blob: 86f9f8b82f8ecfc47b92554e3a82501346b0fcb8 [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/kthread.h>
22#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080023#include <linux/freezer.h>
Dan Williamsf25748e32016-01-15 16:56:43 -080024#include <linux/pfn_t.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080025#include <linux/mman.h>
Dan Williams3565fce2016-01-15 16:56:55 -080026#include <linux/memremap.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020027#include <linux/pagemap.h>
Kirill A. Shutemov49071d42016-01-15 16:54:40 -080028#include <linux/debugfs.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000029#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080030#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070031#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070032#include <linux/page_idle.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
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -080038enum scan_result {
39 SCAN_FAIL,
40 SCAN_SUCCEED,
41 SCAN_PMD_NULL,
42 SCAN_EXCEED_NONE_PTE,
43 SCAN_PTE_NON_PRESENT,
44 SCAN_PAGE_RO,
45 SCAN_NO_REFERENCED_PAGE,
46 SCAN_PAGE_NULL,
47 SCAN_SCAN_ABORT,
48 SCAN_PAGE_COUNT,
49 SCAN_PAGE_LRU,
50 SCAN_PAGE_LOCK,
51 SCAN_PAGE_ANON,
Kirill A. Shutemovb1caa952016-01-15 16:52:39 -080052 SCAN_PAGE_COMPOUND,
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -080053 SCAN_ANY_PROCESS,
54 SCAN_VMA_NULL,
55 SCAN_VMA_CHECK,
56 SCAN_ADDRESS_RANGE,
57 SCAN_SWAP_CACHE_PAGE,
58 SCAN_DEL_PAGE_LRU,
59 SCAN_ALLOC_HUGE_PAGE_FAIL,
60 SCAN_CGROUP_CHARGE_FAIL
61};
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/huge_memory.h>
65
Andrea Arcangeliba761492011-01-13 15:46:58 -080066/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080067 * By default transparent hugepage support is disabled in order that avoid
68 * to risk increase the memory footprint of applications without a guaranteed
69 * benefit. When transparent hugepage support is enabled, is for all mappings,
70 * and khugepaged scans all mappings.
71 * Defrag is invoked by khugepaged hugepage allocations and by page faults
72 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080073 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080074unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080075#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080076 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080077#endif
78#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
79 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
80#endif
Mel Gorman444eb2a42016-03-17 14:19:23 -070081 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080082 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
83 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080084
85/* default scan 8*512 pte (or vmas) every 30 second */
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +053086static unsigned int khugepaged_pages_to_scan __read_mostly;
Andrea Arcangeliba761492011-01-13 15:46:58 -080087static unsigned int khugepaged_pages_collapsed;
88static unsigned int khugepaged_full_scans;
89static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
90/* during fragmentation poll the hugepage allocator once every minute */
91static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
92static struct task_struct *khugepaged_thread __read_mostly;
93static DEFINE_MUTEX(khugepaged_mutex);
94static DEFINE_SPINLOCK(khugepaged_mm_lock);
95static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
96/*
97 * default collapse hugepages if there is at least one pte mapped like
98 * it would have happened if the vma was large enough during page
99 * fault.
100 */
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530101static unsigned int khugepaged_max_ptes_none __read_mostly;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800102
103static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800104static int khugepaged_slab_init(void);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700105static void khugepaged_slab_exit(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800106
Sasha Levin43b5fbb2013-02-22 16:32:27 -0800107#define MM_SLOTS_HASH_BITS 10
108static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
109
Andrea Arcangeliba761492011-01-13 15:46:58 -0800110static struct kmem_cache *mm_slot_cache __read_mostly;
111
112/**
113 * struct mm_slot - hash lookup from mm to mm_slot
114 * @hash: hash collision list
115 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
116 * @mm: the mm that this information is valid for
117 */
118struct mm_slot {
119 struct hlist_node hash;
120 struct list_head mm_node;
121 struct mm_struct *mm;
122};
123
124/**
125 * struct khugepaged_scan - cursor for scanning
126 * @mm_head: the head of the mm list to scan
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
129 *
130 * There is only the one khugepaged_scan instance of this cursor structure.
131 */
132struct khugepaged_scan {
133 struct list_head mm_head;
134 struct mm_slot *mm_slot;
135 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -0700136};
137static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800138 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
139};
140
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800141static struct shrinker deferred_split_shrinker;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800142
Nicholas Krause2c0b80d2015-09-08 15:00:33 -0700143static void set_recommended_min_free_kbytes(void)
Andrea Arcangelif0005652011-01-13 15:47:04 -0800144{
145 struct zone *zone;
146 int nr_zones = 0;
147 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800148
Andrea Arcangelif0005652011-01-13 15:47:04 -0800149 for_each_populated_zone(zone)
150 nr_zones++;
151
Mel Gorman974a7862015-11-06 16:28:34 -0800152 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
Andrea Arcangelif0005652011-01-13 15:47:04 -0800153 recommended_min = pageblock_nr_pages * nr_zones * 2;
154
155 /*
156 * Make sure that on average at least two pageblocks are almost free
157 * of another type, one for a migratetype to fall back to and a
158 * second to avoid subsequent fallbacks of other types There are 3
159 * MIGRATE_TYPES we care about.
160 */
161 recommended_min += pageblock_nr_pages * nr_zones *
162 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
163
164 /* don't ever allow to reserve more than 5% of the lowmem */
165 recommended_min = min(recommended_min,
166 (unsigned long) nr_free_buffer_pages() / 20);
167 recommended_min <<= (PAGE_SHIFT-10);
168
Han Pingtian42aa83c2014-01-23 15:53:28 -0800169 if (recommended_min > min_free_kbytes) {
170 if (user_min_free_kbytes >= 0)
Joe Perches756a025f02016-03-17 14:19:47 -0700171 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
Han Pingtian42aa83c2014-01-23 15:53:28 -0800172 min_free_kbytes, recommended_min);
173
Andrea Arcangelif0005652011-01-13 15:47:04 -0800174 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800175 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800176 setup_per_zone_wmarks();
Andrea Arcangelif0005652011-01-13 15:47:04 -0800177}
Andrea Arcangelif0005652011-01-13 15:47:04 -0800178
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700179static int start_stop_khugepaged(void)
Andrea Arcangeliba761492011-01-13 15:46:58 -0800180{
181 int err = 0;
182 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800183 if (!khugepaged_thread)
184 khugepaged_thread = kthread_run(khugepaged, NULL,
185 "khugepaged");
Viresh Kumar18e8e5c2015-08-12 15:59:46 +0530186 if (IS_ERR(khugepaged_thread)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700187 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
Andrea Arcangeliba761492011-01-13 15:46:58 -0800188 err = PTR_ERR(khugepaged_thread);
189 khugepaged_thread = NULL;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700190 goto fail;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800191 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700192
193 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800194 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800195
196 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700197 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700198 kthread_stop(khugepaged_thread);
199 khugepaged_thread = NULL;
200 }
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700201fail:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800202 return err;
203}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800204
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800205static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -0800206struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800207
Matthew Wilcoxfc437042015-09-08 14:58:51 -0700208struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800209{
210 struct page *zero_page;
211retry:
212 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -0700213 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800214
215 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
216 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800217 if (!zero_page) {
218 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700219 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800220 }
221 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800222 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700223 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800224 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700225 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800226 goto retry;
227 }
228
229 /* We take additional reference here. It will be put back by shrinker */
230 atomic_set(&huge_zero_refcount, 2);
231 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -0700232 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800233}
234
235static void put_huge_zero_page(void)
236{
237 /*
238 * Counter should never go to zero here. Only shrinker can put
239 * last reference.
240 */
241 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
242}
243
Glauber Costa48896462013-08-28 10:18:15 +1000244static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
245 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800246{
Glauber Costa48896462013-08-28 10:18:15 +1000247 /* we can free zero page only if last reference remains */
248 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
249}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800250
Glauber Costa48896462013-08-28 10:18:15 +1000251static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
252 struct shrink_control *sc)
253{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800254 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700255 struct page *zero_page = xchg(&huge_zero_page, NULL);
256 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700257 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000258 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800259 }
260
261 return 0;
262}
263
264static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000265 .count_objects = shrink_huge_zero_page_count,
266 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800267 .seeks = DEFAULT_SEEKS,
268};
269
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800270#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800271
Mel Gorman444eb2a42016-03-17 14:19:23 -0700272static ssize_t triple_flag_store(struct kobject *kobj,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800273 struct kobj_attribute *attr,
274 const char *buf, size_t count,
275 enum transparent_hugepage_flag enabled,
Mel Gorman444eb2a42016-03-17 14:19:23 -0700276 enum transparent_hugepage_flag deferred,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800277 enum transparent_hugepage_flag req_madv)
278{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700279 if (!memcmp("defer", buf,
280 min(sizeof("defer")-1, count))) {
281 if (enabled == deferred)
282 return -EINVAL;
283 clear_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800284 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700285 set_bit(deferred, &transparent_hugepage_flags);
286 } else if (!memcmp("always", buf,
287 min(sizeof("always")-1, count))) {
288 clear_bit(deferred, &transparent_hugepage_flags);
289 clear_bit(req_madv, &transparent_hugepage_flags);
290 set_bit(enabled, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800291 } else if (!memcmp("madvise", buf,
292 min(sizeof("madvise")-1, count))) {
293 clear_bit(enabled, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700294 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800295 set_bit(req_madv, &transparent_hugepage_flags);
296 } else if (!memcmp("never", buf,
297 min(sizeof("never")-1, count))) {
298 clear_bit(enabled, &transparent_hugepage_flags);
299 clear_bit(req_madv, &transparent_hugepage_flags);
Mel Gorman444eb2a42016-03-17 14:19:23 -0700300 clear_bit(deferred, &transparent_hugepage_flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800301 } else
302 return -EINVAL;
303
304 return count;
305}
306
307static ssize_t enabled_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf)
309{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700310 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
311 return sprintf(buf, "[always] madvise never\n");
312 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
313 return sprintf(buf, "always [madvise] never\n");
314 else
315 return sprintf(buf, "always madvise [never]\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800316}
Mel Gorman444eb2a42016-03-17 14:19:23 -0700317
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800318static ssize_t enabled_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count)
321{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800322 ssize_t ret;
323
Mel Gorman444eb2a42016-03-17 14:19:23 -0700324 ret = triple_flag_store(kobj, attr, buf, count,
325 TRANSPARENT_HUGEPAGE_FLAG,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800326 TRANSPARENT_HUGEPAGE_FLAG,
327 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
328
329 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700330 int err;
331
332 mutex_lock(&khugepaged_mutex);
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700333 err = start_stop_khugepaged();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700334 mutex_unlock(&khugepaged_mutex);
335
Andrea Arcangeliba761492011-01-13 15:46:58 -0800336 if (err)
337 ret = err;
338 }
339
340 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800341}
342static struct kobj_attribute enabled_attr =
343 __ATTR(enabled, 0644, enabled_show, enabled_store);
344
345static ssize_t single_flag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf,
347 enum transparent_hugepage_flag flag)
348{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700349 return sprintf(buf, "%d\n",
350 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800351}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700352
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800353static ssize_t single_flag_store(struct kobject *kobj,
354 struct kobj_attribute *attr,
355 const char *buf, size_t count,
356 enum transparent_hugepage_flag flag)
357{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700358 unsigned long value;
359 int ret;
360
361 ret = kstrtoul(buf, 10, &value);
362 if (ret < 0)
363 return ret;
364 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800365 return -EINVAL;
366
Ben Hutchingse27e6152011-04-14 15:22:21 -0700367 if (value)
368 set_bit(flag, &transparent_hugepage_flags);
369 else
370 clear_bit(flag, &transparent_hugepage_flags);
371
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800372 return count;
373}
374
375/*
376 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
377 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
378 * memory just to allocate one more hugepage.
379 */
380static ssize_t defrag_show(struct kobject *kobj,
381 struct kobj_attribute *attr, char *buf)
382{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700383 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
384 return sprintf(buf, "[always] defer madvise never\n");
385 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
386 return sprintf(buf, "always [defer] madvise never\n");
387 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
388 return sprintf(buf, "always defer [madvise] never\n");
389 else
390 return sprintf(buf, "always defer madvise [never]\n");
391
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800392}
393static ssize_t defrag_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700397 return triple_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
399 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800400 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
401}
402static struct kobj_attribute defrag_attr =
403 __ATTR(defrag, 0644, defrag_show, defrag_store);
404
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800405static ssize_t use_zero_page_show(struct kobject *kobj,
406 struct kobj_attribute *attr, char *buf)
407{
408 return single_flag_show(kobj, attr, buf,
409 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
410}
411static ssize_t use_zero_page_store(struct kobject *kobj,
412 struct kobj_attribute *attr, const char *buf, size_t count)
413{
414 return single_flag_store(kobj, attr, buf, count,
415 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
416}
417static struct kobj_attribute use_zero_page_attr =
418 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800419#ifdef CONFIG_DEBUG_VM
420static ssize_t debug_cow_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
422{
423 return single_flag_show(kobj, attr, buf,
424 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
425}
426static ssize_t debug_cow_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
429{
430 return single_flag_store(kobj, attr, buf, count,
431 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
432}
433static struct kobj_attribute debug_cow_attr =
434 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
435#endif /* CONFIG_DEBUG_VM */
436
437static struct attribute *hugepage_attr[] = {
438 &enabled_attr.attr,
439 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800440 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800441#ifdef CONFIG_DEBUG_VM
442 &debug_cow_attr.attr,
443#endif
444 NULL,
445};
446
447static struct attribute_group hugepage_attr_group = {
448 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800449};
450
451static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
452 struct kobj_attribute *attr,
453 char *buf)
454{
455 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
456}
457
458static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
459 struct kobj_attribute *attr,
460 const char *buf, size_t count)
461{
462 unsigned long msecs;
463 int err;
464
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700465 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800466 if (err || msecs > UINT_MAX)
467 return -EINVAL;
468
469 khugepaged_scan_sleep_millisecs = msecs;
470 wake_up_interruptible(&khugepaged_wait);
471
472 return count;
473}
474static struct kobj_attribute scan_sleep_millisecs_attr =
475 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
476 scan_sleep_millisecs_store);
477
478static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 char *buf)
481{
482 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
483}
484
485static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
486 struct kobj_attribute *attr,
487 const char *buf, size_t count)
488{
489 unsigned long msecs;
490 int err;
491
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700492 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800493 if (err || msecs > UINT_MAX)
494 return -EINVAL;
495
496 khugepaged_alloc_sleep_millisecs = msecs;
497 wake_up_interruptible(&khugepaged_wait);
498
499 return count;
500}
501static struct kobj_attribute alloc_sleep_millisecs_attr =
502 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
503 alloc_sleep_millisecs_store);
504
505static ssize_t pages_to_scan_show(struct kobject *kobj,
506 struct kobj_attribute *attr,
507 char *buf)
508{
509 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
510}
511static ssize_t pages_to_scan_store(struct kobject *kobj,
512 struct kobj_attribute *attr,
513 const char *buf, size_t count)
514{
515 int err;
516 unsigned long pages;
517
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700518 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800519 if (err || !pages || pages > UINT_MAX)
520 return -EINVAL;
521
522 khugepaged_pages_to_scan = pages;
523
524 return count;
525}
526static struct kobj_attribute pages_to_scan_attr =
527 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
528 pages_to_scan_store);
529
530static ssize_t pages_collapsed_show(struct kobject *kobj,
531 struct kobj_attribute *attr,
532 char *buf)
533{
534 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
535}
536static struct kobj_attribute pages_collapsed_attr =
537 __ATTR_RO(pages_collapsed);
538
539static ssize_t full_scans_show(struct kobject *kobj,
540 struct kobj_attribute *attr,
541 char *buf)
542{
543 return sprintf(buf, "%u\n", khugepaged_full_scans);
544}
545static struct kobj_attribute full_scans_attr =
546 __ATTR_RO(full_scans);
547
548static ssize_t khugepaged_defrag_show(struct kobject *kobj,
549 struct kobj_attribute *attr, char *buf)
550{
551 return single_flag_show(kobj, attr, buf,
552 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
553}
554static ssize_t khugepaged_defrag_store(struct kobject *kobj,
555 struct kobj_attribute *attr,
556 const char *buf, size_t count)
557{
558 return single_flag_store(kobj, attr, buf, count,
559 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
560}
561static struct kobj_attribute khugepaged_defrag_attr =
562 __ATTR(defrag, 0644, khugepaged_defrag_show,
563 khugepaged_defrag_store);
564
565/*
566 * max_ptes_none controls if khugepaged should collapse hugepages over
567 * any unmapped ptes in turn potentially increasing the memory
568 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
569 * reduce the available free memory in the system as it
570 * runs. Increasing max_ptes_none will instead potentially reduce the
571 * free memory in the system during the khugepaged scan.
572 */
573static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
574 struct kobj_attribute *attr,
575 char *buf)
576{
577 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
578}
579static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
580 struct kobj_attribute *attr,
581 const char *buf, size_t count)
582{
583 int err;
584 unsigned long max_ptes_none;
585
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700586 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800587 if (err || max_ptes_none > HPAGE_PMD_NR-1)
588 return -EINVAL;
589
590 khugepaged_max_ptes_none = max_ptes_none;
591
592 return count;
593}
594static struct kobj_attribute khugepaged_max_ptes_none_attr =
595 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
596 khugepaged_max_ptes_none_store);
597
598static struct attribute *khugepaged_attr[] = {
599 &khugepaged_defrag_attr.attr,
600 &khugepaged_max_ptes_none_attr.attr,
601 &pages_to_scan_attr.attr,
602 &pages_collapsed_attr.attr,
603 &full_scans_attr.attr,
604 &scan_sleep_millisecs_attr.attr,
605 &alloc_sleep_millisecs_attr.attr,
606 NULL,
607};
608
609static struct attribute_group khugepaged_attr_group = {
610 .attrs = khugepaged_attr,
611 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800612};
Shaohua Li569e5592012-01-12 17:19:11 -0800613
614static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
615{
616 int err;
617
618 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
619 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700620 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800621 return -ENOMEM;
622 }
623
624 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
625 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700626 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800627 goto delete_obj;
628 }
629
630 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
631 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700632 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800633 goto remove_hp_group;
634 }
635
636 return 0;
637
638remove_hp_group:
639 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
640delete_obj:
641 kobject_put(*hugepage_kobj);
642 return err;
643}
644
645static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
646{
647 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
648 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
649 kobject_put(hugepage_kobj);
650}
651#else
652static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
653{
654 return 0;
655}
656
657static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
658{
659}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800660#endif /* CONFIG_SYSFS */
661
662static int __init hugepage_init(void)
663{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800664 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800665 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800666
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800667 if (!has_transparent_hugepage()) {
668 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800669 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800670 }
671
Kirill A. Shutemovff20c2e2016-03-01 09:45:14 +0530672 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
673 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
674 /*
675 * hugepages can't be allocated by the buddy allocator
676 */
677 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
678 /*
679 * we use page->mapping and page->index in second tail page
680 * as list_head: assuming THP order >= 2
681 */
682 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
683
Shaohua Li569e5592012-01-12 17:19:11 -0800684 err = hugepage_init_sysfs(&hugepage_kobj);
685 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700686 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800687
688 err = khugepaged_slab_init();
689 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700690 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800691
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700692 err = register_shrinker(&huge_zero_page_shrinker);
693 if (err)
694 goto err_hzp_shrinker;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800695 err = register_shrinker(&deferred_split_shrinker);
696 if (err)
697 goto err_split_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800698
Rik van Riel97562cd2011-01-13 15:47:12 -0800699 /*
700 * By default disable transparent hugepages on smaller systems,
701 * where the extra memory used could hurt more than TLB overhead
702 * is likely to save. The admin can still enable it through /sys.
703 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700704 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800705 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700706 return 0;
707 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800708
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700709 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700710 if (err)
711 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800712
Shaohua Li569e5592012-01-12 17:19:11 -0800713 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700714err_khugepaged:
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800715 unregister_shrinker(&deferred_split_shrinker);
716err_split_shrinker:
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700717 unregister_shrinker(&huge_zero_page_shrinker);
718err_hzp_shrinker:
719 khugepaged_slab_exit();
720err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800721 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700722err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800723 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800724}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800725subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800726
727static int __init setup_transparent_hugepage(char *str)
728{
729 int ret = 0;
730 if (!str)
731 goto out;
732 if (!strcmp(str, "always")) {
733 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
734 &transparent_hugepage_flags);
735 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
736 &transparent_hugepage_flags);
737 ret = 1;
738 } else if (!strcmp(str, "madvise")) {
739 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
740 &transparent_hugepage_flags);
741 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
742 &transparent_hugepage_flags);
743 ret = 1;
744 } else if (!strcmp(str, "never")) {
745 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
746 &transparent_hugepage_flags);
747 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
748 &transparent_hugepage_flags);
749 ret = 1;
750 }
751out:
752 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700753 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800754 return ret;
755}
756__setup("transparent_hugepage=", setup_transparent_hugepage);
757
Mel Gormanb32967f2012-11-19 12:35:47 +0000758pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800759{
760 if (likely(vma->vm_flags & VM_WRITE))
761 pmd = pmd_mkwrite(pmd);
762 return pmd;
763}
764
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700765static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800766{
767 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700768 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800769 entry = pmd_mkhuge(entry);
770 return entry;
771}
772
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800773static inline struct list_head *page_deferred_list(struct page *page)
774{
775 /*
776 * ->lru in the tail pages is occupied by compound_head.
777 * Let's use ->mapping + ->index in the second tail page as list_head.
778 */
779 return (struct list_head *)&page[2].mapping;
780}
781
782void prep_transhuge_page(struct page *page)
783{
784 /*
785 * we use page->mapping and page->indexlru in second tail page
786 * as list_head: assuming THP order >= 2
787 */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800788
789 INIT_LIST_HEAD(page_deferred_list(page));
790 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
791}
792
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800793static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
794 struct vm_area_struct *vma,
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700795 unsigned long address, pmd_t *pmd,
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700796 struct page *page, gfp_t gfp,
797 unsigned int flags)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800798{
Johannes Weiner00501b52014-08-08 14:19:20 -0700799 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800800 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800801 spinlock_t *ptl;
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700802 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800803
Sasha Levin309381fea2014-01-23 15:52:54 -0800804 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700805
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800806 if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700807 put_page(page);
808 count_vm_event(THP_FAULT_FALLBACK);
809 return VM_FAULT_FALLBACK;
810 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800811
Johannes Weiner00501b52014-08-08 14:19:20 -0700812 pgtable = pte_alloc_one(mm, haddr);
813 if (unlikely(!pgtable)) {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800814 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700815 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700816 return VM_FAULT_OOM;
817 }
818
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800819 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700820 /*
821 * The memory barrier inside __SetPageUptodate makes sure that
822 * clear_huge_page writes become visible before the set_pmd_at()
823 * write.
824 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800825 __SetPageUptodate(page);
826
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800827 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800828 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800829 spin_unlock(ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800830 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800831 put_page(page);
832 pte_free(mm, pgtable);
833 } else {
834 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700835
836 /* Deliver the page fault to userland */
837 if (userfaultfd_missing(vma)) {
838 int ret;
839
840 spin_unlock(ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800841 mem_cgroup_cancel_charge(page, memcg, true);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700842 put_page(page);
843 pte_free(mm, pgtable);
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700844 ret = handle_userfault(vma, address, flags,
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700845 VM_UFFD_MISSING);
846 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
847 return ret;
848 }
849
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700850 entry = mk_huge_pmd(page, vma->vm_page_prot);
851 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800852 page_add_new_anon_rmap(page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -0800853 mem_cgroup_commit_charge(page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -0700854 lru_cache_add_active_or_unevictable(page, vma);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700855 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800856 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800857 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800858 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800859 spin_unlock(ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700860 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800861 }
862
David Rientjesaa2e8782012-05-29 15:06:17 -0700863 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800864}
865
Mel Gorman444eb2a42016-03-17 14:19:23 -0700866/*
867 * If THP is set to always then directly reclaim/compact as necessary
868 * If set to defer then do no reclaim and defer to khugepaged
869 * If set to madvise and the VMA is flagged then directly reclaim/compact
870 */
871static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800872{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700873 gfp_t reclaim_flags = 0;
874
875 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) &&
876 (vma->vm_flags & VM_HUGEPAGE))
877 reclaim_flags = __GFP_DIRECT_RECLAIM;
878 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
879 reclaim_flags = __GFP_KSWAPD_RECLAIM;
880 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
881 reclaim_flags = __GFP_DIRECT_RECLAIM;
882
883 return GFP_TRANSHUGE | reclaim_flags;
884}
885
886/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
887static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
888{
889 return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800890}
891
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800892/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700893static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800894 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700895 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800896{
897 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700898 if (!pmd_none(*pmd))
899 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700900 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800901 entry = pmd_mkhuge(entry);
Matthew Wilcox12c9d702016-02-02 16:57:57 -0800902 if (pgtable)
903 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800904 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800905 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700906 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800907}
908
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800909int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
910 unsigned long address, pmd_t *pmd,
911 unsigned int flags)
912{
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800913 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800914 struct page *page;
915 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800916
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700917 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700918 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700919 if (unlikely(anon_vma_prepare(vma)))
920 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700921 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700922 return VM_FAULT_OOM;
Dominik Dingel593befa2014-10-23 12:07:44 +0200923 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700924 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800925 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700926 pgtable_t pgtable;
927 struct page *zero_page;
928 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700929 int ret;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700930 pgtable = pte_alloc_one(mm, haddr);
931 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800932 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700933 zero_page = get_huge_zero_page();
934 if (unlikely(!zero_page)) {
935 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700936 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700937 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700938 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800939 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700940 ret = 0;
941 set = false;
942 if (pmd_none(*pmd)) {
943 if (userfaultfd_missing(vma)) {
944 spin_unlock(ptl);
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700945 ret = handle_userfault(vma, address, flags,
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700946 VM_UFFD_MISSING);
947 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
948 } else {
949 set_huge_zero_page(pgtable, mm, vma,
950 haddr, pmd,
951 zero_page);
952 spin_unlock(ptl);
953 set = true;
954 }
955 } else
956 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700957 if (!set) {
958 pte_free(mm, pgtable);
959 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800960 }
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700961 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800962 }
Mel Gorman444eb2a42016-03-17 14:19:23 -0700963 gfp = alloc_hugepage_direct_gfpmask(vma);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800964 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700965 if (unlikely(!page)) {
966 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700967 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700968 }
Kirill A. Shutemov9a982252016-01-15 16:54:17 -0800969 prep_transhuge_page(page);
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700970 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
971 flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800972}
973
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700974static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800975 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700976{
977 struct mm_struct *mm = vma->vm_mm;
978 pmd_t entry;
979 spinlock_t *ptl;
980
981 ptl = pmd_lock(mm, pmd);
Dan Williamsf25748e32016-01-15 16:56:43 -0800982 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
983 if (pfn_t_devmap(pfn))
984 entry = pmd_mkdevmap(entry);
Ross Zwisler01871e52016-01-15 16:56:02 -0800985 if (write) {
986 entry = pmd_mkyoung(pmd_mkdirty(entry));
987 entry = maybe_pmd_mkwrite(entry, vma);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700988 }
Ross Zwisler01871e52016-01-15 16:56:02 -0800989 set_pmd_at(mm, addr, pmd, entry);
990 update_mmu_cache_pmd(vma, addr, pmd);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700991 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700992}
993
994int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Dan Williamsf25748e32016-01-15 16:56:43 -0800995 pmd_t *pmd, pfn_t pfn, bool write)
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700996{
997 pgprot_t pgprot = vma->vm_page_prot;
998 /*
999 * If we had pmd_special, we could avoid all these restrictions,
1000 * but we need to be consistent with PTEs and architectures that
1001 * can't support a 'special' bit.
1002 */
1003 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1004 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005 (VM_PFNMAP|VM_MIXEDMAP));
1006 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
Dan Williamsf25748e32016-01-15 16:56:43 -08001007 BUG_ON(!pfn_t_devmap(pfn));
Matthew Wilcox5cad4652015-09-08 14:58:54 -07001008
1009 if (addr < vma->vm_start || addr >= vma->vm_end)
1010 return VM_FAULT_SIGBUS;
1011 if (track_pfn_insert(vma, &pgprot, pfn))
1012 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -07001013 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
1014 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -07001015}
1016
Dan Williams3565fce2016-01-15 16:56:55 -08001017static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1018 pmd_t *pmd)
1019{
1020 pmd_t _pmd;
1021
1022 /*
1023 * We should set the dirty bit only for FOLL_WRITE but for now
1024 * the dirty bit in the pmd is meaningless. And if the dirty
1025 * bit will become meaningful and we'll only set it with
1026 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1027 * set the young bit, instead of the current set_pmd_at.
1028 */
1029 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1030 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031 pmd, _pmd, 1))
1032 update_mmu_cache_pmd(vma, addr, pmd);
1033}
1034
1035struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036 pmd_t *pmd, int flags)
1037{
1038 unsigned long pfn = pmd_pfn(*pmd);
1039 struct mm_struct *mm = vma->vm_mm;
1040 struct dev_pagemap *pgmap;
1041 struct page *page;
1042
1043 assert_spin_locked(pmd_lockptr(mm, pmd));
1044
1045 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1046 return NULL;
1047
1048 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1049 /* pass */;
1050 else
1051 return NULL;
1052
1053 if (flags & FOLL_TOUCH)
1054 touch_pmd(vma, addr, pmd);
1055
1056 /*
1057 * device mapped pages can only be returned if the
1058 * caller will manage the page reference count.
1059 */
1060 if (!(flags & FOLL_GET))
1061 return ERR_PTR(-EEXIST);
1062
1063 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1064 pgmap = get_dev_pagemap(pfn, NULL);
1065 if (!pgmap)
1066 return ERR_PTR(-EFAULT);
1067 page = pfn_to_page(pfn);
1068 get_page(page);
1069 put_dev_pagemap(pgmap);
1070
1071 return page;
1072}
1073
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001074int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1075 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1076 struct vm_area_struct *vma)
1077{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001078 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001079 struct page *src_page;
1080 pmd_t pmd;
Matthew Wilcox12c9d702016-02-02 16:57:57 -08001081 pgtable_t pgtable = NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001082 int ret;
1083
Matthew Wilcox12c9d702016-02-02 16:57:57 -08001084 if (!vma_is_dax(vma)) {
1085 ret = -ENOMEM;
1086 pgtable = pte_alloc_one(dst_mm, addr);
1087 if (unlikely(!pgtable))
1088 goto out;
1089 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001090
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001091 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1092 src_ptl = pmd_lockptr(src_mm, src_pmd);
1093 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001094
1095 ret = -EAGAIN;
1096 pmd = *src_pmd;
Dan Williams5c7fb562016-01-15 16:56:52 -08001097 if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001098 pte_free(dst_mm, pgtable);
1099 goto out_unlock;
1100 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001101 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001102 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001103 * under splitting since we don't split the page itself, only pmd to
1104 * a page table.
1105 */
1106 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001107 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001108 /*
1109 * get_huge_zero_page() will never allocate a new page here,
1110 * since we already have a zero page to copy. It just takes a
1111 * reference.
1112 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001113 zero_page = get_huge_zero_page();
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07001114 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001115 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -08001116 ret = 0;
1117 goto out_unlock;
1118 }
Mel Gormande466bd2013-12-18 17:08:42 -08001119
Matthew Wilcox12c9d702016-02-02 16:57:57 -08001120 if (!vma_is_dax(vma)) {
Dan Williams5c7fb562016-01-15 16:56:52 -08001121 /* thp accounting separate from pmd_devmap accounting */
1122 src_page = pmd_page(pmd);
1123 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1124 get_page(src_page);
1125 page_dup_rmap(src_page, true);
1126 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1127 atomic_long_inc(&dst_mm->nr_ptes);
1128 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1129 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001130
1131 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1132 pmd = pmd_mkold(pmd_wrprotect(pmd));
1133 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134
1135 ret = 0;
1136out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001137 spin_unlock(src_ptl);
1138 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001139out:
1140 return ret;
1141}
1142
Will Deacona1dd4502012-12-11 16:01:27 -08001143void huge_pmd_set_accessed(struct mm_struct *mm,
1144 struct vm_area_struct *vma,
1145 unsigned long address,
1146 pmd_t *pmd, pmd_t orig_pmd,
1147 int dirty)
1148{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001149 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -08001150 pmd_t entry;
1151 unsigned long haddr;
1152
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001153 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -08001154 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1155 goto unlock;
1156
1157 entry = pmd_mkyoung(orig_pmd);
1158 haddr = address & HPAGE_PMD_MASK;
1159 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1160 update_mmu_cache_pmd(vma, address, pmd);
1161
1162unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001163 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001164}
1165
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001166static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1167 struct vm_area_struct *vma,
1168 unsigned long address,
1169 pmd_t *pmd, pmd_t orig_pmd,
1170 struct page *page,
1171 unsigned long haddr)
1172{
Johannes Weiner00501b52014-08-08 14:19:20 -07001173 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001174 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001175 pgtable_t pgtable;
1176 pmd_t _pmd;
1177 int ret = 0, i;
1178 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001179 unsigned long mmun_start; /* For mmu_notifiers */
1180 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001181
1182 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1183 GFP_KERNEL);
1184 if (unlikely(!pages)) {
1185 ret |= VM_FAULT_OOM;
1186 goto out;
1187 }
1188
1189 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001190 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1191 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001192 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001193 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -07001194 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001195 &memcg, false))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001196 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001197 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001198 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001201 mem_cgroup_cancel_charge(pages[i], memcg,
1202 false);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001203 put_page(pages[i]);
1204 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205 kfree(pages);
1206 ret |= VM_FAULT_OOM;
1207 goto out;
1208 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001209 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001210 }
1211
1212 for (i = 0; i < HPAGE_PMD_NR; i++) {
1213 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001214 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001215 __SetPageUptodate(pages[i]);
1216 cond_resched();
1217 }
1218
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001219 mmun_start = haddr;
1220 mmun_end = haddr + HPAGE_PMD_SIZE;
1221 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1222
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001223 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001224 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1225 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001226 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001227
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001228 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001229 /* leave pmd empty until pte is filled */
1230
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001231 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001232 pmd_populate(mm, &_pmd, pgtable);
1233
1234 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1235 pte_t *pte, entry;
1236 entry = mk_pte(pages[i], vma->vm_page_prot);
1237 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001238 memcg = (void *)page_private(pages[i]);
1239 set_page_private(pages[i], 0);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001240 page_add_new_anon_rmap(pages[i], vma, haddr, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001241 mem_cgroup_commit_charge(pages[i], memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07001242 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001243 pte = pte_offset_map(&_pmd, haddr);
1244 VM_BUG_ON(!pte_none(*pte));
1245 set_pte_at(mm, haddr, pte, entry);
1246 pte_unmap(pte);
1247 }
1248 kfree(pages);
1249
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001250 smp_wmb(); /* make pte visible before pmd */
1251 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001252 page_remove_rmap(page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001253 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001254
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001255 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1256
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001257 ret |= VM_FAULT_WRITE;
1258 put_page(page);
1259
1260out:
1261 return ret;
1262
1263out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001264 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001265 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001266 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001267 memcg = (void *)page_private(pages[i]);
1268 set_page_private(pages[i], 0);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001269 mem_cgroup_cancel_charge(pages[i], memcg, false);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001270 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001271 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001272 kfree(pages);
1273 goto out;
1274}
1275
1276int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1277 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1278{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001279 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001280 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001281 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001282 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001283 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001284 unsigned long mmun_start; /* For mmu_notifiers */
1285 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001286 gfp_t huge_gfp; /* for allocation and charge */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001287
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001288 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001289 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001290 haddr = address & HPAGE_PMD_MASK;
1291 if (is_huge_zero_pmd(orig_pmd))
1292 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001293 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001294 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1295 goto out_unlock;
1296
1297 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001298 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Kirill A. Shutemov1f25fe22016-01-15 16:52:24 -08001299 /*
1300 * We can only reuse the page if nobody else maps the huge page or it's
1301 * part. We can do it by checking page_mapcount() on each sub-page, but
1302 * it's expensive.
1303 * The cheaper way is to check page_count() to be equal 1: every
1304 * mapcount takes page reference reference, so this way we can
1305 * guarantee, that the PMD is the only mapping.
1306 * This can give false negative if somebody pinned the page, but that's
1307 * fine.
1308 */
1309 if (page_mapcount(page) == 1 && page_count(page) == 1) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001310 pmd_t entry;
1311 entry = pmd_mkyoung(orig_pmd);
1312 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1313 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001314 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001315 ret |= VM_FAULT_WRITE;
1316 goto out_unlock;
1317 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001318 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001319 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001320alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001321 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001322 !transparent_hugepage_debug_cow()) {
Mel Gorman444eb2a42016-03-17 14:19:23 -07001323 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
Michal Hocko3b363692015-04-15 16:13:29 -07001324 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001325 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001326 new_page = NULL;
1327
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001328 if (likely(new_page)) {
1329 prep_transhuge_page(new_page);
1330 } else {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001331 if (!page) {
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001332 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001333 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001334 } else {
1335 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1336 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001337 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001338 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001339 ret |= VM_FAULT_FALLBACK;
1340 }
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001341 put_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001342 }
David Rientjes17766dd2013-09-12 15:14:06 -07001343 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001344 goto out;
1345 }
1346
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001347 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1348 true))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001349 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001350 if (page) {
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001351 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001352 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001353 } else
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001354 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001355 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001356 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001357 goto out;
1358 }
1359
David Rientjes17766dd2013-09-12 15:14:06 -07001360 count_vm_event(THP_FAULT_ALLOC);
1361
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001362 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001363 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1364 else
1365 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001366 __SetPageUptodate(new_page);
1367
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001368 mmun_start = haddr;
1369 mmun_end = haddr + HPAGE_PMD_SIZE;
1370 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1371
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001372 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001373 if (page)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001374 put_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001375 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001376 spin_unlock(ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001377 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001378 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001379 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001380 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001381 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001382 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1383 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001384 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001385 page_add_new_anon_rmap(new_page, vma, haddr, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08001386 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07001387 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001388 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001389 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001390 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001391 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001392 put_huge_zero_page();
1393 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001394 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001395 page_remove_rmap(page, true);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001396 put_page(page);
1397 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001398 ret |= VM_FAULT_WRITE;
1399 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001400 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001401out_mn:
1402 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1403out:
1404 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001405out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001406 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001407 return ret;
1408}
1409
David Rientjesb676b292012-10-08 16:34:03 -07001410struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001411 unsigned long addr,
1412 pmd_t *pmd,
1413 unsigned int flags)
1414{
David Rientjesb676b292012-10-08 16:34:03 -07001415 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001416 struct page *page = NULL;
1417
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001418 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001419
1420 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1421 goto out;
1422
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001423 /* Avoid dumping huge zero page */
1424 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1425 return ERR_PTR(-EFAULT);
1426
Mel Gorman2b4847e2013-12-18 17:08:32 -08001427 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001428 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001429 goto out;
1430
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001431 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001432 VM_BUG_ON_PAGE(!PageHead(page), page);
Dan Williams3565fce2016-01-15 16:56:55 -08001433 if (flags & FOLL_TOUCH)
1434 touch_pmd(vma, addr, pmd);
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001435 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001436 /*
1437 * We don't mlock() pte-mapped THPs. This way we can avoid
1438 * leaking mlocked pages into non-VM_LOCKED VMAs.
1439 *
1440 * In most cases the pmd is the only mapping of the page as we
1441 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1442 * writable private mappings in populate_vma_page_range().
1443 *
1444 * The only scenario when we have the page shared here is if we
1445 * mlocking read-only mapping shared over fork(). We skip
1446 * mlocking such pages.
1447 */
1448 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1449 page->mapping && trylock_page(page)) {
David Rientjesb676b292012-10-08 16:34:03 -07001450 lru_add_drain();
1451 if (page->mapping)
1452 mlock_vma_page(page);
1453 unlock_page(page);
1454 }
1455 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001456 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001457 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001458 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001459 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001460
1461out:
1462 return page;
1463}
1464
Mel Gormand10e63f2012-10-25 14:16:31 +02001465/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001466int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1467 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001468{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001469 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001470 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001471 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001472 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001473 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001474 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001475 bool page_locked;
1476 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001477 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001478 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001479
Mel Gormanc0e7cad2015-02-12 14:58:41 -08001480 /* A PROT_NONE fault should not end up here */
1481 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1482
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001483 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001484 if (unlikely(!pmd_same(pmd, *pmdp)))
1485 goto out_unlock;
1486
Mel Gormande466bd2013-12-18 17:08:42 -08001487 /*
1488 * If there are potential migrations, wait for completion and retry
1489 * without disrupting NUMA hinting information. Do not relock and
1490 * check_same as the page may no longer be mapped.
1491 */
1492 if (unlikely(pmd_trans_migrating(*pmdp))) {
Mel Gorman5d833062015-02-12 14:58:16 -08001493 page = pmd_page(*pmdp);
Mel Gormande466bd2013-12-18 17:08:42 -08001494 spin_unlock(ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001495 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001496 goto out;
1497 }
1498
Mel Gormand10e63f2012-10-25 14:16:31 +02001499 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001500 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001501 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001502 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001503 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001504 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001505 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001506 flags |= TNF_FAULT_LOCAL;
1507 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001508
Mel Gormanbea66fb2015-03-25 15:55:37 -07001509 /* See similar comment in do_numa_page for explanation */
1510 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001511 flags |= TNF_NO_GROUP;
1512
1513 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001514 * Acquire the page lock to serialise THP migrations but avoid dropping
1515 * page_table_lock if at all possible
1516 */
Mel Gormanb8916632013-10-07 11:28:44 +01001517 page_locked = trylock_page(page);
1518 target_nid = mpol_misplaced(page, vma, haddr);
1519 if (target_nid == -1) {
1520 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001521 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001522 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001523 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001524
Mel Gormande466bd2013-12-18 17:08:42 -08001525 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001526 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001527 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001528 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001529 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001530 goto out;
1531 }
1532
Mel Gorman2b4847e2013-12-18 17:08:32 -08001533 /*
1534 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1535 * to serialises splits
1536 */
Mel Gormanb8916632013-10-07 11:28:44 +01001537 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001538 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001539 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001540
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001541 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001542 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001543 if (unlikely(!pmd_same(pmd, *pmdp))) {
1544 unlock_page(page);
1545 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001546 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001547 goto out_unlock;
1548 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001549
Mel Gormanc3a489c2013-12-18 17:08:38 -08001550 /* Bail if we fail to protect against THP splits for any reason */
1551 if (unlikely(!anon_vma)) {
1552 put_page(page);
1553 page_nid = -1;
1554 goto clear_pmdnuma;
1555 }
1556
Mel Gormana54a4072013-10-07 11:28:46 +01001557 /*
1558 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001559 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001560 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001561 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001562 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001563 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001564 if (migrated) {
1565 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001566 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001567 } else
1568 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001569
Mel Gorman8191acb2013-10-07 11:28:45 +01001570 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001571clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001572 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001573 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001574 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001575 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001576 if (was_writable)
1577 pmd = pmd_mkwrite(pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001578 set_pmd_at(mm, haddr, pmdp, pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001579 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001580 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001581out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001582 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001583
1584out:
1585 if (anon_vma)
1586 page_unlock_anon_vma_read(anon_vma);
1587
Mel Gorman8191acb2013-10-07 11:28:45 +01001588 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001589 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001590
Mel Gormand10e63f2012-10-25 14:16:31 +02001591 return 0;
1592}
1593
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001594int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1595 pmd_t *pmd, unsigned long addr, unsigned long next)
1596
1597{
1598 spinlock_t *ptl;
1599 pmd_t orig_pmd;
1600 struct page *page;
1601 struct mm_struct *mm = tlb->mm;
1602 int ret = 0;
1603
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001604 ptl = pmd_trans_huge_lock(pmd, vma);
1605 if (!ptl)
Linus Torvalds25eedab2016-01-17 18:33:15 -08001606 goto out_unlocked;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08001607
1608 orig_pmd = *pmd;
1609 if (is_huge_zero_pmd(orig_pmd)) {
1610 ret = 1;
1611 goto out;
1612 }
1613
1614 page = pmd_page(orig_pmd);
1615 /*
1616 * If other processes are mapping this page, we couldn't discard
1617 * the page unless they all do MADV_FREE so let's skip the page.
1618 */
1619 if (page_mapcount(page) != 1)
1620 goto out;
1621
1622 if (!trylock_page(page))
1623 goto out;
1624
1625 /*
1626 * If user want to discard part-pages of THP, split it so MADV_FREE
1627 * will deactivate only them.
1628 */
1629 if (next - addr != HPAGE_PMD_SIZE) {
1630 get_page(page);
1631 spin_unlock(ptl);
1632 if (split_huge_page(page)) {
1633 put_page(page);
1634 unlock_page(page);
1635 goto out_unlocked;
1636 }
1637 put_page(page);
1638 unlock_page(page);
1639 ret = 1;
1640 goto out_unlocked;
1641 }
1642
1643 if (PageDirty(page))
1644 ClearPageDirty(page);
1645 unlock_page(page);
1646
1647 if (PageActive(page))
1648 deactivate_page(page);
1649
1650 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1651 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1652 tlb->fullmm);
1653 orig_pmd = pmd_mkold(orig_pmd);
1654 orig_pmd = pmd_mkclean(orig_pmd);
1655
1656 set_pmd_at(mm, addr, pmd, orig_pmd);
1657 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1658 }
1659 ret = 1;
1660out:
1661 spin_unlock(ptl);
1662out_unlocked:
1663 return ret;
1664}
1665
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001666int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001667 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001668{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001669 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001670 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001671
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001672 ptl = __pmd_trans_huge_lock(pmd, vma);
1673 if (!ptl)
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001674 return 0;
1675 /*
1676 * For architectures like ppc64 we look at deposited pgtable
1677 * when calling pmdp_huge_get_and_clear. So do the
1678 * pgtable_trans_huge_withdraw after finishing pmdp related
1679 * operations.
1680 */
1681 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1682 tlb->fullmm);
1683 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1684 if (vma_is_dax(vma)) {
1685 spin_unlock(ptl);
1686 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001687 put_huge_zero_page();
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001688 } else if (is_huge_zero_pmd(orig_pmd)) {
1689 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1690 atomic_long_dec(&tlb->mm->nr_ptes);
1691 spin_unlock(ptl);
1692 put_huge_zero_page();
1693 } else {
1694 struct page *page = pmd_page(orig_pmd);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001695 page_remove_rmap(page, true);
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001696 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1697 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1698 VM_BUG_ON_PAGE(!PageHead(page), page);
1699 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1700 atomic_long_dec(&tlb->mm->nr_ptes);
1701 spin_unlock(ptl);
1702 tlb_remove_page(tlb, page);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001703 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001704 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001705}
1706
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001707bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001708 unsigned long old_addr,
1709 unsigned long new_addr, unsigned long old_end,
1710 pmd_t *old_pmd, pmd_t *new_pmd)
1711{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001712 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001713 pmd_t pmd;
1714
1715 struct mm_struct *mm = vma->vm_mm;
1716
1717 if ((old_addr & ~HPAGE_PMD_MASK) ||
1718 (new_addr & ~HPAGE_PMD_MASK) ||
1719 old_end - old_addr < HPAGE_PMD_SIZE ||
1720 (new_vma->vm_flags & VM_NOHUGEPAGE))
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001721 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001722
1723 /*
1724 * The destination pmd shouldn't be established, free_pgtables()
1725 * should have release it.
1726 */
1727 if (WARN_ON(!pmd_none(*new_pmd))) {
1728 VM_BUG_ON(pmd_trans_huge(*new_pmd));
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001729 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001730 }
1731
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001732 /*
1733 * We don't have to worry about the ordering of src and dst
1734 * ptlocks because exclusive mmap_sem prevents deadlock.
1735 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001736 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1737 if (old_ptl) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001738 new_ptl = pmd_lockptr(mm, new_pmd);
1739 if (new_ptl != old_ptl)
1740 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001741 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001742 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001743
Kirill A. Shutemov69a8ec22016-02-17 13:11:12 -08001744 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1745 vma_is_anonymous(vma)) {
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301746 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001747 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1748 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001749 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301750 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1751 if (new_ptl != old_ptl)
1752 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001753 spin_unlock(old_ptl);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001754 return true;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001755 }
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001756 return false;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001757}
1758
Mel Gormanf123d742013-10-07 11:28:49 +01001759/*
1760 * Returns
1761 * - 0 if PMD could not be locked
1762 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1763 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1764 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001765int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001766 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001767{
1768 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001769 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001770 int ret = 0;
1771
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001772 ptl = __pmd_trans_huge_lock(pmd, vma);
1773 if (ptl) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001774 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001775 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001776 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001777
1778 /*
1779 * Avoid trapping faults against the zero page. The read-only
1780 * data is likely to be read-cached on the local CPU and
1781 * local/remote hits to the zero page are not interesting.
1782 */
1783 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1784 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001785 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001786 }
1787
Mel Gorman10c10452015-02-12 14:58:44 -08001788 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001789 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001790 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001791 if (preserve_write)
1792 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001793 ret = HPAGE_PMD_NR;
1794 set_pmd_at(mm, addr, pmd, entry);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001795 BUG_ON(!preserve_write && pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001796 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001797 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001798 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001799
1800 return ret;
1801}
1802
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001803/*
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001804 * Returns true if a given pmd maps a thp, false otherwise.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001805 *
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001806 * Note that if it returns true, this routine returns without unlocking page
1807 * table lock. So callers must unlock it.
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001808 */
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001809spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001810{
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001811 spinlock_t *ptl;
1812 ptl = pmd_lock(vma->vm_mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08001813 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -08001814 return ptl;
1815 spin_unlock(ptl);
1816 return NULL;
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001817}
1818
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001819#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001820
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001821int hugepage_madvise(struct vm_area_struct *vma,
1822 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001823{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001824 switch (advice) {
1825 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001826#ifdef CONFIG_S390
1827 /*
1828 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1829 * can't handle this properly after s390_enable_sie, so we simply
1830 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1831 */
1832 if (mm_has_pgste(vma->vm_mm))
1833 return 0;
1834#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001835 /*
1836 * Be somewhat over-protective like KSM for now!
1837 */
Jason J. Herne1a763612015-11-20 15:57:04 -08001838 if (*vm_flags & VM_NO_THP)
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001839 return -EINVAL;
1840 *vm_flags &= ~VM_NOHUGEPAGE;
1841 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001842 /*
1843 * If the vma become good for khugepaged to scan,
1844 * register it here without waiting a page fault that
1845 * may not happen any time soon.
1846 */
David Rientjes6d50e602014-10-29 14:50:31 -07001847 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001848 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001849 break;
1850 case MADV_NOHUGEPAGE:
1851 /*
1852 * Be somewhat over-protective like KSM for now!
1853 */
Jason J. Herne1a763612015-11-20 15:57:04 -08001854 if (*vm_flags & VM_NO_THP)
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001855 return -EINVAL;
1856 *vm_flags &= ~VM_HUGEPAGE;
1857 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001858 /*
1859 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1860 * this vma even if we leave the mm registered in khugepaged if
1861 * it got registered before VM_NOHUGEPAGE was set.
1862 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001863 break;
1864 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001865
1866 return 0;
1867}
1868
Andrea Arcangeliba761492011-01-13 15:46:58 -08001869static int __init khugepaged_slab_init(void)
1870{
1871 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1872 sizeof(struct mm_slot),
1873 __alignof__(struct mm_slot), 0, NULL);
1874 if (!mm_slot_cache)
1875 return -ENOMEM;
1876
1877 return 0;
1878}
1879
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -07001880static void __init khugepaged_slab_exit(void)
1881{
1882 kmem_cache_destroy(mm_slot_cache);
1883}
1884
Andrea Arcangeliba761492011-01-13 15:46:58 -08001885static inline struct mm_slot *alloc_mm_slot(void)
1886{
1887 if (!mm_slot_cache) /* initialization failed */
1888 return NULL;
1889 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1890}
1891
1892static inline void free_mm_slot(struct mm_slot *mm_slot)
1893{
1894 kmem_cache_free(mm_slot_cache, mm_slot);
1895}
1896
Andrea Arcangeliba761492011-01-13 15:46:58 -08001897static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1898{
1899 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08001900
Sasha Levinb67bfe02013-02-27 17:06:00 -08001901 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001902 if (mm == mm_slot->mm)
1903 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001904
Andrea Arcangeliba761492011-01-13 15:46:58 -08001905 return NULL;
1906}
1907
1908static void insert_to_mm_slots_hash(struct mm_struct *mm,
1909 struct mm_slot *mm_slot)
1910{
Andrea Arcangeliba761492011-01-13 15:46:58 -08001911 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001912 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001913}
1914
1915static inline int khugepaged_test_exit(struct mm_struct *mm)
1916{
1917 return atomic_read(&mm->mm_users) == 0;
1918}
1919
1920int __khugepaged_enter(struct mm_struct *mm)
1921{
1922 struct mm_slot *mm_slot;
1923 int wakeup;
1924
1925 mm_slot = alloc_mm_slot();
1926 if (!mm_slot)
1927 return -ENOMEM;
1928
1929 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07001930 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001931 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1932 free_mm_slot(mm_slot);
1933 return 0;
1934 }
1935
1936 spin_lock(&khugepaged_mm_lock);
1937 insert_to_mm_slots_hash(mm, mm_slot);
1938 /*
1939 * Insert just behind the scanning cursor, to let the area settle
1940 * down a little.
1941 */
1942 wakeup = list_empty(&khugepaged_scan.mm_head);
1943 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1944 spin_unlock(&khugepaged_mm_lock);
1945
1946 atomic_inc(&mm->mm_count);
1947 if (wakeup)
1948 wake_up_interruptible(&khugepaged_wait);
1949
1950 return 0;
1951}
1952
David Rientjes6d50e602014-10-29 14:50:31 -07001953int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1954 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001955{
1956 unsigned long hstart, hend;
1957 if (!vma->anon_vma)
1958 /*
1959 * Not yet faulted in so we will register later in the
1960 * page fault if needed.
1961 */
1962 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001963 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001964 /* khugepaged not yet working on file or special mappings */
1965 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07001966 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001967 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1968 hend = vma->vm_end & HPAGE_PMD_MASK;
1969 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07001970 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001971 return 0;
1972}
1973
1974void __khugepaged_exit(struct mm_struct *mm)
1975{
1976 struct mm_slot *mm_slot;
1977 int free = 0;
1978
1979 spin_lock(&khugepaged_mm_lock);
1980 mm_slot = get_mm_slot(mm);
1981 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001982 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001983 list_del(&mm_slot->mm_node);
1984 free = 1;
1985 }
Chris Wrightd788e802011-07-25 17:12:14 -07001986 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001987
1988 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08001989 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1990 free_mm_slot(mm_slot);
1991 mmdrop(mm);
1992 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08001993 /*
1994 * This is required to serialize against
1995 * khugepaged_test_exit() (which is guaranteed to run
1996 * under mmap sem read mode). Stop here (after we
1997 * return all pagetables will be destroyed) until
1998 * khugepaged has finished working on the pagetables
1999 * under the mmap_sem.
2000 */
2001 down_write(&mm->mmap_sem);
2002 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002003 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002004}
2005
2006static void release_pte_page(struct page *page)
2007{
2008 /* 0 stands for page_is_file_cache(page) == false */
2009 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2010 unlock_page(page);
2011 putback_lru_page(page);
2012}
2013
2014static void release_pte_pages(pte_t *pte, pte_t *_pte)
2015{
2016 while (--_pte >= pte) {
2017 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002018 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002019 release_pte_page(pte_page(pteval));
2020 }
2021}
2022
Andrea Arcangeliba761492011-01-13 15:46:58 -08002023static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2024 unsigned long address,
2025 pte_t *pte)
2026{
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002027 struct page *page = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002028 pte_t *_pte;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002029 int none_or_zero = 0, result = 0;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002030 bool referenced = false, writable = false;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002031
Andrea Arcangeliba761492011-01-13 15:46:58 -08002032 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2033 _pte++, address += PAGE_SIZE) {
2034 pte_t pteval = *_pte;
Minchan Kim47aee4d2015-10-22 13:32:19 -07002035 if (pte_none(pteval) || (pte_present(pteval) &&
2036 is_zero_pfn(pte_pfn(pteval)))) {
Andrea Arcangelic1294d02015-09-04 15:46:27 -07002037 if (!userfaultfd_armed(vma) &&
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002038 ++none_or_zero <= khugepaged_max_ptes_none) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002039 continue;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002040 } else {
2041 result = SCAN_EXCEED_NONE_PTE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002042 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002043 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002044 }
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002045 if (!pte_present(pteval)) {
2046 result = SCAN_PTE_NON_PRESENT;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002047 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002048 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002049 page = vm_normal_page(vma, address, pteval);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002050 if (unlikely(!page)) {
2051 result = SCAN_PAGE_NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002052 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002053 }
Bob Liu344aa352012-12-11 16:00:34 -08002054
Sasha Levin309381fea2014-01-23 15:52:54 -08002055 VM_BUG_ON_PAGE(PageCompound(page), page);
2056 VM_BUG_ON_PAGE(!PageAnon(page), page);
2057 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002058
Andrea Arcangeliba761492011-01-13 15:46:58 -08002059 /*
2060 * We can do it before isolate_lru_page because the
2061 * page can't be freed from under us. NOTE: PG_lock
2062 * is needed to serialize against split_huge_page
2063 * when invoked from the VM.
2064 */
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002065 if (!trylock_page(page)) {
2066 result = SCAN_PAGE_LOCK;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002067 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002068 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002069
2070 /*
2071 * cannot use mapcount: can't collapse if there's a gup pin.
2072 * The page must only be referenced by the scanned process
2073 * and page swap cache.
2074 */
2075 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2076 unlock_page(page);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002077 result = SCAN_PAGE_COUNT;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002078 goto out;
2079 }
2080 if (pte_write(pteval)) {
2081 writable = true;
2082 } else {
2083 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2084 unlock_page(page);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002085 result = SCAN_SWAP_CACHE_PAGE;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002086 goto out;
2087 }
2088 /*
2089 * Page is not in the swap cache. It can be collapsed
2090 * into a THP.
2091 */
2092 }
2093
Andrea Arcangeliba761492011-01-13 15:46:58 -08002094 /*
2095 * Isolate the page to avoid collapsing an hugepage
2096 * currently in use by the VM.
2097 */
2098 if (isolate_lru_page(page)) {
2099 unlock_page(page);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002100 result = SCAN_DEL_PAGE_LRU;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002101 goto out;
2102 }
2103 /* 0 stands for page_is_file_cache(page) == false */
2104 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002105 VM_BUG_ON_PAGE(!PageLocked(page), page);
2106 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002107
2108 /* If there is no mapped pte young don't collapse the page */
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07002109 if (pte_young(pteval) ||
2110 page_is_young(page) || PageReferenced(page) ||
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002111 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002112 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002113 }
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002114 if (likely(writable)) {
2115 if (likely(referenced)) {
2116 result = SCAN_SUCCEED;
yalin wang16fd0fe2016-01-21 16:40:30 -08002117 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002118 referenced, writable, result);
2119 return 1;
2120 }
2121 } else {
2122 result = SCAN_PAGE_RO;
2123 }
2124
Andrea Arcangeliba761492011-01-13 15:46:58 -08002125out:
Bob Liu344aa352012-12-11 16:00:34 -08002126 release_pte_pages(pte, _pte);
yalin wang16fd0fe2016-01-21 16:40:30 -08002127 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002128 referenced, writable, result);
Bob Liu344aa352012-12-11 16:00:34 -08002129 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002130}
2131
2132static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2133 struct vm_area_struct *vma,
2134 unsigned long address,
2135 spinlock_t *ptl)
2136{
2137 pte_t *_pte;
2138 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2139 pte_t pteval = *_pte;
2140 struct page *src_page;
2141
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002142 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002143 clear_user_highpage(page, address);
2144 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002145 if (is_zero_pfn(pte_pfn(pteval))) {
2146 /*
2147 * ptl mostly unnecessary.
2148 */
2149 spin_lock(ptl);
2150 /*
2151 * paravirt calls inside pte_clear here are
2152 * superfluous.
2153 */
2154 pte_clear(vma->vm_mm, address, _pte);
2155 spin_unlock(ptl);
2156 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002157 } else {
2158 src_page = pte_page(pteval);
2159 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002160 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002161 release_pte_page(src_page);
2162 /*
2163 * ptl mostly unnecessary, but preempt has to
2164 * be disabled to update the per-cpu stats
2165 * inside page_remove_rmap().
2166 */
2167 spin_lock(ptl);
2168 /*
2169 * paravirt calls inside pte_clear here are
2170 * superfluous.
2171 */
2172 pte_clear(vma->vm_mm, address, _pte);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002173 page_remove_rmap(src_page, false);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002174 spin_unlock(ptl);
2175 free_page_and_swap_cache(src_page);
2176 }
2177
2178 address += PAGE_SIZE;
2179 page++;
2180 }
2181}
2182
Xiao Guangrong26234f32012-10-08 16:29:51 -07002183static void khugepaged_alloc_sleep(void)
2184{
Petr Mladekbde43c62015-09-08 15:04:05 -07002185 DEFINE_WAIT(wait);
2186
2187 add_wait_queue(&khugepaged_wait, &wait);
2188 freezable_schedule_timeout_interruptible(
2189 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2190 remove_wait_queue(&khugepaged_wait, &wait);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002191}
2192
Bob Liu9f1b8682013-11-12 15:07:37 -08002193static int khugepaged_node_load[MAX_NUMNODES];
2194
David Rientjes14a4e212014-08-06 16:07:29 -07002195static bool khugepaged_scan_abort(int nid)
2196{
2197 int i;
2198
2199 /*
2200 * If zone_reclaim_mode is disabled, then no extra effort is made to
2201 * allocate memory locally.
2202 */
2203 if (!zone_reclaim_mode)
2204 return false;
2205
2206 /* If there is a count for this node already, it must be acceptable */
2207 if (khugepaged_node_load[nid])
2208 return false;
2209
2210 for (i = 0; i < MAX_NUMNODES; i++) {
2211 if (!khugepaged_node_load[i])
2212 continue;
2213 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2214 return true;
2215 }
2216 return false;
2217}
2218
Xiao Guangrong26234f32012-10-08 16:29:51 -07002219#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002220static int khugepaged_find_target_node(void)
2221{
2222 static int last_khugepaged_target_node = NUMA_NO_NODE;
2223 int nid, target_node = 0, max_value = 0;
2224
2225 /* find first node with max normal pages hit */
2226 for (nid = 0; nid < MAX_NUMNODES; nid++)
2227 if (khugepaged_node_load[nid] > max_value) {
2228 max_value = khugepaged_node_load[nid];
2229 target_node = nid;
2230 }
2231
2232 /* do some balance if several nodes have the same hit record */
2233 if (target_node <= last_khugepaged_target_node)
2234 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2235 nid++)
2236 if (max_value == khugepaged_node_load[nid]) {
2237 target_node = nid;
2238 break;
2239 }
2240
2241 last_khugepaged_target_node = target_node;
2242 return target_node;
2243}
2244
Xiao Guangrong26234f32012-10-08 16:29:51 -07002245static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2246{
2247 if (IS_ERR(*hpage)) {
2248 if (!*wait)
2249 return false;
2250
2251 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002252 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002253 khugepaged_alloc_sleep();
2254 } else if (*hpage) {
2255 put_page(*hpage);
2256 *hpage = NULL;
2257 }
2258
2259 return true;
2260}
2261
Michal Hocko3b363692015-04-15 16:13:29 -07002262static struct page *
2263khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Aaron Tomlind6669d62015-11-06 16:28:52 -08002264 unsigned long address, int node)
Xiao Guangrong26234f32012-10-08 16:29:51 -07002265{
Sasha Levin309381fea2014-01-23 15:52:54 -08002266 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002267
Xiao Guangrong26234f32012-10-08 16:29:51 -07002268 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002269 * Before allocating the hugepage, release the mmap_sem read lock.
2270 * The allocation can take potentially a long time if it involves
2271 * sync compaction, and we do not need to hold the mmap_sem during
2272 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002273 */
2274 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002275
Vlastimil Babka96db8002015-09-08 15:03:50 -07002276 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002277 if (unlikely(!*hpage)) {
2278 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2279 *hpage = ERR_PTR(-ENOMEM);
2280 return NULL;
2281 }
2282
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002283 prep_transhuge_page(*hpage);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002284 count_vm_event(THP_COLLAPSE_ALLOC);
2285 return *hpage;
2286}
2287#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002288static int khugepaged_find_target_node(void)
2289{
2290 return 0;
2291}
2292
Mel Gorman444eb2a42016-03-17 14:19:23 -07002293static inline struct page *alloc_khugepaged_hugepage(void)
Bob Liu10dc4152013-11-12 15:07:35 -08002294{
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002295 struct page *page;
2296
Mel Gorman444eb2a42016-03-17 14:19:23 -07002297 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2298 HPAGE_PMD_ORDER);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08002299 if (page)
2300 prep_transhuge_page(page);
2301 return page;
Bob Liu10dc4152013-11-12 15:07:35 -08002302}
2303
Xiao Guangrong26234f32012-10-08 16:29:51 -07002304static struct page *khugepaged_alloc_hugepage(bool *wait)
2305{
2306 struct page *hpage;
2307
2308 do {
Mel Gorman444eb2a42016-03-17 14:19:23 -07002309 hpage = alloc_khugepaged_hugepage();
Xiao Guangrong26234f32012-10-08 16:29:51 -07002310 if (!hpage) {
2311 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2312 if (!*wait)
2313 return NULL;
2314
2315 *wait = false;
2316 khugepaged_alloc_sleep();
2317 } else
2318 count_vm_event(THP_COLLAPSE_ALLOC);
2319 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2320
2321 return hpage;
2322}
2323
2324static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2325{
2326 if (!*hpage)
2327 *hpage = khugepaged_alloc_hugepage(wait);
2328
2329 if (unlikely(!*hpage))
2330 return false;
2331
2332 return true;
2333}
2334
Michal Hocko3b363692015-04-15 16:13:29 -07002335static struct page *
2336khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Aaron Tomlind6669d62015-11-06 16:28:52 -08002337 unsigned long address, int node)
Xiao Guangrong26234f32012-10-08 16:29:51 -07002338{
2339 up_read(&mm->mmap_sem);
2340 VM_BUG_ON(!*hpage);
Michal Hocko3b363692015-04-15 16:13:29 -07002341
Xiao Guangrong26234f32012-10-08 16:29:51 -07002342 return *hpage;
2343}
2344#endif
2345
Bob Liufa475e52012-12-11 16:00:39 -08002346static bool hugepage_vma_check(struct vm_area_struct *vma)
2347{
2348 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2349 (vma->vm_flags & VM_NOHUGEPAGE))
2350 return false;
Bob Liufa475e52012-12-11 16:00:39 -08002351 if (!vma->anon_vma || vma->vm_ops)
2352 return false;
2353 if (is_vma_temporary_stack(vma))
2354 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002355 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002356 return true;
2357}
2358
Andrea Arcangeliba761492011-01-13 15:46:58 -08002359static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002360 unsigned long address,
2361 struct page **hpage,
2362 struct vm_area_struct *vma,
2363 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002364{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002365 pmd_t *pmd, _pmd;
2366 pte_t *pte;
2367 pgtable_t pgtable;
2368 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002369 spinlock_t *pmd_ptl, *pte_ptl;
Arnd Bergmann629d9d12016-01-18 21:50:26 +01002370 int isolated = 0, result = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002371 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002372 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002373 unsigned long mmun_start; /* For mmu_notifiers */
2374 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07002375 gfp_t gfp;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002376
2377 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002378
Michal Hocko3b363692015-04-15 16:13:29 -07002379 /* Only allocate from the target node */
Mel Gorman444eb2a42016-03-17 14:19:23 -07002380 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
Michal Hocko3b363692015-04-15 16:13:29 -07002381
Xiao Guangrong26234f32012-10-08 16:29:51 -07002382 /* release the mmap_sem read lock. */
Aaron Tomlind6669d62015-11-06 16:28:52 -08002383 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002384 if (!new_page) {
2385 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2386 goto out_nolock;
2387 }
Andrea Arcangelice83d212011-01-13 15:47:06 -08002388
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002389 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002390 result = SCAN_CGROUP_CHARGE_FAIL;
2391 goto out_nolock;
2392 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002393
2394 /*
2395 * Prevent all access to pagetables with the exception of
2396 * gup_fast later hanlded by the ptep_clear_flush and the VM
2397 * handled by the anon_vma lock + PG_lock.
2398 */
2399 down_write(&mm->mmap_sem);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002400 if (unlikely(khugepaged_test_exit(mm))) {
2401 result = SCAN_ANY_PROCESS;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002402 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002403 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002404
2405 vma = find_vma(mm, address);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002406 if (!vma) {
2407 result = SCAN_VMA_NULL;
Libina8f531eb2013-09-11 14:20:38 -07002408 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002409 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002410 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2411 hend = vma->vm_end & HPAGE_PMD_MASK;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002412 if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2413 result = SCAN_ADDRESS_RANGE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002414 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002415 }
2416 if (!hugepage_vma_check(vma)) {
2417 result = SCAN_VMA_CHECK;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002418 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002419 }
Bob Liu62190492012-12-11 16:00:37 -08002420 pmd = mm_find_pmd(mm, address);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002421 if (!pmd) {
2422 result = SCAN_PMD_NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002423 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002424 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002425
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002426 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002427
2428 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002429 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002430
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002431 mmun_start = address;
2432 mmun_end = address + HPAGE_PMD_SIZE;
2433 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002434 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002435 /*
2436 * After this gup_fast can't run anymore. This also removes
2437 * any huge TLB entry from the CPU so we won't allow
2438 * huge and small TLB entries for the same virtual address
2439 * to avoid the risk of CPU bugs in that area.
2440 */
Aneesh Kumar K.V15a25b22015-06-24 16:57:39 -07002441 _pmd = pmdp_collapse_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002442 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002443 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002444
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002445 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002446 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002447 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002448
2449 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002450 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002451 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002452 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002453 /*
2454 * We can only use set_pmd_at when establishing
2455 * hugepmds and never for establishing regular pmds that
2456 * points to regular pagetables. Use pmd_populate for that
2457 */
2458 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002459 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002460 anon_vma_unlock_write(vma->anon_vma);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002461 result = SCAN_FAIL;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002462 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002463 }
2464
2465 /*
2466 * All pages are isolated and locked so anon_vma rmap
2467 * can't run anymore.
2468 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002469 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002470
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002471 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002472 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002473 __SetPageUptodate(new_page);
2474 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002475
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002476 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2477 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002478
2479 /*
2480 * spin_lock() below is not the equivalent of smp_wmb(), so
2481 * this is needed to avoid the copy_huge_page writes to become
2482 * visible after the set_pmd_at() write.
2483 */
2484 smp_wmb();
2485
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002486 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002487 BUG_ON(!pmd_none(*pmd));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002488 page_add_new_anon_rmap(new_page, vma, address, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002489 mem_cgroup_commit_charge(new_page, memcg, false, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07002490 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002491 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002492 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002493 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002494 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002495
2496 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002497
Andrea Arcangeliba761492011-01-13 15:46:58 -08002498 khugepaged_pages_collapsed++;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002499 result = SCAN_SUCCEED;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002500out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002501 up_write(&mm->mmap_sem);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002502 trace_mm_collapse_huge_page(mm, isolated, result);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002503 return;
2504
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002505out_nolock:
2506 trace_mm_collapse_huge_page(mm, isolated, result);
2507 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002508out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002509 mem_cgroup_cancel_charge(new_page, memcg, true);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002510 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002511}
2512
2513static int khugepaged_scan_pmd(struct mm_struct *mm,
2514 struct vm_area_struct *vma,
2515 unsigned long address,
2516 struct page **hpage)
2517{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002518 pmd_t *pmd;
2519 pte_t *pte, *_pte;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002520 int ret = 0, none_or_zero = 0, result = 0;
2521 struct page *page = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002522 unsigned long _address;
2523 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002524 int node = NUMA_NO_NODE;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002525 bool writable = false, referenced = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002526
2527 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2528
Bob Liu62190492012-12-11 16:00:37 -08002529 pmd = mm_find_pmd(mm, address);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002530 if (!pmd) {
2531 result = SCAN_PMD_NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002532 goto out;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002533 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002534
Bob Liu9f1b8682013-11-12 15:07:37 -08002535 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002536 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2537 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2538 _pte++, _address += PAGE_SIZE) {
2539 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002540 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangelic1294d02015-09-04 15:46:27 -07002541 if (!userfaultfd_armed(vma) &&
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002542 ++none_or_zero <= khugepaged_max_ptes_none) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002543 continue;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002544 } else {
2545 result = SCAN_EXCEED_NONE_PTE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002546 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002547 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002548 }
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002549 if (!pte_present(pteval)) {
2550 result = SCAN_PTE_NON_PRESENT;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002551 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002552 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002553 if (pte_write(pteval))
2554 writable = true;
2555
Andrea Arcangeliba761492011-01-13 15:46:58 -08002556 page = vm_normal_page(vma, _address, pteval);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002557 if (unlikely(!page)) {
2558 result = SCAN_PAGE_NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002559 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002560 }
Kirill A. Shutemovb1caa952016-01-15 16:52:39 -08002561
2562 /* TODO: teach khugepaged to collapse THP mapped with pte */
2563 if (PageCompound(page)) {
2564 result = SCAN_PAGE_COMPOUND;
2565 goto out_unmap;
2566 }
2567
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002568 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002569 * Record which node the original page is from and save this
2570 * information to khugepaged_node_load[].
2571 * Khupaged will allocate hugepage from the node has the max
2572 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002573 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002574 node = page_to_nid(page);
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002575 if (khugepaged_scan_abort(node)) {
2576 result = SCAN_SCAN_ABORT;
David Rientjes14a4e212014-08-06 16:07:29 -07002577 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002578 }
Bob Liu9f1b8682013-11-12 15:07:37 -08002579 khugepaged_node_load[node]++;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002580 if (!PageLRU(page)) {
Kirill A. Shutemov0fda2782016-03-25 14:22:20 -07002581 result = SCAN_PAGE_LRU;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002582 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002583 }
2584 if (PageLocked(page)) {
2585 result = SCAN_PAGE_LOCK;
2586 goto out_unmap;
2587 }
2588 if (!PageAnon(page)) {
2589 result = SCAN_PAGE_ANON;
2590 goto out_unmap;
2591 }
2592
Ebru Akagunduz10359212015-02-11 15:28:28 -08002593 /*
2594 * cannot use mapcount: can't collapse if there's a gup pin.
2595 * The page must only be referenced by the scanned process
2596 * and page swap cache.
2597 */
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002598 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2599 result = SCAN_PAGE_COUNT;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002600 goto out_unmap;
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002601 }
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07002602 if (pte_young(pteval) ||
2603 page_is_young(page) || PageReferenced(page) ||
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002604 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002605 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002606 }
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002607 if (writable) {
2608 if (referenced) {
2609 result = SCAN_SUCCEED;
2610 ret = 1;
2611 } else {
2612 result = SCAN_NO_REFERENCED_PAGE;
2613 }
2614 } else {
2615 result = SCAN_PAGE_RO;
2616 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002617out_unmap:
2618 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002619 if (ret) {
2620 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002621 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002622 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002623 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002624out:
yalin wang16fd0fe2016-01-21 16:40:30 -08002625 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
Ebru Akagunduz7d2eba02016-01-14 15:22:19 -08002626 none_or_zero, result);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002627 return ret;
2628}
2629
2630static void collect_mm_slot(struct mm_slot *mm_slot)
2631{
2632 struct mm_struct *mm = mm_slot->mm;
2633
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002634 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002635
2636 if (khugepaged_test_exit(mm)) {
2637 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002638 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002639 list_del(&mm_slot->mm_node);
2640
2641 /*
2642 * Not strictly needed because the mm exited already.
2643 *
2644 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2645 */
2646
2647 /* khugepaged_mm_lock actually not necessary for the below */
2648 free_mm_slot(mm_slot);
2649 mmdrop(mm);
2650 }
2651}
2652
2653static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2654 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002655 __releases(&khugepaged_mm_lock)
2656 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002657{
2658 struct mm_slot *mm_slot;
2659 struct mm_struct *mm;
2660 struct vm_area_struct *vma;
2661 int progress = 0;
2662
2663 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002664 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002665
2666 if (khugepaged_scan.mm_slot)
2667 mm_slot = khugepaged_scan.mm_slot;
2668 else {
2669 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2670 struct mm_slot, mm_node);
2671 khugepaged_scan.address = 0;
2672 khugepaged_scan.mm_slot = mm_slot;
2673 }
2674 spin_unlock(&khugepaged_mm_lock);
2675
2676 mm = mm_slot->mm;
2677 down_read(&mm->mmap_sem);
2678 if (unlikely(khugepaged_test_exit(mm)))
2679 vma = NULL;
2680 else
2681 vma = find_vma(mm, khugepaged_scan.address);
2682
2683 progress++;
2684 for (; vma; vma = vma->vm_next) {
2685 unsigned long hstart, hend;
2686
2687 cond_resched();
2688 if (unlikely(khugepaged_test_exit(mm))) {
2689 progress++;
2690 break;
2691 }
Bob Liufa475e52012-12-11 16:00:39 -08002692 if (!hugepage_vma_check(vma)) {
2693skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002694 progress++;
2695 continue;
2696 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002697 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2698 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002699 if (hstart >= hend)
2700 goto skip;
2701 if (khugepaged_scan.address > hend)
2702 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002703 if (khugepaged_scan.address < hstart)
2704 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002705 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002706
2707 while (khugepaged_scan.address < hend) {
2708 int ret;
2709 cond_resched();
2710 if (unlikely(khugepaged_test_exit(mm)))
2711 goto breakouterloop;
2712
2713 VM_BUG_ON(khugepaged_scan.address < hstart ||
2714 khugepaged_scan.address + HPAGE_PMD_SIZE >
2715 hend);
2716 ret = khugepaged_scan_pmd(mm, vma,
2717 khugepaged_scan.address,
2718 hpage);
2719 /* move to next address */
2720 khugepaged_scan.address += HPAGE_PMD_SIZE;
2721 progress += HPAGE_PMD_NR;
2722 if (ret)
2723 /* we released mmap_sem so break loop */
2724 goto breakouterloop_mmap_sem;
2725 if (progress >= pages)
2726 goto breakouterloop;
2727 }
2728 }
2729breakouterloop:
2730 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2731breakouterloop_mmap_sem:
2732
2733 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002734 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002735 /*
2736 * Release the current mm_slot if this mm is about to die, or
2737 * if we scanned all vmas of this mm.
2738 */
2739 if (khugepaged_test_exit(mm) || !vma) {
2740 /*
2741 * Make sure that if mm_users is reaching zero while
2742 * khugepaged runs here, khugepaged_exit will find
2743 * mm_slot not pointing to the exiting mm.
2744 */
2745 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2746 khugepaged_scan.mm_slot = list_entry(
2747 mm_slot->mm_node.next,
2748 struct mm_slot, mm_node);
2749 khugepaged_scan.address = 0;
2750 } else {
2751 khugepaged_scan.mm_slot = NULL;
2752 khugepaged_full_scans++;
2753 }
2754
2755 collect_mm_slot(mm_slot);
2756 }
2757
2758 return progress;
2759}
2760
2761static int khugepaged_has_work(void)
2762{
2763 return !list_empty(&khugepaged_scan.mm_head) &&
2764 khugepaged_enabled();
2765}
2766
2767static int khugepaged_wait_event(void)
2768{
2769 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002770 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002771}
2772
Xiao Guangrongd5169042012-10-08 16:29:48 -07002773static void khugepaged_do_scan(void)
2774{
2775 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002776 unsigned int progress = 0, pass_through_head = 0;
2777 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002778 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002779
2780 barrier(); /* write khugepaged_pages_to_scan to local stack */
2781
2782 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002783 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002784 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002785
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002786 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002787
Jiri Kosinacd092412015-06-24 16:56:07 -07002788 if (unlikely(kthread_should_stop() || try_to_freeze()))
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002789 break;
2790
Andrea Arcangeliba761492011-01-13 15:46:58 -08002791 spin_lock(&khugepaged_mm_lock);
2792 if (!khugepaged_scan.mm_slot)
2793 pass_through_head++;
2794 if (khugepaged_has_work() &&
2795 pass_through_head < 2)
2796 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002797 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002798 else
2799 progress = pages;
2800 spin_unlock(&khugepaged_mm_lock);
2801 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002802
Xiao Guangrongd5169042012-10-08 16:29:48 -07002803 if (!IS_ERR_OR_NULL(hpage))
2804 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002805}
2806
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002807static void khugepaged_wait_work(void)
2808{
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002809 if (khugepaged_has_work()) {
2810 if (!khugepaged_scan_sleep_millisecs)
2811 return;
2812
2813 wait_event_freezable_timeout(khugepaged_wait,
2814 kthread_should_stop(),
2815 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2816 return;
2817 }
2818
2819 if (khugepaged_enabled())
2820 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2821}
2822
Andrea Arcangeliba761492011-01-13 15:46:58 -08002823static int khugepaged(void *none)
2824{
2825 struct mm_slot *mm_slot;
2826
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002827 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002828 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002829
Xiao Guangrongb7231782012-10-08 16:29:54 -07002830 while (!kthread_should_stop()) {
2831 khugepaged_do_scan();
2832 khugepaged_wait_work();
2833 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002834
2835 spin_lock(&khugepaged_mm_lock);
2836 mm_slot = khugepaged_scan.mm_slot;
2837 khugepaged_scan.mm_slot = NULL;
2838 if (mm_slot)
2839 collect_mm_slot(mm_slot);
2840 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002841 return 0;
2842}
2843
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002844static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2845 unsigned long haddr, pmd_t *pmd)
2846{
2847 struct mm_struct *mm = vma->vm_mm;
2848 pgtable_t pgtable;
2849 pmd_t _pmd;
2850 int i;
2851
2852 /* leave pmd empty until pte is filled */
2853 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2854
2855 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2856 pmd_populate(mm, &_pmd, pgtable);
2857
2858 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2859 pte_t *pte, entry;
2860 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2861 entry = pte_mkspecial(entry);
2862 pte = pte_offset_map(&_pmd, haddr);
2863 VM_BUG_ON(!pte_none(*pte));
2864 set_pte_at(mm, haddr, pte, entry);
2865 pte_unmap(pte);
2866 }
2867 smp_wmb(); /* make pte visible before pmd */
2868 pmd_populate(mm, pmd, pgtable);
2869 put_huge_zero_page();
2870}
2871
2872static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002873 unsigned long haddr, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002874{
2875 struct mm_struct *mm = vma->vm_mm;
2876 struct page *page;
2877 pgtable_t pgtable;
2878 pmd_t _pmd;
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002879 bool young, write, dirty;
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002880 unsigned long addr;
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002881 int i;
2882
2883 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2884 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2885 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
Dan Williams5c7fb562016-01-15 16:56:52 -08002886 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002887
2888 count_vm_event(THP_SPLIT_PMD);
2889
2890 if (vma_is_dax(vma)) {
2891 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2892 if (is_huge_zero_pmd(_pmd))
2893 put_huge_zero_page();
2894 return;
2895 } else if (is_huge_zero_pmd(*pmd)) {
2896 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2897 }
2898
2899 page = pmd_page(*pmd);
2900 VM_BUG_ON_PAGE(!page_count(page), page);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07002901 page_ref_add(page, HPAGE_PMD_NR - 1);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002902 write = pmd_write(*pmd);
2903 young = pmd_young(*pmd);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002904 dirty = pmd_dirty(*pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002905
Aneesh Kumar K.Vc777e2a2016-02-09 06:50:31 +05302906 pmdp_huge_split_prepare(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002907 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2908 pmd_populate(mm, &_pmd, pgtable);
2909
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002910 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002911 pte_t entry, *pte;
2912 /*
2913 * Note that NUMA hinting access restrictions are not
2914 * transferred to avoid any possibility of altering
2915 * permissions across VMAs.
2916 */
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002917 if (freeze) {
2918 swp_entry_t swp_entry;
2919 swp_entry = make_migration_entry(page + i, write);
2920 entry = swp_entry_to_pte(swp_entry);
2921 } else {
2922 entry = mk_pte(page + i, vma->vm_page_prot);
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002923 entry = maybe_mkwrite(entry, vma);
Kirill A. Shutemovba988282016-01-15 16:53:56 -08002924 if (!write)
2925 entry = pte_wrprotect(entry);
2926 if (!young)
2927 entry = pte_mkold(entry);
2928 }
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08002929 if (dirty)
2930 SetPageDirty(page + i);
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002931 pte = pte_offset_map(&_pmd, addr);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002932 BUG_ON(!pte_none(*pte));
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002933 set_pte_at(mm, addr, pte, entry);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002934 atomic_inc(&page[i]._mapcount);
2935 pte_unmap(pte);
2936 }
2937
2938 /*
2939 * Set PG_double_map before dropping compound_mapcount to avoid
2940 * false-negative page_mapped().
2941 */
2942 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2943 for (i = 0; i < HPAGE_PMD_NR; i++)
2944 atomic_inc(&page[i]._mapcount);
2945 }
2946
2947 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2948 /* Last compound_mapcount is gone. */
2949 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2950 if (TestClearPageDoubleMap(page)) {
2951 /* No need in mapcount reference anymore */
2952 for (i = 0; i < HPAGE_PMD_NR; i++)
2953 atomic_dec(&page[i]._mapcount);
2954 }
2955 }
2956
2957 smp_wmb(); /* make pte visible before pmd */
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002958 /*
2959 * Up to this point the pmd is present and huge and userland has the
2960 * whole access to the hugepage during the split (which happens in
2961 * place). If we overwrite the pmd with the not-huge version pointing
2962 * to the pte here (which of course we could if all CPUs were bug
2963 * free), userland could trigger a small page size TLB miss on the
2964 * small sized TLB while the hugepage TLB entry is still established in
2965 * the huge TLB. Some CPU doesn't like that.
2966 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2967 * 383 on page 93. Intel should be safe but is also warns that it's
2968 * only safe if the permission and cache attributes of the two entries
2969 * loaded in the two TLB is identical (which should be the case here).
2970 * But it is generally safer to never allow small and huge TLB entries
2971 * for the same virtual address to be loaded simultaneously. So instead
2972 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2973 * current pmd notpresent (atomically because here the pmd_trans_huge
2974 * and pmd_trans_splitting must remain set at all times on the pmd
2975 * until the split is complete for this pmd), then we flush the SMP TLB
2976 * and finally we write the non-huge version of the pmd entry with
2977 * pmd_populate.
2978 */
2979 pmdp_invalidate(vma, haddr, pmd);
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002980 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002981
2982 if (freeze) {
Kirill A. Shutemov2ac015e2016-02-24 18:58:03 +03002983 for (i = 0; i < HPAGE_PMD_NR; i++) {
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08002984 page_remove_rmap(page + i, false);
2985 put_page(page + i);
2986 }
2987 }
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002988}
2989
2990void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07002991 unsigned long address, bool freeze)
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08002992{
2993 spinlock_t *ptl;
2994 struct mm_struct *mm = vma->vm_mm;
2995 unsigned long haddr = address & HPAGE_PMD_MASK;
2996
2997 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2998 ptl = pmd_lock(mm, pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08002999 if (pmd_trans_huge(*pmd)) {
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07003000 struct page *page = pmd_page(*pmd);
Dan Williams5c7fb562016-01-15 16:56:52 -08003001 if (PageMlocked(page))
Kirill A. Shutemov5f737712016-03-17 14:20:13 -07003002 clear_page_mlock(page);
Dan Williams5c7fb562016-01-15 16:56:52 -08003003 } else if (!pmd_devmap(*pmd))
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08003004 goto out;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003005 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08003006out:
Kirill A. Shutemoveef1b3b2016-01-15 16:53:53 -08003007 spin_unlock(ptl);
3008 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3009}
3010
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003011void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3012 bool freeze, struct page *page)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003013{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07003014 pgd_t *pgd;
3015 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003016 pmd_t *pmd;
3017
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08003018 pgd = pgd_offset(vma->vm_mm, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07003019 if (!pgd_present(*pgd))
3020 return;
3021
3022 pud = pud_offset(pgd, address);
3023 if (!pud_present(*pud))
3024 return;
3025
3026 pmd = pmd_offset(pud, address);
Dan Williams5c7fb562016-01-15 16:56:52 -08003027 if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003028 return;
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003029
3030 /*
3031 * If caller asks to setup a migration entries, we need a page to check
3032 * pmd against. Otherwise we can end up replacing wrong page.
3033 */
3034 VM_BUG_ON(freeze && !page);
3035 if (page && page != pmd_page(*pmd))
3036 return;
3037
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003038 /*
3039 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3040 * materialize from under us.
3041 */
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003042 __split_huge_pmd(vma, pmd, address, freeze);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003043}
3044
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07003045void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003046 unsigned long start,
3047 unsigned long end,
3048 long adjust_next)
3049{
3050 /*
3051 * If the new start address isn't hpage aligned and it could
3052 * previously contain an hugepage: check if we need to split
3053 * an huge pmd.
3054 */
3055 if (start & ~HPAGE_PMD_MASK &&
3056 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3057 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003058 split_huge_pmd_address(vma, start, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003059
3060 /*
3061 * If the new end address isn't hpage aligned and it could
3062 * previously contain an hugepage: check if we need to split
3063 * an huge pmd.
3064 */
3065 if (end & ~HPAGE_PMD_MASK &&
3066 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3067 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003068 split_huge_pmd_address(vma, end, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003069
3070 /*
3071 * If we're also updating the vma->vm_next->vm_start, if the new
3072 * vm_next->vm_start isn't page aligned and it could previously
3073 * contain an hugepage: check if we need to split an huge pmd.
3074 */
3075 if (adjust_next > 0) {
3076 struct vm_area_struct *next = vma->vm_next;
3077 unsigned long nstart = next->vm_start;
3078 nstart += adjust_next << PAGE_SHIFT;
3079 if (nstart & ~HPAGE_PMD_MASK &&
3080 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3081 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003082 split_huge_pmd_address(next, nstart, false, NULL);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003083 }
3084}
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003085
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003086static void freeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003087{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003088 enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3089 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3090 int i, ret;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003091
3092 VM_BUG_ON_PAGE(!PageHead(page), page);
3093
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003094 /* We only need TTU_SPLIT_HUGE_PMD once */
3095 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3096 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3097 /* Cut short if the page is unmapped */
3098 if (page_count(page) == 1)
3099 return;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003100
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003101 ret = try_to_unmap(page + i, ttu_flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003102 }
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003103 VM_BUG_ON(ret);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003104}
3105
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003106static void unfreeze_page(struct page *page)
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003107{
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003108 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003109
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003110 for (i = 0; i < HPAGE_PMD_NR; i++)
3111 remove_migration_ptes(page + i, page + i, true);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003112}
3113
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003114static void __split_huge_page_tail(struct page *head, int tail,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003115 struct lruvec *lruvec, struct list_head *list)
3116{
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003117 struct page *page_tail = head + tail;
3118
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003119 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
Joonsoo Kimfe896d12016-03-17 14:19:26 -07003120 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003121
3122 /*
3123 * tail_page->_count is zero and not changing from under us. But
3124 * get_page_unless_zero() may be running from under us on the
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003125 * tail_page. If we used atomic_set() below instead of atomic_inc(), we
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003126 * would then run atomic_set() concurrently with
3127 * get_page_unless_zero(), and atomic_set() is implemented in C not
3128 * using locked ops. spin_unlock on x86 sometime uses locked ops
3129 * because of PPro errata 66, 92, so unless somebody can guarantee
3130 * atomic_set() here would be safe on all archs (and not only on x86),
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003131 * it's safer to use atomic_inc().
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003132 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -07003133 page_ref_inc(page_tail);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003134
3135 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3136 page_tail->flags |= (head->flags &
3137 ((1L << PG_referenced) |
3138 (1L << PG_swapbacked) |
3139 (1L << PG_mlocked) |
3140 (1L << PG_uptodate) |
3141 (1L << PG_active) |
3142 (1L << PG_locked) |
Minchan Kimb8d3c4c2016-01-15 16:55:42 -08003143 (1L << PG_unevictable) |
3144 (1L << PG_dirty)));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003145
3146 /*
3147 * After clearing PageTail the gup refcount can be released.
3148 * Page flags also must be visible before we make the page non-compound.
3149 */
3150 smp_wmb();
3151
3152 clear_compound_head(page_tail);
3153
3154 if (page_is_young(head))
3155 set_page_young(page_tail);
3156 if (page_is_idle(head))
3157 set_page_idle(page_tail);
3158
3159 /* ->mapping in first tail page is compound_mapcount */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003160 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003161 page_tail);
3162 page_tail->mapping = head->mapping;
3163
3164 page_tail->index = head->index + tail;
3165 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3166 lru_add_page_tail(head, page_tail, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003167}
3168
3169static void __split_huge_page(struct page *page, struct list_head *list)
3170{
3171 struct page *head = compound_head(page);
3172 struct zone *zone = page_zone(head);
3173 struct lruvec *lruvec;
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003174 int i;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003175
3176 /* prevent PageLRU to go away from under us, and freeze lru stats */
3177 spin_lock_irq(&zone->lru_lock);
3178 lruvec = mem_cgroup_page_lruvec(head, zone);
3179
3180 /* complete memcg works before add pages to LRU */
3181 mem_cgroup_split_huge_fixup(head);
3182
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003183 for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
Kirill A. Shutemov8df651c2016-03-15 14:57:30 -07003184 __split_huge_page_tail(head, i, lruvec, list);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003185
3186 ClearPageCompound(head);
3187 spin_unlock_irq(&zone->lru_lock);
3188
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003189 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003190
3191 for (i = 0; i < HPAGE_PMD_NR; i++) {
3192 struct page *subpage = head + i;
3193 if (subpage == page)
3194 continue;
3195 unlock_page(subpage);
3196
3197 /*
3198 * Subpages may be freed if there wasn't any mapping
3199 * like if add_to_swap() is running on a lru page that
3200 * had its mapping zapped. And freeing these pages
3201 * requires taking the lru_lock so we do the put_page
3202 * of the tail pages after the split is complete.
3203 */
3204 put_page(subpage);
3205 }
3206}
3207
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -08003208int total_mapcount(struct page *page)
3209{
3210 int i, ret;
3211
3212 VM_BUG_ON_PAGE(PageTail(page), page);
3213
3214 if (likely(!PageCompound(page)))
3215 return atomic_read(&page->_mapcount) + 1;
3216
3217 ret = compound_mapcount(page);
3218 if (PageHuge(page))
3219 return ret;
3220 for (i = 0; i < HPAGE_PMD_NR; i++)
3221 ret += atomic_read(&page[i]._mapcount) + 1;
3222 if (PageDoubleMap(page))
3223 ret -= HPAGE_PMD_NR;
3224 return ret;
3225}
3226
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003227/*
3228 * This function splits huge page into normal pages. @page can point to any
3229 * subpage of huge page to split. Split doesn't change the position of @page.
3230 *
3231 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3232 * The huge page must be locked.
3233 *
3234 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3235 *
3236 * Both head page and tail pages will inherit mapping, flags, and so on from
3237 * the hugepage.
3238 *
3239 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3240 * they are not mapped.
3241 *
3242 * Returns 0 if the hugepage is split successfully.
3243 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3244 * us.
3245 */
3246int split_huge_page_to_list(struct page *page, struct list_head *list)
3247{
3248 struct page *head = compound_head(page);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003249 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003250 struct anon_vma *anon_vma;
3251 int count, mapcount, ret;
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08003252 bool mlocked;
Kirill A. Shutemov0b9b6ff2016-01-20 14:58:09 -08003253 unsigned long flags;
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003254
3255 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3256 VM_BUG_ON_PAGE(!PageAnon(page), page);
3257 VM_BUG_ON_PAGE(!PageLocked(page), page);
3258 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3259 VM_BUG_ON_PAGE(!PageCompound(page), page);
3260
3261 /*
3262 * The caller does not necessarily hold an mmap_sem that would prevent
3263 * the anon_vma disappearing so we first we take a reference to it
3264 * and then lock the anon_vma for write. This is similar to
3265 * page_lock_anon_vma_read except the write lock is taken to serialise
3266 * against parallel split or collapse operations.
3267 */
3268 anon_vma = page_get_anon_vma(head);
3269 if (!anon_vma) {
3270 ret = -EBUSY;
3271 goto out;
3272 }
3273 anon_vma_lock_write(anon_vma);
3274
3275 /*
3276 * Racy check if we can split the page, before freeze_page() will
3277 * split PMDs
3278 */
3279 if (total_mapcount(head) != page_count(head) - 1) {
3280 ret = -EBUSY;
3281 goto out_unlock;
3282 }
3283
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08003284 mlocked = PageMlocked(page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003285 freeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003286 VM_BUG_ON_PAGE(compound_mapcount(head), head);
3287
Kirill A. Shutemovd9654322016-01-15 16:54:43 -08003288 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3289 if (mlocked)
3290 lru_add_drain();
3291
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003292 /* Prevent deferred_split_scan() touching ->_count */
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003293 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003294 count = page_count(head);
3295 mapcount = total_mapcount(head);
Kirill A. Shutemovbd560862016-01-15 16:55:46 -08003296 if (!mapcount && count == 1) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003297 if (!list_empty(page_deferred_list(head))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003298 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003299 list_del(page_deferred_list(head));
3300 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003301 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003302 __split_huge_page(page, list);
3303 ret = 0;
Kirill A. Shutemovbd560862016-01-15 16:55:46 -08003304 } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003305 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003306 pr_alert("total_mapcount: %u, page_count(): %u\n",
3307 mapcount, count);
3308 if (PageTail(page))
3309 dump_page(head, NULL);
Kirill A. Shutemovbd560862016-01-15 16:55:46 -08003310 dump_page(page, "total_mapcount(head) > 0");
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003311 BUG();
3312 } else {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003313 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07003314 unfreeze_page(head);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08003315 ret = -EBUSY;
3316 }
3317
3318out_unlock:
3319 anon_vma_unlock_write(anon_vma);
3320 put_anon_vma(anon_vma);
3321out:
3322 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3323 return ret;
3324}
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003325
3326void free_transhuge_page(struct page *page)
3327{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003328 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003329 unsigned long flags;
3330
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003331 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003332 if (!list_empty(page_deferred_list(page))) {
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003333 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003334 list_del(page_deferred_list(page));
3335 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003336 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003337 free_compound_page(page);
3338}
3339
3340void deferred_split_huge_page(struct page *page)
3341{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003342 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003343 unsigned long flags;
3344
3345 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3346
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003347 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003348 if (list_empty(page_deferred_list(page))) {
Kirill A. Shutemovf9719a02016-03-17 14:18:45 -07003349 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003350 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3351 pgdata->split_queue_len++;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003352 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003353 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003354}
3355
3356static unsigned long deferred_split_count(struct shrinker *shrink,
3357 struct shrink_control *sc)
3358{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003359 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08003360 return ACCESS_ONCE(pgdata->split_queue_len);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003361}
3362
3363static unsigned long deferred_split_scan(struct shrinker *shrink,
3364 struct shrink_control *sc)
3365{
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003366 struct pglist_data *pgdata = NODE_DATA(sc->nid);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003367 unsigned long flags;
3368 LIST_HEAD(list), *pos, *next;
3369 struct page *page;
3370 int split = 0;
3371
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003372 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003373 /* Take pin on all head pages to avoid freeing them under us */
Kirill A. Shutemovae026202016-02-05 15:36:53 -08003374 list_for_each_safe(pos, next, &pgdata->split_queue) {
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003375 page = list_entry((void *)pos, struct page, mapping);
3376 page = compound_head(page);
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08003377 if (get_page_unless_zero(page)) {
3378 list_move(page_deferred_list(page), &list);
3379 } else {
3380 /* We lost race with put_compound_page() */
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003381 list_del_init(page_deferred_list(page));
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003382 pgdata->split_queue_len--;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003383 }
Kirill A. Shutemove3ae1952016-02-02 16:57:15 -08003384 if (!--sc->nr_to_scan)
3385 break;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003386 }
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003387 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003388
3389 list_for_each_safe(pos, next, &list) {
3390 page = list_entry((void *)pos, struct page, mapping);
3391 lock_page(page);
3392 /* split_huge_page() removes page from list on success */
3393 if (!split_huge_page(page))
3394 split++;
3395 unlock_page(page);
3396 put_page(page);
3397 }
3398
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003399 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3400 list_splice_tail(&list, &pgdata->split_queue);
3401 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003402
Kirill A. Shutemovcb8d68e2016-02-02 16:57:12 -08003403 /*
3404 * Stop shrinker if we didn't split any page, but the queue is empty.
3405 * This can happen if pages were freed under us.
3406 */
3407 if (!split && list_empty(&pgdata->split_queue))
3408 return SHRINK_STOP;
3409 return split;
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003410}
3411
3412static struct shrinker deferred_split_shrinker = {
3413 .count_objects = deferred_split_count,
3414 .scan_objects = deferred_split_scan,
3415 .seeks = DEFAULT_SEEKS,
Kirill A. Shutemova3d0a9182016-02-02 16:57:08 -08003416 .flags = SHRINKER_NUMA_AWARE,
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08003417};
Kirill A. Shutemov49071d42016-01-15 16:54:40 -08003418
3419#ifdef CONFIG_DEBUG_FS
3420static int split_huge_pages_set(void *data, u64 val)
3421{
3422 struct zone *zone;
3423 struct page *page;
3424 unsigned long pfn, max_zone_pfn;
3425 unsigned long total = 0, split = 0;
3426
3427 if (val != 1)
3428 return -EINVAL;
3429
3430 for_each_populated_zone(zone) {
3431 max_zone_pfn = zone_end_pfn(zone);
3432 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3433 if (!pfn_valid(pfn))
3434 continue;
3435
3436 page = pfn_to_page(pfn);
3437 if (!get_page_unless_zero(page))
3438 continue;
3439
3440 if (zone != page_zone(page))
3441 goto next;
3442
3443 if (!PageHead(page) || !PageAnon(page) ||
3444 PageHuge(page))
3445 goto next;
3446
3447 total++;
3448 lock_page(page);
3449 if (!split_huge_page(page))
3450 split++;
3451 unlock_page(page);
3452next:
3453 put_page(page);
3454 }
3455 }
3456
3457 pr_info("%lu of %lu THP split", split, total);
3458
3459 return 0;
3460}
3461DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3462 "%llu\n");
3463
3464static int __init split_huge_pages_debugfs(void)
3465{
3466 void *ret;
3467
3468 ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3469 &split_huge_pages_fops);
3470 if (!ret)
3471 pr_warn("Failed to create split_huge_pages in debugfs");
3472 return 0;
3473}
3474late_initcall(split_huge_pages_debugfs);
3475#endif