blob: 4b06b8db9df23c8f33406586507bbaecf7f5444c [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>
Matthew Wilcox4897c762015-09-08 14:58:45 -070019#include <linux/dax.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080020#include <linux/kthread.h>
21#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080022#include <linux/freezer.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080023#include <linux/mman.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020024#include <linux/pagemap.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000025#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080026#include <linux/hashtable.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070027#include <linux/userfaultfd_k.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070028#include <linux/page_idle.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080029
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080030#include <asm/tlb.h>
31#include <asm/pgalloc.h>
32#include "internal.h"
33
Andrea Arcangeliba761492011-01-13 15:46:58 -080034/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080035 * By default transparent hugepage support is disabled in order that avoid
36 * to risk increase the memory footprint of applications without a guaranteed
37 * benefit. When transparent hugepage support is enabled, is for all mappings,
38 * and khugepaged scans all mappings.
39 * Defrag is invoked by khugepaged hugepage allocations and by page faults
40 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080041 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080042unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080043#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080044 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080045#endif
46#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
47 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
48#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080049 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080050 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
51 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080052
53/* default scan 8*512 pte (or vmas) every 30 second */
54static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
55static unsigned int khugepaged_pages_collapsed;
56static unsigned int khugepaged_full_scans;
57static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58/* during fragmentation poll the hugepage allocator once every minute */
59static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60static struct task_struct *khugepaged_thread __read_mostly;
61static DEFINE_MUTEX(khugepaged_mutex);
62static DEFINE_SPINLOCK(khugepaged_mm_lock);
63static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
64/*
65 * default collapse hugepages if there is at least one pte mapped like
66 * it would have happened if the vma was large enough during page
67 * fault.
68 */
69static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
70
71static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -080072static int khugepaged_slab_init(void);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -070073static void khugepaged_slab_exit(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080074
Sasha Levin43b5fbb2013-02-22 16:32:27 -080075#define MM_SLOTS_HASH_BITS 10
76static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
77
Andrea Arcangeliba761492011-01-13 15:46:58 -080078static struct kmem_cache *mm_slot_cache __read_mostly;
79
80/**
81 * struct mm_slot - hash lookup from mm to mm_slot
82 * @hash: hash collision list
83 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
84 * @mm: the mm that this information is valid for
85 */
86struct mm_slot {
87 struct hlist_node hash;
88 struct list_head mm_node;
89 struct mm_struct *mm;
90};
91
92/**
93 * struct khugepaged_scan - cursor for scanning
94 * @mm_head: the head of the mm list to scan
95 * @mm_slot: the current mm_slot we are scanning
96 * @address: the next address inside that to be scanned
97 *
98 * There is only the one khugepaged_scan instance of this cursor structure.
99 */
100struct khugepaged_scan {
101 struct list_head mm_head;
102 struct mm_slot *mm_slot;
103 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -0700104};
105static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800106 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107};
108
Andrea Arcangelif0005652011-01-13 15:47:04 -0800109
Nicholas Krause2c0b80d2015-09-08 15:00:33 -0700110static void set_recommended_min_free_kbytes(void)
Andrea Arcangelif0005652011-01-13 15:47:04 -0800111{
112 struct zone *zone;
113 int nr_zones = 0;
114 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800115
Andrea Arcangelif0005652011-01-13 15:47:04 -0800116 for_each_populated_zone(zone)
117 nr_zones++;
118
119 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
120 recommended_min = pageblock_nr_pages * nr_zones * 2;
121
122 /*
123 * Make sure that on average at least two pageblocks are almost free
124 * of another type, one for a migratetype to fall back to and a
125 * second to avoid subsequent fallbacks of other types There are 3
126 * MIGRATE_TYPES we care about.
127 */
128 recommended_min += pageblock_nr_pages * nr_zones *
129 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
130
131 /* don't ever allow to reserve more than 5% of the lowmem */
132 recommended_min = min(recommended_min,
133 (unsigned long) nr_free_buffer_pages() / 20);
134 recommended_min <<= (PAGE_SHIFT-10);
135
Han Pingtian42aa83c2014-01-23 15:53:28 -0800136 if (recommended_min > min_free_kbytes) {
137 if (user_min_free_kbytes >= 0)
138 pr_info("raising min_free_kbytes from %d to %lu "
139 "to help transparent hugepage allocations\n",
140 min_free_kbytes, recommended_min);
141
Andrea Arcangelif0005652011-01-13 15:47:04 -0800142 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800143 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800144 setup_per_zone_wmarks();
Andrea Arcangelif0005652011-01-13 15:47:04 -0800145}
Andrea Arcangelif0005652011-01-13 15:47:04 -0800146
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700147static int start_stop_khugepaged(void)
Andrea Arcangeliba761492011-01-13 15:46:58 -0800148{
149 int err = 0;
150 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800151 if (!khugepaged_thread)
152 khugepaged_thread = kthread_run(khugepaged, NULL,
153 "khugepaged");
154 if (unlikely(IS_ERR(khugepaged_thread))) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700155 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
Andrea Arcangeliba761492011-01-13 15:46:58 -0800156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700158 goto fail;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800159 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700160
161 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800162 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800163
164 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700165 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700166 kthread_stop(khugepaged_thread);
167 khugepaged_thread = NULL;
168 }
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700169fail:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800170 return err;
171}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800172
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800173static atomic_t huge_zero_refcount;
Wang, Yalin56873f42015-02-11 15:24:51 -0800174struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800175
Matthew Wilcoxfc437042015-09-08 14:58:51 -0700176struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800177{
178 struct page *zero_page;
179retry:
180 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Jason Low4db0c3c2015-04-15 16:14:08 -0700181 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800182
183 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
184 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800185 if (!zero_page) {
186 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700187 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800188 }
189 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800190 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700191 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800192 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700193 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800194 goto retry;
195 }
196
197 /* We take additional reference here. It will be put back by shrinker */
198 atomic_set(&huge_zero_refcount, 2);
199 preempt_enable();
Jason Low4db0c3c2015-04-15 16:14:08 -0700200 return READ_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800201}
202
203static void put_huge_zero_page(void)
204{
205 /*
206 * Counter should never go to zero here. Only shrinker can put
207 * last reference.
208 */
209 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
210}
211
Glauber Costa48896462013-08-28 10:18:15 +1000212static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
213 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800214{
Glauber Costa48896462013-08-28 10:18:15 +1000215 /* we can free zero page only if last reference remains */
216 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
217}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800218
Glauber Costa48896462013-08-28 10:18:15 +1000219static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
220 struct shrink_control *sc)
221{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800222 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700223 struct page *zero_page = xchg(&huge_zero_page, NULL);
224 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700225 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000226 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800227 }
228
229 return 0;
230}
231
232static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000233 .count_objects = shrink_huge_zero_page_count,
234 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800235 .seeks = DEFAULT_SEEKS,
236};
237
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800238#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800239
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800240static ssize_t double_flag_show(struct kobject *kobj,
241 struct kobj_attribute *attr, char *buf,
242 enum transparent_hugepage_flag enabled,
243 enum transparent_hugepage_flag req_madv)
244{
245 if (test_bit(enabled, &transparent_hugepage_flags)) {
246 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
247 return sprintf(buf, "[always] madvise never\n");
248 } else if (test_bit(req_madv, &transparent_hugepage_flags))
249 return sprintf(buf, "always [madvise] never\n");
250 else
251 return sprintf(buf, "always madvise [never]\n");
252}
253static ssize_t double_flag_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count,
256 enum transparent_hugepage_flag enabled,
257 enum transparent_hugepage_flag req_madv)
258{
259 if (!memcmp("always", buf,
260 min(sizeof("always")-1, count))) {
261 set_bit(enabled, &transparent_hugepage_flags);
262 clear_bit(req_madv, &transparent_hugepage_flags);
263 } else if (!memcmp("madvise", buf,
264 min(sizeof("madvise")-1, count))) {
265 clear_bit(enabled, &transparent_hugepage_flags);
266 set_bit(req_madv, &transparent_hugepage_flags);
267 } else if (!memcmp("never", buf,
268 min(sizeof("never")-1, count))) {
269 clear_bit(enabled, &transparent_hugepage_flags);
270 clear_bit(req_madv, &transparent_hugepage_flags);
271 } else
272 return -EINVAL;
273
274 return count;
275}
276
277static ssize_t enabled_show(struct kobject *kobj,
278 struct kobj_attribute *attr, char *buf)
279{
280 return double_flag_show(kobj, attr, buf,
281 TRANSPARENT_HUGEPAGE_FLAG,
282 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
283}
284static ssize_t enabled_store(struct kobject *kobj,
285 struct kobj_attribute *attr,
286 const char *buf, size_t count)
287{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800288 ssize_t ret;
289
290 ret = double_flag_store(kobj, attr, buf, count,
291 TRANSPARENT_HUGEPAGE_FLAG,
292 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
293
294 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700295 int err;
296
297 mutex_lock(&khugepaged_mutex);
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700298 err = start_stop_khugepaged();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700299 mutex_unlock(&khugepaged_mutex);
300
Andrea Arcangeliba761492011-01-13 15:46:58 -0800301 if (err)
302 ret = err;
303 }
304
305 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800306}
307static struct kobj_attribute enabled_attr =
308 __ATTR(enabled, 0644, enabled_show, enabled_store);
309
310static ssize_t single_flag_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf,
312 enum transparent_hugepage_flag flag)
313{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700314 return sprintf(buf, "%d\n",
315 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800316}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700317
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800318static ssize_t single_flag_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count,
321 enum transparent_hugepage_flag flag)
322{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700323 unsigned long value;
324 int ret;
325
326 ret = kstrtoul(buf, 10, &value);
327 if (ret < 0)
328 return ret;
329 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800330 return -EINVAL;
331
Ben Hutchingse27e6152011-04-14 15:22:21 -0700332 if (value)
333 set_bit(flag, &transparent_hugepage_flags);
334 else
335 clear_bit(flag, &transparent_hugepage_flags);
336
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800337 return count;
338}
339
340/*
341 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
342 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
343 * memory just to allocate one more hugepage.
344 */
345static ssize_t defrag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf)
347{
348 return double_flag_show(kobj, attr, buf,
349 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
350 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
351}
352static ssize_t defrag_store(struct kobject *kobj,
353 struct kobj_attribute *attr,
354 const char *buf, size_t count)
355{
356 return double_flag_store(kobj, attr, buf, count,
357 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
358 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
359}
360static struct kobj_attribute defrag_attr =
361 __ATTR(defrag, 0644, defrag_show, defrag_store);
362
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800363static ssize_t use_zero_page_show(struct kobject *kobj,
364 struct kobj_attribute *attr, char *buf)
365{
366 return single_flag_show(kobj, attr, buf,
367 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
368}
369static ssize_t use_zero_page_store(struct kobject *kobj,
370 struct kobj_attribute *attr, const char *buf, size_t count)
371{
372 return single_flag_store(kobj, attr, buf, count,
373 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
374}
375static struct kobj_attribute use_zero_page_attr =
376 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800377#ifdef CONFIG_DEBUG_VM
378static ssize_t debug_cow_show(struct kobject *kobj,
379 struct kobj_attribute *attr, char *buf)
380{
381 return single_flag_show(kobj, attr, buf,
382 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
383}
384static ssize_t debug_cow_store(struct kobject *kobj,
385 struct kobj_attribute *attr,
386 const char *buf, size_t count)
387{
388 return single_flag_store(kobj, attr, buf, count,
389 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
390}
391static struct kobj_attribute debug_cow_attr =
392 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
393#endif /* CONFIG_DEBUG_VM */
394
395static struct attribute *hugepage_attr[] = {
396 &enabled_attr.attr,
397 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800398 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800399#ifdef CONFIG_DEBUG_VM
400 &debug_cow_attr.attr,
401#endif
402 NULL,
403};
404
405static struct attribute_group hugepage_attr_group = {
406 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800407};
408
409static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
410 struct kobj_attribute *attr,
411 char *buf)
412{
413 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
414}
415
416static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
417 struct kobj_attribute *attr,
418 const char *buf, size_t count)
419{
420 unsigned long msecs;
421 int err;
422
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700423 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800424 if (err || msecs > UINT_MAX)
425 return -EINVAL;
426
427 khugepaged_scan_sleep_millisecs = msecs;
428 wake_up_interruptible(&khugepaged_wait);
429
430 return count;
431}
432static struct kobj_attribute scan_sleep_millisecs_attr =
433 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
434 scan_sleep_millisecs_store);
435
436static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
437 struct kobj_attribute *attr,
438 char *buf)
439{
440 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
441}
442
443static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
444 struct kobj_attribute *attr,
445 const char *buf, size_t count)
446{
447 unsigned long msecs;
448 int err;
449
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700450 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800451 if (err || msecs > UINT_MAX)
452 return -EINVAL;
453
454 khugepaged_alloc_sleep_millisecs = msecs;
455 wake_up_interruptible(&khugepaged_wait);
456
457 return count;
458}
459static struct kobj_attribute alloc_sleep_millisecs_attr =
460 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
461 alloc_sleep_millisecs_store);
462
463static ssize_t pages_to_scan_show(struct kobject *kobj,
464 struct kobj_attribute *attr,
465 char *buf)
466{
467 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
468}
469static ssize_t pages_to_scan_store(struct kobject *kobj,
470 struct kobj_attribute *attr,
471 const char *buf, size_t count)
472{
473 int err;
474 unsigned long pages;
475
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700476 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800477 if (err || !pages || pages > UINT_MAX)
478 return -EINVAL;
479
480 khugepaged_pages_to_scan = pages;
481
482 return count;
483}
484static struct kobj_attribute pages_to_scan_attr =
485 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
486 pages_to_scan_store);
487
488static ssize_t pages_collapsed_show(struct kobject *kobj,
489 struct kobj_attribute *attr,
490 char *buf)
491{
492 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
493}
494static struct kobj_attribute pages_collapsed_attr =
495 __ATTR_RO(pages_collapsed);
496
497static ssize_t full_scans_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_full_scans);
502}
503static struct kobj_attribute full_scans_attr =
504 __ATTR_RO(full_scans);
505
506static ssize_t khugepaged_defrag_show(struct kobject *kobj,
507 struct kobj_attribute *attr, char *buf)
508{
509 return single_flag_show(kobj, attr, buf,
510 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
511}
512static ssize_t khugepaged_defrag_store(struct kobject *kobj,
513 struct kobj_attribute *attr,
514 const char *buf, size_t count)
515{
516 return single_flag_store(kobj, attr, buf, count,
517 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
518}
519static struct kobj_attribute khugepaged_defrag_attr =
520 __ATTR(defrag, 0644, khugepaged_defrag_show,
521 khugepaged_defrag_store);
522
523/*
524 * max_ptes_none controls if khugepaged should collapse hugepages over
525 * any unmapped ptes in turn potentially increasing the memory
526 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
527 * reduce the available free memory in the system as it
528 * runs. Increasing max_ptes_none will instead potentially reduce the
529 * free memory in the system during the khugepaged scan.
530 */
531static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
532 struct kobj_attribute *attr,
533 char *buf)
534{
535 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
536}
537static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
538 struct kobj_attribute *attr,
539 const char *buf, size_t count)
540{
541 int err;
542 unsigned long max_ptes_none;
543
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700544 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800545 if (err || max_ptes_none > HPAGE_PMD_NR-1)
546 return -EINVAL;
547
548 khugepaged_max_ptes_none = max_ptes_none;
549
550 return count;
551}
552static struct kobj_attribute khugepaged_max_ptes_none_attr =
553 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
554 khugepaged_max_ptes_none_store);
555
556static struct attribute *khugepaged_attr[] = {
557 &khugepaged_defrag_attr.attr,
558 &khugepaged_max_ptes_none_attr.attr,
559 &pages_to_scan_attr.attr,
560 &pages_collapsed_attr.attr,
561 &full_scans_attr.attr,
562 &scan_sleep_millisecs_attr.attr,
563 &alloc_sleep_millisecs_attr.attr,
564 NULL,
565};
566
567static struct attribute_group khugepaged_attr_group = {
568 .attrs = khugepaged_attr,
569 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800570};
Shaohua Li569e5592012-01-12 17:19:11 -0800571
572static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
573{
574 int err;
575
576 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
577 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700578 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800579 return -ENOMEM;
580 }
581
582 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
583 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700584 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800585 goto delete_obj;
586 }
587
588 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
589 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700590 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800591 goto remove_hp_group;
592 }
593
594 return 0;
595
596remove_hp_group:
597 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
598delete_obj:
599 kobject_put(*hugepage_kobj);
600 return err;
601}
602
603static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
604{
605 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
606 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
607 kobject_put(hugepage_kobj);
608}
609#else
610static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
611{
612 return 0;
613}
614
615static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
616{
617}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800618#endif /* CONFIG_SYSFS */
619
620static int __init hugepage_init(void)
621{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800622 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800623 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800624
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800625 if (!has_transparent_hugepage()) {
626 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800627 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800628 }
629
Shaohua Li569e5592012-01-12 17:19:11 -0800630 err = hugepage_init_sysfs(&hugepage_kobj);
631 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700632 goto err_sysfs;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800633
634 err = khugepaged_slab_init();
635 if (err)
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700636 goto err_slab;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800637
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700638 err = register_shrinker(&huge_zero_page_shrinker);
639 if (err)
640 goto err_hzp_shrinker;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800641
Rik van Riel97562cd2011-01-13 15:47:12 -0800642 /*
643 * By default disable transparent hugepages on smaller systems,
644 * where the extra memory used could hurt more than TLB overhead
645 * is likely to save. The admin can still enable it through /sys.
646 */
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700647 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
Rik van Riel97562cd2011-01-13 15:47:12 -0800648 transparent_hugepage_flags = 0;
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700649 return 0;
650 }
Rik van Riel97562cd2011-01-13 15:47:12 -0800651
Kirill A. Shutemov79553da2015-04-15 16:14:56 -0700652 err = start_stop_khugepaged();
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700653 if (err)
654 goto err_khugepaged;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800655
Shaohua Li569e5592012-01-12 17:19:11 -0800656 return 0;
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700657err_khugepaged:
658 unregister_shrinker(&huge_zero_page_shrinker);
659err_hzp_shrinker:
660 khugepaged_slab_exit();
661err_slab:
Shaohua Li569e5592012-01-12 17:19:11 -0800662 hugepage_exit_sysfs(hugepage_kobj);
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -0700663err_sysfs:
Andrea Arcangeliba761492011-01-13 15:46:58 -0800664 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800665}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800666subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800667
668static int __init setup_transparent_hugepage(char *str)
669{
670 int ret = 0;
671 if (!str)
672 goto out;
673 if (!strcmp(str, "always")) {
674 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
675 &transparent_hugepage_flags);
676 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
677 &transparent_hugepage_flags);
678 ret = 1;
679 } else if (!strcmp(str, "madvise")) {
680 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
681 &transparent_hugepage_flags);
682 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
683 &transparent_hugepage_flags);
684 ret = 1;
685 } else if (!strcmp(str, "never")) {
686 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
687 &transparent_hugepage_flags);
688 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
689 &transparent_hugepage_flags);
690 ret = 1;
691 }
692out:
693 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700694 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800695 return ret;
696}
697__setup("transparent_hugepage=", setup_transparent_hugepage);
698
Mel Gormanb32967f2012-11-19 12:35:47 +0000699pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800700{
701 if (likely(vma->vm_flags & VM_WRITE))
702 pmd = pmd_mkwrite(pmd);
703 return pmd;
704}
705
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700706static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800707{
708 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700709 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800710 entry = pmd_mkhuge(entry);
711 return entry;
712}
713
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800714static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
715 struct vm_area_struct *vma,
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700716 unsigned long address, pmd_t *pmd,
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700717 struct page *page, gfp_t gfp,
718 unsigned int flags)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800719{
Johannes Weiner00501b52014-08-08 14:19:20 -0700720 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800721 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800722 spinlock_t *ptl;
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700723 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800724
Sasha Levin309381fea2014-01-23 15:52:54 -0800725 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700726
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700727 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
728 put_page(page);
729 count_vm_event(THP_FAULT_FALLBACK);
730 return VM_FAULT_FALLBACK;
731 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800732
Johannes Weiner00501b52014-08-08 14:19:20 -0700733 pgtable = pte_alloc_one(mm, haddr);
734 if (unlikely(!pgtable)) {
735 mem_cgroup_cancel_charge(page, memcg);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700736 put_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700737 return VM_FAULT_OOM;
738 }
739
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800740 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700741 /*
742 * The memory barrier inside __SetPageUptodate makes sure that
743 * clear_huge_page writes become visible before the set_pmd_at()
744 * write.
745 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800746 __SetPageUptodate(page);
747
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800748 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800749 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800750 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -0700751 mem_cgroup_cancel_charge(page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800752 put_page(page);
753 pte_free(mm, pgtable);
754 } else {
755 pmd_t entry;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700756
757 /* Deliver the page fault to userland */
758 if (userfaultfd_missing(vma)) {
759 int ret;
760
761 spin_unlock(ptl);
762 mem_cgroup_cancel_charge(page, memcg);
763 put_page(page);
764 pte_free(mm, pgtable);
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700765 ret = handle_userfault(vma, address, flags,
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700766 VM_UFFD_MISSING);
767 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
768 return ret;
769 }
770
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700771 entry = mk_huge_pmd(page, vma->vm_page_prot);
772 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800773 page_add_new_anon_rmap(page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700774 mem_cgroup_commit_charge(page, memcg, false);
775 lru_cache_add_active_or_unevictable(page, vma);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700776 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800777 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800778 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800779 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800780 spin_unlock(ptl);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700781 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800782 }
783
David Rientjesaa2e8782012-05-29 15:06:17 -0700784 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800785}
786
Andi Kleencc5d4622011-03-22 16:33:13 -0700787static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800788{
Andi Kleencc5d4622011-03-22 16:33:13 -0700789 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800790}
791
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800792/* Caller must hold page table lock. */
Kirill A. Shutemovd295e342015-09-08 14:59:34 -0700793static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800794 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700795 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800796{
797 pmd_t entry;
Andrew Morton7c414162015-09-08 14:58:43 -0700798 if (!pmd_none(*pmd))
799 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700800 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800801 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700802 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800803 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800804 atomic_long_inc(&mm->nr_ptes);
Andrew Morton7c414162015-09-08 14:58:43 -0700805 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800806}
807
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800808int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
809 unsigned long address, pmd_t *pmd,
810 unsigned int flags)
811{
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800812 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800813 struct page *page;
814 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800815
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700816 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700817 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700818 if (unlikely(anon_vma_prepare(vma)))
819 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700820 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700821 return VM_FAULT_OOM;
Dominik Dingel593befa2014-10-23 12:07:44 +0200822 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700823 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800824 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700825 pgtable_t pgtable;
826 struct page *zero_page;
827 bool set;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700828 int ret;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700829 pgtable = pte_alloc_one(mm, haddr);
830 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800831 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700832 zero_page = get_huge_zero_page();
833 if (unlikely(!zero_page)) {
834 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700835 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700836 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700837 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800838 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700839 ret = 0;
840 set = false;
841 if (pmd_none(*pmd)) {
842 if (userfaultfd_missing(vma)) {
843 spin_unlock(ptl);
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 } else {
848 set_huge_zero_page(pgtable, mm, vma,
849 haddr, pmd,
850 zero_page);
851 spin_unlock(ptl);
852 set = true;
853 }
854 } else
855 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700856 if (!set) {
857 pte_free(mm, pgtable);
858 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800859 }
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700860 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800861 }
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800862 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
863 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700864 if (unlikely(!page)) {
865 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700866 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700867 }
Andrea Arcangeli230c92a2015-09-04 15:47:20 -0700868 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
869 flags);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800870}
871
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700872static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700873 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
874{
875 struct mm_struct *mm = vma->vm_mm;
876 pmd_t entry;
877 spinlock_t *ptl;
878
879 ptl = pmd_lock(mm, pmd);
880 if (pmd_none(*pmd)) {
881 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
882 if (write) {
883 entry = pmd_mkyoung(pmd_mkdirty(entry));
884 entry = maybe_pmd_mkwrite(entry, vma);
885 }
886 set_pmd_at(mm, addr, pmd, entry);
887 update_mmu_cache_pmd(vma, addr, pmd);
888 }
889 spin_unlock(ptl);
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700890}
891
892int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
893 pmd_t *pmd, unsigned long pfn, bool write)
894{
895 pgprot_t pgprot = vma->vm_page_prot;
896 /*
897 * If we had pmd_special, we could avoid all these restrictions,
898 * but we need to be consistent with PTEs and architectures that
899 * can't support a 'special' bit.
900 */
901 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
902 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
903 (VM_PFNMAP|VM_MIXEDMAP));
904 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
905 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
906
907 if (addr < vma->vm_start || addr >= vma->vm_end)
908 return VM_FAULT_SIGBUS;
909 if (track_pfn_insert(vma, &pgprot, pfn))
910 return VM_FAULT_SIGBUS;
Matthew Wilcoxae18d6d2015-09-08 14:59:14 -0700911 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
912 return VM_FAULT_NOPAGE;
Matthew Wilcox5cad4652015-09-08 14:58:54 -0700913}
914
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800915int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
917 struct vm_area_struct *vma)
918{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800919 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800920 struct page *src_page;
921 pmd_t pmd;
922 pgtable_t pgtable;
923 int ret;
924
925 ret = -ENOMEM;
926 pgtable = pte_alloc_one(dst_mm, addr);
927 if (unlikely(!pgtable))
928 goto out;
929
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800930 dst_ptl = pmd_lock(dst_mm, dst_pmd);
931 src_ptl = pmd_lockptr(src_mm, src_pmd);
932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800933
934 ret = -EAGAIN;
935 pmd = *src_pmd;
936 if (unlikely(!pmd_trans_huge(pmd))) {
937 pte_free(dst_mm, pgtable);
938 goto out_unlock;
939 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800940 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800941 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800942 * under splitting since we don't split the page itself, only pmd to
943 * a page table.
944 */
945 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700946 struct page *zero_page;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800947 /*
948 * get_huge_zero_page() will never allocate a new page here,
949 * since we already have a zero page to copy. It just takes a
950 * reference.
951 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700952 zero_page = get_huge_zero_page();
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -0700953 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700954 zero_page);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800955 ret = 0;
956 goto out_unlock;
957 }
Mel Gormande466bd2013-12-18 17:08:42 -0800958
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800959 if (unlikely(pmd_trans_splitting(pmd))) {
960 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800961 spin_unlock(src_ptl);
962 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800963 pte_free(dst_mm, pgtable);
964
965 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
966 goto out;
967 }
968 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800969 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800970 get_page(src_page);
971 page_dup_rmap(src_page);
972 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
973
974 pmdp_set_wrprotect(src_mm, addr, src_pmd);
975 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700976 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800977 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800978 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800979
980 ret = 0;
981out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800982 spin_unlock(src_ptl);
983 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800984out:
985 return ret;
986}
987
Will Deacona1dd4502012-12-11 16:01:27 -0800988void huge_pmd_set_accessed(struct mm_struct *mm,
989 struct vm_area_struct *vma,
990 unsigned long address,
991 pmd_t *pmd, pmd_t orig_pmd,
992 int dirty)
993{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800994 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800995 pmd_t entry;
996 unsigned long haddr;
997
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800998 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000 goto unlock;
1001
1002 entry = pmd_mkyoung(orig_pmd);
1003 haddr = address & HPAGE_PMD_MASK;
1004 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1005 update_mmu_cache_pmd(vma, address, pmd);
1006
1007unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001008 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -08001009}
1010
Hugh Dickins5338a932014-06-23 13:22:05 -07001011/*
1012 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1013 * during copy_user_huge_page()'s copy_page_rep(): in the case when
1014 * the source page gets split and a tail freed before copy completes.
1015 * Called under pmd_lock of checked pmd, so safe from splitting itself.
1016 */
1017static void get_user_huge_page(struct page *page)
1018{
1019 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1020 struct page *endpage = page + HPAGE_PMD_NR;
1021
1022 atomic_add(HPAGE_PMD_NR, &page->_count);
1023 while (++page < endpage)
1024 get_huge_page_tail(page);
1025 } else {
1026 get_page(page);
1027 }
1028}
1029
1030static void put_user_huge_page(struct page *page)
1031{
1032 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1033 struct page *endpage = page + HPAGE_PMD_NR;
1034
1035 while (page < endpage)
1036 put_page(page++);
1037 } else {
1038 put_page(page);
1039 }
1040}
1041
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001042static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1043 struct vm_area_struct *vma,
1044 unsigned long address,
1045 pmd_t *pmd, pmd_t orig_pmd,
1046 struct page *page,
1047 unsigned long haddr)
1048{
Johannes Weiner00501b52014-08-08 14:19:20 -07001049 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001050 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001051 pgtable_t pgtable;
1052 pmd_t _pmd;
1053 int ret = 0, i;
1054 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001055 unsigned long mmun_start; /* For mmu_notifiers */
1056 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001057
1058 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1059 GFP_KERNEL);
1060 if (unlikely(!pages)) {
1061 ret |= VM_FAULT_OOM;
1062 goto out;
1063 }
1064
1065 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001066 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1067 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001068 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001069 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -07001070 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1071 &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001072 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001073 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001074 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001075 memcg = (void *)page_private(pages[i]);
1076 set_page_private(pages[i], 0);
1077 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001078 put_page(pages[i]);
1079 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 kfree(pages);
1081 ret |= VM_FAULT_OOM;
1082 goto out;
1083 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001084 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001085 }
1086
1087 for (i = 0; i < HPAGE_PMD_NR; i++) {
1088 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001089 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001090 __SetPageUptodate(pages[i]);
1091 cond_resched();
1092 }
1093
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001094 mmun_start = haddr;
1095 mmun_end = haddr + HPAGE_PMD_SIZE;
1096 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1097
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001098 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001099 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1100 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001101 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001102
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001103 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001104 /* leave pmd empty until pte is filled */
1105
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001106 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001107 pmd_populate(mm, &_pmd, pgtable);
1108
1109 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1110 pte_t *pte, entry;
1111 entry = mk_pte(pages[i], vma->vm_page_prot);
1112 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001113 memcg = (void *)page_private(pages[i]);
1114 set_page_private(pages[i], 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001115 page_add_new_anon_rmap(pages[i], vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001116 mem_cgroup_commit_charge(pages[i], memcg, false);
1117 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001118 pte = pte_offset_map(&_pmd, haddr);
1119 VM_BUG_ON(!pte_none(*pte));
1120 set_pte_at(mm, haddr, pte, entry);
1121 pte_unmap(pte);
1122 }
1123 kfree(pages);
1124
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 smp_wmb(); /* make pte visible before pmd */
1126 pmd_populate(mm, pmd, pgtable);
1127 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001128 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001129
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001130 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1131
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001132 ret |= VM_FAULT_WRITE;
1133 put_page(page);
1134
1135out:
1136 return ret;
1137
1138out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001139 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001140 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001141 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001142 memcg = (void *)page_private(pages[i]);
1143 set_page_private(pages[i], 0);
1144 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001145 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001146 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001147 kfree(pages);
1148 goto out;
1149}
1150
1151int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1152 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1153{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001154 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001155 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001156 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001157 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001158 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001159 unsigned long mmun_start; /* For mmu_notifiers */
1160 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07001161 gfp_t huge_gfp; /* for allocation and charge */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001162
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001163 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001164 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001165 haddr = address & HPAGE_PMD_MASK;
1166 if (is_huge_zero_pmd(orig_pmd))
1167 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001168 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001169 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1170 goto out_unlock;
1171
1172 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001173 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001174 if (page_mapcount(page) == 1) {
1175 pmd_t entry;
1176 entry = pmd_mkyoung(orig_pmd);
1177 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1178 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001179 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001180 ret |= VM_FAULT_WRITE;
1181 goto out_unlock;
1182 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001183 get_user_huge_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001184 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001185alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001186 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001187 !transparent_hugepage_debug_cow()) {
Michal Hocko3b363692015-04-15 16:13:29 -07001188 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1189 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001190 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191 new_page = NULL;
1192
1193 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001194 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001195 split_huge_page_pmd(vma, address, pmd);
1196 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001197 } else {
1198 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1199 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001200 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001201 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001202 ret |= VM_FAULT_FALLBACK;
1203 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001204 put_user_huge_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001205 }
David Rientjes17766dd2013-09-12 15:14:06 -07001206 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001207 goto out;
1208 }
1209
Michal Hocko3b363692015-04-15 16:13:29 -07001210 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001211 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001212 if (page) {
1213 split_huge_page(page);
Hugh Dickins5338a932014-06-23 13:22:05 -07001214 put_user_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001215 } else
1216 split_huge_page_pmd(vma, address, pmd);
1217 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001218 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001219 goto out;
1220 }
1221
David Rientjes17766dd2013-09-12 15:14:06 -07001222 count_vm_event(THP_FAULT_ALLOC);
1223
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001224 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001225 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1226 else
1227 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001228 __SetPageUptodate(new_page);
1229
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001230 mmun_start = haddr;
1231 mmun_end = haddr + HPAGE_PMD_SIZE;
1232 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1233
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001234 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001235 if (page)
Hugh Dickins5338a932014-06-23 13:22:05 -07001236 put_user_huge_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001237 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001238 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -07001239 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001240 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001241 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001242 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001243 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001244 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1245 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001246 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001247 page_add_new_anon_rmap(new_page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001248 mem_cgroup_commit_charge(new_page, memcg, false);
1249 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001250 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001251 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001252 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001253 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001254 put_huge_zero_page();
1255 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001256 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001257 page_remove_rmap(page);
1258 put_page(page);
1259 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001260 ret |= VM_FAULT_WRITE;
1261 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001262 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001263out_mn:
1264 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1265out:
1266 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001267out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001268 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001269 return ret;
1270}
1271
David Rientjesb676b292012-10-08 16:34:03 -07001272struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001273 unsigned long addr,
1274 pmd_t *pmd,
1275 unsigned int flags)
1276{
David Rientjesb676b292012-10-08 16:34:03 -07001277 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001278 struct page *page = NULL;
1279
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001280 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001281
1282 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1283 goto out;
1284
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001285 /* Avoid dumping huge zero page */
1286 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1287 return ERR_PTR(-EFAULT);
1288
Mel Gorman2b4847e2013-12-18 17:08:32 -08001289 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001290 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001291 goto out;
1292
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001293 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001294 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001295 if (flags & FOLL_TOUCH) {
1296 pmd_t _pmd;
1297 /*
1298 * We should set the dirty bit only for FOLL_WRITE but
1299 * for now the dirty bit in the pmd is meaningless.
1300 * And if the dirty bit will become meaningful and
1301 * we'll only set it with FOLL_WRITE, an atomic
1302 * set_bit will be required on the pmd to set the
1303 * young bit, instead of the current set_pmd_at.
1304 */
1305 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001306 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1307 pmd, _pmd, 1))
1308 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001309 }
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -07001310 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
David Rientjesb676b292012-10-08 16:34:03 -07001311 if (page->mapping && trylock_page(page)) {
1312 lru_add_drain();
1313 if (page->mapping)
1314 mlock_vma_page(page);
1315 unlock_page(page);
1316 }
1317 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001318 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001319 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001320 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001321 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001322
1323out:
1324 return page;
1325}
1326
Mel Gormand10e63f2012-10-25 14:16:31 +02001327/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001328int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1329 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001330{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001331 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001332 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001333 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001334 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001335 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001336 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001337 bool page_locked;
1338 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001339 bool was_writable;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001340 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001341
Mel Gormanc0e7cad2015-02-12 14:58:41 -08001342 /* A PROT_NONE fault should not end up here */
1343 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1344
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001345 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001346 if (unlikely(!pmd_same(pmd, *pmdp)))
1347 goto out_unlock;
1348
Mel Gormande466bd2013-12-18 17:08:42 -08001349 /*
1350 * If there are potential migrations, wait for completion and retry
1351 * without disrupting NUMA hinting information. Do not relock and
1352 * check_same as the page may no longer be mapped.
1353 */
1354 if (unlikely(pmd_trans_migrating(*pmdp))) {
Mel Gorman5d833062015-02-12 14:58:16 -08001355 page = pmd_page(*pmdp);
Mel Gormande466bd2013-12-18 17:08:42 -08001356 spin_unlock(ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001357 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001358 goto out;
1359 }
1360
Mel Gormand10e63f2012-10-25 14:16:31 +02001361 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001362 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001363 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001364 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001365 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001366 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001367 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001368 flags |= TNF_FAULT_LOCAL;
1369 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001370
Mel Gormanbea66fb2015-03-25 15:55:37 -07001371 /* See similar comment in do_numa_page for explanation */
1372 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001373 flags |= TNF_NO_GROUP;
1374
1375 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001376 * Acquire the page lock to serialise THP migrations but avoid dropping
1377 * page_table_lock if at all possible
1378 */
Mel Gormanb8916632013-10-07 11:28:44 +01001379 page_locked = trylock_page(page);
1380 target_nid = mpol_misplaced(page, vma, haddr);
1381 if (target_nid == -1) {
1382 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001383 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001384 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001385 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001386
Mel Gormande466bd2013-12-18 17:08:42 -08001387 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001388 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001389 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001390 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001391 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001392 goto out;
1393 }
1394
Mel Gorman2b4847e2013-12-18 17:08:32 -08001395 /*
1396 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1397 * to serialises splits
1398 */
Mel Gormanb8916632013-10-07 11:28:44 +01001399 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001400 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001401 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001402
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001403 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001404 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001405 if (unlikely(!pmd_same(pmd, *pmdp))) {
1406 unlock_page(page);
1407 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001408 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001409 goto out_unlock;
1410 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001411
Mel Gormanc3a489c2013-12-18 17:08:38 -08001412 /* Bail if we fail to protect against THP splits for any reason */
1413 if (unlikely(!anon_vma)) {
1414 put_page(page);
1415 page_nid = -1;
1416 goto clear_pmdnuma;
1417 }
1418
Mel Gormana54a4072013-10-07 11:28:46 +01001419 /*
1420 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001421 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001422 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001423 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001424 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001425 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001426 if (migrated) {
1427 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001428 page_nid = target_nid;
Mel Gorman074c2382015-03-25 15:55:42 -07001429 } else
1430 flags |= TNF_MIGRATE_FAIL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001431
Mel Gorman8191acb2013-10-07 11:28:45 +01001432 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001433clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001434 BUG_ON(!PageLocked(page));
Mel Gormanb191f9b2015-03-25 15:55:40 -07001435 was_writable = pmd_write(pmd);
Mel Gorman4d942462015-02-12 14:58:28 -08001436 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormanb7b04002015-03-25 15:55:45 -07001437 pmd = pmd_mkyoung(pmd);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001438 if (was_writable)
1439 pmd = pmd_mkwrite(pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001440 set_pmd_at(mm, haddr, pmdp, pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001441 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001442 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001443out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001444 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001445
1446out:
1447 if (anon_vma)
1448 page_unlock_anon_vma_read(anon_vma);
1449
Mel Gorman8191acb2013-10-07 11:28:45 +01001450 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001451 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001452
Mel Gormand10e63f2012-10-25 14:16:31 +02001453 return 0;
1454}
1455
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001456int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001457 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001458{
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001459 pmd_t orig_pmd;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001460 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001461
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001462 if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1463 return 0;
1464 /*
1465 * For architectures like ppc64 we look at deposited pgtable
1466 * when calling pmdp_huge_get_and_clear. So do the
1467 * pgtable_trans_huge_withdraw after finishing pmdp related
1468 * operations.
1469 */
1470 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1471 tlb->fullmm);
1472 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1473 if (vma_is_dax(vma)) {
1474 spin_unlock(ptl);
1475 if (is_huge_zero_pmd(orig_pmd))
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001476 put_huge_zero_page();
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001477 } else if (is_huge_zero_pmd(orig_pmd)) {
1478 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1479 atomic_long_dec(&tlb->mm->nr_ptes);
1480 spin_unlock(ptl);
1481 put_huge_zero_page();
1482 } else {
1483 struct page *page = pmd_page(orig_pmd);
1484 page_remove_rmap(page);
1485 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1486 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1487 VM_BUG_ON_PAGE(!PageHead(page), page);
1488 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1489 atomic_long_dec(&tlb->mm->nr_ptes);
1490 spin_unlock(ptl);
1491 tlb_remove_page(tlb, page);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001492 }
Kirill A. Shutemovda146762015-09-08 14:59:31 -07001493 return 1;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001494}
1495
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001496int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1497 unsigned long old_addr,
1498 unsigned long new_addr, unsigned long old_end,
1499 pmd_t *old_pmd, pmd_t *new_pmd)
1500{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001501 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001502 int ret = 0;
1503 pmd_t pmd;
1504
1505 struct mm_struct *mm = vma->vm_mm;
1506
1507 if ((old_addr & ~HPAGE_PMD_MASK) ||
1508 (new_addr & ~HPAGE_PMD_MASK) ||
1509 old_end - old_addr < HPAGE_PMD_SIZE ||
1510 (new_vma->vm_flags & VM_NOHUGEPAGE))
1511 goto out;
1512
1513 /*
1514 * The destination pmd shouldn't be established, free_pgtables()
1515 * should have release it.
1516 */
1517 if (WARN_ON(!pmd_none(*new_pmd))) {
1518 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1519 goto out;
1520 }
1521
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001522 /*
1523 * We don't have to worry about the ordering of src and dst
1524 * ptlocks because exclusive mmap_sem prevents deadlock.
1525 */
1526 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001527 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001528 new_ptl = pmd_lockptr(mm, new_pmd);
1529 if (new_ptl != old_ptl)
1530 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001531 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001532 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001533
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301534 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1535 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001536 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1537 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001538 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301539 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1540 if (new_ptl != old_ptl)
1541 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001542 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001543 }
1544out:
1545 return ret;
1546}
1547
Mel Gormanf123d742013-10-07 11:28:49 +01001548/*
1549 * Returns
1550 * - 0 if PMD could not be locked
1551 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1552 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1553 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001554int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001555 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001556{
1557 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001558 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001559 int ret = 0;
1560
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001561 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001562 pmd_t entry;
Mel Gormanb191f9b2015-03-25 15:55:40 -07001563 bool preserve_write = prot_numa && pmd_write(*pmd);
Mel Gormanba68bc02015-03-07 15:20:48 +00001564 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001565
1566 /*
1567 * Avoid trapping faults against the zero page. The read-only
1568 * data is likely to be read-cached on the local CPU and
1569 * local/remote hits to the zero page are not interesting.
1570 */
1571 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1572 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001573 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001574 }
1575
Mel Gorman10c10452015-02-12 14:58:44 -08001576 if (!prot_numa || !pmd_protnone(*pmd)) {
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07001577 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
Mel Gorman10c10452015-02-12 14:58:44 -08001578 entry = pmd_modify(entry, newprot);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001579 if (preserve_write)
1580 entry = pmd_mkwrite(entry);
Mel Gorman10c10452015-02-12 14:58:44 -08001581 ret = HPAGE_PMD_NR;
1582 set_pmd_at(mm, addr, pmd, entry);
Mel Gormanb191f9b2015-03-25 15:55:40 -07001583 BUG_ON(!preserve_write && pmd_write(entry));
Mel Gorman10c10452015-02-12 14:58:44 -08001584 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001585 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001586 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001587
1588 return ret;
1589}
1590
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001591/*
1592 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1593 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1594 *
1595 * Note that if it returns 1, this routine returns without unlocking page
1596 * table locks. So callers must unlock them.
1597 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001598int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1599 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001600{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001601 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001602 if (likely(pmd_trans_huge(*pmd))) {
1603 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001604 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001605 wait_split_huge_page(vma->anon_vma, pmd);
1606 return -1;
1607 } else {
1608 /* Thp mapped by 'pmd' is stable, so we can
1609 * handle it as it is. */
1610 return 1;
1611 }
1612 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001613 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001614 return 0;
1615}
1616
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001617/*
1618 * This function returns whether a given @page is mapped onto the @address
1619 * in the virtual space of @mm.
1620 *
1621 * When it's true, this function returns *pmd with holding the page table lock
1622 * and passing it back to the caller via @ptl.
1623 * If it's false, returns NULL without holding the page table lock.
1624 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001625pmd_t *page_check_address_pmd(struct page *page,
1626 struct mm_struct *mm,
1627 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001628 enum page_check_address_pmd_flag flag,
1629 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001630{
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001631 pgd_t *pgd;
1632 pud_t *pud;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001633 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001634
1635 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001636 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001637
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001638 pgd = pgd_offset(mm, address);
1639 if (!pgd_present(*pgd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001640 return NULL;
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001641 pud = pud_offset(pgd, address);
1642 if (!pud_present(*pud))
1643 return NULL;
1644 pmd = pmd_offset(pud, address);
1645
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001646 *ptl = pmd_lock(mm, pmd);
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001647 if (!pmd_present(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001648 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001649 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001650 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001651 /*
1652 * split_vma() may create temporary aliased mappings. There is
1653 * no risk as long as all huge pmd are found and have their
1654 * splitting bit set before __split_huge_page_refcount
1655 * runs. Finding the same huge pmd more than once during the
1656 * same rmap walk is not a problem.
1657 */
1658 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1659 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001660 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001661 if (pmd_trans_huge(*pmd)) {
1662 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1663 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001664 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001665 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001666unlock:
1667 spin_unlock(*ptl);
1668 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001669}
1670
1671static int __split_huge_page_splitting(struct page *page,
1672 struct vm_area_struct *vma,
1673 unsigned long address)
1674{
1675 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001676 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001677 pmd_t *pmd;
1678 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001679 /* For mmu_notifiers */
1680 const unsigned long mmun_start = address;
1681 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001682
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001683 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001684 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001685 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001686 if (pmd) {
1687 /*
1688 * We can't temporarily set the pmd to null in order
1689 * to split it, the pmd must remain marked huge at all
1690 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001691 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001692 * serialize against split_huge_page*.
1693 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001694 pmdp_splitting_flush(vma, address, pmd);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001695
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001696 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001697 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001698 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001699 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001700
1701 return ret;
1702}
1703
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001704static void __split_huge_page_refcount(struct page *page,
1705 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001706{
1707 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001708 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001709 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001710 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001711
1712 /* prevent PageLRU to go away from under us, and freeze lru stats */
1713 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001714 lruvec = mem_cgroup_page_lruvec(page, zone);
1715
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001716 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001717 /* complete memcg works before add pages to LRU */
1718 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001719
Shaohua Li45676882012-01-12 17:19:18 -08001720 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001721 struct page *page_tail = page + i;
1722
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001723 /* tail_page->_mapcount cannot change */
1724 BUG_ON(page_mapcount(page_tail) < 0);
1725 tail_count += page_mapcount(page_tail);
1726 /* check for overflow */
1727 BUG_ON(tail_count < 0);
1728 BUG_ON(atomic_read(&page_tail->_count) != 0);
1729 /*
1730 * tail_page->_count is zero and not changing from
1731 * under us. But get_page_unless_zero() may be running
1732 * from under us on the tail_page. If we used
1733 * atomic_set() below instead of atomic_add(), we
1734 * would then run atomic_set() concurrently with
1735 * get_page_unless_zero(), and atomic_set() is
1736 * implemented in C not using locked ops. spin_unlock
1737 * on x86 sometime uses locked ops because of PPro
1738 * errata 66, 92, so unless somebody can guarantee
1739 * atomic_set() here would be safe on all archs (and
1740 * not only on x86), it's safer to use atomic_add().
1741 */
1742 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1743 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001744
1745 /* after clearing PageTail the gup refcount can be released */
Waiman Long3a79d522014-08-06 16:05:38 -07001746 smp_mb__after_atomic();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001747
Naoya Horiguchif4c18e62015-08-06 15:47:08 -07001748 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001749 page_tail->flags |= (page->flags &
1750 ((1L << PG_referenced) |
1751 (1L << PG_swapbacked) |
1752 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001753 (1L << PG_uptodate) |
1754 (1L << PG_active) |
1755 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001756 page_tail->flags |= (1L << PG_dirty);
1757
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001758 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001759 smp_wmb();
1760
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07001761 if (page_is_young(page))
1762 set_page_young(page_tail);
1763 if (page_is_idle(page))
1764 set_page_idle(page_tail);
1765
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001766 /*
1767 * __split_huge_page_splitting() already set the
1768 * splitting bit in all pmd that could map this
1769 * hugepage, that will ensure no CPU can alter the
1770 * mapcount on the head page. The mapcount is only
1771 * accounted in the head page and it has to be
1772 * transferred to all tail pages in the below code. So
1773 * for this code to be safe, the split the mapcount
1774 * can't change. But that doesn't mean userland can't
1775 * keep changing and reading the page contents while
1776 * we transfer the mapcount, so the pmd splitting
1777 * status is achieved setting a reserved bit in the
1778 * pmd, not by clearing the present bit.
1779 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001780 page_tail->_mapcount = page->_mapcount;
1781
1782 BUG_ON(page_tail->mapping);
1783 page_tail->mapping = page->mapping;
1784
Shaohua Li45676882012-01-12 17:19:18 -08001785 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001786 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001787
1788 BUG_ON(!PageAnon(page_tail));
1789 BUG_ON(!PageUptodate(page_tail));
1790 BUG_ON(!PageDirty(page_tail));
1791 BUG_ON(!PageSwapBacked(page_tail));
1792
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001793 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001794 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001795 atomic_sub(tail_count, &page->_count);
1796 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001797
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001798 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001799
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001800 ClearPageCompound(page);
1801 compound_unlock(page);
1802 spin_unlock_irq(&zone->lru_lock);
1803
1804 for (i = 1; i < HPAGE_PMD_NR; i++) {
1805 struct page *page_tail = page + i;
1806 BUG_ON(page_count(page_tail) <= 0);
1807 /*
1808 * Tail pages may be freed if there wasn't any mapping
1809 * like if add_to_swap() is running on a lru page that
1810 * had its mapping zapped. And freeing these pages
1811 * requires taking the lru_lock so we do the put_page
1812 * of the tail pages after the split is complete.
1813 */
1814 put_page(page_tail);
1815 }
1816
1817 /*
1818 * Only the head page (now become a regular page) is required
1819 * to be pinned by the caller.
1820 */
1821 BUG_ON(page_count(page) <= 0);
1822}
1823
1824static int __split_huge_page_map(struct page *page,
1825 struct vm_area_struct *vma,
1826 unsigned long address)
1827{
1828 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001829 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001830 pmd_t *pmd, _pmd;
1831 int ret = 0, i;
1832 pgtable_t pgtable;
1833 unsigned long haddr;
1834
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001835 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001836 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001837 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001838 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001839 pmd_populate(mm, &_pmd, pgtable);
Waiman Longf8303c22014-08-06 16:05:36 -07001840 if (pmd_write(*pmd))
1841 BUG_ON(page_mapcount(page) != 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001842
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001843 haddr = address;
1844 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001845 pte_t *pte, entry;
1846 BUG_ON(PageCompound(page+i));
Mel Gormanabc40bd2014-10-02 19:47:42 +01001847 /*
Mel Gorman8a0516e2015-02-12 14:58:22 -08001848 * Note that NUMA hinting access restrictions are not
1849 * transferred to avoid any possibility of altering
1850 * permissions across VMAs.
Mel Gormanabc40bd2014-10-02 19:47:42 +01001851 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001852 entry = mk_pte(page + i, vma->vm_page_prot);
1853 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1854 if (!pmd_write(*pmd))
1855 entry = pte_wrprotect(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001856 if (!pmd_young(*pmd))
1857 entry = pte_mkold(entry);
1858 pte = pte_offset_map(&_pmd, haddr);
1859 BUG_ON(!pte_none(*pte));
1860 set_pte_at(mm, haddr, pte, entry);
1861 pte_unmap(pte);
1862 }
1863
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001864 smp_wmb(); /* make pte visible before pmd */
1865 /*
1866 * Up to this point the pmd is present and huge and
1867 * userland has the whole access to the hugepage
1868 * during the split (which happens in place). If we
1869 * overwrite the pmd with the not-huge version
1870 * pointing to the pte here (which of course we could
1871 * if all CPUs were bug free), userland could trigger
1872 * a small page size TLB miss on the small sized TLB
1873 * while the hugepage TLB entry is still established
1874 * in the huge TLB. Some CPU doesn't like that. See
1875 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1876 * Erratum 383 on page 93. Intel should be safe but is
1877 * also warns that it's only safe if the permission
1878 * and cache attributes of the two entries loaded in
1879 * the two TLB is identical (which should be the case
1880 * here). But it is generally safer to never allow
1881 * small and huge TLB entries for the same virtual
1882 * address to be loaded simultaneously. So instead of
1883 * doing "pmd_populate(); flush_tlb_range();" we first
1884 * mark the current pmd notpresent (atomically because
1885 * here the pmd_trans_huge and pmd_trans_splitting
1886 * must remain set at all times on the pmd until the
1887 * split is complete for this pmd), then we flush the
1888 * SMP TLB and finally we write the non-huge version
1889 * of the pmd entry with pmd_populate.
1890 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001891 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001892 pmd_populate(mm, pmd, pgtable);
1893 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001894 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001895 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001896
1897 return ret;
1898}
1899
Ingo Molnar5a505082012-12-02 19:56:46 +00001900/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001901static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001902 struct anon_vma *anon_vma,
1903 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001904{
1905 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001906 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001907 struct anon_vma_chain *avc;
1908
1909 BUG_ON(!PageHead(page));
1910 BUG_ON(PageTail(page));
1911
1912 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001913 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001914 struct vm_area_struct *vma = avc->vma;
1915 unsigned long addr = vma_address(page, vma);
1916 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001917 mapcount += __split_huge_page_splitting(page, vma, addr);
1918 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001919 /*
1920 * It is critical that new vmas are added to the tail of the
1921 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1922 * and establishes a child pmd before
1923 * __split_huge_page_splitting() freezes the parent pmd (so if
1924 * we fail to prevent copy_huge_pmd() from running until the
1925 * whole __split_huge_page() is complete), we will still see
1926 * the newly established pmd of the child later during the
1927 * walk, to be able to set it as pmd_trans_splitting too.
1928 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001929 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001930 pr_err("mapcount %d page_mapcount %d\n",
1931 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001932 BUG();
1933 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001934
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001935 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001936
1937 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001938 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001939 struct vm_area_struct *vma = avc->vma;
1940 unsigned long addr = vma_address(page, vma);
1941 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001942 mapcount2 += __split_huge_page_map(page, vma, addr);
1943 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001944 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001945 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1946 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001947 BUG();
1948 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001949}
1950
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001951/*
1952 * Split a hugepage into normal pages. This doesn't change the position of head
1953 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1954 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1955 * from the hugepage.
1956 * Return 0 if the hugepage is split successfully otherwise return 1.
1957 */
1958int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001959{
1960 struct anon_vma *anon_vma;
1961 int ret = 1;
1962
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001963 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001964 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001965
1966 /*
1967 * The caller does not necessarily hold an mmap_sem that would prevent
1968 * the anon_vma disappearing so we first we take a reference to it
1969 * and then lock the anon_vma for write. This is similar to
1970 * page_lock_anon_vma_read except the write lock is taken to serialise
1971 * against parallel split or collapse operations.
1972 */
1973 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001974 if (!anon_vma)
1975 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001976 anon_vma_lock_write(anon_vma);
1977
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001978 ret = 0;
1979 if (!PageCompound(page))
1980 goto out_unlock;
1981
1982 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001983 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001984 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001985
1986 BUG_ON(PageCompound(page));
1987out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001988 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001989 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001990out:
1991 return ret;
1992}
1993
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001994#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001995
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001996int hugepage_madvise(struct vm_area_struct *vma,
1997 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001998{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001999 switch (advice) {
2000 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07002001#ifdef CONFIG_S390
2002 /*
2003 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2004 * can't handle this properly after s390_enable_sie, so we simply
2005 * ignore the madvise to prevent qemu from causing a SIGSEGV.
2006 */
2007 if (mm_has_pgste(vma->vm_mm))
2008 return 0;
2009#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002010 /*
2011 * Be somewhat over-protective like KSM for now!
2012 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002013 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002014 return -EINVAL;
2015 *vm_flags &= ~VM_NOHUGEPAGE;
2016 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08002017 /*
2018 * If the vma become good for khugepaged to scan,
2019 * register it here without waiting a page fault that
2020 * may not happen any time soon.
2021 */
David Rientjes6d50e602014-10-29 14:50:31 -07002022 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08002023 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002024 break;
2025 case MADV_NOHUGEPAGE:
2026 /*
2027 * Be somewhat over-protective like KSM for now!
2028 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002029 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002030 return -EINVAL;
2031 *vm_flags &= ~VM_HUGEPAGE;
2032 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08002033 /*
2034 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2035 * this vma even if we leave the mm registered in khugepaged if
2036 * it got registered before VM_NOHUGEPAGE was set.
2037 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002038 break;
2039 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08002040
2041 return 0;
2042}
2043
Andrea Arcangeliba761492011-01-13 15:46:58 -08002044static int __init khugepaged_slab_init(void)
2045{
2046 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2047 sizeof(struct mm_slot),
2048 __alignof__(struct mm_slot), 0, NULL);
2049 if (!mm_slot_cache)
2050 return -ENOMEM;
2051
2052 return 0;
2053}
2054
Kirill A. Shutemov65ebb642015-04-15 16:14:20 -07002055static void __init khugepaged_slab_exit(void)
2056{
2057 kmem_cache_destroy(mm_slot_cache);
2058}
2059
Andrea Arcangeliba761492011-01-13 15:46:58 -08002060static inline struct mm_slot *alloc_mm_slot(void)
2061{
2062 if (!mm_slot_cache) /* initialization failed */
2063 return NULL;
2064 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2065}
2066
2067static inline void free_mm_slot(struct mm_slot *mm_slot)
2068{
2069 kmem_cache_free(mm_slot_cache, mm_slot);
2070}
2071
Andrea Arcangeliba761492011-01-13 15:46:58 -08002072static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2073{
2074 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002075
Sasha Levinb67bfe02013-02-27 17:06:00 -08002076 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077 if (mm == mm_slot->mm)
2078 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002079
Andrea Arcangeliba761492011-01-13 15:46:58 -08002080 return NULL;
2081}
2082
2083static void insert_to_mm_slots_hash(struct mm_struct *mm,
2084 struct mm_slot *mm_slot)
2085{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002086 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002087 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002088}
2089
2090static inline int khugepaged_test_exit(struct mm_struct *mm)
2091{
2092 return atomic_read(&mm->mm_users) == 0;
2093}
2094
2095int __khugepaged_enter(struct mm_struct *mm)
2096{
2097 struct mm_slot *mm_slot;
2098 int wakeup;
2099
2100 mm_slot = alloc_mm_slot();
2101 if (!mm_slot)
2102 return -ENOMEM;
2103
2104 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07002105 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002106 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2107 free_mm_slot(mm_slot);
2108 return 0;
2109 }
2110
2111 spin_lock(&khugepaged_mm_lock);
2112 insert_to_mm_slots_hash(mm, mm_slot);
2113 /*
2114 * Insert just behind the scanning cursor, to let the area settle
2115 * down a little.
2116 */
2117 wakeup = list_empty(&khugepaged_scan.mm_head);
2118 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2119 spin_unlock(&khugepaged_mm_lock);
2120
2121 atomic_inc(&mm->mm_count);
2122 if (wakeup)
2123 wake_up_interruptible(&khugepaged_wait);
2124
2125 return 0;
2126}
2127
David Rientjes6d50e602014-10-29 14:50:31 -07002128int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2129 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002130{
2131 unsigned long hstart, hend;
2132 if (!vma->anon_vma)
2133 /*
2134 * Not yet faulted in so we will register later in the
2135 * page fault if needed.
2136 */
2137 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002138 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002139 /* khugepaged not yet working on file or special mappings */
2140 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07002141 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002142 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2143 hend = vma->vm_end & HPAGE_PMD_MASK;
2144 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07002145 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002146 return 0;
2147}
2148
2149void __khugepaged_exit(struct mm_struct *mm)
2150{
2151 struct mm_slot *mm_slot;
2152 int free = 0;
2153
2154 spin_lock(&khugepaged_mm_lock);
2155 mm_slot = get_mm_slot(mm);
2156 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002157 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002158 list_del(&mm_slot->mm_node);
2159 free = 1;
2160 }
Chris Wrightd788e802011-07-25 17:12:14 -07002161 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002162
2163 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002164 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2165 free_mm_slot(mm_slot);
2166 mmdrop(mm);
2167 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002168 /*
2169 * This is required to serialize against
2170 * khugepaged_test_exit() (which is guaranteed to run
2171 * under mmap sem read mode). Stop here (after we
2172 * return all pagetables will be destroyed) until
2173 * khugepaged has finished working on the pagetables
2174 * under the mmap_sem.
2175 */
2176 down_write(&mm->mmap_sem);
2177 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002178 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002179}
2180
2181static void release_pte_page(struct page *page)
2182{
2183 /* 0 stands for page_is_file_cache(page) == false */
2184 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2185 unlock_page(page);
2186 putback_lru_page(page);
2187}
2188
2189static void release_pte_pages(pte_t *pte, pte_t *_pte)
2190{
2191 while (--_pte >= pte) {
2192 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002193 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002194 release_pte_page(pte_page(pteval));
2195 }
2196}
2197
Andrea Arcangeliba761492011-01-13 15:46:58 -08002198static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2199 unsigned long address,
2200 pte_t *pte)
2201{
2202 struct page *page;
2203 pte_t *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002204 int none_or_zero = 0;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002205 bool referenced = false, writable = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002206 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2207 _pte++, address += PAGE_SIZE) {
2208 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002209 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangelic1294d02015-09-04 15:46:27 -07002210 if (!userfaultfd_armed(vma) &&
2211 ++none_or_zero <= khugepaged_max_ptes_none)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002212 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002213 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002214 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002215 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002216 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002217 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002218 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002219 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002220 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002221
Sasha Levin309381fea2014-01-23 15:52:54 -08002222 VM_BUG_ON_PAGE(PageCompound(page), page);
2223 VM_BUG_ON_PAGE(!PageAnon(page), page);
2224 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002225
Andrea Arcangeliba761492011-01-13 15:46:58 -08002226 /*
2227 * We can do it before isolate_lru_page because the
2228 * page can't be freed from under us. NOTE: PG_lock
2229 * is needed to serialize against split_huge_page
2230 * when invoked from the VM.
2231 */
Bob Liu344aa352012-12-11 16:00:34 -08002232 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002233 goto out;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002234
2235 /*
2236 * cannot use mapcount: can't collapse if there's a gup pin.
2237 * The page must only be referenced by the scanned process
2238 * and page swap cache.
2239 */
2240 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2241 unlock_page(page);
2242 goto out;
2243 }
2244 if (pte_write(pteval)) {
2245 writable = true;
2246 } else {
2247 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2248 unlock_page(page);
2249 goto out;
2250 }
2251 /*
2252 * Page is not in the swap cache. It can be collapsed
2253 * into a THP.
2254 */
2255 }
2256
Andrea Arcangeliba761492011-01-13 15:46:58 -08002257 /*
2258 * Isolate the page to avoid collapsing an hugepage
2259 * currently in use by the VM.
2260 */
2261 if (isolate_lru_page(page)) {
2262 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002263 goto out;
2264 }
2265 /* 0 stands for page_is_file_cache(page) == false */
2266 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002267 VM_BUG_ON_PAGE(!PageLocked(page), page);
2268 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002269
2270 /* If there is no mapped pte young don't collapse the page */
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07002271 if (pte_young(pteval) ||
2272 page_is_young(page) || PageReferenced(page) ||
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002273 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002274 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002275 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002276 if (likely(referenced && writable))
Bob Liu344aa352012-12-11 16:00:34 -08002277 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002278out:
Bob Liu344aa352012-12-11 16:00:34 -08002279 release_pte_pages(pte, _pte);
2280 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002281}
2282
2283static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2284 struct vm_area_struct *vma,
2285 unsigned long address,
2286 spinlock_t *ptl)
2287{
2288 pte_t *_pte;
2289 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2290 pte_t pteval = *_pte;
2291 struct page *src_page;
2292
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002293 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002294 clear_user_highpage(page, address);
2295 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002296 if (is_zero_pfn(pte_pfn(pteval))) {
2297 /*
2298 * ptl mostly unnecessary.
2299 */
2300 spin_lock(ptl);
2301 /*
2302 * paravirt calls inside pte_clear here are
2303 * superfluous.
2304 */
2305 pte_clear(vma->vm_mm, address, _pte);
2306 spin_unlock(ptl);
2307 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002308 } else {
2309 src_page = pte_page(pteval);
2310 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002311 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002312 release_pte_page(src_page);
2313 /*
2314 * ptl mostly unnecessary, but preempt has to
2315 * be disabled to update the per-cpu stats
2316 * inside page_remove_rmap().
2317 */
2318 spin_lock(ptl);
2319 /*
2320 * paravirt calls inside pte_clear here are
2321 * superfluous.
2322 */
2323 pte_clear(vma->vm_mm, address, _pte);
2324 page_remove_rmap(src_page);
2325 spin_unlock(ptl);
2326 free_page_and_swap_cache(src_page);
2327 }
2328
2329 address += PAGE_SIZE;
2330 page++;
2331 }
2332}
2333
Xiao Guangrong26234f32012-10-08 16:29:51 -07002334static void khugepaged_alloc_sleep(void)
2335{
Petr Mladekbde43c62015-09-08 15:04:05 -07002336 DEFINE_WAIT(wait);
2337
2338 add_wait_queue(&khugepaged_wait, &wait);
2339 freezable_schedule_timeout_interruptible(
2340 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2341 remove_wait_queue(&khugepaged_wait, &wait);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002342}
2343
Bob Liu9f1b8682013-11-12 15:07:37 -08002344static int khugepaged_node_load[MAX_NUMNODES];
2345
David Rientjes14a4e212014-08-06 16:07:29 -07002346static bool khugepaged_scan_abort(int nid)
2347{
2348 int i;
2349
2350 /*
2351 * If zone_reclaim_mode is disabled, then no extra effort is made to
2352 * allocate memory locally.
2353 */
2354 if (!zone_reclaim_mode)
2355 return false;
2356
2357 /* If there is a count for this node already, it must be acceptable */
2358 if (khugepaged_node_load[nid])
2359 return false;
2360
2361 for (i = 0; i < MAX_NUMNODES; i++) {
2362 if (!khugepaged_node_load[i])
2363 continue;
2364 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2365 return true;
2366 }
2367 return false;
2368}
2369
Xiao Guangrong26234f32012-10-08 16:29:51 -07002370#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002371static int khugepaged_find_target_node(void)
2372{
2373 static int last_khugepaged_target_node = NUMA_NO_NODE;
2374 int nid, target_node = 0, max_value = 0;
2375
2376 /* find first node with max normal pages hit */
2377 for (nid = 0; nid < MAX_NUMNODES; nid++)
2378 if (khugepaged_node_load[nid] > max_value) {
2379 max_value = khugepaged_node_load[nid];
2380 target_node = nid;
2381 }
2382
2383 /* do some balance if several nodes have the same hit record */
2384 if (target_node <= last_khugepaged_target_node)
2385 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2386 nid++)
2387 if (max_value == khugepaged_node_load[nid]) {
2388 target_node = nid;
2389 break;
2390 }
2391
2392 last_khugepaged_target_node = target_node;
2393 return target_node;
2394}
2395
Xiao Guangrong26234f32012-10-08 16:29:51 -07002396static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2397{
2398 if (IS_ERR(*hpage)) {
2399 if (!*wait)
2400 return false;
2401
2402 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002403 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002404 khugepaged_alloc_sleep();
2405 } else if (*hpage) {
2406 put_page(*hpage);
2407 *hpage = NULL;
2408 }
2409
2410 return true;
2411}
2412
Michal Hocko3b363692015-04-15 16:13:29 -07002413static struct page *
2414khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002415 struct vm_area_struct *vma, unsigned long address,
2416 int node)
2417{
Sasha Levin309381fea2014-01-23 15:52:54 -08002418 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002419
Xiao Guangrong26234f32012-10-08 16:29:51 -07002420 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002421 * Before allocating the hugepage, release the mmap_sem read lock.
2422 * The allocation can take potentially a long time if it involves
2423 * sync compaction, and we do not need to hold the mmap_sem during
2424 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002425 */
2426 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002427
Vlastimil Babka96db8002015-09-08 15:03:50 -07002428 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002429 if (unlikely(!*hpage)) {
2430 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2431 *hpage = ERR_PTR(-ENOMEM);
2432 return NULL;
2433 }
2434
2435 count_vm_event(THP_COLLAPSE_ALLOC);
2436 return *hpage;
2437}
2438#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002439static int khugepaged_find_target_node(void)
2440{
2441 return 0;
2442}
2443
Bob Liu10dc4152013-11-12 15:07:35 -08002444static inline struct page *alloc_hugepage(int defrag)
2445{
2446 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2447 HPAGE_PMD_ORDER);
2448}
2449
Xiao Guangrong26234f32012-10-08 16:29:51 -07002450static struct page *khugepaged_alloc_hugepage(bool *wait)
2451{
2452 struct page *hpage;
2453
2454 do {
2455 hpage = alloc_hugepage(khugepaged_defrag());
2456 if (!hpage) {
2457 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2458 if (!*wait)
2459 return NULL;
2460
2461 *wait = false;
2462 khugepaged_alloc_sleep();
2463 } else
2464 count_vm_event(THP_COLLAPSE_ALLOC);
2465 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2466
2467 return hpage;
2468}
2469
2470static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2471{
2472 if (!*hpage)
2473 *hpage = khugepaged_alloc_hugepage(wait);
2474
2475 if (unlikely(!*hpage))
2476 return false;
2477
2478 return true;
2479}
2480
Michal Hocko3b363692015-04-15 16:13:29 -07002481static struct page *
2482khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002483 struct vm_area_struct *vma, unsigned long address,
2484 int node)
2485{
2486 up_read(&mm->mmap_sem);
2487 VM_BUG_ON(!*hpage);
Michal Hocko3b363692015-04-15 16:13:29 -07002488
Xiao Guangrong26234f32012-10-08 16:29:51 -07002489 return *hpage;
2490}
2491#endif
2492
Bob Liufa475e52012-12-11 16:00:39 -08002493static bool hugepage_vma_check(struct vm_area_struct *vma)
2494{
2495 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2496 (vma->vm_flags & VM_NOHUGEPAGE))
2497 return false;
2498
2499 if (!vma->anon_vma || vma->vm_ops)
2500 return false;
2501 if (is_vma_temporary_stack(vma))
2502 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002503 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002504 return true;
2505}
2506
Andrea Arcangeliba761492011-01-13 15:46:58 -08002507static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002508 unsigned long address,
2509 struct page **hpage,
2510 struct vm_area_struct *vma,
2511 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002512{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002513 pmd_t *pmd, _pmd;
2514 pte_t *pte;
2515 pgtable_t pgtable;
2516 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002517 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002518 int isolated;
2519 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002520 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002521 unsigned long mmun_start; /* For mmu_notifiers */
2522 unsigned long mmun_end; /* For mmu_notifiers */
Michal Hocko3b363692015-04-15 16:13:29 -07002523 gfp_t gfp;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002524
2525 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002526
Michal Hocko3b363692015-04-15 16:13:29 -07002527 /* Only allocate from the target node */
2528 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2529 __GFP_THISNODE;
2530
Xiao Guangrong26234f32012-10-08 16:29:51 -07002531 /* release the mmap_sem read lock. */
Michal Hocko3b363692015-04-15 16:13:29 -07002532 new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002533 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002534 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002535
Johannes Weiner00501b52014-08-08 14:19:20 -07002536 if (unlikely(mem_cgroup_try_charge(new_page, mm,
Michal Hocko3b363692015-04-15 16:13:29 -07002537 gfp, &memcg)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002538 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002539
2540 /*
2541 * Prevent all access to pagetables with the exception of
2542 * gup_fast later hanlded by the ptep_clear_flush and the VM
2543 * handled by the anon_vma lock + PG_lock.
2544 */
2545 down_write(&mm->mmap_sem);
2546 if (unlikely(khugepaged_test_exit(mm)))
2547 goto out;
2548
2549 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002550 if (!vma)
2551 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002552 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2553 hend = vma->vm_end & HPAGE_PMD_MASK;
2554 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2555 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002556 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002557 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002558 pmd = mm_find_pmd(mm, address);
2559 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002560 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002561
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002562 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002563
2564 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002565 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002566
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002567 mmun_start = address;
2568 mmun_end = address + HPAGE_PMD_SIZE;
2569 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002570 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002571 /*
2572 * After this gup_fast can't run anymore. This also removes
2573 * any huge TLB entry from the CPU so we won't allow
2574 * huge and small TLB entries for the same virtual address
2575 * to avoid the risk of CPU bugs in that area.
2576 */
Aneesh Kumar K.V15a25b22015-06-24 16:57:39 -07002577 _pmd = pmdp_collapse_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002578 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002579 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002580
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002581 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002582 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002583 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002584
2585 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002586 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002587 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002588 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002589 /*
2590 * We can only use set_pmd_at when establishing
2591 * hugepmds and never for establishing regular pmds that
2592 * points to regular pagetables. Use pmd_populate for that
2593 */
2594 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002595 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002596 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002597 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002598 }
2599
2600 /*
2601 * All pages are isolated and locked so anon_vma rmap
2602 * can't run anymore.
2603 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002604 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002605
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002606 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002607 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002608 __SetPageUptodate(new_page);
2609 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002610
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002611 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2612 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002613
2614 /*
2615 * spin_lock() below is not the equivalent of smp_wmb(), so
2616 * this is needed to avoid the copy_huge_page writes to become
2617 * visible after the set_pmd_at() write.
2618 */
2619 smp_wmb();
2620
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002621 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002622 BUG_ON(!pmd_none(*pmd));
2623 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002624 mem_cgroup_commit_charge(new_page, memcg, false);
2625 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002626 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002627 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002628 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002629 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002630
2631 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002632
Andrea Arcangeliba761492011-01-13 15:46:58 -08002633 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002634out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002635 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002636 return;
2637
Andrea Arcangelice83d212011-01-13 15:47:06 -08002638out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002639 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002640 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002641}
2642
2643static int khugepaged_scan_pmd(struct mm_struct *mm,
2644 struct vm_area_struct *vma,
2645 unsigned long address,
2646 struct page **hpage)
2647{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002648 pmd_t *pmd;
2649 pte_t *pte, *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002650 int ret = 0, none_or_zero = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002651 struct page *page;
2652 unsigned long _address;
2653 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002654 int node = NUMA_NO_NODE;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002655 bool writable = false, referenced = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002656
2657 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2658
Bob Liu62190492012-12-11 16:00:37 -08002659 pmd = mm_find_pmd(mm, address);
2660 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002661 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002662
Bob Liu9f1b8682013-11-12 15:07:37 -08002663 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002664 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2665 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2666 _pte++, _address += PAGE_SIZE) {
2667 pte_t pteval = *_pte;
Ebru Akagunduzca0984c2015-04-14 15:45:24 -07002668 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
Andrea Arcangelic1294d02015-09-04 15:46:27 -07002669 if (!userfaultfd_armed(vma) &&
2670 ++none_or_zero <= khugepaged_max_ptes_none)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002671 continue;
2672 else
2673 goto out_unmap;
2674 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002675 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002676 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002677 if (pte_write(pteval))
2678 writable = true;
2679
Andrea Arcangeliba761492011-01-13 15:46:58 -08002680 page = vm_normal_page(vma, _address, pteval);
2681 if (unlikely(!page))
2682 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002683 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002684 * Record which node the original page is from and save this
2685 * information to khugepaged_node_load[].
2686 * Khupaged will allocate hugepage from the node has the max
2687 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002688 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002689 node = page_to_nid(page);
David Rientjes14a4e212014-08-06 16:07:29 -07002690 if (khugepaged_scan_abort(node))
2691 goto out_unmap;
Bob Liu9f1b8682013-11-12 15:07:37 -08002692 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002693 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002694 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2695 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002696 /*
2697 * cannot use mapcount: can't collapse if there's a gup pin.
2698 * The page must only be referenced by the scanned process
2699 * and page swap cache.
2700 */
2701 if (page_count(page) != 1 + !!PageSwapCache(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002702 goto out_unmap;
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07002703 if (pte_young(pteval) ||
2704 page_is_young(page) || PageReferenced(page) ||
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002705 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002706 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002707 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002708 if (referenced && writable)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002709 ret = 1;
2710out_unmap:
2711 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002712 if (ret) {
2713 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002714 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002715 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002716 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002717out:
2718 return ret;
2719}
2720
2721static void collect_mm_slot(struct mm_slot *mm_slot)
2722{
2723 struct mm_struct *mm = mm_slot->mm;
2724
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002725 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002726
2727 if (khugepaged_test_exit(mm)) {
2728 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002729 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002730 list_del(&mm_slot->mm_node);
2731
2732 /*
2733 * Not strictly needed because the mm exited already.
2734 *
2735 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2736 */
2737
2738 /* khugepaged_mm_lock actually not necessary for the below */
2739 free_mm_slot(mm_slot);
2740 mmdrop(mm);
2741 }
2742}
2743
2744static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2745 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002746 __releases(&khugepaged_mm_lock)
2747 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002748{
2749 struct mm_slot *mm_slot;
2750 struct mm_struct *mm;
2751 struct vm_area_struct *vma;
2752 int progress = 0;
2753
2754 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002755 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002756
2757 if (khugepaged_scan.mm_slot)
2758 mm_slot = khugepaged_scan.mm_slot;
2759 else {
2760 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2761 struct mm_slot, mm_node);
2762 khugepaged_scan.address = 0;
2763 khugepaged_scan.mm_slot = mm_slot;
2764 }
2765 spin_unlock(&khugepaged_mm_lock);
2766
2767 mm = mm_slot->mm;
2768 down_read(&mm->mmap_sem);
2769 if (unlikely(khugepaged_test_exit(mm)))
2770 vma = NULL;
2771 else
2772 vma = find_vma(mm, khugepaged_scan.address);
2773
2774 progress++;
2775 for (; vma; vma = vma->vm_next) {
2776 unsigned long hstart, hend;
2777
2778 cond_resched();
2779 if (unlikely(khugepaged_test_exit(mm))) {
2780 progress++;
2781 break;
2782 }
Bob Liufa475e52012-12-11 16:00:39 -08002783 if (!hugepage_vma_check(vma)) {
2784skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002785 progress++;
2786 continue;
2787 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002788 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2789 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002790 if (hstart >= hend)
2791 goto skip;
2792 if (khugepaged_scan.address > hend)
2793 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002794 if (khugepaged_scan.address < hstart)
2795 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002796 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002797
2798 while (khugepaged_scan.address < hend) {
2799 int ret;
2800 cond_resched();
2801 if (unlikely(khugepaged_test_exit(mm)))
2802 goto breakouterloop;
2803
2804 VM_BUG_ON(khugepaged_scan.address < hstart ||
2805 khugepaged_scan.address + HPAGE_PMD_SIZE >
2806 hend);
2807 ret = khugepaged_scan_pmd(mm, vma,
2808 khugepaged_scan.address,
2809 hpage);
2810 /* move to next address */
2811 khugepaged_scan.address += HPAGE_PMD_SIZE;
2812 progress += HPAGE_PMD_NR;
2813 if (ret)
2814 /* we released mmap_sem so break loop */
2815 goto breakouterloop_mmap_sem;
2816 if (progress >= pages)
2817 goto breakouterloop;
2818 }
2819 }
2820breakouterloop:
2821 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2822breakouterloop_mmap_sem:
2823
2824 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002825 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002826 /*
2827 * Release the current mm_slot if this mm is about to die, or
2828 * if we scanned all vmas of this mm.
2829 */
2830 if (khugepaged_test_exit(mm) || !vma) {
2831 /*
2832 * Make sure that if mm_users is reaching zero while
2833 * khugepaged runs here, khugepaged_exit will find
2834 * mm_slot not pointing to the exiting mm.
2835 */
2836 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2837 khugepaged_scan.mm_slot = list_entry(
2838 mm_slot->mm_node.next,
2839 struct mm_slot, mm_node);
2840 khugepaged_scan.address = 0;
2841 } else {
2842 khugepaged_scan.mm_slot = NULL;
2843 khugepaged_full_scans++;
2844 }
2845
2846 collect_mm_slot(mm_slot);
2847 }
2848
2849 return progress;
2850}
2851
2852static int khugepaged_has_work(void)
2853{
2854 return !list_empty(&khugepaged_scan.mm_head) &&
2855 khugepaged_enabled();
2856}
2857
2858static int khugepaged_wait_event(void)
2859{
2860 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002861 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002862}
2863
Xiao Guangrongd5169042012-10-08 16:29:48 -07002864static void khugepaged_do_scan(void)
2865{
2866 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002867 unsigned int progress = 0, pass_through_head = 0;
2868 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002869 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002870
2871 barrier(); /* write khugepaged_pages_to_scan to local stack */
2872
2873 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002874 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002875 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002876
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002877 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002878
Jiri Kosinacd092412015-06-24 16:56:07 -07002879 if (unlikely(kthread_should_stop() || try_to_freeze()))
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002880 break;
2881
Andrea Arcangeliba761492011-01-13 15:46:58 -08002882 spin_lock(&khugepaged_mm_lock);
2883 if (!khugepaged_scan.mm_slot)
2884 pass_through_head++;
2885 if (khugepaged_has_work() &&
2886 pass_through_head < 2)
2887 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002888 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002889 else
2890 progress = pages;
2891 spin_unlock(&khugepaged_mm_lock);
2892 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002893
Xiao Guangrongd5169042012-10-08 16:29:48 -07002894 if (!IS_ERR_OR_NULL(hpage))
2895 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002896}
2897
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002898static void khugepaged_wait_work(void)
2899{
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002900 if (khugepaged_has_work()) {
2901 if (!khugepaged_scan_sleep_millisecs)
2902 return;
2903
2904 wait_event_freezable_timeout(khugepaged_wait,
2905 kthread_should_stop(),
2906 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2907 return;
2908 }
2909
2910 if (khugepaged_enabled())
2911 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2912}
2913
Andrea Arcangeliba761492011-01-13 15:46:58 -08002914static int khugepaged(void *none)
2915{
2916 struct mm_slot *mm_slot;
2917
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002918 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002919 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002920
Xiao Guangrongb7231782012-10-08 16:29:54 -07002921 while (!kthread_should_stop()) {
2922 khugepaged_do_scan();
2923 khugepaged_wait_work();
2924 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002925
2926 spin_lock(&khugepaged_mm_lock);
2927 mm_slot = khugepaged_scan.mm_slot;
2928 khugepaged_scan.mm_slot = NULL;
2929 if (mm_slot)
2930 collect_mm_slot(mm_slot);
2931 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002932 return 0;
2933}
2934
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002935static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2936 unsigned long haddr, pmd_t *pmd)
2937{
2938 struct mm_struct *mm = vma->vm_mm;
2939 pgtable_t pgtable;
2940 pmd_t _pmd;
2941 int i;
2942
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -07002943 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002944 /* leave pmd empty until pte is filled */
2945
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002946 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002947 pmd_populate(mm, &_pmd, pgtable);
2948
2949 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2950 pte_t *pte, entry;
2951 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2952 entry = pte_mkspecial(entry);
2953 pte = pte_offset_map(&_pmd, haddr);
2954 VM_BUG_ON(!pte_none(*pte));
2955 set_pte_at(mm, haddr, pte, entry);
2956 pte_unmap(pte);
2957 }
2958 smp_wmb(); /* make pte visible before pmd */
2959 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002960 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002961}
2962
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002963void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2964 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002965{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002966 spinlock_t *ptl;
Matthew Wilcox4897c762015-09-08 14:58:45 -07002967 struct page *page = NULL;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002968 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002969 unsigned long haddr = address & HPAGE_PMD_MASK;
2970 unsigned long mmun_start; /* For mmu_notifiers */
2971 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002972
2973 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002974
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002975 mmun_start = haddr;
2976 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002977again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002978 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002979 ptl = pmd_lock(mm, pmd);
Matthew Wilcox4897c762015-09-08 14:58:45 -07002980 if (unlikely(!pmd_trans_huge(*pmd)))
2981 goto unlock;
2982 if (vma_is_dax(vma)) {
Kirill A. Shutemov5b701b82015-09-08 14:59:28 -07002983 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2984 if (is_huge_zero_pmd(_pmd))
2985 put_huge_zero_page();
Matthew Wilcox4897c762015-09-08 14:58:45 -07002986 } else if (is_huge_zero_pmd(*pmd)) {
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002987 __split_huge_zero_page_pmd(vma, haddr, pmd);
Matthew Wilcox4897c762015-09-08 14:58:45 -07002988 } else {
2989 page = pmd_page(*pmd);
2990 VM_BUG_ON_PAGE(!page_count(page), page);
2991 get_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002992 }
Matthew Wilcox4897c762015-09-08 14:58:45 -07002993 unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002994 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002995 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002996
Matthew Wilcox4897c762015-09-08 14:58:45 -07002997 if (!page)
2998 return;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002999
Matthew Wilcox4897c762015-09-08 14:58:45 -07003000 split_huge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003001 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07003002
3003 /*
3004 * We don't always have down_write of mmap_sem here: a racing
3005 * do_huge_pmd_wp_page() might have copied-on-write to another
3006 * huge page before our split_huge_page() got the anon_vma lock.
3007 */
3008 if (unlikely(pmd_trans_huge(*pmd)))
3009 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003010}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003011
Kirill A. Shutemove1803772012-12-12 13:50:59 -08003012void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3013 pmd_t *pmd)
3014{
3015 struct vm_area_struct *vma;
3016
3017 vma = find_vma(mm, address);
3018 BUG_ON(vma == NULL);
3019 split_huge_page_pmd(vma, address, pmd);
3020}
3021
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003022static void split_huge_page_address(struct mm_struct *mm,
3023 unsigned long address)
3024{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07003025 pgd_t *pgd;
3026 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003027 pmd_t *pmd;
3028
3029 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3030
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07003031 pgd = pgd_offset(mm, address);
3032 if (!pgd_present(*pgd))
3033 return;
3034
3035 pud = pud_offset(pgd, address);
3036 if (!pud_present(*pud))
3037 return;
3038
3039 pmd = pmd_offset(pud, address);
3040 if (!pmd_present(*pmd))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003041 return;
3042 /*
3043 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3044 * materialize from under us.
3045 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08003046 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003047}
3048
Kirill A. Shutemove1b99962015-09-08 14:58:37 -07003049void vma_adjust_trans_huge(struct vm_area_struct *vma,
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08003050 unsigned long start,
3051 unsigned long end,
3052 long adjust_next)
3053{
3054 /*
3055 * If the new start address isn't hpage aligned and it could
3056 * previously contain an hugepage: check if we need to split
3057 * an huge pmd.
3058 */
3059 if (start & ~HPAGE_PMD_MASK &&
3060 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3061 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3062 split_huge_page_address(vma->vm_mm, start);
3063
3064 /*
3065 * If the new end address isn't hpage aligned and it could
3066 * previously contain an hugepage: check if we need to split
3067 * an huge pmd.
3068 */
3069 if (end & ~HPAGE_PMD_MASK &&
3070 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3071 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3072 split_huge_page_address(vma->vm_mm, end);
3073
3074 /*
3075 * If we're also updating the vma->vm_next->vm_start, if the new
3076 * vm_next->vm_start isn't page aligned and it could previously
3077 * contain an hugepage: check if we need to split an huge pmd.
3078 */
3079 if (adjust_next > 0) {
3080 struct vm_area_struct *next = vma->vm_next;
3081 unsigned long nstart = next->vm_start;
3082 nstart += adjust_next << PAGE_SHIFT;
3083 if (nstart & ~HPAGE_PMD_MASK &&
3084 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3085 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3086 split_huge_page_address(next->vm_mm, nstart);
3087 }
3088}