blob: dc12249af6dfda6a0cc2aeb828af839cd80ebad2 [file] [log] [blame]
Thomas Gleixner20c8ccb2019-06-04 10:11:32 +02001// SPDX-License-Identifier: GPL-2.0-only
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002/*
3 * Copyright (C) 2009 Red Hat, Inc.
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08004 */
5
Andrew Mortonae3a8c12014-06-04 16:06:58 -07006#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08008#include <linux/mm.h>
9#include <linux/sched.h>
Ingo Molnarf7ccbae2017-02-08 18:51:30 +010010#include <linux/sched/coredump.h>
Ingo Molnar6a3827d2017-02-08 18:51:31 +010011#include <linux/sched/numa_balancing.h>
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080012#include <linux/highmem.h>
13#include <linux/hugetlb.h>
14#include <linux/mmu_notifier.h>
15#include <linux/rmap.h>
16#include <linux/swap.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080017#include <linux/shrinker.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080018#include <linux/mm_inline.h>
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -080019#include <linux/swapops.h>
Matthew Wilcox4897c762015-09-08 14:58:45 -070020#include <linux/dax.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080021#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080022#include <linux/freezer.h>
Dan Williamsf25748e32016-01-15 16:56:43 -080023#include <linux/pfn_t.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080024#include <linux/mman.h>
Dan Williams3565fce2016-01-15 16:56:55 -080025#include <linux/memremap.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020026#include <linux/pagemap.h>
Kirill A. Shutemov49071d42016-01-15 16:54:40 -080027#include <linux/debugfs.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000028#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080029#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070030#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070031#include <linux/page_idle.h>
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -070032#include <linux/shmem_fs.h>
Michal Hocko6b31d592017-08-18 15:16:15 -070033#include <linux/oom.h>
Anshuman Khandual98fa15f2019-03-05 15:42:58 -080034#include <linux/numa.h>
Vlastimil Babkaf7da6772019-08-24 17:54:59 -070035#include <linux/page_owner.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
Michal Hocko7635d9c2018-12-28 00:38:21 -080065bool transparent_hugepage_enabled(struct vm_area_struct *vma)
66{
Yang Shic0630662019-07-18 15:57:27 -070067 /* The addr is used to check if the vma size fits */
68 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
69
70 if (!transhuge_vma_suitable(vma, addr))
71 return false;
Michal Hocko7635d9c2018-12-28 00:38:21 -080072 if (vma_is_anonymous(vma))
73 return __transparent_hugepage_enabled(vma);
Yang Shic0630662019-07-18 15:57:27 -070074 if (vma_is_shmem(vma))
75 return shmem_huge_enabled(vma);
Michal Hocko7635d9c2018-12-28 00:38:21 -080076
77 return false;
78}
79
Aaron Lu6fcb52a2016-10-07 17:00:08 -070080static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080081{
82 struct page *zero_page;
83retry:
84 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070085 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080086
87 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
88 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080089 if (!zero_page) {
90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070091 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080092 }
93 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080094 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070095 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080096 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070097 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080098 goto retry;
99 }
100
101 /* We take additional reference here. It will be put back by shrinker */
102 atomic_set(&huge_zero_refcount, 2);
103 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -0700104 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800105}
106
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700107static void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800108{
109 /*
110 * Counter should never go to zero here. Only shrinker can put
111 * last reference.
112 */
113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
114}
115
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700116struct page *mm_get_huge_zero_page(struct mm_struct *mm)
117{
118 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 return READ_ONCE(huge_zero_page);
120
121 if (!get_huge_zero_page())
122 return NULL;
123
124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125 put_huge_zero_page();
126
127 return READ_ONCE(huge_zero_page);
128}
129
130void mm_put_huge_zero_page(struct mm_struct *mm)
131{
132 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133 put_huge_zero_page();
134}
135
Glauber Costa48896462013-08-28 10:18:15 +1000136static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800138{
Glauber Costa48896462013-08-28 10:18:15 +1000139 /* we can free zero page only if last reference remains */
140 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
141}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800142
Glauber Costa48896462013-08-28 10:18:15 +1000143static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144 struct shrink_control *sc)
145{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800146 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700147 struct page *zero_page = xchg(&huge_zero_page, NULL);
148 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700149 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000150 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800151 }
152
153 return 0;
154}
155
156static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000157 .count_objects = shrink_huge_zero_page_count,
158 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800159 .seeks = DEFAULT_SEEKS,
160};
161
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800162#ifdef CONFIG_SYSFS
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800163static ssize_t enabled_show(struct kobject *kobj,
164 struct kobj_attribute *attr, char *buf)
165{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
167 return sprintf(buf, "[always] madvise never\n");
168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
169 return sprintf(buf, "always [madvise] never\n");
170 else
171 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800172}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700173
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800174static ssize_t enabled_store(struct kobject *kobj,
175 struct kobj_attribute *attr,
176 const char *buf, size_t count)
177{
David Rientjes21440d72017-02-22 15:45:49 -0800178 ssize_t ret = count;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800179
David Rientjesf42f2552020-01-30 22:14:48 -0800180 if (sysfs_streq(buf, "always")) {
David Rientjes21440d72017-02-22 15:45:49 -0800181 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
182 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800183 } else if (sysfs_streq(buf, "madvise")) {
David Rientjes21440d72017-02-22 15:45:49 -0800184 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
185 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800186 } else if (sysfs_streq(buf, "never")) {
David Rientjes21440d72017-02-22 15:45:49 -0800187 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
188 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
189 } else
190 ret = -EINVAL;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800191
192 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700193 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800194 if (err)
195 ret = err;
196 }
Andrea Arcangeliba761492011-01-13 15:46:58 -0800197 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800198}
199static struct kobj_attribute enabled_attr =
200 __ATTR(enabled, 0644, enabled_show, enabled_store);
201
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700202ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800203 struct kobj_attribute *attr, char *buf,
204 enum transparent_hugepage_flag flag)
205{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700206 return sprintf(buf, "%d\n",
207 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800208}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700209
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700210ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800211 struct kobj_attribute *attr,
212 const char *buf, size_t count,
213 enum transparent_hugepage_flag flag)
214{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700215 unsigned long value;
216 int ret;
217
218 ret = kstrtoul(buf, 10, &value);
219 if (ret < 0)
220 return ret;
221 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800222 return -EINVAL;
223
Ben Hutchingse27e6152011-04-14 15:22:21 -0700224 if (value)
225 set_bit(flag, &transparent_hugepage_flags);
226 else
227 clear_bit(flag, &transparent_hugepage_flags);
228
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800229 return count;
230}
231
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800232static ssize_t defrag_show(struct kobject *kobj,
233 struct kobj_attribute *attr, char *buf)
234{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800236 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
Mel Gorman444eb2a42016-03-17 14:19:23 -0700237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
David Rientjes21440d72017-02-22 15:45:49 -0800238 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
240 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
241 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
242 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
243 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800244}
David Rientjes21440d72017-02-22 15:45:49 -0800245
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800246static ssize_t defrag_store(struct kobject *kobj,
247 struct kobj_attribute *attr,
248 const char *buf, size_t count)
249{
David Rientjesf42f2552020-01-30 22:14:48 -0800250 if (sysfs_streq(buf, "always")) {
David Rientjes21440d72017-02-22 15:45:49 -0800251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800255 } else if (sysfs_streq(buf, "defer+madvise")) {
David Rientjes21440d72017-02-22 15:45:49 -0800256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
259 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800260 } else if (sysfs_streq(buf, "defer")) {
David Rientjes4fad7fb2017-04-07 16:04:54 -0700261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800265 } else if (sysfs_streq(buf, "madvise")) {
David Rientjes21440d72017-02-22 15:45:49 -0800266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
David Rientjesf42f2552020-01-30 22:14:48 -0800270 } else if (sysfs_streq(buf, "never")) {
David Rientjes21440d72017-02-22 15:45:49 -0800271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
272 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
275 } else
276 return -EINVAL;
277
278 return count;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800279}
280static struct kobj_attribute defrag_attr =
281 __ATTR(defrag, 0644, defrag_show, defrag_store);
282
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800283static ssize_t use_zero_page_show(struct kobject *kobj,
284 struct kobj_attribute *attr, char *buf)
285{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700286 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
288}
289static ssize_t use_zero_page_store(struct kobject *kobj,
290 struct kobj_attribute *attr, const char *buf, size_t count)
291{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700292 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800293 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
294}
295static struct kobj_attribute use_zero_page_attr =
296 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Hugh Dickins49920d22016-12-12 16:44:50 -0800297
298static ssize_t hpage_pmd_size_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300{
301 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
302}
303static struct kobj_attribute hpage_pmd_size_attr =
304 __ATTR_RO(hpage_pmd_size);
305
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800306#ifdef CONFIG_DEBUG_VM
307static ssize_t debug_cow_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf)
309{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700310 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312}
313static ssize_t debug_cow_store(struct kobject *kobj,
314 struct kobj_attribute *attr,
315 const char *buf, size_t count)
316{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700317 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800318 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
319}
320static struct kobj_attribute debug_cow_attr =
321 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
322#endif /* CONFIG_DEBUG_VM */
323
324static struct attribute *hugepage_attr[] = {
325 &enabled_attr.attr,
326 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800327 &use_zero_page_attr.attr,
Hugh Dickins49920d22016-12-12 16:44:50 -0800328 &hpage_pmd_size_attr.attr,
Matthew Wilcox (Oracle)396bcc52020-04-06 20:04:35 -0700329#ifdef CONFIG_SHMEM
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700330 &shmem_enabled_attr.attr,
331#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800332#ifdef CONFIG_DEBUG_VM
333 &debug_cow_attr.attr,
334#endif
335 NULL,
336};
337
Arvind Yadav8aa95a22017-09-06 16:22:03 -0700338static const struct attribute_group hugepage_attr_group = {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800339 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800340};
341
Shaohua Li569e5592012-01-12 17:19:11 -0800342static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
343{
344 int err;
345
346 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
347 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700348 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800349 return -ENOMEM;
350 }
351
352 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
353 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700354 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800355 goto delete_obj;
356 }
357
358 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
359 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700360 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800361 goto remove_hp_group;
362 }
363
364 return 0;
365
366remove_hp_group:
367 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
368delete_obj:
369 kobject_put(*hugepage_kobj);
370 return err;
371}
372
373static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
374{
375 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
376 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
377 kobject_put(hugepage_kobj);
378}
379#else
380static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
381{
382 return 0;
383}
384
385static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
386{
387}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800388#endif /* CONFIG_SYSFS */
389
390static int __init hugepage_init(void)
391{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800392 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800393 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800394
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800395 if (!has_transparent_hugepage()) {
396 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800397 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800398 }
399
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530400 /*
401 * hugepages can't be allocated by the buddy allocator
402 */
403 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
404 /*
405 * we use page->mapping and page->index in second tail page
406 * as list_head: assuming THP order >= 2
407 */
408 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
409
Shaohua Li569e5592012-01-12 17:19:11 -0800410 err = hugepage_init_sysfs(&hugepage_kobj);
411 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700412 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800413
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700414 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800415 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700416 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800417
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700418 err = register_shrinker(&huge_zero_page_shrinker);
419 if (err)
420 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800421 err = register_shrinker(&deferred_split_shrinker);
422 if (err)
423 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800424
Rik van Riel97562cd2011-01-13 15:47:12 -0800425 /*
426 * By default disable transparent hugepages on smaller systems,
427 * where the extra memory used could hurt more than TLB overhead
428 * is likely to save. The admin can still enable it through /sys.
429 */
Arun KSca79b0c2018-12-28 00:34:29 -0800430 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800431 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2932015-04-15 16:14:56 -0700432 return 0;
433 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800434
Kirill A. Shutemov79553da2932015-04-15 16:14:56 -0700435 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700436 if (err)
437 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800438
Shaohua Li569e5592012-01-12 17:19:11 -0800439 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700440err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800441 unregister_shrinker(&deferred_split_shrinker);
442err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700443 unregister_shrinker(&huge_zero_page_shrinker);
444err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700445 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700446err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800447 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700448err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800449 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800450}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800451subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800452
453static int __init setup_transparent_hugepage(char *str)
454{
455 int ret = 0;
456 if (!str)
457 goto out;
458 if (!strcmp(str, "always")) {
459 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
460 &transparent_hugepage_flags);
461 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
462 &transparent_hugepage_flags);
463 ret = 1;
464 } else if (!strcmp(str, "madvise")) {
465 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
466 &transparent_hugepage_flags);
467 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
468 &transparent_hugepage_flags);
469 ret = 1;
470 } else if (!strcmp(str, "never")) {
471 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
472 &transparent_hugepage_flags);
473 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
474 &transparent_hugepage_flags);
475 ret = 1;
476 }
477out:
478 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700479 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800480 return ret;
481}
482__setup("transparent_hugepage=", setup_transparent_hugepage);
483
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800484pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800485{
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800486 if (likely(vma->vm_flags & VM_WRITE))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800487 pmd = pmd_mkwrite(pmd);
488 return pmd;
489}
490
Yang Shi87eaceb2019-09-23 15:38:15 -0700491#ifdef CONFIG_MEMCG
492static inline struct deferred_split *get_deferred_split_queue(struct page *page)
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800493{
Yang Shi87eaceb2019-09-23 15:38:15 -0700494 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
495 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
496
497 if (memcg)
498 return &memcg->deferred_split_queue;
499 else
500 return &pgdat->deferred_split_queue;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800501}
Yang Shi87eaceb2019-09-23 15:38:15 -0700502#else
503static inline struct deferred_split *get_deferred_split_queue(struct page *page)
504{
505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506
507 return &pgdat->deferred_split_queue;
508}
509#endif
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800510
511void prep_transhuge_page(struct page *page)
512{
513 /*
514 * we use page->mapping and page->indexlru in second tail page
515 * as list_head: assuming THP order >= 2
516 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800517
518 INIT_LIST_HEAD(page_deferred_list(page));
519 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
520}
521
Sean Christopherson005ba372020-01-08 12:24:36 -0800522bool is_transparent_hugepage(struct page *page)
523{
524 if (!PageCompound(page))
525 return 0;
526
527 page = compound_head(page);
528 return is_huge_zero_page(page) ||
529 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
530}
531EXPORT_SYMBOL_GPL(is_transparent_hugepage);
532
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800533static unsigned long __thp_get_unmapped_area(struct file *filp,
534 unsigned long addr, unsigned long len,
Toshi Kani74d2fad2016-10-07 16:59:56 -0700535 loff_t off, unsigned long flags, unsigned long size)
536{
Toshi Kani74d2fad2016-10-07 16:59:56 -0700537 loff_t off_end = off + len;
538 loff_t off_align = round_up(off, size);
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800539 unsigned long len_pad, ret;
Toshi Kani74d2fad2016-10-07 16:59:56 -0700540
541 if (off_end <= off_align || (off_end - off_align) < size)
542 return 0;
543
544 len_pad = len + size;
545 if (len_pad < len || (off + len_pad) < off)
546 return 0;
547
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800548 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
Toshi Kani74d2fad2016-10-07 16:59:56 -0700549 off >> PAGE_SHIFT, flags);
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800550
551 /*
552 * The failure might be due to length padding. The caller will retry
553 * without the padding.
554 */
555 if (IS_ERR_VALUE(ret))
Toshi Kani74d2fad2016-10-07 16:59:56 -0700556 return 0;
557
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800558 /*
559 * Do not try to align to THP boundary if allocation at the address
560 * hint succeeds.
561 */
562 if (ret == addr)
563 return addr;
564
565 ret += (off - ret) & (size - 1);
566 return ret;
Toshi Kani74d2fad2016-10-07 16:59:56 -0700567}
568
569unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
570 unsigned long len, unsigned long pgoff, unsigned long flags)
571{
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800572 unsigned long ret;
Toshi Kani74d2fad2016-10-07 16:59:56 -0700573 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
574
Toshi Kani74d2fad2016-10-07 16:59:56 -0700575 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
576 goto out;
577
Kirill A. Shutemov97d3d0f2020-01-13 16:29:10 -0800578 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
579 if (ret)
580 return ret;
581out:
Toshi Kani74d2fad2016-10-07 16:59:56 -0700582 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
583}
584EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
585
Souptick Joarder2b740302018-08-23 17:01:36 -0700586static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
587 struct page *page, gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800588{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800589 struct vm_area_struct *vma = vmf->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700590 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800591 pgtable_t pgtable;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800592 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Souptick Joarder2b740302018-08-23 17:01:36 -0700593 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800594
Sasha Levin309381fea2014-01-23 15:52:54 -0800595 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700596
Tejun Heo2cf85582018-07-03 11:14:56 -0400597 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700598 put_page(page);
599 count_vm_event(THP_FAULT_FALLBACK);
David Rientjes85b9f462020-04-06 20:04:28 -0700600 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700601 return VM_FAULT_FALLBACK;
602 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800603
Joel Fernandes (Google)4cf58922019-01-03 15:28:34 -0800604 pgtable = pte_alloc_one(vma->vm_mm);
Johannes Weiner00501b52014-08-08 14:19:20 -0700605 if (unlikely(!pgtable)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700606 ret = VM_FAULT_OOM;
607 goto release;
Johannes Weiner00501b52014-08-08 14:19:20 -0700608 }
609
Huang Yingc79b57e2017-09-06 16:25:04 -0700610 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700611 /*
612 * The memory barrier inside __SetPageUptodate makes sure that
613 * clear_huge_page writes become visible before the set_pmd_at()
614 * write.
615 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800616 __SetPageUptodate(page);
617
Jan Kara82b0f8c2016-12-14 15:06:58 -0800618 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
619 if (unlikely(!pmd_none(*vmf->pmd))) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700620 goto unlock_release;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800621 } else {
622 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700623
Michal Hocko6b31d592017-08-18 15:16:15 -0700624 ret = check_stable_address_space(vma->vm_mm);
625 if (ret)
626 goto unlock_release;
627
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700628 /* Deliver the page fault to userland */
629 if (userfaultfd_missing(vma)) {
Souptick Joarder2b740302018-08-23 17:01:36 -0700630 vm_fault_t ret2;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700631
Jan Kara82b0f8c2016-12-14 15:06:58 -0800632 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800633 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700634 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700635 pte_free(vma->vm_mm, pgtable);
Souptick Joarder2b740302018-08-23 17:01:36 -0700636 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
637 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
638 return ret2;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700639 }
640
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700641 entry = mk_huge_pmd(page, vma->vm_page_prot);
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800642 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800643 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800644 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700645 lru_cache_add_active_or_unevictable(page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800646 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
647 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700648 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800649 mm_inc_nr_ptes(vma->vm_mm);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800650 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700651 count_vm_event(THP_FAULT_ALLOC);
Chris Down1ff9e6e2019-03-05 15:48:09 -0800652 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800653 }
654
David Rientjesaa2e8782012-05-29 15:06:17 -0700655 return 0;
Michal Hocko6b31d592017-08-18 15:16:15 -0700656unlock_release:
657 spin_unlock(vmf->ptl);
658release:
659 if (pgtable)
660 pte_free(vma->vm_mm, pgtable);
661 mem_cgroup_cancel_charge(page, memcg, true);
662 put_page(page);
663 return ret;
664
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800665}
666
Mel Gorman444eb2a42016-03-17 14:19:23 -0700667/*
David Rientjes21440d72017-02-22 15:45:49 -0800668 * always: directly stall for all thp allocations
669 * defer: wake kswapd and fail if not immediately available
670 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
671 * fail if not immediately available
672 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
673 * available
674 * never: never stall for any thp allocation
Mel Gorman444eb2a42016-03-17 14:19:23 -0700675 */
David Rientjes19deb762019-09-04 12:54:20 -0700676static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800677{
David Rientjes21440d72017-02-22 15:45:49 -0800678 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Michal Hocko89c83fb2018-11-02 15:48:31 -0700679
David Rientjesac79f782019-09-04 12:54:18 -0700680 /* Always do synchronous compaction */
David Rientjes21440d72017-02-22 15:45:49 -0800681 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
Andrea Arcangelia8282602019-08-13 15:37:53 -0700682 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
David Rientjesac79f782019-09-04 12:54:18 -0700683
684 /* Kick kcompactd and fail quickly */
David Rientjes21440d72017-02-22 15:45:49 -0800685 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
David Rientjes19deb762019-09-04 12:54:20 -0700686 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
David Rientjesac79f782019-09-04 12:54:18 -0700687
688 /* Synchronous compaction if madvised, otherwise kick kcompactd */
David Rientjes21440d72017-02-22 15:45:49 -0800689 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
David Rientjes19deb762019-09-04 12:54:20 -0700690 return GFP_TRANSHUGE_LIGHT |
691 (vma_madvised ? __GFP_DIRECT_RECLAIM :
692 __GFP_KSWAPD_RECLAIM);
David Rientjesac79f782019-09-04 12:54:18 -0700693
694 /* Only do synchronous compaction if madvised */
David Rientjes21440d72017-02-22 15:45:49 -0800695 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
David Rientjes19deb762019-09-04 12:54:20 -0700696 return GFP_TRANSHUGE_LIGHT |
697 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
David Rientjesac79f782019-09-04 12:54:18 -0700698
David Rientjes19deb762019-09-04 12:54:20 -0700699 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700700}
701
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800702/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700703static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800704 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700705 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800706{
707 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700708 if (!pmd_none(*pmd))
709 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700710 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800711 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800712 if (pgtable)
713 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800714 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800715 mm_inc_nr_ptes(mm);
Andrew Morton7c414162015-09-08 14:58:43 -0700716 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800717}
718
Souptick Joarder2b740302018-08-23 17:01:36 -0700719vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720{
Jan Kara82b0f8c2016-12-14 15:06:58 -0800721 struct vm_area_struct *vma = vmf->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800722 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800723 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800724 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725
Yang Shi43675e62019-07-18 15:57:24 -0700726 if (!transhuge_vma_suitable(vma, haddr))
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700727 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700728 if (unlikely(anon_vma_prepare(vma)))
729 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700730 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700731 return VM_FAULT_OOM;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800732 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700733 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700734 transparent_hugepage_use_zero_page()) {
735 pgtable_t pgtable;
736 struct page *zero_page;
737 bool set;
Souptick Joarder2b740302018-08-23 17:01:36 -0700738 vm_fault_t ret;
Joel Fernandes (Google)4cf58922019-01-03 15:28:34 -0800739 pgtable = pte_alloc_one(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700740 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800741 return VM_FAULT_OOM;
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700742 zero_page = mm_get_huge_zero_page(vma->vm_mm);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700743 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700744 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700745 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700746 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700747 }
Jan Kara82b0f8c2016-12-14 15:06:58 -0800748 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700749 ret = 0;
750 set = false;
Jan Kara82b0f8c2016-12-14 15:06:58 -0800751 if (pmd_none(*vmf->pmd)) {
Michal Hocko6b31d592017-08-18 15:16:15 -0700752 ret = check_stable_address_space(vma->vm_mm);
753 if (ret) {
754 spin_unlock(vmf->ptl);
755 } else if (userfaultfd_missing(vma)) {
Jan Kara82b0f8c2016-12-14 15:06:58 -0800756 spin_unlock(vmf->ptl);
757 ret = handle_userfault(vmf, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700758 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
759 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700760 set_huge_zero_page(pgtable, vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -0800761 haddr, vmf->pmd, zero_page);
762 spin_unlock(vmf->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700763 set = true;
764 }
765 } else
Jan Kara82b0f8c2016-12-14 15:06:58 -0800766 spin_unlock(vmf->ptl);
Aaron Lu6fcb52a2016-10-07 17:00:08 -0700767 if (!set)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700768 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700769 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800770 }
David Rientjes19deb762019-09-04 12:54:20 -0700771 gfp = alloc_hugepage_direct_gfpmask(vma);
772 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700773 if (unlikely(!page)) {
774 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700775 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700776 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800777 prep_transhuge_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -0800778 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800779}
780
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700781static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700782 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
783 pgtable_t pgtable)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700784{
785 struct mm_struct *mm = vma->vm_mm;
786 pmd_t entry;
787 spinlock_t *ptl;
788
789 ptl = pmd_lock(mm, pmd);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700790 if (!pmd_none(*pmd)) {
791 if (write) {
792 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
793 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
794 goto out_unlock;
795 }
796 entry = pmd_mkyoung(*pmd);
797 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
798 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
799 update_mmu_cache_pmd(vma, addr, pmd);
800 }
801
802 goto out_unlock;
803 }
804
Dan Williamsf25748e32016-01-15 16:56:43 -0800805 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
806 if (pfn_t_devmap(pfn))
807 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800808 if (write) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800809 entry = pmd_mkyoung(pmd_mkdirty(entry));
810 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700811 }
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700812
813 if (pgtable) {
814 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -0800815 mm_inc_nr_ptes(mm);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700816 pgtable = NULL;
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700817 }
818
Ross Zwisler01871e52016-01-15 16:56:02 -0800819 set_pmd_at(mm, addr, pmd, entry);
820 update_mmu_cache_pmd(vma, addr, pmd);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700821
822out_unlock:
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700823 spin_unlock(ptl);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700824 if (pgtable)
825 pte_free(mm, pgtable);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700826}
827
Thomas Hellstrom (VMware)9a9731b2020-03-24 18:48:09 +0100828/**
829 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
830 * @vmf: Structure describing the fault
831 * @pfn: pfn to insert
832 * @pgprot: page protection to use
833 * @write: whether it's a write fault
834 *
835 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
836 * also consult the vmf_insert_mixed_prot() documentation when
837 * @pgprot != @vmf->vma->vm_page_prot.
838 *
839 * Return: vm_fault_t value.
840 */
841vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
842 pgprot_t pgprot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700843{
Dan Williamsfce86ff2019-05-13 17:15:33 -0700844 unsigned long addr = vmf->address & PMD_MASK;
845 struct vm_area_struct *vma = vmf->vma;
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700846 pgtable_t pgtable = NULL;
Dan Williamsfce86ff2019-05-13 17:15:33 -0700847
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700848 /*
849 * If we had pmd_special, we could avoid all these restrictions,
850 * but we need to be consistent with PTEs and architectures that
851 * can't support a 'special' bit.
852 */
Dave Jiange1fb4a02018-08-17 15:43:40 -0700853 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
854 !pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700855 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
856 (VM_PFNMAP|VM_MIXEDMAP));
857 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700858
859 if (addr < vma->vm_start || addr >= vma->vm_end)
860 return VM_FAULT_SIGBUS;
Borislav Petkov308a0472016-10-26 19:43:43 +0200861
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700862 if (arch_needs_pgtable_deposit()) {
Joel Fernandes (Google)4cf58922019-01-03 15:28:34 -0800863 pgtable = pte_alloc_one(vma->vm_mm);
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -0700864 if (!pgtable)
865 return VM_FAULT_OOM;
866 }
867
Borislav Petkov308a0472016-10-26 19:43:43 +0200868 track_pfn_insert(vma, &pgprot, pfn);
869
Dan Williamsfce86ff2019-05-13 17:15:33 -0700870 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700871 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700872}
Thomas Hellstrom (VMware)9a9731b2020-03-24 18:48:09 +0100873EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700874
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800875#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800876static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800877{
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800878 if (likely(vma->vm_flags & VM_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800879 pud = pud_mkwrite(pud);
880 return pud;
881}
882
883static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
884 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
885{
886 struct mm_struct *mm = vma->vm_mm;
887 pud_t entry;
888 spinlock_t *ptl;
889
890 ptl = pud_lock(mm, pud);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700891 if (!pud_none(*pud)) {
892 if (write) {
893 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
894 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
895 goto out_unlock;
896 }
897 entry = pud_mkyoung(*pud);
898 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
899 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
900 update_mmu_cache_pud(vma, addr, pud);
901 }
902 goto out_unlock;
903 }
904
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800905 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
906 if (pfn_t_devmap(pfn))
907 entry = pud_mkdevmap(entry);
908 if (write) {
Linus Torvaldsf55e1012017-11-29 09:01:01 -0800909 entry = pud_mkyoung(pud_mkdirty(entry));
910 entry = maybe_pud_mkwrite(entry, vma);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800911 }
912 set_pud_at(mm, addr, pud, entry);
913 update_mmu_cache_pud(vma, addr, pud);
Aneesh Kumar K.Vc6f3c5e2019-04-05 18:39:10 -0700914
915out_unlock:
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800916 spin_unlock(ptl);
917}
918
Thomas Hellstrom (VMware)9a9731b2020-03-24 18:48:09 +0100919/**
920 * vmf_insert_pfn_pud_prot - insert a pud size pfn
921 * @vmf: Structure describing the fault
922 * @pfn: pfn to insert
923 * @pgprot: page protection to use
924 * @write: whether it's a write fault
925 *
926 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
927 * also consult the vmf_insert_mixed_prot() documentation when
928 * @pgprot != @vmf->vma->vm_page_prot.
929 *
930 * Return: vm_fault_t value.
931 */
932vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
933 pgprot_t pgprot, bool write)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800934{
Dan Williamsfce86ff2019-05-13 17:15:33 -0700935 unsigned long addr = vmf->address & PUD_MASK;
936 struct vm_area_struct *vma = vmf->vma;
Dan Williamsfce86ff2019-05-13 17:15:33 -0700937
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800938 /*
939 * If we had pud_special, we could avoid all these restrictions,
940 * but we need to be consistent with PTEs and architectures that
941 * can't support a 'special' bit.
942 */
Dave Jiang62ec0d82018-09-04 15:46:16 -0700943 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
944 !pfn_t_devmap(pfn));
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800945 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
946 (VM_PFNMAP|VM_MIXEDMAP));
947 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800948
949 if (addr < vma->vm_start || addr >= vma->vm_end)
950 return VM_FAULT_SIGBUS;
951
952 track_pfn_insert(vma, &pgprot, pfn);
953
Dan Williamsfce86ff2019-05-13 17:15:33 -0700954 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800955 return VM_FAULT_NOPAGE;
956}
Thomas Hellstrom (VMware)9a9731b2020-03-24 18:48:09 +0100957EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800958#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
959
Dan Williams3565fce2016-01-15 16:56:55 -0800960static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300961 pmd_t *pmd, int flags)
Dan Williams3565fce2016-01-15 16:56:55 -0800962{
963 pmd_t _pmd;
964
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300965 _pmd = pmd_mkyoung(*pmd);
966 if (flags & FOLL_WRITE)
967 _pmd = pmd_mkdirty(_pmd);
Dan Williams3565fce2016-01-15 16:56:55 -0800968 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +0300969 pmd, _pmd, flags & FOLL_WRITE))
Dan Williams3565fce2016-01-15 16:56:55 -0800970 update_mmu_cache_pmd(vma, addr, pmd);
971}
972
973struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
Keith Buschdf06b372018-10-26 15:10:28 -0700974 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
Dan Williams3565fce2016-01-15 16:56:55 -0800975{
976 unsigned long pfn = pmd_pfn(*pmd);
977 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -0800978 struct page *page;
979
980 assert_spin_locked(pmd_lockptr(mm, pmd));
981
Keno Fischer8310d482017-01-24 15:17:48 -0800982 /*
983 * When we COW a devmap PMD entry, we split it into PTEs, so we should
984 * not be in this function with `flags & FOLL_COW` set.
985 */
986 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
987
John Hubbard3faa52c2020-04-01 21:05:29 -0700988 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
989 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
990 (FOLL_PIN | FOLL_GET)))
991 return NULL;
992
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800993 if (flags & FOLL_WRITE && !pmd_write(*pmd))
Dan Williams3565fce2016-01-15 16:56:55 -0800994 return NULL;
995
996 if (pmd_present(*pmd) && pmd_devmap(*pmd))
997 /* pass */;
998 else
999 return NULL;
1000
1001 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001002 touch_pmd(vma, addr, pmd, flags);
Dan Williams3565fce2016-01-15 16:56:55 -08001003
1004 /*
1005 * device mapped pages can only be returned if the
1006 * caller will manage the page reference count.
1007 */
John Hubbard3faa52c2020-04-01 21:05:29 -07001008 if (!(flags & (FOLL_GET | FOLL_PIN)))
Dan Williams3565fce2016-01-15 16:56:55 -08001009 return ERR_PTR(-EEXIST);
1010
1011 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
Keith Buschdf06b372018-10-26 15:10:28 -07001012 *pgmap = get_dev_pagemap(pfn, *pgmap);
1013 if (!*pgmap)
Dan Williams3565fce2016-01-15 16:56:55 -08001014 return ERR_PTR(-EFAULT);
1015 page = pfn_to_page(pfn);
John Hubbard3faa52c2020-04-01 21:05:29 -07001016 if (!try_grab_page(page, flags))
1017 page = ERR_PTR(-ENOMEM);
Dan Williams3565fce2016-01-15 16:56:55 -08001018
1019 return page;
1020}
1021
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001022int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1024 struct vm_area_struct *vma)
1025{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001026 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001027 struct page *src_page;
1028 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -08001029 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001030 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001031
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001032 /* Skip if can be re-fill on fault */
1033 if (!vma_is_anonymous(vma))
1034 return 0;
1035
Joel Fernandes (Google)4cf58922019-01-03 15:28:34 -08001036 pgtable = pte_alloc_one(dst_mm);
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001037 if (unlikely(!pgtable))
1038 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001039
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001040 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1041 src_ptl = pmd_lockptr(src_mm, src_pmd);
1042 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001043
1044 ret = -EAGAIN;
1045 pmd = *src_pmd;
Zi Yan84c3fc42017-09-08 16:11:01 -07001046
1047#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1048 if (unlikely(is_swap_pmd(pmd))) {
1049 swp_entry_t entry = pmd_to_swp_entry(pmd);
1050
1051 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1052 if (is_write_migration_entry(entry)) {
1053 make_migration_entry_read(&entry);
1054 pmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001055 if (pmd_swp_soft_dirty(*src_pmd))
1056 pmd = pmd_swp_mksoft_dirty(pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07001057 set_pmd_at(src_mm, addr, src_pmd, pmd);
1058 }
Zi Yandd8a67f2017-11-02 15:59:47 -07001059 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovaf5b0f62017-11-15 17:35:40 -08001060 mm_inc_nr_ptes(dst_mm);
Zi Yandd8a67f2017-11-02 15:59:47 -07001061 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Zi Yan84c3fc42017-09-08 16:11:01 -07001062 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1063 ret = 0;
1064 goto out_unlock;
1065 }
1066#endif
1067
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001068 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001069 pte_free(dst_mm, pgtable);
1070 goto out_unlock;
1071 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001072 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001073 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001074 * under splitting since we don't split the page itself, only pmd to
1075 * a page table.
1076 */
1077 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001078 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001079 /*
1080 * get_huge_zero_page() will never allocate a new page here,
1081 * since we already have a zero page to copy. It just takes a
1082 * reference.
1083 */
Aaron Lu6fcb52a2016-10-07 17:00:08 -07001084 zero_page = mm_get_huge_zero_page(dst_mm);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07001085 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001086 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001087 ret = 0;
1088 goto out_unlock;
1089 }
Mel Gormande466bd2013-12-18 17:08:42 -08001090
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001091 src_page = pmd_page(pmd);
1092 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1093 get_page(src_page);
1094 page_dup_rmap(src_page, true);
1095 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -08001096 mm_inc_nr_ptes(dst_mm);
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -07001097 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001098
1099 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1100 pmd = pmd_mkold(pmd_wrprotect(pmd));
1101 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001102
1103 ret = 0;
1104out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001105 spin_unlock(src_ptl);
1106 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001107out:
1108 return ret;
1109}
1110
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001111#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1112static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001113 pud_t *pud, int flags)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001114{
1115 pud_t _pud;
1116
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001117 _pud = pud_mkyoung(*pud);
1118 if (flags & FOLL_WRITE)
1119 _pud = pud_mkdirty(_pud);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001120 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001121 pud, _pud, flags & FOLL_WRITE))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001122 update_mmu_cache_pud(vma, addr, pud);
1123}
1124
1125struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
Keith Buschdf06b372018-10-26 15:10:28 -07001126 pud_t *pud, int flags, struct dev_pagemap **pgmap)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001127{
1128 unsigned long pfn = pud_pfn(*pud);
1129 struct mm_struct *mm = vma->vm_mm;
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001130 struct page *page;
1131
1132 assert_spin_locked(pud_lockptr(mm, pud));
1133
Linus Torvaldsf6f37322017-12-15 18:53:22 -08001134 if (flags & FOLL_WRITE && !pud_write(*pud))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001135 return NULL;
1136
John Hubbard3faa52c2020-04-01 21:05:29 -07001137 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1138 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1139 (FOLL_PIN | FOLL_GET)))
1140 return NULL;
1141
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001142 if (pud_present(*pud) && pud_devmap(*pud))
1143 /* pass */;
1144 else
1145 return NULL;
1146
1147 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001148 touch_pud(vma, addr, pud, flags);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001149
1150 /*
1151 * device mapped pages can only be returned if the
1152 * caller will manage the page reference count.
John Hubbard3faa52c2020-04-01 21:05:29 -07001153 *
1154 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001155 */
John Hubbard3faa52c2020-04-01 21:05:29 -07001156 if (!(flags & (FOLL_GET | FOLL_PIN)))
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001157 return ERR_PTR(-EEXIST);
1158
1159 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
Keith Buschdf06b372018-10-26 15:10:28 -07001160 *pgmap = get_dev_pagemap(pfn, *pgmap);
1161 if (!*pgmap)
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001162 return ERR_PTR(-EFAULT);
1163 page = pfn_to_page(pfn);
John Hubbard3faa52c2020-04-01 21:05:29 -07001164 if (!try_grab_page(page, flags))
1165 page = ERR_PTR(-ENOMEM);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08001166
1167 return page;
1168}
1169
1170int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1171 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1172 struct vm_area_struct *vma)
1173{
1174 spinlock_t *dst_ptl, *src_ptl;
1175 pud_t pud;
1176 int ret;
1177
1178 dst_ptl = pud_lock(dst_mm, dst_pud);
1179 src_ptl = pud_lockptr(src_mm, src_pud);
1180 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1181
1182 ret = -EAGAIN;
1183 pud = *src_pud;
1184 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1185 goto out_unlock;
1186
1187 /*
1188 * When page table lock is held, the huge zero pud should not be
1189 * under splitting since we don't split the page itself, only pud to
1190 * a page table.
1191 */
1192 if (is_huge_zero_pud(pud)) {
1193 /* No huge zero pud yet */
1194 }
1195
1196 pudp_set_wrprotect(src_mm, addr, src_pud);
1197 pud = pud_mkold(pud_wrprotect(pud));
1198 set_pud_at(dst_mm, addr, dst_pud, pud);
1199
1200 ret = 0;
1201out_unlock:
1202 spin_unlock(src_ptl);
1203 spin_unlock(dst_ptl);
1204 return ret;
1205}
1206
1207void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1208{
1209 pud_t entry;
1210 unsigned long haddr;
1211 bool write = vmf->flags & FAULT_FLAG_WRITE;
1212
1213 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1214 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1215 goto unlock;
1216
1217 entry = pud_mkyoung(orig_pud);
1218 if (write)
1219 entry = pud_mkdirty(entry);
1220 haddr = vmf->address & HPAGE_PUD_MASK;
1221 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1222 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1223
1224unlock:
1225 spin_unlock(vmf->ptl);
1226}
1227#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1228
Jan Kara82b0f8c2016-12-14 15:06:58 -08001229void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -08001230{
1231 pmd_t entry;
1232 unsigned long haddr;
Minchan Kim20f664a2017-01-10 16:57:51 -08001233 bool write = vmf->flags & FAULT_FLAG_WRITE;
Will Deacona1dd4502012-12-11 16:01:27 -08001234
Jan Kara82b0f8c2016-12-14 15:06:58 -08001235 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1236 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -08001237 goto unlock;
1238
1239 entry = pmd_mkyoung(orig_pmd);
Minchan Kim20f664a2017-01-10 16:57:51 -08001240 if (write)
1241 entry = pmd_mkdirty(entry);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001242 haddr = vmf->address & HPAGE_PMD_MASK;
Minchan Kim20f664a2017-01-10 16:57:51 -08001243 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
Jan Kara82b0f8c2016-12-14 15:06:58 -08001244 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001245
1246unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001247 spin_unlock(vmf->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001248}
1249
Souptick Joarder2b740302018-08-23 17:01:36 -07001250static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1251 pmd_t orig_pmd, struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001252{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001253 struct vm_area_struct *vma = vmf->vma;
1254 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -07001255 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001256 pgtable_t pgtable;
1257 pmd_t _pmd;
Souptick Joarder2b740302018-08-23 17:01:36 -07001258 int i;
1259 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001260 struct page **pages;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001261 struct mmu_notifier_range range;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001262
Kees Cook6da2ec52018-06-12 13:55:00 -07001263 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1264 GFP_KERNEL);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001265 if (unlikely(!pages)) {
1266 ret |= VM_FAULT_OOM;
1267 goto out;
1268 }
1269
1270 for (i = 0; i < HPAGE_PMD_NR; i++) {
Michal Hocko41b61672017-01-10 16:57:42 -08001271 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001272 vmf->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001273 if (unlikely(!pages[i] ||
Tejun Heo2cf85582018-07-03 11:14:56 -04001274 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001275 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001276 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001277 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001278 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001279 memcg = (void *)page_private(pages[i]);
1280 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001281 mem_cgroup_cancel_charge(pages[i], memcg,
1282 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001283 put_page(pages[i]);
1284 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001285 kfree(pages);
1286 ret |= VM_FAULT_OOM;
1287 goto out;
1288 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001289 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001290 }
1291
1292 for (i = 0; i < HPAGE_PMD_NR; i++) {
1293 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001294 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001295 __SetPageUptodate(pages[i]);
1296 cond_resched();
1297 }
1298
Jérôme Glisse7269f992019-05-13 17:20:53 -07001299 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1300 haddr, haddr + HPAGE_PMD_SIZE);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001301 mmu_notifier_invalidate_range_start(&range);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001302
Jan Kara82b0f8c2016-12-14 15:06:58 -08001303 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1304 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001305 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001306 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001307
Jérôme Glisse0f108512017-11-15 17:34:07 -08001308 /*
1309 * Leave pmd empty until pte is filled note we must notify here as
1310 * concurrent CPU thread might write to new page before the call to
1311 * mmu_notifier_invalidate_range_end() happens which can lead to a
1312 * device seeing memory write in different order than CPU.
1313 *
Mike Rapoportad56b732018-03-21 21:22:47 +02001314 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08001315 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001316 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001317
Jan Kara82b0f8c2016-12-14 15:06:58 -08001318 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001319 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001320
1321 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001322 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001323 entry = mk_pte(pages[i], vma->vm_page_prot);
1324 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001325 memcg = (void *)page_private(pages[i]);
1326 set_page_private(pages[i], 0);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001327 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001328 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001329 lru_cache_add_active_or_unevictable(pages[i], vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001330 vmf->pte = pte_offset_map(&_pmd, haddr);
1331 VM_BUG_ON(!pte_none(*vmf->pte));
1332 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1333 pte_unmap(vmf->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001334 }
1335 kfree(pages);
1336
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001337 smp_wmb(); /* make pte visible before pmd */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001338 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001339 page_remove_rmap(page, true);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001340 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001341
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001342 /*
1343 * No need to double call mmu_notifier->invalidate_range() callback as
1344 * the above pmdp_huge_clear_flush_notify() did already call it.
1345 */
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001346 mmu_notifier_invalidate_range_only_end(&range);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001347
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001348 ret |= VM_FAULT_WRITE;
1349 put_page(page);
1350
1351out:
1352 return ret;
1353
1354out_free_pages:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001355 spin_unlock(vmf->ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001356 mmu_notifier_invalidate_range_end(&range);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001357 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001358 memcg = (void *)page_private(pages[i]);
1359 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001360 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001361 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001362 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001363 kfree(pages);
1364 goto out;
1365}
1366
Souptick Joarder2b740302018-08-23 17:01:36 -07001367vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001368{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001369 struct vm_area_struct *vma = vmf->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001370 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001371 struct mem_cgroup *memcg;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001372 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001373 struct mmu_notifier_range range;
Michal Hocko3b3636922015-04-15 16:13:29 -07001374 gfp_t huge_gfp; /* for allocation and charge */
Souptick Joarder2b740302018-08-23 17:01:36 -07001375 vm_fault_t ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001376
Jan Kara82b0f8c2016-12-14 15:06:58 -08001377 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001378 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001379 if (is_huge_zero_pmd(orig_pmd))
1380 goto alloc;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001381 spin_lock(vmf->ptl);
1382 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001383 goto out_unlock;
1384
1385 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001386 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001387 /*
1388 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001389 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001390 */
Huang Yingba3c4ce2017-09-06 16:22:19 -07001391 if (!trylock_page(page)) {
1392 get_page(page);
1393 spin_unlock(vmf->ptl);
1394 lock_page(page);
1395 spin_lock(vmf->ptl);
1396 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1397 unlock_page(page);
1398 put_page(page);
1399 goto out_unlock;
1400 }
1401 put_page(page);
1402 }
1403 if (reuse_swap_page(page, NULL)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001404 pmd_t entry;
1405 entry = pmd_mkyoung(orig_pmd);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001406 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001407 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1408 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001409 ret |= VM_FAULT_WRITE;
Huang Yingba3c4ce2017-09-06 16:22:19 -07001410 unlock_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001411 goto out_unlock;
1412 }
Huang Yingba3c4ce2017-09-06 16:22:19 -07001413 unlock_page(page);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001414 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001415 spin_unlock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001416alloc:
Michal Hocko7635d9c2018-12-28 00:38:21 -08001417 if (__transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001418 !transparent_hugepage_debug_cow()) {
David Rientjes19deb762019-09-04 12:54:20 -07001419 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1420 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001421 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001422 new_page = NULL;
1423
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001424 if (likely(new_page)) {
1425 prep_transhuge_page(new_page);
1426 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001427 if (!page) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001428 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001429 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001430 } else {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001431 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001432 if (ret & VM_FAULT_OOM) {
Jan Kara82b0f8c2016-12-14 15:06:58 -08001433 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001434 ret |= VM_FAULT_FALLBACK;
1435 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001436 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001437 }
David Rientjes17766dd2013-09-12 15:14:06 -07001438 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001439 goto out;
1440 }
1441
Tejun Heo2cf85582018-07-03 11:14:56 -04001442 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
Michal Hocko2a70f6a2018-04-10 16:29:30 -07001443 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001444 put_page(new_page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001445 split_huge_pmd(vma, vmf->pmd, vmf->address);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001446 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001447 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001448 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001449 count_vm_event(THP_FAULT_FALLBACK);
David Rientjes85b9f462020-04-06 20:04:28 -07001450 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001451 goto out;
1452 }
1453
David Rientjes17766dd2013-09-12 15:14:06 -07001454 count_vm_event(THP_FAULT_ALLOC);
Chris Down1ff9e6e2019-03-05 15:48:09 -08001455 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
David Rientjes17766dd2013-09-12 15:14:06 -07001456
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001457 if (!page)
Huang Yingc79b57e2017-09-06 16:25:04 -07001458 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001459 else
Huang Yingc9f4cd72018-08-17 15:45:49 -07001460 copy_user_huge_page(new_page, page, vmf->address,
1461 vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001462 __SetPageUptodate(new_page);
1463
Jérôme Glisse7269f992019-05-13 17:20:53 -07001464 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1465 haddr, haddr + HPAGE_PMD_SIZE);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001466 mmu_notifier_invalidate_range_start(&range);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001467
Jan Kara82b0f8c2016-12-14 15:06:58 -08001468 spin_lock(vmf->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001469 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001470 put_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001471 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1472 spin_unlock(vmf->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001473 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001474 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001475 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001476 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001477 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001478 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
Linus Torvaldsf55e1012017-11-29 09:01:01 -08001479 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001480 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001481 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001482 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001483 lru_cache_add_active_or_unevictable(new_page, vma);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001484 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1485 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001486 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001487 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001488 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001489 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001490 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001491 put_page(page);
1492 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001493 ret |= VM_FAULT_WRITE;
1494 }
Jan Kara82b0f8c2016-12-14 15:06:58 -08001495 spin_unlock(vmf->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001496out_mn:
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08001497 /*
1498 * No need to double call mmu_notifier->invalidate_range() callback as
1499 * the above pmdp_huge_clear_flush_notify() did already call it.
1500 */
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08001501 mmu_notifier_invalidate_range_only_end(&range);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001502out:
1503 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001504out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001505 spin_unlock(vmf->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001506 return ret;
1507}
1508
Keno Fischer8310d482017-01-24 15:17:48 -08001509/*
1510 * FOLL_FORCE can write to even unwritable pmd's, but only
1511 * after we've gone through a COW cycle and they are dirty.
1512 */
1513static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1514{
Linus Torvaldsf6f37322017-12-15 18:53:22 -08001515 return pmd_write(pmd) ||
Keno Fischer8310d482017-01-24 15:17:48 -08001516 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1517}
1518
David Rientjesb676b292012-10-08 16:34:03 -07001519struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001520 unsigned long addr,
1521 pmd_t *pmd,
1522 unsigned int flags)
1523{
David Rientjesb676b292012-10-08 16:34:03 -07001524 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001525 struct page *page = NULL;
1526
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001527 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001528
Keno Fischer8310d482017-01-24 15:17:48 -08001529 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001530 goto out;
1531
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001532 /* Avoid dumping huge zero page */
1533 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1534 return ERR_PTR(-EFAULT);
1535
Mel Gorman2b4847e2013-12-18 17:08:32 -08001536 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001537 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001538 goto out;
1539
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001540 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001541 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
John Hubbard3faa52c2020-04-01 21:05:29 -07001542
1543 if (!try_grab_page(page, flags))
1544 return ERR_PTR(-ENOMEM);
1545
Dan Williams3565fce2016-01-15 16:56:55 -08001546 if (flags & FOLL_TOUCH)
Kirill A. Shutemova8f97362017-11-27 06:21:25 +03001547 touch_pmd(vma, addr, pmd, flags);
John Hubbard3faa52c2020-04-01 21:05:29 -07001548
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001549 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001550 /*
1551 * We don't mlock() pte-mapped THPs. This way we can avoid
1552 * leaking mlocked pages into non-VM_LOCKED VMAs.
1553 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001554 * For anon THP:
1555 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001556 * In most cases the pmd is the only mapping of the page as we
1557 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1558 * writable private mappings in populate_vma_page_range().
1559 *
1560 * The only scenario when we have the page shared here is if we
1561 * mlocking read-only mapping shared over fork(). We skip
1562 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001563 *
1564 * For file THP:
1565 *
1566 * We can expect PageDoubleMap() to be stable under page lock:
1567 * for file pages we set it in page_add_file_rmap(), which
1568 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001569 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001570
1571 if (PageAnon(page) && compound_mapcount(page) != 1)
1572 goto skip_mlock;
1573 if (PageDoubleMap(page) || !page->mapping)
1574 goto skip_mlock;
1575 if (!trylock_page(page))
1576 goto skip_mlock;
1577 lru_add_drain();
1578 if (page->mapping && !PageDoubleMap(page))
1579 mlock_vma_page(page);
1580 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001581 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001582skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001583 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001584 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001585
1586out:
1587 return page;
1588}
1589
Mel Gormand10e63f2012-10-25 14:16:31 +02001590/* NUMA hinting page fault entry point for trans huge pmds */
Souptick Joarder2b740302018-08-23 17:01:36 -07001591vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001592{
Jan Kara82b0f8c2016-12-14 15:06:58 -08001593 struct vm_area_struct *vma = vmf->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001594 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001595 struct page *page;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001596 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001597 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001598 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001599 bool page_locked;
1600 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001601 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001602 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001603
Jan Kara82b0f8c2016-12-14 15:06:58 -08001604 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1605 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001606 goto out_unlock;
1607
Mel Gormande466bd2013-12-18 17:08:42 -08001608 /*
1609 * If there are potential migrations, wait for completion and retry
1610 * without disrupting NUMA hinting information. Do not relock and
1611 * check_same as the page may no longer be mapped.
1612 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001613 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1614 page = pmd_page(*vmf->pmd);
Mark Rutland3c226c62017-06-16 14:02:34 -07001615 if (!get_page_unless_zero(page))
1616 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001617 spin_unlock(vmf->ptl);
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001618 put_and_wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001619 goto out;
1620 }
1621
Mel Gormand10e63f2012-10-25 14:16:31 +02001622 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001623 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001624 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001625 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001626 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001627 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001628 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001629 flags |= TNF_FAULT_LOCAL;
1630 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001631
Mel Gormanbea66fb2015-03-25 15:55:37 -07001632 /* See similar comment in do_numa_page for explanation */
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001633 if (!pmd_savedwrite(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001634 flags |= TNF_NO_GROUP;
1635
1636 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001637 * Acquire the page lock to serialise THP migrations but avoid dropping
1638 * page_table_lock if at all possible
1639 */
Mel Gormanb8916632013-10-07 11:28:44 +01001640 page_locked = trylock_page(page);
1641 target_nid = mpol_misplaced(page, vma, haddr);
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001642 if (target_nid == NUMA_NO_NODE) {
Mel Gormanb8916632013-10-07 11:28:44 +01001643 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001644 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001645 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001646 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001647
Mel Gormande466bd2013-12-18 17:08:42 -08001648 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001649 if (!page_locked) {
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001650 page_nid = NUMA_NO_NODE;
Mark Rutland3c226c62017-06-16 14:02:34 -07001651 if (!get_page_unless_zero(page))
1652 goto out_unlock;
Jan Kara82b0f8c2016-12-14 15:06:58 -08001653 spin_unlock(vmf->ptl);
Hugh Dickins9a1ea432018-12-28 00:36:14 -08001654 put_and_wait_on_page_locked(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001655 goto out;
1656 }
1657
Mel Gorman2b4847e2013-12-18 17:08:32 -08001658 /*
1659 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1660 * to serialises splits
1661 */
Mel Gormanb8916632013-10-07 11:28:44 +01001662 get_page(page);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001663 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001664 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001665
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001666 /* Confirm the PMD did not change while page_table_lock was released */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001667 spin_lock(vmf->ptl);
1668 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001669 unlock_page(page);
1670 put_page(page);
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001671 page_nid = NUMA_NO_NODE;
Mel Gormanb32967f2012-11-19 12:35:47 +00001672 goto out_unlock;
1673 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001674
Mel Gormanc3a489c2013-12-18 17:08:38 -08001675 /* Bail if we fail to protect against THP splits for any reason */
1676 if (unlikely(!anon_vma)) {
1677 put_page(page);
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001678 page_nid = NUMA_NO_NODE;
Mel Gormanc3a489c2013-12-18 17:08:38 -08001679 goto clear_pmdnuma;
1680 }
1681
Mel Gormana54a4072013-10-07 11:28:46 +01001682 /*
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001683 * Since we took the NUMA fault, we must have observed the !accessible
1684 * bit. Make sure all other CPUs agree with that, to avoid them
1685 * modifying the page we're about to migrate.
1686 *
1687 * Must be done under PTL such that we'll observe the relevant
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001688 * inc_tlb_flush_pending().
1689 *
1690 * We are not sure a pending tlb flush here is for a huge page
1691 * mapping or not. Hence use the tlb range variant
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001692 */
Andrea Arcangeli7066f0f2018-10-26 15:10:40 -07001693 if (mm_tlb_flush_pending(vma->vm_mm)) {
Peter Zijlstraccde85b2017-08-11 14:29:01 +02001694 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli7066f0f2018-10-26 15:10:40 -07001695 /*
1696 * change_huge_pmd() released the pmd lock before
1697 * invalidating the secondary MMUs sharing the primary
1698 * MMU pagetables (with ->invalidate_range()). The
1699 * mmu_notifier_invalidate_range_end() (which
1700 * internally calls ->invalidate_range()) in
1701 * change_pmd_range() will run after us, so we can't
1702 * rely on it here and we need an explicit invalidate.
1703 */
1704 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1705 haddr + HPAGE_PMD_SIZE);
1706 }
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001707
1708 /*
Mel Gormana54a4072013-10-07 11:28:46 +01001709 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001710 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001711 */
Jan Kara82b0f8c2016-12-14 15:06:58 -08001712 spin_unlock(vmf->ptl);
Peter Zijlstra8b1b4362017-06-07 18:05:07 +02001713
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001714 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
Jan Kara82b0f8c2016-12-14 15:06:58 -08001715 vmf->pmd, pmd, vmf->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001716 if (migrated) {
1717 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001718 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001719 } else
1720 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001721
Mel Gorman8191acb2013-10-07 11:28:45 +01001722 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001723clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001724 BUG_ON(!PageLocked(page));
Aneesh Kumar K.V288bc542017-02-24 14:59:16 -08001725 was_writable = pmd_savedwrite(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001726 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001727 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001728 if (was_writable)
1729 pmd = pmd_mkwrite(pmd);
Jan Kara82b0f8c2016-12-14 15:06:58 -08001730 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1731 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001732 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001733out_unlock:
Jan Kara82b0f8c2016-12-14 15:06:58 -08001734 spin_unlock(vmf->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001735
1736out:
1737 if (anon_vma)
1738 page_unlock_anon_vma_read(anon_vma);
1739
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001740 if (page_nid != NUMA_NO_NODE)
Jan Kara82b0f8c2016-12-14 15:06:58 -08001741 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
Aneesh Kumar K.V9a8b3002017-02-24 14:59:56 -08001742 flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001743
Mel Gormand10e63f2012-10-25 14:16:31 +02001744 return 0;
1745}
1746
Huang Ying319904a2016-07-28 15:48:03 -07001747/*
1748 * Return true if we do MADV_FREE successfully on entire pmd page.
1749 * Otherwise, return false.
1750 */
1751bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001752 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001753{
1754 spinlock_t *ptl;
1755 pmd_t orig_pmd;
1756 struct page *page;
1757 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001758 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001759
Peter Zijlstraed6a7932018-08-31 14:46:08 +02001760 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001761
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001762 ptl = pmd_trans_huge_lock(pmd, vma);
1763 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001764 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001765
1766 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001767 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001768 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001769
Zi Yan84c3fc42017-09-08 16:11:01 -07001770 if (unlikely(!pmd_present(orig_pmd))) {
1771 VM_BUG_ON(thp_migration_supported() &&
1772 !is_pmd_migration_entry(orig_pmd));
1773 goto out;
1774 }
1775
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001776 page = pmd_page(orig_pmd);
1777 /*
1778 * If other processes are mapping this page, we couldn't discard
1779 * the page unless they all do MADV_FREE so let's skip the page.
1780 */
1781 if (page_mapcount(page) != 1)
1782 goto out;
1783
1784 if (!trylock_page(page))
1785 goto out;
1786
1787 /*
1788 * If user want to discard part-pages of THP, split it so MADV_FREE
1789 * will deactivate only them.
1790 */
1791 if (next - addr != HPAGE_PMD_SIZE) {
1792 get_page(page);
1793 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001794 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001795 unlock_page(page);
Kirill A. Shutemovbbf29ff2017-07-06 15:35:28 -07001796 put_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001797 goto out_unlocked;
1798 }
1799
1800 if (PageDirty(page))
1801 ClearPageDirty(page);
1802 unlock_page(page);
1803
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001804 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
Kirill A. Shutemov58ceeb62017-04-13 14:56:26 -07001805 pmdp_invalidate(vma, addr, pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001806 orig_pmd = pmd_mkold(orig_pmd);
1807 orig_pmd = pmd_mkclean(orig_pmd);
1808
1809 set_pmd_at(mm, addr, pmd, orig_pmd);
1810 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1811 }
Shaohua Li802a3a92017-05-03 14:52:32 -07001812
1813 mark_page_lazyfree(page);
Huang Ying319904a2016-07-28 15:48:03 -07001814 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001815out:
1816 spin_unlock(ptl);
1817out_unlocked:
1818 return ret;
1819}
1820
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001821static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1822{
1823 pgtable_t pgtable;
1824
1825 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1826 pte_free(mm, pgtable);
Kirill A. Shutemovc4812902017-11-15 17:35:37 -08001827 mm_dec_nr_ptes(mm);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001828}
1829
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001830int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001831 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001832{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001833 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001834 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001835
Peter Zijlstraed6a7932018-08-31 14:46:08 +02001836 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08001837
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001838 ptl = __pmd_trans_huge_lock(pmd, vma);
1839 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001840 return 0;
1841 /*
1842 * For architectures like ppc64 we look at deposited pgtable
1843 * when calling pmdp_huge_get_and_clear. So do the
1844 * pgtable_trans_huge_withdraw after finishing pmdp related
1845 * operations.
1846 */
1847 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1848 tlb->fullmm);
1849 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Thomas Hellstrom (VMware)2484ca92020-03-24 18:47:17 +01001850 if (vma_is_special_huge(vma)) {
Oliver O'Halloran3b6521f2017-05-08 15:59:43 -07001851 if (arch_needs_pgtable_deposit())
1852 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001853 spin_unlock(ptl);
1854 if (is_huge_zero_pmd(orig_pmd))
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001855 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001856 } else if (is_huge_zero_pmd(orig_pmd)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001857 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001858 spin_unlock(ptl);
Aneesh Kumar K.Vc0f2e172016-12-12 16:42:31 -08001859 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001860 } else {
Zi Yan616b8372017-09-08 16:10:57 -07001861 struct page *page = NULL;
1862 int flush_needed = 1;
1863
1864 if (pmd_present(orig_pmd)) {
1865 page = pmd_page(orig_pmd);
1866 page_remove_rmap(page, true);
1867 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1868 VM_BUG_ON_PAGE(!PageHead(page), page);
1869 } else if (thp_migration_supported()) {
1870 swp_entry_t entry;
1871
1872 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1873 entry = pmd_to_swp_entry(orig_pmd);
1874 page = pfn_to_page(swp_offset(entry));
1875 flush_needed = 0;
1876 } else
1877 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1878
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001879 if (PageAnon(page)) {
Oliver O'Halloranc14a6eb2017-05-08 15:59:40 -07001880 zap_deposited_table(tlb->mm, pmd);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001881 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1882 } else {
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08001883 if (arch_needs_pgtable_deposit())
1884 zap_deposited_table(tlb->mm, pmd);
Yang Shifadae292018-08-17 15:44:55 -07001885 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001886 }
Zi Yan616b8372017-09-08 16:10:57 -07001887
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001888 spin_unlock(ptl);
Zi Yan616b8372017-09-08 16:10:57 -07001889 if (flush_needed)
1890 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001891 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001892 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001893}
1894
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001895#ifndef pmd_move_must_withdraw
1896static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1897 spinlock_t *old_pmd_ptl,
1898 struct vm_area_struct *vma)
1899{
1900 /*
1901 * With split pmd lock we also need to move preallocated
1902 * PTE page table if new_pmd is on different PMD page table.
1903 *
1904 * We also don't deposit and withdraw tables for file pages.
1905 */
1906 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1907}
1908#endif
1909
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001910static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1911{
1912#ifdef CONFIG_MEM_SOFT_DIRTY
1913 if (unlikely(is_pmd_migration_entry(pmd)))
1914 pmd = pmd_swp_mksoft_dirty(pmd);
1915 else if (pmd_present(pmd))
1916 pmd = pmd_mksoft_dirty(pmd);
1917#endif
1918 return pmd;
1919}
1920
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001921bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001922 unsigned long new_addr, unsigned long old_end,
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001923 pmd_t *old_pmd, pmd_t *new_pmd)
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001924{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001925 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001926 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001927 struct mm_struct *mm = vma->vm_mm;
Aaron Lu5d190422016-11-10 17:16:33 +08001928 bool force_flush = false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001929
1930 if ((old_addr & ~HPAGE_PMD_MASK) ||
1931 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001932 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001933 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001934
1935 /*
1936 * The destination pmd shouldn't be established, free_pgtables()
1937 * should have release it.
1938 */
1939 if (WARN_ON(!pmd_none(*new_pmd))) {
1940 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001941 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001942 }
1943
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001944 /*
1945 * We don't have to worry about the ordering of src and dst
1946 * ptlocks because exclusive mmap_sem prevents deadlock.
1947 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001948 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1949 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001950 new_ptl = pmd_lockptr(mm, new_pmd);
1951 if (new_ptl != old_ptl)
1952 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001953 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001954 if (pmd_present(pmd))
Aaron Lua2ce2662016-11-29 13:27:31 +08001955 force_flush = true;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001956 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001957
Aneesh Kumar K.V1dd38b62016-12-12 16:44:29 -08001958 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301959 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001960 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1961 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001962 }
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07001963 pmd = move_soft_dirty_pmd(pmd);
1964 set_pmd_at(mm, new_addr, new_pmd, pmd);
Aaron Lu5d190422016-11-10 17:16:33 +08001965 if (force_flush)
1966 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
Linus Torvaldseb66ae02018-10-12 15:22:59 -07001967 if (new_ptl != old_ptl)
1968 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001969 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001970 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001971 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001972 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001973}
1974
Mel Gormanf123d742013-10-07 11:28:49 +01001975/*
1976 * Returns
1977 * - 0 if PMD could not be locked
1978 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1979 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1980 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001981int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Peter Xu58705442020-04-06 20:05:45 -07001982 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001983{
1984 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001985 spinlock_t *ptl;
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001986 pmd_t entry;
1987 bool preserve_write;
1988 int ret;
Peter Xu58705442020-04-06 20:05:45 -07001989 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001990
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001991 ptl = __pmd_trans_huge_lock(pmd, vma);
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001992 if (!ptl)
1993 return 0;
Mel Gormane944fd62015-02-12 14:58:35 -08001994
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07001995 preserve_write = prot_numa && pmd_write(*pmd);
1996 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001997
Zi Yan84c3fc42017-09-08 16:11:01 -07001998#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1999 if (is_swap_pmd(*pmd)) {
2000 swp_entry_t entry = pmd_to_swp_entry(*pmd);
2001
2002 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2003 if (is_write_migration_entry(entry)) {
2004 pmd_t newpmd;
2005 /*
2006 * A protection check is difficult so
2007 * just be safe and disable write
2008 */
2009 make_migration_entry_read(&entry);
2010 newpmd = swp_entry_to_pmd(entry);
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07002011 if (pmd_swp_soft_dirty(*pmd))
2012 newpmd = pmd_swp_mksoft_dirty(newpmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07002013 set_pmd_at(mm, addr, pmd, newpmd);
2014 }
2015 goto unlock;
2016 }
2017#endif
2018
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07002019 /*
2020 * Avoid trapping faults against the zero page. The read-only
2021 * data is likely to be read-cached on the local CPU and
2022 * local/remote hits to the zero page are not interesting.
2023 */
2024 if (prot_numa && is_huge_zero_pmd(*pmd))
2025 goto unlock;
Johannes Weinercd7548a2011-01-13 15:47:04 -08002026
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07002027 if (prot_numa && pmd_protnone(*pmd))
2028 goto unlock;
2029
Kirill A. Shutemovced10802017-04-13 14:56:20 -07002030 /*
2031 * In case prot_numa, we are under down_read(mmap_sem). It's critical
2032 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2033 * which is also under down_read(mmap_sem):
2034 *
2035 * CPU0: CPU1:
2036 * change_huge_pmd(prot_numa=1)
2037 * pmdp_huge_get_and_clear_notify()
2038 * madvise_dontneed()
2039 * zap_pmd_range()
2040 * pmd_trans_huge(*pmd) == 0 (without ptl)
2041 * // skip the pmd
2042 * set_pmd_at();
2043 * // pmd is re-established
2044 *
2045 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2046 * which may break userspace.
2047 *
2048 * pmdp_invalidate() is required to make sure we don't miss
2049 * dirty/young flags set by hardware.
2050 */
Kirill A. Shutemova3cf9882018-01-31 16:18:20 -08002051 entry = pmdp_invalidate(vma, addr, pmd);
Kirill A. Shutemovced10802017-04-13 14:56:20 -07002052
Kirill A. Shutemov0a85e51d2017-04-13 14:56:17 -07002053 entry = pmd_modify(entry, newprot);
2054 if (preserve_write)
2055 entry = pmd_mk_savedwrite(entry);
2056 ret = HPAGE_PMD_NR;
2057 set_pmd_at(mm, addr, pmd, entry);
2058 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2059unlock:
2060 spin_unlock(ptl);
Johannes Weinercd7548a2011-01-13 15:47:04 -08002061 return ret;
2062}
2063
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07002064/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07002065 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07002066 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07002067 * Note that if it returns page table lock pointer, this routine returns without
2068 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07002069 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08002070spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07002071{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08002072 spinlock_t *ptl;
2073 ptl = pmd_lock(vma->vm_mm, pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07002074 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2075 pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08002076 return ptl;
2077 spin_unlock(ptl);
2078 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07002079}
2080
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002081/*
2082 * Returns true if a given pud maps a thp, false otherwise.
2083 *
2084 * Note that if it returns true, this routine returns without unlocking page
2085 * table lock. So callers must unlock it.
2086 */
2087spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2088{
2089 spinlock_t *ptl;
2090
2091 ptl = pud_lock(vma->vm_mm, pud);
2092 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2093 return ptl;
2094 spin_unlock(ptl);
2095 return NULL;
2096}
2097
2098#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2099int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2100 pud_t *pud, unsigned long addr)
2101{
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002102 spinlock_t *ptl;
2103
2104 ptl = __pud_trans_huge_lock(pud, vma);
2105 if (!ptl)
2106 return 0;
2107 /*
2108 * For architectures like ppc64 we look at deposited pgtable
2109 * when calling pudp_huge_get_and_clear. So do the
2110 * pgtable_trans_huge_withdraw after finishing pudp related
2111 * operations.
2112 */
Qian Cai70516b92019-03-05 15:50:00 -08002113 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002114 tlb_remove_pud_tlb_entry(tlb, pud, addr);
Thomas Hellstrom (VMware)2484ca92020-03-24 18:47:17 +01002115 if (vma_is_special_huge(vma)) {
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002116 spin_unlock(ptl);
2117 /* No zero page support yet */
2118 } else {
2119 /* No support for anonymous PUD pages yet */
2120 BUG();
2121 }
2122 return 1;
2123}
2124
2125static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2126 unsigned long haddr)
2127{
2128 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2129 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2130 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2131 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2132
Yisheng Xiece9311c2017-03-09 16:17:00 -08002133 count_vm_event(THP_SPLIT_PUD);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002134
2135 pudp_huge_clear_flush_notify(vma, haddr, pud);
2136}
2137
2138void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2139 unsigned long address)
2140{
2141 spinlock_t *ptl;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002142 struct mmu_notifier_range range;
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002143
Jérôme Glisse7269f992019-05-13 17:20:53 -07002144 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07002145 address & HPAGE_PUD_MASK,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002146 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2147 mmu_notifier_invalidate_range_start(&range);
2148 ptl = pud_lock(vma->vm_mm, pud);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002149 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2150 goto out;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002151 __split_huge_pud_locked(vma, pud, range.start);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002152
2153out:
2154 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002155 /*
2156 * No need to double call mmu_notifier->invalidate_range() callback as
2157 * the above pudp_huge_clear_flush_notify() did already call it.
2158 */
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002159 mmu_notifier_invalidate_range_only_end(&range);
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -08002160}
2161#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2162
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002163static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2164 unsigned long haddr, pmd_t *pmd)
2165{
2166 struct mm_struct *mm = vma->vm_mm;
2167 pgtable_t pgtable;
2168 pmd_t _pmd;
2169 int i;
2170
Jérôme Glisse0f108512017-11-15 17:34:07 -08002171 /*
2172 * Leave pmd empty until pte is filled note that it is fine to delay
2173 * notification until mmu_notifier_invalidate_range_end() as we are
2174 * replacing a zero pmd write protected page with a zero pte write
2175 * protected page.
2176 *
Mike Rapoportad56b732018-03-21 21:22:47 +02002177 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08002178 */
2179 pmdp_huge_clear_flush(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002180
2181 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2182 pmd_populate(mm, &_pmd, pgtable);
2183
2184 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2185 pte_t *pte, entry;
2186 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2187 entry = pte_mkspecial(entry);
2188 pte = pte_offset_map(&_pmd, haddr);
2189 VM_BUG_ON(!pte_none(*pte));
2190 set_pte_at(mm, haddr, pte, entry);
2191 pte_unmap(pte);
2192 }
2193 smp_wmb(); /* make pte visible before pmd */
2194 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002195}
2196
2197static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002198 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002199{
2200 struct mm_struct *mm = vma->vm_mm;
2201 struct page *page;
2202 pgtable_t pgtable;
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002203 pmd_t old_pmd, _pmd;
Kirill A. Shutemova3cf9882018-01-31 16:18:20 -08002204 bool young, write, soft_dirty, pmd_migration = false;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002205 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002206 int i;
2207
2208 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2209 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2210 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Zi Yan84c3fc42017-09-08 16:11:01 -07002211 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2212 && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002213
2214 count_vm_event(THP_SPLIT_PMD);
2215
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002216 if (!vma_is_anonymous(vma)) {
2217 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Aneesh Kumar K.V953c66c2016-12-12 16:44:32 -08002218 /*
2219 * We are going to unmap this huge page. So
2220 * just go ahead and zap it
2221 */
2222 if (arch_needs_pgtable_deposit())
2223 zap_deposited_table(mm, pmd);
Thomas Hellstrom (VMware)2484ca92020-03-24 18:47:17 +01002224 if (vma_is_special_huge(vma))
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002225 return;
2226 page = pmd_page(_pmd);
Hugh Dickinse1f1b152018-07-20 17:53:45 -07002227 if (!PageDirty(page) && pmd_dirty(_pmd))
2228 set_page_dirty(page);
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07002229 if (!PageReferenced(page) && pmd_young(_pmd))
2230 SetPageReferenced(page);
2231 page_remove_rmap(page, true);
2232 put_page(page);
Yang Shifadae292018-08-17 15:44:55 -07002233 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002234 return;
2235 } else if (is_huge_zero_pmd(*pmd)) {
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002236 /*
2237 * FIXME: Do we want to invalidate secondary mmu by calling
2238 * mmu_notifier_invalidate_range() see comments below inside
2239 * __split_huge_pmd() ?
2240 *
2241 * We are going from a zero huge page write protected to zero
2242 * small page also write protected so it does not seems useful
2243 * to invalidate secondary mmu at this time.
2244 */
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002245 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2246 }
2247
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002248 /*
2249 * Up to this point the pmd is present and huge and userland has the
2250 * whole access to the hugepage during the split (which happens in
2251 * place). If we overwrite the pmd with the not-huge version pointing
2252 * to the pte here (which of course we could if all CPUs were bug
2253 * free), userland could trigger a small page size TLB miss on the
2254 * small sized TLB while the hugepage TLB entry is still established in
2255 * the huge TLB. Some CPU doesn't like that.
2256 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2257 * 383 on page 93. Intel should be safe but is also warns that it's
2258 * only safe if the permission and cache attributes of the two entries
2259 * loaded in the two TLB is identical (which should be the case here).
2260 * But it is generally safer to never allow small and huge TLB entries
2261 * for the same virtual address to be loaded simultaneously. So instead
2262 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2263 * current pmd notpresent (atomically because here the pmd_trans_huge
2264 * must remain set at all times on the pmd until the split is complete
2265 * for this pmd), then we flush the SMP TLB and finally we write the
2266 * non-huge version of the pmd entry with pmd_populate.
2267 */
2268 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2269
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002270 pmd_migration = is_pmd_migration_entry(old_pmd);
Peter Xu2e83ee12018-12-21 14:30:50 -08002271 if (unlikely(pmd_migration)) {
Zi Yan84c3fc42017-09-08 16:11:01 -07002272 swp_entry_t entry;
2273
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002274 entry = pmd_to_swp_entry(old_pmd);
Zi Yan84c3fc42017-09-08 16:11:01 -07002275 page = pfn_to_page(swp_offset(entry));
Peter Xu2e83ee12018-12-21 14:30:50 -08002276 write = is_write_migration_entry(entry);
2277 young = false;
2278 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2279 } else {
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002280 page = pmd_page(old_pmd);
Peter Xu2e83ee12018-12-21 14:30:50 -08002281 if (pmd_dirty(old_pmd))
2282 SetPageDirty(page);
2283 write = pmd_write(old_pmd);
2284 young = pmd_young(old_pmd);
2285 soft_dirty = pmd_soft_dirty(old_pmd);
2286 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002287 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002288 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002289
Aneesh Kumar K.V423ac9a2018-01-31 16:18:24 -08002290 /*
2291 * Withdraw the table only after we mark the pmd entry invalid.
2292 * This's critical for some architectures (Power).
2293 */
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002294 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2295 pmd_populate(mm, &_pmd, pgtable);
2296
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002297 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002298 pte_t entry, *pte;
2299 /*
2300 * Note that NUMA hinting access restrictions are not
2301 * transferred to avoid any possibility of altering
2302 * permissions across VMAs.
2303 */
Zi Yan84c3fc42017-09-08 16:11:01 -07002304 if (freeze || pmd_migration) {
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002305 swp_entry_t swp_entry;
2306 swp_entry = make_migration_entry(page + i, write);
2307 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002308 if (soft_dirty)
2309 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002310 } else {
Andrea Arcangeli6d2329f2016-10-07 17:01:22 -07002311 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002312 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002313 if (!write)
2314 entry = pte_wrprotect(entry);
2315 if (!young)
2316 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07002317 if (soft_dirty)
2318 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002319 }
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002320 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002321 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002322 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002323 atomic_inc(&page[i]._mapcount);
2324 pte_unmap(pte);
2325 }
2326
2327 /*
2328 * Set PG_double_map before dropping compound_mapcount to avoid
2329 * false-negative page_mapped().
2330 */
2331 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2332 for (i = 0; i < HPAGE_PMD_NR; i++)
2333 atomic_inc(&page[i]._mapcount);
2334 }
2335
2336 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2337 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07002338 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002339 if (TestClearPageDoubleMap(page)) {
2340 /* No need in mapcount reference anymore */
2341 for (i = 0; i < HPAGE_PMD_NR; i++)
2342 atomic_dec(&page[i]._mapcount);
2343 }
2344 }
2345
2346 smp_wmb(); /* make pte visible before pmd */
2347 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002348
2349 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002350 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002351 page_remove_rmap(page + i, false);
2352 put_page(page + i);
2353 }
2354 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002355}
2356
2357void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002358 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002359{
2360 spinlock_t *ptl;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002361 struct mmu_notifier_range range;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002362
Jérôme Glisse7269f992019-05-13 17:20:53 -07002363 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07002364 address & HPAGE_PMD_MASK,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002365 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2366 mmu_notifier_invalidate_range_start(&range);
2367 ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002368
2369 /*
2370 * If caller asks to setup a migration entries, we need a page to check
2371 * pmd against. Otherwise we can end up replacing wrong page.
2372 */
2373 VM_BUG_ON(freeze && !page);
2374 if (page && page != pmd_page(*pmd))
2375 goto out;
2376
Dan Williams5c7fb562016-01-15 16:56:52 -08002377 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002378 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002379 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07002380 clear_page_mlock(page);
Zi Yan84c3fc42017-09-08 16:11:01 -07002381 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002382 goto out;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002383 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002384out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002385 spin_unlock(ptl);
Jérôme Glisse4645b9f2017-11-15 17:34:11 -08002386 /*
2387 * No need to double call mmu_notifier->invalidate_range() callback.
2388 * They are 3 cases to consider inside __split_huge_pmd_locked():
2389 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2390 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2391 * fault will trigger a flush_notify before pointing to a new page
2392 * (it is fine if the secondary mmu keeps pointing to the old zero
2393 * page in the meantime)
2394 * 3) Split a huge pmd into pte pointing to the same page. No need
2395 * to invalidate secondary tlb entry they are all still valid.
2396 * any further changes to individual pte will notify. So no need
2397 * to call mmu_notifier->invalidate_range()
2398 */
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08002399 mmu_notifier_invalidate_range_only_end(&range);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002400}
2401
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002402void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2403 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002404{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002405 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002406 p4d_t *p4d;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002407 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002408 pmd_t *pmd;
2409
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08002410 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002411 if (!pgd_present(*pgd))
2412 return;
2413
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03002414 p4d = p4d_offset(pgd, address);
2415 if (!p4d_present(*p4d))
2416 return;
2417
2418 pud = pud_offset(p4d, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002419 if (!pud_present(*pud))
2420 return;
2421
2422 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002423
Naoya Horiguchi33f47512016-07-14 12:07:32 -07002424 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002425}
2426
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07002427void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002428 unsigned long start,
2429 unsigned long end,
2430 long adjust_next)
2431{
2432 /*
2433 * If the new start address isn't hpage aligned and it could
2434 * previously contain an hugepage: check if we need to split
2435 * an huge pmd.
2436 */
2437 if (start & ~HPAGE_PMD_MASK &&
2438 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2439 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002440 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002441
2442 /*
2443 * If the new end address isn't hpage aligned and it could
2444 * previously contain an hugepage: check if we need to split
2445 * an huge pmd.
2446 */
2447 if (end & ~HPAGE_PMD_MASK &&
2448 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2449 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002450 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002451
2452 /*
2453 * If we're also updating the vma->vm_next->vm_start, if the new
2454 * vm_next->vm_start isn't page aligned and it could previously
2455 * contain an hugepage: check if we need to split an huge pmd.
2456 */
2457 if (adjust_next > 0) {
2458 struct vm_area_struct *next = vma->vm_next;
2459 unsigned long nstart = next->vm_start;
2460 nstart += adjust_next << PAGE_SHIFT;
2461 if (nstart & ~HPAGE_PMD_MASK &&
2462 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2463 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002464 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002465 }
2466}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002467
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002468static void unmap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002469{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002470 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08002471 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
Minchan Kim666e5a42017-05-03 14:54:20 -07002472 bool unmap_success;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002473
2474 VM_BUG_ON_PAGE(!PageHead(page), page);
2475
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002476 if (PageAnon(page))
Naoya Horiguchib5ff8162017-09-08 16:10:49 -07002477 ttu_flags |= TTU_SPLIT_FREEZE;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002478
Minchan Kim666e5a42017-05-03 14:54:20 -07002479 unmap_success = try_to_unmap(page, ttu_flags);
2480 VM_BUG_ON_PAGE(!unmap_success, page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002481}
2482
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002483static void remap_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002484{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002485 int i;
Kirill A. Shutemovace71a12017-02-24 14:57:45 -08002486 if (PageTransHuge(page)) {
2487 remove_migration_ptes(page, page, true);
2488 } else {
2489 for (i = 0; i < HPAGE_PMD_NR; i++)
2490 remove_migration_ptes(page + i, page + i, true);
2491 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002492}
2493
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002494static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002495 struct lruvec *lruvec, struct list_head *list)
2496{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002497 struct page *page_tail = head + tail;
2498
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002499 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002500
2501 /*
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002502 * Clone page flags before unfreezing refcount.
2503 *
2504 * After successful get_page_unless_zero() might follow flags change,
2505 * for exmaple lock_page() which set PG_waiters.
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002506 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002507 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2508 page_tail->flags |= (head->flags &
2509 ((1L << PG_referenced) |
2510 (1L << PG_swapbacked) |
Huang Ying38d8b4e2017-07-06 15:37:18 -07002511 (1L << PG_swapcache) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002512 (1L << PG_mlocked) |
2513 (1L << PG_uptodate) |
2514 (1L << PG_active) |
Johannes Weiner1899ad12018-10-26 15:06:04 -07002515 (1L << PG_workingset) |
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002516 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002517 (1L << PG_unevictable) |
2518 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002519
Hugh Dickins173d9d92018-11-30 14:10:16 -08002520 /* ->mapping in first tail page is compound_mapcount */
2521 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2522 page_tail);
2523 page_tail->mapping = head->mapping;
2524 page_tail->index = head->index + tail;
2525
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002526 /* Page flags must be visible before we make the page non-compound. */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002527 smp_wmb();
2528
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002529 /*
2530 * Clear PageTail before unfreezing page refcount.
2531 *
2532 * After successful get_page_unless_zero() might follow put_page()
2533 * which needs correct compound_head().
2534 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002535 clear_compound_head(page_tail);
2536
Konstantin Khlebnikov605ca5e2018-04-05 16:23:28 -07002537 /* Finally unfreeze refcount. Additional reference from page cache. */
2538 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2539 PageSwapCache(head)));
2540
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002541 if (page_is_young(head))
2542 set_page_young(page_tail);
2543 if (page_is_idle(head))
2544 set_page_idle(page_tail);
2545
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002546 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
Michal Hocko94723aa2018-04-10 16:30:07 -07002547
2548 /*
2549 * always add to the tail because some iterators expect new
2550 * pages to show after the currently processed elements - e.g.
2551 * migrate_pages
2552 */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002553 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002554}
2555
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002556static void __split_huge_page(struct page *page, struct list_head *list,
Hugh Dickins006d3ff2018-11-30 14:10:21 -08002557 pgoff_t end, unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002558{
2559 struct page *head = compound_head(page);
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08002560 pg_data_t *pgdat = page_pgdat(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002561 struct lruvec *lruvec;
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002562 struct address_space *swap_cache = NULL;
2563 unsigned long offset = 0;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002564 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002565
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08002566 lruvec = mem_cgroup_page_lruvec(head, pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002567
2568 /* complete memcg works before add pages to LRU */
2569 mem_cgroup_split_huge_fixup(head);
2570
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002571 if (PageAnon(head) && PageSwapCache(head)) {
2572 swp_entry_t entry = { .val = page_private(head) };
2573
2574 offset = swp_offset(entry);
2575 swap_cache = swap_address_space(entry);
2576 xa_lock(&swap_cache->i_pages);
2577 }
2578
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002579 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07002580 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002581 /* Some pages can be beyond i_size: drop them from page cache */
2582 if (head[i].index >= end) {
Hugh Dickins2d077d42018-06-01 16:50:45 -07002583 ClearPageDirty(head + i);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002584 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002585 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2586 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002587 put_page(head + i);
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002588 } else if (!PageAnon(page)) {
2589 __xa_store(&head->mapping->i_pages, head[i].index,
2590 head + i, 0);
2591 } else if (swap_cache) {
2592 __xa_store(&swap_cache->i_pages, offset + i,
2593 head + i, 0);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002594 }
2595 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002596
2597 ClearPageCompound(head);
Vlastimil Babkaf7da6772019-08-24 17:54:59 -07002598
2599 split_page_owner(head, HPAGE_PMD_ORDER);
2600
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002601 /* See comment in __split_huge_page_tail() */
2602 if (PageAnon(head)) {
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002603 /* Additional pin to swap cache */
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002604 if (PageSwapCache(head)) {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002605 page_ref_add(head, 2);
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002606 xa_unlock(&swap_cache->i_pages);
2607 } else {
Huang Ying38d8b4e2017-07-06 15:37:18 -07002608 page_ref_inc(head);
Matthew Wilcox (Oracle)41011962019-09-23 15:34:52 -07002609 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002610 } else {
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002611 /* Additional pin to page cache */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002612 page_ref_add(head, 2);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002613 xa_unlock(&head->mapping->i_pages);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002614 }
2615
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08002616 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002617
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002618 remap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002619
2620 for (i = 0; i < HPAGE_PMD_NR; i++) {
2621 struct page *subpage = head + i;
2622 if (subpage == page)
2623 continue;
2624 unlock_page(subpage);
2625
2626 /*
2627 * Subpages may be freed if there wasn't any mapping
2628 * like if add_to_swap() is running on a lru page that
2629 * had its mapping zapped. And freeing these pages
2630 * requires taking the lru_lock so we do the put_page
2631 * of the tail pages after the split is complete.
2632 */
2633 put_page(subpage);
2634 }
2635}
2636
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002637int total_mapcount(struct page *page)
2638{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002639 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002640
2641 VM_BUG_ON_PAGE(PageTail(page), page);
2642
2643 if (likely(!PageCompound(page)))
2644 return atomic_read(&page->_mapcount) + 1;
2645
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002646 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002647 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002648 return compound;
2649 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002650 for (i = 0; i < HPAGE_PMD_NR; i++)
2651 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002652 /* File pages has compound_mapcount included in _mapcount */
2653 if (!PageAnon(page))
2654 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08002655 if (PageDoubleMap(page))
2656 ret -= HPAGE_PMD_NR;
2657 return ret;
2658}
2659
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002660/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002661 * This calculates accurately how many mappings a transparent hugepage
2662 * has (unlike page_mapcount() which isn't fully accurate). This full
2663 * accuracy is primarily needed to know if copy-on-write faults can
2664 * reuse the page and change the mapping to read-write instead of
2665 * copying them. At the same time this returns the total_mapcount too.
2666 *
2667 * The function returns the highest mapcount any one of the subpages
2668 * has. If the return value is one, even if different processes are
2669 * mapping different subpages of the transparent hugepage, they can
2670 * all reuse it, because each process is reusing a different subpage.
2671 *
2672 * The total_mapcount is instead counting all virtual mappings of the
2673 * subpages. If the total_mapcount is equal to "one", it tells the
2674 * caller all mappings belong to the same "mm" and in turn the
2675 * anon_vma of the transparent hugepage can become the vma->anon_vma
2676 * local one as no other process may be mapping any of the subpages.
2677 *
2678 * It would be more accurate to replace page_mapcount() with
2679 * page_trans_huge_mapcount(), however we only use
2680 * page_trans_huge_mapcount() in the copy-on-write faults where we
2681 * need full accuracy to avoid breaking page pinning, because
2682 * page_trans_huge_mapcount() is slower than page_mapcount().
2683 */
2684int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2685{
2686 int i, ret, _total_mapcount, mapcount;
2687
2688 /* hugetlbfs shouldn't call it */
2689 VM_BUG_ON_PAGE(PageHuge(page), page);
2690
2691 if (likely(!PageTransCompound(page))) {
2692 mapcount = atomic_read(&page->_mapcount) + 1;
2693 if (total_mapcount)
2694 *total_mapcount = mapcount;
2695 return mapcount;
2696 }
2697
2698 page = compound_head(page);
2699
2700 _total_mapcount = ret = 0;
2701 for (i = 0; i < HPAGE_PMD_NR; i++) {
2702 mapcount = atomic_read(&page[i]._mapcount) + 1;
2703 ret = max(ret, mapcount);
2704 _total_mapcount += mapcount;
2705 }
2706 if (PageDoubleMap(page)) {
2707 ret -= 1;
2708 _total_mapcount -= HPAGE_PMD_NR;
2709 }
2710 mapcount = compound_mapcount(page);
2711 ret += mapcount;
2712 _total_mapcount += mapcount;
2713 if (total_mapcount)
2714 *total_mapcount = _total_mapcount;
2715 return ret;
2716}
2717
Huang Yingb8f593c2017-07-06 15:37:28 -07002718/* Racy check whether the huge page can be split */
2719bool can_split_huge_page(struct page *page, int *pextra_pins)
2720{
2721 int extra_pins;
2722
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002723 /* Additional pins from page cache */
Huang Yingb8f593c2017-07-06 15:37:28 -07002724 if (PageAnon(page))
2725 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2726 else
2727 extra_pins = HPAGE_PMD_NR;
2728 if (pextra_pins)
2729 *pextra_pins = extra_pins;
2730 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2731}
2732
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002733/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002734 * This function splits huge page into normal pages. @page can point to any
2735 * subpage of huge page to split. Split doesn't change the position of @page.
2736 *
2737 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2738 * The huge page must be locked.
2739 *
2740 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2741 *
2742 * Both head page and tail pages will inherit mapping, flags, and so on from
2743 * the hugepage.
2744 *
2745 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2746 * they are not mapped.
2747 *
2748 * Returns 0 if the hugepage is split successfully.
2749 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2750 * us.
2751 */
2752int split_huge_page_to_list(struct page *page, struct list_head *list)
2753{
2754 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002755 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Wei Yanga8803e62020-01-30 22:14:32 -08002756 struct deferred_split *ds_queue = get_deferred_split_queue(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002757 struct anon_vma *anon_vma = NULL;
2758 struct address_space *mapping = NULL;
2759 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002760 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08002761 unsigned long flags;
Hugh Dickins006d3ff2018-11-30 14:10:21 -08002762 pgoff_t end;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002763
Wei Yangcb829622020-01-30 22:14:29 -08002764 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
Wei Yanga8803e62020-01-30 22:14:32 -08002765 VM_BUG_ON_PAGE(!PageLocked(head), head);
2766 VM_BUG_ON_PAGE(!PageCompound(head), head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002767
Wei Yanga8803e62020-01-30 22:14:32 -08002768 if (PageWriteback(head))
Huang Ying59807682017-09-06 16:22:34 -07002769 return -EBUSY;
2770
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002771 if (PageAnon(head)) {
2772 /*
2773 * The caller does not necessarily hold an mmap_sem that would
2774 * prevent the anon_vma disappearing so we first we take a
2775 * reference to it and then lock the anon_vma for write. This
2776 * is similar to page_lock_anon_vma_read except the write lock
2777 * is taken to serialise against parallel split or collapse
2778 * operations.
2779 */
2780 anon_vma = page_get_anon_vma(head);
2781 if (!anon_vma) {
2782 ret = -EBUSY;
2783 goto out;
2784 }
Hugh Dickins006d3ff2018-11-30 14:10:21 -08002785 end = -1;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002786 mapping = NULL;
2787 anon_vma_lock_write(anon_vma);
2788 } else {
2789 mapping = head->mapping;
2790
2791 /* Truncated ? */
2792 if (!mapping) {
2793 ret = -EBUSY;
2794 goto out;
2795 }
2796
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002797 anon_vma = NULL;
2798 i_mmap_lock_read(mapping);
Hugh Dickins006d3ff2018-11-30 14:10:21 -08002799
2800 /*
2801 *__split_huge_page() may need to trim off pages beyond EOF:
2802 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2803 * which cannot be nested inside the page tree lock. So note
2804 * end now: i_size itself may be changed at any moment, but
2805 * head page lock is good enough to serialize the trimming.
2806 */
2807 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002808 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002809
2810 /*
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002811 * Racy check if we can split the page, before unmap_page() will
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002812 * split PMDs
2813 */
Huang Yingb8f593c2017-07-06 15:37:28 -07002814 if (!can_split_huge_page(head, &extra_pins)) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002815 ret = -EBUSY;
2816 goto out_unlock;
2817 }
2818
Wei Yanga8803e62020-01-30 22:14:32 -08002819 mlocked = PageMlocked(head);
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002820 unmap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002821 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2822
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002823 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2824 if (mlocked)
2825 lru_add_drain();
2826
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002827 /* prevent PageLRU to go away from under us, and freeze lru stats */
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08002828 spin_lock_irqsave(&pgdata->lru_lock, flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002829
2830 if (mapping) {
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002831 XA_STATE(xas, &mapping->i_pages, page_index(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002832
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002833 /*
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002834 * Check if the head page is present in page cache.
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002835 * We assume all tail are present too, if head is there.
2836 */
Matthew Wilcoxaa5dc072017-12-04 10:16:10 -05002837 xa_lock(&mapping->i_pages);
2838 if (xas_load(&xas) != head)
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002839 goto fail;
2840 }
2841
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002842 /* Prevent deferred_split_scan() touching ->_refcount */
Yang Shi364c1ee2019-09-23 15:38:06 -07002843 spin_lock(&ds_queue->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002844 count = page_count(head);
2845 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002846 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002847 if (!list_empty(page_deferred_list(head))) {
Yang Shi364c1ee2019-09-23 15:38:06 -07002848 ds_queue->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002849 list_del(page_deferred_list(head));
2850 }
Wei Yangafb97172020-01-30 22:14:35 -08002851 spin_unlock(&ds_queue->split_queue_lock);
Kirill A. Shutemov06d3eff2019-10-18 20:20:30 -07002852 if (mapping) {
Wei Yanga8803e62020-01-30 22:14:32 -08002853 if (PageSwapBacked(head))
2854 __dec_node_page_state(head, NR_SHMEM_THPS);
Kirill A. Shutemov06d3eff2019-10-18 20:20:30 -07002855 else
Wei Yanga8803e62020-01-30 22:14:32 -08002856 __dec_node_page_state(head, NR_FILE_THPS);
Kirill A. Shutemov06d3eff2019-10-18 20:20:30 -07002857 }
2858
Hugh Dickins006d3ff2018-11-30 14:10:21 -08002859 __split_huge_page(page, list, end, flags);
Huang Ying59807682017-09-06 16:22:34 -07002860 if (PageSwapCache(head)) {
2861 swp_entry_t entry = { .val = page_private(head) };
2862
2863 ret = split_swap_cluster(entry);
2864 } else
2865 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002866 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002867 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2868 pr_alert("total_mapcount: %u, page_count(): %u\n",
2869 mapcount, count);
2870 if (PageTail(page))
2871 dump_page(head, NULL);
2872 dump_page(page, "total_mapcount(head) > 0");
2873 BUG();
2874 }
Yang Shi364c1ee2019-09-23 15:38:06 -07002875 spin_unlock(&ds_queue->split_queue_lock);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002876fail: if (mapping)
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002877 xa_unlock(&mapping->i_pages);
Andrey Ryabininf4b7e272019-03-05 15:49:39 -08002878 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
Hugh Dickins906f9cd2018-11-30 14:10:13 -08002879 remap_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002880 ret = -EBUSY;
2881 }
2882
2883out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002884 if (anon_vma) {
2885 anon_vma_unlock_write(anon_vma);
2886 put_anon_vma(anon_vma);
2887 }
2888 if (mapping)
2889 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002890out:
2891 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2892 return ret;
2893}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002894
2895void free_transhuge_page(struct page *page)
2896{
Yang Shi87eaceb2019-09-23 15:38:15 -07002897 struct deferred_split *ds_queue = get_deferred_split_queue(page);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002898 unsigned long flags;
2899
Yang Shi364c1ee2019-09-23 15:38:06 -07002900 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002901 if (!list_empty(page_deferred_list(page))) {
Yang Shi364c1ee2019-09-23 15:38:06 -07002902 ds_queue->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002903 list_del(page_deferred_list(page));
2904 }
Yang Shi364c1ee2019-09-23 15:38:06 -07002905 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002906 free_compound_page(page);
2907}
2908
2909void deferred_split_huge_page(struct page *page)
2910{
Yang Shi87eaceb2019-09-23 15:38:15 -07002911 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2912#ifdef CONFIG_MEMCG
2913 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2914#endif
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002915 unsigned long flags;
2916
2917 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2918
Yang Shi87eaceb2019-09-23 15:38:15 -07002919 /*
2920 * The try_to_unmap() in page reclaim path might reach here too,
2921 * this may cause a race condition to corrupt deferred split queue.
2922 * And, if page reclaim is already handling the same page, it is
2923 * unnecessary to handle it again in shrinker.
2924 *
2925 * Check PageSwapCache to determine if the page is being
2926 * handled by page reclaim since THP swap would add the page into
2927 * swap cache before calling try_to_unmap().
2928 */
2929 if (PageSwapCache(page))
2930 return;
2931
Yang Shi364c1ee2019-09-23 15:38:06 -07002932 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002933 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002934 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Yang Shi364c1ee2019-09-23 15:38:06 -07002935 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2936 ds_queue->split_queue_len++;
Yang Shi87eaceb2019-09-23 15:38:15 -07002937#ifdef CONFIG_MEMCG
2938 if (memcg)
2939 memcg_set_shrinker_bit(memcg, page_to_nid(page),
2940 deferred_split_shrinker.id);
2941#endif
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002942 }
Yang Shi364c1ee2019-09-23 15:38:06 -07002943 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002944}
2945
2946static unsigned long deferred_split_count(struct shrinker *shrink,
2947 struct shrink_control *sc)
2948{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002949 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Yang Shi364c1ee2019-09-23 15:38:06 -07002950 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
Yang Shi87eaceb2019-09-23 15:38:15 -07002951
2952#ifdef CONFIG_MEMCG
2953 if (sc->memcg)
2954 ds_queue = &sc->memcg->deferred_split_queue;
2955#endif
Yang Shi364c1ee2019-09-23 15:38:06 -07002956 return READ_ONCE(ds_queue->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002957}
2958
2959static unsigned long deferred_split_scan(struct shrinker *shrink,
2960 struct shrink_control *sc)
2961{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002962 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Yang Shi364c1ee2019-09-23 15:38:06 -07002963 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002964 unsigned long flags;
2965 LIST_HEAD(list), *pos, *next;
2966 struct page *page;
2967 int split = 0;
2968
Yang Shi87eaceb2019-09-23 15:38:15 -07002969#ifdef CONFIG_MEMCG
2970 if (sc->memcg)
2971 ds_queue = &sc->memcg->deferred_split_queue;
2972#endif
2973
Yang Shi364c1ee2019-09-23 15:38:06 -07002974 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002975 /* Take pin on all head pages to avoid freeing them under us */
Yang Shi364c1ee2019-09-23 15:38:06 -07002976 list_for_each_safe(pos, next, &ds_queue->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002977 page = list_entry((void *)pos, struct page, mapping);
2978 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002979 if (get_page_unless_zero(page)) {
2980 list_move(page_deferred_list(page), &list);
2981 } else {
2982 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002983 list_del_init(page_deferred_list(page));
Yang Shi364c1ee2019-09-23 15:38:06 -07002984 ds_queue->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002985 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002986 if (!--sc->nr_to_scan)
2987 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002988 }
Yang Shi364c1ee2019-09-23 15:38:06 -07002989 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002990
2991 list_for_each_safe(pos, next, &list) {
2992 page = list_entry((void *)pos, struct page, mapping);
Kirill A. Shutemovfa41b902018-03-22 16:17:31 -07002993 if (!trylock_page(page))
2994 goto next;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002995 /* split_huge_page() removes page from list on success */
2996 if (!split_huge_page(page))
2997 split++;
2998 unlock_page(page);
Kirill A. Shutemovfa41b902018-03-22 16:17:31 -07002999next:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003000 put_page(page);
3001 }
3002
Yang Shi364c1ee2019-09-23 15:38:06 -07003003 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3004 list_splice_tail(&list, &ds_queue->split_queue);
3005 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003006
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08003007 /*
3008 * Stop shrinker if we didn't split any page, but the queue is empty.
3009 * This can happen if pages were freed under us.
3010 */
Yang Shi364c1ee2019-09-23 15:38:06 -07003011 if (!split && list_empty(&ds_queue->split_queue))
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08003012 return SHRINK_STOP;
3013 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003014}
3015
3016static struct shrinker deferred_split_shrinker = {
3017 .count_objects = deferred_split_count,
3018 .scan_objects = deferred_split_scan,
3019 .seeks = DEFAULT_SEEKS,
Yang Shi87eaceb2019-09-23 15:38:15 -07003020 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
3021 SHRINKER_NONSLAB,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003022};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003023
3024#ifdef CONFIG_DEBUG_FS
3025static int split_huge_pages_set(void *data, u64 val)
3026{
3027 struct zone *zone;
3028 struct page *page;
3029 unsigned long pfn, max_zone_pfn;
3030 unsigned long total = 0, split = 0;
3031
3032 if (val != 1)
3033 return -EINVAL;
3034
3035 for_each_populated_zone(zone) {
3036 max_zone_pfn = zone_end_pfn(zone);
3037 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3038 if (!pfn_valid(pfn))
3039 continue;
3040
3041 page = pfn_to_page(pfn);
3042 if (!get_page_unless_zero(page))
3043 continue;
3044
3045 if (zone != page_zone(page))
3046 goto next;
3047
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07003048 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003049 goto next;
3050
3051 total++;
3052 lock_page(page);
3053 if (!split_huge_page(page))
3054 split++;
3055 unlock_page(page);
3056next:
3057 put_page(page);
3058 }
3059 }
3060
Yang Shi145bdaa2016-05-05 16:22:00 -07003061 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003062
3063 return 0;
3064}
zhong jiangf1287862019-11-30 17:57:09 -08003065DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003066 "%llu\n");
3067
3068static int __init split_huge_pages_debugfs(void)
3069{
Greg Kroah-Hartmand9f79792019-03-05 15:46:09 -08003070 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3071 &split_huge_pages_fops);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003072 return 0;
3073}
3074late_initcall(split_huge_pages_debugfs);
3075#endif
Zi Yan616b8372017-09-08 16:10:57 -07003076
3077#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3078void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3079 struct page *page)
3080{
3081 struct vm_area_struct *vma = pvmw->vma;
3082 struct mm_struct *mm = vma->vm_mm;
3083 unsigned long address = pvmw->address;
3084 pmd_t pmdval;
3085 swp_entry_t entry;
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07003086 pmd_t pmdswp;
Zi Yan616b8372017-09-08 16:10:57 -07003087
3088 if (!(pvmw->pmd && !pvmw->pte))
3089 return;
3090
Zi Yan616b8372017-09-08 16:10:57 -07003091 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
Huang Ying8a8683a2020-03-05 22:28:29 -08003092 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
Zi Yan616b8372017-09-08 16:10:57 -07003093 if (pmd_dirty(pmdval))
3094 set_page_dirty(page);
3095 entry = make_migration_entry(page, pmd_write(pmdval));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07003096 pmdswp = swp_entry_to_pmd(entry);
3097 if (pmd_soft_dirty(pmdval))
3098 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3099 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
Zi Yan616b8372017-09-08 16:10:57 -07003100 page_remove_rmap(page, true);
3101 put_page(page);
Zi Yan616b8372017-09-08 16:10:57 -07003102}
3103
3104void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3105{
3106 struct vm_area_struct *vma = pvmw->vma;
3107 struct mm_struct *mm = vma->vm_mm;
3108 unsigned long address = pvmw->address;
3109 unsigned long mmun_start = address & HPAGE_PMD_MASK;
3110 pmd_t pmde;
3111 swp_entry_t entry;
3112
3113 if (!(pvmw->pmd && !pvmw->pte))
3114 return;
3115
3116 entry = pmd_to_swp_entry(*pvmw->pmd);
3117 get_page(new);
3118 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
Naoya Horiguchiab6e3d02017-09-08 16:11:04 -07003119 if (pmd_swp_soft_dirty(*pvmw->pmd))
3120 pmde = pmd_mksoft_dirty(pmde);
Zi Yan616b8372017-09-08 16:10:57 -07003121 if (is_write_migration_entry(entry))
Linus Torvaldsf55e1012017-11-29 09:01:01 -08003122 pmde = maybe_pmd_mkwrite(pmde, vma);
Zi Yan616b8372017-09-08 16:10:57 -07003123
3124 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
Naoya Horiguchie71769a2018-04-20 14:55:45 -07003125 if (PageAnon(new))
3126 page_add_anon_rmap(new, vma, mmun_start, true);
3127 else
3128 page_add_file_rmap(new, true);
Zi Yan616b8372017-09-08 16:10:57 -07003129 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
Kirill A. Shutemove125fe42018-10-05 15:51:41 -07003130 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
Zi Yan616b8372017-09-08 16:10:57 -07003131 mlock_vma_page(new);
3132 update_mmu_cache_pmd(vma, address, pvmw->pmd);
3133}
3134#endif