xref: /openbmc/linux/mm/huge_memory.c (revision 05bcf503)
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>
15 #include <linux/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
20 #include <linux/pagemap.h>
21 #include <asm/tlb.h>
22 #include <asm/pgalloc.h>
23 #include "internal.h"
24 
25 /*
26  * By default transparent hugepage support is enabled for all mappings
27  * and khugepaged scans all mappings. Defrag is only invoked by
28  * khugepaged hugepage allocations and by page faults inside
29  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
30  * allocations.
31  */
32 unsigned long transparent_hugepage_flags __read_mostly =
33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
34 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
35 #endif
36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
37 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
38 #endif
39 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
40 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
41 
42 /* default scan 8*512 pte (or vmas) every 30 second */
43 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
44 static unsigned int khugepaged_pages_collapsed;
45 static unsigned int khugepaged_full_scans;
46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
47 /* during fragmentation poll the hugepage allocator once every minute */
48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
49 static struct task_struct *khugepaged_thread __read_mostly;
50 static DEFINE_MUTEX(khugepaged_mutex);
51 static DEFINE_SPINLOCK(khugepaged_mm_lock);
52 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
53 /*
54  * default collapse hugepages if there is at least one pte mapped like
55  * it would have happened if the vma was large enough during page
56  * fault.
57  */
58 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
59 
60 static int khugepaged(void *none);
61 static int mm_slots_hash_init(void);
62 static int khugepaged_slab_init(void);
63 static void khugepaged_slab_free(void);
64 
65 #define MM_SLOTS_HASH_HEADS 1024
66 static struct hlist_head *mm_slots_hash __read_mostly;
67 static struct kmem_cache *mm_slot_cache __read_mostly;
68 
69 /**
70  * struct mm_slot - hash lookup from mm to mm_slot
71  * @hash: hash collision list
72  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
73  * @mm: the mm that this information is valid for
74  */
75 struct mm_slot {
76 	struct hlist_node hash;
77 	struct list_head mm_node;
78 	struct mm_struct *mm;
79 };
80 
81 /**
82  * struct khugepaged_scan - cursor for scanning
83  * @mm_head: the head of the mm list to scan
84  * @mm_slot: the current mm_slot we are scanning
85  * @address: the next address inside that to be scanned
86  *
87  * There is only the one khugepaged_scan instance of this cursor structure.
88  */
89 struct khugepaged_scan {
90 	struct list_head mm_head;
91 	struct mm_slot *mm_slot;
92 	unsigned long address;
93 };
94 static struct khugepaged_scan khugepaged_scan = {
95 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
96 };
97 
98 
99 static int set_recommended_min_free_kbytes(void)
100 {
101 	struct zone *zone;
102 	int nr_zones = 0;
103 	unsigned long recommended_min;
104 	extern int min_free_kbytes;
105 
106 	if (!khugepaged_enabled())
107 		return 0;
108 
109 	for_each_populated_zone(zone)
110 		nr_zones++;
111 
112 	/* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
113 	recommended_min = pageblock_nr_pages * nr_zones * 2;
114 
115 	/*
116 	 * Make sure that on average at least two pageblocks are almost free
117 	 * of another type, one for a migratetype to fall back to and a
118 	 * second to avoid subsequent fallbacks of other types There are 3
119 	 * MIGRATE_TYPES we care about.
120 	 */
121 	recommended_min += pageblock_nr_pages * nr_zones *
122 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
123 
124 	/* don't ever allow to reserve more than 5% of the lowmem */
125 	recommended_min = min(recommended_min,
126 			      (unsigned long) nr_free_buffer_pages() / 20);
127 	recommended_min <<= (PAGE_SHIFT-10);
128 
129 	if (recommended_min > min_free_kbytes)
130 		min_free_kbytes = recommended_min;
131 	setup_per_zone_wmarks();
132 	return 0;
133 }
134 late_initcall(set_recommended_min_free_kbytes);
135 
136 static int start_khugepaged(void)
137 {
138 	int err = 0;
139 	if (khugepaged_enabled()) {
140 		if (!khugepaged_thread)
141 			khugepaged_thread = kthread_run(khugepaged, NULL,
142 							"khugepaged");
143 		if (unlikely(IS_ERR(khugepaged_thread))) {
144 			printk(KERN_ERR
145 			       "khugepaged: kthread_run(khugepaged) failed\n");
146 			err = PTR_ERR(khugepaged_thread);
147 			khugepaged_thread = NULL;
148 		}
149 
150 		if (!list_empty(&khugepaged_scan.mm_head))
151 			wake_up_interruptible(&khugepaged_wait);
152 
153 		set_recommended_min_free_kbytes();
154 	} else if (khugepaged_thread) {
155 		kthread_stop(khugepaged_thread);
156 		khugepaged_thread = NULL;
157 	}
158 
159 	return err;
160 }
161 
162 #ifdef CONFIG_SYSFS
163 
164 static ssize_t double_flag_show(struct kobject *kobj,
165 				struct kobj_attribute *attr, char *buf,
166 				enum transparent_hugepage_flag enabled,
167 				enum transparent_hugepage_flag req_madv)
168 {
169 	if (test_bit(enabled, &transparent_hugepage_flags)) {
170 		VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
171 		return sprintf(buf, "[always] madvise never\n");
172 	} else if (test_bit(req_madv, &transparent_hugepage_flags))
173 		return sprintf(buf, "always [madvise] never\n");
174 	else
175 		return sprintf(buf, "always madvise [never]\n");
176 }
177 static ssize_t double_flag_store(struct kobject *kobj,
178 				 struct kobj_attribute *attr,
179 				 const char *buf, size_t count,
180 				 enum transparent_hugepage_flag enabled,
181 				 enum transparent_hugepage_flag req_madv)
182 {
183 	if (!memcmp("always", buf,
184 		    min(sizeof("always")-1, count))) {
185 		set_bit(enabled, &transparent_hugepage_flags);
186 		clear_bit(req_madv, &transparent_hugepage_flags);
187 	} else if (!memcmp("madvise", buf,
188 			   min(sizeof("madvise")-1, count))) {
189 		clear_bit(enabled, &transparent_hugepage_flags);
190 		set_bit(req_madv, &transparent_hugepage_flags);
191 	} else if (!memcmp("never", buf,
192 			   min(sizeof("never")-1, count))) {
193 		clear_bit(enabled, &transparent_hugepage_flags);
194 		clear_bit(req_madv, &transparent_hugepage_flags);
195 	} else
196 		return -EINVAL;
197 
198 	return count;
199 }
200 
201 static ssize_t enabled_show(struct kobject *kobj,
202 			    struct kobj_attribute *attr, char *buf)
203 {
204 	return double_flag_show(kobj, attr, buf,
205 				TRANSPARENT_HUGEPAGE_FLAG,
206 				TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
207 }
208 static ssize_t enabled_store(struct kobject *kobj,
209 			     struct kobj_attribute *attr,
210 			     const char *buf, size_t count)
211 {
212 	ssize_t ret;
213 
214 	ret = double_flag_store(kobj, attr, buf, count,
215 				TRANSPARENT_HUGEPAGE_FLAG,
216 				TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
217 
218 	if (ret > 0) {
219 		int err;
220 
221 		mutex_lock(&khugepaged_mutex);
222 		err = start_khugepaged();
223 		mutex_unlock(&khugepaged_mutex);
224 
225 		if (err)
226 			ret = err;
227 	}
228 
229 	return ret;
230 }
231 static struct kobj_attribute enabled_attr =
232 	__ATTR(enabled, 0644, enabled_show, enabled_store);
233 
234 static ssize_t single_flag_show(struct kobject *kobj,
235 				struct kobj_attribute *attr, char *buf,
236 				enum transparent_hugepage_flag flag)
237 {
238 	return sprintf(buf, "%d\n",
239 		       !!test_bit(flag, &transparent_hugepage_flags));
240 }
241 
242 static ssize_t single_flag_store(struct kobject *kobj,
243 				 struct kobj_attribute *attr,
244 				 const char *buf, size_t count,
245 				 enum transparent_hugepage_flag flag)
246 {
247 	unsigned long value;
248 	int ret;
249 
250 	ret = kstrtoul(buf, 10, &value);
251 	if (ret < 0)
252 		return ret;
253 	if (value > 1)
254 		return -EINVAL;
255 
256 	if (value)
257 		set_bit(flag, &transparent_hugepage_flags);
258 	else
259 		clear_bit(flag, &transparent_hugepage_flags);
260 
261 	return count;
262 }
263 
264 /*
265  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
266  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
267  * memory just to allocate one more hugepage.
268  */
269 static ssize_t defrag_show(struct kobject *kobj,
270 			   struct kobj_attribute *attr, char *buf)
271 {
272 	return double_flag_show(kobj, attr, buf,
273 				TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
274 				TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
275 }
276 static ssize_t defrag_store(struct kobject *kobj,
277 			    struct kobj_attribute *attr,
278 			    const char *buf, size_t count)
279 {
280 	return double_flag_store(kobj, attr, buf, count,
281 				 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
282 				 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
283 }
284 static struct kobj_attribute defrag_attr =
285 	__ATTR(defrag, 0644, defrag_show, defrag_store);
286 
287 #ifdef CONFIG_DEBUG_VM
288 static ssize_t debug_cow_show(struct kobject *kobj,
289 				struct kobj_attribute *attr, char *buf)
290 {
291 	return single_flag_show(kobj, attr, buf,
292 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
293 }
294 static ssize_t debug_cow_store(struct kobject *kobj,
295 			       struct kobj_attribute *attr,
296 			       const char *buf, size_t count)
297 {
298 	return single_flag_store(kobj, attr, buf, count,
299 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
300 }
301 static struct kobj_attribute debug_cow_attr =
302 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
303 #endif /* CONFIG_DEBUG_VM */
304 
305 static struct attribute *hugepage_attr[] = {
306 	&enabled_attr.attr,
307 	&defrag_attr.attr,
308 #ifdef CONFIG_DEBUG_VM
309 	&debug_cow_attr.attr,
310 #endif
311 	NULL,
312 };
313 
314 static struct attribute_group hugepage_attr_group = {
315 	.attrs = hugepage_attr,
316 };
317 
318 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
319 					 struct kobj_attribute *attr,
320 					 char *buf)
321 {
322 	return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
323 }
324 
325 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
326 					  struct kobj_attribute *attr,
327 					  const char *buf, size_t count)
328 {
329 	unsigned long msecs;
330 	int err;
331 
332 	err = strict_strtoul(buf, 10, &msecs);
333 	if (err || msecs > UINT_MAX)
334 		return -EINVAL;
335 
336 	khugepaged_scan_sleep_millisecs = msecs;
337 	wake_up_interruptible(&khugepaged_wait);
338 
339 	return count;
340 }
341 static struct kobj_attribute scan_sleep_millisecs_attr =
342 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
343 	       scan_sleep_millisecs_store);
344 
345 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
346 					  struct kobj_attribute *attr,
347 					  char *buf)
348 {
349 	return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
350 }
351 
352 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
353 					   struct kobj_attribute *attr,
354 					   const char *buf, size_t count)
355 {
356 	unsigned long msecs;
357 	int err;
358 
359 	err = strict_strtoul(buf, 10, &msecs);
360 	if (err || msecs > UINT_MAX)
361 		return -EINVAL;
362 
363 	khugepaged_alloc_sleep_millisecs = msecs;
364 	wake_up_interruptible(&khugepaged_wait);
365 
366 	return count;
367 }
368 static struct kobj_attribute alloc_sleep_millisecs_attr =
369 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
370 	       alloc_sleep_millisecs_store);
371 
372 static ssize_t pages_to_scan_show(struct kobject *kobj,
373 				  struct kobj_attribute *attr,
374 				  char *buf)
375 {
376 	return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
377 }
378 static ssize_t pages_to_scan_store(struct kobject *kobj,
379 				   struct kobj_attribute *attr,
380 				   const char *buf, size_t count)
381 {
382 	int err;
383 	unsigned long pages;
384 
385 	err = strict_strtoul(buf, 10, &pages);
386 	if (err || !pages || pages > UINT_MAX)
387 		return -EINVAL;
388 
389 	khugepaged_pages_to_scan = pages;
390 
391 	return count;
392 }
393 static struct kobj_attribute pages_to_scan_attr =
394 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
395 	       pages_to_scan_store);
396 
397 static ssize_t pages_collapsed_show(struct kobject *kobj,
398 				    struct kobj_attribute *attr,
399 				    char *buf)
400 {
401 	return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
402 }
403 static struct kobj_attribute pages_collapsed_attr =
404 	__ATTR_RO(pages_collapsed);
405 
406 static ssize_t full_scans_show(struct kobject *kobj,
407 			       struct kobj_attribute *attr,
408 			       char *buf)
409 {
410 	return sprintf(buf, "%u\n", khugepaged_full_scans);
411 }
412 static struct kobj_attribute full_scans_attr =
413 	__ATTR_RO(full_scans);
414 
415 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
416 				      struct kobj_attribute *attr, char *buf)
417 {
418 	return single_flag_show(kobj, attr, buf,
419 				TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
420 }
421 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
422 				       struct kobj_attribute *attr,
423 				       const char *buf, size_t count)
424 {
425 	return single_flag_store(kobj, attr, buf, count,
426 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
427 }
428 static struct kobj_attribute khugepaged_defrag_attr =
429 	__ATTR(defrag, 0644, khugepaged_defrag_show,
430 	       khugepaged_defrag_store);
431 
432 /*
433  * max_ptes_none controls if khugepaged should collapse hugepages over
434  * any unmapped ptes in turn potentially increasing the memory
435  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
436  * reduce the available free memory in the system as it
437  * runs. Increasing max_ptes_none will instead potentially reduce the
438  * free memory in the system during the khugepaged scan.
439  */
440 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
441 					     struct kobj_attribute *attr,
442 					     char *buf)
443 {
444 	return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
445 }
446 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
447 					      struct kobj_attribute *attr,
448 					      const char *buf, size_t count)
449 {
450 	int err;
451 	unsigned long max_ptes_none;
452 
453 	err = strict_strtoul(buf, 10, &max_ptes_none);
454 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
455 		return -EINVAL;
456 
457 	khugepaged_max_ptes_none = max_ptes_none;
458 
459 	return count;
460 }
461 static struct kobj_attribute khugepaged_max_ptes_none_attr =
462 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
463 	       khugepaged_max_ptes_none_store);
464 
465 static struct attribute *khugepaged_attr[] = {
466 	&khugepaged_defrag_attr.attr,
467 	&khugepaged_max_ptes_none_attr.attr,
468 	&pages_to_scan_attr.attr,
469 	&pages_collapsed_attr.attr,
470 	&full_scans_attr.attr,
471 	&scan_sleep_millisecs_attr.attr,
472 	&alloc_sleep_millisecs_attr.attr,
473 	NULL,
474 };
475 
476 static struct attribute_group khugepaged_attr_group = {
477 	.attrs = khugepaged_attr,
478 	.name = "khugepaged",
479 };
480 
481 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
482 {
483 	int err;
484 
485 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
486 	if (unlikely(!*hugepage_kobj)) {
487 		printk(KERN_ERR "hugepage: failed kobject create\n");
488 		return -ENOMEM;
489 	}
490 
491 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
492 	if (err) {
493 		printk(KERN_ERR "hugepage: failed register hugeage group\n");
494 		goto delete_obj;
495 	}
496 
497 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
498 	if (err) {
499 		printk(KERN_ERR "hugepage: failed register hugeage group\n");
500 		goto remove_hp_group;
501 	}
502 
503 	return 0;
504 
505 remove_hp_group:
506 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
507 delete_obj:
508 	kobject_put(*hugepage_kobj);
509 	return err;
510 }
511 
512 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
513 {
514 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
515 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
516 	kobject_put(hugepage_kobj);
517 }
518 #else
519 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
520 {
521 	return 0;
522 }
523 
524 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
525 {
526 }
527 #endif /* CONFIG_SYSFS */
528 
529 static int __init hugepage_init(void)
530 {
531 	int err;
532 	struct kobject *hugepage_kobj;
533 
534 	if (!has_transparent_hugepage()) {
535 		transparent_hugepage_flags = 0;
536 		return -EINVAL;
537 	}
538 
539 	err = hugepage_init_sysfs(&hugepage_kobj);
540 	if (err)
541 		return err;
542 
543 	err = khugepaged_slab_init();
544 	if (err)
545 		goto out;
546 
547 	err = mm_slots_hash_init();
548 	if (err) {
549 		khugepaged_slab_free();
550 		goto out;
551 	}
552 
553 	/*
554 	 * By default disable transparent hugepages on smaller systems,
555 	 * where the extra memory used could hurt more than TLB overhead
556 	 * is likely to save.  The admin can still enable it through /sys.
557 	 */
558 	if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
559 		transparent_hugepage_flags = 0;
560 
561 	start_khugepaged();
562 
563 	return 0;
564 out:
565 	hugepage_exit_sysfs(hugepage_kobj);
566 	return err;
567 }
568 module_init(hugepage_init)
569 
570 static int __init setup_transparent_hugepage(char *str)
571 {
572 	int ret = 0;
573 	if (!str)
574 		goto out;
575 	if (!strcmp(str, "always")) {
576 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
577 			&transparent_hugepage_flags);
578 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
579 			  &transparent_hugepage_flags);
580 		ret = 1;
581 	} else if (!strcmp(str, "madvise")) {
582 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
583 			  &transparent_hugepage_flags);
584 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
585 			&transparent_hugepage_flags);
586 		ret = 1;
587 	} else if (!strcmp(str, "never")) {
588 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
589 			  &transparent_hugepage_flags);
590 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
591 			  &transparent_hugepage_flags);
592 		ret = 1;
593 	}
594 out:
595 	if (!ret)
596 		printk(KERN_WARNING
597 		       "transparent_hugepage= cannot parse, ignored\n");
598 	return ret;
599 }
600 __setup("transparent_hugepage=", setup_transparent_hugepage);
601 
602 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
603 {
604 	if (likely(vma->vm_flags & VM_WRITE))
605 		pmd = pmd_mkwrite(pmd);
606 	return pmd;
607 }
608 
609 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
610 					struct vm_area_struct *vma,
611 					unsigned long haddr, pmd_t *pmd,
612 					struct page *page)
613 {
614 	pgtable_t pgtable;
615 
616 	VM_BUG_ON(!PageCompound(page));
617 	pgtable = pte_alloc_one(mm, haddr);
618 	if (unlikely(!pgtable))
619 		return VM_FAULT_OOM;
620 
621 	clear_huge_page(page, haddr, HPAGE_PMD_NR);
622 	__SetPageUptodate(page);
623 
624 	spin_lock(&mm->page_table_lock);
625 	if (unlikely(!pmd_none(*pmd))) {
626 		spin_unlock(&mm->page_table_lock);
627 		mem_cgroup_uncharge_page(page);
628 		put_page(page);
629 		pte_free(mm, pgtable);
630 	} else {
631 		pmd_t entry;
632 		entry = mk_pmd(page, vma->vm_page_prot);
633 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
634 		entry = pmd_mkhuge(entry);
635 		/*
636 		 * The spinlocking to take the lru_lock inside
637 		 * page_add_new_anon_rmap() acts as a full memory
638 		 * barrier to be sure clear_huge_page writes become
639 		 * visible after the set_pmd_at() write.
640 		 */
641 		page_add_new_anon_rmap(page, vma, haddr);
642 		set_pmd_at(mm, haddr, pmd, entry);
643 		pgtable_trans_huge_deposit(mm, pgtable);
644 		add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
645 		mm->nr_ptes++;
646 		spin_unlock(&mm->page_table_lock);
647 	}
648 
649 	return 0;
650 }
651 
652 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
653 {
654 	return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
655 }
656 
657 static inline struct page *alloc_hugepage_vma(int defrag,
658 					      struct vm_area_struct *vma,
659 					      unsigned long haddr, int nd,
660 					      gfp_t extra_gfp)
661 {
662 	return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
663 			       HPAGE_PMD_ORDER, vma, haddr, nd);
664 }
665 
666 #ifndef CONFIG_NUMA
667 static inline struct page *alloc_hugepage(int defrag)
668 {
669 	return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
670 			   HPAGE_PMD_ORDER);
671 }
672 #endif
673 
674 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
675 			       unsigned long address, pmd_t *pmd,
676 			       unsigned int flags)
677 {
678 	struct page *page;
679 	unsigned long haddr = address & HPAGE_PMD_MASK;
680 	pte_t *pte;
681 
682 	if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
683 		if (unlikely(anon_vma_prepare(vma)))
684 			return VM_FAULT_OOM;
685 		if (unlikely(khugepaged_enter(vma)))
686 			return VM_FAULT_OOM;
687 		page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
688 					  vma, haddr, numa_node_id(), 0);
689 		if (unlikely(!page)) {
690 			count_vm_event(THP_FAULT_FALLBACK);
691 			goto out;
692 		}
693 		count_vm_event(THP_FAULT_ALLOC);
694 		if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
695 			put_page(page);
696 			goto out;
697 		}
698 		if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
699 							  page))) {
700 			mem_cgroup_uncharge_page(page);
701 			put_page(page);
702 			goto out;
703 		}
704 
705 		return 0;
706 	}
707 out:
708 	/*
709 	 * Use __pte_alloc instead of pte_alloc_map, because we can't
710 	 * run pte_offset_map on the pmd, if an huge pmd could
711 	 * materialize from under us from a different thread.
712 	 */
713 	if (unlikely(__pte_alloc(mm, vma, pmd, address)))
714 		return VM_FAULT_OOM;
715 	/* if an huge pmd materialized from under us just retry later */
716 	if (unlikely(pmd_trans_huge(*pmd)))
717 		return 0;
718 	/*
719 	 * A regular pmd is established and it can't morph into a huge pmd
720 	 * from under us anymore at this point because we hold the mmap_sem
721 	 * read mode and khugepaged takes it in write mode. So now it's
722 	 * safe to run pte_offset_map().
723 	 */
724 	pte = pte_offset_map(pmd, address);
725 	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
726 }
727 
728 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
729 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
730 		  struct vm_area_struct *vma)
731 {
732 	struct page *src_page;
733 	pmd_t pmd;
734 	pgtable_t pgtable;
735 	int ret;
736 
737 	ret = -ENOMEM;
738 	pgtable = pte_alloc_one(dst_mm, addr);
739 	if (unlikely(!pgtable))
740 		goto out;
741 
742 	spin_lock(&dst_mm->page_table_lock);
743 	spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
744 
745 	ret = -EAGAIN;
746 	pmd = *src_pmd;
747 	if (unlikely(!pmd_trans_huge(pmd))) {
748 		pte_free(dst_mm, pgtable);
749 		goto out_unlock;
750 	}
751 	if (unlikely(pmd_trans_splitting(pmd))) {
752 		/* split huge page running from under us */
753 		spin_unlock(&src_mm->page_table_lock);
754 		spin_unlock(&dst_mm->page_table_lock);
755 		pte_free(dst_mm, pgtable);
756 
757 		wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
758 		goto out;
759 	}
760 	src_page = pmd_page(pmd);
761 	VM_BUG_ON(!PageHead(src_page));
762 	get_page(src_page);
763 	page_dup_rmap(src_page);
764 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
765 
766 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
767 	pmd = pmd_mkold(pmd_wrprotect(pmd));
768 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
769 	pgtable_trans_huge_deposit(dst_mm, pgtable);
770 	dst_mm->nr_ptes++;
771 
772 	ret = 0;
773 out_unlock:
774 	spin_unlock(&src_mm->page_table_lock);
775 	spin_unlock(&dst_mm->page_table_lock);
776 out:
777 	return ret;
778 }
779 
780 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
781 					struct vm_area_struct *vma,
782 					unsigned long address,
783 					pmd_t *pmd, pmd_t orig_pmd,
784 					struct page *page,
785 					unsigned long haddr)
786 {
787 	pgtable_t pgtable;
788 	pmd_t _pmd;
789 	int ret = 0, i;
790 	struct page **pages;
791 	unsigned long mmun_start;	/* For mmu_notifiers */
792 	unsigned long mmun_end;		/* For mmu_notifiers */
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_node(GFP_HIGHUSER_MOVABLE |
803 					       __GFP_OTHER_NODE,
804 					       vma, address, page_to_nid(page));
805 		if (unlikely(!pages[i] ||
806 			     mem_cgroup_newpage_charge(pages[i], mm,
807 						       GFP_KERNEL))) {
808 			if (pages[i])
809 				put_page(pages[i]);
810 			mem_cgroup_uncharge_start();
811 			while (--i >= 0) {
812 				mem_cgroup_uncharge_page(pages[i]);
813 				put_page(pages[i]);
814 			}
815 			mem_cgroup_uncharge_end();
816 			kfree(pages);
817 			ret |= VM_FAULT_OOM;
818 			goto out;
819 		}
820 	}
821 
822 	for (i = 0; i < HPAGE_PMD_NR; i++) {
823 		copy_user_highpage(pages[i], page + i,
824 				   haddr + PAGE_SIZE * i, vma);
825 		__SetPageUptodate(pages[i]);
826 		cond_resched();
827 	}
828 
829 	mmun_start = haddr;
830 	mmun_end   = haddr + HPAGE_PMD_SIZE;
831 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
832 
833 	spin_lock(&mm->page_table_lock);
834 	if (unlikely(!pmd_same(*pmd, orig_pmd)))
835 		goto out_free_pages;
836 	VM_BUG_ON(!PageHead(page));
837 
838 	pmdp_clear_flush(vma, haddr, pmd);
839 	/* leave pmd empty until pte is filled */
840 
841 	pgtable = pgtable_trans_huge_withdraw(mm);
842 	pmd_populate(mm, &_pmd, pgtable);
843 
844 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
845 		pte_t *pte, entry;
846 		entry = mk_pte(pages[i], vma->vm_page_prot);
847 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
848 		page_add_new_anon_rmap(pages[i], vma, haddr);
849 		pte = pte_offset_map(&_pmd, haddr);
850 		VM_BUG_ON(!pte_none(*pte));
851 		set_pte_at(mm, haddr, pte, entry);
852 		pte_unmap(pte);
853 	}
854 	kfree(pages);
855 
856 	smp_wmb(); /* make pte visible before pmd */
857 	pmd_populate(mm, pmd, pgtable);
858 	page_remove_rmap(page);
859 	spin_unlock(&mm->page_table_lock);
860 
861 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
862 
863 	ret |= VM_FAULT_WRITE;
864 	put_page(page);
865 
866 out:
867 	return ret;
868 
869 out_free_pages:
870 	spin_unlock(&mm->page_table_lock);
871 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
872 	mem_cgroup_uncharge_start();
873 	for (i = 0; i < HPAGE_PMD_NR; i++) {
874 		mem_cgroup_uncharge_page(pages[i]);
875 		put_page(pages[i]);
876 	}
877 	mem_cgroup_uncharge_end();
878 	kfree(pages);
879 	goto out;
880 }
881 
882 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
883 			unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
884 {
885 	int ret = 0;
886 	struct page *page, *new_page;
887 	unsigned long haddr;
888 	unsigned long mmun_start;	/* For mmu_notifiers */
889 	unsigned long mmun_end;		/* For mmu_notifiers */
890 
891 	VM_BUG_ON(!vma->anon_vma);
892 	spin_lock(&mm->page_table_lock);
893 	if (unlikely(!pmd_same(*pmd, orig_pmd)))
894 		goto out_unlock;
895 
896 	page = pmd_page(orig_pmd);
897 	VM_BUG_ON(!PageCompound(page) || !PageHead(page));
898 	haddr = address & HPAGE_PMD_MASK;
899 	if (page_mapcount(page) == 1) {
900 		pmd_t entry;
901 		entry = pmd_mkyoung(orig_pmd);
902 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
903 		if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
904 			update_mmu_cache_pmd(vma, address, pmd);
905 		ret |= VM_FAULT_WRITE;
906 		goto out_unlock;
907 	}
908 	get_page(page);
909 	spin_unlock(&mm->page_table_lock);
910 
911 	if (transparent_hugepage_enabled(vma) &&
912 	    !transparent_hugepage_debug_cow())
913 		new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
914 					      vma, haddr, numa_node_id(), 0);
915 	else
916 		new_page = NULL;
917 
918 	if (unlikely(!new_page)) {
919 		count_vm_event(THP_FAULT_FALLBACK);
920 		ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
921 						   pmd, orig_pmd, page, haddr);
922 		if (ret & VM_FAULT_OOM)
923 			split_huge_page(page);
924 		put_page(page);
925 		goto out;
926 	}
927 	count_vm_event(THP_FAULT_ALLOC);
928 
929 	if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
930 		put_page(new_page);
931 		split_huge_page(page);
932 		put_page(page);
933 		ret |= VM_FAULT_OOM;
934 		goto out;
935 	}
936 
937 	copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
938 	__SetPageUptodate(new_page);
939 
940 	mmun_start = haddr;
941 	mmun_end   = haddr + HPAGE_PMD_SIZE;
942 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
943 
944 	spin_lock(&mm->page_table_lock);
945 	put_page(page);
946 	if (unlikely(!pmd_same(*pmd, orig_pmd))) {
947 		spin_unlock(&mm->page_table_lock);
948 		mem_cgroup_uncharge_page(new_page);
949 		put_page(new_page);
950 		goto out_mn;
951 	} else {
952 		pmd_t entry;
953 		VM_BUG_ON(!PageHead(page));
954 		entry = mk_pmd(new_page, vma->vm_page_prot);
955 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
956 		entry = pmd_mkhuge(entry);
957 		pmdp_clear_flush(vma, haddr, pmd);
958 		page_add_new_anon_rmap(new_page, vma, haddr);
959 		set_pmd_at(mm, haddr, pmd, entry);
960 		update_mmu_cache_pmd(vma, address, pmd);
961 		page_remove_rmap(page);
962 		put_page(page);
963 		ret |= VM_FAULT_WRITE;
964 	}
965 	spin_unlock(&mm->page_table_lock);
966 out_mn:
967 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
968 out:
969 	return ret;
970 out_unlock:
971 	spin_unlock(&mm->page_table_lock);
972 	return ret;
973 }
974 
975 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
976 				   unsigned long addr,
977 				   pmd_t *pmd,
978 				   unsigned int flags)
979 {
980 	struct mm_struct *mm = vma->vm_mm;
981 	struct page *page = NULL;
982 
983 	assert_spin_locked(&mm->page_table_lock);
984 
985 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
986 		goto out;
987 
988 	page = pmd_page(*pmd);
989 	VM_BUG_ON(!PageHead(page));
990 	if (flags & FOLL_TOUCH) {
991 		pmd_t _pmd;
992 		/*
993 		 * We should set the dirty bit only for FOLL_WRITE but
994 		 * for now the dirty bit in the pmd is meaningless.
995 		 * And if the dirty bit will become meaningful and
996 		 * we'll only set it with FOLL_WRITE, an atomic
997 		 * set_bit will be required on the pmd to set the
998 		 * young bit, instead of the current set_pmd_at.
999 		 */
1000 		_pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1001 		set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1002 	}
1003 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1004 		if (page->mapping && trylock_page(page)) {
1005 			lru_add_drain();
1006 			if (page->mapping)
1007 				mlock_vma_page(page);
1008 			unlock_page(page);
1009 		}
1010 	}
1011 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1012 	VM_BUG_ON(!PageCompound(page));
1013 	if (flags & FOLL_GET)
1014 		get_page_foll(page);
1015 
1016 out:
1017 	return page;
1018 }
1019 
1020 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1021 		 pmd_t *pmd, unsigned long addr)
1022 {
1023 	int ret = 0;
1024 
1025 	if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1026 		struct page *page;
1027 		pgtable_t pgtable;
1028 		pmd_t orig_pmd;
1029 		pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1030 		orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1031 		page = pmd_page(orig_pmd);
1032 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1033 		page_remove_rmap(page);
1034 		VM_BUG_ON(page_mapcount(page) < 0);
1035 		add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1036 		VM_BUG_ON(!PageHead(page));
1037 		tlb->mm->nr_ptes--;
1038 		spin_unlock(&tlb->mm->page_table_lock);
1039 		tlb_remove_page(tlb, page);
1040 		pte_free(tlb->mm, pgtable);
1041 		ret = 1;
1042 	}
1043 	return ret;
1044 }
1045 
1046 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1047 		unsigned long addr, unsigned long end,
1048 		unsigned char *vec)
1049 {
1050 	int ret = 0;
1051 
1052 	if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1053 		/*
1054 		 * All logical pages in the range are present
1055 		 * if backed by a huge page.
1056 		 */
1057 		spin_unlock(&vma->vm_mm->page_table_lock);
1058 		memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1059 		ret = 1;
1060 	}
1061 
1062 	return ret;
1063 }
1064 
1065 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1066 		  unsigned long old_addr,
1067 		  unsigned long new_addr, unsigned long old_end,
1068 		  pmd_t *old_pmd, pmd_t *new_pmd)
1069 {
1070 	int ret = 0;
1071 	pmd_t pmd;
1072 
1073 	struct mm_struct *mm = vma->vm_mm;
1074 
1075 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1076 	    (new_addr & ~HPAGE_PMD_MASK) ||
1077 	    old_end - old_addr < HPAGE_PMD_SIZE ||
1078 	    (new_vma->vm_flags & VM_NOHUGEPAGE))
1079 		goto out;
1080 
1081 	/*
1082 	 * The destination pmd shouldn't be established, free_pgtables()
1083 	 * should have release it.
1084 	 */
1085 	if (WARN_ON(!pmd_none(*new_pmd))) {
1086 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1087 		goto out;
1088 	}
1089 
1090 	ret = __pmd_trans_huge_lock(old_pmd, vma);
1091 	if (ret == 1) {
1092 		pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1093 		VM_BUG_ON(!pmd_none(*new_pmd));
1094 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1095 		spin_unlock(&mm->page_table_lock);
1096 	}
1097 out:
1098 	return ret;
1099 }
1100 
1101 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1102 		unsigned long addr, pgprot_t newprot)
1103 {
1104 	struct mm_struct *mm = vma->vm_mm;
1105 	int ret = 0;
1106 
1107 	if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1108 		pmd_t entry;
1109 		entry = pmdp_get_and_clear(mm, addr, pmd);
1110 		entry = pmd_modify(entry, newprot);
1111 		set_pmd_at(mm, addr, pmd, entry);
1112 		spin_unlock(&vma->vm_mm->page_table_lock);
1113 		ret = 1;
1114 	}
1115 
1116 	return ret;
1117 }
1118 
1119 /*
1120  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1121  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1122  *
1123  * Note that if it returns 1, this routine returns without unlocking page
1124  * table locks. So callers must unlock them.
1125  */
1126 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1127 {
1128 	spin_lock(&vma->vm_mm->page_table_lock);
1129 	if (likely(pmd_trans_huge(*pmd))) {
1130 		if (unlikely(pmd_trans_splitting(*pmd))) {
1131 			spin_unlock(&vma->vm_mm->page_table_lock);
1132 			wait_split_huge_page(vma->anon_vma, pmd);
1133 			return -1;
1134 		} else {
1135 			/* Thp mapped by 'pmd' is stable, so we can
1136 			 * handle it as it is. */
1137 			return 1;
1138 		}
1139 	}
1140 	spin_unlock(&vma->vm_mm->page_table_lock);
1141 	return 0;
1142 }
1143 
1144 pmd_t *page_check_address_pmd(struct page *page,
1145 			      struct mm_struct *mm,
1146 			      unsigned long address,
1147 			      enum page_check_address_pmd_flag flag)
1148 {
1149 	pgd_t *pgd;
1150 	pud_t *pud;
1151 	pmd_t *pmd, *ret = NULL;
1152 
1153 	if (address & ~HPAGE_PMD_MASK)
1154 		goto out;
1155 
1156 	pgd = pgd_offset(mm, address);
1157 	if (!pgd_present(*pgd))
1158 		goto out;
1159 
1160 	pud = pud_offset(pgd, address);
1161 	if (!pud_present(*pud))
1162 		goto out;
1163 
1164 	pmd = pmd_offset(pud, address);
1165 	if (pmd_none(*pmd))
1166 		goto out;
1167 	if (pmd_page(*pmd) != page)
1168 		goto out;
1169 	/*
1170 	 * split_vma() may create temporary aliased mappings. There is
1171 	 * no risk as long as all huge pmd are found and have their
1172 	 * splitting bit set before __split_huge_page_refcount
1173 	 * runs. Finding the same huge pmd more than once during the
1174 	 * same rmap walk is not a problem.
1175 	 */
1176 	if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1177 	    pmd_trans_splitting(*pmd))
1178 		goto out;
1179 	if (pmd_trans_huge(*pmd)) {
1180 		VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1181 			  !pmd_trans_splitting(*pmd));
1182 		ret = pmd;
1183 	}
1184 out:
1185 	return ret;
1186 }
1187 
1188 static int __split_huge_page_splitting(struct page *page,
1189 				       struct vm_area_struct *vma,
1190 				       unsigned long address)
1191 {
1192 	struct mm_struct *mm = vma->vm_mm;
1193 	pmd_t *pmd;
1194 	int ret = 0;
1195 	/* For mmu_notifiers */
1196 	const unsigned long mmun_start = address;
1197 	const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1198 
1199 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1200 	spin_lock(&mm->page_table_lock);
1201 	pmd = page_check_address_pmd(page, mm, address,
1202 				     PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1203 	if (pmd) {
1204 		/*
1205 		 * We can't temporarily set the pmd to null in order
1206 		 * to split it, the pmd must remain marked huge at all
1207 		 * times or the VM won't take the pmd_trans_huge paths
1208 		 * and it won't wait on the anon_vma->root->mutex to
1209 		 * serialize against split_huge_page*.
1210 		 */
1211 		pmdp_splitting_flush(vma, address, pmd);
1212 		ret = 1;
1213 	}
1214 	spin_unlock(&mm->page_table_lock);
1215 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1216 
1217 	return ret;
1218 }
1219 
1220 static void __split_huge_page_refcount(struct page *page)
1221 {
1222 	int i;
1223 	struct zone *zone = page_zone(page);
1224 	struct lruvec *lruvec;
1225 	int tail_count = 0;
1226 
1227 	/* prevent PageLRU to go away from under us, and freeze lru stats */
1228 	spin_lock_irq(&zone->lru_lock);
1229 	lruvec = mem_cgroup_page_lruvec(page, zone);
1230 
1231 	compound_lock(page);
1232 	/* complete memcg works before add pages to LRU */
1233 	mem_cgroup_split_huge_fixup(page);
1234 
1235 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1236 		struct page *page_tail = page + i;
1237 
1238 		/* tail_page->_mapcount cannot change */
1239 		BUG_ON(page_mapcount(page_tail) < 0);
1240 		tail_count += page_mapcount(page_tail);
1241 		/* check for overflow */
1242 		BUG_ON(tail_count < 0);
1243 		BUG_ON(atomic_read(&page_tail->_count) != 0);
1244 		/*
1245 		 * tail_page->_count is zero and not changing from
1246 		 * under us. But get_page_unless_zero() may be running
1247 		 * from under us on the tail_page. If we used
1248 		 * atomic_set() below instead of atomic_add(), we
1249 		 * would then run atomic_set() concurrently with
1250 		 * get_page_unless_zero(), and atomic_set() is
1251 		 * implemented in C not using locked ops. spin_unlock
1252 		 * on x86 sometime uses locked ops because of PPro
1253 		 * errata 66, 92, so unless somebody can guarantee
1254 		 * atomic_set() here would be safe on all archs (and
1255 		 * not only on x86), it's safer to use atomic_add().
1256 		 */
1257 		atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1258 			   &page_tail->_count);
1259 
1260 		/* after clearing PageTail the gup refcount can be released */
1261 		smp_mb();
1262 
1263 		/*
1264 		 * retain hwpoison flag of the poisoned tail page:
1265 		 *   fix for the unsuitable process killed on Guest Machine(KVM)
1266 		 *   by the memory-failure.
1267 		 */
1268 		page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1269 		page_tail->flags |= (page->flags &
1270 				     ((1L << PG_referenced) |
1271 				      (1L << PG_swapbacked) |
1272 				      (1L << PG_mlocked) |
1273 				      (1L << PG_uptodate)));
1274 		page_tail->flags |= (1L << PG_dirty);
1275 
1276 		/* clear PageTail before overwriting first_page */
1277 		smp_wmb();
1278 
1279 		/*
1280 		 * __split_huge_page_splitting() already set the
1281 		 * splitting bit in all pmd that could map this
1282 		 * hugepage, that will ensure no CPU can alter the
1283 		 * mapcount on the head page. The mapcount is only
1284 		 * accounted in the head page and it has to be
1285 		 * transferred to all tail pages in the below code. So
1286 		 * for this code to be safe, the split the mapcount
1287 		 * can't change. But that doesn't mean userland can't
1288 		 * keep changing and reading the page contents while
1289 		 * we transfer the mapcount, so the pmd splitting
1290 		 * status is achieved setting a reserved bit in the
1291 		 * pmd, not by clearing the present bit.
1292 		*/
1293 		page_tail->_mapcount = page->_mapcount;
1294 
1295 		BUG_ON(page_tail->mapping);
1296 		page_tail->mapping = page->mapping;
1297 
1298 		page_tail->index = page->index + i;
1299 
1300 		BUG_ON(!PageAnon(page_tail));
1301 		BUG_ON(!PageUptodate(page_tail));
1302 		BUG_ON(!PageDirty(page_tail));
1303 		BUG_ON(!PageSwapBacked(page_tail));
1304 
1305 		lru_add_page_tail(page, page_tail, lruvec);
1306 	}
1307 	atomic_sub(tail_count, &page->_count);
1308 	BUG_ON(atomic_read(&page->_count) <= 0);
1309 
1310 	__mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1311 	__mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1312 
1313 	ClearPageCompound(page);
1314 	compound_unlock(page);
1315 	spin_unlock_irq(&zone->lru_lock);
1316 
1317 	for (i = 1; i < HPAGE_PMD_NR; i++) {
1318 		struct page *page_tail = page + i;
1319 		BUG_ON(page_count(page_tail) <= 0);
1320 		/*
1321 		 * Tail pages may be freed if there wasn't any mapping
1322 		 * like if add_to_swap() is running on a lru page that
1323 		 * had its mapping zapped. And freeing these pages
1324 		 * requires taking the lru_lock so we do the put_page
1325 		 * of the tail pages after the split is complete.
1326 		 */
1327 		put_page(page_tail);
1328 	}
1329 
1330 	/*
1331 	 * Only the head page (now become a regular page) is required
1332 	 * to be pinned by the caller.
1333 	 */
1334 	BUG_ON(page_count(page) <= 0);
1335 }
1336 
1337 static int __split_huge_page_map(struct page *page,
1338 				 struct vm_area_struct *vma,
1339 				 unsigned long address)
1340 {
1341 	struct mm_struct *mm = vma->vm_mm;
1342 	pmd_t *pmd, _pmd;
1343 	int ret = 0, i;
1344 	pgtable_t pgtable;
1345 	unsigned long haddr;
1346 
1347 	spin_lock(&mm->page_table_lock);
1348 	pmd = page_check_address_pmd(page, mm, address,
1349 				     PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1350 	if (pmd) {
1351 		pgtable = pgtable_trans_huge_withdraw(mm);
1352 		pmd_populate(mm, &_pmd, pgtable);
1353 
1354 		haddr = address;
1355 		for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1356 			pte_t *pte, entry;
1357 			BUG_ON(PageCompound(page+i));
1358 			entry = mk_pte(page + i, vma->vm_page_prot);
1359 			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1360 			if (!pmd_write(*pmd))
1361 				entry = pte_wrprotect(entry);
1362 			else
1363 				BUG_ON(page_mapcount(page) != 1);
1364 			if (!pmd_young(*pmd))
1365 				entry = pte_mkold(entry);
1366 			pte = pte_offset_map(&_pmd, haddr);
1367 			BUG_ON(!pte_none(*pte));
1368 			set_pte_at(mm, haddr, pte, entry);
1369 			pte_unmap(pte);
1370 		}
1371 
1372 		smp_wmb(); /* make pte visible before pmd */
1373 		/*
1374 		 * Up to this point the pmd is present and huge and
1375 		 * userland has the whole access to the hugepage
1376 		 * during the split (which happens in place). If we
1377 		 * overwrite the pmd with the not-huge version
1378 		 * pointing to the pte here (which of course we could
1379 		 * if all CPUs were bug free), userland could trigger
1380 		 * a small page size TLB miss on the small sized TLB
1381 		 * while the hugepage TLB entry is still established
1382 		 * in the huge TLB. Some CPU doesn't like that. See
1383 		 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1384 		 * Erratum 383 on page 93. Intel should be safe but is
1385 		 * also warns that it's only safe if the permission
1386 		 * and cache attributes of the two entries loaded in
1387 		 * the two TLB is identical (which should be the case
1388 		 * here). But it is generally safer to never allow
1389 		 * small and huge TLB entries for the same virtual
1390 		 * address to be loaded simultaneously. So instead of
1391 		 * doing "pmd_populate(); flush_tlb_range();" we first
1392 		 * mark the current pmd notpresent (atomically because
1393 		 * here the pmd_trans_huge and pmd_trans_splitting
1394 		 * must remain set at all times on the pmd until the
1395 		 * split is complete for this pmd), then we flush the
1396 		 * SMP TLB and finally we write the non-huge version
1397 		 * of the pmd entry with pmd_populate.
1398 		 */
1399 		pmdp_invalidate(vma, address, pmd);
1400 		pmd_populate(mm, pmd, pgtable);
1401 		ret = 1;
1402 	}
1403 	spin_unlock(&mm->page_table_lock);
1404 
1405 	return ret;
1406 }
1407 
1408 /* must be called with anon_vma->root->mutex hold */
1409 static void __split_huge_page(struct page *page,
1410 			      struct anon_vma *anon_vma)
1411 {
1412 	int mapcount, mapcount2;
1413 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1414 	struct anon_vma_chain *avc;
1415 
1416 	BUG_ON(!PageHead(page));
1417 	BUG_ON(PageTail(page));
1418 
1419 	mapcount = 0;
1420 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1421 		struct vm_area_struct *vma = avc->vma;
1422 		unsigned long addr = vma_address(page, vma);
1423 		BUG_ON(is_vma_temporary_stack(vma));
1424 		mapcount += __split_huge_page_splitting(page, vma, addr);
1425 	}
1426 	/*
1427 	 * It is critical that new vmas are added to the tail of the
1428 	 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1429 	 * and establishes a child pmd before
1430 	 * __split_huge_page_splitting() freezes the parent pmd (so if
1431 	 * we fail to prevent copy_huge_pmd() from running until the
1432 	 * whole __split_huge_page() is complete), we will still see
1433 	 * the newly established pmd of the child later during the
1434 	 * walk, to be able to set it as pmd_trans_splitting too.
1435 	 */
1436 	if (mapcount != page_mapcount(page))
1437 		printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1438 		       mapcount, page_mapcount(page));
1439 	BUG_ON(mapcount != page_mapcount(page));
1440 
1441 	__split_huge_page_refcount(page);
1442 
1443 	mapcount2 = 0;
1444 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1445 		struct vm_area_struct *vma = avc->vma;
1446 		unsigned long addr = vma_address(page, vma);
1447 		BUG_ON(is_vma_temporary_stack(vma));
1448 		mapcount2 += __split_huge_page_map(page, vma, addr);
1449 	}
1450 	if (mapcount != mapcount2)
1451 		printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1452 		       mapcount, mapcount2, page_mapcount(page));
1453 	BUG_ON(mapcount != mapcount2);
1454 }
1455 
1456 int split_huge_page(struct page *page)
1457 {
1458 	struct anon_vma *anon_vma;
1459 	int ret = 1;
1460 
1461 	BUG_ON(!PageAnon(page));
1462 	anon_vma = page_lock_anon_vma(page);
1463 	if (!anon_vma)
1464 		goto out;
1465 	ret = 0;
1466 	if (!PageCompound(page))
1467 		goto out_unlock;
1468 
1469 	BUG_ON(!PageSwapBacked(page));
1470 	__split_huge_page(page, anon_vma);
1471 	count_vm_event(THP_SPLIT);
1472 
1473 	BUG_ON(PageCompound(page));
1474 out_unlock:
1475 	page_unlock_anon_vma(anon_vma);
1476 out:
1477 	return ret;
1478 }
1479 
1480 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1481 
1482 int hugepage_madvise(struct vm_area_struct *vma,
1483 		     unsigned long *vm_flags, int advice)
1484 {
1485 	struct mm_struct *mm = vma->vm_mm;
1486 
1487 	switch (advice) {
1488 	case MADV_HUGEPAGE:
1489 		/*
1490 		 * Be somewhat over-protective like KSM for now!
1491 		 */
1492 		if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1493 			return -EINVAL;
1494 		if (mm->def_flags & VM_NOHUGEPAGE)
1495 			return -EINVAL;
1496 		*vm_flags &= ~VM_NOHUGEPAGE;
1497 		*vm_flags |= VM_HUGEPAGE;
1498 		/*
1499 		 * If the vma become good for khugepaged to scan,
1500 		 * register it here without waiting a page fault that
1501 		 * may not happen any time soon.
1502 		 */
1503 		if (unlikely(khugepaged_enter_vma_merge(vma)))
1504 			return -ENOMEM;
1505 		break;
1506 	case MADV_NOHUGEPAGE:
1507 		/*
1508 		 * Be somewhat over-protective like KSM for now!
1509 		 */
1510 		if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1511 			return -EINVAL;
1512 		*vm_flags &= ~VM_HUGEPAGE;
1513 		*vm_flags |= VM_NOHUGEPAGE;
1514 		/*
1515 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1516 		 * this vma even if we leave the mm registered in khugepaged if
1517 		 * it got registered before VM_NOHUGEPAGE was set.
1518 		 */
1519 		break;
1520 	}
1521 
1522 	return 0;
1523 }
1524 
1525 static int __init khugepaged_slab_init(void)
1526 {
1527 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1528 					  sizeof(struct mm_slot),
1529 					  __alignof__(struct mm_slot), 0, NULL);
1530 	if (!mm_slot_cache)
1531 		return -ENOMEM;
1532 
1533 	return 0;
1534 }
1535 
1536 static void __init khugepaged_slab_free(void)
1537 {
1538 	kmem_cache_destroy(mm_slot_cache);
1539 	mm_slot_cache = NULL;
1540 }
1541 
1542 static inline struct mm_slot *alloc_mm_slot(void)
1543 {
1544 	if (!mm_slot_cache)	/* initialization failed */
1545 		return NULL;
1546 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1547 }
1548 
1549 static inline void free_mm_slot(struct mm_slot *mm_slot)
1550 {
1551 	kmem_cache_free(mm_slot_cache, mm_slot);
1552 }
1553 
1554 static int __init mm_slots_hash_init(void)
1555 {
1556 	mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1557 				GFP_KERNEL);
1558 	if (!mm_slots_hash)
1559 		return -ENOMEM;
1560 	return 0;
1561 }
1562 
1563 #if 0
1564 static void __init mm_slots_hash_free(void)
1565 {
1566 	kfree(mm_slots_hash);
1567 	mm_slots_hash = NULL;
1568 }
1569 #endif
1570 
1571 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1572 {
1573 	struct mm_slot *mm_slot;
1574 	struct hlist_head *bucket;
1575 	struct hlist_node *node;
1576 
1577 	bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1578 				% MM_SLOTS_HASH_HEADS];
1579 	hlist_for_each_entry(mm_slot, node, bucket, hash) {
1580 		if (mm == mm_slot->mm)
1581 			return mm_slot;
1582 	}
1583 	return NULL;
1584 }
1585 
1586 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1587 				    struct mm_slot *mm_slot)
1588 {
1589 	struct hlist_head *bucket;
1590 
1591 	bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1592 				% MM_SLOTS_HASH_HEADS];
1593 	mm_slot->mm = mm;
1594 	hlist_add_head(&mm_slot->hash, bucket);
1595 }
1596 
1597 static inline int khugepaged_test_exit(struct mm_struct *mm)
1598 {
1599 	return atomic_read(&mm->mm_users) == 0;
1600 }
1601 
1602 int __khugepaged_enter(struct mm_struct *mm)
1603 {
1604 	struct mm_slot *mm_slot;
1605 	int wakeup;
1606 
1607 	mm_slot = alloc_mm_slot();
1608 	if (!mm_slot)
1609 		return -ENOMEM;
1610 
1611 	/* __khugepaged_exit() must not run from under us */
1612 	VM_BUG_ON(khugepaged_test_exit(mm));
1613 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1614 		free_mm_slot(mm_slot);
1615 		return 0;
1616 	}
1617 
1618 	spin_lock(&khugepaged_mm_lock);
1619 	insert_to_mm_slots_hash(mm, mm_slot);
1620 	/*
1621 	 * Insert just behind the scanning cursor, to let the area settle
1622 	 * down a little.
1623 	 */
1624 	wakeup = list_empty(&khugepaged_scan.mm_head);
1625 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1626 	spin_unlock(&khugepaged_mm_lock);
1627 
1628 	atomic_inc(&mm->mm_count);
1629 	if (wakeup)
1630 		wake_up_interruptible(&khugepaged_wait);
1631 
1632 	return 0;
1633 }
1634 
1635 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1636 {
1637 	unsigned long hstart, hend;
1638 	if (!vma->anon_vma)
1639 		/*
1640 		 * Not yet faulted in so we will register later in the
1641 		 * page fault if needed.
1642 		 */
1643 		return 0;
1644 	if (vma->vm_ops)
1645 		/* khugepaged not yet working on file or special mappings */
1646 		return 0;
1647 	VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1648 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1649 	hend = vma->vm_end & HPAGE_PMD_MASK;
1650 	if (hstart < hend)
1651 		return khugepaged_enter(vma);
1652 	return 0;
1653 }
1654 
1655 void __khugepaged_exit(struct mm_struct *mm)
1656 {
1657 	struct mm_slot *mm_slot;
1658 	int free = 0;
1659 
1660 	spin_lock(&khugepaged_mm_lock);
1661 	mm_slot = get_mm_slot(mm);
1662 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1663 		hlist_del(&mm_slot->hash);
1664 		list_del(&mm_slot->mm_node);
1665 		free = 1;
1666 	}
1667 	spin_unlock(&khugepaged_mm_lock);
1668 
1669 	if (free) {
1670 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1671 		free_mm_slot(mm_slot);
1672 		mmdrop(mm);
1673 	} else if (mm_slot) {
1674 		/*
1675 		 * This is required to serialize against
1676 		 * khugepaged_test_exit() (which is guaranteed to run
1677 		 * under mmap sem read mode). Stop here (after we
1678 		 * return all pagetables will be destroyed) until
1679 		 * khugepaged has finished working on the pagetables
1680 		 * under the mmap_sem.
1681 		 */
1682 		down_write(&mm->mmap_sem);
1683 		up_write(&mm->mmap_sem);
1684 	}
1685 }
1686 
1687 static void release_pte_page(struct page *page)
1688 {
1689 	/* 0 stands for page_is_file_cache(page) == false */
1690 	dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1691 	unlock_page(page);
1692 	putback_lru_page(page);
1693 }
1694 
1695 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1696 {
1697 	while (--_pte >= pte) {
1698 		pte_t pteval = *_pte;
1699 		if (!pte_none(pteval))
1700 			release_pte_page(pte_page(pteval));
1701 	}
1702 }
1703 
1704 static void release_all_pte_pages(pte_t *pte)
1705 {
1706 	release_pte_pages(pte, pte + HPAGE_PMD_NR);
1707 }
1708 
1709 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1710 					unsigned long address,
1711 					pte_t *pte)
1712 {
1713 	struct page *page;
1714 	pte_t *_pte;
1715 	int referenced = 0, isolated = 0, none = 0;
1716 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1717 	     _pte++, address += PAGE_SIZE) {
1718 		pte_t pteval = *_pte;
1719 		if (pte_none(pteval)) {
1720 			if (++none <= khugepaged_max_ptes_none)
1721 				continue;
1722 			else {
1723 				release_pte_pages(pte, _pte);
1724 				goto out;
1725 			}
1726 		}
1727 		if (!pte_present(pteval) || !pte_write(pteval)) {
1728 			release_pte_pages(pte, _pte);
1729 			goto out;
1730 		}
1731 		page = vm_normal_page(vma, address, pteval);
1732 		if (unlikely(!page)) {
1733 			release_pte_pages(pte, _pte);
1734 			goto out;
1735 		}
1736 		VM_BUG_ON(PageCompound(page));
1737 		BUG_ON(!PageAnon(page));
1738 		VM_BUG_ON(!PageSwapBacked(page));
1739 
1740 		/* cannot use mapcount: can't collapse if there's a gup pin */
1741 		if (page_count(page) != 1) {
1742 			release_pte_pages(pte, _pte);
1743 			goto out;
1744 		}
1745 		/*
1746 		 * We can do it before isolate_lru_page because the
1747 		 * page can't be freed from under us. NOTE: PG_lock
1748 		 * is needed to serialize against split_huge_page
1749 		 * when invoked from the VM.
1750 		 */
1751 		if (!trylock_page(page)) {
1752 			release_pte_pages(pte, _pte);
1753 			goto out;
1754 		}
1755 		/*
1756 		 * Isolate the page to avoid collapsing an hugepage
1757 		 * currently in use by the VM.
1758 		 */
1759 		if (isolate_lru_page(page)) {
1760 			unlock_page(page);
1761 			release_pte_pages(pte, _pte);
1762 			goto out;
1763 		}
1764 		/* 0 stands for page_is_file_cache(page) == false */
1765 		inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1766 		VM_BUG_ON(!PageLocked(page));
1767 		VM_BUG_ON(PageLRU(page));
1768 
1769 		/* If there is no mapped pte young don't collapse the page */
1770 		if (pte_young(pteval) || PageReferenced(page) ||
1771 		    mmu_notifier_test_young(vma->vm_mm, address))
1772 			referenced = 1;
1773 	}
1774 	if (unlikely(!referenced))
1775 		release_all_pte_pages(pte);
1776 	else
1777 		isolated = 1;
1778 out:
1779 	return isolated;
1780 }
1781 
1782 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1783 				      struct vm_area_struct *vma,
1784 				      unsigned long address,
1785 				      spinlock_t *ptl)
1786 {
1787 	pte_t *_pte;
1788 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1789 		pte_t pteval = *_pte;
1790 		struct page *src_page;
1791 
1792 		if (pte_none(pteval)) {
1793 			clear_user_highpage(page, address);
1794 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1795 		} else {
1796 			src_page = pte_page(pteval);
1797 			copy_user_highpage(page, src_page, address, vma);
1798 			VM_BUG_ON(page_mapcount(src_page) != 1);
1799 			release_pte_page(src_page);
1800 			/*
1801 			 * ptl mostly unnecessary, but preempt has to
1802 			 * be disabled to update the per-cpu stats
1803 			 * inside page_remove_rmap().
1804 			 */
1805 			spin_lock(ptl);
1806 			/*
1807 			 * paravirt calls inside pte_clear here are
1808 			 * superfluous.
1809 			 */
1810 			pte_clear(vma->vm_mm, address, _pte);
1811 			page_remove_rmap(src_page);
1812 			spin_unlock(ptl);
1813 			free_page_and_swap_cache(src_page);
1814 		}
1815 
1816 		address += PAGE_SIZE;
1817 		page++;
1818 	}
1819 }
1820 
1821 static void khugepaged_alloc_sleep(void)
1822 {
1823 	wait_event_freezable_timeout(khugepaged_wait, false,
1824 			msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1825 }
1826 
1827 #ifdef CONFIG_NUMA
1828 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1829 {
1830 	if (IS_ERR(*hpage)) {
1831 		if (!*wait)
1832 			return false;
1833 
1834 		*wait = false;
1835 		*hpage = NULL;
1836 		khugepaged_alloc_sleep();
1837 	} else if (*hpage) {
1838 		put_page(*hpage);
1839 		*hpage = NULL;
1840 	}
1841 
1842 	return true;
1843 }
1844 
1845 static struct page
1846 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1847 		       struct vm_area_struct *vma, unsigned long address,
1848 		       int node)
1849 {
1850 	VM_BUG_ON(*hpage);
1851 	/*
1852 	 * Allocate the page while the vma is still valid and under
1853 	 * the mmap_sem read mode so there is no memory allocation
1854 	 * later when we take the mmap_sem in write mode. This is more
1855 	 * friendly behavior (OTOH it may actually hide bugs) to
1856 	 * filesystems in userland with daemons allocating memory in
1857 	 * the userland I/O paths.  Allocating memory with the
1858 	 * mmap_sem in read mode is good idea also to allow greater
1859 	 * scalability.
1860 	 */
1861 	*hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1862 				      node, __GFP_OTHER_NODE);
1863 
1864 	/*
1865 	 * After allocating the hugepage, release the mmap_sem read lock in
1866 	 * preparation for taking it in write mode.
1867 	 */
1868 	up_read(&mm->mmap_sem);
1869 	if (unlikely(!*hpage)) {
1870 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1871 		*hpage = ERR_PTR(-ENOMEM);
1872 		return NULL;
1873 	}
1874 
1875 	count_vm_event(THP_COLLAPSE_ALLOC);
1876 	return *hpage;
1877 }
1878 #else
1879 static struct page *khugepaged_alloc_hugepage(bool *wait)
1880 {
1881 	struct page *hpage;
1882 
1883 	do {
1884 		hpage = alloc_hugepage(khugepaged_defrag());
1885 		if (!hpage) {
1886 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1887 			if (!*wait)
1888 				return NULL;
1889 
1890 			*wait = false;
1891 			khugepaged_alloc_sleep();
1892 		} else
1893 			count_vm_event(THP_COLLAPSE_ALLOC);
1894 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
1895 
1896 	return hpage;
1897 }
1898 
1899 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1900 {
1901 	if (!*hpage)
1902 		*hpage = khugepaged_alloc_hugepage(wait);
1903 
1904 	if (unlikely(!*hpage))
1905 		return false;
1906 
1907 	return true;
1908 }
1909 
1910 static struct page
1911 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1912 		       struct vm_area_struct *vma, unsigned long address,
1913 		       int node)
1914 {
1915 	up_read(&mm->mmap_sem);
1916 	VM_BUG_ON(!*hpage);
1917 	return  *hpage;
1918 }
1919 #endif
1920 
1921 static void collapse_huge_page(struct mm_struct *mm,
1922 				   unsigned long address,
1923 				   struct page **hpage,
1924 				   struct vm_area_struct *vma,
1925 				   int node)
1926 {
1927 	pgd_t *pgd;
1928 	pud_t *pud;
1929 	pmd_t *pmd, _pmd;
1930 	pte_t *pte;
1931 	pgtable_t pgtable;
1932 	struct page *new_page;
1933 	spinlock_t *ptl;
1934 	int isolated;
1935 	unsigned long hstart, hend;
1936 	unsigned long mmun_start;	/* For mmu_notifiers */
1937 	unsigned long mmun_end;		/* For mmu_notifiers */
1938 
1939 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1940 
1941 	/* release the mmap_sem read lock. */
1942 	new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1943 	if (!new_page)
1944 		return;
1945 
1946 	if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1947 		return;
1948 
1949 	/*
1950 	 * Prevent all access to pagetables with the exception of
1951 	 * gup_fast later hanlded by the ptep_clear_flush and the VM
1952 	 * handled by the anon_vma lock + PG_lock.
1953 	 */
1954 	down_write(&mm->mmap_sem);
1955 	if (unlikely(khugepaged_test_exit(mm)))
1956 		goto out;
1957 
1958 	vma = find_vma(mm, address);
1959 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1960 	hend = vma->vm_end & HPAGE_PMD_MASK;
1961 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1962 		goto out;
1963 
1964 	if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1965 	    (vma->vm_flags & VM_NOHUGEPAGE))
1966 		goto out;
1967 
1968 	if (!vma->anon_vma || vma->vm_ops)
1969 		goto out;
1970 	if (is_vma_temporary_stack(vma))
1971 		goto out;
1972 	VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1973 
1974 	pgd = pgd_offset(mm, address);
1975 	if (!pgd_present(*pgd))
1976 		goto out;
1977 
1978 	pud = pud_offset(pgd, address);
1979 	if (!pud_present(*pud))
1980 		goto out;
1981 
1982 	pmd = pmd_offset(pud, address);
1983 	/* pmd can't go away or become huge under us */
1984 	if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1985 		goto out;
1986 
1987 	anon_vma_lock(vma->anon_vma);
1988 
1989 	pte = pte_offset_map(pmd, address);
1990 	ptl = pte_lockptr(mm, pmd);
1991 
1992 	mmun_start = address;
1993 	mmun_end   = address + HPAGE_PMD_SIZE;
1994 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1995 	spin_lock(&mm->page_table_lock); /* probably unnecessary */
1996 	/*
1997 	 * After this gup_fast can't run anymore. This also removes
1998 	 * any huge TLB entry from the CPU so we won't allow
1999 	 * huge and small TLB entries for the same virtual address
2000 	 * to avoid the risk of CPU bugs in that area.
2001 	 */
2002 	_pmd = pmdp_clear_flush(vma, address, pmd);
2003 	spin_unlock(&mm->page_table_lock);
2004 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2005 
2006 	spin_lock(ptl);
2007 	isolated = __collapse_huge_page_isolate(vma, address, pte);
2008 	spin_unlock(ptl);
2009 
2010 	if (unlikely(!isolated)) {
2011 		pte_unmap(pte);
2012 		spin_lock(&mm->page_table_lock);
2013 		BUG_ON(!pmd_none(*pmd));
2014 		set_pmd_at(mm, address, pmd, _pmd);
2015 		spin_unlock(&mm->page_table_lock);
2016 		anon_vma_unlock(vma->anon_vma);
2017 		goto out;
2018 	}
2019 
2020 	/*
2021 	 * All pages are isolated and locked so anon_vma rmap
2022 	 * can't run anymore.
2023 	 */
2024 	anon_vma_unlock(vma->anon_vma);
2025 
2026 	__collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2027 	pte_unmap(pte);
2028 	__SetPageUptodate(new_page);
2029 	pgtable = pmd_pgtable(_pmd);
2030 
2031 	_pmd = mk_pmd(new_page, vma->vm_page_prot);
2032 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2033 	_pmd = pmd_mkhuge(_pmd);
2034 
2035 	/*
2036 	 * spin_lock() below is not the equivalent of smp_wmb(), so
2037 	 * this is needed to avoid the copy_huge_page writes to become
2038 	 * visible after the set_pmd_at() write.
2039 	 */
2040 	smp_wmb();
2041 
2042 	spin_lock(&mm->page_table_lock);
2043 	BUG_ON(!pmd_none(*pmd));
2044 	page_add_new_anon_rmap(new_page, vma, address);
2045 	set_pmd_at(mm, address, pmd, _pmd);
2046 	update_mmu_cache_pmd(vma, address, pmd);
2047 	pgtable_trans_huge_deposit(mm, pgtable);
2048 	spin_unlock(&mm->page_table_lock);
2049 
2050 	*hpage = NULL;
2051 
2052 	khugepaged_pages_collapsed++;
2053 out_up_write:
2054 	up_write(&mm->mmap_sem);
2055 	return;
2056 
2057 out:
2058 	mem_cgroup_uncharge_page(new_page);
2059 	goto out_up_write;
2060 }
2061 
2062 static int khugepaged_scan_pmd(struct mm_struct *mm,
2063 			       struct vm_area_struct *vma,
2064 			       unsigned long address,
2065 			       struct page **hpage)
2066 {
2067 	pgd_t *pgd;
2068 	pud_t *pud;
2069 	pmd_t *pmd;
2070 	pte_t *pte, *_pte;
2071 	int ret = 0, referenced = 0, none = 0;
2072 	struct page *page;
2073 	unsigned long _address;
2074 	spinlock_t *ptl;
2075 	int node = -1;
2076 
2077 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2078 
2079 	pgd = pgd_offset(mm, address);
2080 	if (!pgd_present(*pgd))
2081 		goto out;
2082 
2083 	pud = pud_offset(pgd, address);
2084 	if (!pud_present(*pud))
2085 		goto out;
2086 
2087 	pmd = pmd_offset(pud, address);
2088 	if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2089 		goto out;
2090 
2091 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2092 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2093 	     _pte++, _address += PAGE_SIZE) {
2094 		pte_t pteval = *_pte;
2095 		if (pte_none(pteval)) {
2096 			if (++none <= khugepaged_max_ptes_none)
2097 				continue;
2098 			else
2099 				goto out_unmap;
2100 		}
2101 		if (!pte_present(pteval) || !pte_write(pteval))
2102 			goto out_unmap;
2103 		page = vm_normal_page(vma, _address, pteval);
2104 		if (unlikely(!page))
2105 			goto out_unmap;
2106 		/*
2107 		 * Chose the node of the first page. This could
2108 		 * be more sophisticated and look at more pages,
2109 		 * but isn't for now.
2110 		 */
2111 		if (node == -1)
2112 			node = page_to_nid(page);
2113 		VM_BUG_ON(PageCompound(page));
2114 		if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2115 			goto out_unmap;
2116 		/* cannot use mapcount: can't collapse if there's a gup pin */
2117 		if (page_count(page) != 1)
2118 			goto out_unmap;
2119 		if (pte_young(pteval) || PageReferenced(page) ||
2120 		    mmu_notifier_test_young(vma->vm_mm, address))
2121 			referenced = 1;
2122 	}
2123 	if (referenced)
2124 		ret = 1;
2125 out_unmap:
2126 	pte_unmap_unlock(pte, ptl);
2127 	if (ret)
2128 		/* collapse_huge_page will return with the mmap_sem released */
2129 		collapse_huge_page(mm, address, hpage, vma, node);
2130 out:
2131 	return ret;
2132 }
2133 
2134 static void collect_mm_slot(struct mm_slot *mm_slot)
2135 {
2136 	struct mm_struct *mm = mm_slot->mm;
2137 
2138 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2139 
2140 	if (khugepaged_test_exit(mm)) {
2141 		/* free mm_slot */
2142 		hlist_del(&mm_slot->hash);
2143 		list_del(&mm_slot->mm_node);
2144 
2145 		/*
2146 		 * Not strictly needed because the mm exited already.
2147 		 *
2148 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2149 		 */
2150 
2151 		/* khugepaged_mm_lock actually not necessary for the below */
2152 		free_mm_slot(mm_slot);
2153 		mmdrop(mm);
2154 	}
2155 }
2156 
2157 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2158 					    struct page **hpage)
2159 	__releases(&khugepaged_mm_lock)
2160 	__acquires(&khugepaged_mm_lock)
2161 {
2162 	struct mm_slot *mm_slot;
2163 	struct mm_struct *mm;
2164 	struct vm_area_struct *vma;
2165 	int progress = 0;
2166 
2167 	VM_BUG_ON(!pages);
2168 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2169 
2170 	if (khugepaged_scan.mm_slot)
2171 		mm_slot = khugepaged_scan.mm_slot;
2172 	else {
2173 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
2174 				     struct mm_slot, mm_node);
2175 		khugepaged_scan.address = 0;
2176 		khugepaged_scan.mm_slot = mm_slot;
2177 	}
2178 	spin_unlock(&khugepaged_mm_lock);
2179 
2180 	mm = mm_slot->mm;
2181 	down_read(&mm->mmap_sem);
2182 	if (unlikely(khugepaged_test_exit(mm)))
2183 		vma = NULL;
2184 	else
2185 		vma = find_vma(mm, khugepaged_scan.address);
2186 
2187 	progress++;
2188 	for (; vma; vma = vma->vm_next) {
2189 		unsigned long hstart, hend;
2190 
2191 		cond_resched();
2192 		if (unlikely(khugepaged_test_exit(mm))) {
2193 			progress++;
2194 			break;
2195 		}
2196 
2197 		if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2198 		     !khugepaged_always()) ||
2199 		    (vma->vm_flags & VM_NOHUGEPAGE)) {
2200 		skip:
2201 			progress++;
2202 			continue;
2203 		}
2204 		if (!vma->anon_vma || vma->vm_ops)
2205 			goto skip;
2206 		if (is_vma_temporary_stack(vma))
2207 			goto skip;
2208 		VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2209 
2210 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2211 		hend = vma->vm_end & HPAGE_PMD_MASK;
2212 		if (hstart >= hend)
2213 			goto skip;
2214 		if (khugepaged_scan.address > hend)
2215 			goto skip;
2216 		if (khugepaged_scan.address < hstart)
2217 			khugepaged_scan.address = hstart;
2218 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2219 
2220 		while (khugepaged_scan.address < hend) {
2221 			int ret;
2222 			cond_resched();
2223 			if (unlikely(khugepaged_test_exit(mm)))
2224 				goto breakouterloop;
2225 
2226 			VM_BUG_ON(khugepaged_scan.address < hstart ||
2227 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2228 				  hend);
2229 			ret = khugepaged_scan_pmd(mm, vma,
2230 						  khugepaged_scan.address,
2231 						  hpage);
2232 			/* move to next address */
2233 			khugepaged_scan.address += HPAGE_PMD_SIZE;
2234 			progress += HPAGE_PMD_NR;
2235 			if (ret)
2236 				/* we released mmap_sem so break loop */
2237 				goto breakouterloop_mmap_sem;
2238 			if (progress >= pages)
2239 				goto breakouterloop;
2240 		}
2241 	}
2242 breakouterloop:
2243 	up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2244 breakouterloop_mmap_sem:
2245 
2246 	spin_lock(&khugepaged_mm_lock);
2247 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2248 	/*
2249 	 * Release the current mm_slot if this mm is about to die, or
2250 	 * if we scanned all vmas of this mm.
2251 	 */
2252 	if (khugepaged_test_exit(mm) || !vma) {
2253 		/*
2254 		 * Make sure that if mm_users is reaching zero while
2255 		 * khugepaged runs here, khugepaged_exit will find
2256 		 * mm_slot not pointing to the exiting mm.
2257 		 */
2258 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2259 			khugepaged_scan.mm_slot = list_entry(
2260 				mm_slot->mm_node.next,
2261 				struct mm_slot, mm_node);
2262 			khugepaged_scan.address = 0;
2263 		} else {
2264 			khugepaged_scan.mm_slot = NULL;
2265 			khugepaged_full_scans++;
2266 		}
2267 
2268 		collect_mm_slot(mm_slot);
2269 	}
2270 
2271 	return progress;
2272 }
2273 
2274 static int khugepaged_has_work(void)
2275 {
2276 	return !list_empty(&khugepaged_scan.mm_head) &&
2277 		khugepaged_enabled();
2278 }
2279 
2280 static int khugepaged_wait_event(void)
2281 {
2282 	return !list_empty(&khugepaged_scan.mm_head) ||
2283 		kthread_should_stop();
2284 }
2285 
2286 static void khugepaged_do_scan(void)
2287 {
2288 	struct page *hpage = NULL;
2289 	unsigned int progress = 0, pass_through_head = 0;
2290 	unsigned int pages = khugepaged_pages_to_scan;
2291 	bool wait = true;
2292 
2293 	barrier(); /* write khugepaged_pages_to_scan to local stack */
2294 
2295 	while (progress < pages) {
2296 		if (!khugepaged_prealloc_page(&hpage, &wait))
2297 			break;
2298 
2299 		cond_resched();
2300 
2301 		if (unlikely(kthread_should_stop() || freezing(current)))
2302 			break;
2303 
2304 		spin_lock(&khugepaged_mm_lock);
2305 		if (!khugepaged_scan.mm_slot)
2306 			pass_through_head++;
2307 		if (khugepaged_has_work() &&
2308 		    pass_through_head < 2)
2309 			progress += khugepaged_scan_mm_slot(pages - progress,
2310 							    &hpage);
2311 		else
2312 			progress = pages;
2313 		spin_unlock(&khugepaged_mm_lock);
2314 	}
2315 
2316 	if (!IS_ERR_OR_NULL(hpage))
2317 		put_page(hpage);
2318 }
2319 
2320 static void khugepaged_wait_work(void)
2321 {
2322 	try_to_freeze();
2323 
2324 	if (khugepaged_has_work()) {
2325 		if (!khugepaged_scan_sleep_millisecs)
2326 			return;
2327 
2328 		wait_event_freezable_timeout(khugepaged_wait,
2329 					     kthread_should_stop(),
2330 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2331 		return;
2332 	}
2333 
2334 	if (khugepaged_enabled())
2335 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2336 }
2337 
2338 static int khugepaged(void *none)
2339 {
2340 	struct mm_slot *mm_slot;
2341 
2342 	set_freezable();
2343 	set_user_nice(current, 19);
2344 
2345 	while (!kthread_should_stop()) {
2346 		khugepaged_do_scan();
2347 		khugepaged_wait_work();
2348 	}
2349 
2350 	spin_lock(&khugepaged_mm_lock);
2351 	mm_slot = khugepaged_scan.mm_slot;
2352 	khugepaged_scan.mm_slot = NULL;
2353 	if (mm_slot)
2354 		collect_mm_slot(mm_slot);
2355 	spin_unlock(&khugepaged_mm_lock);
2356 	return 0;
2357 }
2358 
2359 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2360 {
2361 	struct page *page;
2362 
2363 	spin_lock(&mm->page_table_lock);
2364 	if (unlikely(!pmd_trans_huge(*pmd))) {
2365 		spin_unlock(&mm->page_table_lock);
2366 		return;
2367 	}
2368 	page = pmd_page(*pmd);
2369 	VM_BUG_ON(!page_count(page));
2370 	get_page(page);
2371 	spin_unlock(&mm->page_table_lock);
2372 
2373 	split_huge_page(page);
2374 
2375 	put_page(page);
2376 	BUG_ON(pmd_trans_huge(*pmd));
2377 }
2378 
2379 static void split_huge_page_address(struct mm_struct *mm,
2380 				    unsigned long address)
2381 {
2382 	pgd_t *pgd;
2383 	pud_t *pud;
2384 	pmd_t *pmd;
2385 
2386 	VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2387 
2388 	pgd = pgd_offset(mm, address);
2389 	if (!pgd_present(*pgd))
2390 		return;
2391 
2392 	pud = pud_offset(pgd, address);
2393 	if (!pud_present(*pud))
2394 		return;
2395 
2396 	pmd = pmd_offset(pud, address);
2397 	if (!pmd_present(*pmd))
2398 		return;
2399 	/*
2400 	 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2401 	 * materialize from under us.
2402 	 */
2403 	split_huge_page_pmd(mm, pmd);
2404 }
2405 
2406 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2407 			     unsigned long start,
2408 			     unsigned long end,
2409 			     long adjust_next)
2410 {
2411 	/*
2412 	 * If the new start address isn't hpage aligned and it could
2413 	 * previously contain an hugepage: check if we need to split
2414 	 * an huge pmd.
2415 	 */
2416 	if (start & ~HPAGE_PMD_MASK &&
2417 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2418 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2419 		split_huge_page_address(vma->vm_mm, start);
2420 
2421 	/*
2422 	 * If the new end address isn't hpage aligned and it could
2423 	 * previously contain an hugepage: check if we need to split
2424 	 * an huge pmd.
2425 	 */
2426 	if (end & ~HPAGE_PMD_MASK &&
2427 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2428 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2429 		split_huge_page_address(vma->vm_mm, end);
2430 
2431 	/*
2432 	 * If we're also updating the vma->vm_next->vm_start, if the new
2433 	 * vm_next->vm_start isn't page aligned and it could previously
2434 	 * contain an hugepage: check if we need to split an huge pmd.
2435 	 */
2436 	if (adjust_next > 0) {
2437 		struct vm_area_struct *next = vma->vm_next;
2438 		unsigned long nstart = next->vm_start;
2439 		nstart += adjust_next << PAGE_SHIFT;
2440 		if (nstart & ~HPAGE_PMD_MASK &&
2441 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2442 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2443 			split_huge_page_address(next->vm_mm, nstart);
2444 	}
2445 }
2446