blob: a6abd76baa725d56eb1e9adeb8ee1b6fca86cb80 [file] [log] [blame]
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
Andrew Mortonae3a8c12014-06-04 16:06:58 -07008#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080010#include <linux/mm.h>
11#include <linux/sched.h>
12#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>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080033
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080034#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
Andrea Arcangeliba761492011-01-13 15:46:58 -080038/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080039 * By default transparent hugepage support is disabled in order that avoid
40 * to risk increase the memory footprint of applications without a guaranteed
41 * benefit. When transparent hugepage support is enabled, is for all mappings,
42 * and khugepaged scans all mappings.
43 * Defrag is invoked by khugepaged hugepage allocations and by page faults
44 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080045 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080046unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080047#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080048 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080049#endif
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070053 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080054 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080056
Kirill A. Shutemov9a982252016-01-15 16:54:17 -080057static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -080058
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080059static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -080060struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -080061
Matthew Wilcoxfc437042015-09-08 14:58:51 -070062struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080063{
64 struct page *zero_page;
65retry:
66 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -070067 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080068
69 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080071 if (!zero_page) {
72 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070073 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -080074 }
75 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080076 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -070077 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080078 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -070079 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080080 goto retry;
81 }
82
83 /* We take additional reference here. It will be put back by shrinker */
84 atomic_set(&huge_zero_refcount, 2);
85 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -070086 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080087}
88
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -070089void put_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080090{
91 /*
92 * Counter should never go to zero here. Only shrinker can put
93 * last reference.
94 */
95 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96}
97
Glauber Costa48896462013-08-28 10:18:15 +100098static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
99 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800100{
Glauber Costa48896462013-08-28 10:18:15 +1000101 /* we can free zero page only if last reference remains */
102 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
103}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800104
Glauber Costa48896462013-08-28 10:18:15 +1000105static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
106 struct shrink_control *sc)
107{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800108 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700109 struct page *zero_page = xchg(&huge_zero_page, NULL);
110 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700111 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000112 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800113 }
114
115 return 0;
116}
117
118static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000119 .count_objects = shrink_huge_zero_page_count,
120 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800121 .seeks = DEFAULT_SEEKS,
122};
123
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800124#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800125
Mel Gorman444eb2a42016-03-17 14:19:23 -0700126static ssize_t triple_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800127 struct kobj_attribute *attr,
128 const char *buf, size_t count,
129 enum transparent_hugepage_flag enabled,
Mel Gorman444eb2a42016-03-17 14:19:23 -0700130 enum transparent_hugepage_flag deferred,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800131 enum transparent_hugepage_flag req_madv)
132{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700133 if (!memcmp("defer", buf,
134 min(sizeof("defer")-1, count))) {
135 if (enabled == deferred)
136 return -EINVAL;
137 clear_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800138 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700139 set_bit(deferred, &transparent_hugepage_flags);
140 } else if (!memcmp("always", buf,
141 min(sizeof("always")-1, count))) {
142 clear_bit(deferred, &transparent_hugepage_flags);
143 clear_bit(req_madv, &transparent_hugepage_flags);
144 set_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800145 } else if (!memcmp("madvise", buf,
146 min(sizeof("madvise")-1, count))) {
147 clear_bit(enabled, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700148 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800149 set_bit(req_madv, &transparent_hugepage_flags);
150 } else if (!memcmp("never", buf,
151 min(sizeof("never")-1, count))) {
152 clear_bit(enabled, &transparent_hugepage_flags);
153 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700154 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800155 } else
156 return -EINVAL;
157
158 return count;
159}
160
161static ssize_t enabled_show(struct kobject *kobj,
162 struct kobj_attribute *attr, char *buf)
163{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700164 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
165 return sprintf(buf, "[always] madvise never\n");
166 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
167 return sprintf(buf, "always [madvise] never\n");
168 else
169 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800170}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700171
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800172static ssize_t enabled_store(struct kobject *kobj,
173 struct kobj_attribute *attr,
174 const char *buf, size_t count)
175{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800176 ssize_t ret;
177
Mel Gorman444eb2a42016-03-17 14:19:23 -0700178 ret = triple_flag_store(kobj, attr, buf, count,
179 TRANSPARENT_HUGEPAGE_FLAG,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800180 TRANSPARENT_HUGEPAGE_FLAG,
181 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
182
183 if (ret > 0) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700184 int err = start_stop_khugepaged();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800185 if (err)
186 ret = err;
187 }
188
189 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800190}
191static struct kobj_attribute enabled_attr =
192 __ATTR(enabled, 0644, enabled_show, enabled_store);
193
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700194ssize_t single_hugepage_flag_show(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800195 struct kobj_attribute *attr, char *buf,
196 enum transparent_hugepage_flag flag)
197{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700198 return sprintf(buf, "%d\n",
199 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800200}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700201
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700202ssize_t single_hugepage_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800203 struct kobj_attribute *attr,
204 const char *buf, size_t count,
205 enum transparent_hugepage_flag flag)
206{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700207 unsigned long value;
208 int ret;
209
210 ret = kstrtoul(buf, 10, &value);
211 if (ret < 0)
212 return ret;
213 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800214 return -EINVAL;
215
Ben Hutchingse27e6152011-04-14 15:22:21 -0700216 if (value)
217 set_bit(flag, &transparent_hugepage_flags);
218 else
219 clear_bit(flag, &transparent_hugepage_flags);
220
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800221 return count;
222}
223
224/*
225 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
226 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
227 * memory just to allocate one more hugepage.
228 */
229static ssize_t defrag_show(struct kobject *kobj,
230 struct kobj_attribute *attr, char *buf)
231{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700232 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
233 return sprintf(buf, "[always] defer madvise never\n");
234 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
235 return sprintf(buf, "always [defer] madvise never\n");
236 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
237 return sprintf(buf, "always defer [madvise] never\n");
238 else
239 return sprintf(buf, "always defer madvise [never]\n");
240
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800241}
242static ssize_t defrag_store(struct kobject *kobj,
243 struct kobj_attribute *attr,
244 const char *buf, size_t count)
245{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700246 return triple_flag_store(kobj, attr, buf, count,
247 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
248 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800249 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
250}
251static struct kobj_attribute defrag_attr =
252 __ATTR(defrag, 0644, defrag_show, defrag_store);
253
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800254static ssize_t use_zero_page_show(struct kobject *kobj,
255 struct kobj_attribute *attr, char *buf)
256{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700257 return single_hugepage_flag_show(kobj, attr, buf,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800258 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
259}
260static ssize_t use_zero_page_store(struct kobject *kobj,
261 struct kobj_attribute *attr, const char *buf, size_t count)
262{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700263 return single_hugepage_flag_store(kobj, attr, buf, count,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800264 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
265}
266static struct kobj_attribute use_zero_page_attr =
267 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800268#ifdef CONFIG_DEBUG_VM
269static ssize_t debug_cow_show(struct kobject *kobj,
270 struct kobj_attribute *attr, char *buf)
271{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700272 return single_hugepage_flag_show(kobj, attr, buf,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800273 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
274}
275static ssize_t debug_cow_store(struct kobject *kobj,
276 struct kobj_attribute *attr,
277 const char *buf, size_t count)
278{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700279 return single_hugepage_flag_store(kobj, attr, buf, count,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800280 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
281}
282static struct kobj_attribute debug_cow_attr =
283 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
284#endif /* CONFIG_DEBUG_VM */
285
286static struct attribute *hugepage_attr[] = {
287 &enabled_attr.attr,
288 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800289 &use_zero_page_attr.attr,
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700290#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemov5a6e75f2016-07-26 15:26:13 -0700291 &shmem_enabled_attr.attr,
292#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800293#ifdef CONFIG_DEBUG_VM
294 &debug_cow_attr.attr,
295#endif
296 NULL,
297};
298
299static struct attribute_group hugepage_attr_group = {
300 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800301};
302
Shaohua Li569e5592012-01-12 17:19:11 -0800303static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
304{
305 int err;
306
307 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
308 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700309 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800310 return -ENOMEM;
311 }
312
313 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
314 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700315 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800316 goto delete_obj;
317 }
318
319 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
320 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700321 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800322 goto remove_hp_group;
323 }
324
325 return 0;
326
327remove_hp_group:
328 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
329delete_obj:
330 kobject_put(*hugepage_kobj);
331 return err;
332}
333
334static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
335{
336 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
337 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
338 kobject_put(hugepage_kobj);
339}
340#else
341static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
342{
343 return 0;
344}
345
346static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
347{
348}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800349#endif /* CONFIG_SYSFS */
350
351static int __init hugepage_init(void)
352{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800353 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800354 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800355
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800356 if (!has_transparent_hugepage()) {
357 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800358 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800359 }
360
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530361 /*
362 * hugepages can't be allocated by the buddy allocator
363 */
364 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
365 /*
366 * we use page->mapping and page->index in second tail page
367 * as list_head: assuming THP order >= 2
368 */
369 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
370
Shaohua Li569e5592012-01-12 17:19:11 -0800371 err = hugepage_init_sysfs(&hugepage_kobj);
372 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700373 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800374
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700375 err = khugepaged_init();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800376 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700377 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800378
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700379 err = register_shrinker(&huge_zero_page_shrinker);
380 if (err)
381 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800382 err = register_shrinker(&deferred_split_shrinker);
383 if (err)
384 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800385
Rik van Riel97562cd2011-01-13 15:47:12 -0800386 /*
387 * By default disable transparent hugepages on smaller systems,
388 * where the extra memory used could hurt more than TLB overhead
389 * is likely to save. The admin can still enable it through /sys.
390 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700391 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800392 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700393 return 0;
394 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800395
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700396 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700397 if (err)
398 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800399
Shaohua Li569e5592012-01-12 17:19:11 -0800400 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700401err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800402 unregister_shrinker(&deferred_split_shrinker);
403err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700404 unregister_shrinker(&huge_zero_page_shrinker);
405err_hzp_shrinker:
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700406 khugepaged_destroy();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700407err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800408 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700409err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800410 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800411}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800412subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800413
414static int __init setup_transparent_hugepage(char *str)
415{
416 int ret = 0;
417 if (!str)
418 goto out;
419 if (!strcmp(str, "always")) {
420 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
421 &transparent_hugepage_flags);
422 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
423 &transparent_hugepage_flags);
424 ret = 1;
425 } else if (!strcmp(str, "madvise")) {
426 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
427 &transparent_hugepage_flags);
428 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
429 &transparent_hugepage_flags);
430 ret = 1;
431 } else if (!strcmp(str, "never")) {
432 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
433 &transparent_hugepage_flags);
434 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
435 &transparent_hugepage_flags);
436 ret = 1;
437 }
438out:
439 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700440 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800441 return ret;
442}
443__setup("transparent_hugepage=", setup_transparent_hugepage);
444
Mel Gormanb32967f2012-11-19 12:35:47 +0000445pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800446{
447 if (likely(vma->vm_flags & VM_WRITE))
448 pmd = pmd_mkwrite(pmd);
449 return pmd;
450}
451
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800452static inline struct list_head *page_deferred_list(struct page *page)
453{
454 /*
455 * ->lru in the tail pages is occupied by compound_head.
456 * Let's use ->mapping + ->index in the second tail page as list_head.
457 */
458 return (struct list_head *)&page[2].mapping;
459}
460
461void prep_transhuge_page(struct page *page)
462{
463 /*
464 * we use page->mapping and page->indexlru in second tail page
465 * as list_head: assuming THP order >= 2
466 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800467
468 INIT_LIST_HEAD(page_deferred_list(page));
469 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
470}
471
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700472static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
473 gfp_t gfp)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800474{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700475 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -0700476 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800477 pgtable_t pgtable;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700478 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800479
Sasha Levin309381fea2014-01-23 15:52:54 -0800480 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700481
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700482 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700483 put_page(page);
484 count_vm_event(THP_FAULT_FALLBACK);
485 return VM_FAULT_FALLBACK;
486 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800487
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700488 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700489 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800490 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700491 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700492 return VM_FAULT_OOM;
493 }
494
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800495 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700496 /*
497 * The memory barrier inside __SetPageUptodate makes sure that
498 * clear_huge_page writes become visible before the set_pmd_at()
499 * write.
500 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800501 __SetPageUptodate(page);
502
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700503 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
504 if (unlikely(!pmd_none(*fe->pmd))) {
505 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800506 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800507 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700508 pte_free(vma->vm_mm, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800509 } else {
510 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700511
512 /* Deliver the page fault to userland */
513 if (userfaultfd_missing(vma)) {
514 int ret;
515
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700516 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800517 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700518 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700519 pte_free(vma->vm_mm, pgtable);
520 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700521 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
522 return ret;
523 }
524
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700525 entry = mk_huge_pmd(page, vma->vm_page_prot);
526 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800527 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800528 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700529 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700530 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
531 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
532 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
533 atomic_long_inc(&vma->vm_mm->nr_ptes);
534 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700535 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800536 }
537
David Rientjesaa2e8782012-05-29 15:06:17 -0700538 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800539}
540
Mel Gorman444eb2a42016-03-17 14:19:23 -0700541/*
Vlastimil Babka25160352016-07-28 15:49:25 -0700542 * If THP defrag is set to always then directly reclaim/compact as necessary
543 * If set to defer then do only background reclaim/compact and defer to khugepaged
Mel Gorman444eb2a42016-03-17 14:19:23 -0700544 * If set to madvise and the VMA is flagged then directly reclaim/compact
Vlastimil Babka25160352016-07-28 15:49:25 -0700545 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
Mel Gorman444eb2a42016-03-17 14:19:23 -0700546 */
547static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800548{
Vlastimil Babka25160352016-07-28 15:49:25 -0700549 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700550
Vlastimil Babka25160352016-07-28 15:49:25 -0700551 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
552 &transparent_hugepage_flags) && vma_madvised)
553 return GFP_TRANSHUGE;
554 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
555 &transparent_hugepage_flags))
556 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
557 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
558 &transparent_hugepage_flags))
559 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700560
Vlastimil Babka25160352016-07-28 15:49:25 -0700561 return GFP_TRANSHUGE_LIGHT;
Mel Gorman444eb2a42016-03-17 14:19:23 -0700562}
563
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800564/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700565static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800566 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700567 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800568{
569 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700570 if (!pmd_none(*pmd))
571 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700572 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800573 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800574 if (pgtable)
575 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800576 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800577 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700578 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800579}
580
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700581int do_huge_pmd_anonymous_page(struct fault_env *fe)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800582{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700583 struct vm_area_struct *vma = fe->vma;
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800584 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800585 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700586 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800587
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700588 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700589 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700590 if (unlikely(anon_vma_prepare(vma)))
591 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700592 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700593 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700594 if (!(fe->flags & FAULT_FLAG_WRITE) &&
595 !mm_forbids_zeropage(vma->vm_mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700596 transparent_hugepage_use_zero_page()) {
597 pgtable_t pgtable;
598 struct page *zero_page;
599 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700600 int ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700601 pgtable = pte_alloc_one(vma->vm_mm, haddr);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700602 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800603 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700604 zero_page = get_huge_zero_page();
605 if (unlikely(!zero_page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700606 pte_free(vma->vm_mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700607 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700608 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700609 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700610 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700611 ret = 0;
612 set = false;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700613 if (pmd_none(*fe->pmd)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700614 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700615 spin_unlock(fe->ptl);
616 ret = handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700617 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
618 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700619 set_huge_zero_page(pgtable, vma->vm_mm, vma,
620 haddr, fe->pmd, zero_page);
621 spin_unlock(fe->ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700622 set = true;
623 }
624 } else
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700625 spin_unlock(fe->ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700626 if (!set) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700627 pte_free(vma->vm_mm, pgtable);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700628 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800629 }
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700630 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800631 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700632 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800633 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700634 if (unlikely(!page)) {
635 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700636 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700637 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800638 prep_transhuge_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700639 return __do_huge_pmd_anonymous_page(fe, page, gfp);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800640}
641
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700642static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800643 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700644{
645 struct mm_struct *mm = vma->vm_mm;
646 pmd_t entry;
647 spinlock_t *ptl;
648
649 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800650 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
651 if (pfn_t_devmap(pfn))
652 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800653 if (write) {
654 entry = pmd_mkyoung(pmd_mkdirty(entry));
655 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700656 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800657 set_pmd_at(mm, addr, pmd, entry);
658 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700659 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700660}
661
662int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800663 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700664{
665 pgprot_t pgprot = vma->vm_page_prot;
666 /*
667 * If we had pmd_special, we could avoid all these restrictions,
668 * but we need to be consistent with PTEs and architectures that
669 * can't support a 'special' bit.
670 */
671 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
672 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
673 (VM_PFNMAP|VM_MIXEDMAP));
674 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -0800675 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700676
677 if (addr < vma->vm_start || addr >= vma->vm_end)
678 return VM_FAULT_SIGBUS;
679 if (track_pfn_insert(vma, &pgprot, pfn))
680 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700681 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
682 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700683}
Dan Williamsdee41072016-05-14 12:20:44 -0700684EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700685
Dan Williams3565fce2016-01-15 16:56:55 -0800686static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
687 pmd_t *pmd)
688{
689 pmd_t _pmd;
690
691 /*
692 * We should set the dirty bit only for FOLL_WRITE but for now
693 * the dirty bit in the pmd is meaningless. And if the dirty
694 * bit will become meaningful and we'll only set it with
695 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
696 * set the young bit, instead of the current set_pmd_at.
697 */
698 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
699 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
700 pmd, _pmd, 1))
701 update_mmu_cache_pmd(vma, addr, pmd);
702}
703
704struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
705 pmd_t *pmd, int flags)
706{
707 unsigned long pfn = pmd_pfn(*pmd);
708 struct mm_struct *mm = vma->vm_mm;
709 struct dev_pagemap *pgmap;
710 struct page *page;
711
712 assert_spin_locked(pmd_lockptr(mm, pmd));
713
714 if (flags & FOLL_WRITE && !pmd_write(*pmd))
715 return NULL;
716
717 if (pmd_present(*pmd) && pmd_devmap(*pmd))
718 /* pass */;
719 else
720 return NULL;
721
722 if (flags & FOLL_TOUCH)
723 touch_pmd(vma, addr, pmd);
724
725 /*
726 * device mapped pages can only be returned if the
727 * caller will manage the page reference count.
728 */
729 if (!(flags & FOLL_GET))
730 return ERR_PTR(-EEXIST);
731
732 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
733 pgmap = get_dev_pagemap(pfn, NULL);
734 if (!pgmap)
735 return ERR_PTR(-EFAULT);
736 page = pfn_to_page(pfn);
737 get_page(page);
738 put_dev_pagemap(pgmap);
739
740 return page;
741}
742
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800743int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
744 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
745 struct vm_area_struct *vma)
746{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800747 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800748 struct page *src_page;
749 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800750 pgtable_t pgtable = NULL;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700751 int ret = -ENOMEM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800752
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700753 /* Skip if can be re-fill on fault */
754 if (!vma_is_anonymous(vma))
755 return 0;
756
757 pgtable = pte_alloc_one(dst_mm, addr);
758 if (unlikely(!pgtable))
759 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800760
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800761 dst_ptl = pmd_lock(dst_mm, dst_pmd);
762 src_ptl = pmd_lockptr(src_mm, src_pmd);
763 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800764
765 ret = -EAGAIN;
766 pmd = *src_pmd;
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700767 if (unlikely(!pmd_trans_huge(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800768 pte_free(dst_mm, pgtable);
769 goto out_unlock;
770 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800771 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800772 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800773 * under splitting since we don't split the page itself, only pmd to
774 * a page table.
775 */
776 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700777 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800778 /*
779 * get_huge_zero_page() will never allocate a new page here,
780 * since we already have a zero page to copy. It just takes a
781 * reference.
782 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700783 zero_page = get_huge_zero_page();
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700784 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700785 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800786 ret = 0;
787 goto out_unlock;
788 }
Mel Gormande466bd2013-12-18 17:08:42 -0800789
Kirill A. Shutemov628d47c2016-07-26 15:25:42 -0700790 src_page = pmd_page(pmd);
791 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
792 get_page(src_page);
793 page_dup_rmap(src_page, true);
794 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
795 atomic_long_inc(&dst_mm->nr_ptes);
796 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800797
798 pmdp_set_wrprotect(src_mm, addr, src_pmd);
799 pmd = pmd_mkold(pmd_wrprotect(pmd));
800 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800801
802 ret = 0;
803out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800804 spin_unlock(src_ptl);
805 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800806out:
807 return ret;
808}
809
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700810void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
Will Deacona1dd4502012-12-11 16:01:27 -0800811{
812 pmd_t entry;
813 unsigned long haddr;
814
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700815 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
816 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Will Deacona1dd4502012-12-11 16:01:27 -0800817 goto unlock;
818
819 entry = pmd_mkyoung(orig_pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700820 haddr = fe->address & HPAGE_PMD_MASK;
821 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry,
822 fe->flags & FAULT_FLAG_WRITE))
823 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800824
825unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700826 spin_unlock(fe->ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800827}
828
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700829static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
830 struct page *page)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800831{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700832 struct vm_area_struct *vma = fe->vma;
833 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Johannes Weiner00501b52014-08-08 14:19:20 -0700834 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800835 pgtable_t pgtable;
836 pmd_t _pmd;
837 int ret = 0, i;
838 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700839 unsigned long mmun_start; /* For mmu_notifiers */
840 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800841
842 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
843 GFP_KERNEL);
844 if (unlikely(!pages)) {
845 ret |= VM_FAULT_OOM;
846 goto out;
847 }
848
849 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700850 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700851 __GFP_OTHER_NODE, vma,
852 fe->address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800853 if (unlikely(!pages[i] ||
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700854 mem_cgroup_try_charge(pages[i], vma->vm_mm,
855 GFP_KERNEL, &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800856 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800857 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800858 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700859 memcg = (void *)page_private(pages[i]);
860 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800861 mem_cgroup_cancel_charge(pages[i], memcg,
862 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800863 put_page(pages[i]);
864 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800865 kfree(pages);
866 ret |= VM_FAULT_OOM;
867 goto out;
868 }
Johannes Weiner00501b52014-08-08 14:19:20 -0700869 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800870 }
871
872 for (i = 0; i < HPAGE_PMD_NR; i++) {
873 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700874 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800875 __SetPageUptodate(pages[i]);
876 cond_resched();
877 }
878
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700879 mmun_start = haddr;
880 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700881 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700882
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700883 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
884 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800885 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -0800886 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800887
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700888 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800889 /* leave pmd empty until pte is filled */
890
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700891 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
892 pmd_populate(vma->vm_mm, &_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800893
894 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700895 pte_t entry;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896 entry = mk_pte(pages[i], vma->vm_page_prot);
897 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -0700898 memcg = (void *)page_private(pages[i]);
899 set_page_private(pages[i], 0);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700900 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800901 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -0700902 lru_cache_add_active_or_unevictable(pages[i], vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700903 fe->pte = pte_offset_map(&_pmd, haddr);
904 VM_BUG_ON(!pte_none(*fe->pte));
905 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
906 pte_unmap(fe->pte);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800907 }
908 kfree(pages);
909
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700911 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800912 page_remove_rmap(page, true);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700913 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800914
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700915 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700916
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800917 ret |= VM_FAULT_WRITE;
918 put_page(page);
919
920out:
921 return ret;
922
923out_free_pages:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700924 spin_unlock(fe->ptl);
925 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800926 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700927 memcg = (void *)page_private(pages[i]);
928 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800929 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800930 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800931 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800932 kfree(pages);
933 goto out;
934}
935
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700936int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800937{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700938 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800939 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -0700940 struct mem_cgroup *memcg;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700941 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700942 unsigned long mmun_start; /* For mmu_notifiers */
943 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -0700944 gfp_t huge_gfp; /* for allocation and charge */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700945 int ret = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800946
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700947 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -0700948 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800949 if (is_huge_zero_pmd(orig_pmd))
950 goto alloc;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700951 spin_lock(fe->ptl);
952 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800953 goto out_unlock;
954
955 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800956 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -0800957 /*
958 * We can only reuse the page if nobody else maps the huge page or it's
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -0700959 * part.
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -0800960 */
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -0700961 if (page_trans_huge_mapcount(page, NULL) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800962 pmd_t entry;
963 entry = pmd_mkyoung(orig_pmd);
964 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700965 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
966 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800967 ret |= VM_FAULT_WRITE;
968 goto out_unlock;
969 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800970 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700971 spin_unlock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800972alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800973 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800974 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -0700975 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -0700976 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800977 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800978 new_page = NULL;
979
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800980 if (likely(new_page)) {
981 prep_transhuge_page(new_page);
982 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -0800983 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700984 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -0700985 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800986 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700987 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -0800988 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700989 split_huge_pmd(vma, fe->pmd, fe->address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -0800990 ret |= VM_FAULT_FALLBACK;
991 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800992 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800993 }
David Rientjes17766dd2013-09-12 15:14:06 -0700994 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800995 goto out;
996 }
997
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -0700998 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
999 huge_gfp, &memcg, true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001000 put_page(new_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001001 split_huge_pmd(vma, fe->pmd, fe->address);
1002 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001003 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001004 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001005 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001006 goto out;
1007 }
1008
David Rientjes17766dd2013-09-12 15:14:06 -07001009 count_vm_event(THP_FAULT_ALLOC);
1010
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001011 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001012 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1013 else
1014 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001015 __SetPageUptodate(new_page);
1016
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001017 mmun_start = haddr;
1018 mmun_end = haddr + HPAGE_PMD_SIZE;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001019 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001020
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001021 spin_lock(fe->ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001022 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001023 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001024 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1025 spin_unlock(fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001026 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001027 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001028 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001029 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001030 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001031 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1032 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001033 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001034 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001035 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001036 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001037 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1038 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001039 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001040 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001041 put_huge_zero_page();
1042 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001043 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001044 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001045 put_page(page);
1046 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001047 ret |= VM_FAULT_WRITE;
1048 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001049 spin_unlock(fe->ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001050out_mn:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001051 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001052out:
1053 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001054out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001055 spin_unlock(fe->ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056 return ret;
1057}
1058
David Rientjesb676b292012-10-08 16:34:03 -07001059struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001060 unsigned long addr,
1061 pmd_t *pmd,
1062 unsigned int flags)
1063{
David Rientjesb676b292012-10-08 16:34:03 -07001064 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001065 struct page *page = NULL;
1066
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001067 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001068
1069 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1070 goto out;
1071
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001072 /* Avoid dumping huge zero page */
1073 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1074 return ERR_PTR(-EFAULT);
1075
Mel Gorman2b4847e2013-12-18 17:08:32 -08001076 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001077 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001078 goto out;
1079
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 page = pmd_page(*pmd);
Dan Williamsca120cf2016-09-03 10:38:03 -07001081 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001082 if (flags & FOLL_TOUCH)
1083 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001084 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001085 /*
1086 * We don't mlock() pte-mapped THPs. This way we can avoid
1087 * leaking mlocked pages into non-VM_LOCKED VMAs.
1088 *
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001089 * For anon THP:
1090 *
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001091 * In most cases the pmd is the only mapping of the page as we
1092 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1093 * writable private mappings in populate_vma_page_range().
1094 *
1095 * The only scenario when we have the page shared here is if we
1096 * mlocking read-only mapping shared over fork(). We skip
1097 * mlocking such pages.
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001098 *
1099 * For file THP:
1100 *
1101 * We can expect PageDoubleMap() to be stable under page lock:
1102 * for file pages we set it in page_add_file_rmap(), which
1103 * requires page to be locked.
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001104 */
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001105
1106 if (PageAnon(page) && compound_mapcount(page) != 1)
1107 goto skip_mlock;
1108 if (PageDoubleMap(page) || !page->mapping)
1109 goto skip_mlock;
1110 if (!trylock_page(page))
1111 goto skip_mlock;
1112 lru_add_drain();
1113 if (page->mapping && !PageDoubleMap(page))
1114 mlock_vma_page(page);
1115 unlock_page(page);
David Rientjesb676b292012-10-08 16:34:03 -07001116 }
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001117skip_mlock:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001118 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Dan Williamsca120cf2016-09-03 10:38:03 -07001119 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001120 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001121 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001122
1123out:
1124 return page;
1125}
1126
Mel Gormand10e63f2012-10-25 14:16:31 +02001127/* NUMA hinting page fault entry point for trans huge pmds */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001128int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
Mel Gormand10e63f2012-10-25 14:16:31 +02001129{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001130 struct vm_area_struct *vma = fe->vma;
Mel Gormanb8916632013-10-07 11:28:44 +01001131 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001132 struct page *page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001133 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001134 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001135 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001136 bool page_locked;
1137 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001138 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001139 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001140
Mel Gormanc0e7cad2015-02-12 14:58:41 -08001141 /* A PROT_NONE fault should not end up here */
1142 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1143
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001144 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1145 if (unlikely(!pmd_same(pmd, *fe->pmd)))
Mel Gormand10e63f2012-10-25 14:16:31 +02001146 goto out_unlock;
1147
Mel Gormande466bd2013-12-18 17:08:42 -08001148 /*
1149 * If there are potential migrations, wait for completion and retry
1150 * without disrupting NUMA hinting information. Do not relock and
1151 * check_same as the page may no longer be mapped.
1152 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001153 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1154 page = pmd_page(*fe->pmd);
1155 spin_unlock(fe->ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001156 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001157 goto out;
1158 }
1159
Mel Gormand10e63f2012-10-25 14:16:31 +02001160 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001161 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001162 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001163 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001164 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001165 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001166 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001167 flags |= TNF_FAULT_LOCAL;
1168 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001169
Mel Gormanbea66fb2015-03-25 15:55:37 -07001170 /* See similar comment in do_numa_page for explanation */
1171 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001172 flags |= TNF_NO_GROUP;
1173
1174 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001175 * Acquire the page lock to serialise THP migrations but avoid dropping
1176 * page_table_lock if at all possible
1177 */
Mel Gormanb8916632013-10-07 11:28:44 +01001178 page_locked = trylock_page(page);
1179 target_nid = mpol_misplaced(page, vma, haddr);
1180 if (target_nid == -1) {
1181 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001182 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001183 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001184 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001185
Mel Gormande466bd2013-12-18 17:08:42 -08001186 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001187 if (!page_locked) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001188 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001189 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001190 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001191 goto out;
1192 }
1193
Mel Gorman2b4847e2013-12-18 17:08:32 -08001194 /*
1195 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1196 * to serialises splits
1197 */
Mel Gormanb8916632013-10-07 11:28:44 +01001198 get_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001199 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001200 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001201
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001202 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001203 spin_lock(fe->ptl);
1204 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001205 unlock_page(page);
1206 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001207 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001208 goto out_unlock;
1209 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001210
Mel Gormanc3a489c2013-12-18 17:08:38 -08001211 /* Bail if we fail to protect against THP splits for any reason */
1212 if (unlikely(!anon_vma)) {
1213 put_page(page);
1214 page_nid = -1;
1215 goto clear_pmdnuma;
1216 }
1217
Mel Gormana54a4072013-10-07 11:28:46 +01001218 /*
1219 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001220 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001221 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001222 spin_unlock(fe->ptl);
1223 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1224 fe->pmd, pmd, fe->address, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001225 if (migrated) {
1226 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001227 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001228 } else
1229 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001230
Mel Gorman8191acb2013-10-07 11:28:45 +01001231 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001232clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001233 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001234 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001235 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001236 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001237 if (was_writable)
1238 pmd = pmd_mkwrite(pmd);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001239 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1240 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
Mel Gormana54a4072013-10-07 11:28:46 +01001241 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001242out_unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001243 spin_unlock(fe->ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001244
1245out:
1246 if (anon_vma)
1247 page_unlock_anon_vma_read(anon_vma);
1248
Mel Gorman8191acb2013-10-07 11:28:45 +01001249 if (page_nid != -1)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07001250 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001251
Mel Gormand10e63f2012-10-25 14:16:31 +02001252 return 0;
1253}
1254
Huang Ying319904a2016-07-28 15:48:03 -07001255/*
1256 * Return true if we do MADV_FREE successfully on entire pmd page.
1257 * Otherwise, return false.
1258 */
1259bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001260 pmd_t *pmd, unsigned long addr, unsigned long next)
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001261{
1262 spinlock_t *ptl;
1263 pmd_t orig_pmd;
1264 struct page *page;
1265 struct mm_struct *mm = tlb->mm;
Huang Ying319904a2016-07-28 15:48:03 -07001266 bool ret = false;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001267
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001268 ptl = pmd_trans_huge_lock(pmd, vma);
1269 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001270 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001271
1272 orig_pmd = *pmd;
Huang Ying319904a2016-07-28 15:48:03 -07001273 if (is_huge_zero_pmd(orig_pmd))
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001274 goto out;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001275
1276 page = pmd_page(orig_pmd);
1277 /*
1278 * If other processes are mapping this page, we couldn't discard
1279 * the page unless they all do MADV_FREE so let's skip the page.
1280 */
1281 if (page_mapcount(page) != 1)
1282 goto out;
1283
1284 if (!trylock_page(page))
1285 goto out;
1286
1287 /*
1288 * If user want to discard part-pages of THP, split it so MADV_FREE
1289 * will deactivate only them.
1290 */
1291 if (next - addr != HPAGE_PMD_SIZE) {
1292 get_page(page);
1293 spin_unlock(ptl);
Huang Ying9818b8c2016-07-14 12:07:12 -07001294 split_huge_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001295 put_page(page);
1296 unlock_page(page);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001297 goto out_unlocked;
1298 }
1299
1300 if (PageDirty(page))
1301 ClearPageDirty(page);
1302 unlock_page(page);
1303
1304 if (PageActive(page))
1305 deactivate_page(page);
1306
1307 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1308 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1309 tlb->fullmm);
1310 orig_pmd = pmd_mkold(orig_pmd);
1311 orig_pmd = pmd_mkclean(orig_pmd);
1312
1313 set_pmd_at(mm, addr, pmd, orig_pmd);
1314 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1315 }
Huang Ying319904a2016-07-28 15:48:03 -07001316 ret = true;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001317out:
1318 spin_unlock(ptl);
1319out_unlocked:
1320 return ret;
1321}
1322
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001323int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001324 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001325{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001326 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001327 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001328
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001329 ptl = __pmd_trans_huge_lock(pmd, vma);
1330 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001331 return 0;
1332 /*
1333 * For architectures like ppc64 we look at deposited pgtable
1334 * when calling pmdp_huge_get_and_clear. So do the
1335 * pgtable_trans_huge_withdraw after finishing pmdp related
1336 * operations.
1337 */
1338 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1339 tlb->fullmm);
1340 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1341 if (vma_is_dax(vma)) {
1342 spin_unlock(ptl);
1343 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001344 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001345 } else if (is_huge_zero_pmd(orig_pmd)) {
1346 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1347 atomic_long_dec(&tlb->mm->nr_ptes);
1348 spin_unlock(ptl);
Kirill A. Shutemovaa88b682016-04-28 16:18:27 -07001349 tlb_remove_page(tlb, pmd_page(orig_pmd));
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001350 } else {
1351 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001352 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001353 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001354 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovb5072382016-07-26 15:25:34 -07001355 if (PageAnon(page)) {
1356 pgtable_t pgtable;
1357 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1358 pte_free(tlb->mm, pgtable);
1359 atomic_long_dec(&tlb->mm->nr_ptes);
1360 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1361 } else {
1362 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1363 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001364 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07001365 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001366 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001367 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001368}
1369
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001370bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001371 unsigned long new_addr, unsigned long old_end,
1372 pmd_t *old_pmd, pmd_t *new_pmd)
1373{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001374 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001375 pmd_t pmd;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001376 struct mm_struct *mm = vma->vm_mm;
1377
1378 if ((old_addr & ~HPAGE_PMD_MASK) ||
1379 (new_addr & ~HPAGE_PMD_MASK) ||
Hugh Dickinsbf8616d2016-05-19 17:12:54 -07001380 old_end - old_addr < HPAGE_PMD_SIZE)
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001381 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001382
1383 /*
1384 * The destination pmd shouldn't be established, free_pgtables()
1385 * should have release it.
1386 */
1387 if (WARN_ON(!pmd_none(*new_pmd))) {
1388 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001389 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001390 }
1391
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001392 /*
1393 * We don't have to worry about the ordering of src and dst
1394 * ptlocks because exclusive mmap_sem prevents deadlock.
1395 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001396 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1397 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001398 new_ptl = pmd_lockptr(mm, new_pmd);
1399 if (new_ptl != old_ptl)
1400 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001401 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001402 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001403
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001404 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1405 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301406 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001407 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1408 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001409 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301410 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1411 if (new_ptl != old_ptl)
1412 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001413 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001414 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001415 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001416 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001417}
1418
Mel Gormanf123d742013-10-07 11:28:49 +01001419/*
1420 * Returns
1421 * - 0 if PMD could not be locked
1422 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1423 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1424 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001425int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001426 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001427{
1428 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001429 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001430 int ret = 0;
1431
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001432 ptl = __pmd_trans_huge_lock(pmd, vma);
1433 if (ptl) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001434 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001435 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001436 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001437
1438 /*
1439 * Avoid trapping faults against the zero page. The read-only
1440 * data is likely to be read-cached on the local CPU and
1441 * local/remote hits to the zero page are not interesting.
1442 */
1443 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1444 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001445 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001446 }
1447
Mel Gorman10c10452015-02-12 14:58:44 -08001448 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001449 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001450 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001451 if (preserve_write)
1452 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001453 ret = HPAGE_PMD_NR;
1454 set_pmd_at(mm, addr, pmd, entry);
Kirill A. Shutemovb237ade2016-07-26 15:25:45 -07001455 BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1456 pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001457 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001458 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001459 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001460
1461 return ret;
1462}
1463
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001464/*
Huang Ying8f19b0c2016-07-26 15:27:04 -07001465 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001466 *
Huang Ying8f19b0c2016-07-26 15:27:04 -07001467 * Note that if it returns page table lock pointer, this routine returns without
1468 * unlocking page table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001469 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001470spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001471{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001472 spinlock_t *ptl;
1473 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001474 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001475 return ptl;
1476 spin_unlock(ptl);
1477 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001478}
1479
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001480static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1481 unsigned long haddr, pmd_t *pmd)
1482{
1483 struct mm_struct *mm = vma->vm_mm;
1484 pgtable_t pgtable;
1485 pmd_t _pmd;
1486 int i;
1487
1488 /* leave pmd empty until pte is filled */
1489 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1490
1491 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1492 pmd_populate(mm, &_pmd, pgtable);
1493
1494 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1495 pte_t *pte, entry;
1496 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1497 entry = pte_mkspecial(entry);
1498 pte = pte_offset_map(&_pmd, haddr);
1499 VM_BUG_ON(!pte_none(*pte));
1500 set_pte_at(mm, haddr, pte, entry);
1501 pte_unmap(pte);
1502 }
1503 smp_wmb(); /* make pte visible before pmd */
1504 pmd_populate(mm, pmd, pgtable);
1505 put_huge_zero_page();
1506}
1507
1508static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001509 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001510{
1511 struct mm_struct *mm = vma->vm_mm;
1512 struct page *page;
1513 pgtable_t pgtable;
1514 pmd_t _pmd;
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001515 bool young, write, dirty, soft_dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001516 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001517 int i;
1518
1519 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1520 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1521 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08001522 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001523
1524 count_vm_event(THP_SPLIT_PMD);
1525
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001526 if (!vma_is_anonymous(vma)) {
1527 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001528 if (is_huge_zero_pmd(_pmd))
1529 put_huge_zero_page();
Kirill A. Shutemovd21b9e52016-07-26 15:25:37 -07001530 if (vma_is_dax(vma))
1531 return;
1532 page = pmd_page(_pmd);
1533 if (!PageReferenced(page) && pmd_young(_pmd))
1534 SetPageReferenced(page);
1535 page_remove_rmap(page, true);
1536 put_page(page);
1537 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001538 return;
1539 } else if (is_huge_zero_pmd(*pmd)) {
1540 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1541 }
1542
1543 page = pmd_page(*pmd);
1544 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001545 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001546 write = pmd_write(*pmd);
1547 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001548 dirty = pmd_dirty(*pmd);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001549 soft_dirty = pmd_soft_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001550
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05301551 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001552 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1553 pmd_populate(mm, &_pmd, pgtable);
1554
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001555 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001556 pte_t entry, *pte;
1557 /*
1558 * Note that NUMA hinting access restrictions are not
1559 * transferred to avoid any possibility of altering
1560 * permissions across VMAs.
1561 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001562 if (freeze) {
1563 swp_entry_t swp_entry;
1564 swp_entry = make_migration_entry(page + i, write);
1565 entry = swp_entry_to_pte(swp_entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001566 if (soft_dirty)
1567 entry = pte_swp_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001568 } else {
1569 entry = mk_pte(page + i, vma->vm_page_prot);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001570 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001571 if (!write)
1572 entry = pte_wrprotect(entry);
1573 if (!young)
1574 entry = pte_mkold(entry);
Andrea Arcangeli804dd152016-08-25 15:16:57 -07001575 if (soft_dirty)
1576 entry = pte_mksoft_dirty(entry);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08001577 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001578 if (dirty)
1579 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001580 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001581 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001582 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001583 atomic_inc(&page[i]._mapcount);
1584 pte_unmap(pte);
1585 }
1586
1587 /*
1588 * Set PG_double_map before dropping compound_mapcount to avoid
1589 * false-negative page_mapped().
1590 */
1591 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1592 for (i = 0; i < HPAGE_PMD_NR; i++)
1593 atomic_inc(&page[i]._mapcount);
1594 }
1595
1596 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1597 /* Last compound_mapcount is gone. */
Mel Gorman11fb9982016-07-28 15:46:20 -07001598 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001599 if (TestClearPageDoubleMap(page)) {
1600 /* No need in mapcount reference anymore */
1601 for (i = 0; i < HPAGE_PMD_NR; i++)
1602 atomic_dec(&page[i]._mapcount);
1603 }
1604 }
1605
1606 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001607 /*
1608 * Up to this point the pmd is present and huge and userland has the
1609 * whole access to the hugepage during the split (which happens in
1610 * place). If we overwrite the pmd with the not-huge version pointing
1611 * to the pte here (which of course we could if all CPUs were bug
1612 * free), userland could trigger a small page size TLB miss on the
1613 * small sized TLB while the hugepage TLB entry is still established in
1614 * the huge TLB. Some CPU doesn't like that.
1615 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1616 * 383 on page 93. Intel should be safe but is also warns that it's
1617 * only safe if the permission and cache attributes of the two entries
1618 * loaded in the two TLB is identical (which should be the case here).
1619 * But it is generally safer to never allow small and huge TLB entries
1620 * for the same virtual address to be loaded simultaneously. So instead
1621 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1622 * current pmd notpresent (atomically because here the pmd_trans_huge
1623 * and pmd_trans_splitting must remain set at all times on the pmd
1624 * until the split is complete for this pmd), then we flush the SMP TLB
1625 * and finally we write the non-huge version of the pmd entry with
1626 * pmd_populate.
1627 */
1628 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001629 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001630
1631 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03001632 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001633 page_remove_rmap(page + i, false);
1634 put_page(page + i);
1635 }
1636 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001637}
1638
1639void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001640 unsigned long address, bool freeze, struct page *page)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001641{
1642 spinlock_t *ptl;
1643 struct mm_struct *mm = vma->vm_mm;
1644 unsigned long haddr = address & HPAGE_PMD_MASK;
1645
1646 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1647 ptl = pmd_lock(mm, pmd);
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001648
1649 /*
1650 * If caller asks to setup a migration entries, we need a page to check
1651 * pmd against. Otherwise we can end up replacing wrong page.
1652 */
1653 VM_BUG_ON(freeze && !page);
1654 if (page && page != pmd_page(*pmd))
1655 goto out;
1656
Dan Williams5c7fb562016-01-15 16:56:52 -08001657 if (pmd_trans_huge(*pmd)) {
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001658 page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001659 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07001660 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08001661 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001662 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001663 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001664out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08001665 spin_unlock(ptl);
1666 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1667}
1668
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001669void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1670 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001671{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001672 pgd_t *pgd;
1673 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001674 pmd_t *pmd;
1675
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001676 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07001677 if (!pgd_present(*pgd))
1678 return;
1679
1680 pud = pud_offset(pgd, address);
1681 if (!pud_present(*pud))
1682 return;
1683
1684 pmd = pmd_offset(pud, address);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001685
Naoya Horiguchi33f47512016-07-14 12:07:32 -07001686 __split_huge_pmd(vma, pmd, address, freeze, page);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001687}
1688
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07001689void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001690 unsigned long start,
1691 unsigned long end,
1692 long adjust_next)
1693{
1694 /*
1695 * If the new start address isn't hpage aligned and it could
1696 * previously contain an hugepage: check if we need to split
1697 * an huge pmd.
1698 */
1699 if (start & ~HPAGE_PMD_MASK &&
1700 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1701 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001702 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001703
1704 /*
1705 * If the new end address isn't hpage aligned and it could
1706 * previously contain an hugepage: check if we need to split
1707 * an huge pmd.
1708 */
1709 if (end & ~HPAGE_PMD_MASK &&
1710 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1711 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001712 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001713
1714 /*
1715 * If we're also updating the vma->vm_next->vm_start, if the new
1716 * vm_next->vm_start isn't page aligned and it could previously
1717 * contain an hugepage: check if we need to split an huge pmd.
1718 */
1719 if (adjust_next > 0) {
1720 struct vm_area_struct *next = vma->vm_next;
1721 unsigned long nstart = next->vm_start;
1722 nstart += adjust_next << PAGE_SHIFT;
1723 if (nstart & ~HPAGE_PMD_MASK &&
1724 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1725 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001726 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001727 }
1728}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001729
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001730static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001731{
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001732 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1733 TTU_RMAP_LOCKED;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001734 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001735
1736 VM_BUG_ON_PAGE(!PageHead(page), page);
1737
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001738 if (PageAnon(page))
1739 ttu_flags |= TTU_MIGRATION;
1740
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001741 /* We only need TTU_SPLIT_HUGE_PMD once */
1742 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1743 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1744 /* Cut short if the page is unmapped */
1745 if (page_count(page) == 1)
1746 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001747
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001748 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001749 }
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001750 VM_BUG_ON_PAGE(ret, page + i - 1);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001751}
1752
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001753static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001754{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001755 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001756
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001757 for (i = 0; i < HPAGE_PMD_NR; i++)
1758 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001759}
1760
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001761static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001762 struct lruvec *lruvec, struct list_head *list)
1763{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001764 struct page *page_tail = head + tail;
1765
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001766 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07001767 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001768
1769 /*
Joonsoo Kim0139aa72016-05-19 17:10:49 -07001770 * tail_page->_refcount is zero and not changing from under us. But
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001771 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001772 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1773 * atomic_add(), we would then run atomic_set() concurrently with
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001774 * get_page_unless_zero(), and atomic_set() is implemented in C not
1775 * using locked ops. spin_unlock on x86 sometime uses locked ops
1776 * because of PPro errata 66, 92, so unless somebody can guarantee
1777 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001778 * it's safer to use atomic_inc()/atomic_add().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001779 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001780 if (PageAnon(head)) {
1781 page_ref_inc(page_tail);
1782 } else {
1783 /* Additional pin to radix tree */
1784 page_ref_add(page_tail, 2);
1785 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001786
1787 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1788 page_tail->flags |= (head->flags &
1789 ((1L << PG_referenced) |
1790 (1L << PG_swapbacked) |
1791 (1L << PG_mlocked) |
1792 (1L << PG_uptodate) |
1793 (1L << PG_active) |
1794 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001795 (1L << PG_unevictable) |
1796 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001797
1798 /*
1799 * After clearing PageTail the gup refcount can be released.
1800 * Page flags also must be visible before we make the page non-compound.
1801 */
1802 smp_wmb();
1803
1804 clear_compound_head(page_tail);
1805
1806 if (page_is_young(head))
1807 set_page_young(page_tail);
1808 if (page_is_idle(head))
1809 set_page_idle(page_tail);
1810
1811 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001812 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001813 page_tail);
1814 page_tail->mapping = head->mapping;
1815
1816 page_tail->index = head->index + tail;
1817 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1818 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001819}
1820
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001821static void __split_huge_page(struct page *page, struct list_head *list,
1822 unsigned long flags)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001823{
1824 struct page *head = compound_head(page);
1825 struct zone *zone = page_zone(head);
1826 struct lruvec *lruvec;
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001827 pgoff_t end = -1;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001828 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001829
Mel Gorman599d0c92016-07-28 15:45:31 -07001830 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001831
1832 /* complete memcg works before add pages to LRU */
1833 mem_cgroup_split_huge_fixup(head);
1834
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001835 if (!PageAnon(page))
1836 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1837
1838 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07001839 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001840 /* Some pages can be beyond i_size: drop them from page cache */
1841 if (head[i].index >= end) {
1842 __ClearPageDirty(head + i);
1843 __delete_from_page_cache(head + i, NULL);
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001844 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1845 shmem_uncharge(head->mapping->host, 1);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001846 put_page(head + i);
1847 }
1848 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001849
1850 ClearPageCompound(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001851 /* See comment in __split_huge_page_tail() */
1852 if (PageAnon(head)) {
1853 page_ref_inc(head);
1854 } else {
1855 /* Additional pin to radix tree */
1856 page_ref_add(head, 2);
1857 spin_unlock(&head->mapping->tree_lock);
1858 }
1859
Mel Gormana52633d2016-07-28 15:45:28 -07001860 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001861
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001862 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001863
1864 for (i = 0; i < HPAGE_PMD_NR; i++) {
1865 struct page *subpage = head + i;
1866 if (subpage == page)
1867 continue;
1868 unlock_page(subpage);
1869
1870 /*
1871 * Subpages may be freed if there wasn't any mapping
1872 * like if add_to_swap() is running on a lru page that
1873 * had its mapping zapped. And freeing these pages
1874 * requires taking the lru_lock so we do the put_page
1875 * of the tail pages after the split is complete.
1876 */
1877 put_page(subpage);
1878 }
1879}
1880
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001881int total_mapcount(struct page *page)
1882{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001883 int i, compound, ret;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001884
1885 VM_BUG_ON_PAGE(PageTail(page), page);
1886
1887 if (likely(!PageCompound(page)))
1888 return atomic_read(&page->_mapcount) + 1;
1889
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001890 compound = compound_mapcount(page);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001891 if (PageHuge(page))
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001892 return compound;
1893 ret = compound;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001894 for (i = 0; i < HPAGE_PMD_NR; i++)
1895 ret += atomic_read(&page[i]._mapcount) + 1;
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001896 /* File pages has compound_mapcount included in _mapcount */
1897 if (!PageAnon(page))
1898 return ret - compound * HPAGE_PMD_NR;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08001899 if (PageDoubleMap(page))
1900 ret -= HPAGE_PMD_NR;
1901 return ret;
1902}
1903
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001904/*
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07001905 * This calculates accurately how many mappings a transparent hugepage
1906 * has (unlike page_mapcount() which isn't fully accurate). This full
1907 * accuracy is primarily needed to know if copy-on-write faults can
1908 * reuse the page and change the mapping to read-write instead of
1909 * copying them. At the same time this returns the total_mapcount too.
1910 *
1911 * The function returns the highest mapcount any one of the subpages
1912 * has. If the return value is one, even if different processes are
1913 * mapping different subpages of the transparent hugepage, they can
1914 * all reuse it, because each process is reusing a different subpage.
1915 *
1916 * The total_mapcount is instead counting all virtual mappings of the
1917 * subpages. If the total_mapcount is equal to "one", it tells the
1918 * caller all mappings belong to the same "mm" and in turn the
1919 * anon_vma of the transparent hugepage can become the vma->anon_vma
1920 * local one as no other process may be mapping any of the subpages.
1921 *
1922 * It would be more accurate to replace page_mapcount() with
1923 * page_trans_huge_mapcount(), however we only use
1924 * page_trans_huge_mapcount() in the copy-on-write faults where we
1925 * need full accuracy to avoid breaking page pinning, because
1926 * page_trans_huge_mapcount() is slower than page_mapcount().
1927 */
1928int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
1929{
1930 int i, ret, _total_mapcount, mapcount;
1931
1932 /* hugetlbfs shouldn't call it */
1933 VM_BUG_ON_PAGE(PageHuge(page), page);
1934
1935 if (likely(!PageTransCompound(page))) {
1936 mapcount = atomic_read(&page->_mapcount) + 1;
1937 if (total_mapcount)
1938 *total_mapcount = mapcount;
1939 return mapcount;
1940 }
1941
1942 page = compound_head(page);
1943
1944 _total_mapcount = ret = 0;
1945 for (i = 0; i < HPAGE_PMD_NR; i++) {
1946 mapcount = atomic_read(&page[i]._mapcount) + 1;
1947 ret = max(ret, mapcount);
1948 _total_mapcount += mapcount;
1949 }
1950 if (PageDoubleMap(page)) {
1951 ret -= 1;
1952 _total_mapcount -= HPAGE_PMD_NR;
1953 }
1954 mapcount = compound_mapcount(page);
1955 ret += mapcount;
1956 _total_mapcount += mapcount;
1957 if (total_mapcount)
1958 *total_mapcount = _total_mapcount;
1959 return ret;
1960}
1961
1962/*
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001963 * This function splits huge page into normal pages. @page can point to any
1964 * subpage of huge page to split. Split doesn't change the position of @page.
1965 *
1966 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
1967 * The huge page must be locked.
1968 *
1969 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
1970 *
1971 * Both head page and tail pages will inherit mapping, flags, and so on from
1972 * the hugepage.
1973 *
1974 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
1975 * they are not mapped.
1976 *
1977 * Returns 0 if the hugepage is split successfully.
1978 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
1979 * us.
1980 */
1981int split_huge_page_to_list(struct page *page, struct list_head *list)
1982{
1983 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08001984 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001985 struct anon_vma *anon_vma = NULL;
1986 struct address_space *mapping = NULL;
1987 int count, mapcount, extra_pins, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08001988 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08001989 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001990
1991 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001992 VM_BUG_ON_PAGE(!PageLocked(page), page);
1993 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1994 VM_BUG_ON_PAGE(!PageCompound(page), page);
1995
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07001996 if (PageAnon(head)) {
1997 /*
1998 * The caller does not necessarily hold an mmap_sem that would
1999 * prevent the anon_vma disappearing so we first we take a
2000 * reference to it and then lock the anon_vma for write. This
2001 * is similar to page_lock_anon_vma_read except the write lock
2002 * is taken to serialise against parallel split or collapse
2003 * operations.
2004 */
2005 anon_vma = page_get_anon_vma(head);
2006 if (!anon_vma) {
2007 ret = -EBUSY;
2008 goto out;
2009 }
2010 extra_pins = 0;
2011 mapping = NULL;
2012 anon_vma_lock_write(anon_vma);
2013 } else {
2014 mapping = head->mapping;
2015
2016 /* Truncated ? */
2017 if (!mapping) {
2018 ret = -EBUSY;
2019 goto out;
2020 }
2021
2022 /* Addidional pins from radix tree */
2023 extra_pins = HPAGE_PMD_NR;
2024 anon_vma = NULL;
2025 i_mmap_lock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002026 }
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002027
2028 /*
2029 * Racy check if we can split the page, before freeze_page() will
2030 * split PMDs
2031 */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002032 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002033 ret = -EBUSY;
2034 goto out_unlock;
2035 }
2036
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002037 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002038 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002039 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2040
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08002041 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2042 if (mlocked)
2043 lru_add_drain();
2044
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002045 /* prevent PageLRU to go away from under us, and freeze lru stats */
Mel Gormana52633d2016-07-28 15:45:28 -07002046 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002047
2048 if (mapping) {
2049 void **pslot;
2050
2051 spin_lock(&mapping->tree_lock);
2052 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2053 page_index(head));
2054 /*
2055 * Check if the head page is present in radix tree.
2056 * We assume all tail are present too, if head is there.
2057 */
2058 if (radix_tree_deref_slot_protected(pslot,
2059 &mapping->tree_lock) != head)
2060 goto fail;
2061 }
2062
Joonsoo Kim0139aa72016-05-19 17:10:49 -07002063 /* Prevent deferred_split_scan() touching ->_refcount */
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002064 spin_lock(&pgdata->split_queue_lock);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002065 count = page_count(head);
2066 mapcount = total_mapcount(head);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002067 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002068 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002069 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002070 list_del(page_deferred_list(head));
2071 }
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07002072 if (mapping)
Mel Gorman11fb9982016-07-28 15:46:20 -07002073 __dec_node_page_state(page, NR_SHMEM_THPS);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002074 spin_unlock(&pgdata->split_queue_lock);
2075 __split_huge_page(page, list, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002076 ret = 0;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002077 } else {
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002078 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2079 pr_alert("total_mapcount: %u, page_count(): %u\n",
2080 mapcount, count);
2081 if (PageTail(page))
2082 dump_page(head, NULL);
2083 dump_page(page, "total_mapcount(head) > 0");
2084 BUG();
2085 }
2086 spin_unlock(&pgdata->split_queue_lock);
2087fail: if (mapping)
2088 spin_unlock(&mapping->tree_lock);
Mel Gormana52633d2016-07-28 15:45:28 -07002089 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002090 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002091 ret = -EBUSY;
2092 }
2093
2094out_unlock:
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002095 if (anon_vma) {
2096 anon_vma_unlock_write(anon_vma);
2097 put_anon_vma(anon_vma);
2098 }
2099 if (mapping)
2100 i_mmap_unlock_read(mapping);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002101out:
2102 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2103 return ret;
2104}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002105
2106void free_transhuge_page(struct page *page)
2107{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002108 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002109 unsigned long flags;
2110
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002111 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002112 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002113 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002114 list_del(page_deferred_list(page));
2115 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002116 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002117 free_compound_page(page);
2118}
2119
2120void deferred_split_huge_page(struct page *page)
2121{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002122 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002123 unsigned long flags;
2124
2125 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2126
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002127 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002128 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07002129 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002130 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2131 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002132 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002133 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002134}
2135
2136static unsigned long deferred_split_count(struct shrinker *shrink,
2137 struct shrink_control *sc)
2138{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002139 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002140 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002141}
2142
2143static unsigned long deferred_split_scan(struct shrinker *shrink,
2144 struct shrink_control *sc)
2145{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002146 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002147 unsigned long flags;
2148 LIST_HEAD(list), *pos, *next;
2149 struct page *page;
2150 int split = 0;
2151
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002152 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002153 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08002154 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002155 page = list_entry((void *)pos, struct page, mapping);
2156 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002157 if (get_page_unless_zero(page)) {
2158 list_move(page_deferred_list(page), &list);
2159 } else {
2160 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002161 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002162 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002163 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08002164 if (!--sc->nr_to_scan)
2165 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002166 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002167 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002168
2169 list_for_each_safe(pos, next, &list) {
2170 page = list_entry((void *)pos, struct page, mapping);
2171 lock_page(page);
2172 /* split_huge_page() removes page from list on success */
2173 if (!split_huge_page(page))
2174 split++;
2175 unlock_page(page);
2176 put_page(page);
2177 }
2178
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002179 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2180 list_splice_tail(&list, &pgdata->split_queue);
2181 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002182
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08002183 /*
2184 * Stop shrinker if we didn't split any page, but the queue is empty.
2185 * This can happen if pages were freed under us.
2186 */
2187 if (!split && list_empty(&pgdata->split_queue))
2188 return SHRINK_STOP;
2189 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002190}
2191
2192static struct shrinker deferred_split_shrinker = {
2193 .count_objects = deferred_split_count,
2194 .scan_objects = deferred_split_scan,
2195 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08002196 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002197};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002198
2199#ifdef CONFIG_DEBUG_FS
2200static int split_huge_pages_set(void *data, u64 val)
2201{
2202 struct zone *zone;
2203 struct page *page;
2204 unsigned long pfn, max_zone_pfn;
2205 unsigned long total = 0, split = 0;
2206
2207 if (val != 1)
2208 return -EINVAL;
2209
2210 for_each_populated_zone(zone) {
2211 max_zone_pfn = zone_end_pfn(zone);
2212 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2213 if (!pfn_valid(pfn))
2214 continue;
2215
2216 page = pfn_to_page(pfn);
2217 if (!get_page_unless_zero(page))
2218 continue;
2219
2220 if (zone != page_zone(page))
2221 goto next;
2222
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -07002223 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002224 goto next;
2225
2226 total++;
2227 lock_page(page);
2228 if (!split_huge_page(page))
2229 split++;
2230 unlock_page(page);
2231next:
2232 put_page(page);
2233 }
2234 }
2235
Yang Shi145bdaa2016-05-05 16:22:00 -07002236 pr_info("%lu of %lu THP split\n", split, total);
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002237
2238 return 0;
2239}
2240DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2241 "%llu\n");
2242
2243static int __init split_huge_pages_debugfs(void)
2244{
2245 void *ret;
2246
Yang Shi145bdaa2016-05-05 16:22:00 -07002247 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08002248 &split_huge_pages_fops);
2249 if (!ret)
2250 pr_warn("Failed to create split_huge_pages in debugfs");
2251 return 0;
2252}
2253late_initcall(split_huge_pages_debugfs);
2254#endif