projects / GitHub / exynos8895 / android_kernel_samsung_universal8895.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge 4.4.88 into android-4.4
[GitHub/exynos8895/android_kernel_samsung_universal8895.git] / mm / huge_memory.c
CommitLineData
71e3aac0
AA		 1/*
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
ae3a8c1c
AM		 8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
71e3aac0
AA		 10#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>
97ae1749 17#include <linux/shrinker.h>
ba76149f 18#include <linux/mm_inline.h>
4897c765 19#include <linux/dax.h>
ba76149f
AA		 20#include <linux/kthread.h>
21#include <linux/khugepaged.h>
878aee7d 22#include <linux/freezer.h>
a664b2d8 23#include <linux/mman.h>
325adeb5 24#include <linux/pagemap.h>
4daae3b4 25#include <linux/migrate.h>
43b5fbbd 26#include <linux/hashtable.h>
6b251fc9 27#include <linux/userfaultfd_k.h>
33c3fc71 28#include <linux/page_idle.h>
97ae1749 29
71e3aac0
AA		 30#include <asm/tlb.h>
31#include <asm/pgalloc.h>
32#include "internal.h"
33
ba76149f 34/*
8bfa3f9a
JW		 35 * By default transparent hugepage support is disabled in order that avoid
36 * to risk increase the memory footprint of applications without a guaranteed
37 * benefit. When transparent hugepage support is enabled, is for all mappings,
38 * and khugepaged scans all mappings.
39 * Defrag is invoked by khugepaged hugepage allocations and by page faults
40 * for all hugepage allocations.
ba76149f 41 */
71e3aac0 42unsigned long transparent_hugepage_flags __read_mostly =
13ece886 43#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
ba76149f 44 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
13ece886
AA		 45#endif
46#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
47 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
48#endif
d39d33c3 49 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
79da5407
KS		 50 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
51 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
ba76149f
AA		 52
53/* default scan 8*512 pte (or vmas) every 30 second */
54static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
55static unsigned int khugepaged_pages_collapsed;
56static unsigned int khugepaged_full_scans;
57static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58/* during fragmentation poll the hugepage allocator once every minute */
59static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60static struct task_struct *khugepaged_thread __read_mostly;
61static DEFINE_MUTEX(khugepaged_mutex);
62static DEFINE_SPINLOCK(khugepaged_mm_lock);
63static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
64/*
65 * default collapse hugepages if there is at least one pte mapped like
66 * it would have happened if the vma was large enough during page
67 * fault.
68 */
69static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
70
71static int khugepaged(void *none);
ba76149f 72static int khugepaged_slab_init(void);
65ebb64f 73static void khugepaged_slab_exit(void);
ba76149f 74
43b5fbbd
SL		 75#define MM_SLOTS_HASH_BITS 10
76static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
77
ba76149f
AA		 78static struct kmem_cache *mm_slot_cache __read_mostly;
79
80/**
81 * struct mm_slot - hash lookup from mm to mm_slot
82 * @hash: hash collision list
83 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
84 * @mm: the mm that this information is valid for
85 */
86struct mm_slot {
87 struct hlist_node hash;
88 struct list_head mm_node;
89 struct mm_struct *mm;
90};
91
92/**
93 * struct khugepaged_scan - cursor for scanning
94 * @mm_head: the head of the mm list to scan
95 * @mm_slot: the current mm_slot we are scanning
96 * @address: the next address inside that to be scanned
97 *
98 * There is only the one khugepaged_scan instance of this cursor structure.
99 */
100struct khugepaged_scan {
101 struct list_head mm_head;
102 struct mm_slot *mm_slot;
103 unsigned long address;
2f1da642
HS		 104};
105static struct khugepaged_scan khugepaged_scan = {
ba76149f
AA		 106 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107};
108
f000565a 109
2c0b80d4 110static void set_recommended_min_free_kbytes(void)
f000565a
AA		 111{
112 struct zone *zone;
113 int nr_zones = 0;
114 unsigned long recommended_min;
f000565a 115
f000565a
AA		 116 for_each_populated_zone(zone)
117 nr_zones++;
118
974a786e 119 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
f000565a
AA		 120 recommended_min = pageblock_nr_pages * nr_zones * 2;
121
122 /*
123 * Make sure that on average at least two pageblocks are almost free
124 * of another type, one for a migratetype to fall back to and a
125 * second to avoid subsequent fallbacks of other types There are 3
126 * MIGRATE_TYPES we care about.
127 */
128 recommended_min += pageblock_nr_pages * nr_zones *
129 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
130
131 /* don't ever allow to reserve more than 5% of the lowmem */
132 recommended_min = min(recommended_min,
133 (unsigned long) nr_free_buffer_pages() / 20);
134 recommended_min <<= (PAGE_SHIFT-10);
135
42aa83cb
HP		 136 if (recommended_min > min_free_kbytes) {
137 if (user_min_free_kbytes >= 0)
138 pr_info("raising min_free_kbytes from %d to %lu "
139 "to help transparent hugepage allocations\n",
140 min_free_kbytes, recommended_min);
141
f000565a 142 min_free_kbytes = recommended_min;
42aa83cb 143 }
f000565a 144 setup_per_zone_wmarks();
f000565a 145}
f000565a 146
79553da2 147static int start_stop_khugepaged(void)
ba76149f
AA		 148{
149 int err = 0;
150 if (khugepaged_enabled()) {
ba76149f
AA		 151 if (!khugepaged_thread)
152 khugepaged_thread = kthread_run(khugepaged, NULL,
153 "khugepaged");
18e8e5c7 154 if (IS_ERR(khugepaged_thread)) {
ae3a8c1c 155 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
ba76149f
AA		 156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
79553da2 158 goto fail;
ba76149f 159 }
911891af
XG		 160
161 if (!list_empty(&khugepaged_scan.mm_head))
ba76149f 162 wake_up_interruptible(&khugepaged_wait);
f000565a
AA		 163
164 set_recommended_min_free_kbytes();
911891af 165 } else if (khugepaged_thread) {
911891af
XG		 166 kthread_stop(khugepaged_thread);
167 khugepaged_thread = NULL;
168 }
79553da2 169fail:
ba76149f
AA		 170 return err;
171}
71e3aac0 172
97ae1749 173static atomic_t huge_zero_refcount;
56873f43 174struct page *huge_zero_page __read_mostly;
4a6c1297 175
fc437044 176struct page *get_huge_zero_page(void)
97ae1749
KS		 177{
178 struct page *zero_page;
179retry:
180 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
4db0c3c2 181 return READ_ONCE(huge_zero_page);
97ae1749
KS		 182
183 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
4a6c1297 184 HPAGE_PMD_ORDER);
d8a8e1f0
KS		 185 if (!zero_page) {
186 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
5918d10a 187 return NULL;
d8a8e1f0
KS		 188 }
189 count_vm_event(THP_ZERO_PAGE_ALLOC);
97ae1749 190 preempt_disable();
5918d10a 191 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
97ae1749 192 preempt_enable();
5ddacbe9 193 __free_pages(zero_page, compound_order(zero_page));
97ae1749
KS		 194 goto retry;
195 }
196
197 /* We take additional reference here. It will be put back by shrinker */
198 atomic_set(&huge_zero_refcount, 2);
199 preempt_enable();
4db0c3c2 200 return READ_ONCE(huge_zero_page);
4a6c1297
KS		 201}
202
97ae1749 203static void put_huge_zero_page(void)
4a6c1297 204{
97ae1749
KS		 205 /*
206 * Counter should never go to zero here. Only shrinker can put
207 * last reference.
208 */
209 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
4a6c1297
KS		 210}
211
48896466
GC		 212static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
213 struct shrink_control *sc)
4a6c1297 214{
48896466
GC		 215 /* we can free zero page only if last reference remains */
216 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
217}
97ae1749 218
48896466
GC		 219static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
220 struct shrink_control *sc)
221{
97ae1749 222 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
5918d10a
KS		 223 struct page *zero_page = xchg(&huge_zero_page, NULL);
224 BUG_ON(zero_page == NULL);
5ddacbe9 225 __free_pages(zero_page, compound_order(zero_page));
48896466 226 return HPAGE_PMD_NR;
97ae1749
KS		 227 }
228
229 return 0;
4a6c1297
KS		 230}
231
97ae1749 232static struct shrinker huge_zero_page_shrinker = {
48896466
GC		 233 .count_objects = shrink_huge_zero_page_count,
234 .scan_objects = shrink_huge_zero_page_scan,
97ae1749
KS		 235 .seeks = DEFAULT_SEEKS,
236};
237
71e3aac0 238#ifdef CONFIG_SYSFS
ba76149f 239
71e3aac0
AA		 240static ssize_t double_flag_show(struct kobject *kobj,
241 struct kobj_attribute *attr, char *buf,
242 enum transparent_hugepage_flag enabled,
243 enum transparent_hugepage_flag req_madv)
244{
245 if (test_bit(enabled, &transparent_hugepage_flags)) {
246 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
247 return sprintf(buf, "[always] madvise never\n");
248 } else if (test_bit(req_madv, &transparent_hugepage_flags))
249 return sprintf(buf, "always [madvise] never\n");
250 else
251 return sprintf(buf, "always madvise [never]\n");
252}
253static ssize_t double_flag_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count,
256 enum transparent_hugepage_flag enabled,
257 enum transparent_hugepage_flag req_madv)
258{
259 if (!memcmp("always", buf,
260 min(sizeof("always")-1, count))) {
261 set_bit(enabled, &transparent_hugepage_flags);
262 clear_bit(req_madv, &transparent_hugepage_flags);
263 } else if (!memcmp("madvise", buf,
264 min(sizeof("madvise")-1, count))) {
265 clear_bit(enabled, &transparent_hugepage_flags);
266 set_bit(req_madv, &transparent_hugepage_flags);
267 } else if (!memcmp("never", buf,
268 min(sizeof("never")-1, count))) {
269 clear_bit(enabled, &transparent_hugepage_flags);
270 clear_bit(req_madv, &transparent_hugepage_flags);
271 } else
272 return -EINVAL;
273
274 return count;
275}
276
277static ssize_t enabled_show(struct kobject *kobj,
278 struct kobj_attribute *attr, char *buf)
279{
280 return double_flag_show(kobj, attr, buf,
281 TRANSPARENT_HUGEPAGE_FLAG,
282 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
283}
284static ssize_t enabled_store(struct kobject *kobj,
285 struct kobj_attribute *attr,
286 const char *buf, size_t count)
287{
ba76149f
AA		 288 ssize_t ret;
289
290 ret = double_flag_store(kobj, attr, buf, count,
291 TRANSPARENT_HUGEPAGE_FLAG,
292 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
293
294 if (ret > 0) {
911891af
XG		 295 int err;
296
297 mutex_lock(&khugepaged_mutex);
79553da2 298 err = start_stop_khugepaged();
911891af
XG		 299 mutex_unlock(&khugepaged_mutex);
300
ba76149f
AA		 301 if (err)
302 ret = err;
303 }
304
305 return ret;
71e3aac0
AA		 306}
307static struct kobj_attribute enabled_attr =
308 __ATTR(enabled, 0644, enabled_show, enabled_store);
309
310static ssize_t single_flag_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf,
312 enum transparent_hugepage_flag flag)
313{
e27e6151
BH		 314 return sprintf(buf, "%d\n",
315 !!test_bit(flag, &transparent_hugepage_flags));
71e3aac0 316}
e27e6151 317
71e3aac0
AA		 318static ssize_t single_flag_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count,
321 enum transparent_hugepage_flag flag)
322{
e27e6151
BH		 323 unsigned long value;
324 int ret;
325
326 ret = kstrtoul(buf, 10, &value);
327 if (ret < 0)
328 return ret;
329 if (value > 1)
330 return -EINVAL;
331
332 if (value)
71e3aac0 333 set_bit(flag, &transparent_hugepage_flags);
e27e6151 334 else
71e3aac0 335 clear_bit(flag, &transparent_hugepage_flags);
71e3aac0
AA		 336
337 return count;
338}
339
340/*
341 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
342 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
343 * memory just to allocate one more hugepage.
344 */
345static ssize_t defrag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf)
347{
348 return double_flag_show(kobj, attr, buf,
349 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
350 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
351}
352static ssize_t defrag_store(struct kobject *kobj,
353 struct kobj_attribute *attr,
354 const char *buf, size_t count)
355{
356 return double_flag_store(kobj, attr, buf, count,
357 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
358 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
359}
360static struct kobj_attribute defrag_attr =
361 __ATTR(defrag, 0644, defrag_show, defrag_store);
362
79da5407
KS		 363static ssize_t use_zero_page_show(struct kobject *kobj,
364 struct kobj_attribute *attr, char *buf)
365{
366 return single_flag_show(kobj, attr, buf,
367 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
368}
369static ssize_t use_zero_page_store(struct kobject *kobj,
370 struct kobj_attribute *attr, const char *buf, size_t count)
371{
372 return single_flag_store(kobj, attr, buf, count,
373 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
374}
375static struct kobj_attribute use_zero_page_attr =
376 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
71e3aac0
AA		 377#ifdef CONFIG_DEBUG_VM
378static ssize_t debug_cow_show(struct kobject *kobj,
379 struct kobj_attribute *attr, char *buf)
380{
381 return single_flag_show(kobj, attr, buf,
382 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
383}
384static ssize_t debug_cow_store(struct kobject *kobj,
385 struct kobj_attribute *attr,
386 const char *buf, size_t count)
387{
388 return single_flag_store(kobj, attr, buf, count,
389 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
390}
391static struct kobj_attribute debug_cow_attr =
392 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
393#endif /* CONFIG_DEBUG_VM */
394
395static struct attribute *hugepage_attr[] = {
396 &enabled_attr.attr,
397 &defrag_attr.attr,
79da5407 398 &use_zero_page_attr.attr,
71e3aac0
AA		 399#ifdef CONFIG_DEBUG_VM
400 &debug_cow_attr.attr,
401#endif
402 NULL,
403};
404
405static struct attribute_group hugepage_attr_group = {
406 .attrs = hugepage_attr,
ba76149f
AA		 407};
408
409static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
410 struct kobj_attribute *attr,
411 char *buf)
412{
413 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
414}
415
416static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
417 struct kobj_attribute *attr,
418 const char *buf, size_t count)
419{
420 unsigned long msecs;
421 int err;
422
3dbb95f7 423 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 424 if (err || msecs > UINT_MAX)
425 return -EINVAL;
426
427 khugepaged_scan_sleep_millisecs = msecs;
428 wake_up_interruptible(&khugepaged_wait);
429
430 return count;
431}
432static struct kobj_attribute scan_sleep_millisecs_attr =
433 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
434 scan_sleep_millisecs_store);
435
436static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
437 struct kobj_attribute *attr,
438 char *buf)
439{
440 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
441}
442
443static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
444 struct kobj_attribute *attr,
445 const char *buf, size_t count)
446{
447 unsigned long msecs;
448 int err;
449
3dbb95f7 450 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 451 if (err || msecs > UINT_MAX)
452 return -EINVAL;
453
454 khugepaged_alloc_sleep_millisecs = msecs;
455 wake_up_interruptible(&khugepaged_wait);
456
457 return count;
458}
459static struct kobj_attribute alloc_sleep_millisecs_attr =
460 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
461 alloc_sleep_millisecs_store);
462
463static ssize_t pages_to_scan_show(struct kobject *kobj,
464 struct kobj_attribute *attr,
465 char *buf)
466{
467 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
468}
469static ssize_t pages_to_scan_store(struct kobject *kobj,
470 struct kobj_attribute *attr,
471 const char *buf, size_t count)
472{
473 int err;
474 unsigned long pages;
475
3dbb95f7 476 err = kstrtoul(buf, 10, &pages);
ba76149f
AA		 477 if (err || !pages || pages > UINT_MAX)
478 return -EINVAL;
479
480 khugepaged_pages_to_scan = pages;
481
482 return count;
483}
484static struct kobj_attribute pages_to_scan_attr =
485 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
486 pages_to_scan_store);
487
488static ssize_t pages_collapsed_show(struct kobject *kobj,
489 struct kobj_attribute *attr,
490 char *buf)
491{
492 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
493}
494static struct kobj_attribute pages_collapsed_attr =
495 __ATTR_RO(pages_collapsed);
496
497static ssize_t full_scans_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_full_scans);
502}
503static struct kobj_attribute full_scans_attr =
504 __ATTR_RO(full_scans);
505
506static ssize_t khugepaged_defrag_show(struct kobject *kobj,
507 struct kobj_attribute *attr, char *buf)
508{
509 return single_flag_show(kobj, attr, buf,
510 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
511}
512static ssize_t khugepaged_defrag_store(struct kobject *kobj,
513 struct kobj_attribute *attr,
514 const char *buf, size_t count)
515{
516 return single_flag_store(kobj, attr, buf, count,
517 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
518}
519static struct kobj_attribute khugepaged_defrag_attr =
520 __ATTR(defrag, 0644, khugepaged_defrag_show,
521 khugepaged_defrag_store);
522
523/*
524 * max_ptes_none controls if khugepaged should collapse hugepages over
525 * any unmapped ptes in turn potentially increasing the memory
526 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
527 * reduce the available free memory in the system as it
528 * runs. Increasing max_ptes_none will instead potentially reduce the
529 * free memory in the system during the khugepaged scan.
530 */
531static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
532 struct kobj_attribute *attr,
533 char *buf)
534{
535 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
536}
537static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
538 struct kobj_attribute *attr,
539 const char *buf, size_t count)
540{
541 int err;
542 unsigned long max_ptes_none;
543
3dbb95f7 544 err = kstrtoul(buf, 10, &max_ptes_none);
ba76149f
AA		 545 if (err || max_ptes_none > HPAGE_PMD_NR-1)
546 return -EINVAL;
547
548 khugepaged_max_ptes_none = max_ptes_none;
549
550 return count;
551}
552static struct kobj_attribute khugepaged_max_ptes_none_attr =
553 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
554 khugepaged_max_ptes_none_store);
555
556static struct attribute *khugepaged_attr[] = {
557 &khugepaged_defrag_attr.attr,
558 &khugepaged_max_ptes_none_attr.attr,
559 &pages_to_scan_attr.attr,
560 &pages_collapsed_attr.attr,
561 &full_scans_attr.attr,
562 &scan_sleep_millisecs_attr.attr,
563 &alloc_sleep_millisecs_attr.attr,
564 NULL,
565};
566
567static struct attribute_group khugepaged_attr_group = {
568 .attrs = khugepaged_attr,
569 .name = "khugepaged",
71e3aac0 570};
71e3aac0 571
569e5590 572static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
71e3aac0 573{
71e3aac0
AA		 574 int err;
575
569e5590
SL		 576 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
577 if (unlikely(!*hugepage_kobj)) {
ae3a8c1c 578 pr_err("failed to create transparent hugepage kobject\n");
569e5590 579 return -ENOMEM;
ba76149f
AA		 580 }
581
569e5590 582 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
ba76149f 583 if (err) {
ae3a8c1c 584 pr_err("failed to register transparent hugepage group\n");
569e5590 585 goto delete_obj;
ba76149f
AA		 586 }
587
569e5590 588 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
ba76149f 589 if (err) {
ae3a8c1c 590 pr_err("failed to register transparent hugepage group\n");
569e5590 591 goto remove_hp_group;
ba76149f 592 }
569e5590
SL		 593
594 return 0;
595
596remove_hp_group:
597 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
598delete_obj:
599 kobject_put(*hugepage_kobj);
600 return err;
601}
602
603static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
604{
605 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
606 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
607 kobject_put(hugepage_kobj);
608}
609#else
610static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
611{
612 return 0;
613}
614
615static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
616{
617}
618#endif /* CONFIG_SYSFS */
619
620static int __init hugepage_init(void)
621{
622 int err;
623 struct kobject *hugepage_kobj;
624
625 if (!has_transparent_hugepage()) {
626 transparent_hugepage_flags = 0;
627 return -EINVAL;
628 }
629
630 err = hugepage_init_sysfs(&hugepage_kobj);
631 if (err)
65ebb64f 632 goto err_sysfs;
ba76149f
AA		 633
634 err = khugepaged_slab_init();
635 if (err)
65ebb64f 636 goto err_slab;
ba76149f 637
65ebb64f
KS		 638 err = register_shrinker(&huge_zero_page_shrinker);
639 if (err)
640 goto err_hzp_shrinker;
97ae1749 641
97562cd2
RR		 642 /*
643 * By default disable transparent hugepages on smaller systems,
644 * where the extra memory used could hurt more than TLB overhead
645 * is likely to save. The admin can still enable it through /sys.
646 */
79553da2 647 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
97562cd2 648 transparent_hugepage_flags = 0;
79553da2
KS		 649 return 0;
650 }
97562cd2 651
79553da2 652 err = start_stop_khugepaged();
65ebb64f
KS		 653 if (err)
654 goto err_khugepaged;
ba76149f 655
569e5590 656 return 0;
65ebb64f
KS		 657err_khugepaged:
658 unregister_shrinker(&huge_zero_page_shrinker);
659err_hzp_shrinker:
660 khugepaged_slab_exit();
661err_slab:
569e5590 662 hugepage_exit_sysfs(hugepage_kobj);
65ebb64f 663err_sysfs:
ba76149f 664 return err;
71e3aac0 665}
a64fb3cd 666subsys_initcall(hugepage_init);
71e3aac0
AA		 667
668static int __init setup_transparent_hugepage(char *str)
669{
670 int ret = 0;
671 if (!str)
672 goto out;
673 if (!strcmp(str, "always")) {
674 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
675 &transparent_hugepage_flags);
676 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
677 &transparent_hugepage_flags);
678 ret = 1;
679 } else if (!strcmp(str, "madvise")) {
680 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
681 &transparent_hugepage_flags);
682 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
683 &transparent_hugepage_flags);
684 ret = 1;
685 } else if (!strcmp(str, "never")) {
686 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
687 &transparent_hugepage_flags);
688 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
689 &transparent_hugepage_flags);
690 ret = 1;
691 }
692out:
693 if (!ret)
ae3a8c1c 694 pr_warn("transparent_hugepage= cannot parse, ignored\n");
71e3aac0
AA		 695 return ret;
696}
697__setup("transparent_hugepage=", setup_transparent_hugepage);
698
b32967ff 699pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
71e3aac0
AA		 700{
701 if (likely(vma->vm_flags & VM_WRITE))
702 pmd = pmd_mkwrite(pmd);
703 return pmd;
704}
705
3122359a 706static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
b3092b3b
BL		 707{
708 pmd_t entry;
3122359a 709 entry = mk_pmd(page, prot);
b3092b3b
BL		 710 entry = pmd_mkhuge(entry);
711 return entry;
712}
713
71e3aac0
AA		 714static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
715 struct vm_area_struct *vma,
230c92a8 716 unsigned long address, pmd_t *pmd,
6b251fc9
AA		 717 struct page *page, gfp_t gfp,
718 unsigned int flags)
71e3aac0 719{
00501b53 720 struct mem_cgroup *memcg;
71e3aac0 721 pgtable_t pgtable;
c4088ebd 722 spinlock_t *ptl;
230c92a8 723 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 724
309381fe 725 VM_BUG_ON_PAGE(!PageCompound(page), page);
00501b53 726
6b251fc9
AA		 727 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
728 put_page(page);
729 count_vm_event(THP_FAULT_FALLBACK);
730 return VM_FAULT_FALLBACK;
731 }
00501b53 732
71e3aac0 733 pgtable = pte_alloc_one(mm, haddr);
00501b53
JW		 734 if (unlikely(!pgtable)) {
735 mem_cgroup_cancel_charge(page, memcg);
6b251fc9 736 put_page(page);
71e3aac0 737 return VM_FAULT_OOM;
00501b53 738 }
71e3aac0
AA		 739
740 clear_huge_page(page, haddr, HPAGE_PMD_NR);
52f37629
MK		 741 /*
742 * The memory barrier inside __SetPageUptodate makes sure that
743 * clear_huge_page writes become visible before the set_pmd_at()
744 * write.
745 */
71e3aac0
AA		 746 __SetPageUptodate(page);
747
c4088ebd 748 ptl = pmd_lock(mm, pmd);
71e3aac0 749 if (unlikely(!pmd_none(*pmd))) {
c4088ebd 750 spin_unlock(ptl);
00501b53 751 mem_cgroup_cancel_charge(page, memcg);
71e3aac0
AA		 752 put_page(page);
753 pte_free(mm, pgtable);
754 } else {
755 pmd_t entry;
6b251fc9
AA		 756
757 /* Deliver the page fault to userland */
758 if (userfaultfd_missing(vma)) {
759 int ret;
760
761 spin_unlock(ptl);
762 mem_cgroup_cancel_charge(page, memcg);
763 put_page(page);
764 pte_free(mm, pgtable);
230c92a8 765 ret = handle_userfault(vma, address, flags,
6b251fc9
AA		 766 VM_UFFD_MISSING);
767 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
768 return ret;
769 }
770
3122359a
KS		 771 entry = mk_huge_pmd(page, vma->vm_page_prot);
772 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
71e3aac0 773 page_add_new_anon_rmap(page, vma, haddr);
00501b53
JW		 774 mem_cgroup_commit_charge(page, memcg, false);
775 lru_cache_add_active_or_unevictable(page, vma);
6b0b50b0 776 pgtable_trans_huge_deposit(mm, pmd, pgtable);
71e3aac0 777 set_pmd_at(mm, haddr, pmd, entry);
71e3aac0 778 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
e1f56c89 779 atomic_long_inc(&mm->nr_ptes);
c4088ebd 780 spin_unlock(ptl);
6b251fc9 781 count_vm_event(THP_FAULT_ALLOC);
71e3aac0
AA		 782 }
783
aa2e878e 784 return 0;
71e3aac0
AA		 785}
786
cc5d462f 787static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
0bbbc0b3 788{
71baba4b 789 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
0bbbc0b3
AA		 790}
791
c4088ebd 792/* Caller must hold page table lock. */
d295e341 793static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
97ae1749 794 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
5918d10a 795 struct page *zero_page)
fc9fe822
KS		 796{
797 pmd_t entry;
7c414164
AM		 798 if (!pmd_none(*pmd))
799 return false;
5918d10a 800 entry = mk_pmd(zero_page, vma->vm_page_prot);
fc9fe822 801 entry = pmd_mkhuge(entry);
6b0b50b0 802 pgtable_trans_huge_deposit(mm, pmd, pgtable);
fc9fe822 803 set_pmd_at(mm, haddr, pmd, entry);
e1f56c89 804 atomic_long_inc(&mm->nr_ptes);
7c414164 805 return true;
fc9fe822
KS		 806}
807
71e3aac0
AA		 808int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
809 unsigned long address, pmd_t *pmd,
810 unsigned int flags)
811{
077fcf11 812 gfp_t gfp;
71e3aac0
AA		 813 struct page *page;
814 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 815
128ec037 816 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
c0292554 817 return VM_FAULT_FALLBACK;
128ec037
KS		 818 if (unlikely(anon_vma_prepare(vma)))
819 return VM_FAULT_OOM;
6d50e60c 820 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
128ec037 821 return VM_FAULT_OOM;
593befa6 822 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
128ec037 823 transparent_hugepage_use_zero_page()) {
c4088ebd 824 spinlock_t *ptl;
128ec037
KS		 825 pgtable_t pgtable;
826 struct page *zero_page;
827 bool set;
6b251fc9 828 int ret;
128ec037
KS		 829 pgtable = pte_alloc_one(mm, haddr);
830 if (unlikely(!pgtable))
ba76149f 831 return VM_FAULT_OOM;
128ec037
KS		 832 zero_page = get_huge_zero_page();
833 if (unlikely(!zero_page)) {
834 pte_free(mm, pgtable);
81ab4201 835 count_vm_event(THP_FAULT_FALLBACK);
c0292554 836 return VM_FAULT_FALLBACK;
b9bbfbe3 837 }
c4088ebd 838 ptl = pmd_lock(mm, pmd);
6b251fc9
AA		 839 ret = 0;
840 set = false;
841 if (pmd_none(*pmd)) {
842 if (userfaultfd_missing(vma)) {
843 spin_unlock(ptl);
230c92a8 844 ret = handle_userfault(vma, address, flags,
6b251fc9
AA		 845 VM_UFFD_MISSING);
846 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
847 } else {
848 set_huge_zero_page(pgtable, mm, vma,
849 haddr, pmd,
850 zero_page);
851 spin_unlock(ptl);
852 set = true;
853 }
854 } else
855 spin_unlock(ptl);
128ec037
KS		 856 if (!set) {
857 pte_free(mm, pgtable);
858 put_huge_zero_page();
edad9d2c 859 }
6b251fc9 860 return ret;
71e3aac0 861 }
077fcf11
AK		 862 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
863 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
128ec037
KS		 864 if (unlikely(!page)) {
865 count_vm_event(THP_FAULT_FALLBACK);
c0292554 866 return VM_FAULT_FALLBACK;
128ec037 867 }
230c92a8
AA		 868 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
869 flags);
71e3aac0
AA		 870}
871
ae18d6dc 872static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
5cad465d
MW		 873 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
874{
875 struct mm_struct *mm = vma->vm_mm;
876 pmd_t entry;
877 spinlock_t *ptl;
878
879 ptl = pmd_lock(mm, pmd);
880 if (pmd_none(*pmd)) {
881 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
882 if (write) {
883 entry = pmd_mkyoung(pmd_mkdirty(entry));
884 entry = maybe_pmd_mkwrite(entry, vma);
885 }
886 set_pmd_at(mm, addr, pmd, entry);
887 update_mmu_cache_pmd(vma, addr, pmd);
888 }
889 spin_unlock(ptl);
5cad465d
MW		 890}
891
892int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
893 pmd_t *pmd, unsigned long pfn, bool write)
894{
895 pgprot_t pgprot = vma->vm_page_prot;
896 /*
897 * If we had pmd_special, we could avoid all these restrictions,
898 * but we need to be consistent with PTEs and architectures that
899 * can't support a 'special' bit.
900 */
901 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
902 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
903 (VM_PFNMAP|VM_MIXEDMAP));
904 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
905 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
906
907 if (addr < vma->vm_start || addr >= vma->vm_end)
908 return VM_FAULT_SIGBUS;
909 if (track_pfn_insert(vma, &pgprot, pfn))
910 return VM_FAULT_SIGBUS;
ae18d6dc
MW		 911 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
912 return VM_FAULT_NOPAGE;
5cad465d
MW		 913}
914
71e3aac0
AA		 915int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
917 struct vm_area_struct *vma)
918{
c4088ebd 919 spinlock_t *dst_ptl, *src_ptl;
71e3aac0
AA		 920 struct page *src_page;
921 pmd_t pmd;
922 pgtable_t pgtable;
923 int ret;
924
925 ret = -ENOMEM;
926 pgtable = pte_alloc_one(dst_mm, addr);
927 if (unlikely(!pgtable))
928 goto out;
929
c4088ebd
KS		 930 dst_ptl = pmd_lock(dst_mm, dst_pmd);
931 src_ptl = pmd_lockptr(src_mm, src_pmd);
932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
71e3aac0
AA		 933
934 ret = -EAGAIN;
935 pmd = *src_pmd;
936 if (unlikely(!pmd_trans_huge(pmd))) {
937 pte_free(dst_mm, pgtable);
938 goto out_unlock;
939 }
fc9fe822 940 /*
c4088ebd 941 * When page table lock is held, the huge zero pmd should not be
fc9fe822
KS		 942 * under splitting since we don't split the page itself, only pmd to
943 * a page table.
944 */
945 if (is_huge_zero_pmd(pmd)) {
5918d10a 946 struct page *zero_page;
97ae1749
KS		 947 /*
948 * get_huge_zero_page() will never allocate a new page here,
949 * since we already have a zero page to copy. It just takes a
950 * reference.
951 */
5918d10a 952 zero_page = get_huge_zero_page();
6b251fc9 953 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
5918d10a 954 zero_page);
fc9fe822
KS		 955 ret = 0;
956 goto out_unlock;
957 }
de466bd6 958
71e3aac0
AA		 959 if (unlikely(pmd_trans_splitting(pmd))) {
960 /* split huge page running from under us */
c4088ebd
KS		 961 spin_unlock(src_ptl);
962 spin_unlock(dst_ptl);
71e3aac0
AA		 963 pte_free(dst_mm, pgtable);
964
965 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
966 goto out;
967 }
968 src_page = pmd_page(pmd);
309381fe 969 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
71e3aac0
AA		 970 get_page(src_page);
971 page_dup_rmap(src_page);
972 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
973
974 pmdp_set_wrprotect(src_mm, addr, src_pmd);
975 pmd = pmd_mkold(pmd_wrprotect(pmd));
6b0b50b0 976 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
71e3aac0 977 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
e1f56c89 978 atomic_long_inc(&dst_mm->nr_ptes);
71e3aac0
AA		 979
980 ret = 0;
981out_unlock:
c4088ebd
KS		 982 spin_unlock(src_ptl);
983 spin_unlock(dst_ptl);
71e3aac0
AA		 984out:
985 return ret;
986}
987
a1dd450b
WD		 988void huge_pmd_set_accessed(struct mm_struct *mm,
989 struct vm_area_struct *vma,
990 unsigned long address,
991 pmd_t *pmd, pmd_t orig_pmd,
992 int dirty)
993{
c4088ebd 994 spinlock_t *ptl;
a1dd450b
WD		 995 pmd_t entry;
996 unsigned long haddr;
997
c4088ebd 998 ptl = pmd_lock(mm, pmd);
a1dd450b
WD		 999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000 goto unlock;
1001
1002 entry = pmd_mkyoung(orig_pmd);
1003 haddr = address & HPAGE_PMD_MASK;
1004 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1005 update_mmu_cache_pmd(vma, address, pmd);
1006
1007unlock:
c4088ebd 1008 spin_unlock(ptl);
a1dd450b
WD		 1009}
1010
5338a937
HD		 1011/*
1012 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1013 * during copy_user_huge_page()'s copy_page_rep(): in the case when
1014 * the source page gets split and a tail freed before copy completes.
1015 * Called under pmd_lock of checked pmd, so safe from splitting itself.
1016 */
1017static void get_user_huge_page(struct page *page)
1018{
1019 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1020 struct page *endpage = page + HPAGE_PMD_NR;
1021
1022 atomic_add(HPAGE_PMD_NR, &page->_count);
1023 while (++page < endpage)
1024 get_huge_page_tail(page);
1025 } else {
1026 get_page(page);
1027 }
1028}
1029
1030static void put_user_huge_page(struct page *page)
1031{
1032 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1033 struct page *endpage = page + HPAGE_PMD_NR;
1034
1035 while (page < endpage)
1036 put_page(page++);
1037 } else {
1038 put_page(page);
1039 }
1040}
1041
71e3aac0
AA		 1042static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1043 struct vm_area_struct *vma,
1044 unsigned long address,
1045 pmd_t *pmd, pmd_t orig_pmd,
1046 struct page *page,
1047 unsigned long haddr)
1048{
00501b53 1049 struct mem_cgroup *memcg;
c4088ebd 1050 spinlock_t *ptl;
71e3aac0
AA		 1051 pgtable_t pgtable;
1052 pmd_t _pmd;
1053 int ret = 0, i;
1054 struct page **pages;
2ec74c3e
SG		 1055 unsigned long mmun_start; /* For mmu_notifiers */
1056 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1057
1058 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1059 GFP_KERNEL);
1060 if (unlikely(!pages)) {
1061 ret |= VM_FAULT_OOM;
1062 goto out;
1063 }
1064
1065 for (i = 0; i < HPAGE_PMD_NR; i++) {
cc5d462f
AK		 1066 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1067 __GFP_OTHER_NODE,
19ee151e 1068 vma, address, page_to_nid(page));
b9bbfbe3 1069 if (unlikely(!pages[i] ||
00501b53
JW		 1070 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1071 &memcg))) {
b9bbfbe3 1072 if (pages[i])
71e3aac0 1073 put_page(pages[i]);
b9bbfbe3 1074 while (--i >= 0) {
00501b53
JW		 1075 memcg = (void *)page_private(pages[i]);
1076 set_page_private(pages[i], 0);
1077 mem_cgroup_cancel_charge(pages[i], memcg);
b9bbfbe3
AA		 1078 put_page(pages[i]);
1079 }
71e3aac0
AA		 1080 kfree(pages);
1081 ret |= VM_FAULT_OOM;
1082 goto out;
1083 }
00501b53 1084 set_page_private(pages[i], (unsigned long)memcg);
71e3aac0
AA		 1085 }
1086
1087 for (i = 0; i < HPAGE_PMD_NR; i++) {
1088 copy_user_highpage(pages[i], page + i,
0089e485 1089 haddr + PAGE_SIZE * i, vma);
71e3aac0
AA		 1090 __SetPageUptodate(pages[i]);
1091 cond_resched();
1092 }
1093
2ec74c3e
SG		 1094 mmun_start = haddr;
1095 mmun_end = haddr + HPAGE_PMD_SIZE;
1096 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1097
c4088ebd 1098 ptl = pmd_lock(mm, pmd);
71e3aac0
AA		 1099 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1100 goto out_free_pages;
309381fe 1101 VM_BUG_ON_PAGE(!PageHead(page), page);
71e3aac0 1102
8809aa2d 1103 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
71e3aac0
AA		 1104 /* leave pmd empty until pte is filled */
1105
6b0b50b0 1106 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1107 pmd_populate(mm, &_pmd, pgtable);
1108
1109 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1110 pte_t *pte, entry;
1111 entry = mk_pte(pages[i], vma->vm_page_prot);
1112 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
00501b53
JW		 1113 memcg = (void *)page_private(pages[i]);
1114 set_page_private(pages[i], 0);
71e3aac0 1115 page_add_new_anon_rmap(pages[i], vma, haddr);
00501b53
JW		 1116 mem_cgroup_commit_charge(pages[i], memcg, false);
1117 lru_cache_add_active_or_unevictable(pages[i], vma);
71e3aac0
AA		 1118 pte = pte_offset_map(&_pmd, haddr);
1119 VM_BUG_ON(!pte_none(*pte));
1120 set_pte_at(mm, haddr, pte, entry);
1121 pte_unmap(pte);
1122 }
1123 kfree(pages);
1124
71e3aac0
AA		 1125 smp_wmb(); /* make pte visible before pmd */
1126 pmd_populate(mm, pmd, pgtable);
1127 page_remove_rmap(page);
c4088ebd 1128 spin_unlock(ptl);
71e3aac0 1129
2ec74c3e
SG		 1130 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1131
71e3aac0
AA		 1132 ret |= VM_FAULT_WRITE;
1133 put_page(page);
1134
1135out:
1136 return ret;
1137
1138out_free_pages:
c4088ebd 1139 spin_unlock(ptl);
2ec74c3e 1140 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b9bbfbe3 1141 for (i = 0; i < HPAGE_PMD_NR; i++) {
00501b53
JW		 1142 memcg = (void *)page_private(pages[i]);
1143 set_page_private(pages[i], 0);
1144 mem_cgroup_cancel_charge(pages[i], memcg);
71e3aac0 1145 put_page(pages[i]);
b9bbfbe3 1146 }
71e3aac0
AA		 1147 kfree(pages);
1148 goto out;
1149}
1150
1151int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1152 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1153{
c4088ebd 1154 spinlock_t *ptl;
71e3aac0 1155 int ret = 0;
93b4796d 1156 struct page *page = NULL, *new_page;
00501b53 1157 struct mem_cgroup *memcg;
71e3aac0 1158 unsigned long haddr;
2ec74c3e
SG		 1159 unsigned long mmun_start; /* For mmu_notifiers */
1160 unsigned long mmun_end; /* For mmu_notifiers */
3b363692 1161 gfp_t huge_gfp; /* for allocation and charge */
71e3aac0 1162
c4088ebd 1163 ptl = pmd_lockptr(mm, pmd);
81d1b09c 1164 VM_BUG_ON_VMA(!vma->anon_vma, vma);
93b4796d
KS		 1165 haddr = address & HPAGE_PMD_MASK;
1166 if (is_huge_zero_pmd(orig_pmd))
1167 goto alloc;
c4088ebd 1168 spin_lock(ptl);
71e3aac0
AA		 1169 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1170 goto out_unlock;
1171
1172 page = pmd_page(orig_pmd);
309381fe 1173 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
71e3aac0
AA		 1174 if (page_mapcount(page) == 1) {
1175 pmd_t entry;
1176 entry = pmd_mkyoung(orig_pmd);
1177 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1178 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
b113da65 1179 update_mmu_cache_pmd(vma, address, pmd);
71e3aac0
AA		 1180 ret |= VM_FAULT_WRITE;
1181 goto out_unlock;
1182 }
5338a937 1183 get_user_huge_page(page);
c4088ebd 1184 spin_unlock(ptl);
93b4796d 1185alloc:
71e3aac0 1186 if (transparent_hugepage_enabled(vma) &&
077fcf11 1187 !transparent_hugepage_debug_cow()) {
3b363692
MH		 1188 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1189 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
077fcf11 1190 } else
71e3aac0
AA		 1191 new_page = NULL;
1192
1193 if (unlikely(!new_page)) {
eecc1e42 1194 if (!page) {
e9b71ca9
KS		 1195 split_huge_page_pmd(vma, address, pmd);
1196 ret |= VM_FAULT_FALLBACK;
93b4796d
KS		 1197 } else {
1198 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1199 pmd, orig_pmd, page, haddr);
9845cbbd 1200 if (ret & VM_FAULT_OOM) {
93b4796d 1201 split_huge_page(page);
9845cbbd
KS		 1202 ret |= VM_FAULT_FALLBACK;
1203 }
5338a937 1204 put_user_huge_page(page);
93b4796d 1205 }
17766dde 1206 count_vm_event(THP_FAULT_FALLBACK);
71e3aac0
AA		 1207 goto out;
1208 }
1209
3b363692 1210 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
b9bbfbe3 1211 put_page(new_page);
93b4796d
KS		 1212 if (page) {
1213 split_huge_page(page);
5338a937 1214 put_user_huge_page(page);
9845cbbd
KS		 1215 } else
1216 split_huge_page_pmd(vma, address, pmd);
1217 ret |= VM_FAULT_FALLBACK;
17766dde 1218 count_vm_event(THP_FAULT_FALLBACK);
b9bbfbe3
AA		 1219 goto out;
1220 }
1221
17766dde
DR		 1222 count_vm_event(THP_FAULT_ALLOC);
1223
eecc1e42 1224 if (!page)
93b4796d
KS		 1225 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1226 else
1227 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
71e3aac0
AA		 1228 __SetPageUptodate(new_page);
1229
2ec74c3e
SG		 1230 mmun_start = haddr;
1231 mmun_end = haddr + HPAGE_PMD_SIZE;
1232 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1233
c4088ebd 1234 spin_lock(ptl);
93b4796d 1235 if (page)
5338a937 1236 put_user_huge_page(page);
b9bbfbe3 1237 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
c4088ebd 1238 spin_unlock(ptl);
00501b53 1239 mem_cgroup_cancel_charge(new_page, memcg);
71e3aac0 1240 put_page(new_page);
2ec74c3e 1241 goto out_mn;
b9bbfbe3 1242 } else {
71e3aac0 1243 pmd_t entry;
3122359a
KS		 1244 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1245 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
8809aa2d 1246 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
71e3aac0 1247 page_add_new_anon_rmap(new_page, vma, haddr);
00501b53
JW		 1248 mem_cgroup_commit_charge(new_page, memcg, false);
1249 lru_cache_add_active_or_unevictable(new_page, vma);
71e3aac0 1250 set_pmd_at(mm, haddr, pmd, entry);
b113da65 1251 update_mmu_cache_pmd(vma, address, pmd);
eecc1e42 1252 if (!page) {
93b4796d 1253 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
97ae1749
KS		 1254 put_huge_zero_page();
1255 } else {
309381fe 1256 VM_BUG_ON_PAGE(!PageHead(page), page);
93b4796d
KS		 1257 page_remove_rmap(page);
1258 put_page(page);
1259 }
71e3aac0
AA		 1260 ret |= VM_FAULT_WRITE;
1261 }
c4088ebd 1262 spin_unlock(ptl);
2ec74c3e
SG		 1263out_mn:
1264 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1265out:
1266 return ret;
2ec74c3e 1267out_unlock:
c4088ebd 1268 spin_unlock(ptl);
2ec74c3e 1269 return ret;
71e3aac0
AA		 1270}
1271
8b26f53b
KF		 1272/*
1273 * FOLL_FORCE can write to even unwritable pmd's, but only
1274 * after we've gone through a COW cycle and they are dirty.
1275 */
1276static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1277{
1278 return pmd_write(pmd) ||
1279 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1280}
1281
b676b293 1282struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
71e3aac0
AA		 1283 unsigned long addr,
1284 pmd_t *pmd,
1285 unsigned int flags)
1286{
b676b293 1287 struct mm_struct *mm = vma->vm_mm;
71e3aac0
AA		 1288 struct page *page = NULL;
1289
c4088ebd 1290 assert_spin_locked(pmd_lockptr(mm, pmd));
71e3aac0 1291
8b26f53b 1292 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
71e3aac0
AA		 1293 goto out;
1294
85facf25
KS		 1295 /* Avoid dumping huge zero page */
1296 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1297 return ERR_PTR(-EFAULT);
1298
2b4847e7 1299 /* Full NUMA hinting faults to serialise migration in fault paths */
8a0516ed 1300 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
2b4847e7
MG		 1301 goto out;
1302
71e3aac0 1303 page = pmd_page(*pmd);
309381fe 1304 VM_BUG_ON_PAGE(!PageHead(page), page);
71e3aac0
AA		 1305 if (flags & FOLL_TOUCH) {
1306 pmd_t _pmd;
1307 /*
1308 * We should set the dirty bit only for FOLL_WRITE but
1309 * for now the dirty bit in the pmd is meaningless.
1310 * And if the dirty bit will become meaningful and
1311 * we'll only set it with FOLL_WRITE, an atomic
1312 * set_bit will be required on the pmd to set the
1313 * young bit, instead of the current set_pmd_at.
1314 */
1315 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
8663890a
AK		 1316 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1317 pmd, _pmd, 1))
1318 update_mmu_cache_pmd(vma, addr, pmd);
71e3aac0 1319 }
de60f5f1 1320 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
b676b293
DR		 1321 if (page->mapping && trylock_page(page)) {
1322 lru_add_drain();
1323 if (page->mapping)
1324 mlock_vma_page(page);
1325 unlock_page(page);
1326 }
1327 }
71e3aac0 1328 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
309381fe 1329 VM_BUG_ON_PAGE(!PageCompound(page), page);
71e3aac0 1330 if (flags & FOLL_GET)
70b50f94 1331 get_page_foll(page);
71e3aac0
AA		 1332
1333out:
1334 return page;
1335}
1336
d10e63f2 1337/* NUMA hinting page fault entry point for trans huge pmds */
4daae3b4
MG		 1338int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1339 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
d10e63f2 1340{
c4088ebd 1341 spinlock_t *ptl;
b8916634 1342 struct anon_vma *anon_vma = NULL;
b32967ff 1343 struct page *page;
d10e63f2 1344 unsigned long haddr = addr & HPAGE_PMD_MASK;
8191acbd 1345 int page_nid = -1, this_nid = numa_node_id();
90572890 1346 int target_nid, last_cpupid = -1;
8191acbd
MG		 1347 bool page_locked;
1348 bool migrated = false;
b191f9b1 1349 bool was_writable;
6688cc05 1350 int flags = 0;
d10e63f2 1351
c0e7cad9
MG		 1352 /* A PROT_NONE fault should not end up here */
1353 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1354
c4088ebd 1355 ptl = pmd_lock(mm, pmdp);
d10e63f2
MG		 1356 if (unlikely(!pmd_same(pmd, *pmdp)))
1357 goto out_unlock;
1358
de466bd6
MG		 1359 /*
1360 * If there are potential migrations, wait for completion and retry
1361 * without disrupting NUMA hinting information. Do not relock and
1362 * check_same as the page may no longer be mapped.
1363 */
1364 if (unlikely(pmd_trans_migrating(*pmdp))) {
5d833062 1365 page = pmd_page(*pmdp);
cdbf9267
MR		 1366 if (!get_page_unless_zero(page))
1367 goto out_unlock;
de466bd6 1368 spin_unlock(ptl);
5d833062 1369 wait_on_page_locked(page);
cdbf9267 1370 put_page(page);
de466bd6
MG		 1371 goto out;
1372 }
1373
d10e63f2 1374 page = pmd_page(pmd);
a1a46184 1375 BUG_ON(is_huge_zero_page(page));
8191acbd 1376 page_nid = page_to_nid(page);
90572890 1377 last_cpupid = page_cpupid_last(page);
03c5a6e1 1378 count_vm_numa_event(NUMA_HINT_FAULTS);
04bb2f94 1379 if (page_nid == this_nid) {
03c5a6e1 1380 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
04bb2f94
RR		 1381 flags |= TNF_FAULT_LOCAL;
1382 }
4daae3b4 1383
bea66fbd
MG		 1384 /* See similar comment in do_numa_page for explanation */
1385 if (!(vma->vm_flags & VM_WRITE))
6688cc05
PZ		 1386 flags |= TNF_NO_GROUP;
1387
ff9042b1
MG		 1388 /*
1389 * Acquire the page lock to serialise THP migrations but avoid dropping
1390 * page_table_lock if at all possible
1391 */
b8916634
MG		 1392 page_locked = trylock_page(page);
1393 target_nid = mpol_misplaced(page, vma, haddr);
1394 if (target_nid == -1) {
1395 /* If the page was locked, there are no parallel migrations */
a54a407f 1396 if (page_locked)
b8916634 1397 goto clear_pmdnuma;
2b4847e7 1398 }
4daae3b4 1399
de466bd6 1400 /* Migration could have started since the pmd_trans_migrating check */
2b4847e7 1401 if (!page_locked) {
cdbf9267
MR		 1402 if (!get_page_unless_zero(page))
1403 goto out_unlock;
c4088ebd 1404 spin_unlock(ptl);
b8916634 1405 wait_on_page_locked(page);
cdbf9267 1406 put_page(page);
a54a407f 1407 page_nid = -1;
b8916634
MG		 1408 goto out;
1409 }
1410
2b4847e7
MG		 1411 /*
1412 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1413 * to serialises splits
1414 */
b8916634 1415 get_page(page);
c4088ebd 1416 spin_unlock(ptl);
b8916634 1417 anon_vma = page_lock_anon_vma_read(page);
4daae3b4 1418
c69307d5 1419 /* Confirm the PMD did not change while page_table_lock was released */
c4088ebd 1420 spin_lock(ptl);
b32967ff
MG		 1421 if (unlikely(!pmd_same(pmd, *pmdp))) {
1422 unlock_page(page);
1423 put_page(page);
a54a407f 1424 page_nid = -1;
4daae3b4 1425 goto out_unlock;
b32967ff 1426 }
ff9042b1 1427
c3a489ca
MG		 1428 /* Bail if we fail to protect against THP splits for any reason */
1429 if (unlikely(!anon_vma)) {
1430 put_page(page);
1431 page_nid = -1;
1432 goto clear_pmdnuma;
1433 }
1434
a54a407f
MG		 1435 /*
1436 * Migrate the THP to the requested node, returns with page unlocked
8a0516ed 1437 * and access rights restored.
a54a407f 1438 */
c4088ebd 1439 spin_unlock(ptl);
b32967ff 1440 migrated = migrate_misplaced_transhuge_page(mm, vma,
340ef390 1441 pmdp, pmd, addr, page, target_nid);
6688cc05
PZ		 1442 if (migrated) {
1443 flags |= TNF_MIGRATED;
8191acbd 1444 page_nid = target_nid;
074c2381
MG		 1445 } else
1446 flags |= TNF_MIGRATE_FAIL;
b32967ff 1447
8191acbd 1448 goto out;
b32967ff 1449clear_pmdnuma:
a54a407f 1450 BUG_ON(!PageLocked(page));
b191f9b1 1451 was_writable = pmd_write(pmd);
4d942466 1452 pmd = pmd_modify(pmd, vma->vm_page_prot);
b7b04004 1453 pmd = pmd_mkyoung(pmd);
b191f9b1
MG		 1454 if (was_writable)
1455 pmd = pmd_mkwrite(pmd);
d10e63f2 1456 set_pmd_at(mm, haddr, pmdp, pmd);
d10e63f2 1457 update_mmu_cache_pmd(vma, addr, pmdp);
a54a407f 1458 unlock_page(page);
d10e63f2 1459out_unlock:
c4088ebd 1460 spin_unlock(ptl);
b8916634
MG		 1461
1462out:
1463 if (anon_vma)
1464 page_unlock_anon_vma_read(anon_vma);
1465
8191acbd 1466 if (page_nid != -1)
6688cc05 1467 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
8191acbd 1468
d10e63f2
MG		 1469 return 0;
1470}
1471
71e3aac0 1472int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
f21760b1 1473 pmd_t *pmd, unsigned long addr)
71e3aac0 1474{
da146769 1475 pmd_t orig_pmd;
bf929152 1476 spinlock_t *ptl;
71e3aac0 1477
da146769
KS		 1478 if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1479 return 0;
1480 /*
1481 * For architectures like ppc64 we look at deposited pgtable
1482 * when calling pmdp_huge_get_and_clear. So do the
1483 * pgtable_trans_huge_withdraw after finishing pmdp related
1484 * operations.
1485 */
1486 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1487 tlb->fullmm);
1488 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1489 if (vma_is_dax(vma)) {
1490 spin_unlock(ptl);
1491 if (is_huge_zero_pmd(orig_pmd))
97ae1749 1492 put_huge_zero_page();
da146769
KS		 1493 } else if (is_huge_zero_pmd(orig_pmd)) {
1494 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1495 atomic_long_dec(&tlb->mm->nr_ptes);
1496 spin_unlock(ptl);
1497 put_huge_zero_page();
1498 } else {
1499 struct page *page = pmd_page(orig_pmd);
1500 page_remove_rmap(page);
1501 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1502 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1503 VM_BUG_ON_PAGE(!PageHead(page), page);
1504 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1505 atomic_long_dec(&tlb->mm->nr_ptes);
1506 spin_unlock(ptl);
1507 tlb_remove_page(tlb, page);
025c5b24 1508 }
da146769 1509 return 1;
71e3aac0
AA		 1510}
1511
37a1c49a
AA		 1512int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1513 unsigned long old_addr,
1514 unsigned long new_addr, unsigned long old_end,
1515 pmd_t *old_pmd, pmd_t *new_pmd)
1516{
bf929152 1517 spinlock_t *old_ptl, *new_ptl;
37a1c49a
AA		 1518 int ret = 0;
1519 pmd_t pmd;
1520
1521 struct mm_struct *mm = vma->vm_mm;
1522
1523 if ((old_addr & ~HPAGE_PMD_MASK) ||
1524 (new_addr & ~HPAGE_PMD_MASK) ||
1525 old_end - old_addr < HPAGE_PMD_SIZE ||
1526 (new_vma->vm_flags & VM_NOHUGEPAGE))
1527 goto out;
1528
1529 /*
1530 * The destination pmd shouldn't be established, free_pgtables()
1531 * should have release it.
1532 */
1533 if (WARN_ON(!pmd_none(*new_pmd))) {
1534 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1535 goto out;
1536 }
1537
bf929152
KS		 1538 /*
1539 * We don't have to worry about the ordering of src and dst
1540 * ptlocks because exclusive mmap_sem prevents deadlock.
1541 */
1542 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
025c5b24 1543 if (ret == 1) {
bf929152
KS		 1544 new_ptl = pmd_lockptr(mm, new_pmd);
1545 if (new_ptl != old_ptl)
1546 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
8809aa2d 1547 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
025c5b24 1548 VM_BUG_ON(!pmd_none(*new_pmd));
3592806c 1549
b3084f4d
AK		 1550 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1551 pgtable_t pgtable;
3592806c
KS		 1552 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1553 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
3592806c 1554 }
b3084f4d
AK		 1555 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1556 if (new_ptl != old_ptl)
1557 spin_unlock(new_ptl);
bf929152 1558 spin_unlock(old_ptl);
37a1c49a
AA		 1559 }
1560out:
1561 return ret;
1562}
1563
f123d74a
MG		 1564/*
1565 * Returns
1566 * - 0 if PMD could not be locked
1567 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1568 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1569 */
cd7548ab 1570int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
e944fd67 1571 unsigned long addr, pgprot_t newprot, int prot_numa)
cd7548ab
JW		 1572{
1573 struct mm_struct *mm = vma->vm_mm;
bf929152 1574 spinlock_t *ptl;
cd7548ab
JW		 1575 int ret = 0;
1576
bf929152 1577 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24 1578 pmd_t entry;
b191f9b1 1579 bool preserve_write = prot_numa && pmd_write(*pmd);
ba68bc01 1580 ret = 1;
e944fd67
MG		 1581
1582 /*
1583 * Avoid trapping faults against the zero page. The read-only
1584 * data is likely to be read-cached on the local CPU and
1585 * local/remote hits to the zero page are not interesting.
1586 */
1587 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1588 spin_unlock(ptl);
ba68bc01 1589 return ret;
e944fd67
MG		 1590 }
1591
10c1045f 1592 if (!prot_numa || !pmd_protnone(*pmd)) {
8809aa2d 1593 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
10c1045f 1594 entry = pmd_modify(entry, newprot);
b191f9b1
MG		 1595 if (preserve_write)
1596 entry = pmd_mkwrite(entry);
10c1045f
MG		 1597 ret = HPAGE_PMD_NR;
1598 set_pmd_at(mm, addr, pmd, entry);
b191f9b1 1599 BUG_ON(!preserve_write && pmd_write(entry));
10c1045f 1600 }
bf929152 1601 spin_unlock(ptl);
025c5b24
NH		 1602 }
1603
1604 return ret;
1605}
1606
1607/*
1608 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1609 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1610 *
1611 * Note that if it returns 1, this routine returns without unlocking page
1612 * table locks. So callers must unlock them.
1613 */
bf929152
KS		 1614int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1615 spinlock_t **ptl)
025c5b24 1616{
bf929152 1617 *ptl = pmd_lock(vma->vm_mm, pmd);
cd7548ab
JW		 1618 if (likely(pmd_trans_huge(*pmd))) {
1619 if (unlikely(pmd_trans_splitting(*pmd))) {
bf929152 1620 spin_unlock(*ptl);
cd7548ab 1621 wait_split_huge_page(vma->anon_vma, pmd);
025c5b24 1622 return -1;
cd7548ab 1623 } else {
025c5b24
NH		 1624 /* Thp mapped by 'pmd' is stable, so we can
1625 * handle it as it is. */
1626 return 1;
cd7548ab 1627 }
025c5b24 1628 }
bf929152 1629 spin_unlock(*ptl);
025c5b24 1630 return 0;
cd7548ab
JW		 1631}
1632
117b0791
KS		 1633/*
1634 * This function returns whether a given @page is mapped onto the @address
1635 * in the virtual space of @mm.
1636 *
1637 * When it's true, this function returns *pmd with holding the page table lock
1638 * and passing it back to the caller via @ptl.
1639 * If it's false, returns NULL without holding the page table lock.
1640 */
71e3aac0
AA		 1641pmd_t *page_check_address_pmd(struct page *page,
1642 struct mm_struct *mm,
1643 unsigned long address,
117b0791
KS		 1644 enum page_check_address_pmd_flag flag,
1645 spinlock_t **ptl)
71e3aac0 1646{
b5a8cad3
KS		 1647 pgd_t *pgd;
1648 pud_t *pud;
117b0791 1649 pmd_t *pmd;
71e3aac0
AA		 1650
1651 if (address & ~HPAGE_PMD_MASK)
117b0791 1652 return NULL;
71e3aac0 1653
b5a8cad3
KS		 1654 pgd = pgd_offset(mm, address);
1655 if (!pgd_present(*pgd))
117b0791 1656 return NULL;
b5a8cad3
KS		 1657 pud = pud_offset(pgd, address);
1658 if (!pud_present(*pud))
1659 return NULL;
1660 pmd = pmd_offset(pud, address);
1661
117b0791 1662 *ptl = pmd_lock(mm, pmd);
b5a8cad3 1663 if (!pmd_present(*pmd))
117b0791 1664 goto unlock;
71e3aac0 1665 if (pmd_page(*pmd) != page)
117b0791 1666 goto unlock;
94fcc585
AA		 1667 /*
1668 * split_vma() may create temporary aliased mappings. There is
1669 * no risk as long as all huge pmd are found and have their
1670 * splitting bit set before __split_huge_page_refcount
1671 * runs. Finding the same huge pmd more than once during the
1672 * same rmap walk is not a problem.
1673 */
1674 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1675 pmd_trans_splitting(*pmd))
117b0791 1676 goto unlock;
71e3aac0
AA		 1677 if (pmd_trans_huge(*pmd)) {
1678 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1679 !pmd_trans_splitting(*pmd));
117b0791 1680 return pmd;
71e3aac0 1681 }
117b0791
KS		 1682unlock:
1683 spin_unlock(*ptl);
1684 return NULL;
71e3aac0
AA		 1685}
1686
1687static int __split_huge_page_splitting(struct page *page,
1688 struct vm_area_struct *vma,
1689 unsigned long address)
1690{
1691 struct mm_struct *mm = vma->vm_mm;
117b0791 1692 spinlock_t *ptl;
71e3aac0
AA		 1693 pmd_t *pmd;
1694 int ret = 0;
2ec74c3e
SG		 1695 /* For mmu_notifiers */
1696 const unsigned long mmun_start = address;
1697 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
71e3aac0 1698
2ec74c3e 1699 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0 1700 pmd = page_check_address_pmd(page, mm, address,
117b0791 1701 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
71e3aac0
AA		 1702 if (pmd) {
1703 /*
1704 * We can't temporarily set the pmd to null in order
1705 * to split it, the pmd must remain marked huge at all
1706 * times or the VM won't take the pmd_trans_huge paths
5a505085 1707 * and it won't wait on the anon_vma->root->rwsem to
71e3aac0
AA		 1708 * serialize against split_huge_page*.
1709 */
2ec74c3e 1710 pmdp_splitting_flush(vma, address, pmd);
34ee645e 1711
71e3aac0 1712 ret = 1;
117b0791 1713 spin_unlock(ptl);
71e3aac0 1714 }
2ec74c3e 1715 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1716
1717 return ret;
1718}
1719
5bc7b8ac
SL		 1720static void __split_huge_page_refcount(struct page *page,
1721 struct list_head *list)
71e3aac0
AA		 1722{
1723 int i;
71e3aac0 1724 struct zone *zone = page_zone(page);
fa9add64 1725 struct lruvec *lruvec;
70b50f94 1726 int tail_count = 0;
71e3aac0
AA		 1727
1728 /* prevent PageLRU to go away from under us, and freeze lru stats */
1729 spin_lock_irq(&zone->lru_lock);
fa9add64
HD		 1730 lruvec = mem_cgroup_page_lruvec(page, zone);
1731
71e3aac0 1732 compound_lock(page);
e94c8a9c
KH		 1733 /* complete memcg works before add pages to LRU */
1734 mem_cgroup_split_huge_fixup(page);
71e3aac0 1735
45676885 1736 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
71e3aac0
AA		 1737 struct page *page_tail = page + i;
1738
70b50f94
AA		 1739 /* tail_page->_mapcount cannot change */
1740 BUG_ON(page_mapcount(page_tail) < 0);
1741 tail_count += page_mapcount(page_tail);
1742 /* check for overflow */
1743 BUG_ON(tail_count < 0);
1744 BUG_ON(atomic_read(&page_tail->_count) != 0);
1745 /*
1746 * tail_page->_count is zero and not changing from
1747 * under us. But get_page_unless_zero() may be running
1748 * from under us on the tail_page. If we used
1749 * atomic_set() below instead of atomic_add(), we
1750 * would then run atomic_set() concurrently with
1751 * get_page_unless_zero(), and atomic_set() is
1752 * implemented in C not using locked ops. spin_unlock
1753 * on x86 sometime uses locked ops because of PPro
1754 * errata 66, 92, so unless somebody can guarantee
1755 * atomic_set() here would be safe on all archs (and
1756 * not only on x86), it's safer to use atomic_add().
1757 */
1758 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1759 &page_tail->_count);
71e3aac0
AA		 1760
1761 /* after clearing PageTail the gup refcount can be released */
3a79d52a 1762 smp_mb__after_atomic();
71e3aac0 1763
f4c18e6f 1764 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
71e3aac0
AA		 1765 page_tail->flags |= (page->flags &
1766 ((1L << PG_referenced) |
1767 (1L << PG_swapbacked) |
1768 (1L << PG_mlocked) |
e180cf80
KS		 1769 (1L << PG_uptodate) |
1770 (1L << PG_active) |
1771 (1L << PG_unevictable)));
71e3aac0
AA		 1772 page_tail->flags |= (1L << PG_dirty);
1773
1d798ca3 1774 clear_compound_head(page_tail);
71e3aac0 1775
33c3fc71
VD		 1776 if (page_is_young(page))
1777 set_page_young(page_tail);
1778 if (page_is_idle(page))
1779 set_page_idle(page_tail);
1780
71e3aac0
AA		 1781 /*
1782 * __split_huge_page_splitting() already set the
1783 * splitting bit in all pmd that could map this
1784 * hugepage, that will ensure no CPU can alter the
1785 * mapcount on the head page. The mapcount is only
1786 * accounted in the head page and it has to be
1787 * transferred to all tail pages in the below code. So
1788 * for this code to be safe, the split the mapcount
1789 * can't change. But that doesn't mean userland can't
1790 * keep changing and reading the page contents while
1791 * we transfer the mapcount, so the pmd splitting
1792 * status is achieved setting a reserved bit in the
1793 * pmd, not by clearing the present bit.
1794 */
71e3aac0
AA		 1795 page_tail->_mapcount = page->_mapcount;
1796
1797 BUG_ON(page_tail->mapping);
1798 page_tail->mapping = page->mapping;
1799
45676885 1800 page_tail->index = page->index + i;
90572890 1801 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
71e3aac0
AA		 1802
1803 BUG_ON(!PageAnon(page_tail));
1804 BUG_ON(!PageUptodate(page_tail));
1805 BUG_ON(!PageDirty(page_tail));
1806 BUG_ON(!PageSwapBacked(page_tail));
1807
5bc7b8ac 1808 lru_add_page_tail(page, page_tail, lruvec, list);
71e3aac0 1809 }
70b50f94
AA		 1810 atomic_sub(tail_count, &page->_count);
1811 BUG_ON(atomic_read(&page->_count) <= 0);
71e3aac0 1812
fa9add64 1813 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
79134171 1814
71e3aac0
AA		 1815 ClearPageCompound(page);
1816 compound_unlock(page);
1817 spin_unlock_irq(&zone->lru_lock);
1818
1819 for (i = 1; i < HPAGE_PMD_NR; i++) {
1820 struct page *page_tail = page + i;
1821 BUG_ON(page_count(page_tail) <= 0);
1822 /*
1823 * Tail pages may be freed if there wasn't any mapping
1824 * like if add_to_swap() is running on a lru page that
1825 * had its mapping zapped. And freeing these pages
1826 * requires taking the lru_lock so we do the put_page
1827 * of the tail pages after the split is complete.
1828 */
1829 put_page(page_tail);
1830 }
1831
1832 /*
1833 * Only the head page (now become a regular page) is required
1834 * to be pinned by the caller.
1835 */
1836 BUG_ON(page_count(page) <= 0);
1837}
1838
1839static int __split_huge_page_map(struct page *page,
1840 struct vm_area_struct *vma,
1841 unsigned long address)
1842{
1843 struct mm_struct *mm = vma->vm_mm;
117b0791 1844 spinlock_t *ptl;
71e3aac0
AA		 1845 pmd_t *pmd, _pmd;
1846 int ret = 0, i;
1847 pgtable_t pgtable;
1848 unsigned long haddr;
1849
71e3aac0 1850 pmd = page_check_address_pmd(page, mm, address,
117b0791 1851 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
71e3aac0 1852 if (pmd) {
6b0b50b0 1853 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0 1854 pmd_populate(mm, &_pmd, pgtable);
f8303c25
WL		 1855 if (pmd_write(*pmd))
1856 BUG_ON(page_mapcount(page) != 1);
71e3aac0 1857
e3ebcf64
GS		 1858 haddr = address;
1859 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
71e3aac0
AA		 1860 pte_t *pte, entry;
1861 BUG_ON(PageCompound(page+i));
abc40bd2 1862 /*
8a0516ed
MG		 1863 * Note that NUMA hinting access restrictions are not
1864 * transferred to avoid any possibility of altering
1865 * permissions across VMAs.
abc40bd2 1866 */
71e3aac0
AA		 1867 entry = mk_pte(page + i, vma->vm_page_prot);
1868 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1869 if (!pmd_write(*pmd))
1870 entry = pte_wrprotect(entry);
71e3aac0
AA		 1871 if (!pmd_young(*pmd))
1872 entry = pte_mkold(entry);
1873 pte = pte_offset_map(&_pmd, haddr);
1874 BUG_ON(!pte_none(*pte));
1875 set_pte_at(mm, haddr, pte, entry);
1876 pte_unmap(pte);
1877 }
1878
71e3aac0
AA		 1879 smp_wmb(); /* make pte visible before pmd */
1880 /*
1881 * Up to this point the pmd is present and huge and
1882 * userland has the whole access to the hugepage
1883 * during the split (which happens in place). If we
1884 * overwrite the pmd with the not-huge version
1885 * pointing to the pte here (which of course we could
1886 * if all CPUs were bug free), userland could trigger
1887 * a small page size TLB miss on the small sized TLB
1888 * while the hugepage TLB entry is still established
1889 * in the huge TLB. Some CPU doesn't like that. See
1890 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1891 * Erratum 383 on page 93. Intel should be safe but is
1892 * also warns that it's only safe if the permission
1893 * and cache attributes of the two entries loaded in
1894 * the two TLB is identical (which should be the case
1895 * here). But it is generally safer to never allow
1896 * small and huge TLB entries for the same virtual
1897 * address to be loaded simultaneously. So instead of
12ebc158 1898 * doing "pmd_populate(); flush_pmd_tlb_range();" we first
71e3aac0
AA		 1899 * mark the current pmd notpresent (atomically because
1900 * here the pmd_trans_huge and pmd_trans_splitting
1901 * must remain set at all times on the pmd until the
1902 * split is complete for this pmd), then we flush the
1903 * SMP TLB and finally we write the non-huge version
1904 * of the pmd entry with pmd_populate.
1905 */
46dcde73 1906 pmdp_invalidate(vma, address, pmd);
71e3aac0
AA		 1907 pmd_populate(mm, pmd, pgtable);
1908 ret = 1;
117b0791 1909 spin_unlock(ptl);
71e3aac0 1910 }
71e3aac0
AA		 1911
1912 return ret;
1913}
1914
5a505085 1915/* must be called with anon_vma->root->rwsem held */
71e3aac0 1916static void __split_huge_page(struct page *page,
5bc7b8ac
SL		 1917 struct anon_vma *anon_vma,
1918 struct list_head *list)
71e3aac0
AA		 1919{
1920 int mapcount, mapcount2;
bf181b9f 1921 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
71e3aac0
AA		 1922 struct anon_vma_chain *avc;
1923
1924 BUG_ON(!PageHead(page));
1925 BUG_ON(PageTail(page));
1926
1927 mapcount = 0;
bf181b9f 1928 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1929 struct vm_area_struct *vma = avc->vma;
1930 unsigned long addr = vma_address(page, vma);
1931 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1932 mapcount += __split_huge_page_splitting(page, vma, addr);
1933 }
05759d38
AA		 1934 /*
1935 * It is critical that new vmas are added to the tail of the
1936 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1937 * and establishes a child pmd before
1938 * __split_huge_page_splitting() freezes the parent pmd (so if
1939 * we fail to prevent copy_huge_pmd() from running until the
1940 * whole __split_huge_page() is complete), we will still see
1941 * the newly established pmd of the child later during the
1942 * walk, to be able to set it as pmd_trans_splitting too.
1943 */
ff9e43eb 1944 if (mapcount != page_mapcount(page)) {
ae3a8c1c
AM		 1945 pr_err("mapcount %d page_mapcount %d\n",
1946 mapcount, page_mapcount(page));
ff9e43eb
KS		 1947 BUG();
1948 }
71e3aac0 1949
5bc7b8ac 1950 __split_huge_page_refcount(page, list);
71e3aac0
AA		 1951
1952 mapcount2 = 0;
bf181b9f 1953 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1954 struct vm_area_struct *vma = avc->vma;
1955 unsigned long addr = vma_address(page, vma);
1956 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1957 mapcount2 += __split_huge_page_map(page, vma, addr);
1958 }
ff9e43eb 1959 if (mapcount != mapcount2) {
ae3a8c1c
AM		 1960 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1961 mapcount, mapcount2, page_mapcount(page));
ff9e43eb
KS		 1962 BUG();
1963 }
71e3aac0
AA		 1964}
1965
5bc7b8ac
SL		 1966/*
1967 * Split a hugepage into normal pages. This doesn't change the position of head
1968 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1969 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1970 * from the hugepage.
1971 * Return 0 if the hugepage is split successfully otherwise return 1.
1972 */
1973int split_huge_page_to_list(struct page *page, struct list_head *list)
71e3aac0
AA		 1974{
1975 struct anon_vma *anon_vma;
1976 int ret = 1;
1977
5918d10a 1978 BUG_ON(is_huge_zero_page(page));
71e3aac0 1979 BUG_ON(!PageAnon(page));
062f1af2
MG		 1980
1981 /*
1982 * The caller does not necessarily hold an mmap_sem that would prevent
1983 * the anon_vma disappearing so we first we take a reference to it
1984 * and then lock the anon_vma for write. This is similar to
1985 * page_lock_anon_vma_read except the write lock is taken to serialise
1986 * against parallel split or collapse operations.
1987 */
1988 anon_vma = page_get_anon_vma(page);
71e3aac0
AA		 1989 if (!anon_vma)
1990 goto out;
062f1af2
MG		 1991 anon_vma_lock_write(anon_vma);
1992
71e3aac0
AA		 1993 ret = 0;
1994 if (!PageCompound(page))
1995 goto out_unlock;
1996
1997 BUG_ON(!PageSwapBacked(page));
5bc7b8ac 1998 __split_huge_page(page, anon_vma, list);
81ab4201 1999 count_vm_event(THP_SPLIT);
71e3aac0
AA		 2000
2001 BUG_ON(PageCompound(page));
2002out_unlock:
08b52706 2003 anon_vma_unlock_write(anon_vma);
062f1af2 2004 put_anon_vma(anon_vma);
71e3aac0
AA		 2005out:
2006 return ret;
2007}
2008
9050d7eb 2009#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
78f11a25 2010
60ab3244
AA		 2011int hugepage_madvise(struct vm_area_struct *vma,
2012 unsigned long *vm_flags, int advice)
0af4e98b 2013{
a664b2d8
AA		 2014 switch (advice) {
2015 case MADV_HUGEPAGE:
1e1836e8
AT		 2016#ifdef CONFIG_S390
2017 /*
2018 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2019 * can't handle this properly after s390_enable_sie, so we simply
2020 * ignore the madvise to prevent qemu from causing a SIGSEGV.
2021 */
2022 if (mm_has_pgste(vma->vm_mm))
2023 return 0;
2024#endif
a664b2d8
AA		 2025 /*
2026 * Be somewhat over-protective like KSM for now!
2027 */
1a763615 2028 if (*vm_flags & VM_NO_THP)
a664b2d8
AA		 2029 return -EINVAL;
2030 *vm_flags &= ~VM_NOHUGEPAGE;
2031 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 2032 /*
2033 * If the vma become good for khugepaged to scan,
2034 * register it here without waiting a page fault that
2035 * may not happen any time soon.
2036 */
6d50e60c 2037 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
60ab3244 2038 return -ENOMEM;
a664b2d8
AA		 2039 break;
2040 case MADV_NOHUGEPAGE:
2041 /*
2042 * Be somewhat over-protective like KSM for now!
2043 */
1a763615 2044 if (*vm_flags & VM_NO_THP)
a664b2d8
AA		 2045 return -EINVAL;
2046 *vm_flags &= ~VM_HUGEPAGE;
2047 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 2048 /*
2049 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2050 * this vma even if we leave the mm registered in khugepaged if
2051 * it got registered before VM_NOHUGEPAGE was set.
2052 */
a664b2d8
AA		 2053 break;
2054 }
0af4e98b
AA		 2055
2056 return 0;
2057}
2058
ba76149f
AA		 2059static int __init khugepaged_slab_init(void)
2060{
2061 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2062 sizeof(struct mm_slot),
2063 __alignof__(struct mm_slot), 0, NULL);
2064 if (!mm_slot_cache)
2065 return -ENOMEM;
2066
2067 return 0;
2068}
2069
65ebb64f
KS		 2070static void __init khugepaged_slab_exit(void)
2071{
2072 kmem_cache_destroy(mm_slot_cache);
2073}
2074
ba76149f
AA		 2075static inline struct mm_slot *alloc_mm_slot(void)
2076{
2077 if (!mm_slot_cache) /* initialization failed */
2078 return NULL;
2079 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2080}
2081
2082static inline void free_mm_slot(struct mm_slot *mm_slot)
2083{
2084 kmem_cache_free(mm_slot_cache, mm_slot);
2085}
2086
ba76149f
AA		 2087static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2088{
2089 struct mm_slot *mm_slot;
ba76149f 2090
b67bfe0d 2091 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
ba76149f
AA		 2092 if (mm == mm_slot->mm)
2093 return mm_slot;
43b5fbbd 2094
ba76149f
AA		 2095 return NULL;
2096}
2097
2098static void insert_to_mm_slots_hash(struct mm_struct *mm,
2099 struct mm_slot *mm_slot)
2100{
ba76149f 2101 mm_slot->mm = mm;
43b5fbbd 2102 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
ba76149f
AA		 2103}
2104
2105static inline int khugepaged_test_exit(struct mm_struct *mm)
2106{
2107 return atomic_read(&mm->mm_users) == 0;
2108}
2109
2110int __khugepaged_enter(struct mm_struct *mm)
2111{
2112 struct mm_slot *mm_slot;
2113 int wakeup;
2114
2115 mm_slot = alloc_mm_slot();
2116 if (!mm_slot)
2117 return -ENOMEM;
2118
2119 /* __khugepaged_exit() must not run from under us */
96dad67f 2120 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
ba76149f
AA		 2121 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2122 free_mm_slot(mm_slot);
2123 return 0;
2124 }
2125
2126 spin_lock(&khugepaged_mm_lock);
2127 insert_to_mm_slots_hash(mm, mm_slot);
2128 /*
2129 * Insert just behind the scanning cursor, to let the area settle
2130 * down a little.
2131 */
2132 wakeup = list_empty(&khugepaged_scan.mm_head);
2133 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2134 spin_unlock(&khugepaged_mm_lock);
2135
2136 atomic_inc(&mm->mm_count);
2137 if (wakeup)
2138 wake_up_interruptible(&khugepaged_wait);
2139
2140 return 0;
2141}
2142
6d50e60c
DR		 2143int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2144 unsigned long vm_flags)
ba76149f
AA		 2145{
2146 unsigned long hstart, hend;
2147 if (!vma->anon_vma)
2148 /*
2149 * Not yet faulted in so we will register later in the
2150 * page fault if needed.
2151 */
2152 return 0;
be591a68 2153 if (vma->vm_ops || (vm_flags & VM_NO_THP))
ba76149f
AA		 2154 /* khugepaged not yet working on file or special mappings */
2155 return 0;
ba76149f
AA		 2156 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2157 hend = vma->vm_end & HPAGE_PMD_MASK;
2158 if (hstart < hend)
6d50e60c 2159 return khugepaged_enter(vma, vm_flags);
ba76149f
AA		 2160 return 0;
2161}
2162
2163void __khugepaged_exit(struct mm_struct *mm)
2164{
2165 struct mm_slot *mm_slot;
2166 int free = 0;
2167
2168 spin_lock(&khugepaged_mm_lock);
2169 mm_slot = get_mm_slot(mm);
2170 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
43b5fbbd 2171 hash_del(&mm_slot->hash);
ba76149f
AA		 2172 list_del(&mm_slot->mm_node);
2173 free = 1;
2174 }
d788e80a 2175 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2176
2177 if (free) {
ba76149f
AA		 2178 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2179 free_mm_slot(mm_slot);
2180 mmdrop(mm);
2181 } else if (mm_slot) {
ba76149f
AA		 2182 /*
2183 * This is required to serialize against
2184 * khugepaged_test_exit() (which is guaranteed to run
2185 * under mmap sem read mode). Stop here (after we
2186 * return all pagetables will be destroyed) until
2187 * khugepaged has finished working on the pagetables
2188 * under the mmap_sem.
2189 */
2190 down_write(&mm->mmap_sem);
2191 up_write(&mm->mmap_sem);
d788e80a 2192 }
ba76149f
AA		 2193}
2194
2195static void release_pte_page(struct page *page)
2196{
2197 /* 0 stands for page_is_file_cache(page) == false */
2198 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2199 unlock_page(page);
2200 putback_lru_page(page);
2201}
2202
2203static void release_pte_pages(pte_t *pte, pte_t *_pte)
2204{
2205 while (--_pte >= pte) {
2206 pte_t pteval = *_pte;
ca0984ca 2207 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
ba76149f
AA		 2208 release_pte_page(pte_page(pteval));
2209 }
2210}
2211
ba76149f
AA		 2212static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2213 unsigned long address,
2214 pte_t *pte)
2215{
2216 struct page *page;
2217 pte_t *_pte;
ca0984ca 2218 int none_or_zero = 0;
10359213 2219 bool referenced = false, writable = false;
ba76149f
AA		 2220 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2221 _pte++, address += PAGE_SIZE) {
2222 pte_t pteval = *_pte;
47aee4d8
MK		 2223 if (pte_none(pteval) || (pte_present(pteval) &&
2224 is_zero_pfn(pte_pfn(pteval)))) {
c1294d05
AA		 2225 if (!userfaultfd_armed(vma) &&
2226 ++none_or_zero <= khugepaged_max_ptes_none)
ba76149f 2227 continue;
344aa35c 2228 else
ba76149f 2229 goto out;
ba76149f 2230 }
10359213 2231 if (!pte_present(pteval))
ba76149f 2232 goto out;
ba76149f 2233 page = vm_normal_page(vma, address, pteval);
344aa35c 2234 if (unlikely(!page))
ba76149f 2235 goto out;
344aa35c 2236
309381fe
SL		 2237 VM_BUG_ON_PAGE(PageCompound(page), page);
2238 VM_BUG_ON_PAGE(!PageAnon(page), page);
2239 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
ba76149f 2240
ba76149f
AA		 2241 /*
2242 * We can do it before isolate_lru_page because the
2243 * page can't be freed from under us. NOTE: PG_lock
2244 * is needed to serialize against split_huge_page
2245 * when invoked from the VM.
2246 */
344aa35c 2247 if (!trylock_page(page))
ba76149f 2248 goto out;
10359213
EA		 2249
2250 /*
2251 * cannot use mapcount: can't collapse if there's a gup pin.
2252 * The page must only be referenced by the scanned process
2253 * and page swap cache.
2254 */
2255 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2256 unlock_page(page);
2257 goto out;
2258 }
2259 if (pte_write(pteval)) {
2260 writable = true;
2261 } else {
2262 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2263 unlock_page(page);
2264 goto out;
2265 }
2266 /*
2267 * Page is not in the swap cache. It can be collapsed
2268 * into a THP.
2269 */
2270 }
2271
ba76149f
AA		 2272 /*
2273 * Isolate the page to avoid collapsing an hugepage
2274 * currently in use by the VM.
2275 */
2276 if (isolate_lru_page(page)) {
2277 unlock_page(page);
ba76149f
AA		 2278 goto out;
2279 }
2280 /* 0 stands for page_is_file_cache(page) == false */
2281 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
309381fe
SL		 2282 VM_BUG_ON_PAGE(!PageLocked(page), page);
2283 VM_BUG_ON_PAGE(PageLRU(page), page);
ba76149f
AA		 2284
2285 /* If there is no mapped pte young don't collapse the page */
33c3fc71
VD		 2286 if (pte_young(pteval) ||
2287 page_is_young(page) || PageReferenced(page) ||
8ee53820 2288 mmu_notifier_test_young(vma->vm_mm, address))
10359213 2289 referenced = true;
ba76149f 2290 }
10359213 2291 if (likely(referenced && writable))
344aa35c 2292 return 1;
ba76149f 2293out:
344aa35c
BL		 2294 release_pte_pages(pte, _pte);
2295 return 0;
ba76149f
AA		 2296}
2297
2298static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2299 struct vm_area_struct *vma,
2300 unsigned long address,
2301 spinlock_t *ptl)
2302{
2303 pte_t *_pte;
2304 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2305 pte_t pteval = *_pte;
2306 struct page *src_page;
2307
ca0984ca 2308 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
ba76149f
AA		 2309 clear_user_highpage(page, address);
2310 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
ca0984ca
EA		 2311 if (is_zero_pfn(pte_pfn(pteval))) {
2312 /*
2313 * ptl mostly unnecessary.
2314 */
2315 spin_lock(ptl);
2316 /*
2317 * paravirt calls inside pte_clear here are
2318 * superfluous.
2319 */
2320 pte_clear(vma->vm_mm, address, _pte);
2321 spin_unlock(ptl);
2322 }
ba76149f
AA		 2323 } else {
2324 src_page = pte_page(pteval);
2325 copy_user_highpage(page, src_page, address, vma);
309381fe 2326 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
ba76149f
AA		 2327 release_pte_page(src_page);
2328 /*
2329 * ptl mostly unnecessary, but preempt has to
2330 * be disabled to update the per-cpu stats
2331 * inside page_remove_rmap().
2332 */
2333 spin_lock(ptl);
2334 /*
2335 * paravirt calls inside pte_clear here are
2336 * superfluous.
2337 */
2338 pte_clear(vma->vm_mm, address, _pte);
2339 page_remove_rmap(src_page);
2340 spin_unlock(ptl);
2341 free_page_and_swap_cache(src_page);
2342 }
2343
2344 address += PAGE_SIZE;
2345 page++;
2346 }
2347}
2348
26234f36 2349static void khugepaged_alloc_sleep(void)
ba76149f 2350{
bde43c6c
PM		 2351 DEFINE_WAIT(wait);
2352
2353 add_wait_queue(&khugepaged_wait, &wait);
2354 freezable_schedule_timeout_interruptible(
2355 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2356 remove_wait_queue(&khugepaged_wait, &wait);
26234f36 2357}
ba76149f 2358
9f1b868a
BL		 2359static int khugepaged_node_load[MAX_NUMNODES];
2360
14a4e214
DR		 2361static bool khugepaged_scan_abort(int nid)
2362{
2363 int i;
2364
2365 /*
2366 * If zone_reclaim_mode is disabled, then no extra effort is made to
2367 * allocate memory locally.
2368 */
2369 if (!zone_reclaim_mode)
2370 return false;
2371
2372 /* If there is a count for this node already, it must be acceptable */
2373 if (khugepaged_node_load[nid])
2374 return false;
2375
2376 for (i = 0; i < MAX_NUMNODES; i++) {
2377 if (!khugepaged_node_load[i])
2378 continue;
2379 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2380 return true;
2381 }
2382 return false;
2383}
2384
26234f36 2385#ifdef CONFIG_NUMA
9f1b868a
BL		 2386static int khugepaged_find_target_node(void)
2387{
2388 static int last_khugepaged_target_node = NUMA_NO_NODE;
2389 int nid, target_node = 0, max_value = 0;
2390
2391 /* find first node with max normal pages hit */
2392 for (nid = 0; nid < MAX_NUMNODES; nid++)
2393 if (khugepaged_node_load[nid] > max_value) {
2394 max_value = khugepaged_node_load[nid];
2395 target_node = nid;
2396 }
2397
2398 /* do some balance if several nodes have the same hit record */
2399 if (target_node <= last_khugepaged_target_node)
2400 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2401 nid++)
2402 if (max_value == khugepaged_node_load[nid]) {
2403 target_node = nid;
2404 break;
2405 }
2406
2407 last_khugepaged_target_node = target_node;
2408 return target_node;
2409}
2410
26234f36
XG		 2411static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2412{
2413 if (IS_ERR(*hpage)) {
2414 if (!*wait)
2415 return false;
2416
2417 *wait = false;
e3b4126c 2418 *hpage = NULL;
26234f36
XG		 2419 khugepaged_alloc_sleep();
2420 } else if (*hpage) {
2421 put_page(*hpage);
2422 *hpage = NULL;
2423 }
2424
2425 return true;
2426}
2427
3b363692
MH		 2428static struct page *
2429khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
d6669d68 2430 unsigned long address, int node)
26234f36 2431{
309381fe 2432 VM_BUG_ON_PAGE(*hpage, *hpage);
8b164568 2433
ce83d217 2434 /*
8b164568
VB		 2435 * Before allocating the hugepage, release the mmap_sem read lock.
2436 * The allocation can take potentially a long time if it involves
2437 * sync compaction, and we do not need to hold the mmap_sem during
2438 * that. We will recheck the vma after taking it again in write mode.
ce83d217 2439 */
8b164568
VB		 2440 up_read(&mm->mmap_sem);
2441
96db800f 2442 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
26234f36 2443 if (unlikely(!*hpage)) {
81ab4201 2444 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
ce83d217 2445 *hpage = ERR_PTR(-ENOMEM);
26234f36 2446 return NULL;
ce83d217 2447 }
26234f36 2448
65b3c07b 2449 count_vm_event(THP_COLLAPSE_ALLOC);
26234f36
XG		 2450 return *hpage;
2451}
2452#else
9f1b868a
BL		 2453static int khugepaged_find_target_node(void)
2454{
2455 return 0;
2456}
2457
10dc4155
BL		 2458static inline struct page *alloc_hugepage(int defrag)
2459{
2460 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2461 HPAGE_PMD_ORDER);
2462}
2463
26234f36
XG		 2464static struct page *khugepaged_alloc_hugepage(bool *wait)
2465{
2466 struct page *hpage;
2467
2468 do {
2469 hpage = alloc_hugepage(khugepaged_defrag());
2470 if (!hpage) {
2471 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2472 if (!*wait)
2473 return NULL;
2474
2475 *wait = false;
2476 khugepaged_alloc_sleep();
2477 } else
2478 count_vm_event(THP_COLLAPSE_ALLOC);
2479 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2480
2481 return hpage;
2482}
2483
2484static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2485{
2486 if (!*hpage)
2487 *hpage = khugepaged_alloc_hugepage(wait);
2488
2489 if (unlikely(!*hpage))
2490 return false;
2491
2492 return true;
2493}
2494
3b363692
MH		 2495static struct page *
2496khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
d6669d68 2497 unsigned long address, int node)
26234f36
XG		 2498{
2499 up_read(&mm->mmap_sem);
2500 VM_BUG_ON(!*hpage);
3b363692 2501
26234f36
XG		 2502 return *hpage;
2503}
692e0b35
AA		 2504#endif
2505
fa475e51
BL		 2506static bool hugepage_vma_check(struct vm_area_struct *vma)
2507{
2508 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2509 (vma->vm_flags & VM_NOHUGEPAGE))
2510 return false;
2511
2512 if (!vma->anon_vma || vma->vm_ops)
2513 return false;
2514 if (is_vma_temporary_stack(vma))
2515 return false;
be591a68 2516 return !(vma->vm_flags & VM_NO_THP);
fa475e51
BL		 2517}
2518
26234f36
XG		 2519static void collapse_huge_page(struct mm_struct *mm,
2520 unsigned long address,
2521 struct page **hpage,
2522 struct vm_area_struct *vma,
2523 int node)
2524{
26234f36
XG		 2525 pmd_t *pmd, _pmd;
2526 pte_t *pte;
2527 pgtable_t pgtable;
2528 struct page *new_page;
c4088ebd 2529 spinlock_t *pmd_ptl, *pte_ptl;
26234f36
XG		 2530 int isolated;
2531 unsigned long hstart, hend;
00501b53 2532 struct mem_cgroup *memcg;
2ec74c3e
SG		 2533 unsigned long mmun_start; /* For mmu_notifiers */
2534 unsigned long mmun_end; /* For mmu_notifiers */
3b363692 2535 gfp_t gfp;
26234f36
XG		 2536
2537 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2538
3b363692
MH		 2539 /* Only allocate from the target node */
2540 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2541 __GFP_THISNODE;
2542
26234f36 2543 /* release the mmap_sem read lock. */
d6669d68 2544 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
26234f36
XG		 2545 if (!new_page)
2546 return;
2547
00501b53 2548 if (unlikely(mem_cgroup_try_charge(new_page, mm,
3b363692 2549 gfp, &memcg)))
ce83d217 2550 return;
ba76149f
AA		 2551
2552 /*
2553 * Prevent all access to pagetables with the exception of
2554 * gup_fast later hanlded by the ptep_clear_flush and the VM
2555 * handled by the anon_vma lock + PG_lock.
2556 */
2557 down_write(&mm->mmap_sem);
2558 if (unlikely(khugepaged_test_exit(mm)))
2559 goto out;
2560
2561 vma = find_vma(mm, address);
a8f531eb
L		 2562 if (!vma)
2563 goto out;
ba76149f
AA		 2564 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2565 hend = vma->vm_end & HPAGE_PMD_MASK;
2566 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2567 goto out;
fa475e51 2568 if (!hugepage_vma_check(vma))
a7d6e4ec 2569 goto out;
6219049a
BL		 2570 pmd = mm_find_pmd(mm, address);
2571 if (!pmd)
ba76149f 2572 goto out;
ba76149f 2573
4fc3f1d6 2574 anon_vma_lock_write(vma->anon_vma);
ba76149f
AA		 2575
2576 pte = pte_offset_map(pmd, address);
c4088ebd 2577 pte_ptl = pte_lockptr(mm, pmd);
ba76149f 2578
2ec74c3e
SG		 2579 mmun_start = address;
2580 mmun_end = address + HPAGE_PMD_SIZE;
2581 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2582 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
ba76149f
AA		 2583 /*
2584 * After this gup_fast can't run anymore. This also removes
2585 * any huge TLB entry from the CPU so we won't allow
2586 * huge and small TLB entries for the same virtual address
2587 * to avoid the risk of CPU bugs in that area.
2588 */
15a25b2e 2589 _pmd = pmdp_collapse_flush(vma, address, pmd);
c4088ebd 2590 spin_unlock(pmd_ptl);
2ec74c3e 2591 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ba76149f 2592
c4088ebd 2593 spin_lock(pte_ptl);
ba76149f 2594 isolated = __collapse_huge_page_isolate(vma, address, pte);
c4088ebd 2595 spin_unlock(pte_ptl);
ba76149f
AA		 2596
2597 if (unlikely(!isolated)) {
453c7192 2598 pte_unmap(pte);
c4088ebd 2599 spin_lock(pmd_ptl);
ba76149f 2600 BUG_ON(!pmd_none(*pmd));
7c342512
AK		 2601 /*
2602 * We can only use set_pmd_at when establishing
2603 * hugepmds and never for establishing regular pmds that
2604 * points to regular pagetables. Use pmd_populate for that
2605 */
2606 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
c4088ebd 2607 spin_unlock(pmd_ptl);
08b52706 2608 anon_vma_unlock_write(vma->anon_vma);
ce83d217 2609 goto out;
ba76149f
AA		 2610 }
2611
2612 /*
2613 * All pages are isolated and locked so anon_vma rmap
2614 * can't run anymore.
2615 */
08b52706 2616 anon_vma_unlock_write(vma->anon_vma);
ba76149f 2617
c4088ebd 2618 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
453c7192 2619 pte_unmap(pte);
ba76149f
AA		 2620 __SetPageUptodate(new_page);
2621 pgtable = pmd_pgtable(_pmd);
ba76149f 2622
3122359a
KS		 2623 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2624 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
ba76149f
AA		 2625
2626 /*
2627 * spin_lock() below is not the equivalent of smp_wmb(), so
2628 * this is needed to avoid the copy_huge_page writes to become
2629 * visible after the set_pmd_at() write.
2630 */
2631 smp_wmb();
2632
c4088ebd 2633 spin_lock(pmd_ptl);
ba76149f
AA		 2634 BUG_ON(!pmd_none(*pmd));
2635 page_add_new_anon_rmap(new_page, vma, address);
00501b53
JW		 2636 mem_cgroup_commit_charge(new_page, memcg, false);
2637 lru_cache_add_active_or_unevictable(new_page, vma);
fce144b4 2638 pgtable_trans_huge_deposit(mm, pmd, pgtable);
ba76149f 2639 set_pmd_at(mm, address, pmd, _pmd);
b113da65 2640 update_mmu_cache_pmd(vma, address, pmd);
c4088ebd 2641 spin_unlock(pmd_ptl);
ba76149f
AA		 2642
2643 *hpage = NULL;
420256ef 2644
ba76149f 2645 khugepaged_pages_collapsed++;
ce83d217 2646out_up_write:
ba76149f 2647 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 2648 return;
2649
ce83d217 2650out:
00501b53 2651 mem_cgroup_cancel_charge(new_page, memcg);
ce83d217 2652 goto out_up_write;
ba76149f
AA		 2653}
2654
2655static int khugepaged_scan_pmd(struct mm_struct *mm,
2656 struct vm_area_struct *vma,
2657 unsigned long address,
2658 struct page **hpage)
2659{
ba76149f
AA		 2660 pmd_t *pmd;
2661 pte_t *pte, *_pte;
ca0984ca 2662 int ret = 0, none_or_zero = 0;
ba76149f
AA		 2663 struct page *page;
2664 unsigned long _address;
2665 spinlock_t *ptl;
00ef2d2f 2666 int node = NUMA_NO_NODE;
10359213 2667 bool writable = false, referenced = false;
ba76149f
AA		 2668
2669 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2670
6219049a
BL		 2671 pmd = mm_find_pmd(mm, address);
2672 if (!pmd)
ba76149f 2673 goto out;
ba76149f 2674
9f1b868a 2675 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
ba76149f
AA		 2676 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2677 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2678 _pte++, _address += PAGE_SIZE) {
2679 pte_t pteval = *_pte;
ca0984ca 2680 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
c1294d05
AA		 2681 if (!userfaultfd_armed(vma) &&
2682 ++none_or_zero <= khugepaged_max_ptes_none)
ba76149f
AA		 2683 continue;
2684 else
2685 goto out_unmap;
2686 }
10359213 2687 if (!pte_present(pteval))
ba76149f 2688 goto out_unmap;
10359213
EA		 2689 if (pte_write(pteval))
2690 writable = true;
2691
ba76149f
AA		 2692 page = vm_normal_page(vma, _address, pteval);
2693 if (unlikely(!page))
2694 goto out_unmap;
5c4b4be3 2695 /*
9f1b868a
BL		 2696 * Record which node the original page is from and save this
2697 * information to khugepaged_node_load[].
2698 * Khupaged will allocate hugepage from the node has the max
2699 * hit record.
5c4b4be3 2700 */
9f1b868a 2701 node = page_to_nid(page);
14a4e214
DR		 2702 if (khugepaged_scan_abort(node))
2703 goto out_unmap;
9f1b868a 2704 khugepaged_node_load[node]++;
309381fe 2705 VM_BUG_ON_PAGE(PageCompound(page), page);
ba76149f
AA		 2706 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2707 goto out_unmap;
10359213
EA		 2708 /*
2709 * cannot use mapcount: can't collapse if there's a gup pin.
2710 * The page must only be referenced by the scanned process
2711 * and page swap cache.
2712 */
2713 if (page_count(page) != 1 + !!PageSwapCache(page))
ba76149f 2714 goto out_unmap;
33c3fc71
VD		 2715 if (pte_young(pteval) ||
2716 page_is_young(page) || PageReferenced(page) ||
8ee53820 2717 mmu_notifier_test_young(vma->vm_mm, address))
10359213 2718 referenced = true;
ba76149f 2719 }
10359213 2720 if (referenced && writable)
ba76149f
AA		 2721 ret = 1;
2722out_unmap:
2723 pte_unmap_unlock(pte, ptl);
9f1b868a
BL		 2724 if (ret) {
2725 node = khugepaged_find_target_node();
ce83d217 2726 /* collapse_huge_page will return with the mmap_sem released */
5c4b4be3 2727 collapse_huge_page(mm, address, hpage, vma, node);
9f1b868a 2728 }
ba76149f
AA		 2729out:
2730 return ret;
2731}
2732
2733static void collect_mm_slot(struct mm_slot *mm_slot)
2734{
2735 struct mm_struct *mm = mm_slot->mm;
2736
b9980cdc 2737 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2738
2739 if (khugepaged_test_exit(mm)) {
2740 /* free mm_slot */
43b5fbbd 2741 hash_del(&mm_slot->hash);
ba76149f
AA		 2742 list_del(&mm_slot->mm_node);
2743
2744 /*
2745 * Not strictly needed because the mm exited already.
2746 *
2747 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2748 */
2749
2750 /* khugepaged_mm_lock actually not necessary for the below */
2751 free_mm_slot(mm_slot);
2752 mmdrop(mm);
2753 }
2754}
2755
2756static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2757 struct page **hpage)
2f1da642
HS		 2758 __releases(&khugepaged_mm_lock)
2759 __acquires(&khugepaged_mm_lock)
ba76149f
AA		 2760{
2761 struct mm_slot *mm_slot;
2762 struct mm_struct *mm;
2763 struct vm_area_struct *vma;
2764 int progress = 0;
2765
2766 VM_BUG_ON(!pages);
b9980cdc 2767 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2768
2769 if (khugepaged_scan.mm_slot)
2770 mm_slot = khugepaged_scan.mm_slot;
2771 else {
2772 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2773 struct mm_slot, mm_node);
2774 khugepaged_scan.address = 0;
2775 khugepaged_scan.mm_slot = mm_slot;
2776 }
2777 spin_unlock(&khugepaged_mm_lock);
2778
2779 mm = mm_slot->mm;
2780 down_read(&mm->mmap_sem);
2781 if (unlikely(khugepaged_test_exit(mm)))
2782 vma = NULL;
2783 else
2784 vma = find_vma(mm, khugepaged_scan.address);
2785
2786 progress++;
2787 for (; vma; vma = vma->vm_next) {
2788 unsigned long hstart, hend;
2789
2790 cond_resched();
2791 if (unlikely(khugepaged_test_exit(mm))) {
2792 progress++;
2793 break;
2794 }
fa475e51
BL		 2795 if (!hugepage_vma_check(vma)) {
2796skip:
ba76149f
AA		 2797 progress++;
2798 continue;
2799 }
ba76149f
AA		 2800 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2801 hend = vma->vm_end & HPAGE_PMD_MASK;
a7d6e4ec
AA		 2802 if (hstart >= hend)
2803 goto skip;
2804 if (khugepaged_scan.address > hend)
2805 goto skip;
ba76149f
AA		 2806 if (khugepaged_scan.address < hstart)
2807 khugepaged_scan.address = hstart;
a7d6e4ec 2808 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
ba76149f
AA		 2809
2810 while (khugepaged_scan.address < hend) {
2811 int ret;
2812 cond_resched();
2813 if (unlikely(khugepaged_test_exit(mm)))
2814 goto breakouterloop;
2815
2816 VM_BUG_ON(khugepaged_scan.address < hstart ||
2817 khugepaged_scan.address + HPAGE_PMD_SIZE >
2818 hend);
2819 ret = khugepaged_scan_pmd(mm, vma,
2820 khugepaged_scan.address,
2821 hpage);
2822 /* move to next address */
2823 khugepaged_scan.address += HPAGE_PMD_SIZE;
2824 progress += HPAGE_PMD_NR;
2825 if (ret)
2826 /* we released mmap_sem so break loop */
2827 goto breakouterloop_mmap_sem;
2828 if (progress >= pages)
2829 goto breakouterloop;
2830 }
2831 }
2832breakouterloop:
2833 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2834breakouterloop_mmap_sem:
2835
2836 spin_lock(&khugepaged_mm_lock);
a7d6e4ec 2837 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
ba76149f
AA		 2838 /*
2839 * Release the current mm_slot if this mm is about to die, or
2840 * if we scanned all vmas of this mm.
2841 */
2842 if (khugepaged_test_exit(mm) || !vma) {
2843 /*
2844 * Make sure that if mm_users is reaching zero while
2845 * khugepaged runs here, khugepaged_exit will find
2846 * mm_slot not pointing to the exiting mm.
2847 */
2848 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2849 khugepaged_scan.mm_slot = list_entry(
2850 mm_slot->mm_node.next,
2851 struct mm_slot, mm_node);
2852 khugepaged_scan.address = 0;
2853 } else {
2854 khugepaged_scan.mm_slot = NULL;
2855 khugepaged_full_scans++;
2856 }
2857
2858 collect_mm_slot(mm_slot);
2859 }
2860
2861 return progress;
2862}
2863
2864static int khugepaged_has_work(void)
2865{
2866 return !list_empty(&khugepaged_scan.mm_head) &&
2867 khugepaged_enabled();
2868}
2869
2870static int khugepaged_wait_event(void)
2871{
2872 return !list_empty(&khugepaged_scan.mm_head) ||
2017c0bf 2873 kthread_should_stop();
ba76149f
AA		 2874}
2875
d516904b 2876static void khugepaged_do_scan(void)
ba76149f 2877{
d516904b 2878 struct page *hpage = NULL;
ba76149f
AA		 2879 unsigned int progress = 0, pass_through_head = 0;
2880 unsigned int pages = khugepaged_pages_to_scan;
d516904b 2881 bool wait = true;
ba76149f
AA		 2882
2883 barrier(); /* write khugepaged_pages_to_scan to local stack */
2884
2885 while (progress < pages) {
26234f36 2886 if (!khugepaged_prealloc_page(&hpage, &wait))
d516904b 2887 break;
26234f36 2888
420256ef 2889 cond_resched();
ba76149f 2890
cd092411 2891 if (unlikely(kthread_should_stop() || try_to_freeze()))
878aee7d
AA		 2892 break;
2893
ba76149f
AA		 2894 spin_lock(&khugepaged_mm_lock);
2895 if (!khugepaged_scan.mm_slot)
2896 pass_through_head++;
2897 if (khugepaged_has_work() &&
2898 pass_through_head < 2)
2899 progress += khugepaged_scan_mm_slot(pages - progress,
d516904b 2900 &hpage);
ba76149f
AA		 2901 else
2902 progress = pages;
2903 spin_unlock(&khugepaged_mm_lock);
2904 }
ba76149f 2905
d516904b
XG		 2906 if (!IS_ERR_OR_NULL(hpage))
2907 put_page(hpage);
0bbbc0b3
AA		 2908}
2909
2017c0bf
XG		 2910static void khugepaged_wait_work(void)
2911{
2017c0bf
XG		 2912 if (khugepaged_has_work()) {
2913 if (!khugepaged_scan_sleep_millisecs)
2914 return;
2915
2916 wait_event_freezable_timeout(khugepaged_wait,
2917 kthread_should_stop(),
2918 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2919 return;
2920 }
2921
2922 if (khugepaged_enabled())
2923 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2924}
2925
ba76149f
AA		 2926static int khugepaged(void *none)
2927{
2928 struct mm_slot *mm_slot;
2929
878aee7d 2930 set_freezable();
8698a745 2931 set_user_nice(current, MAX_NICE);
ba76149f 2932
b7231789
XG		 2933 while (!kthread_should_stop()) {
2934 khugepaged_do_scan();
2935 khugepaged_wait_work();
2936 }
ba76149f
AA		 2937
2938 spin_lock(&khugepaged_mm_lock);
2939 mm_slot = khugepaged_scan.mm_slot;
2940 khugepaged_scan.mm_slot = NULL;
2941 if (mm_slot)
2942 collect_mm_slot(mm_slot);
2943 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2944 return 0;
2945}
2946
c5a647d0
KS		 2947static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2948 unsigned long haddr, pmd_t *pmd)
2949{
2950 struct mm_struct *mm = vma->vm_mm;
2951 pgtable_t pgtable;
2952 pmd_t _pmd;
2953 int i;
2954
8809aa2d 2955 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
c5a647d0
KS		 2956 /* leave pmd empty until pte is filled */
2957
6b0b50b0 2958 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
c5a647d0
KS		 2959 pmd_populate(mm, &_pmd, pgtable);
2960
2961 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2962 pte_t *pte, entry;
2963 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2964 entry = pte_mkspecial(entry);
2965 pte = pte_offset_map(&_pmd, haddr);
2966 VM_BUG_ON(!pte_none(*pte));
2967 set_pte_at(mm, haddr, pte, entry);
2968 pte_unmap(pte);
2969 }
2970 smp_wmb(); /* make pte visible before pmd */
2971 pmd_populate(mm, pmd, pgtable);
97ae1749 2972 put_huge_zero_page();
c5a647d0
KS		 2973}
2974
e180377f
KS		 2975void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2976 pmd_t *pmd)
71e3aac0 2977{
c4088ebd 2978 spinlock_t *ptl;
4897c765 2979 struct page *page = NULL;
e180377f 2980 struct mm_struct *mm = vma->vm_mm;
c5a647d0
KS		 2981 unsigned long haddr = address & HPAGE_PMD_MASK;
2982 unsigned long mmun_start; /* For mmu_notifiers */
2983 unsigned long mmun_end; /* For mmu_notifiers */
e180377f
KS		 2984
2985 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
71e3aac0 2986
c5a647d0
KS		 2987 mmun_start = haddr;
2988 mmun_end = haddr + HPAGE_PMD_SIZE;
750e8165 2989again:
c5a647d0 2990 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2991 ptl = pmd_lock(mm, pmd);
4897c765
MW		 2992 if (unlikely(!pmd_trans_huge(*pmd)))
2993 goto unlock;
2994 if (vma_is_dax(vma)) {
5b701b84
KS		 2995 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2996 if (is_huge_zero_pmd(_pmd))
2997 put_huge_zero_page();
4897c765 2998 } else if (is_huge_zero_pmd(*pmd)) {
c5a647d0 2999 __split_huge_zero_page_pmd(vma, haddr, pmd);
4897c765
MW		 3000 } else {
3001 page = pmd_page(*pmd);
3002 VM_BUG_ON_PAGE(!page_count(page), page);
3003 get_page(page);
71e3aac0 3004 }
4897c765 3005 unlock:
c4088ebd 3006 spin_unlock(ptl);
c5a647d0 3007 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0 3008
4897c765
MW		 3009 if (!page)
3010 return;
71e3aac0 3011
4897c765 3012 split_huge_page(page);
71e3aac0 3013 put_page(page);
750e8165
HD		 3014
3015 /*
3016 * We don't always have down_write of mmap_sem here: a racing
3017 * do_huge_pmd_wp_page() might have copied-on-write to another
3018 * huge page before our split_huge_page() got the anon_vma lock.
3019 */
3020 if (unlikely(pmd_trans_huge(*pmd)))
3021 goto again;
71e3aac0 3022}
94fcc585 3023
e180377f
KS		 3024void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3025 pmd_t *pmd)
3026{
3027 struct vm_area_struct *vma;
3028
3029 vma = find_vma(mm, address);
3030 BUG_ON(vma == NULL);
3031 split_huge_page_pmd(vma, address, pmd);
3032}
3033
94fcc585
AA		 3034static void split_huge_page_address(struct mm_struct *mm,
3035 unsigned long address)
3036{
f72e7dcd
HD		 3037 pgd_t *pgd;
3038 pud_t *pud;
94fcc585
AA		 3039 pmd_t *pmd;
3040
3041 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3042
f72e7dcd
HD		 3043 pgd = pgd_offset(mm, address);
3044 if (!pgd_present(*pgd))
3045 return;
3046
3047 pud = pud_offset(pgd, address);
3048 if (!pud_present(*pud))
3049 return;
3050
3051 pmd = pmd_offset(pud, address);
3052 if (!pmd_present(*pmd))
94fcc585
AA		 3053 return;
3054 /*
3055 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3056 * materialize from under us.
3057 */
e180377f 3058 split_huge_page_pmd_mm(mm, address, pmd);
94fcc585
AA		 3059}
3060
e1b9996b 3061void vma_adjust_trans_huge(struct vm_area_struct *vma,
94fcc585
AA		 3062 unsigned long start,
3063 unsigned long end,
3064 long adjust_next)
3065{
3066 /*
3067 * If the new start address isn't hpage aligned and it could
3068 * previously contain an hugepage: check if we need to split
3069 * an huge pmd.
3070 */
3071 if (start & ~HPAGE_PMD_MASK &&
3072 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3073 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3074 split_huge_page_address(vma->vm_mm, start);
3075
3076 /*
3077 * If the new end address isn't hpage aligned and it could
3078 * previously contain an hugepage: check if we need to split
3079 * an huge pmd.
3080 */
3081 if (end & ~HPAGE_PMD_MASK &&
3082 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3083 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3084 split_huge_page_address(vma->vm_mm, end);
3085
3086 /*
3087 * If we're also updating the vma->vm_next->vm_start, if the new
3088 * vm_next->vm_start isn't page aligned and it could previously
3089 * contain an hugepage: check if we need to split an huge pmd.
3090 */
3091 if (adjust_next > 0) {
3092 struct vm_area_struct *next = vma->vm_next;
3093 unsigned long nstart = next->vm_start;
3094 nstart += adjust_next << PAGE_SHIFT;
3095 if (nstart & ~HPAGE_PMD_MASK &&
3096 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3097 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3098 split_huge_page_address(next->vm_mm, nstart);
3099 }
3100}