blob: 626e93db28ba162d11e7d286985604bbc523981c [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>
19#include <linux/kthread.h>
20#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080021#include <linux/freezer.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080022#include <linux/mman.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020023#include <linux/pagemap.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000024#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080025#include <linux/hashtable.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080026
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080027#include <asm/tlb.h>
28#include <asm/pgalloc.h>
29#include "internal.h"
30
Andrea Arcangeliba761492011-01-13 15:46:58 -080031/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080032 * By default transparent hugepage support is disabled in order that avoid
33 * to risk increase the memory footprint of applications without a guaranteed
34 * benefit. When transparent hugepage support is enabled, is for all mappings,
35 * and khugepaged scans all mappings.
36 * Defrag is invoked by khugepaged hugepage allocations and by page faults
37 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080038 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080039unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080040#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080041 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080042#endif
43#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
44 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
45#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080046 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080047 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
48 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080049
50/* default scan 8*512 pte (or vmas) every 30 second */
51static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
52static unsigned int khugepaged_pages_collapsed;
53static unsigned int khugepaged_full_scans;
54static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
55/* during fragmentation poll the hugepage allocator once every minute */
56static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
57static struct task_struct *khugepaged_thread __read_mostly;
58static DEFINE_MUTEX(khugepaged_mutex);
59static DEFINE_SPINLOCK(khugepaged_mm_lock);
60static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
61/*
62 * default collapse hugepages if there is at least one pte mapped like
63 * it would have happened if the vma was large enough during page
64 * fault.
65 */
66static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
67
68static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -080069static int khugepaged_slab_init(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080070
Sasha Levin43b5fbb2013-02-22 16:32:27 -080071#define MM_SLOTS_HASH_BITS 10
72static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73
Andrea Arcangeliba761492011-01-13 15:46:58 -080074static struct kmem_cache *mm_slot_cache __read_mostly;
75
76/**
77 * struct mm_slot - hash lookup from mm to mm_slot
78 * @hash: hash collision list
79 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80 * @mm: the mm that this information is valid for
81 */
82struct mm_slot {
83 struct hlist_node hash;
84 struct list_head mm_node;
85 struct mm_struct *mm;
86};
87
88/**
89 * struct khugepaged_scan - cursor for scanning
90 * @mm_head: the head of the mm list to scan
91 * @mm_slot: the current mm_slot we are scanning
92 * @address: the next address inside that to be scanned
93 *
94 * There is only the one khugepaged_scan instance of this cursor structure.
95 */
96struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -0700100};
101static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800102 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
103};
104
Andrea Arcangelif0005652011-01-13 15:47:04 -0800105
106static int set_recommended_min_free_kbytes(void)
107{
108 struct zone *zone;
109 int nr_zones = 0;
110 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800111
Xiao Guangrong17c230a2012-10-08 16:29:56 -0700112 if (!khugepaged_enabled())
Andrea Arcangelif0005652011-01-13 15:47:04 -0800113 return 0;
114
115 for_each_populated_zone(zone)
116 nr_zones++;
117
118 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
119 recommended_min = pageblock_nr_pages * nr_zones * 2;
120
121 /*
122 * Make sure that on average at least two pageblocks are almost free
123 * of another type, one for a migratetype to fall back to and a
124 * second to avoid subsequent fallbacks of other types There are 3
125 * MIGRATE_TYPES we care about.
126 */
127 recommended_min += pageblock_nr_pages * nr_zones *
128 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
129
130 /* don't ever allow to reserve more than 5% of the lowmem */
131 recommended_min = min(recommended_min,
132 (unsigned long) nr_free_buffer_pages() / 20);
133 recommended_min <<= (PAGE_SHIFT-10);
134
Han Pingtian42aa83c2014-01-23 15:53:28 -0800135 if (recommended_min > min_free_kbytes) {
136 if (user_min_free_kbytes >= 0)
137 pr_info("raising min_free_kbytes from %d to %lu "
138 "to help transparent hugepage allocations\n",
139 min_free_kbytes, recommended_min);
140
Andrea Arcangelif0005652011-01-13 15:47:04 -0800141 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800142 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800143 setup_per_zone_wmarks();
144 return 0;
145}
146late_initcall(set_recommended_min_free_kbytes);
147
Andrea Arcangeliba761492011-01-13 15:46:58 -0800148static int start_khugepaged(void)
149{
150 int err = 0;
151 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800152 if (!khugepaged_thread)
153 khugepaged_thread = kthread_run(khugepaged, NULL,
154 "khugepaged");
155 if (unlikely(IS_ERR(khugepaged_thread))) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700156 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
Andrea Arcangeliba761492011-01-13 15:46:58 -0800157 err = PTR_ERR(khugepaged_thread);
158 khugepaged_thread = NULL;
159 }
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 }
Xiao Guangrong637e3a22012-10-08 16:29:38 -0700169
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
176static inline bool is_huge_zero_pmd(pmd_t pmd)
177{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700178 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800179}
180
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700181static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800182{
183 struct page *zero_page;
184retry:
185 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700186 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800187
188 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
189 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800190 if (!zero_page) {
191 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700192 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800193 }
194 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800195 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700196 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800197 preempt_enable();
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700198 __free_pages(zero_page, compound_order(zero_page));
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800199 goto retry;
200 }
201
202 /* We take additional reference here. It will be put back by shrinker */
203 atomic_set(&huge_zero_refcount, 2);
204 preempt_enable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700205 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800206}
207
208static void put_huge_zero_page(void)
209{
210 /*
211 * Counter should never go to zero here. Only shrinker can put
212 * last reference.
213 */
214 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
215}
216
Glauber Costa48896462013-08-28 10:18:15 +1000217static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
218 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800219{
Glauber Costa48896462013-08-28 10:18:15 +1000220 /* we can free zero page only if last reference remains */
221 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
222}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800223
Glauber Costa48896462013-08-28 10:18:15 +1000224static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
225 struct shrink_control *sc)
226{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800227 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700228 struct page *zero_page = xchg(&huge_zero_page, NULL);
229 BUG_ON(zero_page == NULL);
Yu Zhao5ddacbe2014-10-29 14:50:26 -0700230 __free_pages(zero_page, compound_order(zero_page));
Glauber Costa48896462013-08-28 10:18:15 +1000231 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800232 }
233
234 return 0;
235}
236
237static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000238 .count_objects = shrink_huge_zero_page_count,
239 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800240 .seeks = DEFAULT_SEEKS,
241};
242
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800243#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800244
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800245static ssize_t double_flag_show(struct kobject *kobj,
246 struct kobj_attribute *attr, char *buf,
247 enum transparent_hugepage_flag enabled,
248 enum transparent_hugepage_flag req_madv)
249{
250 if (test_bit(enabled, &transparent_hugepage_flags)) {
251 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
252 return sprintf(buf, "[always] madvise never\n");
253 } else if (test_bit(req_madv, &transparent_hugepage_flags))
254 return sprintf(buf, "always [madvise] never\n");
255 else
256 return sprintf(buf, "always madvise [never]\n");
257}
258static ssize_t double_flag_store(struct kobject *kobj,
259 struct kobj_attribute *attr,
260 const char *buf, size_t count,
261 enum transparent_hugepage_flag enabled,
262 enum transparent_hugepage_flag req_madv)
263{
264 if (!memcmp("always", buf,
265 min(sizeof("always")-1, count))) {
266 set_bit(enabled, &transparent_hugepage_flags);
267 clear_bit(req_madv, &transparent_hugepage_flags);
268 } else if (!memcmp("madvise", buf,
269 min(sizeof("madvise")-1, count))) {
270 clear_bit(enabled, &transparent_hugepage_flags);
271 set_bit(req_madv, &transparent_hugepage_flags);
272 } else if (!memcmp("never", buf,
273 min(sizeof("never")-1, count))) {
274 clear_bit(enabled, &transparent_hugepage_flags);
275 clear_bit(req_madv, &transparent_hugepage_flags);
276 } else
277 return -EINVAL;
278
279 return count;
280}
281
282static ssize_t enabled_show(struct kobject *kobj,
283 struct kobj_attribute *attr, char *buf)
284{
285 return double_flag_show(kobj, attr, buf,
286 TRANSPARENT_HUGEPAGE_FLAG,
287 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
288}
289static ssize_t enabled_store(struct kobject *kobj,
290 struct kobj_attribute *attr,
291 const char *buf, size_t count)
292{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800293 ssize_t ret;
294
295 ret = double_flag_store(kobj, attr, buf, count,
296 TRANSPARENT_HUGEPAGE_FLAG,
297 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
298
299 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700300 int err;
301
302 mutex_lock(&khugepaged_mutex);
303 err = start_khugepaged();
304 mutex_unlock(&khugepaged_mutex);
305
Andrea Arcangeliba761492011-01-13 15:46:58 -0800306 if (err)
307 ret = err;
308 }
309
310 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800311}
312static struct kobj_attribute enabled_attr =
313 __ATTR(enabled, 0644, enabled_show, enabled_store);
314
315static ssize_t single_flag_show(struct kobject *kobj,
316 struct kobj_attribute *attr, char *buf,
317 enum transparent_hugepage_flag flag)
318{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700319 return sprintf(buf, "%d\n",
320 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800321}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700322
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800323static ssize_t single_flag_store(struct kobject *kobj,
324 struct kobj_attribute *attr,
325 const char *buf, size_t count,
326 enum transparent_hugepage_flag flag)
327{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700328 unsigned long value;
329 int ret;
330
331 ret = kstrtoul(buf, 10, &value);
332 if (ret < 0)
333 return ret;
334 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800335 return -EINVAL;
336
Ben Hutchingse27e6152011-04-14 15:22:21 -0700337 if (value)
338 set_bit(flag, &transparent_hugepage_flags);
339 else
340 clear_bit(flag, &transparent_hugepage_flags);
341
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800342 return count;
343}
344
345/*
346 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
347 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
348 * memory just to allocate one more hugepage.
349 */
350static ssize_t defrag_show(struct kobject *kobj,
351 struct kobj_attribute *attr, char *buf)
352{
353 return double_flag_show(kobj, attr, buf,
354 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
355 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
356}
357static ssize_t defrag_store(struct kobject *kobj,
358 struct kobj_attribute *attr,
359 const char *buf, size_t count)
360{
361 return double_flag_store(kobj, attr, buf, count,
362 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
363 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
364}
365static struct kobj_attribute defrag_attr =
366 __ATTR(defrag, 0644, defrag_show, defrag_store);
367
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800368static ssize_t use_zero_page_show(struct kobject *kobj,
369 struct kobj_attribute *attr, char *buf)
370{
371 return single_flag_show(kobj, attr, buf,
372 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
373}
374static ssize_t use_zero_page_store(struct kobject *kobj,
375 struct kobj_attribute *attr, const char *buf, size_t count)
376{
377 return single_flag_store(kobj, attr, buf, count,
378 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
379}
380static struct kobj_attribute use_zero_page_attr =
381 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800382#ifdef CONFIG_DEBUG_VM
383static ssize_t debug_cow_show(struct kobject *kobj,
384 struct kobj_attribute *attr, char *buf)
385{
386 return single_flag_show(kobj, attr, buf,
387 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
388}
389static ssize_t debug_cow_store(struct kobject *kobj,
390 struct kobj_attribute *attr,
391 const char *buf, size_t count)
392{
393 return single_flag_store(kobj, attr, buf, count,
394 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
395}
396static struct kobj_attribute debug_cow_attr =
397 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
398#endif /* CONFIG_DEBUG_VM */
399
400static struct attribute *hugepage_attr[] = {
401 &enabled_attr.attr,
402 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800403 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800404#ifdef CONFIG_DEBUG_VM
405 &debug_cow_attr.attr,
406#endif
407 NULL,
408};
409
410static struct attribute_group hugepage_attr_group = {
411 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800412};
413
414static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
415 struct kobj_attribute *attr,
416 char *buf)
417{
418 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
419}
420
421static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
422 struct kobj_attribute *attr,
423 const char *buf, size_t count)
424{
425 unsigned long msecs;
426 int err;
427
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700428 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800429 if (err || msecs > UINT_MAX)
430 return -EINVAL;
431
432 khugepaged_scan_sleep_millisecs = msecs;
433 wake_up_interruptible(&khugepaged_wait);
434
435 return count;
436}
437static struct kobj_attribute scan_sleep_millisecs_attr =
438 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
439 scan_sleep_millisecs_store);
440
441static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
442 struct kobj_attribute *attr,
443 char *buf)
444{
445 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
446}
447
448static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
449 struct kobj_attribute *attr,
450 const char *buf, size_t count)
451{
452 unsigned long msecs;
453 int err;
454
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700455 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800456 if (err || msecs > UINT_MAX)
457 return -EINVAL;
458
459 khugepaged_alloc_sleep_millisecs = msecs;
460 wake_up_interruptible(&khugepaged_wait);
461
462 return count;
463}
464static struct kobj_attribute alloc_sleep_millisecs_attr =
465 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
466 alloc_sleep_millisecs_store);
467
468static ssize_t pages_to_scan_show(struct kobject *kobj,
469 struct kobj_attribute *attr,
470 char *buf)
471{
472 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
473}
474static ssize_t pages_to_scan_store(struct kobject *kobj,
475 struct kobj_attribute *attr,
476 const char *buf, size_t count)
477{
478 int err;
479 unsigned long pages;
480
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700481 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800482 if (err || !pages || pages > UINT_MAX)
483 return -EINVAL;
484
485 khugepaged_pages_to_scan = pages;
486
487 return count;
488}
489static struct kobj_attribute pages_to_scan_attr =
490 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
491 pages_to_scan_store);
492
493static ssize_t pages_collapsed_show(struct kobject *kobj,
494 struct kobj_attribute *attr,
495 char *buf)
496{
497 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
498}
499static struct kobj_attribute pages_collapsed_attr =
500 __ATTR_RO(pages_collapsed);
501
502static ssize_t full_scans_show(struct kobject *kobj,
503 struct kobj_attribute *attr,
504 char *buf)
505{
506 return sprintf(buf, "%u\n", khugepaged_full_scans);
507}
508static struct kobj_attribute full_scans_attr =
509 __ATTR_RO(full_scans);
510
511static ssize_t khugepaged_defrag_show(struct kobject *kobj,
512 struct kobj_attribute *attr, char *buf)
513{
514 return single_flag_show(kobj, attr, buf,
515 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
516}
517static ssize_t khugepaged_defrag_store(struct kobject *kobj,
518 struct kobj_attribute *attr,
519 const char *buf, size_t count)
520{
521 return single_flag_store(kobj, attr, buf, count,
522 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
523}
524static struct kobj_attribute khugepaged_defrag_attr =
525 __ATTR(defrag, 0644, khugepaged_defrag_show,
526 khugepaged_defrag_store);
527
528/*
529 * max_ptes_none controls if khugepaged should collapse hugepages over
530 * any unmapped ptes in turn potentially increasing the memory
531 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
532 * reduce the available free memory in the system as it
533 * runs. Increasing max_ptes_none will instead potentially reduce the
534 * free memory in the system during the khugepaged scan.
535 */
536static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
537 struct kobj_attribute *attr,
538 char *buf)
539{
540 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
541}
542static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
543 struct kobj_attribute *attr,
544 const char *buf, size_t count)
545{
546 int err;
547 unsigned long max_ptes_none;
548
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700549 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800550 if (err || max_ptes_none > HPAGE_PMD_NR-1)
551 return -EINVAL;
552
553 khugepaged_max_ptes_none = max_ptes_none;
554
555 return count;
556}
557static struct kobj_attribute khugepaged_max_ptes_none_attr =
558 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
559 khugepaged_max_ptes_none_store);
560
561static struct attribute *khugepaged_attr[] = {
562 &khugepaged_defrag_attr.attr,
563 &khugepaged_max_ptes_none_attr.attr,
564 &pages_to_scan_attr.attr,
565 &pages_collapsed_attr.attr,
566 &full_scans_attr.attr,
567 &scan_sleep_millisecs_attr.attr,
568 &alloc_sleep_millisecs_attr.attr,
569 NULL,
570};
571
572static struct attribute_group khugepaged_attr_group = {
573 .attrs = khugepaged_attr,
574 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800575};
Shaohua Li569e5592012-01-12 17:19:11 -0800576
577static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
578{
579 int err;
580
581 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
582 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700583 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800584 return -ENOMEM;
585 }
586
587 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
588 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700589 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800590 goto delete_obj;
591 }
592
593 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
594 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700595 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800596 goto remove_hp_group;
597 }
598
599 return 0;
600
601remove_hp_group:
602 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
603delete_obj:
604 kobject_put(*hugepage_kobj);
605 return err;
606}
607
608static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
609{
610 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
611 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
612 kobject_put(hugepage_kobj);
613}
614#else
615static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
616{
617 return 0;
618}
619
620static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
621{
622}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800623#endif /* CONFIG_SYSFS */
624
625static int __init hugepage_init(void)
626{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800627 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800628 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800629
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800630 if (!has_transparent_hugepage()) {
631 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800632 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800633 }
634
Shaohua Li569e5592012-01-12 17:19:11 -0800635 err = hugepage_init_sysfs(&hugepage_kobj);
636 if (err)
637 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800638
639 err = khugepaged_slab_init();
640 if (err)
641 goto out;
642
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800643 register_shrinker(&huge_zero_page_shrinker);
644
Rik van Riel97562cd2011-01-13 15:47:12 -0800645 /*
646 * By default disable transparent hugepages on smaller systems,
647 * where the extra memory used could hurt more than TLB overhead
648 * is likely to save. The admin can still enable it through /sys.
649 */
650 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
651 transparent_hugepage_flags = 0;
652
Andrea Arcangeliba761492011-01-13 15:46:58 -0800653 start_khugepaged();
654
Shaohua Li569e5592012-01-12 17:19:11 -0800655 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800656out:
Shaohua Li569e5592012-01-12 17:19:11 -0800657 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800658 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800659}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800660subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800661
662static int __init setup_transparent_hugepage(char *str)
663{
664 int ret = 0;
665 if (!str)
666 goto out;
667 if (!strcmp(str, "always")) {
668 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
669 &transparent_hugepage_flags);
670 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
671 &transparent_hugepage_flags);
672 ret = 1;
673 } else if (!strcmp(str, "madvise")) {
674 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
675 &transparent_hugepage_flags);
676 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
677 &transparent_hugepage_flags);
678 ret = 1;
679 } else if (!strcmp(str, "never")) {
680 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
681 &transparent_hugepage_flags);
682 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
683 &transparent_hugepage_flags);
684 ret = 1;
685 }
686out:
687 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700688 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800689 return ret;
690}
691__setup("transparent_hugepage=", setup_transparent_hugepage);
692
Mel Gormanb32967f2012-11-19 12:35:47 +0000693pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800694{
695 if (likely(vma->vm_flags & VM_WRITE))
696 pmd = pmd_mkwrite(pmd);
697 return pmd;
698}
699
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700700static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800701{
702 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700703 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800704 entry = pmd_mkhuge(entry);
705 return entry;
706}
707
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800708static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
709 struct vm_area_struct *vma,
710 unsigned long haddr, pmd_t *pmd,
711 struct page *page)
712{
Johannes Weiner00501b52014-08-08 14:19:20 -0700713 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800714 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800715 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800716
Sasha Levin309381fea2014-01-23 15:52:54 -0800717 VM_BUG_ON_PAGE(!PageCompound(page), page);
Johannes Weiner00501b52014-08-08 14:19:20 -0700718
719 if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800721
Johannes Weiner00501b52014-08-08 14:19:20 -0700722 pgtable = pte_alloc_one(mm, haddr);
723 if (unlikely(!pgtable)) {
724 mem_cgroup_cancel_charge(page, memcg);
725 return VM_FAULT_OOM;
726 }
727
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800728 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700729 /*
730 * The memory barrier inside __SetPageUptodate makes sure that
731 * clear_huge_page writes become visible before the set_pmd_at()
732 * write.
733 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800734 __SetPageUptodate(page);
735
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800736 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800737 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800738 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -0700739 mem_cgroup_cancel_charge(page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800740 put_page(page);
741 pte_free(mm, pgtable);
742 } else {
743 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700744 entry = mk_huge_pmd(page, vma->vm_page_prot);
745 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800746 page_add_new_anon_rmap(page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -0700747 mem_cgroup_commit_charge(page, memcg, false);
748 lru_cache_add_active_or_unevictable(page, vma);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700749 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800750 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800751 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800752 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800753 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800754 }
755
David Rientjesaa2e8782012-05-29 15:06:17 -0700756 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800757}
758
Andi Kleencc5d4622011-03-22 16:33:13 -0700759static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800760{
Andi Kleencc5d4622011-03-22 16:33:13 -0700761 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800762}
763
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800764/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800765static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800766 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700767 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800768{
769 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800770 if (!pmd_none(*pmd))
771 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700772 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800773 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800775 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800776 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800777 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800778}
779
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800780int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
783{
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800784 gfp_t gfp;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800785 struct page *page;
786 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800787
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700788 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700789 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700790 if (unlikely(anon_vma_prepare(vma)))
791 return VM_FAULT_OOM;
David Rientjes6d50e602014-10-29 14:50:31 -0700792 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700793 return VM_FAULT_OOM;
Dominik Dingel593befa2014-10-23 12:07:44 +0200794 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700795 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800796 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700797 pgtable_t pgtable;
798 struct page *zero_page;
799 bool set;
800 pgtable = pte_alloc_one(mm, haddr);
801 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800802 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700803 zero_page = get_huge_zero_page();
804 if (unlikely(!zero_page)) {
805 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700806 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700807 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700808 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800809 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700810 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
811 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800812 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700813 if (!set) {
814 pte_free(mm, pgtable);
815 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800816 }
David Rientjesedad9d22012-05-29 15:06:17 -0700817 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800818 }
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -0800819 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
820 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700821 if (unlikely(!page)) {
822 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700823 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700824 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700825 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700826 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700827 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700828 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700829 }
830
David Rientjes17766dd2013-09-12 15:14:06 -0700831 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700832 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800833}
834
835int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
836 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
837 struct vm_area_struct *vma)
838{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800839 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800840 struct page *src_page;
841 pmd_t pmd;
842 pgtable_t pgtable;
843 int ret;
844
845 ret = -ENOMEM;
846 pgtable = pte_alloc_one(dst_mm, addr);
847 if (unlikely(!pgtable))
848 goto out;
849
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800850 dst_ptl = pmd_lock(dst_mm, dst_pmd);
851 src_ptl = pmd_lockptr(src_mm, src_pmd);
852 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800853
854 ret = -EAGAIN;
855 pmd = *src_pmd;
856 if (unlikely(!pmd_trans_huge(pmd))) {
857 pte_free(dst_mm, pgtable);
858 goto out_unlock;
859 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800860 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800861 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800862 * under splitting since we don't split the page itself, only pmd to
863 * a page table.
864 */
865 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700866 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800867 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800868 /*
869 * get_huge_zero_page() will never allocate a new page here,
870 * since we already have a zero page to copy. It just takes a
871 * reference.
872 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700873 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800874 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700875 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800876 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800877 ret = 0;
878 goto out_unlock;
879 }
Mel Gormande466bd2013-12-18 17:08:42 -0800880
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800881 if (unlikely(pmd_trans_splitting(pmd))) {
882 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800883 spin_unlock(src_ptl);
884 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800885 pte_free(dst_mm, pgtable);
886
887 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
888 goto out;
889 }
890 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800891 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800892 get_page(src_page);
893 page_dup_rmap(src_page);
894 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
895
896 pmdp_set_wrprotect(src_mm, addr, src_pmd);
897 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700898 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800899 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800900 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800901
902 ret = 0;
903out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800904 spin_unlock(src_ptl);
905 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800906out:
907 return ret;
908}
909
Will Deacona1dd4502012-12-11 16:01:27 -0800910void huge_pmd_set_accessed(struct mm_struct *mm,
911 struct vm_area_struct *vma,
912 unsigned long address,
913 pmd_t *pmd, pmd_t orig_pmd,
914 int dirty)
915{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800916 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800917 pmd_t entry;
918 unsigned long haddr;
919
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800920 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800921 if (unlikely(!pmd_same(*pmd, orig_pmd)))
922 goto unlock;
923
924 entry = pmd_mkyoung(orig_pmd);
925 haddr = address & HPAGE_PMD_MASK;
926 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
927 update_mmu_cache_pmd(vma, address, pmd);
928
929unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800930 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800931}
932
Hugh Dickins5338a932014-06-23 13:22:05 -0700933/*
934 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
935 * during copy_user_huge_page()'s copy_page_rep(): in the case when
936 * the source page gets split and a tail freed before copy completes.
937 * Called under pmd_lock of checked pmd, so safe from splitting itself.
938 */
939static void get_user_huge_page(struct page *page)
940{
941 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
942 struct page *endpage = page + HPAGE_PMD_NR;
943
944 atomic_add(HPAGE_PMD_NR, &page->_count);
945 while (++page < endpage)
946 get_huge_page_tail(page);
947 } else {
948 get_page(page);
949 }
950}
951
952static void put_user_huge_page(struct page *page)
953{
954 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
955 struct page *endpage = page + HPAGE_PMD_NR;
956
957 while (page < endpage)
958 put_page(page++);
959 } else {
960 put_page(page);
961 }
962}
963
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800964static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
965 struct vm_area_struct *vma,
966 unsigned long address,
967 pmd_t *pmd, pmd_t orig_pmd,
968 struct page *page,
969 unsigned long haddr)
970{
Johannes Weiner00501b52014-08-08 14:19:20 -0700971 struct mem_cgroup *memcg;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800972 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800973 pgtable_t pgtable;
974 pmd_t _pmd;
975 int ret = 0, i;
976 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700977 unsigned long mmun_start; /* For mmu_notifiers */
978 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800979
980 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
981 GFP_KERNEL);
982 if (unlikely(!pages)) {
983 ret |= VM_FAULT_OOM;
984 goto out;
985 }
986
987 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700988 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
989 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -0800990 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800991 if (unlikely(!pages[i] ||
Johannes Weiner00501b52014-08-08 14:19:20 -0700992 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
993 &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800994 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800995 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800996 while (--i >= 0) {
Johannes Weiner00501b52014-08-08 14:19:20 -0700997 memcg = (void *)page_private(pages[i]);
998 set_page_private(pages[i], 0);
999 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001000 put_page(pages[i]);
1001 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001002 kfree(pages);
1003 ret |= VM_FAULT_OOM;
1004 goto out;
1005 }
Johannes Weiner00501b52014-08-08 14:19:20 -07001006 set_page_private(pages[i], (unsigned long)memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001007 }
1008
1009 for (i = 0; i < HPAGE_PMD_NR; i++) {
1010 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001011 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001012 __SetPageUptodate(pages[i]);
1013 cond_resched();
1014 }
1015
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001016 mmun_start = haddr;
1017 mmun_end = haddr + HPAGE_PMD_SIZE;
1018 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1019
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001020 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001021 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1022 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001023 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001024
Joerg Roedel34ee6452014-11-13 13:46:09 +11001025 pmdp_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001026 /* leave pmd empty until pte is filled */
1027
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001028 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001029 pmd_populate(mm, &_pmd, pgtable);
1030
1031 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1032 pte_t *pte, entry;
1033 entry = mk_pte(pages[i], vma->vm_page_prot);
1034 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Johannes Weiner00501b52014-08-08 14:19:20 -07001035 memcg = (void *)page_private(pages[i]);
1036 set_page_private(pages[i], 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001037 page_add_new_anon_rmap(pages[i], vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001038 mem_cgroup_commit_charge(pages[i], memcg, false);
1039 lru_cache_add_active_or_unevictable(pages[i], vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001040 pte = pte_offset_map(&_pmd, haddr);
1041 VM_BUG_ON(!pte_none(*pte));
1042 set_pte_at(mm, haddr, pte, entry);
1043 pte_unmap(pte);
1044 }
1045 kfree(pages);
1046
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001047 smp_wmb(); /* make pte visible before pmd */
1048 pmd_populate(mm, pmd, pgtable);
1049 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001050 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001051
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001052 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1053
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001054 ret |= VM_FAULT_WRITE;
1055 put_page(page);
1056
1057out:
1058 return ret;
1059
1060out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001061 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001062 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001063 for (i = 0; i < HPAGE_PMD_NR; i++) {
Johannes Weiner00501b52014-08-08 14:19:20 -07001064 memcg = (void *)page_private(pages[i]);
1065 set_page_private(pages[i], 0);
1066 mem_cgroup_cancel_charge(pages[i], memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001067 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001068 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001069 kfree(pages);
1070 goto out;
1071}
1072
1073int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1074 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1075{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001076 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001077 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001078 struct page *page = NULL, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07001079 struct mem_cgroup *memcg;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001081 unsigned long mmun_start; /* For mmu_notifiers */
1082 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001083
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001084 ptl = pmd_lockptr(mm, pmd);
Sasha Levin81d1b092014-10-09 15:28:10 -07001085 VM_BUG_ON_VMA(!vma->anon_vma, vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001086 haddr = address & HPAGE_PMD_MASK;
1087 if (is_huge_zero_pmd(orig_pmd))
1088 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001089 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001090 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1091 goto out_unlock;
1092
1093 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001094 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001095 if (page_mapcount(page) == 1) {
1096 pmd_t entry;
1097 entry = pmd_mkyoung(orig_pmd);
1098 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1099 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001100 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001101 ret |= VM_FAULT_WRITE;
1102 goto out_unlock;
1103 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001104 get_user_huge_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001105 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001106alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001107 if (transparent_hugepage_enabled(vma) &&
Aneesh Kumar K.V077fcf12015-02-11 15:27:12 -08001108 !transparent_hugepage_debug_cow()) {
1109 gfp_t gfp;
1110
1111 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1112 new_page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
1113 } else
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001114 new_page = NULL;
1115
1116 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001117 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001118 split_huge_page_pmd(vma, address, pmd);
1119 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001120 } else {
1121 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1122 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001123 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001124 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001125 ret |= VM_FAULT_FALLBACK;
1126 }
Hugh Dickins5338a932014-06-23 13:22:05 -07001127 put_user_huge_page(page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001128 }
David Rientjes17766dd2013-09-12 15:14:06 -07001129 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001130 goto out;
1131 }
1132
Johannes Weiner00501b52014-08-08 14:19:20 -07001133 if (unlikely(mem_cgroup_try_charge(new_page, mm,
1134 GFP_TRANSHUGE, &memcg))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001135 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001136 if (page) {
1137 split_huge_page(page);
Hugh Dickins5338a932014-06-23 13:22:05 -07001138 put_user_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001139 } else
1140 split_huge_page_pmd(vma, address, pmd);
1141 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001142 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001143 goto out;
1144 }
1145
David Rientjes17766dd2013-09-12 15:14:06 -07001146 count_vm_event(THP_FAULT_ALLOC);
1147
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001148 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001149 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1150 else
1151 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001152 __SetPageUptodate(new_page);
1153
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001154 mmun_start = haddr;
1155 mmun_end = haddr + HPAGE_PMD_SIZE;
1156 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1157
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001158 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001159 if (page)
Hugh Dickins5338a932014-06-23 13:22:05 -07001160 put_user_huge_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001161 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001162 spin_unlock(ptl);
Johannes Weiner00501b52014-08-08 14:19:20 -07001163 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001165 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001166 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001167 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001168 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1169 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001170 pmdp_clear_flush_notify(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001171 page_add_new_anon_rmap(new_page, vma, haddr);
Johannes Weiner00501b52014-08-08 14:19:20 -07001172 mem_cgroup_commit_charge(new_page, memcg, false);
1173 lru_cache_add_active_or_unevictable(new_page, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001174 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001175 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001176 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001177 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001178 put_huge_zero_page();
1179 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001180 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001181 page_remove_rmap(page);
1182 put_page(page);
1183 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001184 ret |= VM_FAULT_WRITE;
1185 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001186 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001187out_mn:
1188 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1189out:
1190 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001191out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001192 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001193 return ret;
1194}
1195
David Rientjesb676b292012-10-08 16:34:03 -07001196struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001197 unsigned long addr,
1198 pmd_t *pmd,
1199 unsigned int flags)
1200{
David Rientjesb676b292012-10-08 16:34:03 -07001201 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001202 struct page *page = NULL;
1203
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001204 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001205
1206 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1207 goto out;
1208
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001209 /* Avoid dumping huge zero page */
1210 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1211 return ERR_PTR(-EFAULT);
1212
Mel Gorman2b4847e2013-12-18 17:08:32 -08001213 /* Full NUMA hinting faults to serialise migration in fault paths */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001214 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Mel Gorman2b4847e2013-12-18 17:08:32 -08001215 goto out;
1216
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001217 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001218 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001219 if (flags & FOLL_TOUCH) {
1220 pmd_t _pmd;
1221 /*
1222 * We should set the dirty bit only for FOLL_WRITE but
1223 * for now the dirty bit in the pmd is meaningless.
1224 * And if the dirty bit will become meaningful and
1225 * we'll only set it with FOLL_WRITE, an atomic
1226 * set_bit will be required on the pmd to set the
1227 * young bit, instead of the current set_pmd_at.
1228 */
1229 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001230 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1231 pmd, _pmd, 1))
1232 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 }
David Rientjesb676b292012-10-08 16:34:03 -07001234 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1235 if (page->mapping && trylock_page(page)) {
1236 lru_add_drain();
1237 if (page->mapping)
1238 mlock_vma_page(page);
1239 unlock_page(page);
1240 }
1241 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001242 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001243 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001244 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001245 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246
1247out:
1248 return page;
1249}
1250
Mel Gormand10e63f2012-10-25 14:16:31 +02001251/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001252int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1253 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001254{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001255 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001256 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001257 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001258 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001259 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001260 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001261 bool page_locked;
1262 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001263 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001264
Mel Gormanc0e7cad2015-02-12 14:58:41 -08001265 /* A PROT_NONE fault should not end up here */
1266 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1267
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001268 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001269 if (unlikely(!pmd_same(pmd, *pmdp)))
1270 goto out_unlock;
1271
Mel Gormande466bd2013-12-18 17:08:42 -08001272 /*
1273 * If there are potential migrations, wait for completion and retry
1274 * without disrupting NUMA hinting information. Do not relock and
1275 * check_same as the page may no longer be mapped.
1276 */
1277 if (unlikely(pmd_trans_migrating(*pmdp))) {
Mel Gorman5d833062015-02-12 14:58:16 -08001278 page = pmd_page(*pmdp);
Mel Gormande466bd2013-12-18 17:08:42 -08001279 spin_unlock(ptl);
Mel Gorman5d833062015-02-12 14:58:16 -08001280 wait_on_page_locked(page);
Mel Gormande466bd2013-12-18 17:08:42 -08001281 goto out;
1282 }
1283
Mel Gormand10e63f2012-10-25 14:16:31 +02001284 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001285 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001286 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001287 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001288 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001289 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001290 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001291 flags |= TNF_FAULT_LOCAL;
1292 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001293
Mel Gormanff9042b2013-10-07 11:28:43 +01001294 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001295 * Avoid grouping on DSO/COW pages in specific and RO pages
1296 * in general, RO pages shouldn't hurt as much anyway since
1297 * they can be in shared cache state.
Linus Torvalds53da3bc2015-03-12 08:45:46 -07001298 *
1299 * FIXME! This checks "pmd_dirty()" as an approximation of
1300 * "is this a read-only page", since checking "pmd_write()"
1301 * is even more broken. We haven't actually turned this into
1302 * a writable page, so pmd_write() will always be false.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001303 */
Linus Torvalds53da3bc2015-03-12 08:45:46 -07001304 if (!pmd_dirty(pmd))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001305 flags |= TNF_NO_GROUP;
1306
1307 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001308 * Acquire the page lock to serialise THP migrations but avoid dropping
1309 * page_table_lock if at all possible
1310 */
Mel Gormanb8916632013-10-07 11:28:44 +01001311 page_locked = trylock_page(page);
1312 target_nid = mpol_misplaced(page, vma, haddr);
1313 if (target_nid == -1) {
1314 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001315 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001316 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001317 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001318
Mel Gormande466bd2013-12-18 17:08:42 -08001319 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001320 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001321 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001322 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001323 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001324 goto out;
1325 }
1326
Mel Gorman2b4847e2013-12-18 17:08:32 -08001327 /*
1328 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1329 * to serialises splits
1330 */
Mel Gormanb8916632013-10-07 11:28:44 +01001331 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001332 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001333 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001334
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001335 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001336 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001337 if (unlikely(!pmd_same(pmd, *pmdp))) {
1338 unlock_page(page);
1339 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001340 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001341 goto out_unlock;
1342 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001343
Mel Gormanc3a489c2013-12-18 17:08:38 -08001344 /* Bail if we fail to protect against THP splits for any reason */
1345 if (unlikely(!anon_vma)) {
1346 put_page(page);
1347 page_nid = -1;
1348 goto clear_pmdnuma;
1349 }
1350
Mel Gormana54a4072013-10-07 11:28:46 +01001351 /*
1352 * Migrate the THP to the requested node, returns with page unlocked
Mel Gorman8a0516e2015-02-12 14:58:22 -08001353 * and access rights restored.
Mel Gormana54a4072013-10-07 11:28:46 +01001354 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001355 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001356 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001357 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001358 if (migrated) {
1359 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001360 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001361 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001362
Mel Gorman8191acb2013-10-07 11:28:45 +01001363 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001364clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001365 BUG_ON(!PageLocked(page));
Mel Gorman4d942462015-02-12 14:58:28 -08001366 pmd = pmd_modify(pmd, vma->vm_page_prot);
Mel Gormand10e63f2012-10-25 14:16:31 +02001367 set_pmd_at(mm, haddr, pmdp, pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02001368 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001369 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001370out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001371 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001372
1373out:
1374 if (anon_vma)
1375 page_unlock_anon_vma_read(anon_vma);
1376
Mel Gorman8191acb2013-10-07 11:28:45 +01001377 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001378 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001379
Mel Gormand10e63f2012-10-25 14:16:31 +02001380 return 0;
1381}
1382
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001383int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001384 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001385{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001386 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001387 int ret = 0;
1388
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001389 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001390 struct page *page;
1391 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001392 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001393 /*
1394 * For architectures like ppc64 we look at deposited pgtable
1395 * when calling pmdp_get_and_clear. So do the
1396 * pgtable_trans_huge_withdraw after finishing pmdp related
1397 * operations.
1398 */
Martin Schwidefskyfcbe08d62014-10-24 10:52:29 +02001399 orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd,
1400 tlb->fullmm);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001401 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001402 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001403 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001404 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001405 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001406 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001407 } else {
1408 page = pmd_page(orig_pmd);
1409 page_remove_rmap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001410 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001411 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
Sasha Levin309381fea2014-01-23 15:52:54 -08001412 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001413 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001414 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001415 tlb_remove_page(tlb, page);
1416 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001417 pte_free(tlb->mm, pgtable);
1418 ret = 1;
1419 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001420 return ret;
1421}
1422
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001423int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1424 unsigned long old_addr,
1425 unsigned long new_addr, unsigned long old_end,
1426 pmd_t *old_pmd, pmd_t *new_pmd)
1427{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001428 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001429 int ret = 0;
1430 pmd_t pmd;
1431
1432 struct mm_struct *mm = vma->vm_mm;
1433
1434 if ((old_addr & ~HPAGE_PMD_MASK) ||
1435 (new_addr & ~HPAGE_PMD_MASK) ||
1436 old_end - old_addr < HPAGE_PMD_SIZE ||
1437 (new_vma->vm_flags & VM_NOHUGEPAGE))
1438 goto out;
1439
1440 /*
1441 * The destination pmd shouldn't be established, free_pgtables()
1442 * should have release it.
1443 */
1444 if (WARN_ON(!pmd_none(*new_pmd))) {
1445 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1446 goto out;
1447 }
1448
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001449 /*
1450 * We don't have to worry about the ordering of src and dst
1451 * ptlocks because exclusive mmap_sem prevents deadlock.
1452 */
1453 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001454 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001455 new_ptl = pmd_lockptr(mm, new_pmd);
1456 if (new_ptl != old_ptl)
1457 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001458 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1459 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001460
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301461 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1462 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001463 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1464 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001465 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301466 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1467 if (new_ptl != old_ptl)
1468 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001469 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001470 }
1471out:
1472 return ret;
1473}
1474
Mel Gormanf123d742013-10-07 11:28:49 +01001475/*
1476 * Returns
1477 * - 0 if PMD could not be locked
1478 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1479 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1480 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001481int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gormane944fd62015-02-12 14:58:35 -08001482 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001483{
1484 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001485 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001486 int ret = 0;
1487
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001488 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001489 pmd_t entry;
Mel Gormanba68bc02015-03-07 15:20:48 +00001490 ret = 1;
Mel Gormane944fd62015-02-12 14:58:35 -08001491
1492 /*
1493 * Avoid trapping faults against the zero page. The read-only
1494 * data is likely to be read-cached on the local CPU and
1495 * local/remote hits to the zero page are not interesting.
1496 */
1497 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1498 spin_unlock(ptl);
Mel Gormanba68bc02015-03-07 15:20:48 +00001499 return ret;
Mel Gormane944fd62015-02-12 14:58:35 -08001500 }
1501
Mel Gorman10c10452015-02-12 14:58:44 -08001502 if (!prot_numa || !pmd_protnone(*pmd)) {
Mel Gorman10c10452015-02-12 14:58:44 -08001503 entry = pmdp_get_and_clear_notify(mm, addr, pmd);
1504 entry = pmd_modify(entry, newprot);
1505 ret = HPAGE_PMD_NR;
1506 set_pmd_at(mm, addr, pmd, entry);
1507 BUG_ON(pmd_write(entry));
1508 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001509 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001510 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001511
1512 return ret;
1513}
1514
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001515/*
1516 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1517 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1518 *
1519 * Note that if it returns 1, this routine returns without unlocking page
1520 * table locks. So callers must unlock them.
1521 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001522int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1523 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001524{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001525 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001526 if (likely(pmd_trans_huge(*pmd))) {
1527 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001528 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001529 wait_split_huge_page(vma->anon_vma, pmd);
1530 return -1;
1531 } else {
1532 /* Thp mapped by 'pmd' is stable, so we can
1533 * handle it as it is. */
1534 return 1;
1535 }
1536 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001537 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001538 return 0;
1539}
1540
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001541/*
1542 * This function returns whether a given @page is mapped onto the @address
1543 * in the virtual space of @mm.
1544 *
1545 * When it's true, this function returns *pmd with holding the page table lock
1546 * and passing it back to the caller via @ptl.
1547 * If it's false, returns NULL without holding the page table lock.
1548 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001549pmd_t *page_check_address_pmd(struct page *page,
1550 struct mm_struct *mm,
1551 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001552 enum page_check_address_pmd_flag flag,
1553 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001554{
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001555 pgd_t *pgd;
1556 pud_t *pud;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001557 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001558
1559 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001560 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001561
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001562 pgd = pgd_offset(mm, address);
1563 if (!pgd_present(*pgd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001564 return NULL;
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001565 pud = pud_offset(pgd, address);
1566 if (!pud_present(*pud))
1567 return NULL;
1568 pmd = pmd_offset(pud, address);
1569
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001570 *ptl = pmd_lock(mm, pmd);
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001571 if (!pmd_present(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001572 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001573 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001574 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001575 /*
1576 * split_vma() may create temporary aliased mappings. There is
1577 * no risk as long as all huge pmd are found and have their
1578 * splitting bit set before __split_huge_page_refcount
1579 * runs. Finding the same huge pmd more than once during the
1580 * same rmap walk is not a problem.
1581 */
1582 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1583 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001584 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001585 if (pmd_trans_huge(*pmd)) {
1586 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1587 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001588 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001589 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001590unlock:
1591 spin_unlock(*ptl);
1592 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001593}
1594
1595static int __split_huge_page_splitting(struct page *page,
1596 struct vm_area_struct *vma,
1597 unsigned long address)
1598{
1599 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001600 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001601 pmd_t *pmd;
1602 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001603 /* For mmu_notifiers */
1604 const unsigned long mmun_start = address;
1605 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001606
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001607 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001608 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001609 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001610 if (pmd) {
1611 /*
1612 * We can't temporarily set the pmd to null in order
1613 * to split it, the pmd must remain marked huge at all
1614 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001615 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001616 * serialize against split_huge_page*.
1617 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001618 pmdp_splitting_flush(vma, address, pmd);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001619
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001620 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001621 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001622 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001623 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001624
1625 return ret;
1626}
1627
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001628static void __split_huge_page_refcount(struct page *page,
1629 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001630{
1631 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001632 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001633 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001634 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001635
1636 /* prevent PageLRU to go away from under us, and freeze lru stats */
1637 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001638 lruvec = mem_cgroup_page_lruvec(page, zone);
1639
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001640 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001641 /* complete memcg works before add pages to LRU */
1642 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001643
Shaohua Li45676882012-01-12 17:19:18 -08001644 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001645 struct page *page_tail = page + i;
1646
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001647 /* tail_page->_mapcount cannot change */
1648 BUG_ON(page_mapcount(page_tail) < 0);
1649 tail_count += page_mapcount(page_tail);
1650 /* check for overflow */
1651 BUG_ON(tail_count < 0);
1652 BUG_ON(atomic_read(&page_tail->_count) != 0);
1653 /*
1654 * tail_page->_count is zero and not changing from
1655 * under us. But get_page_unless_zero() may be running
1656 * from under us on the tail_page. If we used
1657 * atomic_set() below instead of atomic_add(), we
1658 * would then run atomic_set() concurrently with
1659 * get_page_unless_zero(), and atomic_set() is
1660 * implemented in C not using locked ops. spin_unlock
1661 * on x86 sometime uses locked ops because of PPro
1662 * errata 66, 92, so unless somebody can guarantee
1663 * atomic_set() here would be safe on all archs (and
1664 * not only on x86), it's safer to use atomic_add().
1665 */
1666 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1667 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001668
1669 /* after clearing PageTail the gup refcount can be released */
Waiman Long3a79d522014-08-06 16:05:38 -07001670 smp_mb__after_atomic();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001671
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001672 /*
1673 * retain hwpoison flag of the poisoned tail page:
1674 * fix for the unsuitable process killed on Guest Machine(KVM)
1675 * by the memory-failure.
1676 */
1677 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001678 page_tail->flags |= (page->flags &
1679 ((1L << PG_referenced) |
1680 (1L << PG_swapbacked) |
1681 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001682 (1L << PG_uptodate) |
1683 (1L << PG_active) |
1684 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001685 page_tail->flags |= (1L << PG_dirty);
1686
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001687 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001688 smp_wmb();
1689
1690 /*
1691 * __split_huge_page_splitting() already set the
1692 * splitting bit in all pmd that could map this
1693 * hugepage, that will ensure no CPU can alter the
1694 * mapcount on the head page. The mapcount is only
1695 * accounted in the head page and it has to be
1696 * transferred to all tail pages in the below code. So
1697 * for this code to be safe, the split the mapcount
1698 * can't change. But that doesn't mean userland can't
1699 * keep changing and reading the page contents while
1700 * we transfer the mapcount, so the pmd splitting
1701 * status is achieved setting a reserved bit in the
1702 * pmd, not by clearing the present bit.
1703 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001704 page_tail->_mapcount = page->_mapcount;
1705
1706 BUG_ON(page_tail->mapping);
1707 page_tail->mapping = page->mapping;
1708
Shaohua Li45676882012-01-12 17:19:18 -08001709 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001710 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001711
1712 BUG_ON(!PageAnon(page_tail));
1713 BUG_ON(!PageUptodate(page_tail));
1714 BUG_ON(!PageDirty(page_tail));
1715 BUG_ON(!PageSwapBacked(page_tail));
1716
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001717 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001718 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001719 atomic_sub(tail_count, &page->_count);
1720 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001721
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001722 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001723
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001724 ClearPageCompound(page);
1725 compound_unlock(page);
1726 spin_unlock_irq(&zone->lru_lock);
1727
1728 for (i = 1; i < HPAGE_PMD_NR; i++) {
1729 struct page *page_tail = page + i;
1730 BUG_ON(page_count(page_tail) <= 0);
1731 /*
1732 * Tail pages may be freed if there wasn't any mapping
1733 * like if add_to_swap() is running on a lru page that
1734 * had its mapping zapped. And freeing these pages
1735 * requires taking the lru_lock so we do the put_page
1736 * of the tail pages after the split is complete.
1737 */
1738 put_page(page_tail);
1739 }
1740
1741 /*
1742 * Only the head page (now become a regular page) is required
1743 * to be pinned by the caller.
1744 */
1745 BUG_ON(page_count(page) <= 0);
1746}
1747
1748static int __split_huge_page_map(struct page *page,
1749 struct vm_area_struct *vma,
1750 unsigned long address)
1751{
1752 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001753 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001754 pmd_t *pmd, _pmd;
1755 int ret = 0, i;
1756 pgtable_t pgtable;
1757 unsigned long haddr;
1758
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001759 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001760 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001761 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001762 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001763 pmd_populate(mm, &_pmd, pgtable);
Waiman Longf8303c22014-08-06 16:05:36 -07001764 if (pmd_write(*pmd))
1765 BUG_ON(page_mapcount(page) != 1);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001766
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001767 haddr = address;
1768 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001769 pte_t *pte, entry;
1770 BUG_ON(PageCompound(page+i));
Mel Gormanabc40bd2014-10-02 19:47:42 +01001771 /*
Mel Gorman8a0516e2015-02-12 14:58:22 -08001772 * Note that NUMA hinting access restrictions are not
1773 * transferred to avoid any possibility of altering
1774 * permissions across VMAs.
Mel Gormanabc40bd2014-10-02 19:47:42 +01001775 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001776 entry = mk_pte(page + i, vma->vm_page_prot);
1777 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1778 if (!pmd_write(*pmd))
1779 entry = pte_wrprotect(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001780 if (!pmd_young(*pmd))
1781 entry = pte_mkold(entry);
1782 pte = pte_offset_map(&_pmd, haddr);
1783 BUG_ON(!pte_none(*pte));
1784 set_pte_at(mm, haddr, pte, entry);
1785 pte_unmap(pte);
1786 }
1787
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001788 smp_wmb(); /* make pte visible before pmd */
1789 /*
1790 * Up to this point the pmd is present and huge and
1791 * userland has the whole access to the hugepage
1792 * during the split (which happens in place). If we
1793 * overwrite the pmd with the not-huge version
1794 * pointing to the pte here (which of course we could
1795 * if all CPUs were bug free), userland could trigger
1796 * a small page size TLB miss on the small sized TLB
1797 * while the hugepage TLB entry is still established
1798 * in the huge TLB. Some CPU doesn't like that. See
1799 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1800 * Erratum 383 on page 93. Intel should be safe but is
1801 * also warns that it's only safe if the permission
1802 * and cache attributes of the two entries loaded in
1803 * the two TLB is identical (which should be the case
1804 * here). But it is generally safer to never allow
1805 * small and huge TLB entries for the same virtual
1806 * address to be loaded simultaneously. So instead of
1807 * doing "pmd_populate(); flush_tlb_range();" we first
1808 * mark the current pmd notpresent (atomically because
1809 * here the pmd_trans_huge and pmd_trans_splitting
1810 * must remain set at all times on the pmd until the
1811 * split is complete for this pmd), then we flush the
1812 * SMP TLB and finally we write the non-huge version
1813 * of the pmd entry with pmd_populate.
1814 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001815 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001816 pmd_populate(mm, pmd, pgtable);
1817 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001818 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001819 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001820
1821 return ret;
1822}
1823
Ingo Molnar5a505082012-12-02 19:56:46 +00001824/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001825static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001826 struct anon_vma *anon_vma,
1827 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001828{
1829 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001830 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001831 struct anon_vma_chain *avc;
1832
1833 BUG_ON(!PageHead(page));
1834 BUG_ON(PageTail(page));
1835
1836 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001837 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001838 struct vm_area_struct *vma = avc->vma;
1839 unsigned long addr = vma_address(page, vma);
1840 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001841 mapcount += __split_huge_page_splitting(page, vma, addr);
1842 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001843 /*
1844 * It is critical that new vmas are added to the tail of the
1845 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1846 * and establishes a child pmd before
1847 * __split_huge_page_splitting() freezes the parent pmd (so if
1848 * we fail to prevent copy_huge_pmd() from running until the
1849 * whole __split_huge_page() is complete), we will still see
1850 * the newly established pmd of the child later during the
1851 * walk, to be able to set it as pmd_trans_splitting too.
1852 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001853 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001854 pr_err("mapcount %d page_mapcount %d\n",
1855 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001856 BUG();
1857 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001858
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001859 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001860
1861 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001862 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001863 struct vm_area_struct *vma = avc->vma;
1864 unsigned long addr = vma_address(page, vma);
1865 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001866 mapcount2 += __split_huge_page_map(page, vma, addr);
1867 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001868 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001869 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1870 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001871 BUG();
1872 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001873}
1874
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001875/*
1876 * Split a hugepage into normal pages. This doesn't change the position of head
1877 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1878 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1879 * from the hugepage.
1880 * Return 0 if the hugepage is split successfully otherwise return 1.
1881 */
1882int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001883{
1884 struct anon_vma *anon_vma;
1885 int ret = 1;
1886
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001887 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001888 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001889
1890 /*
1891 * The caller does not necessarily hold an mmap_sem that would prevent
1892 * the anon_vma disappearing so we first we take a reference to it
1893 * and then lock the anon_vma for write. This is similar to
1894 * page_lock_anon_vma_read except the write lock is taken to serialise
1895 * against parallel split or collapse operations.
1896 */
1897 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001898 if (!anon_vma)
1899 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001900 anon_vma_lock_write(anon_vma);
1901
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001902 ret = 0;
1903 if (!PageCompound(page))
1904 goto out_unlock;
1905
1906 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001907 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001908 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001909
1910 BUG_ON(PageCompound(page));
1911out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001912 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001913 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001914out:
1915 return ret;
1916}
1917
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001918#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001919
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001920int hugepage_madvise(struct vm_area_struct *vma,
1921 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001922{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001923 switch (advice) {
1924 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001925#ifdef CONFIG_S390
1926 /*
1927 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1928 * can't handle this properly after s390_enable_sie, so we simply
1929 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1930 */
1931 if (mm_has_pgste(vma->vm_mm))
1932 return 0;
1933#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001934 /*
1935 * Be somewhat over-protective like KSM for now!
1936 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001937 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001938 return -EINVAL;
1939 *vm_flags &= ~VM_NOHUGEPAGE;
1940 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001941 /*
1942 * If the vma become good for khugepaged to scan,
1943 * register it here without waiting a page fault that
1944 * may not happen any time soon.
1945 */
David Rientjes6d50e602014-10-29 14:50:31 -07001946 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001947 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001948 break;
1949 case MADV_NOHUGEPAGE:
1950 /*
1951 * Be somewhat over-protective like KSM for now!
1952 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001953 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001954 return -EINVAL;
1955 *vm_flags &= ~VM_HUGEPAGE;
1956 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001957 /*
1958 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1959 * this vma even if we leave the mm registered in khugepaged if
1960 * it got registered before VM_NOHUGEPAGE was set.
1961 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001962 break;
1963 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001964
1965 return 0;
1966}
1967
Andrea Arcangeliba761492011-01-13 15:46:58 -08001968static int __init khugepaged_slab_init(void)
1969{
1970 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1971 sizeof(struct mm_slot),
1972 __alignof__(struct mm_slot), 0, NULL);
1973 if (!mm_slot_cache)
1974 return -ENOMEM;
1975
1976 return 0;
1977}
1978
Andrea Arcangeliba761492011-01-13 15:46:58 -08001979static inline struct mm_slot *alloc_mm_slot(void)
1980{
1981 if (!mm_slot_cache) /* initialization failed */
1982 return NULL;
1983 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1984}
1985
1986static inline void free_mm_slot(struct mm_slot *mm_slot)
1987{
1988 kmem_cache_free(mm_slot_cache, mm_slot);
1989}
1990
Andrea Arcangeliba761492011-01-13 15:46:58 -08001991static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1992{
1993 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08001994
Sasha Levinb67bfe02013-02-27 17:06:00 -08001995 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001996 if (mm == mm_slot->mm)
1997 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001998
Andrea Arcangeliba761492011-01-13 15:46:58 -08001999 return NULL;
2000}
2001
2002static void insert_to_mm_slots_hash(struct mm_struct *mm,
2003 struct mm_slot *mm_slot)
2004{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002005 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002006 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002007}
2008
2009static inline int khugepaged_test_exit(struct mm_struct *mm)
2010{
2011 return atomic_read(&mm->mm_users) == 0;
2012}
2013
2014int __khugepaged_enter(struct mm_struct *mm)
2015{
2016 struct mm_slot *mm_slot;
2017 int wakeup;
2018
2019 mm_slot = alloc_mm_slot();
2020 if (!mm_slot)
2021 return -ENOMEM;
2022
2023 /* __khugepaged_exit() must not run from under us */
Sasha Levin96dad672014-10-09 15:28:39 -07002024 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002025 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2026 free_mm_slot(mm_slot);
2027 return 0;
2028 }
2029
2030 spin_lock(&khugepaged_mm_lock);
2031 insert_to_mm_slots_hash(mm, mm_slot);
2032 /*
2033 * Insert just behind the scanning cursor, to let the area settle
2034 * down a little.
2035 */
2036 wakeup = list_empty(&khugepaged_scan.mm_head);
2037 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2038 spin_unlock(&khugepaged_mm_lock);
2039
2040 atomic_inc(&mm->mm_count);
2041 if (wakeup)
2042 wake_up_interruptible(&khugepaged_wait);
2043
2044 return 0;
2045}
2046
David Rientjes6d50e602014-10-29 14:50:31 -07002047int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2048 unsigned long vm_flags)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002049{
2050 unsigned long hstart, hend;
2051 if (!vma->anon_vma)
2052 /*
2053 * Not yet faulted in so we will register later in the
2054 * page fault if needed.
2055 */
2056 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002057 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002058 /* khugepaged not yet working on file or special mappings */
2059 return 0;
David Rientjes6d50e602014-10-29 14:50:31 -07002060 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002061 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2062 hend = vma->vm_end & HPAGE_PMD_MASK;
2063 if (hstart < hend)
David Rientjes6d50e602014-10-29 14:50:31 -07002064 return khugepaged_enter(vma, vm_flags);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002065 return 0;
2066}
2067
2068void __khugepaged_exit(struct mm_struct *mm)
2069{
2070 struct mm_slot *mm_slot;
2071 int free = 0;
2072
2073 spin_lock(&khugepaged_mm_lock);
2074 mm_slot = get_mm_slot(mm);
2075 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002076 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077 list_del(&mm_slot->mm_node);
2078 free = 1;
2079 }
Chris Wrightd788e802011-07-25 17:12:14 -07002080 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002081
2082 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002083 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2084 free_mm_slot(mm_slot);
2085 mmdrop(mm);
2086 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002087 /*
2088 * This is required to serialize against
2089 * khugepaged_test_exit() (which is guaranteed to run
2090 * under mmap sem read mode). Stop here (after we
2091 * return all pagetables will be destroyed) until
2092 * khugepaged has finished working on the pagetables
2093 * under the mmap_sem.
2094 */
2095 down_write(&mm->mmap_sem);
2096 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002097 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002098}
2099
2100static void release_pte_page(struct page *page)
2101{
2102 /* 0 stands for page_is_file_cache(page) == false */
2103 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2104 unlock_page(page);
2105 putback_lru_page(page);
2106}
2107
2108static void release_pte_pages(pte_t *pte, pte_t *_pte)
2109{
2110 while (--_pte >= pte) {
2111 pte_t pteval = *_pte;
2112 if (!pte_none(pteval))
2113 release_pte_page(pte_page(pteval));
2114 }
2115}
2116
Andrea Arcangeliba761492011-01-13 15:46:58 -08002117static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2118 unsigned long address,
2119 pte_t *pte)
2120{
2121 struct page *page;
2122 pte_t *_pte;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002123 int none = 0;
2124 bool referenced = false, writable = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002125 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2126 _pte++, address += PAGE_SIZE) {
2127 pte_t pteval = *_pte;
2128 if (pte_none(pteval)) {
2129 if (++none <= khugepaged_max_ptes_none)
2130 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002131 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002132 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002133 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002134 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002135 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002136 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002137 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002138 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002139
Sasha Levin309381fea2014-01-23 15:52:54 -08002140 VM_BUG_ON_PAGE(PageCompound(page), page);
2141 VM_BUG_ON_PAGE(!PageAnon(page), page);
2142 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002143
Andrea Arcangeliba761492011-01-13 15:46:58 -08002144 /*
2145 * We can do it before isolate_lru_page because the
2146 * page can't be freed from under us. NOTE: PG_lock
2147 * is needed to serialize against split_huge_page
2148 * when invoked from the VM.
2149 */
Bob Liu344aa352012-12-11 16:00:34 -08002150 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002151 goto out;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002152
2153 /*
2154 * cannot use mapcount: can't collapse if there's a gup pin.
2155 * The page must only be referenced by the scanned process
2156 * and page swap cache.
2157 */
2158 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2159 unlock_page(page);
2160 goto out;
2161 }
2162 if (pte_write(pteval)) {
2163 writable = true;
2164 } else {
2165 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2166 unlock_page(page);
2167 goto out;
2168 }
2169 /*
2170 * Page is not in the swap cache. It can be collapsed
2171 * into a THP.
2172 */
2173 }
2174
Andrea Arcangeliba761492011-01-13 15:46:58 -08002175 /*
2176 * Isolate the page to avoid collapsing an hugepage
2177 * currently in use by the VM.
2178 */
2179 if (isolate_lru_page(page)) {
2180 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002181 goto out;
2182 }
2183 /* 0 stands for page_is_file_cache(page) == false */
2184 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002185 VM_BUG_ON_PAGE(!PageLocked(page), page);
2186 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002187
2188 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002189 if (pte_young(pteval) || PageReferenced(page) ||
2190 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002191 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002192 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002193 if (likely(referenced && writable))
Bob Liu344aa352012-12-11 16:00:34 -08002194 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002195out:
Bob Liu344aa352012-12-11 16:00:34 -08002196 release_pte_pages(pte, _pte);
2197 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002198}
2199
2200static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2201 struct vm_area_struct *vma,
2202 unsigned long address,
2203 spinlock_t *ptl)
2204{
2205 pte_t *_pte;
2206 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2207 pte_t pteval = *_pte;
2208 struct page *src_page;
2209
2210 if (pte_none(pteval)) {
2211 clear_user_highpage(page, address);
2212 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2213 } else {
2214 src_page = pte_page(pteval);
2215 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002216 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002217 release_pte_page(src_page);
2218 /*
2219 * ptl mostly unnecessary, but preempt has to
2220 * be disabled to update the per-cpu stats
2221 * inside page_remove_rmap().
2222 */
2223 spin_lock(ptl);
2224 /*
2225 * paravirt calls inside pte_clear here are
2226 * superfluous.
2227 */
2228 pte_clear(vma->vm_mm, address, _pte);
2229 page_remove_rmap(src_page);
2230 spin_unlock(ptl);
2231 free_page_and_swap_cache(src_page);
2232 }
2233
2234 address += PAGE_SIZE;
2235 page++;
2236 }
2237}
2238
Xiao Guangrong26234f32012-10-08 16:29:51 -07002239static void khugepaged_alloc_sleep(void)
2240{
2241 wait_event_freezable_timeout(khugepaged_wait, false,
2242 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2243}
2244
Bob Liu9f1b8682013-11-12 15:07:37 -08002245static int khugepaged_node_load[MAX_NUMNODES];
2246
David Rientjes14a4e212014-08-06 16:07:29 -07002247static bool khugepaged_scan_abort(int nid)
2248{
2249 int i;
2250
2251 /*
2252 * If zone_reclaim_mode is disabled, then no extra effort is made to
2253 * allocate memory locally.
2254 */
2255 if (!zone_reclaim_mode)
2256 return false;
2257
2258 /* If there is a count for this node already, it must be acceptable */
2259 if (khugepaged_node_load[nid])
2260 return false;
2261
2262 for (i = 0; i < MAX_NUMNODES; i++) {
2263 if (!khugepaged_node_load[i])
2264 continue;
2265 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2266 return true;
2267 }
2268 return false;
2269}
2270
Xiao Guangrong26234f32012-10-08 16:29:51 -07002271#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002272static int khugepaged_find_target_node(void)
2273{
2274 static int last_khugepaged_target_node = NUMA_NO_NODE;
2275 int nid, target_node = 0, max_value = 0;
2276
2277 /* find first node with max normal pages hit */
2278 for (nid = 0; nid < MAX_NUMNODES; nid++)
2279 if (khugepaged_node_load[nid] > max_value) {
2280 max_value = khugepaged_node_load[nid];
2281 target_node = nid;
2282 }
2283
2284 /* do some balance if several nodes have the same hit record */
2285 if (target_node <= last_khugepaged_target_node)
2286 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2287 nid++)
2288 if (max_value == khugepaged_node_load[nid]) {
2289 target_node = nid;
2290 break;
2291 }
2292
2293 last_khugepaged_target_node = target_node;
2294 return target_node;
2295}
2296
Xiao Guangrong26234f32012-10-08 16:29:51 -07002297static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2298{
2299 if (IS_ERR(*hpage)) {
2300 if (!*wait)
2301 return false;
2302
2303 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002304 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002305 khugepaged_alloc_sleep();
2306 } else if (*hpage) {
2307 put_page(*hpage);
2308 *hpage = NULL;
2309 }
2310
2311 return true;
2312}
2313
2314static struct page
2315*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2316 struct vm_area_struct *vma, unsigned long address,
2317 int node)
2318{
Sasha Levin309381fea2014-01-23 15:52:54 -08002319 VM_BUG_ON_PAGE(*hpage, *hpage);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002320
Xiao Guangrong26234f32012-10-08 16:29:51 -07002321 /*
Vlastimil Babka8b164562014-10-09 15:27:00 -07002322 * Before allocating the hugepage, release the mmap_sem read lock.
2323 * The allocation can take potentially a long time if it involves
2324 * sync compaction, and we do not need to hold the mmap_sem during
2325 * that. We will recheck the vma after taking it again in write mode.
Xiao Guangrong26234f32012-10-08 16:29:51 -07002326 */
2327 up_read(&mm->mmap_sem);
Vlastimil Babka8b164562014-10-09 15:27:00 -07002328
2329 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2330 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002331 if (unlikely(!*hpage)) {
2332 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2333 *hpage = ERR_PTR(-ENOMEM);
2334 return NULL;
2335 }
2336
2337 count_vm_event(THP_COLLAPSE_ALLOC);
2338 return *hpage;
2339}
2340#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002341static int khugepaged_find_target_node(void)
2342{
2343 return 0;
2344}
2345
Bob Liu10dc4152013-11-12 15:07:35 -08002346static inline struct page *alloc_hugepage(int defrag)
2347{
2348 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2349 HPAGE_PMD_ORDER);
2350}
2351
Xiao Guangrong26234f32012-10-08 16:29:51 -07002352static struct page *khugepaged_alloc_hugepage(bool *wait)
2353{
2354 struct page *hpage;
2355
2356 do {
2357 hpage = alloc_hugepage(khugepaged_defrag());
2358 if (!hpage) {
2359 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2360 if (!*wait)
2361 return NULL;
2362
2363 *wait = false;
2364 khugepaged_alloc_sleep();
2365 } else
2366 count_vm_event(THP_COLLAPSE_ALLOC);
2367 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2368
2369 return hpage;
2370}
2371
2372static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2373{
2374 if (!*hpage)
2375 *hpage = khugepaged_alloc_hugepage(wait);
2376
2377 if (unlikely(!*hpage))
2378 return false;
2379
2380 return true;
2381}
2382
2383static struct page
2384*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2385 struct vm_area_struct *vma, unsigned long address,
2386 int node)
2387{
2388 up_read(&mm->mmap_sem);
2389 VM_BUG_ON(!*hpage);
2390 return *hpage;
2391}
2392#endif
2393
Bob Liufa475e52012-12-11 16:00:39 -08002394static bool hugepage_vma_check(struct vm_area_struct *vma)
2395{
2396 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2397 (vma->vm_flags & VM_NOHUGEPAGE))
2398 return false;
2399
2400 if (!vma->anon_vma || vma->vm_ops)
2401 return false;
2402 if (is_vma_temporary_stack(vma))
2403 return false;
Sasha Levin81d1b092014-10-09 15:28:10 -07002404 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
Bob Liufa475e52012-12-11 16:00:39 -08002405 return true;
2406}
2407
Andrea Arcangeliba761492011-01-13 15:46:58 -08002408static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002409 unsigned long address,
2410 struct page **hpage,
2411 struct vm_area_struct *vma,
2412 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002413{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002414 pmd_t *pmd, _pmd;
2415 pte_t *pte;
2416 pgtable_t pgtable;
2417 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002418 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002419 int isolated;
2420 unsigned long hstart, hend;
Johannes Weiner00501b52014-08-08 14:19:20 -07002421 struct mem_cgroup *memcg;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002422 unsigned long mmun_start; /* For mmu_notifiers */
2423 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002424
2425 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002426
Xiao Guangrong26234f32012-10-08 16:29:51 -07002427 /* release the mmap_sem read lock. */
2428 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2429 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002430 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002431
Johannes Weiner00501b52014-08-08 14:19:20 -07002432 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2433 GFP_TRANSHUGE, &memcg)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002434 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002435
2436 /*
2437 * Prevent all access to pagetables with the exception of
2438 * gup_fast later hanlded by the ptep_clear_flush and the VM
2439 * handled by the anon_vma lock + PG_lock.
2440 */
2441 down_write(&mm->mmap_sem);
2442 if (unlikely(khugepaged_test_exit(mm)))
2443 goto out;
2444
2445 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002446 if (!vma)
2447 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002448 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2449 hend = vma->vm_end & HPAGE_PMD_MASK;
2450 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2451 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002452 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002453 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002454 pmd = mm_find_pmd(mm, address);
2455 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002456 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002457
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002458 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002459
2460 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002461 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002462
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002463 mmun_start = address;
2464 mmun_end = address + HPAGE_PMD_SIZE;
2465 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002466 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002467 /*
2468 * After this gup_fast can't run anymore. This also removes
2469 * any huge TLB entry from the CPU so we won't allow
2470 * huge and small TLB entries for the same virtual address
2471 * to avoid the risk of CPU bugs in that area.
2472 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002473 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002474 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002475 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002476
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002477 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002478 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002479 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002480
2481 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002482 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002483 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002484 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002485 /*
2486 * We can only use set_pmd_at when establishing
2487 * hugepmds and never for establishing regular pmds that
2488 * points to regular pagetables. Use pmd_populate for that
2489 */
2490 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002491 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002492 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002493 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002494 }
2495
2496 /*
2497 * All pages are isolated and locked so anon_vma rmap
2498 * can't run anymore.
2499 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002500 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002501
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002502 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002503 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002504 __SetPageUptodate(new_page);
2505 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002506
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002507 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2508 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002509
2510 /*
2511 * spin_lock() below is not the equivalent of smp_wmb(), so
2512 * this is needed to avoid the copy_huge_page writes to become
2513 * visible after the set_pmd_at() write.
2514 */
2515 smp_wmb();
2516
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002517 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002518 BUG_ON(!pmd_none(*pmd));
2519 page_add_new_anon_rmap(new_page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002520 mem_cgroup_commit_charge(new_page, memcg, false);
2521 lru_cache_add_active_or_unevictable(new_page, vma);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002522 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002523 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002524 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002525 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002526
2527 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002528
Andrea Arcangeliba761492011-01-13 15:46:58 -08002529 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002530out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002531 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002532 return;
2533
Andrea Arcangelice83d212011-01-13 15:47:06 -08002534out:
Johannes Weiner00501b52014-08-08 14:19:20 -07002535 mem_cgroup_cancel_charge(new_page, memcg);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002536 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002537}
2538
2539static int khugepaged_scan_pmd(struct mm_struct *mm,
2540 struct vm_area_struct *vma,
2541 unsigned long address,
2542 struct page **hpage)
2543{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002544 pmd_t *pmd;
2545 pte_t *pte, *_pte;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002546 int ret = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002547 struct page *page;
2548 unsigned long _address;
2549 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002550 int node = NUMA_NO_NODE;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002551 bool writable = false, referenced = false;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002552
2553 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2554
Bob Liu62190492012-12-11 16:00:37 -08002555 pmd = mm_find_pmd(mm, address);
2556 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002557 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002558
Bob Liu9f1b8682013-11-12 15:07:37 -08002559 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002560 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2561 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2562 _pte++, _address += PAGE_SIZE) {
2563 pte_t pteval = *_pte;
2564 if (pte_none(pteval)) {
2565 if (++none <= khugepaged_max_ptes_none)
2566 continue;
2567 else
2568 goto out_unmap;
2569 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002570 if (!pte_present(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002571 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002572 if (pte_write(pteval))
2573 writable = true;
2574
Andrea Arcangeliba761492011-01-13 15:46:58 -08002575 page = vm_normal_page(vma, _address, pteval);
2576 if (unlikely(!page))
2577 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002578 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002579 * Record which node the original page is from and save this
2580 * information to khugepaged_node_load[].
2581 * Khupaged will allocate hugepage from the node has the max
2582 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002583 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002584 node = page_to_nid(page);
David Rientjes14a4e212014-08-06 16:07:29 -07002585 if (khugepaged_scan_abort(node))
2586 goto out_unmap;
Bob Liu9f1b8682013-11-12 15:07:37 -08002587 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002588 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002589 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2590 goto out_unmap;
Ebru Akagunduz10359212015-02-11 15:28:28 -08002591 /*
2592 * cannot use mapcount: can't collapse if there's a gup pin.
2593 * The page must only be referenced by the scanned process
2594 * and page swap cache.
2595 */
2596 if (page_count(page) != 1 + !!PageSwapCache(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002597 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002598 if (pte_young(pteval) || PageReferenced(page) ||
2599 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz10359212015-02-11 15:28:28 -08002600 referenced = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002601 }
Ebru Akagunduz10359212015-02-11 15:28:28 -08002602 if (referenced && writable)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002603 ret = 1;
2604out_unmap:
2605 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002606 if (ret) {
2607 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002608 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002609 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002610 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002611out:
2612 return ret;
2613}
2614
2615static void collect_mm_slot(struct mm_slot *mm_slot)
2616{
2617 struct mm_struct *mm = mm_slot->mm;
2618
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002619 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002620
2621 if (khugepaged_test_exit(mm)) {
2622 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002623 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002624 list_del(&mm_slot->mm_node);
2625
2626 /*
2627 * Not strictly needed because the mm exited already.
2628 *
2629 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2630 */
2631
2632 /* khugepaged_mm_lock actually not necessary for the below */
2633 free_mm_slot(mm_slot);
2634 mmdrop(mm);
2635 }
2636}
2637
2638static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2639 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002640 __releases(&khugepaged_mm_lock)
2641 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002642{
2643 struct mm_slot *mm_slot;
2644 struct mm_struct *mm;
2645 struct vm_area_struct *vma;
2646 int progress = 0;
2647
2648 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002649 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002650
2651 if (khugepaged_scan.mm_slot)
2652 mm_slot = khugepaged_scan.mm_slot;
2653 else {
2654 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2655 struct mm_slot, mm_node);
2656 khugepaged_scan.address = 0;
2657 khugepaged_scan.mm_slot = mm_slot;
2658 }
2659 spin_unlock(&khugepaged_mm_lock);
2660
2661 mm = mm_slot->mm;
2662 down_read(&mm->mmap_sem);
2663 if (unlikely(khugepaged_test_exit(mm)))
2664 vma = NULL;
2665 else
2666 vma = find_vma(mm, khugepaged_scan.address);
2667
2668 progress++;
2669 for (; vma; vma = vma->vm_next) {
2670 unsigned long hstart, hend;
2671
2672 cond_resched();
2673 if (unlikely(khugepaged_test_exit(mm))) {
2674 progress++;
2675 break;
2676 }
Bob Liufa475e52012-12-11 16:00:39 -08002677 if (!hugepage_vma_check(vma)) {
2678skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002679 progress++;
2680 continue;
2681 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002682 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2683 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002684 if (hstart >= hend)
2685 goto skip;
2686 if (khugepaged_scan.address > hend)
2687 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002688 if (khugepaged_scan.address < hstart)
2689 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002690 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002691
2692 while (khugepaged_scan.address < hend) {
2693 int ret;
2694 cond_resched();
2695 if (unlikely(khugepaged_test_exit(mm)))
2696 goto breakouterloop;
2697
2698 VM_BUG_ON(khugepaged_scan.address < hstart ||
2699 khugepaged_scan.address + HPAGE_PMD_SIZE >
2700 hend);
2701 ret = khugepaged_scan_pmd(mm, vma,
2702 khugepaged_scan.address,
2703 hpage);
2704 /* move to next address */
2705 khugepaged_scan.address += HPAGE_PMD_SIZE;
2706 progress += HPAGE_PMD_NR;
2707 if (ret)
2708 /* we released mmap_sem so break loop */
2709 goto breakouterloop_mmap_sem;
2710 if (progress >= pages)
2711 goto breakouterloop;
2712 }
2713 }
2714breakouterloop:
2715 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2716breakouterloop_mmap_sem:
2717
2718 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002719 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002720 /*
2721 * Release the current mm_slot if this mm is about to die, or
2722 * if we scanned all vmas of this mm.
2723 */
2724 if (khugepaged_test_exit(mm) || !vma) {
2725 /*
2726 * Make sure that if mm_users is reaching zero while
2727 * khugepaged runs here, khugepaged_exit will find
2728 * mm_slot not pointing to the exiting mm.
2729 */
2730 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2731 khugepaged_scan.mm_slot = list_entry(
2732 mm_slot->mm_node.next,
2733 struct mm_slot, mm_node);
2734 khugepaged_scan.address = 0;
2735 } else {
2736 khugepaged_scan.mm_slot = NULL;
2737 khugepaged_full_scans++;
2738 }
2739
2740 collect_mm_slot(mm_slot);
2741 }
2742
2743 return progress;
2744}
2745
2746static int khugepaged_has_work(void)
2747{
2748 return !list_empty(&khugepaged_scan.mm_head) &&
2749 khugepaged_enabled();
2750}
2751
2752static int khugepaged_wait_event(void)
2753{
2754 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002755 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002756}
2757
Xiao Guangrongd5169042012-10-08 16:29:48 -07002758static void khugepaged_do_scan(void)
2759{
2760 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002761 unsigned int progress = 0, pass_through_head = 0;
2762 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002763 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002764
2765 barrier(); /* write khugepaged_pages_to_scan to local stack */
2766
2767 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002768 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002769 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002770
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002771 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002772
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002773 if (unlikely(kthread_should_stop() || freezing(current)))
2774 break;
2775
Andrea Arcangeliba761492011-01-13 15:46:58 -08002776 spin_lock(&khugepaged_mm_lock);
2777 if (!khugepaged_scan.mm_slot)
2778 pass_through_head++;
2779 if (khugepaged_has_work() &&
2780 pass_through_head < 2)
2781 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002782 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002783 else
2784 progress = pages;
2785 spin_unlock(&khugepaged_mm_lock);
2786 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002787
Xiao Guangrongd5169042012-10-08 16:29:48 -07002788 if (!IS_ERR_OR_NULL(hpage))
2789 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002790}
2791
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002792static void khugepaged_wait_work(void)
2793{
2794 try_to_freeze();
2795
2796 if (khugepaged_has_work()) {
2797 if (!khugepaged_scan_sleep_millisecs)
2798 return;
2799
2800 wait_event_freezable_timeout(khugepaged_wait,
2801 kthread_should_stop(),
2802 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2803 return;
2804 }
2805
2806 if (khugepaged_enabled())
2807 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2808}
2809
Andrea Arcangeliba761492011-01-13 15:46:58 -08002810static int khugepaged(void *none)
2811{
2812 struct mm_slot *mm_slot;
2813
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002814 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002815 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002816
Xiao Guangrongb7231782012-10-08 16:29:54 -07002817 while (!kthread_should_stop()) {
2818 khugepaged_do_scan();
2819 khugepaged_wait_work();
2820 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002821
2822 spin_lock(&khugepaged_mm_lock);
2823 mm_slot = khugepaged_scan.mm_slot;
2824 khugepaged_scan.mm_slot = NULL;
2825 if (mm_slot)
2826 collect_mm_slot(mm_slot);
2827 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002828 return 0;
2829}
2830
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002831static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2832 unsigned long haddr, pmd_t *pmd)
2833{
2834 struct mm_struct *mm = vma->vm_mm;
2835 pgtable_t pgtable;
2836 pmd_t _pmd;
2837 int i;
2838
Joerg Roedel34ee6452014-11-13 13:46:09 +11002839 pmdp_clear_flush_notify(vma, haddr, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002840 /* leave pmd empty until pte is filled */
2841
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002842 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002843 pmd_populate(mm, &_pmd, pgtable);
2844
2845 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2846 pte_t *pte, entry;
2847 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2848 entry = pte_mkspecial(entry);
2849 pte = pte_offset_map(&_pmd, haddr);
2850 VM_BUG_ON(!pte_none(*pte));
2851 set_pte_at(mm, haddr, pte, entry);
2852 pte_unmap(pte);
2853 }
2854 smp_wmb(); /* make pte visible before pmd */
2855 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002856 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002857}
2858
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002859void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2860 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002861{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002862 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002863 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002864 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002865 unsigned long haddr = address & HPAGE_PMD_MASK;
2866 unsigned long mmun_start; /* For mmu_notifiers */
2867 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002868
2869 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002870
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002871 mmun_start = haddr;
2872 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002873again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002874 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002875 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002876 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002877 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002878 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2879 return;
2880 }
2881 if (is_huge_zero_pmd(*pmd)) {
2882 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002883 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002884 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002885 return;
2886 }
2887 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002888 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002889 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002890 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002891 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002892
2893 split_huge_page(page);
2894
2895 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002896
2897 /*
2898 * We don't always have down_write of mmap_sem here: a racing
2899 * do_huge_pmd_wp_page() might have copied-on-write to another
2900 * huge page before our split_huge_page() got the anon_vma lock.
2901 */
2902 if (unlikely(pmd_trans_huge(*pmd)))
2903 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002904}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002905
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002906void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2907 pmd_t *pmd)
2908{
2909 struct vm_area_struct *vma;
2910
2911 vma = find_vma(mm, address);
2912 BUG_ON(vma == NULL);
2913 split_huge_page_pmd(vma, address, pmd);
2914}
2915
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002916static void split_huge_page_address(struct mm_struct *mm,
2917 unsigned long address)
2918{
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002919 pgd_t *pgd;
2920 pud_t *pud;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002921 pmd_t *pmd;
2922
2923 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2924
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -07002925 pgd = pgd_offset(mm, address);
2926 if (!pgd_present(*pgd))
2927 return;
2928
2929 pud = pud_offset(pgd, address);
2930 if (!pud_present(*pud))
2931 return;
2932
2933 pmd = pmd_offset(pud, address);
2934 if (!pmd_present(*pmd))
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002935 return;
2936 /*
2937 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2938 * materialize from under us.
2939 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002940 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002941}
2942
2943void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2944 unsigned long start,
2945 unsigned long end,
2946 long adjust_next)
2947{
2948 /*
2949 * If the new start address isn't hpage aligned and it could
2950 * previously contain an hugepage: check if we need to split
2951 * an huge pmd.
2952 */
2953 if (start & ~HPAGE_PMD_MASK &&
2954 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2955 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2956 split_huge_page_address(vma->vm_mm, start);
2957
2958 /*
2959 * If the new end address isn't hpage aligned and it could
2960 * previously contain an hugepage: check if we need to split
2961 * an huge pmd.
2962 */
2963 if (end & ~HPAGE_PMD_MASK &&
2964 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2965 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2966 split_huge_page_address(vma->vm_mm, end);
2967
2968 /*
2969 * If we're also updating the vma->vm_next->vm_start, if the new
2970 * vm_next->vm_start isn't page aligned and it could previously
2971 * contain an hugepage: check if we need to split an huge pmd.
2972 */
2973 if (adjust_next > 0) {
2974 struct vm_area_struct *next = vma->vm_next;
2975 unsigned long nstart = next->vm_start;
2976 nstart += adjust_next << PAGE_SHIFT;
2977 if (nstart & ~HPAGE_PMD_MASK &&
2978 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2979 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2980 split_huge_page_address(next->vm_mm, nstart);
2981 }
2982}