projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
kconfig: rename CONFIG_EMBEDDED to CONFIG_EXPERT
[linux-2.6-block.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
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/highmem.h>
11#include <linux/hugetlb.h>
12#include <linux/mmu_notifier.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
ba76149f
AA		 15#include <linux/mm_inline.h>
16#include <linux/kthread.h>
17#include <linux/khugepaged.h>
878aee7d 18#include <linux/freezer.h>
a664b2d8 19#include <linux/mman.h>
71e3aac0
AA		 20#include <asm/tlb.h>
21#include <asm/pgalloc.h>
22#include "internal.h"
23
ba76149f
AA		 24/*
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29 * allocations.
30 */
71e3aac0 31unsigned long transparent_hugepage_flags __read_mostly =
13ece886 32#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
ba76149f 33 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
13ece886
AA		 34#endif
35#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37#endif
d39d33c3 38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
ba76149f
AA		 39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41/* default scan 8*512 pte (or vmas) every 30 second */
42static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43static unsigned int khugepaged_pages_collapsed;
44static unsigned int khugepaged_full_scans;
45static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46/* during fragmentation poll the hugepage allocator once every minute */
47static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48static struct task_struct *khugepaged_thread __read_mostly;
49static DEFINE_MUTEX(khugepaged_mutex);
50static DEFINE_SPINLOCK(khugepaged_mm_lock);
51static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52/*
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
55 * fault.
56 */
57static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59static int khugepaged(void *none);
60static int mm_slots_hash_init(void);
61static int khugepaged_slab_init(void);
62static void khugepaged_slab_free(void);
63
64#define MM_SLOTS_HASH_HEADS 1024
65static struct hlist_head *mm_slots_hash __read_mostly;
66static struct kmem_cache *mm_slot_cache __read_mostly;
67
68/**
69 * struct mm_slot - hash lookup from mm to mm_slot
70 * @hash: hash collision list
71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72 * @mm: the mm that this information is valid for
73 */
74struct mm_slot {
75 struct hlist_node hash;
76 struct list_head mm_node;
77 struct mm_struct *mm;
78};
79
80/**
81 * struct khugepaged_scan - cursor for scanning
82 * @mm_head: the head of the mm list to scan
83 * @mm_slot: the current mm_slot we are scanning
84 * @address: the next address inside that to be scanned
85 *
86 * There is only the one khugepaged_scan instance of this cursor structure.
87 */
88struct khugepaged_scan {
89 struct list_head mm_head;
90 struct mm_slot *mm_slot;
91 unsigned long address;
92} khugepaged_scan = {
93 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
94};
95
f000565a
AA		 96
97static int set_recommended_min_free_kbytes(void)
98{
99 struct zone *zone;
100 int nr_zones = 0;
101 unsigned long recommended_min;
102 extern int min_free_kbytes;
103
104 if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
105 &transparent_hugepage_flags) &&
106 !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
107 &transparent_hugepage_flags))
108 return 0;
109
110 for_each_populated_zone(zone)
111 nr_zones++;
112
113 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
114 recommended_min = pageblock_nr_pages * nr_zones * 2;
115
116 /*
117 * Make sure that on average at least two pageblocks are almost free
118 * of another type, one for a migratetype to fall back to and a
119 * second to avoid subsequent fallbacks of other types There are 3
120 * MIGRATE_TYPES we care about.
121 */
122 recommended_min += pageblock_nr_pages * nr_zones *
123 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
124
125 /* don't ever allow to reserve more than 5% of the lowmem */
126 recommended_min = min(recommended_min,
127 (unsigned long) nr_free_buffer_pages() / 20);
128 recommended_min <<= (PAGE_SHIFT-10);
129
130 if (recommended_min > min_free_kbytes)
131 min_free_kbytes = recommended_min;
132 setup_per_zone_wmarks();
133 return 0;
134}
135late_initcall(set_recommended_min_free_kbytes);
136
ba76149f
AA		 137static int start_khugepaged(void)
138{
139 int err = 0;
140 if (khugepaged_enabled()) {
141 int wakeup;
142 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
143 err = -ENOMEM;
144 goto out;
145 }
146 mutex_lock(&khugepaged_mutex);
147 if (!khugepaged_thread)
148 khugepaged_thread = kthread_run(khugepaged, NULL,
149 "khugepaged");
150 if (unlikely(IS_ERR(khugepaged_thread))) {
151 printk(KERN_ERR
152 "khugepaged: kthread_run(khugepaged) failed\n");
153 err = PTR_ERR(khugepaged_thread);
154 khugepaged_thread = NULL;
155 }
156 wakeup = !list_empty(&khugepaged_scan.mm_head);
157 mutex_unlock(&khugepaged_mutex);
158 if (wakeup)
159 wake_up_interruptible(&khugepaged_wait);
f000565a
AA		 160
161 set_recommended_min_free_kbytes();
ba76149f
AA		 162 } else
163 /* wakeup to exit */
164 wake_up_interruptible(&khugepaged_wait);
165out:
166 return err;
167}
71e3aac0
AA		 168
169#ifdef CONFIG_SYSFS
ba76149f 170
71e3aac0
AA		 171static ssize_t double_flag_show(struct kobject *kobj,
172 struct kobj_attribute *attr, char *buf,
173 enum transparent_hugepage_flag enabled,
174 enum transparent_hugepage_flag req_madv)
175{
176 if (test_bit(enabled, &transparent_hugepage_flags)) {
177 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
178 return sprintf(buf, "[always] madvise never\n");
179 } else if (test_bit(req_madv, &transparent_hugepage_flags))
180 return sprintf(buf, "always [madvise] never\n");
181 else
182 return sprintf(buf, "always madvise [never]\n");
183}
184static ssize_t double_flag_store(struct kobject *kobj,
185 struct kobj_attribute *attr,
186 const char *buf, size_t count,
187 enum transparent_hugepage_flag enabled,
188 enum transparent_hugepage_flag req_madv)
189{
190 if (!memcmp("always", buf,
191 min(sizeof("always")-1, count))) {
192 set_bit(enabled, &transparent_hugepage_flags);
193 clear_bit(req_madv, &transparent_hugepage_flags);
194 } else if (!memcmp("madvise", buf,
195 min(sizeof("madvise")-1, count))) {
196 clear_bit(enabled, &transparent_hugepage_flags);
197 set_bit(req_madv, &transparent_hugepage_flags);
198 } else if (!memcmp("never", buf,
199 min(sizeof("never")-1, count))) {
200 clear_bit(enabled, &transparent_hugepage_flags);
201 clear_bit(req_madv, &transparent_hugepage_flags);
202 } else
203 return -EINVAL;
204
205 return count;
206}
207
208static ssize_t enabled_show(struct kobject *kobj,
209 struct kobj_attribute *attr, char *buf)
210{
211 return double_flag_show(kobj, attr, buf,
212 TRANSPARENT_HUGEPAGE_FLAG,
213 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
214}
215static ssize_t enabled_store(struct kobject *kobj,
216 struct kobj_attribute *attr,
217 const char *buf, size_t count)
218{
ba76149f
AA		 219 ssize_t ret;
220
221 ret = double_flag_store(kobj, attr, buf, count,
222 TRANSPARENT_HUGEPAGE_FLAG,
223 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
224
225 if (ret > 0) {
226 int err = start_khugepaged();
227 if (err)
228 ret = err;
229 }
230
f000565a
AA		 231 if (ret > 0 &&
232 (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
233 &transparent_hugepage_flags) ||
234 test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
235 &transparent_hugepage_flags)))
236 set_recommended_min_free_kbytes();
237
ba76149f 238 return ret;
71e3aac0
AA		 239}
240static struct kobj_attribute enabled_attr =
241 __ATTR(enabled, 0644, enabled_show, enabled_store);
242
243static ssize_t single_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag flag)
246{
247 if (test_bit(flag, &transparent_hugepage_flags))
248 return sprintf(buf, "[yes] no\n");
249 else
250 return sprintf(buf, "yes [no]\n");
251}
252static ssize_t single_flag_store(struct kobject *kobj,
253 struct kobj_attribute *attr,
254 const char *buf, size_t count,
255 enum transparent_hugepage_flag flag)
256{
257 if (!memcmp("yes", buf,
258 min(sizeof("yes")-1, count))) {
259 set_bit(flag, &transparent_hugepage_flags);
260 } else if (!memcmp("no", buf,
261 min(sizeof("no")-1, count))) {
262 clear_bit(flag, &transparent_hugepage_flags);
263 } else
264 return -EINVAL;
265
266 return count;
267}
268
269/*
270 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
271 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
272 * memory just to allocate one more hugepage.
273 */
274static ssize_t defrag_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
276{
277 return double_flag_show(kobj, attr, buf,
278 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
279 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
280}
281static ssize_t defrag_store(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 const char *buf, size_t count)
284{
285 return double_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
287 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
288}
289static struct kobj_attribute defrag_attr =
290 __ATTR(defrag, 0644, defrag_show, defrag_store);
291
292#ifdef CONFIG_DEBUG_VM
293static ssize_t debug_cow_show(struct kobject *kobj,
294 struct kobj_attribute *attr, char *buf)
295{
296 return single_flag_show(kobj, attr, buf,
297 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
298}
299static ssize_t debug_cow_store(struct kobject *kobj,
300 struct kobj_attribute *attr,
301 const char *buf, size_t count)
302{
303 return single_flag_store(kobj, attr, buf, count,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305}
306static struct kobj_attribute debug_cow_attr =
307 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
308#endif /* CONFIG_DEBUG_VM */
309
310static struct attribute *hugepage_attr[] = {
311 &enabled_attr.attr,
312 &defrag_attr.attr,
313#ifdef CONFIG_DEBUG_VM
314 &debug_cow_attr.attr,
315#endif
316 NULL,
317};
318
319static struct attribute_group hugepage_attr_group = {
320 .attrs = hugepage_attr,
ba76149f
AA		 321};
322
323static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
324 struct kobj_attribute *attr,
325 char *buf)
326{
327 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
328}
329
330static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
331 struct kobj_attribute *attr,
332 const char *buf, size_t count)
333{
334 unsigned long msecs;
335 int err;
336
337 err = strict_strtoul(buf, 10, &msecs);
338 if (err || msecs > UINT_MAX)
339 return -EINVAL;
340
341 khugepaged_scan_sleep_millisecs = msecs;
342 wake_up_interruptible(&khugepaged_wait);
343
344 return count;
345}
346static struct kobj_attribute scan_sleep_millisecs_attr =
347 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
348 scan_sleep_millisecs_store);
349
350static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
351 struct kobj_attribute *attr,
352 char *buf)
353{
354 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
355}
356
357static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
358 struct kobj_attribute *attr,
359 const char *buf, size_t count)
360{
361 unsigned long msecs;
362 int err;
363
364 err = strict_strtoul(buf, 10, &msecs);
365 if (err || msecs > UINT_MAX)
366 return -EINVAL;
367
368 khugepaged_alloc_sleep_millisecs = msecs;
369 wake_up_interruptible(&khugepaged_wait);
370
371 return count;
372}
373static struct kobj_attribute alloc_sleep_millisecs_attr =
374 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
375 alloc_sleep_millisecs_store);
376
377static ssize_t pages_to_scan_show(struct kobject *kobj,
378 struct kobj_attribute *attr,
379 char *buf)
380{
381 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
382}
383static ssize_t pages_to_scan_store(struct kobject *kobj,
384 struct kobj_attribute *attr,
385 const char *buf, size_t count)
386{
387 int err;
388 unsigned long pages;
389
390 err = strict_strtoul(buf, 10, &pages);
391 if (err || !pages || pages > UINT_MAX)
392 return -EINVAL;
393
394 khugepaged_pages_to_scan = pages;
395
396 return count;
397}
398static struct kobj_attribute pages_to_scan_attr =
399 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
400 pages_to_scan_store);
401
402static ssize_t pages_collapsed_show(struct kobject *kobj,
403 struct kobj_attribute *attr,
404 char *buf)
405{
406 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
407}
408static struct kobj_attribute pages_collapsed_attr =
409 __ATTR_RO(pages_collapsed);
410
411static ssize_t full_scans_show(struct kobject *kobj,
412 struct kobj_attribute *attr,
413 char *buf)
414{
415 return sprintf(buf, "%u\n", khugepaged_full_scans);
416}
417static struct kobj_attribute full_scans_attr =
418 __ATTR_RO(full_scans);
419
420static ssize_t khugepaged_defrag_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
422{
423 return single_flag_show(kobj, attr, buf,
424 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
425}
426static ssize_t khugepaged_defrag_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
429{
430 return single_flag_store(kobj, attr, buf, count,
431 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
432}
433static struct kobj_attribute khugepaged_defrag_attr =
434 __ATTR(defrag, 0644, khugepaged_defrag_show,
435 khugepaged_defrag_store);
436
437/*
438 * max_ptes_none controls if khugepaged should collapse hugepages over
439 * any unmapped ptes in turn potentially increasing the memory
440 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
441 * reduce the available free memory in the system as it
442 * runs. Increasing max_ptes_none will instead potentially reduce the
443 * free memory in the system during the khugepaged scan.
444 */
445static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 char *buf)
448{
449 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
450}
451static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
452 struct kobj_attribute *attr,
453 const char *buf, size_t count)
454{
455 int err;
456 unsigned long max_ptes_none;
457
458 err = strict_strtoul(buf, 10, &max_ptes_none);
459 if (err || max_ptes_none > HPAGE_PMD_NR-1)
460 return -EINVAL;
461
462 khugepaged_max_ptes_none = max_ptes_none;
463
464 return count;
465}
466static struct kobj_attribute khugepaged_max_ptes_none_attr =
467 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
468 khugepaged_max_ptes_none_store);
469
470static struct attribute *khugepaged_attr[] = {
471 &khugepaged_defrag_attr.attr,
472 &khugepaged_max_ptes_none_attr.attr,
473 &pages_to_scan_attr.attr,
474 &pages_collapsed_attr.attr,
475 &full_scans_attr.attr,
476 &scan_sleep_millisecs_attr.attr,
477 &alloc_sleep_millisecs_attr.attr,
478 NULL,
479};
480
481static struct attribute_group khugepaged_attr_group = {
482 .attrs = khugepaged_attr,
483 .name = "khugepaged",
71e3aac0
AA		 484};
485#endif /* CONFIG_SYSFS */
486
487static int __init hugepage_init(void)
488{
71e3aac0 489 int err;
ba76149f
AA		 490#ifdef CONFIG_SYSFS
491 static struct kobject *hugepage_kobj;
4b7167b9 492#endif
71e3aac0 493
4b7167b9
AA		 494 err = -EINVAL;
495 if (!has_transparent_hugepage()) {
496 transparent_hugepage_flags = 0;
497 goto out;
498 }
499
500#ifdef CONFIG_SYSFS
ba76149f
AA		 501 err = -ENOMEM;
502 hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
503 if (unlikely(!hugepage_kobj)) {
504 printk(KERN_ERR "hugepage: failed kobject create\n");
505 goto out;
506 }
507
508 err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
509 if (err) {
510 printk(KERN_ERR "hugepage: failed register hugeage group\n");
511 goto out;
512 }
513
514 err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
515 if (err) {
516 printk(KERN_ERR "hugepage: failed register hugeage group\n");
517 goto out;
518 }
71e3aac0 519#endif
ba76149f
AA		 520
521 err = khugepaged_slab_init();
522 if (err)
523 goto out;
524
525 err = mm_slots_hash_init();
526 if (err) {
527 khugepaged_slab_free();
528 goto out;
529 }
530
97562cd2
RR		 531 /*
532 * By default disable transparent hugepages on smaller systems,
533 * where the extra memory used could hurt more than TLB overhead
534 * is likely to save. The admin can still enable it through /sys.
535 */
536 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
537 transparent_hugepage_flags = 0;
538
ba76149f
AA		 539 start_khugepaged();
540
f000565a
AA		 541 set_recommended_min_free_kbytes();
542
ba76149f
AA		 543out:
544 return err;
71e3aac0
AA		 545}
546module_init(hugepage_init)
547
548static int __init setup_transparent_hugepage(char *str)
549{
550 int ret = 0;
551 if (!str)
552 goto out;
553 if (!strcmp(str, "always")) {
554 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
555 &transparent_hugepage_flags);
556 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
557 &transparent_hugepage_flags);
558 ret = 1;
559 } else if (!strcmp(str, "madvise")) {
560 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
561 &transparent_hugepage_flags);
562 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
563 &transparent_hugepage_flags);
564 ret = 1;
565 } else if (!strcmp(str, "never")) {
566 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
567 &transparent_hugepage_flags);
568 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
569 &transparent_hugepage_flags);
570 ret = 1;
571 }
572out:
573 if (!ret)
574 printk(KERN_WARNING
575 "transparent_hugepage= cannot parse, ignored\n");
576 return ret;
577}
578__setup("transparent_hugepage=", setup_transparent_hugepage);
579
580static void prepare_pmd_huge_pte(pgtable_t pgtable,
581 struct mm_struct *mm)
582{
583 assert_spin_locked(&mm->page_table_lock);
584
585 /* FIFO */
586 if (!mm->pmd_huge_pte)
587 INIT_LIST_HEAD(&pgtable->lru);
588 else
589 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
590 mm->pmd_huge_pte = pgtable;
591}
592
593static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
594{
595 if (likely(vma->vm_flags & VM_WRITE))
596 pmd = pmd_mkwrite(pmd);
597 return pmd;
598}
599
600static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
601 struct vm_area_struct *vma,
602 unsigned long haddr, pmd_t *pmd,
603 struct page *page)
604{
605 int ret = 0;
606 pgtable_t pgtable;
607
608 VM_BUG_ON(!PageCompound(page));
609 pgtable = pte_alloc_one(mm, haddr);
610 if (unlikely(!pgtable)) {
b9bbfbe3 611 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 612 put_page(page);
613 return VM_FAULT_OOM;
614 }
615
616 clear_huge_page(page, haddr, HPAGE_PMD_NR);
617 __SetPageUptodate(page);
618
619 spin_lock(&mm->page_table_lock);
620 if (unlikely(!pmd_none(*pmd))) {
621 spin_unlock(&mm->page_table_lock);
b9bbfbe3 622 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 623 put_page(page);
624 pte_free(mm, pgtable);
625 } else {
626 pmd_t entry;
627 entry = mk_pmd(page, vma->vm_page_prot);
628 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
629 entry = pmd_mkhuge(entry);
630 /*
631 * The spinlocking to take the lru_lock inside
632 * page_add_new_anon_rmap() acts as a full memory
633 * barrier to be sure clear_huge_page writes become
634 * visible after the set_pmd_at() write.
635 */
636 page_add_new_anon_rmap(page, vma, haddr);
637 set_pmd_at(mm, haddr, pmd, entry);
638 prepare_pmd_huge_pte(pgtable, mm);
639 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
640 spin_unlock(&mm->page_table_lock);
641 }
642
643 return ret;
644}
645
0bbbc0b3
AA		 646static inline gfp_t alloc_hugepage_gfpmask(int defrag)
647{
648 return GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT);
649}
650
651static inline struct page *alloc_hugepage_vma(int defrag,
652 struct vm_area_struct *vma,
653 unsigned long haddr)
654{
655 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag),
656 HPAGE_PMD_ORDER, vma, haddr);
657}
658
659#ifndef CONFIG_NUMA
71e3aac0
AA		 660static inline struct page *alloc_hugepage(int defrag)
661{
0bbbc0b3 662 return alloc_pages(alloc_hugepage_gfpmask(defrag),
71e3aac0
AA		 663 HPAGE_PMD_ORDER);
664}
0bbbc0b3 665#endif
71e3aac0
AA		 666
667int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
668 unsigned long address, pmd_t *pmd,
669 unsigned int flags)
670{
671 struct page *page;
672 unsigned long haddr = address & HPAGE_PMD_MASK;
673 pte_t *pte;
674
675 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
676 if (unlikely(anon_vma_prepare(vma)))
677 return VM_FAULT_OOM;
ba76149f
AA		 678 if (unlikely(khugepaged_enter(vma)))
679 return VM_FAULT_OOM;
0bbbc0b3
AA		 680 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
681 vma, haddr);
71e3aac0
AA		 682 if (unlikely(!page))
683 goto out;
b9bbfbe3
AA		 684 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
685 put_page(page);
686 goto out;
687 }
71e3aac0
AA		 688
689 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
690 }
691out:
692 /*
693 * Use __pte_alloc instead of pte_alloc_map, because we can't
694 * run pte_offset_map on the pmd, if an huge pmd could
695 * materialize from under us from a different thread.
696 */
697 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
698 return VM_FAULT_OOM;
699 /* if an huge pmd materialized from under us just retry later */
700 if (unlikely(pmd_trans_huge(*pmd)))
701 return 0;
702 /*
703 * A regular pmd is established and it can't morph into a huge pmd
704 * from under us anymore at this point because we hold the mmap_sem
705 * read mode and khugepaged takes it in write mode. So now it's
706 * safe to run pte_offset_map().
707 */
708 pte = pte_offset_map(pmd, address);
709 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
710}
711
712int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
713 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
714 struct vm_area_struct *vma)
715{
716 struct page *src_page;
717 pmd_t pmd;
718 pgtable_t pgtable;
719 int ret;
720
721 ret = -ENOMEM;
722 pgtable = pte_alloc_one(dst_mm, addr);
723 if (unlikely(!pgtable))
724 goto out;
725
726 spin_lock(&dst_mm->page_table_lock);
727 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
728
729 ret = -EAGAIN;
730 pmd = *src_pmd;
731 if (unlikely(!pmd_trans_huge(pmd))) {
732 pte_free(dst_mm, pgtable);
733 goto out_unlock;
734 }
735 if (unlikely(pmd_trans_splitting(pmd))) {
736 /* split huge page running from under us */
737 spin_unlock(&src_mm->page_table_lock);
738 spin_unlock(&dst_mm->page_table_lock);
739 pte_free(dst_mm, pgtable);
740
741 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
742 goto out;
743 }
744 src_page = pmd_page(pmd);
745 VM_BUG_ON(!PageHead(src_page));
746 get_page(src_page);
747 page_dup_rmap(src_page);
748 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
749
750 pmdp_set_wrprotect(src_mm, addr, src_pmd);
751 pmd = pmd_mkold(pmd_wrprotect(pmd));
752 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
753 prepare_pmd_huge_pte(pgtable, dst_mm);
754
755 ret = 0;
756out_unlock:
757 spin_unlock(&src_mm->page_table_lock);
758 spin_unlock(&dst_mm->page_table_lock);
759out:
760 return ret;
761}
762
763/* no "address" argument so destroys page coloring of some arch */
764pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
765{
766 pgtable_t pgtable;
767
768 assert_spin_locked(&mm->page_table_lock);
769
770 /* FIFO */
771 pgtable = mm->pmd_huge_pte;
772 if (list_empty(&pgtable->lru))
773 mm->pmd_huge_pte = NULL;
774 else {
775 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
776 struct page, lru);
777 list_del(&pgtable->lru);
778 }
779 return pgtable;
780}
781
782static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
783 struct vm_area_struct *vma,
784 unsigned long address,
785 pmd_t *pmd, pmd_t orig_pmd,
786 struct page *page,
787 unsigned long haddr)
788{
789 pgtable_t pgtable;
790 pmd_t _pmd;
791 int ret = 0, i;
792 struct page **pages;
793
794 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
795 GFP_KERNEL);
796 if (unlikely(!pages)) {
797 ret |= VM_FAULT_OOM;
798 goto out;
799 }
800
801 for (i = 0; i < HPAGE_PMD_NR; i++) {
802 pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
803 vma, address);
b9bbfbe3
AA		 804 if (unlikely(!pages[i] ||
805 mem_cgroup_newpage_charge(pages[i], mm,
806 GFP_KERNEL))) {
807 if (pages[i])
71e3aac0 808 put_page(pages[i]);
b9bbfbe3
AA		 809 mem_cgroup_uncharge_start();
810 while (--i >= 0) {
811 mem_cgroup_uncharge_page(pages[i]);
812 put_page(pages[i]);
813 }
814 mem_cgroup_uncharge_end();
71e3aac0
AA		 815 kfree(pages);
816 ret |= VM_FAULT_OOM;
817 goto out;
818 }
819 }
820
821 for (i = 0; i < HPAGE_PMD_NR; i++) {
822 copy_user_highpage(pages[i], page + i,
823 haddr + PAGE_SHIFT*i, vma);
824 __SetPageUptodate(pages[i]);
825 cond_resched();
826 }
827
828 spin_lock(&mm->page_table_lock);
829 if (unlikely(!pmd_same(*pmd, orig_pmd)))
830 goto out_free_pages;
831 VM_BUG_ON(!PageHead(page));
832
833 pmdp_clear_flush_notify(vma, haddr, pmd);
834 /* leave pmd empty until pte is filled */
835
836 pgtable = get_pmd_huge_pte(mm);
837 pmd_populate(mm, &_pmd, pgtable);
838
839 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
840 pte_t *pte, entry;
841 entry = mk_pte(pages[i], vma->vm_page_prot);
842 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
843 page_add_new_anon_rmap(pages[i], vma, haddr);
844 pte = pte_offset_map(&_pmd, haddr);
845 VM_BUG_ON(!pte_none(*pte));
846 set_pte_at(mm, haddr, pte, entry);
847 pte_unmap(pte);
848 }
849 kfree(pages);
850
851 mm->nr_ptes++;
852 smp_wmb(); /* make pte visible before pmd */
853 pmd_populate(mm, pmd, pgtable);
854 page_remove_rmap(page);
855 spin_unlock(&mm->page_table_lock);
856
857 ret |= VM_FAULT_WRITE;
858 put_page(page);
859
860out:
861 return ret;
862
863out_free_pages:
864 spin_unlock(&mm->page_table_lock);
b9bbfbe3
AA		 865 mem_cgroup_uncharge_start();
866 for (i = 0; i < HPAGE_PMD_NR; i++) {
867 mem_cgroup_uncharge_page(pages[i]);
71e3aac0 868 put_page(pages[i]);
b9bbfbe3
AA		 869 }
870 mem_cgroup_uncharge_end();
71e3aac0
AA		 871 kfree(pages);
872 goto out;
873}
874
875int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
876 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
877{
878 int ret = 0;
879 struct page *page, *new_page;
880 unsigned long haddr;
881
882 VM_BUG_ON(!vma->anon_vma);
883 spin_lock(&mm->page_table_lock);
884 if (unlikely(!pmd_same(*pmd, orig_pmd)))
885 goto out_unlock;
886
887 page = pmd_page(orig_pmd);
888 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
889 haddr = address & HPAGE_PMD_MASK;
890 if (page_mapcount(page) == 1) {
891 pmd_t entry;
892 entry = pmd_mkyoung(orig_pmd);
893 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
894 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
895 update_mmu_cache(vma, address, entry);
896 ret |= VM_FAULT_WRITE;
897 goto out_unlock;
898 }
899 get_page(page);
900 spin_unlock(&mm->page_table_lock);
901
902 if (transparent_hugepage_enabled(vma) &&
903 !transparent_hugepage_debug_cow())
0bbbc0b3
AA		 904 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
905 vma, haddr);
71e3aac0
AA		 906 else
907 new_page = NULL;
908
909 if (unlikely(!new_page)) {
910 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
911 pmd, orig_pmd, page, haddr);
912 put_page(page);
913 goto out;
914 }
915
b9bbfbe3
AA		 916 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
917 put_page(new_page);
918 put_page(page);
919 ret |= VM_FAULT_OOM;
920 goto out;
921 }
922
71e3aac0
AA		 923 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
924 __SetPageUptodate(new_page);
925
926 spin_lock(&mm->page_table_lock);
927 put_page(page);
b9bbfbe3
AA		 928 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
929 mem_cgroup_uncharge_page(new_page);
71e3aac0 930 put_page(new_page);
b9bbfbe3 931 } else {
71e3aac0
AA		 932 pmd_t entry;
933 VM_BUG_ON(!PageHead(page));
934 entry = mk_pmd(new_page, vma->vm_page_prot);
935 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
936 entry = pmd_mkhuge(entry);
937 pmdp_clear_flush_notify(vma, haddr, pmd);
938 page_add_new_anon_rmap(new_page, vma, haddr);
939 set_pmd_at(mm, haddr, pmd, entry);
940 update_mmu_cache(vma, address, entry);
941 page_remove_rmap(page);
942 put_page(page);
943 ret |= VM_FAULT_WRITE;
944 }
945out_unlock:
946 spin_unlock(&mm->page_table_lock);
947out:
948 return ret;
949}
950
951struct page *follow_trans_huge_pmd(struct mm_struct *mm,
952 unsigned long addr,
953 pmd_t *pmd,
954 unsigned int flags)
955{
956 struct page *page = NULL;
957
958 assert_spin_locked(&mm->page_table_lock);
959
960 if (flags & FOLL_WRITE && !pmd_write(*pmd))
961 goto out;
962
963 page = pmd_page(*pmd);
964 VM_BUG_ON(!PageHead(page));
965 if (flags & FOLL_TOUCH) {
966 pmd_t _pmd;
967 /*
968 * We should set the dirty bit only for FOLL_WRITE but
969 * for now the dirty bit in the pmd is meaningless.
970 * And if the dirty bit will become meaningful and
971 * we'll only set it with FOLL_WRITE, an atomic
972 * set_bit will be required on the pmd to set the
973 * young bit, instead of the current set_pmd_at.
974 */
975 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
976 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
977 }
978 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
979 VM_BUG_ON(!PageCompound(page));
980 if (flags & FOLL_GET)
981 get_page(page);
982
983out:
984 return page;
985}
986
987int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
988 pmd_t *pmd)
989{
990 int ret = 0;
991
992 spin_lock(&tlb->mm->page_table_lock);
993 if (likely(pmd_trans_huge(*pmd))) {
994 if (unlikely(pmd_trans_splitting(*pmd))) {
995 spin_unlock(&tlb->mm->page_table_lock);
996 wait_split_huge_page(vma->anon_vma,
997 pmd);
998 } else {
999 struct page *page;
1000 pgtable_t pgtable;
1001 pgtable = get_pmd_huge_pte(tlb->mm);
1002 page = pmd_page(*pmd);
1003 pmd_clear(pmd);
1004 page_remove_rmap(page);
1005 VM_BUG_ON(page_mapcount(page) < 0);
1006 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1007 VM_BUG_ON(!PageHead(page));
1008 spin_unlock(&tlb->mm->page_table_lock);
1009 tlb_remove_page(tlb, page);
1010 pte_free(tlb->mm, pgtable);
1011 ret = 1;
1012 }
1013 } else
1014 spin_unlock(&tlb->mm->page_table_lock);
1015
1016 return ret;
1017}
1018
0ca1634d
JW		 1019int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1020 unsigned long addr, unsigned long end,
1021 unsigned char *vec)
1022{
1023 int ret = 0;
1024
1025 spin_lock(&vma->vm_mm->page_table_lock);
1026 if (likely(pmd_trans_huge(*pmd))) {
1027 ret = !pmd_trans_splitting(*pmd);
1028 spin_unlock(&vma->vm_mm->page_table_lock);
1029 if (unlikely(!ret))
1030 wait_split_huge_page(vma->anon_vma, pmd);
1031 else {
1032 /*
1033 * All logical pages in the range are present
1034 * if backed by a huge page.
1035 */
1036 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1037 }
1038 } else
1039 spin_unlock(&vma->vm_mm->page_table_lock);
1040
1041 return ret;
1042}
1043
cd7548ab
JW		 1044int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1045 unsigned long addr, pgprot_t newprot)
1046{
1047 struct mm_struct *mm = vma->vm_mm;
1048 int ret = 0;
1049
1050 spin_lock(&mm->page_table_lock);
1051 if (likely(pmd_trans_huge(*pmd))) {
1052 if (unlikely(pmd_trans_splitting(*pmd))) {
1053 spin_unlock(&mm->page_table_lock);
1054 wait_split_huge_page(vma->anon_vma, pmd);
1055 } else {
1056 pmd_t entry;
1057
1058 entry = pmdp_get_and_clear(mm, addr, pmd);
1059 entry = pmd_modify(entry, newprot);
1060 set_pmd_at(mm, addr, pmd, entry);
1061 spin_unlock(&vma->vm_mm->page_table_lock);
1062 flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1063 ret = 1;
1064 }
1065 } else
1066 spin_unlock(&vma->vm_mm->page_table_lock);
1067
1068 return ret;
1069}
1070
71e3aac0
AA		 1071pmd_t *page_check_address_pmd(struct page *page,
1072 struct mm_struct *mm,
1073 unsigned long address,
1074 enum page_check_address_pmd_flag flag)
1075{
1076 pgd_t *pgd;
1077 pud_t *pud;
1078 pmd_t *pmd, *ret = NULL;
1079
1080 if (address & ~HPAGE_PMD_MASK)
1081 goto out;
1082
1083 pgd = pgd_offset(mm, address);
1084 if (!pgd_present(*pgd))
1085 goto out;
1086
1087 pud = pud_offset(pgd, address);
1088 if (!pud_present(*pud))
1089 goto out;
1090
1091 pmd = pmd_offset(pud, address);
1092 if (pmd_none(*pmd))
1093 goto out;
1094 if (pmd_page(*pmd) != page)
1095 goto out;
94fcc585
AA		 1096 /*
1097 * split_vma() may create temporary aliased mappings. There is
1098 * no risk as long as all huge pmd are found and have their
1099 * splitting bit set before __split_huge_page_refcount
1100 * runs. Finding the same huge pmd more than once during the
1101 * same rmap walk is not a problem.
1102 */
1103 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1104 pmd_trans_splitting(*pmd))
1105 goto out;
71e3aac0
AA		 1106 if (pmd_trans_huge(*pmd)) {
1107 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1108 !pmd_trans_splitting(*pmd));
1109 ret = pmd;
1110 }
1111out:
1112 return ret;
1113}
1114
1115static int __split_huge_page_splitting(struct page *page,
1116 struct vm_area_struct *vma,
1117 unsigned long address)
1118{
1119 struct mm_struct *mm = vma->vm_mm;
1120 pmd_t *pmd;
1121 int ret = 0;
1122
1123 spin_lock(&mm->page_table_lock);
1124 pmd = page_check_address_pmd(page, mm, address,
1125 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1126 if (pmd) {
1127 /*
1128 * We can't temporarily set the pmd to null in order
1129 * to split it, the pmd must remain marked huge at all
1130 * times or the VM won't take the pmd_trans_huge paths
1131 * and it won't wait on the anon_vma->root->lock to
1132 * serialize against split_huge_page*.
1133 */
1134 pmdp_splitting_flush_notify(vma, address, pmd);
1135 ret = 1;
1136 }
1137 spin_unlock(&mm->page_table_lock);
1138
1139 return ret;
1140}
1141
1142static void __split_huge_page_refcount(struct page *page)
1143{
1144 int i;
1145 unsigned long head_index = page->index;
1146 struct zone *zone = page_zone(page);
2c888cfb 1147 int zonestat;
71e3aac0
AA		 1148
1149 /* prevent PageLRU to go away from under us, and freeze lru stats */
1150 spin_lock_irq(&zone->lru_lock);
1151 compound_lock(page);
1152
1153 for (i = 1; i < HPAGE_PMD_NR; i++) {
1154 struct page *page_tail = page + i;
1155
1156 /* tail_page->_count cannot change */
1157 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1158 BUG_ON(page_count(page) <= 0);
1159 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1160 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1161
1162 /* after clearing PageTail the gup refcount can be released */
1163 smp_mb();
1164
1165 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1166 page_tail->flags |= (page->flags &
1167 ((1L << PG_referenced) |
1168 (1L << PG_swapbacked) |
1169 (1L << PG_mlocked) |
1170 (1L << PG_uptodate)));
1171 page_tail->flags |= (1L << PG_dirty);
1172
1173 /*
1174 * 1) clear PageTail before overwriting first_page
1175 * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1176 */
1177 smp_wmb();
1178
1179 /*
1180 * __split_huge_page_splitting() already set the
1181 * splitting bit in all pmd that could map this
1182 * hugepage, that will ensure no CPU can alter the
1183 * mapcount on the head page. The mapcount is only
1184 * accounted in the head page and it has to be
1185 * transferred to all tail pages in the below code. So
1186 * for this code to be safe, the split the mapcount
1187 * can't change. But that doesn't mean userland can't
1188 * keep changing and reading the page contents while
1189 * we transfer the mapcount, so the pmd splitting
1190 * status is achieved setting a reserved bit in the
1191 * pmd, not by clearing the present bit.
1192 */
1193 BUG_ON(page_mapcount(page_tail));
1194 page_tail->_mapcount = page->_mapcount;
1195
1196 BUG_ON(page_tail->mapping);
1197 page_tail->mapping = page->mapping;
1198
1199 page_tail->index = ++head_index;
1200
1201 BUG_ON(!PageAnon(page_tail));
1202 BUG_ON(!PageUptodate(page_tail));
1203 BUG_ON(!PageDirty(page_tail));
1204 BUG_ON(!PageSwapBacked(page_tail));
1205
1206 lru_add_page_tail(zone, page, page_tail);
1207 }
1208
79134171
AA		 1209 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1210 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1211
2c888cfb
RR		 1212 /*
1213 * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1214 * so adjust those appropriately if this page is on the LRU.
1215 */
1216 if (PageLRU(page)) {
1217 zonestat = NR_LRU_BASE + page_lru(page);
1218 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1219 }
1220
71e3aac0
AA		 1221 ClearPageCompound(page);
1222 compound_unlock(page);
1223 spin_unlock_irq(&zone->lru_lock);
1224
1225 for (i = 1; i < HPAGE_PMD_NR; i++) {
1226 struct page *page_tail = page + i;
1227 BUG_ON(page_count(page_tail) <= 0);
1228 /*
1229 * Tail pages may be freed if there wasn't any mapping
1230 * like if add_to_swap() is running on a lru page that
1231 * had its mapping zapped. And freeing these pages
1232 * requires taking the lru_lock so we do the put_page
1233 * of the tail pages after the split is complete.
1234 */
1235 put_page(page_tail);
1236 }
1237
1238 /*
1239 * Only the head page (now become a regular page) is required
1240 * to be pinned by the caller.
1241 */
1242 BUG_ON(page_count(page) <= 0);
1243}
1244
1245static int __split_huge_page_map(struct page *page,
1246 struct vm_area_struct *vma,
1247 unsigned long address)
1248{
1249 struct mm_struct *mm = vma->vm_mm;
1250 pmd_t *pmd, _pmd;
1251 int ret = 0, i;
1252 pgtable_t pgtable;
1253 unsigned long haddr;
1254
1255 spin_lock(&mm->page_table_lock);
1256 pmd = page_check_address_pmd(page, mm, address,
1257 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1258 if (pmd) {
1259 pgtable = get_pmd_huge_pte(mm);
1260 pmd_populate(mm, &_pmd, pgtable);
1261
1262 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1263 i++, haddr += PAGE_SIZE) {
1264 pte_t *pte, entry;
1265 BUG_ON(PageCompound(page+i));
1266 entry = mk_pte(page + i, vma->vm_page_prot);
1267 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1268 if (!pmd_write(*pmd))
1269 entry = pte_wrprotect(entry);
1270 else
1271 BUG_ON(page_mapcount(page) != 1);
1272 if (!pmd_young(*pmd))
1273 entry = pte_mkold(entry);
1274 pte = pte_offset_map(&_pmd, haddr);
1275 BUG_ON(!pte_none(*pte));
1276 set_pte_at(mm, haddr, pte, entry);
1277 pte_unmap(pte);
1278 }
1279
1280 mm->nr_ptes++;
1281 smp_wmb(); /* make pte visible before pmd */
1282 /*
1283 * Up to this point the pmd is present and huge and
1284 * userland has the whole access to the hugepage
1285 * during the split (which happens in place). If we
1286 * overwrite the pmd with the not-huge version
1287 * pointing to the pte here (which of course we could
1288 * if all CPUs were bug free), userland could trigger
1289 * a small page size TLB miss on the small sized TLB
1290 * while the hugepage TLB entry is still established
1291 * in the huge TLB. Some CPU doesn't like that. See
1292 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1293 * Erratum 383 on page 93. Intel should be safe but is
1294 * also warns that it's only safe if the permission
1295 * and cache attributes of the two entries loaded in
1296 * the two TLB is identical (which should be the case
1297 * here). But it is generally safer to never allow
1298 * small and huge TLB entries for the same virtual
1299 * address to be loaded simultaneously. So instead of
1300 * doing "pmd_populate(); flush_tlb_range();" we first
1301 * mark the current pmd notpresent (atomically because
1302 * here the pmd_trans_huge and pmd_trans_splitting
1303 * must remain set at all times on the pmd until the
1304 * split is complete for this pmd), then we flush the
1305 * SMP TLB and finally we write the non-huge version
1306 * of the pmd entry with pmd_populate.
1307 */
1308 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1309 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1310 pmd_populate(mm, pmd, pgtable);
1311 ret = 1;
1312 }
1313 spin_unlock(&mm->page_table_lock);
1314
1315 return ret;
1316}
1317
1318/* must be called with anon_vma->root->lock hold */
1319static void __split_huge_page(struct page *page,
1320 struct anon_vma *anon_vma)
1321{
1322 int mapcount, mapcount2;
1323 struct anon_vma_chain *avc;
1324
1325 BUG_ON(!PageHead(page));
1326 BUG_ON(PageTail(page));
1327
1328 mapcount = 0;
1329 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1330 struct vm_area_struct *vma = avc->vma;
1331 unsigned long addr = vma_address(page, vma);
1332 BUG_ON(is_vma_temporary_stack(vma));
1333 if (addr == -EFAULT)
1334 continue;
1335 mapcount += __split_huge_page_splitting(page, vma, addr);
1336 }
05759d38
AA		 1337 /*
1338 * It is critical that new vmas are added to the tail of the
1339 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1340 * and establishes a child pmd before
1341 * __split_huge_page_splitting() freezes the parent pmd (so if
1342 * we fail to prevent copy_huge_pmd() from running until the
1343 * whole __split_huge_page() is complete), we will still see
1344 * the newly established pmd of the child later during the
1345 * walk, to be able to set it as pmd_trans_splitting too.
1346 */
1347 if (mapcount != page_mapcount(page))
1348 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1349 mapcount, page_mapcount(page));
71e3aac0
AA		 1350 BUG_ON(mapcount != page_mapcount(page));
1351
1352 __split_huge_page_refcount(page);
1353
1354 mapcount2 = 0;
1355 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1356 struct vm_area_struct *vma = avc->vma;
1357 unsigned long addr = vma_address(page, vma);
1358 BUG_ON(is_vma_temporary_stack(vma));
1359 if (addr == -EFAULT)
1360 continue;
1361 mapcount2 += __split_huge_page_map(page, vma, addr);
1362 }
05759d38
AA		 1363 if (mapcount != mapcount2)
1364 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1365 mapcount, mapcount2, page_mapcount(page));
71e3aac0
AA		 1366 BUG_ON(mapcount != mapcount2);
1367}
1368
1369int split_huge_page(struct page *page)
1370{
1371 struct anon_vma *anon_vma;
1372 int ret = 1;
1373
1374 BUG_ON(!PageAnon(page));
1375 anon_vma = page_lock_anon_vma(page);
1376 if (!anon_vma)
1377 goto out;
1378 ret = 0;
1379 if (!PageCompound(page))
1380 goto out_unlock;
1381
1382 BUG_ON(!PageSwapBacked(page));
1383 __split_huge_page(page, anon_vma);
1384
1385 BUG_ON(PageCompound(page));
1386out_unlock:
1387 page_unlock_anon_vma(anon_vma);
1388out:
1389 return ret;
1390}
1391
60ab3244
AA		 1392int hugepage_madvise(struct vm_area_struct *vma,
1393 unsigned long *vm_flags, int advice)
0af4e98b 1394{
a664b2d8
AA		 1395 switch (advice) {
1396 case MADV_HUGEPAGE:
1397 /*
1398 * Be somewhat over-protective like KSM for now!
1399 */
1400 if (*vm_flags & (VM_HUGEPAGE |
1401 VM_SHARED | VM_MAYSHARE |
1402 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1403 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1404 VM_MIXEDMAP | VM_SAO))
1405 return -EINVAL;
1406 *vm_flags &= ~VM_NOHUGEPAGE;
1407 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 1408 /*
1409 * If the vma become good for khugepaged to scan,
1410 * register it here without waiting a page fault that
1411 * may not happen any time soon.
1412 */
1413 if (unlikely(khugepaged_enter_vma_merge(vma)))
1414 return -ENOMEM;
a664b2d8
AA		 1415 break;
1416 case MADV_NOHUGEPAGE:
1417 /*
1418 * Be somewhat over-protective like KSM for now!
1419 */
1420 if (*vm_flags & (VM_NOHUGEPAGE |
1421 VM_SHARED | VM_MAYSHARE |
1422 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1423 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1424 VM_MIXEDMAP | VM_SAO))
1425 return -EINVAL;
1426 *vm_flags &= ~VM_HUGEPAGE;
1427 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 1428 /*
1429 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1430 * this vma even if we leave the mm registered in khugepaged if
1431 * it got registered before VM_NOHUGEPAGE was set.
1432 */
a664b2d8
AA		 1433 break;
1434 }
0af4e98b
AA		 1435
1436 return 0;
1437}
1438
ba76149f
AA		 1439static int __init khugepaged_slab_init(void)
1440{
1441 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1442 sizeof(struct mm_slot),
1443 __alignof__(struct mm_slot), 0, NULL);
1444 if (!mm_slot_cache)
1445 return -ENOMEM;
1446
1447 return 0;
1448}
1449
1450static void __init khugepaged_slab_free(void)
1451{
1452 kmem_cache_destroy(mm_slot_cache);
1453 mm_slot_cache = NULL;
1454}
1455
1456static inline struct mm_slot *alloc_mm_slot(void)
1457{
1458 if (!mm_slot_cache) /* initialization failed */
1459 return NULL;
1460 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1461}
1462
1463static inline void free_mm_slot(struct mm_slot *mm_slot)
1464{
1465 kmem_cache_free(mm_slot_cache, mm_slot);
1466}
1467
1468static int __init mm_slots_hash_init(void)
1469{
1470 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1471 GFP_KERNEL);
1472 if (!mm_slots_hash)
1473 return -ENOMEM;
1474 return 0;
1475}
1476
1477#if 0
1478static void __init mm_slots_hash_free(void)
1479{
1480 kfree(mm_slots_hash);
1481 mm_slots_hash = NULL;
1482}
1483#endif
1484
1485static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1486{
1487 struct mm_slot *mm_slot;
1488 struct hlist_head *bucket;
1489 struct hlist_node *node;
1490
1491 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1492 % MM_SLOTS_HASH_HEADS];
1493 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1494 if (mm == mm_slot->mm)
1495 return mm_slot;
1496 }
1497 return NULL;
1498}
1499
1500static void insert_to_mm_slots_hash(struct mm_struct *mm,
1501 struct mm_slot *mm_slot)
1502{
1503 struct hlist_head *bucket;
1504
1505 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1506 % MM_SLOTS_HASH_HEADS];
1507 mm_slot->mm = mm;
1508 hlist_add_head(&mm_slot->hash, bucket);
1509}
1510
1511static inline int khugepaged_test_exit(struct mm_struct *mm)
1512{
1513 return atomic_read(&mm->mm_users) == 0;
1514}
1515
1516int __khugepaged_enter(struct mm_struct *mm)
1517{
1518 struct mm_slot *mm_slot;
1519 int wakeup;
1520
1521 mm_slot = alloc_mm_slot();
1522 if (!mm_slot)
1523 return -ENOMEM;
1524
1525 /* __khugepaged_exit() must not run from under us */
1526 VM_BUG_ON(khugepaged_test_exit(mm));
1527 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1528 free_mm_slot(mm_slot);
1529 return 0;
1530 }
1531
1532 spin_lock(&khugepaged_mm_lock);
1533 insert_to_mm_slots_hash(mm, mm_slot);
1534 /*
1535 * Insert just behind the scanning cursor, to let the area settle
1536 * down a little.
1537 */
1538 wakeup = list_empty(&khugepaged_scan.mm_head);
1539 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1540 spin_unlock(&khugepaged_mm_lock);
1541
1542 atomic_inc(&mm->mm_count);
1543 if (wakeup)
1544 wake_up_interruptible(&khugepaged_wait);
1545
1546 return 0;
1547}
1548
1549int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1550{
1551 unsigned long hstart, hend;
1552 if (!vma->anon_vma)
1553 /*
1554 * Not yet faulted in so we will register later in the
1555 * page fault if needed.
1556 */
1557 return 0;
1558 if (vma->vm_file || vma->vm_ops)
1559 /* khugepaged not yet working on file or special mappings */
1560 return 0;
1561 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1562 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1563 hend = vma->vm_end & HPAGE_PMD_MASK;
1564 if (hstart < hend)
1565 return khugepaged_enter(vma);
1566 return 0;
1567}
1568
1569void __khugepaged_exit(struct mm_struct *mm)
1570{
1571 struct mm_slot *mm_slot;
1572 int free = 0;
1573
1574 spin_lock(&khugepaged_mm_lock);
1575 mm_slot = get_mm_slot(mm);
1576 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1577 hlist_del(&mm_slot->hash);
1578 list_del(&mm_slot->mm_node);
1579 free = 1;
1580 }
1581
1582 if (free) {
1583 spin_unlock(&khugepaged_mm_lock);
1584 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1585 free_mm_slot(mm_slot);
1586 mmdrop(mm);
1587 } else if (mm_slot) {
1588 spin_unlock(&khugepaged_mm_lock);
1589 /*
1590 * This is required to serialize against
1591 * khugepaged_test_exit() (which is guaranteed to run
1592 * under mmap sem read mode). Stop here (after we
1593 * return all pagetables will be destroyed) until
1594 * khugepaged has finished working on the pagetables
1595 * under the mmap_sem.
1596 */
1597 down_write(&mm->mmap_sem);
1598 up_write(&mm->mmap_sem);
1599 } else
1600 spin_unlock(&khugepaged_mm_lock);
1601}
1602
1603static void release_pte_page(struct page *page)
1604{
1605 /* 0 stands for page_is_file_cache(page) == false */
1606 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1607 unlock_page(page);
1608 putback_lru_page(page);
1609}
1610
1611static void release_pte_pages(pte_t *pte, pte_t *_pte)
1612{
1613 while (--_pte >= pte) {
1614 pte_t pteval = *_pte;
1615 if (!pte_none(pteval))
1616 release_pte_page(pte_page(pteval));
1617 }
1618}
1619
1620static void release_all_pte_pages(pte_t *pte)
1621{
1622 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1623}
1624
1625static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1626 unsigned long address,
1627 pte_t *pte)
1628{
1629 struct page *page;
1630 pte_t *_pte;
1631 int referenced = 0, isolated = 0, none = 0;
1632 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1633 _pte++, address += PAGE_SIZE) {
1634 pte_t pteval = *_pte;
1635 if (pte_none(pteval)) {
1636 if (++none <= khugepaged_max_ptes_none)
1637 continue;
1638 else {
1639 release_pte_pages(pte, _pte);
1640 goto out;
1641 }
1642 }
1643 if (!pte_present(pteval) || !pte_write(pteval)) {
1644 release_pte_pages(pte, _pte);
1645 goto out;
1646 }
1647 page = vm_normal_page(vma, address, pteval);
1648 if (unlikely(!page)) {
1649 release_pte_pages(pte, _pte);
1650 goto out;
1651 }
1652 VM_BUG_ON(PageCompound(page));
1653 BUG_ON(!PageAnon(page));
1654 VM_BUG_ON(!PageSwapBacked(page));
1655
1656 /* cannot use mapcount: can't collapse if there's a gup pin */
1657 if (page_count(page) != 1) {
1658 release_pte_pages(pte, _pte);
1659 goto out;
1660 }
1661 /*
1662 * We can do it before isolate_lru_page because the
1663 * page can't be freed from under us. NOTE: PG_lock
1664 * is needed to serialize against split_huge_page
1665 * when invoked from the VM.
1666 */
1667 if (!trylock_page(page)) {
1668 release_pte_pages(pte, _pte);
1669 goto out;
1670 }
1671 /*
1672 * Isolate the page to avoid collapsing an hugepage
1673 * currently in use by the VM.
1674 */
1675 if (isolate_lru_page(page)) {
1676 unlock_page(page);
1677 release_pte_pages(pte, _pte);
1678 goto out;
1679 }
1680 /* 0 stands for page_is_file_cache(page) == false */
1681 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1682 VM_BUG_ON(!PageLocked(page));
1683 VM_BUG_ON(PageLRU(page));
1684
1685 /* If there is no mapped pte young don't collapse the page */
8ee53820
AA		 1686 if (pte_young(pteval) || PageReferenced(page) ||
1687 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 1688 referenced = 1;
1689 }
1690 if (unlikely(!referenced))
1691 release_all_pte_pages(pte);
1692 else
1693 isolated = 1;
1694out:
1695 return isolated;
1696}
1697
1698static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1699 struct vm_area_struct *vma,
1700 unsigned long address,
1701 spinlock_t *ptl)
1702{
1703 pte_t *_pte;
1704 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1705 pte_t pteval = *_pte;
1706 struct page *src_page;
1707
1708 if (pte_none(pteval)) {
1709 clear_user_highpage(page, address);
1710 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1711 } else {
1712 src_page = pte_page(pteval);
1713 copy_user_highpage(page, src_page, address, vma);
1714 VM_BUG_ON(page_mapcount(src_page) != 1);
1715 VM_BUG_ON(page_count(src_page) != 2);
1716 release_pte_page(src_page);
1717 /*
1718 * ptl mostly unnecessary, but preempt has to
1719 * be disabled to update the per-cpu stats
1720 * inside page_remove_rmap().
1721 */
1722 spin_lock(ptl);
1723 /*
1724 * paravirt calls inside pte_clear here are
1725 * superfluous.
1726 */
1727 pte_clear(vma->vm_mm, address, _pte);
1728 page_remove_rmap(src_page);
1729 spin_unlock(ptl);
1730 free_page_and_swap_cache(src_page);
1731 }
1732
1733 address += PAGE_SIZE;
1734 page++;
1735 }
1736}
1737
1738static void collapse_huge_page(struct mm_struct *mm,
1739 unsigned long address,
ce83d217
AA		 1740 struct page **hpage,
1741 struct vm_area_struct *vma)
ba76149f 1742{
ba76149f
AA		 1743 pgd_t *pgd;
1744 pud_t *pud;
1745 pmd_t *pmd, _pmd;
1746 pte_t *pte;
1747 pgtable_t pgtable;
1748 struct page *new_page;
1749 spinlock_t *ptl;
1750 int isolated;
1751 unsigned long hstart, hend;
1752
1753 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
0bbbc0b3 1754#ifndef CONFIG_NUMA
ba76149f 1755 VM_BUG_ON(!*hpage);
ce83d217 1756 new_page = *hpage;
0bbbc0b3
AA		 1757#else
1758 VM_BUG_ON(*hpage);
ce83d217
AA		 1759 /*
1760 * Allocate the page while the vma is still valid and under
1761 * the mmap_sem read mode so there is no memory allocation
1762 * later when we take the mmap_sem in write mode. This is more
1763 * friendly behavior (OTOH it may actually hide bugs) to
1764 * filesystems in userland with daemons allocating memory in
1765 * the userland I/O paths. Allocating memory with the
1766 * mmap_sem in read mode is good idea also to allow greater
1767 * scalability.
1768 */
1769 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address);
1770 if (unlikely(!new_page)) {
1771 up_read(&mm->mmap_sem);
1772 *hpage = ERR_PTR(-ENOMEM);
1773 return;
1774 }
0bbbc0b3 1775#endif
ce83d217
AA		 1776 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1777 up_read(&mm->mmap_sem);
1778 put_page(new_page);
1779 return;
1780 }
1781
1782 /* after allocating the hugepage upgrade to mmap_sem write mode */
1783 up_read(&mm->mmap_sem);
ba76149f
AA		 1784
1785 /*
1786 * Prevent all access to pagetables with the exception of
1787 * gup_fast later hanlded by the ptep_clear_flush and the VM
1788 * handled by the anon_vma lock + PG_lock.
1789 */
1790 down_write(&mm->mmap_sem);
1791 if (unlikely(khugepaged_test_exit(mm)))
1792 goto out;
1793
1794 vma = find_vma(mm, address);
1795 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1796 hend = vma->vm_end & HPAGE_PMD_MASK;
1797 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1798 goto out;
1799
60ab3244
AA		 1800 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1801 (vma->vm_flags & VM_NOHUGEPAGE))
ba76149f
AA		 1802 goto out;
1803
1804 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1805 if (!vma->anon_vma || vma->vm_ops || vma->vm_file)
1806 goto out;
1807 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1808
1809 pgd = pgd_offset(mm, address);
1810 if (!pgd_present(*pgd))
1811 goto out;
1812
1813 pud = pud_offset(pgd, address);
1814 if (!pud_present(*pud))
1815 goto out;
1816
1817 pmd = pmd_offset(pud, address);
1818 /* pmd can't go away or become huge under us */
1819 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1820 goto out;
1821
ba76149f
AA		 1822 anon_vma_lock(vma->anon_vma);
1823
1824 pte = pte_offset_map(pmd, address);
1825 ptl = pte_lockptr(mm, pmd);
1826
1827 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1828 /*
1829 * After this gup_fast can't run anymore. This also removes
1830 * any huge TLB entry from the CPU so we won't allow
1831 * huge and small TLB entries for the same virtual address
1832 * to avoid the risk of CPU bugs in that area.
1833 */
1834 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1835 spin_unlock(&mm->page_table_lock);
1836
1837 spin_lock(ptl);
1838 isolated = __collapse_huge_page_isolate(vma, address, pte);
1839 spin_unlock(ptl);
1840 pte_unmap(pte);
1841
1842 if (unlikely(!isolated)) {
1843 spin_lock(&mm->page_table_lock);
1844 BUG_ON(!pmd_none(*pmd));
1845 set_pmd_at(mm, address, pmd, _pmd);
1846 spin_unlock(&mm->page_table_lock);
1847 anon_vma_unlock(vma->anon_vma);
1848 mem_cgroup_uncharge_page(new_page);
ce83d217 1849 goto out;
ba76149f
AA		 1850 }
1851
1852 /*
1853 * All pages are isolated and locked so anon_vma rmap
1854 * can't run anymore.
1855 */
1856 anon_vma_unlock(vma->anon_vma);
1857
1858 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1859 __SetPageUptodate(new_page);
1860 pgtable = pmd_pgtable(_pmd);
1861 VM_BUG_ON(page_count(pgtable) != 1);
1862 VM_BUG_ON(page_mapcount(pgtable) != 0);
1863
1864 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1865 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1866 _pmd = pmd_mkhuge(_pmd);
1867
1868 /*
1869 * spin_lock() below is not the equivalent of smp_wmb(), so
1870 * this is needed to avoid the copy_huge_page writes to become
1871 * visible after the set_pmd_at() write.
1872 */
1873 smp_wmb();
1874
1875 spin_lock(&mm->page_table_lock);
1876 BUG_ON(!pmd_none(*pmd));
1877 page_add_new_anon_rmap(new_page, vma, address);
1878 set_pmd_at(mm, address, pmd, _pmd);
1879 update_mmu_cache(vma, address, entry);
1880 prepare_pmd_huge_pte(pgtable, mm);
1881 mm->nr_ptes--;
1882 spin_unlock(&mm->page_table_lock);
1883
0bbbc0b3 1884#ifndef CONFIG_NUMA
ba76149f 1885 *hpage = NULL;
0bbbc0b3 1886#endif
ba76149f 1887 khugepaged_pages_collapsed++;
ce83d217 1888out_up_write:
ba76149f 1889 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 1890 return;
1891
ce83d217 1892out:
0bbbc0b3
AA		 1893#ifdef CONFIG_NUMA
1894 put_page(new_page);
1895#endif
ce83d217 1896 goto out_up_write;
ba76149f
AA		 1897}
1898
1899static int khugepaged_scan_pmd(struct mm_struct *mm,
1900 struct vm_area_struct *vma,
1901 unsigned long address,
1902 struct page **hpage)
1903{
1904 pgd_t *pgd;
1905 pud_t *pud;
1906 pmd_t *pmd;
1907 pte_t *pte, *_pte;
1908 int ret = 0, referenced = 0, none = 0;
1909 struct page *page;
1910 unsigned long _address;
1911 spinlock_t *ptl;
1912
1913 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1914
1915 pgd = pgd_offset(mm, address);
1916 if (!pgd_present(*pgd))
1917 goto out;
1918
1919 pud = pud_offset(pgd, address);
1920 if (!pud_present(*pud))
1921 goto out;
1922
1923 pmd = pmd_offset(pud, address);
1924 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1925 goto out;
1926
1927 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1928 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1929 _pte++, _address += PAGE_SIZE) {
1930 pte_t pteval = *_pte;
1931 if (pte_none(pteval)) {
1932 if (++none <= khugepaged_max_ptes_none)
1933 continue;
1934 else
1935 goto out_unmap;
1936 }
1937 if (!pte_present(pteval) || !pte_write(pteval))
1938 goto out_unmap;
1939 page = vm_normal_page(vma, _address, pteval);
1940 if (unlikely(!page))
1941 goto out_unmap;
1942 VM_BUG_ON(PageCompound(page));
1943 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1944 goto out_unmap;
1945 /* cannot use mapcount: can't collapse if there's a gup pin */
1946 if (page_count(page) != 1)
1947 goto out_unmap;
8ee53820
AA		 1948 if (pte_young(pteval) || PageReferenced(page) ||
1949 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 1950 referenced = 1;
1951 }
1952 if (referenced)
1953 ret = 1;
1954out_unmap:
1955 pte_unmap_unlock(pte, ptl);
ce83d217
AA		 1956 if (ret)
1957 /* collapse_huge_page will return with the mmap_sem released */
1958 collapse_huge_page(mm, address, hpage, vma);
ba76149f
AA		 1959out:
1960 return ret;
1961}
1962
1963static void collect_mm_slot(struct mm_slot *mm_slot)
1964{
1965 struct mm_struct *mm = mm_slot->mm;
1966
1967 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1968
1969 if (khugepaged_test_exit(mm)) {
1970 /* free mm_slot */
1971 hlist_del(&mm_slot->hash);
1972 list_del(&mm_slot->mm_node);
1973
1974 /*
1975 * Not strictly needed because the mm exited already.
1976 *
1977 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1978 */
1979
1980 /* khugepaged_mm_lock actually not necessary for the below */
1981 free_mm_slot(mm_slot);
1982 mmdrop(mm);
1983 }
1984}
1985
1986static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1987 struct page **hpage)
1988{
1989 struct mm_slot *mm_slot;
1990 struct mm_struct *mm;
1991 struct vm_area_struct *vma;
1992 int progress = 0;
1993
1994 VM_BUG_ON(!pages);
1995 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1996
1997 if (khugepaged_scan.mm_slot)
1998 mm_slot = khugepaged_scan.mm_slot;
1999 else {
2000 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2001 struct mm_slot, mm_node);
2002 khugepaged_scan.address = 0;
2003 khugepaged_scan.mm_slot = mm_slot;
2004 }
2005 spin_unlock(&khugepaged_mm_lock);
2006
2007 mm = mm_slot->mm;
2008 down_read(&mm->mmap_sem);
2009 if (unlikely(khugepaged_test_exit(mm)))
2010 vma = NULL;
2011 else
2012 vma = find_vma(mm, khugepaged_scan.address);
2013
2014 progress++;
2015 for (; vma; vma = vma->vm_next) {
2016 unsigned long hstart, hend;
2017
2018 cond_resched();
2019 if (unlikely(khugepaged_test_exit(mm))) {
2020 progress++;
2021 break;
2022 }
2023
60ab3244
AA		 2024 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2025 !khugepaged_always()) ||
2026 (vma->vm_flags & VM_NOHUGEPAGE)) {
ba76149f
AA		 2027 progress++;
2028 continue;
2029 }
2030
2031 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
2032 if (!vma->anon_vma || vma->vm_ops || vma->vm_file) {
2033 khugepaged_scan.address = vma->vm_end;
2034 progress++;
2035 continue;
2036 }
2037 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
2038
2039 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2040 hend = vma->vm_end & HPAGE_PMD_MASK;
2041 if (hstart >= hend) {
2042 progress++;
2043 continue;
2044 }
2045 if (khugepaged_scan.address < hstart)
2046 khugepaged_scan.address = hstart;
2047 if (khugepaged_scan.address > hend) {
2048 khugepaged_scan.address = hend + HPAGE_PMD_SIZE;
2049 progress++;
2050 continue;
2051 }
2052 BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2053
2054 while (khugepaged_scan.address < hend) {
2055 int ret;
2056 cond_resched();
2057 if (unlikely(khugepaged_test_exit(mm)))
2058 goto breakouterloop;
2059
2060 VM_BUG_ON(khugepaged_scan.address < hstart ||
2061 khugepaged_scan.address + HPAGE_PMD_SIZE >
2062 hend);
2063 ret = khugepaged_scan_pmd(mm, vma,
2064 khugepaged_scan.address,
2065 hpage);
2066 /* move to next address */
2067 khugepaged_scan.address += HPAGE_PMD_SIZE;
2068 progress += HPAGE_PMD_NR;
2069 if (ret)
2070 /* we released mmap_sem so break loop */
2071 goto breakouterloop_mmap_sem;
2072 if (progress >= pages)
2073 goto breakouterloop;
2074 }
2075 }
2076breakouterloop:
2077 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2078breakouterloop_mmap_sem:
2079
2080 spin_lock(&khugepaged_mm_lock);
2081 BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2082 /*
2083 * Release the current mm_slot if this mm is about to die, or
2084 * if we scanned all vmas of this mm.
2085 */
2086 if (khugepaged_test_exit(mm) || !vma) {
2087 /*
2088 * Make sure that if mm_users is reaching zero while
2089 * khugepaged runs here, khugepaged_exit will find
2090 * mm_slot not pointing to the exiting mm.
2091 */
2092 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2093 khugepaged_scan.mm_slot = list_entry(
2094 mm_slot->mm_node.next,
2095 struct mm_slot, mm_node);
2096 khugepaged_scan.address = 0;
2097 } else {
2098 khugepaged_scan.mm_slot = NULL;
2099 khugepaged_full_scans++;
2100 }
2101
2102 collect_mm_slot(mm_slot);
2103 }
2104
2105 return progress;
2106}
2107
2108static int khugepaged_has_work(void)
2109{
2110 return !list_empty(&khugepaged_scan.mm_head) &&
2111 khugepaged_enabled();
2112}
2113
2114static int khugepaged_wait_event(void)
2115{
2116 return !list_empty(&khugepaged_scan.mm_head) ||
2117 !khugepaged_enabled();
2118}
2119
2120static void khugepaged_do_scan(struct page **hpage)
2121{
2122 unsigned int progress = 0, pass_through_head = 0;
2123 unsigned int pages = khugepaged_pages_to_scan;
2124
2125 barrier(); /* write khugepaged_pages_to_scan to local stack */
2126
2127 while (progress < pages) {
2128 cond_resched();
2129
0bbbc0b3 2130#ifndef CONFIG_NUMA
ba76149f
AA		 2131 if (!*hpage) {
2132 *hpage = alloc_hugepage(khugepaged_defrag());
2133 if (unlikely(!*hpage))
2134 break;
2135 }
0bbbc0b3
AA		 2136#else
2137 if (IS_ERR(*hpage))
2138 break;
2139#endif
ba76149f 2140
878aee7d
AA		 2141 if (unlikely(kthread_should_stop() || freezing(current)))
2142 break;
2143
ba76149f
AA		 2144 spin_lock(&khugepaged_mm_lock);
2145 if (!khugepaged_scan.mm_slot)
2146 pass_through_head++;
2147 if (khugepaged_has_work() &&
2148 pass_through_head < 2)
2149 progress += khugepaged_scan_mm_slot(pages - progress,
2150 hpage);
2151 else
2152 progress = pages;
2153 spin_unlock(&khugepaged_mm_lock);
2154 }
2155}
2156
0bbbc0b3
AA		 2157static void khugepaged_alloc_sleep(void)
2158{
2159 DEFINE_WAIT(wait);
2160 add_wait_queue(&khugepaged_wait, &wait);
2161 schedule_timeout_interruptible(
2162 msecs_to_jiffies(
2163 khugepaged_alloc_sleep_millisecs));
2164 remove_wait_queue(&khugepaged_wait, &wait);
2165}
2166
2167#ifndef CONFIG_NUMA
ba76149f
AA		 2168static struct page *khugepaged_alloc_hugepage(void)
2169{
2170 struct page *hpage;
2171
2172 do {
2173 hpage = alloc_hugepage(khugepaged_defrag());
0bbbc0b3
AA		 2174 if (!hpage)
2175 khugepaged_alloc_sleep();
ba76149f
AA		 2176 } while (unlikely(!hpage) &&
2177 likely(khugepaged_enabled()));
2178 return hpage;
2179}
0bbbc0b3 2180#endif
ba76149f
AA		 2181
2182static void khugepaged_loop(void)
2183{
2184 struct page *hpage;
2185
0bbbc0b3
AA		 2186#ifdef CONFIG_NUMA
2187 hpage = NULL;
2188#endif
ba76149f 2189 while (likely(khugepaged_enabled())) {
0bbbc0b3 2190#ifndef CONFIG_NUMA
ba76149f
AA		 2191 hpage = khugepaged_alloc_hugepage();
2192 if (unlikely(!hpage))
2193 break;
0bbbc0b3
AA		 2194#else
2195 if (IS_ERR(hpage)) {
2196 khugepaged_alloc_sleep();
2197 hpage = NULL;
2198 }
2199#endif
ba76149f
AA		 2200
2201 khugepaged_do_scan(&hpage);
0bbbc0b3 2202#ifndef CONFIG_NUMA
ba76149f
AA		 2203 if (hpage)
2204 put_page(hpage);
0bbbc0b3 2205#endif
878aee7d
AA		 2206 try_to_freeze();
2207 if (unlikely(kthread_should_stop()))
2208 break;
ba76149f
AA		 2209 if (khugepaged_has_work()) {
2210 DEFINE_WAIT(wait);
2211 if (!khugepaged_scan_sleep_millisecs)
2212 continue;
2213 add_wait_queue(&khugepaged_wait, &wait);
2214 schedule_timeout_interruptible(
2215 msecs_to_jiffies(
2216 khugepaged_scan_sleep_millisecs));
2217 remove_wait_queue(&khugepaged_wait, &wait);
2218 } else if (khugepaged_enabled())
878aee7d
AA		 2219 wait_event_freezable(khugepaged_wait,
2220 khugepaged_wait_event());
ba76149f
AA		 2221 }
2222}
2223
2224static int khugepaged(void *none)
2225{
2226 struct mm_slot *mm_slot;
2227
878aee7d 2228 set_freezable();
ba76149f
AA		 2229 set_user_nice(current, 19);
2230
2231 /* serialize with start_khugepaged() */
2232 mutex_lock(&khugepaged_mutex);
2233
2234 for (;;) {
2235 mutex_unlock(&khugepaged_mutex);
2236 BUG_ON(khugepaged_thread != current);
2237 khugepaged_loop();
2238 BUG_ON(khugepaged_thread != current);
2239
2240 mutex_lock(&khugepaged_mutex);
2241 if (!khugepaged_enabled())
2242 break;
878aee7d
AA		 2243 if (unlikely(kthread_should_stop()))
2244 break;
ba76149f
AA		 2245 }
2246
2247 spin_lock(&khugepaged_mm_lock);
2248 mm_slot = khugepaged_scan.mm_slot;
2249 khugepaged_scan.mm_slot = NULL;
2250 if (mm_slot)
2251 collect_mm_slot(mm_slot);
2252 spin_unlock(&khugepaged_mm_lock);
2253
2254 khugepaged_thread = NULL;
2255 mutex_unlock(&khugepaged_mutex);
2256
2257 return 0;
2258}
2259
71e3aac0
AA		 2260void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2261{
2262 struct page *page;
2263
2264 spin_lock(&mm->page_table_lock);
2265 if (unlikely(!pmd_trans_huge(*pmd))) {
2266 spin_unlock(&mm->page_table_lock);
2267 return;
2268 }
2269 page = pmd_page(*pmd);
2270 VM_BUG_ON(!page_count(page));
2271 get_page(page);
2272 spin_unlock(&mm->page_table_lock);
2273
2274 split_huge_page(page);
2275
2276 put_page(page);
2277 BUG_ON(pmd_trans_huge(*pmd));
2278}
94fcc585
AA		 2279
2280static void split_huge_page_address(struct mm_struct *mm,
2281 unsigned long address)
2282{
2283 pgd_t *pgd;
2284 pud_t *pud;
2285 pmd_t *pmd;
2286
2287 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2288
2289 pgd = pgd_offset(mm, address);
2290 if (!pgd_present(*pgd))
2291 return;
2292
2293 pud = pud_offset(pgd, address);
2294 if (!pud_present(*pud))
2295 return;
2296
2297 pmd = pmd_offset(pud, address);
2298 if (!pmd_present(*pmd))
2299 return;
2300 /*
2301 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2302 * materialize from under us.
2303 */
2304 split_huge_page_pmd(mm, pmd);
2305}
2306
2307void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2308 unsigned long start,
2309 unsigned long end,
2310 long adjust_next)
2311{
2312 /*
2313 * If the new start address isn't hpage aligned and it could
2314 * previously contain an hugepage: check if we need to split
2315 * an huge pmd.
2316 */
2317 if (start & ~HPAGE_PMD_MASK &&
2318 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2319 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2320 split_huge_page_address(vma->vm_mm, start);
2321
2322 /*
2323 * If the new end address isn't hpage aligned and it could
2324 * previously contain an hugepage: check if we need to split
2325 * an huge pmd.
2326 */
2327 if (end & ~HPAGE_PMD_MASK &&
2328 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2329 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2330 split_huge_page_address(vma->vm_mm, end);
2331
2332 /*
2333 * If we're also updating the vma->vm_next->vm_start, if the new
2334 * vm_next->vm_start isn't page aligned and it could previously
2335 * contain an hugepage: check if we need to split an huge pmd.
2336 */
2337 if (adjust_next > 0) {
2338 struct vm_area_struct *next = vma->vm_next;
2339 unsigned long nstart = next->vm_start;
2340 nstart += adjust_next << PAGE_SHIFT;
2341 if (nstart & ~HPAGE_PMD_MASK &&
2342 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2343 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2344 split_huge_page_address(next->vm_mm, nstart);
2345 }
2346}
Linux 6.x block layer and io_uring tree(s)