759static unsigned long deferred_split_count(struct shrinker *shrink,
2760 struct shrink_control *sc)
2761{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002762 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Mark Rutland6aa7de02017-10-23 14:07:29 -07002763 return READ_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002764}
2765
2766static unsigned long deferred_split_scan(struct shrinker *shrink,
2767 struct shrink_control *sc)
2768{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002769 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002770 unsigned long flags;
2771 LIST_HEAD(list), *pos, *next;
2772 struct page *page;
2773 int split = 0;
2774
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002775 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002776 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002777 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002778 page = list_entry((void *)pos, struct page, mapping);
2779 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002780 if (get_page_unless_zero(page)) {
2781 list_move(page_deferred_list(page), &list);
2782 } else {
2783 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002784 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002785 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002786 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002787 if (!--sc->nr_to_scan)
2788 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002789 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002790 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002791
2792 list_for_each_safe(pos, next, &list) {
2793 page = list_entry((void *)pos, struct page, mapping);
2794 lock_page(page);
2795 /* split_huge_page() removes page from list on success */
2796 if (!split_huge_page(page))
2797 split++;
2798 unlock_page(page);
2799 put_page(page);
2800 }
2801
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002802 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2803 list_splice_tail(&list, &pgdata->split_queue);
2804 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002805
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002806 /*
2807 * Stop shrinker if we didn't split any page, but the queue is empty.
2808 * This can happen if pages were freed under us.
2809 */
2810 if (!split && list_empty(&pgdata->split_queue))
2811 return SHRINK_STOP;
2812 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002813}
2814
2815static struct shrinker deferred_split_shrinker = {
2816 .count_objects = deferred_split_count,
2817 .scan_objects = deferred_split_scan,
2818 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002819 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002820};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002821
2822#ifdef CONFIG_DEBUG_FS
2823static int split_huge_pages_set(void *data, u64 val)
2824{
2825 struct zone *zone;
2826 struct page *page;
2827 unsigned long pfn, max_zone_pfn;
2828 unsigned long total = 0, split = 0;
2829
2830 if (val != 1)
2831 return -EINVAL;
2832
2833 for_each_populated_zone(zone) {
2834 max_zone_pfn = zone_end_pfn(zone);
2835 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2836 if (!pfn_valid(pfn))
2837 continue;
2838
2839 page = pfn_to_page(pfn);
2840 if (!get_page_unless_zero(page))
2841 continue;
2842
2843 if (zone != page_zone(page))
2844 goto next;
2845
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002846 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002847 goto next;
2848
2849 total++;
2850 lock_page(page);
2851 if (!split_huge_page(page))
2852 split++;
2853 unlock_page(page);
2854next:
2855 put_page(page);
2856 }
2857 }
2858
Yang Shi145bdaa2016-05-05 16:22:00 -07002859 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002860
2861 return 0;
2862}
2863DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2864 "%llu\n");
2865
2866static int __init split_huge_pages_debugfs(void)
2867{
2868 void *ret;
2869
Yang Shi145bdaa2016-05-05 16:22:00 -07002870 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002871 &split_huge_pages_fops);
2872 if (!ret)
2873 pr_warn("Failed to create split_huge_pages in debugfs");
2874 return 0;
2875}
2876late_initcall(split_huge_pages_debugfs);
2877#endif
Zi Yan616b8372017-09-08 16:10:57 -07002878
2879#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2880void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2881 struct page *page)
2882{
2883 struct vm_area_struct *vma = pvmw->vma;
2884 struct mm_struct *mm = vma->vm_mm;
2885 unsigned long address = pvmw->address;
2886 pmd_t pmdval;
2887 swp_entry_t entry;
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002888 pmd_t pmdswp;
Zi Yan616b8372017-09-08 16:10:57 -07002889
2890 if (!(pvmw->pmd && !pvmw->pte))
2891 return;
2892
2893 mmu_notifier_invalidate_range_start(mm, address,
2894 address + HPAGE_PMD_SIZE);
2895
2896 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2897 pmdval = *pvmw->pmd;
2898 pmdp_invalidate(vma, address, pvmw->pmd);
2899 if (pmd_dirty(pmdval))
2900 set_page_dirty(page);
2901 entry = make_migration_entry(page, pmd_write(pmdval));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002902 pmdswp = swp_entry_to_pmd(entry);
2903 if (pmd_soft_dirty(pmdval))
2904 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2905 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
Zi Yan616b8372017-09-08 16:10:57 -07002906 page_remove_rmap(page, true);
2907 put_page(page);
2908
2909 mmu_notifier_invalidate_range_end(mm, address,
2910 address + HPAGE_PMD_SIZE);
2911}
2912
2913void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2914{
2915 struct vm_area_struct *vma = pvmw->vma;
2916 struct mm_struct *mm = vma->vm_mm;
2917 unsigned long address = pvmw->address;
2918 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2919 pmd_t pmde;
2920 swp_entry_t entry;
2921
2922 if (!(pvmw->pmd && !pvmw->pte))
2923 return;
2924
2925 entry = pmd_to_swp_entry(*pvmw->pmd);
2926 get_page(new);
2927 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002928 if (pmd_swp_soft_dirty(*pvmw->pmd))
2929 pmde = pmd_mksoft_dirty(pmde);
Zi Yan616b8372017-09-08 16:10:57 -07002930 if (is_write_migration_entry(entry))
Linus Torvaldsf55e1012017-11-29 09:01:01 -08002931 pmde = maybe_pmd_mkwrite(pmde, vma);
Zi Yan616b8372017-09-08 16:10:57 -07002932
2933 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2934 page_add_anon_rmap(new, vma, mmun_start, true);
2935 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2936 if (vma->vm_flags & VM_LOCKED)
2937 mlock_vma_page(new);
2938 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2939}
2940#endif