blob: d2cd70cfaa901a7dc1b31f52d2315410545999d4 [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{
Matthew Wilcoxfa3015b2018-06-07 17:08:42 -0700486 /* ->lru in the tail pages is occupied by compound_head. */
487 return &page[2].deferred_list;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800488}
489
490void prep_transhuge_page(struct page *page)
491{
492 /*
493 * we use page->mapping and page->indexlru in second tail page
494 * as list_head: assuming THP order >= 2
495 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800496
497 INIT_LIST_HEAD(page_deferred_list(page));
498 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
499}
500
Toshi Kani74d2fad2016-10-07 16:59:56 -0700501unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
502 loff_t off, unsigned long flags, unsigned long size)
503{
504 unsigned long addr;
505 loff_t off_end = off + len;
506 loff_t off_align = round_up(off, size);
507 unsigned long len_pad;
508
509 if (off_end <= off_align || (off_end - off_align) < size)
510 return 0;
511
512 len_pad = len + size;
513 if (len_pad < len || (off + len_pad) < off)
514 return 0;
515
516 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
517 off >> PAGE_SHIFT, flags);
518 if (IS_ERR_VALUE(addr))
519 return 0;
520
521 addr += (off - addr) & (size - 1);
522 return addr;
523}
524
525unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
526 unsigned long len, unsigned long pgoff, unsigned long flags)
527{
528 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
529
530 if (addr)
531 goto out;
532 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
533 goto out;
534
535 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
536 if (addr)
537 return addr;
538
539 out:
540 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
541}
542EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
543
Souptick Joarder2b740302018-08-23 17:01:36 -0700544static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
545 struct page *page, gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800546{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800547 struct vm_area_struct *vma = vmf->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700548 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800549 pgtable_t pgtable;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800550 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Souptick Joarder2b740302018-08-23 17:01:36 -0700551 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800552
Sasha Levin309381fea2014-01-23 15:52:54 -0800553 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700554
Tejun Heo2cf85582018-07-03 11:14:56 -0400555 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700556 put_page(page);
557 count_vm_event(THP_FAULT_FALLBACK);
558 return VM_FAULT_FALLBACK;
559 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800560
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700561 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700562 if (unlikely(!pgtable)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700563 ret = VM_FAULT_OOM;
564 goto release;
Johannes Weiner00501b52014-08-08 14:19:20 -0700565 }
566
Huang Yingc79b57e2017-09-06 16:25:04 -0700567 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700568 /*
569 * The memory barrier inside __SetPageUptodate makes sure that
570 * clear_huge_page writes become visible before the set_pmd_at()
571 * write.
572 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800573 __SetPageUptodate(page);
574
Jan Kara82b0f8c2016-12-14 15:06:58 -0800575 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
576 if (unlikely(!pmd_none(*vmf->pmd))) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700577 goto unlock_release;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800578 } else {
579 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700580
Michal Hocko6b31d592017-08-18 15:16:15 -0700581 ret = check_stable_address_space(vma->vm_mm);
582 if (ret)
583 goto unlock_release;
584
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700585 /* Deliver the page fault to userland */
586 if (userfaultfd_missing(vma)) {
Souptick Joarder2b740302018-08-23 17:01:36 -0700587 vm_fault_t ret2;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700588
Jan Kara82b0f8c2016-12-14 15:06:58 -0800589 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800590 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700591 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700592 pte_free(vma->vm_mm, pgtable);
Souptick Joarder2b740302018-08-23 17:01:36 -0700593 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
594 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
595 return ret2;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700596 }
597
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700598 entry = mk_huge_pmd(page, vma->vm_page_prot);
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800599 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800600 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800601 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700602 lru_cache_add_active_or_unevictable(page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800603 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700605 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800606 mm_inc_nr_ptes(vma->vm_mm);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800607 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700608 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800609 }
610
David Rientjesaa2e8782012-05-29 15:06:17 -0700611 return 0;
Michal Hocko6b31d592017-08-18 15:16:15 -0700612unlock_release:
613 spin_unlock(vmf->ptl);
614release:
615 if (pgtable)
616 pte_free(vma->vm_mm, pgtable);
617 mem_cgroup_cancel_charge(page, memcg, true);
618 put_page(page);
619 return ret;
620
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800621}
622
Mel Gorman444eb2a42016-03-17 14:19:23 -0700623/*
David Rientjes21440d72017-02-22 15:45:49 -0800624 * always: directly stall for all thp allocations
625 * defer: wake kswapd and fail if not immediately available
626 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
627 * fail if not immediately available
628 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
629 * available
630 * never: never stall for any thp allocation
Mel Gorman444eb2a42016-03-17 14:19:23 -0700631 */
632static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800633{
David Rientjes21440d72017-02-22 15:45:49 -0800634 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700635
David Rientjes21440d72017-02-22 15:45:49 -0800636 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
Vlastimil Babka25160352016-07-28 15:49:25 -0700637 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
David Rientjes21440d72017-02-22 15:45:49 -0800638 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
639 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
640 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
641 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
642 __GFP_KSWAPD_RECLAIM);
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645 0);
Vlastimil Babka25160352016-07-28 15:49:25 -0700646 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700647}
648
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800649/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700650static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800651 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700652 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800653{
654 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700655 if (!pmd_none(*pmd))
656 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700657 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800658 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800659 if (pgtable)
660 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800661 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800662 mm_inc_nr_ptes(mm);
Andrew Morton7c414162015-09-08 14:58:43 -0700663 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800664}
665
Souptick Joarder2b740302018-08-23 17:01:36 -0700666vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800667{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800668 struct vm_area_struct *vma = vmf->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800669 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800670 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800671 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800672
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700673 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700674 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700675 if (unlikely(anon_vma_prepare(vma)))
676 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700677 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700678 return VM_FAULT_OOM;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800679 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700680 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700681 transparent_hugepage_use_zero_page()) {
682 pgtable_t pgtable;
683 struct page *zero_page;
684 bool set;
Souptick Joarder2b740302018-08-23 17:01:36 -0700685 vm_fault_t ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700686 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700687 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800688 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700689 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700690 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700691 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700692 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700693 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700694 }
Jan Kara82b0f8c2016-12-14 15:06:58 -0800695 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700696 ret = 0;
697 set = false;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800698 if (pmd_none(*vmf->pmd)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700699 ret = check_stable_address_space(vma->vm_mm);
700 if (ret) {
701 spin_unlock(vmf->ptl);
702 } else if (userfaultfd_missing(vma)) {
Jan Kara82b0f8c2016-12-14 15:06:58 -0800703 spin_unlock(vmf->ptl);
704 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700705 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
706 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700707 set_huge_zero_page(pgtable, vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -0800708 haddr, vmf->pmd, zero_page);
709 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700710 set = true;
711 }
712 } else
Jan Kara82b0f8c2016-12-14 15:06:58 -0800713 spin_unlock(vmf->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700714 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700715 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700716 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800717 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700718 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800719 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700720 if (unlikely(!page)) {
721 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700722 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700723 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800724 prep_transhuge_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800725 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800726}
727
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700728static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700729 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
730 pgtable_t pgtable)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700731{
732 struct mm_struct *mm = vma->vm_mm;
733 pmd_t entry;
734 spinlock_t *ptl;
735
736 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800737 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
738 if (pfn_t_devmap(pfn))
739 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800740 if (write) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800741 entry = pmd_mkyoung(pmd_mkdirty(entry));
742 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700743 }
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700744
745 if (pgtable) {
746 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800747 mm_inc_nr_ptes(mm);
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700748 }
749
Ross Zwisler01871e52016-01-15 16:56:02 -0800750 set_pmd_at(mm, addr, pmd, entry);
751 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700752 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700753}
754
Dan Williams226ab562018-07-13 21:49:34 -0700755vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800756 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700757{
758 pgprot_t pgprot = vma->vm_page_prot;
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700759 pgtable_t pgtable = NULL;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700760 /*
761 * If we had pmd_special, we could avoid all these restrictions,
762 * but we need to be consistent with PTEs and architectures that
763 * can't support a 'special' bit.
764 */
Dave Jiange1fb4a02018-08-17 15:43:40 -0700765 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
766 !pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700767 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
768 (VM_PFNMAP|VM_MIXEDMAP));
769 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700770
771 if (addr < vma->vm_start || addr >= vma->vm_end)
772 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200773
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700774 if (arch_needs_pgtable_deposit()) {
775 pgtable = pte_alloc_one(vma->vm_mm, addr);
776 if (!pgtable)
777 return VM_FAULT_OOM;
778 }
779
Borislav Petkov308a0472016-10-26 19:43:43 +0200780 track_pfn_insert(vma, &pgprot, pfn);
781
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700782 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700783 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700784}
Dan Williamsdee41072016-05-14 12:20:44 -0700785EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700786
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800787#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800788static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800789{
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800790 if (likely(vma->vm_flags & VM_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800791 pud = pud_mkwrite(pud);
792 return pud;
793}
794
795static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
796 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
797{
798 struct mm_struct *mm = vma->vm_mm;
799 pud_t entry;
800 spinlock_t *ptl;
801
802 ptl = pud_lock(mm, pud);
803 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
804 if (pfn_t_devmap(pfn))
805 entry = pud_mkdevmap(entry);
806 if (write) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800807 entry = pud_mkyoung(pud_mkdirty(entry));
808 entry = maybe_pud_mkwrite(entry, vma);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800809 }
810 set_pud_at(mm, addr, pud, entry);
811 update_mmu_cache_pud(vma, addr, pud);
812 spin_unlock(ptl);
813}
814
Dan Williams226ab562018-07-13 21:49:34 -0700815vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800816 pud_t *pud, pfn_t pfn, bool write)
817{
818 pgprot_t pgprot = vma->vm_page_prot;
819 /*
820 * If we had pud_special, we could avoid all these restrictions,
821 * but we need to be consistent with PTEs and architectures that
822 * can't support a 'special' bit.
823 */
Dave Jiang62ec0d82018-09-04 15:46:16 -0700824 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
825 !pfn_t_devmap(pfn));
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800826 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
827 (VM_PFNMAP|VM_MIXEDMAP));
828 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800829
830 if (addr < vma->vm_start || addr >= vma->vm_end)
831 return VM_FAULT_SIGBUS;
832
833 track_pfn_insert(vma, &pgprot, pfn);
834
835 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
836 return VM_FAULT_NOPAGE;
837}
838EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
839#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
840
Dan Williams3565fce2016-01-15 16:56:55 -0800841static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300842 pmd_t *pmd, int flags)
Dan Williams3565fce2016-01-15 16:56:55 -0800843{
844 pmd_t _pmd;
845
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300846 _pmd = pmd_mkyoung(*pmd);
847 if (flags & FOLL_WRITE)
848 _pmd = pmd_mkdirty(_pmd);
Dan Williams3565fce2016-01-15 16:56:55 -0800849 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300850 pmd, _pmd, flags & FOLL_WRITE))
Dan Williams3565fce2016-01-15 16:56:55 -0800851 update_mmu_cache_pmd(vma, addr, pmd);
852}
853
854struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
855 pmd_t *pmd, int flags)
856{
857 unsigned long pfn = pmd_pfn(*pmd);
858 struct mm_struct *mm = vma->vm_mm;
859 struct dev_pagemap *pgmap;
860 struct page *page;
861
862 assert_spin_locked(pmd_lockptr(mm, pmd));
863
Keno Fischer8310d482017-01-24 15:17:48 -0800864 /*
865 * When we COW a devmap PMD entry, we split it into PTEs, so we should
866 * not be in this function with `flags & FOLL_COW` set.
867 */
868 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
869
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800870 if (flags & FOLL_WRITE && !pmd_write(*pmd))
Dan Williams3565fce2016-01-15 16:56:55 -0800871 return NULL;
872
873 if (pmd_present(*pmd) && pmd_devmap(*pmd))
874 /* pass */;
875 else
876 return NULL;
877
878 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300879 touch_pmd(vma, addr, pmd, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800880
881 /*
882 * device mapped pages can only be returned if the
883 * caller will manage the page reference count.
884 */
885 if (!(flags & FOLL_GET))
886 return ERR_PTR(-EEXIST);
887
888 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
889 pgmap = get_dev_pagemap(pfn, NULL);
890 if (!pgmap)
891 return ERR_PTR(-EFAULT);
892 page = pfn_to_page(pfn);
893 get_page(page);
894 put_dev_pagemap(pgmap);
895
896 return page;
897}
898
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800899int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
900 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
901 struct vm_area_struct *vma)
902{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800903 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800904 struct page *src_page;
905 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800906 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700907 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800908
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700909 /* Skip if can be re-fill on fault */
910 if (!vma_is_anonymous(vma))
911 return 0;
912
913 pgtable = pte_alloc_one(dst_mm, addr);
914 if (unlikely(!pgtable))
915 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800916
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800917 dst_ptl = pmd_lock(dst_mm, dst_pmd);
918 src_ptl = pmd_lockptr(src_mm, src_pmd);
919 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800920
921 ret = -EAGAIN;
922 pmd = *src_pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -0700923
924#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
925 if (unlikely(is_swap_pmd(pmd))) {
926 swp_entry_t entry = pmd_to_swp_entry(pmd);
927
928 VM_BUG_ON(!is_pmd_migration_entry(pmd));
929 if (is_write_migration_entry(entry)) {
930 make_migration_entry_read(&entry);
931 pmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -0700932 if (pmd_swp_soft_dirty(*src_pmd))
933 pmd = pmd_swp_mksoft_dirty(pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -0700934 set_pmd_at(src_mm, addr, src_pmd, pmd);
935 }
Zi Yandd8a67f2017-11-02 15:59:47 -0700936 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovaf5b0f62017-11-15 17:35:40 -0800937 mm_inc_nr_ptes(dst_mm);
Zi Yandd8a67f2017-11-02 15:59:47 -0700938 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Zi Yan84c3fc42017-09-08 16:11:01 -0700939 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
940 ret = 0;
941 goto out_unlock;
942 }
943#endif
944
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700945 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800946 pte_free(dst_mm, pgtable);
947 goto out_unlock;
948 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800949 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800950 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800951 * under splitting since we don't split the page itself, only pmd to
952 * a page table.
953 */
954 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700955 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800956 /*
957 * get_huge_zero_page() will never allocate a new page here,
958 * since we already have a zero page to copy. It just takes a
959 * reference.
960 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700961 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700962 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700963 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800964 ret = 0;
965 goto out_unlock;
966 }
Mel Gormande466bd2013-12-18 17:08:42 -0800967
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700968 src_page = pmd_page(pmd);
969 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
970 get_page(src_page);
971 page_dup_rmap(src_page, true);
972 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800973 mm_inc_nr_ptes(dst_mm);
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700974 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800975
976 pmdp_set_wrprotect(src_mm, addr, src_pmd);
977 pmd = pmd_mkold(pmd_wrprotect(pmd));
978 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800979
980 ret = 0;
981out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800982 spin_unlock(src_ptl);
983 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800984out:
985 return ret;
986}
987
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800988#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
989static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300990 pud_t *pud, int flags)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800991{
992 pud_t _pud;
993
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300994 _pud = pud_mkyoung(*pud);
995 if (flags & FOLL_WRITE)
996 _pud = pud_mkdirty(_pud);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800997 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300998 pud, _pud, flags & FOLL_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800999 update_mmu_cache_pud(vma, addr, pud);
1000}
1001
1002struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1003 pud_t *pud, int flags)
1004{
1005 unsigned long pfn = pud_pfn(*pud);
1006 struct mm_struct *mm = vma->vm_mm;
1007 struct dev_pagemap *pgmap;
1008 struct page *page;
1009
1010 assert_spin_locked(pud_lockptr(mm, pud));
1011
Linus Torvaldsf6f37322017-12-15 18:53:22 -08001012 if (flags & FOLL_WRITE && !pud_write(*pud))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001013 return NULL;
1014
1015 if (pud_present(*pud) && pud_devmap(*pud))
1016 /* pass */;
1017 else
1018 return NULL;
1019
1020 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001021 touch_pud(vma, addr, pud, flags);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001022
1023 /*
1024 * device mapped pages can only be returned if the
1025 * caller will manage the page reference count.
1026 */
1027 if (!(flags & FOLL_GET))
1028 return ERR_PTR(-EEXIST);
1029
1030 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1031 pgmap = get_dev_pagemap(pfn, NULL);
1032 if (!pgmap)
1033 return ERR_PTR(-EFAULT);
1034 page = pfn_to_page(pfn);
1035 get_page(page);
1036 put_dev_pagemap(pgmap);
1037
1038 return page;
1039}
1040
1041int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1042 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1043 struct vm_area_struct *vma)
1044{
1045 spinlock_t *dst_ptl, *src_ptl;
1046 pud_t pud;
1047 int ret;
1048
1049 dst_ptl = pud_lock(dst_mm, dst_pud);
1050 src_ptl = pud_lockptr(src_mm, src_pud);
1051 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1052
1053 ret = -EAGAIN;
1054 pud = *src_pud;
1055 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1056 goto out_unlock;
1057
1058 /*
1059 * When page table lock is held, the huge zero pud should not be
1060 * under splitting since we don't split the page itself, only pud to
1061 * a page table.
1062 */
1063 if (is_huge_zero_pud(pud)) {
1064 /* No huge zero pud yet */
1065 }
1066
1067 pudp_set_wrprotect(src_mm, addr, src_pud);
1068 pud = pud_mkold(pud_wrprotect(pud));
1069 set_pud_at(dst_mm, addr, dst_pud, pud);
1070
1071 ret = 0;
1072out_unlock:
1073 spin_unlock(src_ptl);
1074 spin_unlock(dst_ptl);
1075 return ret;
1076}
1077
1078void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1079{
1080 pud_t entry;
1081 unsigned long haddr;
1082 bool write = vmf->flags & FAULT_FLAG_WRITE;
1083
1084 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1085 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1086 goto unlock;
1087
1088 entry = pud_mkyoung(orig_pud);
1089 if (write)
1090 entry = pud_mkdirty(entry);
1091 haddr = vmf->address & HPAGE_PUD_MASK;
1092 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1093 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1094
1095unlock:
1096 spin_unlock(vmf->ptl);
1097}
1098#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1099
Jan Kara82b0f8c2016-12-14 15:06:58 -08001100void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001101{
1102 pmd_t entry;
1103 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001104 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001105
Jan Kara82b0f8c2016-12-14 15:06:58 -08001106 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1107 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001108 goto unlock;
1109
1110 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001111 if (write)
1112 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001113 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001114 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001115 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001116
1117unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001118 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001119}
1120
Souptick Joarder2b740302018-08-23 17:01:36 -07001121static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1122 pmd_t orig_pmd, struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001123{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001124 struct vm_area_struct *vma = vmf->vma;
1125 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001126 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001127 pgtable_t pgtable;
1128 pmd_t _pmd;
Souptick Joarder2b740302018-08-23 17:01:36 -07001129 int i;
1130 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001131 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001132 unsigned long mmun_start; /* For mmu_notifiers */
1133 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134
Kees Cook6da2ec52018-06-12 13:55:00 -07001135 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1136 GFP_KERNEL);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001137 if (unlikely(!pages)) {
1138 ret |= VM_FAULT_OOM;
1139 goto out;
1140 }
1141
1142 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001143 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001144 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001145 if (unlikely(!pages[i] ||
Tejun Heo2cf85582018-07-03 11:14:56 -04001146 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001147 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001148 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001149 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001150 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001151 memcg = (void *)page_private(pages[i]);
1152 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001153 mem_cgroup_cancel_charge(pages[i], memcg,
1154 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001155 put_page(pages[i]);
1156 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001157 kfree(pages);
1158 ret |= VM_FAULT_OOM;
1159 goto out;
1160 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001161 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001162 }
1163
1164 for (i = 0; i < HPAGE_PMD_NR; i++) {
1165 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001166 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001167 __SetPageUptodate(pages[i]);
1168 cond_resched();
1169 }
1170
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001171 mmun_start = haddr;
1172 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001173 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001174
Jan Kara82b0f8c2016-12-14 15:06:58 -08001175 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1176 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001177 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001178 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001179
Jérôme Glisse0f108512017-11-15 17:34:07 -08001180 /*
1181 * Leave pmd empty until pte is filled note we must notify here as
1182 * concurrent CPU thread might write to new page before the call to
1183 * mmu_notifier_invalidate_range_end() happens which can lead to a
1184 * device seeing memory write in different order than CPU.
1185 *
Mike Rapoportad56b732018-03-21 21:22:47 +02001186 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08001187 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001188 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001189
Jan Kara82b0f8c2016-12-14 15:06:58 -08001190 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001191 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001192
1193 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001194 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001195 entry = mk_pte(pages[i], vma->vm_page_prot);
1196 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001197 memcg = (void *)page_private(pages[i]);
1198 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001199 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001200 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001201 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001202 vmf->pte = pte_offset_map(&_pmd, haddr);
1203 VM_BUG_ON(!pte_none(*vmf->pte));
1204 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1205 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001206 }
1207 kfree(pages);
1208
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001210 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001211 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001212 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001213
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001214 /*
1215 * No need to double call mmu_notifier->invalidate_range() callback as
1216 * the above pmdp_huge_clear_flush_notify() did already call it.
1217 */
1218 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1219 mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001220
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001221 ret |= VM_FAULT_WRITE;
1222 put_page(page);
1223
1224out:
1225 return ret;
1226
1227out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001228 spin_unlock(vmf->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001229 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001230 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001231 memcg = (void *)page_private(pages[i]);
1232 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001233 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001234 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001235 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001236 kfree(pages);
1237 goto out;
1238}
1239
Souptick Joarder2b740302018-08-23 17:01:36 -07001240vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001241{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001242 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001243 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001244 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001245 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001246 unsigned long mmun_start; /* For mmu_notifiers */
1247 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001248 gfp_t huge_gfp; /* for allocation and charge */
Souptick Joarder2b740302018-08-23 17:01:36 -07001249 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001250
Jan Kara82b0f8c2016-12-14 15:06:58 -08001251 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001252 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001253 if (is_huge_zero_pmd(orig_pmd))
1254 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001255 spin_lock(vmf->ptl);
1256 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001257 goto out_unlock;
1258
1259 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001260 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001261 /*
1262 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001263 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001264 */
Huang Yingba3c4ce2017-09-06 16:22:19 -07001265 if (!trylock_page(page)) {
1266 get_page(page);
1267 spin_unlock(vmf->ptl);
1268 lock_page(page);
1269 spin_lock(vmf->ptl);
1270 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1271 unlock_page(page);
1272 put_page(page);
1273 goto out_unlock;
1274 }
1275 put_page(page);
1276 }
1277 if (reuse_swap_page(page, NULL)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001278 pmd_t entry;
1279 entry = pmd_mkyoung(orig_pmd);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001280 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001281 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1282 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001283 ret |= VM_FAULT_WRITE;
Huang Yingba3c4ce2017-09-06 16:22:19 -07001284 unlock_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001285 goto out_unlock;
1286 }
Huang Yingba3c4ce2017-09-06 16:22:19 -07001287 unlock_page(page);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001288 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001289 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001290alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001291 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001292 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001293 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001294 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001295 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001296 new_page = NULL;
1297
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001298 if (likely(new_page)) {
1299 prep_transhuge_page(new_page);
1300 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001301 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001302 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001303 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001304 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001305 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001306 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001307 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001308 ret |= VM_FAULT_FALLBACK;
1309 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001310 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001311 }
David Rientjes17766dd2013-09-12 15:14:06 -07001312 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001313 goto out;
1314 }
1315
Tejun Heo2cf85582018-07-03 11:14:56 -04001316 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
Michal Hocko2a70f6a2018-04-10 16:29:30 -07001317 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001318 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001319 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001320 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001321 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001322 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001323 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001324 goto out;
1325 }
1326
David Rientjes17766dd2013-09-12 15:14:06 -07001327 count_vm_event(THP_FAULT_ALLOC);
1328
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001329 if (!page)
Huang Yingc79b57e2017-09-06 16:25:04 -07001330 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001331 else
Huang Yingc9f4cd72018-08-17 15:45:49 -07001332 copy_user_huge_page(new_page, page, vmf->address,
1333 vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001334 __SetPageUptodate(new_page);
1335
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001336 mmun_start = haddr;
1337 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001338 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001339
Jan Kara82b0f8c2016-12-14 15:06:58 -08001340 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001341 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001342 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001343 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1344 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001345 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001346 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001347 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001348 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001349 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001350 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001351 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001352 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001353 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001354 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001355 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001356 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1357 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001358 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001359 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001360 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001361 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001362 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001363 put_page(page);
1364 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001365 ret |= VM_FAULT_WRITE;
1366 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001367 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001368out_mn:
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001369 /*
1370 * No need to double call mmu_notifier->invalidate_range() callback as
1371 * the above pmdp_huge_clear_flush_notify() did already call it.
1372 */
1373 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1374 mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001375out:
1376 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001377out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001378 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001379 return ret;
1380}
1381
Keno Fischer8310d482017-01-24 15:17:48 -08001382/*
1383 * FOLL_FORCE can write to even unwritable pmd's, but only
1384 * after we've gone through a COW cycle and they are dirty.
1385 */
1386static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1387{
Linus Torvaldsf6f37322017-12-15 18:53:22 -08001388 return pmd_write(pmd) ||
Keno Fischer8310d482017-01-24 15:17:48 -08001389 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1390}
1391
David Rientjesb676b292012-10-08 16:34:03 -07001392struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001393 unsigned long addr,
1394 pmd_t *pmd,
1395 unsigned int flags)
1396{
David Rientjesb676b292012-10-08 16:34:03 -07001397 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001398 struct page *page = NULL;
1399
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001400 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001401
Keno Fischer8310d482017-01-24 15:17:48 -08001402 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001403 goto out;
1404
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001405 /* Avoid dumping huge zero page */
1406 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1407 return ERR_PTR(-EFAULT);
1408
Mel Gorman2b4847e2013-12-18 17:08:32 -08001409 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001410 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001411 goto out;
1412
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001413 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001414 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001415 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001416 touch_pmd(vma, addr, pmd, flags);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001417 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001418 /*
1419 * We don't mlock() pte-mapped THPs. This way we can avoid
1420 * leaking mlocked pages into non-VM_LOCKED VMAs.
1421 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001422 * For anon THP:
1423 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001424 * In most cases the pmd is the only mapping of the page as we
1425 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1426 * writable private mappings in populate_vma_page_range().
1427 *
1428 * The only scenario when we have the page shared here is if we
1429 * mlocking read-only mapping shared over fork(). We skip
1430 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001431 *
1432 * For file THP:
1433 *
1434 * We can expect PageDoubleMap() to be stable under page lock:
1435 * for file pages we set it in page_add_file_rmap(), which
1436 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001437 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001438
1439 if (PageAnon(page) && compound_mapcount(page) != 1)
1440 goto skip_mlock;
1441 if (PageDoubleMap(page) || !page->mapping)
1442 goto skip_mlock;
1443 if (!trylock_page(page))
1444 goto skip_mlock;
1445 lru_add_drain();
1446 if (page->mapping && !PageDoubleMap(page))
1447 mlock_vma_page(page);
1448 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001449 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001450skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001451 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001452 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001453 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001454 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001455
1456out:
1457 return page;
1458}
1459
Mel Gormand10e63f2012-10-25 14:16:31 +02001460/* NUMA hinting page fault entry point for trans huge pmds */
Souptick Joarder2b740302018-08-23 17:01:36 -07001461vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001462{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001463 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001464 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001465 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001466 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001467 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001468 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001469 bool page_locked;
1470 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001471 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001472 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001473
Jan Kara82b0f8c2016-12-14 15:06:58 -08001474 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1475 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001476 goto out_unlock;
1477
Mel Gormande466bd2013-12-18 17:08:42 -08001478 /*
1479 * If there are potential migrations, wait for completion and retry
1480 * without disrupting NUMA hinting information. Do not relock and
1481 * check_same as the page may no longer be mapped.
1482 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001483 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1484 page = pmd_page(*vmf->pmd);
Mark Rutland3c226c62017-06-16 14:02:34 -07001485 if (!get_page_unless_zero(page))
1486 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001487 spin_unlock(vmf->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001488 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001489 put_page(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001490 goto out;
1491 }
1492
Mel Gormand10e63f2012-10-25 14:16:31 +02001493 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001494 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001495 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001496 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001497 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001498 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001499 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001500 flags |= TNF_FAULT_LOCAL;
1501 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001502
Mel Gormanbea66fb2015-03-25 15:55:37 -07001503 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001504 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001505 flags |= TNF_NO_GROUP;
1506
1507 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001508 * Acquire the page lock to serialise THP migrations but avoid dropping
1509 * page_table_lock if at all possible
1510 */
Mel Gormanb8916632013-10-07 11:28:44 +01001511 page_locked = trylock_page(page);
1512 target_nid = mpol_misplaced(page, vma, haddr);
1513 if (target_nid == -1) {
1514 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001515 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001516 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001517 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001518
Mel Gormande466bd2013-12-18 17:08:42 -08001519 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001520 if (!page_locked) {
Mark Rutland3c226c62017-06-16 14:02:34 -07001521 page_nid = -1;
1522 if (!get_page_unless_zero(page))
1523 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001524 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001525 wait_on_page_locked(page);
Mark Rutland3c226c62017-06-16 14:02:34 -07001526 put_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001527 goto out;
1528 }
1529
Mel Gorman2b4847e2013-12-18 17:08:32 -08001530 /*
1531 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1532 * to serialises splits
1533 */
Mel Gormanb8916632013-10-07 11:28:44 +01001534 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001535 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001536 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001537
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001538 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001539 spin_lock(vmf->ptl);
1540 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001541 unlock_page(page);
1542 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001543 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001544 goto out_unlock;
1545 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001546
Mel Gormanc3a489c2013-12-18 17:08:38 -08001547 /* Bail if we fail to protect against THP splits for any reason */
1548 if (unlikely(!anon_vma)) {
1549 put_page(page);
1550 page_nid = -1;
1551 goto clear_pmdnuma;
1552 }
1553
Mel Gormana54a4072013-10-07 11:28:46 +01001554 /*
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001555 * Since we took the NUMA fault, we must have observed the !accessible
1556 * bit. Make sure all other CPUs agree with that, to avoid them
1557 * modifying the page we're about to migrate.
1558 *
1559 * Must be done under PTL such that we'll observe the relevant
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001560 * inc_tlb_flush_pending().
1561 *
1562 * We are not sure a pending tlb flush here is for a huge page
1563 * mapping or not. Hence use the tlb range variant
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001564 */
1565 if (mm_tlb_flush_pending(vma->vm_mm))
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001566 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001567
1568 /*
Mel Gormana54a4072013-10-07 11:28:46 +01001569 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001570 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001571 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001572 spin_unlock(vmf->ptl);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001573
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001574 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001575 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001576 if (migrated) {
1577 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001578 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001579 } else
1580 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001581
Mel Gorman8191acb2013-10-07 11:28:45 +01001582 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001583clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001584 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001585 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001586 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001587 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001588 if (was_writable)
1589 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001590 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1591 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001592 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001593out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001594 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001595
1596out:
1597 if (anon_vma)
1598 page_unlock_anon_vma_read(anon_vma);
1599
Mel Gorman8191acb2013-10-07 11:28:45 +01001600 if (page_nid != -1)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001601 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001602 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001603
Mel Gormand10e63f2012-10-25 14:16:31 +02001604 return 0;
1605}
1606
Huang Ying319904a2016-07-28 15:48:03 -07001607/*
1608 * Return true if we do MADV_FREE successfully on entire pmd page.
1609 * Otherwise, return false.
1610 */
1611bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001612 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001613{
1614 spinlock_t *ptl;
1615 pmd_t orig_pmd;
1616 struct page *page;
1617 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001618 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001619
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001620 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1621
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001622 ptl = pmd_trans_huge_lock(pmd, vma);
1623 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001624 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001625
1626 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001627 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001628 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001629
Zi Yan84c3fc42017-09-08 16:11:01 -07001630 if (unlikely(!pmd_present(orig_pmd))) {
1631 VM_BUG_ON(thp_migration_supported() &&
1632 !is_pmd_migration_entry(orig_pmd));
1633 goto out;
1634 }
1635
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001636 page = pmd_page(orig_pmd);
1637 /*
1638 * If other processes are mapping this page, we couldn't discard
1639 * the page unless they all do MADV_FREE so let's skip the page.
1640 */
1641 if (page_mapcount(page) != 1)
1642 goto out;
1643
1644 if (!trylock_page(page))
1645 goto out;
1646
1647 /*
1648 * If user want to discard part-pages of THP, split it so MADV_FREE
1649 * will deactivate only them.
1650 */
1651 if (next - addr != HPAGE_PMD_SIZE) {
1652 get_page(page);
1653 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001654 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001655 unlock_page(page);
Kirill A. Shutemovbbf29ff2017-07-06 15:35:28 -07001656 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001657 goto out_unlocked;
1658 }
1659
1660 if (PageDirty(page))
1661 ClearPageDirty(page);
1662 unlock_page(page);
1663
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001664 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemov58ceeb62017-04-13 14:56:26 -07001665 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001666 orig_pmd = pmd_mkold(orig_pmd);
1667 orig_pmd = pmd_mkclean(orig_pmd);
1668
1669 set_pmd_at(mm, addr, pmd, orig_pmd);
1670 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1671 }
Shaohua Li802a3a92017-05-03 14:52:32 -07001672
1673 mark_page_lazyfree(page);
Huang Ying319904a2016-07-28 15:48:03 -07001674 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001675out:
1676 spin_unlock(ptl);
1677out_unlocked:
1678 return ret;
1679}
1680
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001681static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1682{
1683 pgtable_t pgtable;
1684
1685 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1686 pte_free(mm, pgtable);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -08001687 mm_dec_nr_ptes(mm);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001688}
1689
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001690int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001691 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001692{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001693 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001694 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001695
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001696 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1697
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001698 ptl = __pmd_trans_huge_lock(pmd, vma);
1699 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001700 return 0;
1701 /*
1702 * For architectures like ppc64 we look at deposited pgtable
1703 * when calling pmdp_huge_get_and_clear. So do the
1704 * pgtable_trans_huge_withdraw after finishing pmdp related
1705 * operations.
1706 */
1707 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1708 tlb->fullmm);
1709 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1710 if (vma_is_dax(vma)) {
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -07001711 if (arch_needs_pgtable_deposit())
1712 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001713 spin_unlock(ptl);
1714 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001715 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001716 } else if (is_huge_zero_pmd(orig_pmd)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001717 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001718 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001719 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001720 } else {
Zi Yan616b8372017-09-08 16:10:57 -07001721 struct page *page = NULL;
1722 int flush_needed = 1;
1723
1724 if (pmd_present(orig_pmd)) {
1725 page = pmd_page(orig_pmd);
1726 page_remove_rmap(page, true);
1727 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1728 VM_BUG_ON_PAGE(!PageHead(page), page);
1729 } else if (thp_migration_supported()) {
1730 swp_entry_t entry;
1731
1732 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1733 entry = pmd_to_swp_entry(orig_pmd);
1734 page = pfn_to_page(swp_offset(entry));
1735 flush_needed = 0;
1736 } else
1737 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1738
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001739 if (PageAnon(page)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001740 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001741 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1742 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001743 if (arch_needs_pgtable_deposit())
1744 zap_deposited_table(tlb->mm, pmd);
Yang Shifadae292018-08-17 15:44:55 -07001745 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001746 }
Zi Yan616b8372017-09-08 16:10:57 -07001747
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001748 spin_unlock(ptl);
Zi Yan616b8372017-09-08 16:10:57 -07001749 if (flush_needed)
1750 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001751 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001752 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001753}
1754
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001755#ifndef pmd_move_must_withdraw
1756static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1757 spinlock_t *old_pmd_ptl,
1758 struct vm_area_struct *vma)
1759{
1760 /*
1761 * With split pmd lock we also need to move preallocated
1762 * PTE page table if new_pmd is on different PMD page table.
1763 *
1764 * We also don't deposit and withdraw tables for file pages.
1765 */
1766 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1767}
1768#endif
1769
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001770static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1771{
1772#ifdef CONFIG_MEM_SOFT_DIRTY
1773 if (unlikely(is_pmd_migration_entry(pmd)))
1774 pmd = pmd_swp_mksoft_dirty(pmd);
1775 else if (pmd_present(pmd))
1776 pmd = pmd_mksoft_dirty(pmd);
1777#endif
1778 return pmd;
1779}
1780
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001781bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001782 unsigned long new_addr, unsigned long old_end,
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001783 pmd_t *old_pmd, pmd_t *new_pmd)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001784{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001785 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001786 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001787 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001788 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001789
1790 if ((old_addr & ~HPAGE_PMD_MASK) ||
1791 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001792 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001793 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001794
1795 /*
1796 * The destination pmd shouldn't be established, free_pgtables()
1797 * should have release it.
1798 */
1799 if (WARN_ON(!pmd_none(*new_pmd))) {
1800 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001801 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001802 }
1803
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001804 /*
1805 * We don't have to worry about the ordering of src and dst
1806 * ptlocks because exclusive mmap_sem prevents deadlock.
1807 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001808 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1809 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001810 new_ptl = pmd_lockptr(mm, new_pmd);
1811 if (new_ptl != old_ptl)
1812 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001813 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001814 if (pmd_present(pmd))
Aaron Lua2ce2662016-11-29 13:27:31 +08001815 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001816 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001817
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001818 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301819 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001820 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1821 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001822 }
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001823 pmd = move_soft_dirty_pmd(pmd);
1824 set_pmd_at(mm, new_addr, new_pmd, pmd);
Aaron Lu5d190422016-11-10 17:16:33 +08001825 if (force_flush)
1826 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001827 if (new_ptl != old_ptl)
1828 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001829 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001830 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001831 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001832 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001833}
1834
Mel Gormanf123d742013-10-07 11:28:49 +01001835/*
1836 * Returns
1837 * - 0 if PMD could not be locked
1838 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1839 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1840 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001841int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001842 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001843{
1844 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001845 spinlock_t *ptl;
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001846 pmd_t entry;
1847 bool preserve_write;
1848 int ret;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001849
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001850 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001851 if (!ptl)
1852 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001853
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001854 preserve_write = prot_numa && pmd_write(*pmd);
1855 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001856
Zi Yan84c3fc42017-09-08 16:11:01 -07001857#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1858 if (is_swap_pmd(*pmd)) {
1859 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1860
1861 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1862 if (is_write_migration_entry(entry)) {
1863 pmd_t newpmd;
1864 /*
1865 * A protection check is difficult so
1866 * just be safe and disable write
1867 */
1868 make_migration_entry_read(&entry);
1869 newpmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001870 if (pmd_swp_soft_dirty(*pmd))
1871 newpmd = pmd_swp_mksoft_dirty(newpmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001872 set_pmd_at(mm, addr, pmd, newpmd);
1873 }
1874 goto unlock;
1875 }
1876#endif
1877
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001878 /*
1879 * Avoid trapping faults against the zero page. The read-only
1880 * data is likely to be read-cached on the local CPU and
1881 * local/remote hits to the zero page are not interesting.
1882 */
1883 if (prot_numa && is_huge_zero_pmd(*pmd))
1884 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001885
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001886 if (prot_numa && pmd_protnone(*pmd))
1887 goto unlock;
1888
Kirill A. Shutemovced10802017-04-13 14:56:20 -07001889 /*
1890 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1891 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1892 * which is also under down_read(mmap_sem):
1893 *
1894 * CPU0: CPU1:
1895 * change_huge_pmd(prot_numa=1)
1896 * pmdp_huge_get_and_clear_notify()
1897 * madvise_dontneed()
1898 * zap_pmd_range()
1899 * pmd_trans_huge(*pmd) == 0 (without ptl)
1900 * // skip the pmd
1901 * set_pmd_at();
1902 * // pmd is re-established
1903 *
1904 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1905 * which may break userspace.
1906 *
1907 * pmdp_invalidate() is required to make sure we don't miss
1908 * dirty/young flags set by hardware.
1909 */
Kirill A. Shutemova3cf9882018-01-31 16:18:20 -08001910 entry = pmdp_invalidate(vma, addr, pmd);
Kirill A. Shutemovced10802017-04-13 14:56:20 -07001911
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001912 entry = pmd_modify(entry, newprot);
1913 if (preserve_write)
1914 entry = pmd_mk_savedwrite(entry);
1915 ret = HPAGE_PMD_NR;
1916 set_pmd_at(mm, addr, pmd, entry);
1917 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1918unlock:
1919 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001920 return ret;
1921}
1922
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001923/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001924 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001925 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001926 * Note that if it returns page table lock pointer, this routine returns without
1927 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001928 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001929spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001930{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001931 spinlock_t *ptl;
1932 ptl = pmd_lock(vma->vm_mm, pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001933 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1934 pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001935 return ptl;
1936 spin_unlock(ptl);
1937 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001938}
1939
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001940/*
1941 * Returns true if a given pud maps a thp, false otherwise.
1942 *
1943 * Note that if it returns true, this routine returns without unlocking page
1944 * table lock. So callers must unlock it.
1945 */
1946spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1947{
1948 spinlock_t *ptl;
1949
1950 ptl = pud_lock(vma->vm_mm, pud);
1951 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1952 return ptl;
1953 spin_unlock(ptl);
1954 return NULL;
1955}
1956
1957#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1958int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1959 pud_t *pud, unsigned long addr)
1960{
1961 pud_t orig_pud;
1962 spinlock_t *ptl;
1963
1964 ptl = __pud_trans_huge_lock(pud, vma);
1965 if (!ptl)
1966 return 0;
1967 /*
1968 * For architectures like ppc64 we look at deposited pgtable
1969 * when calling pudp_huge_get_and_clear. So do the
1970 * pgtable_trans_huge_withdraw after finishing pudp related
1971 * operations.
1972 */
1973 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1974 tlb->fullmm);
1975 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1976 if (vma_is_dax(vma)) {
1977 spin_unlock(ptl);
1978 /* No zero page support yet */
1979 } else {
1980 /* No support for anonymous PUD pages yet */
1981 BUG();
1982 }
1983 return 1;
1984}
1985
1986static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1987 unsigned long haddr)
1988{
1989 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1990 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1991 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1992 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1993
Yisheng Xiece9311c2017-03-09 16:17:00 -08001994 count_vm_event(THP_SPLIT_PUD);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001995
1996 pudp_huge_clear_flush_notify(vma, haddr, pud);
1997}
1998
1999void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2000 unsigned long address)
2001{
2002 spinlock_t *ptl;
2003 struct mm_struct *mm = vma->vm_mm;
2004 unsigned long haddr = address & HPAGE_PUD_MASK;
2005
2006 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2007 ptl = pud_lock(mm, pud);
2008 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2009 goto out;
2010 __split_huge_pud_locked(vma, pud, haddr);
2011
2012out:
2013 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002014 /*
2015 * No need to double call mmu_notifier->invalidate_range() callback as
2016 * the above pudp_huge_clear_flush_notify() did already call it.
2017 */
2018 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2019 HPAGE_PUD_SIZE);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002020}
2021#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2022
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002023static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2024 unsigned long haddr, pmd_t *pmd)
2025{
2026 struct mm_struct *mm = vma->vm_mm;
2027 pgtable_t pgtable;
2028 pmd_t _pmd;
2029 int i;
2030
Jérôme Glisse0f108512017-11-15 17:34:07 -08002031 /*
2032 * Leave pmd empty until pte is filled note that it is fine to delay
2033 * notification until mmu_notifier_invalidate_range_end() as we are
2034 * replacing a zero pmd write protected page with a zero pte write
2035 * protected page.
2036 *
Mike Rapoportad56b732018-03-21 21:22:47 +02002037 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08002038 */
2039 pmdp_huge_clear_flush(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002040
2041 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2042 pmd_populate(mm, &_pmd, pgtable);
2043
2044 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2045 pte_t *pte, entry;
2046 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2047 entry = pte_mkspecial(entry);
2048 pte = pte_offset_map(&_pmd, haddr);
2049 VM_BUG_ON(!pte_none(*pte));
2050 set_pte_at(mm, haddr, pte, entry);
2051 pte_unmap(pte);
2052 }
2053 smp_wmb(); /* make pte visible before pmd */
2054 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002055}
2056
2057static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002058 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002059{
2060 struct mm_struct *mm = vma->vm_mm;
2061 struct page *page;
2062 pgtable_t pgtable;
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002063 pmd_t old_pmd, _pmd;
Kirill A. Shutemova3cf9882018-01-31 16:18:20 -08002064 bool young, write, soft_dirty, pmd_migration = false;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002065 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002066 int i;
2067
2068 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2069 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2070 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Zi Yan84c3fc42017-09-08 16:11:01 -07002071 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2072 && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002073
2074 count_vm_event(THP_SPLIT_PMD);
2075
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002076 if (!vma_is_anonymous(vma)) {
2077 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08002078 /*
2079 * We are going to unmap this huge page. So
2080 * just go ahead and zap it
2081 */
2082 if (arch_needs_pgtable_deposit())
2083 zap_deposited_table(mm, pmd);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002084 if (vma_is_dax(vma))
2085 return;
2086 page = pmd_page(_pmd);
Hugh Dickinse1f1b152018-07-20 17:53:45 -07002087 if (!PageDirty(page) && pmd_dirty(_pmd))
2088 set_page_dirty(page);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002089 if (!PageReferenced(page) && pmd_young(_pmd))
2090 SetPageReferenced(page);
2091 page_remove_rmap(page, true);
2092 put_page(page);
Yang Shifadae292018-08-17 15:44:55 -07002093 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002094 return;
2095 } else if (is_huge_zero_pmd(*pmd)) {
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002096 /*
2097 * FIXME: Do we want to invalidate secondary mmu by calling
2098 * mmu_notifier_invalidate_range() see comments below inside
2099 * __split_huge_pmd() ?
2100 *
2101 * We are going from a zero huge page write protected to zero
2102 * small page also write protected so it does not seems useful
2103 * to invalidate secondary mmu at this time.
2104 */
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002105 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2106 }
2107
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002108 /*
2109 * Up to this point the pmd is present and huge and userland has the
2110 * whole access to the hugepage during the split (which happens in
2111 * place). If we overwrite the pmd with the not-huge version pointing
2112 * to the pte here (which of course we could if all CPUs were bug
2113 * free), userland could trigger a small page size TLB miss on the
2114 * small sized TLB while the hugepage TLB entry is still established in
2115 * the huge TLB. Some CPU doesn't like that.
2116 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2117 * 383 on page 93. Intel should be safe but is also warns that it's
2118 * only safe if the permission and cache attributes of the two entries
2119 * loaded in the two TLB is identical (which should be the case here).
2120 * But it is generally safer to never allow small and huge TLB entries
2121 * for the same virtual address to be loaded simultaneously. So instead
2122 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2123 * current pmd notpresent (atomically because here the pmd_trans_huge
2124 * must remain set at all times on the pmd until the split is complete
2125 * for this pmd), then we flush the SMP TLB and finally we write the
2126 * non-huge version of the pmd entry with pmd_populate.
2127 */
2128 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2129
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002130 pmd_migration = is_pmd_migration_entry(old_pmd);
Peter Xu161a5652018-12-21 14:30:50 -08002131 if (unlikely(pmd_migration)) {
Zi Yan84c3fc42017-09-08 16:11:01 -07002132 swp_entry_t entry;
2133
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002134 entry = pmd_to_swp_entry(old_pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07002135 page = pfn_to_page(swp_offset(entry));
Peter Xu161a5652018-12-21 14:30:50 -08002136 write = is_write_migration_entry(entry);
2137 young = false;
2138 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2139 } else {
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002140 page = pmd_page(old_pmd);
Peter Xu161a5652018-12-21 14:30:50 -08002141 if (pmd_dirty(old_pmd))
2142 SetPageDirty(page);
2143 write = pmd_write(old_pmd);
2144 young = pmd_young(old_pmd);
2145 soft_dirty = pmd_soft_dirty(old_pmd);
2146 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002147 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002148 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002149
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002150 /*
2151 * Withdraw the table only after we mark the pmd entry invalid.
2152 * This's critical for some architectures (Power).
2153 */
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002154 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2155 pmd_populate(mm, &_pmd, pgtable);
2156
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002157 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002158 pte_t entry, *pte;
2159 /*
2160 * Note that NUMA hinting access restrictions are not
2161 * transferred to avoid any possibility of altering
2162 * permissions across VMAs.
2163 */
Zi Yan84c3fc42017-09-08 16:11:01 -07002164 if (freeze || pmd_migration) {
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002165 swp_entry_t swp_entry;
2166 swp_entry = make_migration_entry(page + i, write);
2167 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002168 if (soft_dirty)
2169 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002170 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07002171 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002172 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002173 if (!write)
2174 entry = pte_wrprotect(entry);
2175 if (!young)
2176 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002177 if (soft_dirty)
2178 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002179 }
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002180 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002181 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002182 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002183 atomic_inc(&page[i]._mapcount);
2184 pte_unmap(pte);
2185 }
2186
2187 /*
2188 * Set PG_double_map before dropping compound_mapcount to avoid
2189 * false-negative page_mapped().
2190 */
2191 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2192 for (i = 0; i < HPAGE_PMD_NR; i++)
2193 atomic_inc(&page[i]._mapcount);
2194 }
2195
2196 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2197 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07002198 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002199 if (TestClearPageDoubleMap(page)) {
2200 /* No need in mapcount reference anymore */
2201 for (i = 0; i < HPAGE_PMD_NR; i++)
2202 atomic_dec(&page[i]._mapcount);
2203 }
2204 }
2205
2206 smp_wmb(); /* make pte visible before pmd */
2207 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002208
2209 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002210 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002211 page_remove_rmap(page + i, false);
2212 put_page(page + i);
2213 }
2214 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002215}
2216
2217void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002218 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002219{
2220 spinlock_t *ptl;
2221 struct mm_struct *mm = vma->vm_mm;
2222 unsigned long haddr = address & HPAGE_PMD_MASK;
2223
2224 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2225 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002226
2227 /*
2228 * If caller asks to setup a migration entries, we need a page to check
2229 * pmd against. Otherwise we can end up replacing wrong page.
2230 */
2231 VM_BUG_ON(freeze && !page);
2232 if (page && page != pmd_page(*pmd))
2233 goto out;
2234
Dan Williams5c7fb562016-01-15 16:56:52 -08002235 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002236 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002237 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002238 clear_page_mlock(page);
Zi Yan84c3fc42017-09-08 16:11:01 -07002239 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002240 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002241 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002242out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002243 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002244 /*
2245 * No need to double call mmu_notifier->invalidate_range() callback.
2246 * They are 3 cases to consider inside __split_huge_pmd_locked():
2247 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2248 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2249 * fault will trigger a flush_notify before pointing to a new page
2250 * (it is fine if the secondary mmu keeps pointing to the old zero
2251 * page in the meantime)
2252 * 3) Split a huge pmd into pte pointing to the same page. No need
2253 * to invalidate secondary tlb entry they are all still valid.
2254 * any further changes to individual pte will notify. So no need
2255 * to call mmu_notifier->invalidate_range()
2256 */
2257 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2258 HPAGE_PMD_SIZE);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002259}
2260
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002261void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2262 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002263{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002264 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002265 p4d_t *p4d;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002266 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002267 pmd_t *pmd;
2268
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002269 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002270 if (!pgd_present(*pgd))
2271 return;
2272
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002273 p4d = p4d_offset(pgd, address);
2274 if (!p4d_present(*p4d))
2275 return;
2276
2277 pud = pud_offset(p4d, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002278 if (!pud_present(*pud))
2279 return;
2280
2281 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002282
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002283 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002284}
2285
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002286void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002287 unsigned long start,
2288 unsigned long end,
2289 long adjust_next)
2290{
2291 /*
2292 * If the new start address isn't hpage aligned and it could
2293 * previously contain an hugepage: check if we need to split
2294 * an huge pmd.
2295 */
2296 if (start & ~HPAGE_PMD_MASK &&
2297 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2298 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002299 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002300
2301 /*
2302 * If the new end address isn't hpage aligned and it could
2303 * previously contain an hugepage: check if we need to split
2304 * an huge pmd.
2305 */
2306 if (end & ~HPAGE_PMD_MASK &&
2307 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2308 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002309 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002310
2311 /*
2312 * If we're also updating the vma->vm_next->vm_start, if the new
2313 * vm_next->vm_start isn't page aligned and it could previously
2314 * contain an hugepage: check if we need to split an huge pmd.
2315 */
2316 if (adjust_next > 0) {
2317 struct vm_area_struct *next = vma->vm_next;
2318 unsigned long nstart = next->vm_start;
2319 nstart += adjust_next << PAGE_SHIFT;
2320 if (nstart & ~HPAGE_PMD_MASK &&
2321 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2322 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002323 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002324 }
2325}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002326
Hugh Dickins69697e62018-11-30 14:10:13 -08002327static void unmap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002328{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002329 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002330 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
Minchan Kim666e5a42017-05-03 14:54:20 -07002331 bool unmap_success;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002332
2333 VM_BUG_ON_PAGE(!PageHead(page), page);
2334
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002335 if (PageAnon(page))
Naoya Horiguchib5ff8162017-09-08 16:10:49 -07002336 ttu_flags |= TTU_SPLIT_FREEZE;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002337
Minchan Kim666e5a42017-05-03 14:54:20 -07002338 unmap_success = try_to_unmap(page, ttu_flags);
2339 VM_BUG_ON_PAGE(!unmap_success, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002340}
2341
Hugh Dickins69697e62018-11-30 14:10:13 -08002342static void remap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002343{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002344 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002345 if (PageTransHuge(page)) {
2346 remove_migration_ptes(page, page, true);
2347 } else {
2348 for (i = 0; i < HPAGE_PMD_NR; i++)
2349 remove_migration_ptes(page + i, page + i, true);
2350 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002351}
2352
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002353static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002354 struct lruvec *lruvec, struct list_head *list)
2355{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002356 struct page *page_tail = head + tail;
2357
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002358 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002359
2360 /*
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002361 * Clone page flags before unfreezing refcount.
2362 *
2363 * After successful get_page_unless_zero() might follow flags change,
2364 * for exmaple lock_page() which set PG_waiters.
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002365 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002366 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2367 page_tail->flags |= (head->flags &
2368 ((1L << PG_referenced) |
2369 (1L << PG_swapbacked) |
Huang Ying38d8b4e2017-07-06 15:37:18 -07002370 (1L << PG_swapcache) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002371 (1L << PG_mlocked) |
2372 (1L << PG_uptodate) |
2373 (1L << PG_active) |
2374 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002375 (1L << PG_unevictable) |
2376 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002377
Hugh Dickins7e186562018-11-30 14:10:16 -08002378 /* ->mapping in first tail page is compound_mapcount */
2379 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2380 page_tail);
2381 page_tail->mapping = head->mapping;
2382 page_tail->index = head->index + tail;
2383
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002384 /* Page flags must be visible before we make the page non-compound. */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002385 smp_wmb();
2386
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002387 /*
2388 * Clear PageTail before unfreezing page refcount.
2389 *
2390 * After successful get_page_unless_zero() might follow put_page()
2391 * which needs correct compound_head().
2392 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002393 clear_compound_head(page_tail);
2394
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002395 /* Finally unfreeze refcount. Additional reference from page cache. */
2396 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2397 PageSwapCache(head)));
2398
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002399 if (page_is_young(head))
2400 set_page_young(page_tail);
2401 if (page_is_idle(head))
2402 set_page_idle(page_tail);
2403
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002404 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
Michal Hocko94723aa2018-04-10 16:30:07 -07002405
2406 /*
2407 * always add to the tail because some iterators expect new
2408 * pages to show after the currently processed elements - e.g.
2409 * migrate_pages
2410 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002411 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002412}
2413
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002414static void __split_huge_page(struct page *page, struct list_head *list,
Hugh Dickinsd31ff472018-11-30 14:10:21 -08002415 pgoff_t end, unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002416{
2417 struct page *head = compound_head(page);
2418 struct zone *zone = page_zone(head);
2419 struct lruvec *lruvec;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002420 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002421
Mel Gorman599d0c92016-07-28 15:45:31 -07002422 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002423
2424 /* complete memcg works before add pages to LRU */
2425 mem_cgroup_split_huge_fixup(head);
2426
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002427 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002428 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002429 /* Some pages can be beyond i_size: drop them from page cache */
2430 if (head[i].index >= end) {
Hugh Dickins2d077d42018-06-01 16:50:45 -07002431 ClearPageDirty(head + i);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002432 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002433 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2434 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002435 put_page(head + i);
2436 }
2437 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002438
2439 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002440 /* See comment in __split_huge_page_tail() */
2441 if (PageAnon(head)) {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002442 /* Additional pin to radix tree of swap cache */
2443 if (PageSwapCache(head))
2444 page_ref_add(head, 2);
2445 else
2446 page_ref_inc(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002447 } else {
2448 /* Additional pin to radix tree */
2449 page_ref_add(head, 2);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002450 xa_unlock(&head->mapping->i_pages);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002451 }
2452
Mel Gormana52633d2016-07-28 15:45:28 -07002453 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002454
Hugh Dickins69697e62018-11-30 14:10:13 -08002455 remap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002456
2457 for (i = 0; i < HPAGE_PMD_NR; i++) {
2458 struct page *subpage = head + i;
2459 if (subpage == page)
2460 continue;
2461 unlock_page(subpage);
2462
2463 /*
2464 * Subpages may be freed if there wasn't any mapping
2465 * like if add_to_swap() is running on a lru page that
2466 * had its mapping zapped. And freeing these pages
2467 * requires taking the lru_lock so we do the put_page
2468 * of the tail pages after the split is complete.
2469 */
2470 put_page(subpage);
2471 }
2472}
2473
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002474int total_mapcount(struct page *page)
2475{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002476 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002477
2478 VM_BUG_ON_PAGE(PageTail(page), page);
2479
2480 if (likely(!PageCompound(page)))
2481 return atomic_read(&page->_mapcount) + 1;
2482
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002483 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002484 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002485 return compound;
2486 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002487 for (i = 0; i < HPAGE_PMD_NR; i++)
2488 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002489 /* File pages has compound_mapcount included in _mapcount */
2490 if (!PageAnon(page))
2491 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002492 if (PageDoubleMap(page))
2493 ret -= HPAGE_PMD_NR;
2494 return ret;
2495}
2496
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002497/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002498 * This calculates accurately how many mappings a transparent hugepage
2499 * has (unlike page_mapcount() which isn't fully accurate). This full
2500 * accuracy is primarily needed to know if copy-on-write faults can
2501 * reuse the page and change the mapping to read-write instead of
2502 * copying them. At the same time this returns the total_mapcount too.
2503 *
2504 * The function returns the highest mapcount any one of the subpages
2505 * has. If the return value is one, even if different processes are
2506 * mapping different subpages of the transparent hugepage, they can
2507 * all reuse it, because each process is reusing a different subpage.
2508 *
2509 * The total_mapcount is instead counting all virtual mappings of the
2510 * subpages. If the total_mapcount is equal to "one", it tells the
2511 * caller all mappings belong to the same "mm" and in turn the
2512 * anon_vma of the transparent hugepage can become the vma->anon_vma
2513 * local one as no other process may be mapping any of the subpages.
2514 *
2515 * It would be more accurate to replace page_mapcount() with
2516 * page_trans_huge_mapcount(), however we only use
2517 * page_trans_huge_mapcount() in the copy-on-write faults where we
2518 * need full accuracy to avoid breaking page pinning, because
2519 * page_trans_huge_mapcount() is slower than page_mapcount().
2520 */
2521int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2522{
2523 int i, ret, _total_mapcount, mapcount;
2524
2525 /* hugetlbfs shouldn't call it */
2526 VM_BUG_ON_PAGE(PageHuge(page), page);
2527
2528 if (likely(!PageTransCompound(page))) {
2529 mapcount = atomic_read(&page->_mapcount) + 1;
2530 if (total_mapcount)
2531 *total_mapcount = mapcount;
2532 return mapcount;
2533 }
2534
2535 page = compound_head(page);
2536
2537 _total_mapcount = ret = 0;
2538 for (i = 0; i < HPAGE_PMD_NR; i++) {
2539 mapcount = atomic_read(&page[i]._mapcount) + 1;
2540 ret = max(ret, mapcount);
2541 _total_mapcount += mapcount;
2542 }
2543 if (PageDoubleMap(page)) {
2544 ret -= 1;
2545 _total_mapcount -= HPAGE_PMD_NR;
2546 }
2547 mapcount = compound_mapcount(page);
2548 ret += mapcount;
2549 _total_mapcount += mapcount;
2550 if (total_mapcount)
2551 *total_mapcount = _total_mapcount;
2552 return ret;
2553}
2554
Huang Yingb8f593c2017-07-06 15:37:28 -07002555/* Racy check whether the huge page can be split */
2556bool can_split_huge_page(struct page *page, int *pextra_pins)
2557{
2558 int extra_pins;
2559
2560 /* Additional pins from radix tree */
2561 if (PageAnon(page))
2562 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2563 else
2564 extra_pins = HPAGE_PMD_NR;
2565 if (pextra_pins)
2566 *pextra_pins = extra_pins;
2567 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2568}
2569
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002570/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002571 * This function splits huge page into normal pages. @page can point to any
2572 * subpage of huge page to split. Split doesn't change the position of @page.
2573 *
2574 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2575 * The huge page must be locked.
2576 *
2577 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2578 *
2579 * Both head page and tail pages will inherit mapping, flags, and so on from
2580 * the hugepage.
2581 *
2582 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2583 * they are not mapped.
2584 *
2585 * Returns 0 if the hugepage is split successfully.
2586 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2587 * us.
2588 */
2589int split_huge_page_to_list(struct page *page, struct list_head *list)
2590{
2591 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002592 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002593 struct anon_vma *anon_vma = NULL;
2594 struct address_space *mapping = NULL;
2595 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002596 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002597 unsigned long flags;
Hugh Dickinsd31ff472018-11-30 14:10:21 -08002598 pgoff_t end;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002599
2600 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002601 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002602 VM_BUG_ON_PAGE(!PageCompound(page), page);
2603
Huang Ying59807682017-09-06 16:22:34 -07002604 if (PageWriteback(page))
2605 return -EBUSY;
2606
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002607 if (PageAnon(head)) {
2608 /*
2609 * The caller does not necessarily hold an mmap_sem that would
2610 * prevent the anon_vma disappearing so we first we take a
2611 * reference to it and then lock the anon_vma for write. This
2612 * is similar to page_lock_anon_vma_read except the write lock
2613 * is taken to serialise against parallel split or collapse
2614 * operations.
2615 */
2616 anon_vma = page_get_anon_vma(head);
2617 if (!anon_vma) {
2618 ret = -EBUSY;
2619 goto out;
2620 }
Hugh Dickinsd31ff472018-11-30 14:10:21 -08002621 end = -1;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002622 mapping = NULL;
2623 anon_vma_lock_write(anon_vma);
2624 } else {
2625 mapping = head->mapping;
2626
2627 /* Truncated ? */
2628 if (!mapping) {
2629 ret = -EBUSY;
2630 goto out;
2631 }
2632
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002633 anon_vma = NULL;
2634 i_mmap_lock_read(mapping);
Hugh Dickinsd31ff472018-11-30 14:10:21 -08002635
2636 /*
2637 *__split_huge_page() may need to trim off pages beyond EOF:
2638 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2639 * which cannot be nested inside the page tree lock. So note
2640 * end now: i_size itself may be changed at any moment, but
2641 * head page lock is good enough to serialize the trimming.
2642 */
2643 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002644 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002645
2646 /*
Hugh Dickins69697e62018-11-30 14:10:13 -08002647 * Racy check if we can split the page, before unmap_page() will
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002648 * 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);
Hugh Dickins69697e62018-11-30 14:10:13 -08002656 unmap_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
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002669 xa_lock(&mapping->i_pages);
2670 pslot = radix_tree_lookup_slot(&mapping->i_pages,
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002671 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,
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002677 &mapping->i_pages.xa_lock) != head)
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002678 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);
Hugh Dickinsd31ff472018-11-30 14:10:21 -08002693 __split_huge_page(page, list, end, 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)
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002711 xa_unlock(&mapping->i_pages);
Mel Gormana52633d2016-07-28 15:45:28 -07002712 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Hugh Dickins69697e62018-11-30 14:10:13 -08002713 remap_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);
Kirill A. Shutemovfa41b902018-03-22 16:17:31 -07002794 if (!trylock_page(page))
2795 goto next;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002796 /* split_huge_page() removes page from list on success */
2797 if (!split_huge_page(page))
2798 split++;
2799 unlock_page(page);
Kirill A. Shutemovfa41b902018-03-22 16:17:31 -07002800next:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002801 put_page(page);
2802 }
2803
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002804 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2805 list_splice_tail(&list, &pgdata->split_queue);
2806 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002807
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002808 /*
2809 * Stop shrinker if we didn't split any page, but the queue is empty.
2810 * This can happen if pages were freed under us.
2811 */
2812 if (!split && list_empty(&pgdata->split_queue))
2813 return SHRINK_STOP;
2814 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002815}
2816
2817static struct shrinker deferred_split_shrinker = {
2818 .count_objects = deferred_split_count,
2819 .scan_objects = deferred_split_scan,
2820 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002821 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002822};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002823
2824#ifdef CONFIG_DEBUG_FS
2825static int split_huge_pages_set(void *data, u64 val)
2826{
2827 struct zone *zone;
2828 struct page *page;
2829 unsigned long pfn, max_zone_pfn;
2830 unsigned long total = 0, split = 0;
2831
2832 if (val != 1)
2833 return -EINVAL;
2834
2835 for_each_populated_zone(zone) {
2836 max_zone_pfn = zone_end_pfn(zone);
2837 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2838 if (!pfn_valid(pfn))
2839 continue;
2840
2841 page = pfn_to_page(pfn);
2842 if (!get_page_unless_zero(page))
2843 continue;
2844
2845 if (zone != page_zone(page))
2846 goto next;
2847
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002848 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002849 goto next;
2850
2851 total++;
2852 lock_page(page);
2853 if (!split_huge_page(page))
2854 split++;
2855 unlock_page(page);
2856next:
2857 put_page(page);
2858 }
2859 }
2860
Yang Shi145bdaa2016-05-05 16:22:00 -07002861 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002862
2863 return 0;
2864}
2865DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2866 "%llu\n");
2867
2868static int __init split_huge_pages_debugfs(void)
2869{
2870 void *ret;
2871
Yang Shi145bdaa2016-05-05 16:22:00 -07002872 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002873 &split_huge_pages_fops);
2874 if (!ret)
2875 pr_warn("Failed to create split_huge_pages in debugfs");
2876 return 0;
2877}
2878late_initcall(split_huge_pages_debugfs);
2879#endif
Zi Yan616b8372017-09-08 16:10:57 -07002880
2881#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2882void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2883 struct page *page)
2884{
2885 struct vm_area_struct *vma = pvmw->vma;
2886 struct mm_struct *mm = vma->vm_mm;
2887 unsigned long address = pvmw->address;
2888 pmd_t pmdval;
2889 swp_entry_t entry;
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002890 pmd_t pmdswp;
Zi Yan616b8372017-09-08 16:10:57 -07002891
2892 if (!(pvmw->pmd && !pvmw->pte))
2893 return;
2894
Zi Yan616b8372017-09-08 16:10:57 -07002895 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2896 pmdval = *pvmw->pmd;
2897 pmdp_invalidate(vma, address, pvmw->pmd);
2898 if (pmd_dirty(pmdval))
2899 set_page_dirty(page);
2900 entry = make_migration_entry(page, pmd_write(pmdval));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002901 pmdswp = swp_entry_to_pmd(entry);
2902 if (pmd_soft_dirty(pmdval))
2903 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2904 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
Zi Yan616b8372017-09-08 16:10:57 -07002905 page_remove_rmap(page, true);
2906 put_page(page);
Zi Yan616b8372017-09-08 16:10:57 -07002907}
2908
2909void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2910{
2911 struct vm_area_struct *vma = pvmw->vma;
2912 struct mm_struct *mm = vma->vm_mm;
2913 unsigned long address = pvmw->address;
2914 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2915 pmd_t pmde;
2916 swp_entry_t entry;
2917
2918 if (!(pvmw->pmd && !pvmw->pte))
2919 return;
2920
2921 entry = pmd_to_swp_entry(*pvmw->pmd);
2922 get_page(new);
2923 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002924 if (pmd_swp_soft_dirty(*pvmw->pmd))
2925 pmde = pmd_mksoft_dirty(pmde);
Zi Yan616b8372017-09-08 16:10:57 -07002926 if (is_write_migration_entry(entry))
Linus Torvaldsf55e1012017-11-29 09:01:01 -08002927 pmde = maybe_pmd_mkwrite(pmde, vma);
Zi Yan616b8372017-09-08 16:10:57 -07002928
2929 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
Naoya Horiguchie71769a2018-04-20 14:55:45 -07002930 if (PageAnon(new))
2931 page_add_anon_rmap(new, vma, mmun_start, true);
2932 else
2933 page_add_file_rmap(new, true);
Zi Yan616b8372017-09-08 16:10:57 -07002934 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
Kirill A. Shutemove125fe42018-10-05 15:51:41 -07002935 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
Zi Yan616b8372017-09-08 16:10:57 -07002936 mlock_vma_page(new);
2937 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2938}
2939#endif