path: root/mm/migrate.c (plain)
blob: 24485f5ec498dc48fb40e8d4f23d689cde5da357

1	/*
2	* Memory Migration functionality - linux/mm/migrate.c
3	*
4	* Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5	*
6	* Page migration was first developed in the context of the memory hotplug
7	* project. The main authors of the migration code are:
8	*
9	* IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10	* Hirokazu Takahashi <taka@valinux.co.jp>
11	* Dave Hansen <haveblue@us.ibm.com>
12	* Christoph Lameter
13	*/
14
15	#include <linux/migrate.h>
16	#include <linux/export.h>
17	#include <linux/swap.h>
18	#include <linux/swapops.h>
19	#include <linux/pagemap.h>
20	#include <linux/buffer_head.h>
21	#include <linux/mm_inline.h>
22	#include <linux/nsproxy.h>
23	#include <linux/pagevec.h>
24	#include <linux/ksm.h>
25	#include <linux/rmap.h>
26	#include <linux/topology.h>
27	#include <linux/cpu.h>
28	#include <linux/cpuset.h>
29	#include <linux/writeback.h>
30	#include <linux/mempolicy.h>
31	#include <linux/vmalloc.h>
32	#include <linux/security.h>
33	#include <linux/backing-dev.h>
34	#include <linux/compaction.h>
35	#include <linux/syscalls.h>
36	#include <linux/hugetlb.h>
37	#include <linux/hugetlb_cgroup.h>
38	#include <linux/gfp.h>
39	#include <linux/balloon_compaction.h>
40	#include <linux/mmu_notifier.h>
41	#include <linux/page_idle.h>
42	#include <linux/page_owner.h>
43	#include <linux/ptrace.h>
44	#ifdef CONFIG_AMLOGIC_CMA
45	#include <linux/delay.h>
46	#endif
47
48	#include <asm/tlbflush.h>
49
50	#define CREATE_TRACE_POINTS
51	#include <trace/events/migrate.h>
52
53	#include "internal.h"
54
55	/*
56	* migrate_prep() needs to be called before we start compiling a list of pages
57	* to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
58	* undesirable, use migrate_prep_local()
59	*/
60	int migrate_prep(void)
61	{
62	/*
63	* Clear the LRU lists so pages can be isolated.
64	* Note that pages may be moved off the LRU after we have
65	* drained them. Those pages will fail to migrate like other
66	* pages that may be busy.
67	*/
68	lru_add_drain_all();
69
70	return 0;
71	}
72
73	/* Do the necessary work of migrate_prep but not if it involves other CPUs */
74	int migrate_prep_local(void)
75	{
76	lru_add_drain();
77
78	return 0;
79	}
80
81	bool isolate_movable_page(struct page *page, isolate_mode_t mode)
82	{
83	struct address_space *mapping;
84
85	/*
86	* Avoid burning cycles with pages that are yet under __free_pages(),
87	* or just got freed under us.
88	*
89	* In case we 'win' a race for a movable page being freed under us and
90	* raise its refcount preventing __free_pages() from doing its job
91	* the put_page() at the end of this block will take care of
92	* release this page, thus avoiding a nasty leakage.
93	*/
94	if (unlikely(!get_page_unless_zero(page)))
95	goto out;
96
97	/*
98	* Check PageMovable before holding a PG_lock because page's owner
99	* assumes anybody doesn't touch PG_lock of newly allocated page
100	* so unconditionally grapping the lock ruins page's owner side.
101	*/
102	if (unlikely(!__PageMovable(page)))
103	goto out_putpage;
104	/*
105	* As movable pages are not isolated from LRU lists, concurrent
106	* compaction threads can race against page migration functions
107	* as well as race against the releasing a page.
108	*
109	* In order to avoid having an already isolated movable page
110	* being (wrongly) re-isolated while it is under migration,
111	* or to avoid attempting to isolate pages being released,
112	* lets be sure we have the page lock
113	* before proceeding with the movable page isolation steps.
114	*/
115	if (unlikely(!trylock_page(page)))
116	goto out_putpage;
117
118	if (!PageMovable(page) || PageIsolated(page))
119	goto out_no_isolated;
120
121	mapping = page_mapping(page);
122	VM_BUG_ON_PAGE(!mapping, page);
123
124	if (!mapping->a_ops->isolate_page(page, mode))
125	goto out_no_isolated;
126
127	/* Driver shouldn't use PG_isolated bit of page->flags */
128	WARN_ON_ONCE(PageIsolated(page));
129	__SetPageIsolated(page);
130	unlock_page(page);
131
132	return true;
133
134	out_no_isolated:
135	unlock_page(page);
136	out_putpage:
137	put_page(page);
138	out:
139	return false;
140	}
141
142	/* It should be called on page which is PG_movable */
143	void putback_movable_page(struct page *page)
144	{
145	struct address_space *mapping;
146
147	VM_BUG_ON_PAGE(!PageLocked(page), page);
148	VM_BUG_ON_PAGE(!PageMovable(page), page);
149	VM_BUG_ON_PAGE(!PageIsolated(page), page);
150
151	mapping = page_mapping(page);
152	mapping->a_ops->putback_page(page);
153	__ClearPageIsolated(page);
154	}
155
156	/*
157	* Put previously isolated pages back onto the appropriate lists
158	* from where they were once taken off for compaction/migration.
159	*
160	* This function shall be used whenever the isolated pageset has been
161	* built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
162	* and isolate_huge_page().
163	*/
164	void putback_movable_pages(struct list_head *l)
165	{
166	struct page *page;
167	struct page *page2;
168
169	list_for_each_entry_safe(page, page2, l, lru) {
170	#ifdef CONFIG_AMLOGIC_CMA
171	if (PageCmaAllocating(page)) /* migrate/reclaim failed */
172	ClearPageCmaAllocating(page);
173	#endif
174	if (unlikely(PageHuge(page))) {
175	putback_active_hugepage(page);
176	continue;
177	}
178	list_del(&page->lru);
179	/*
180	* We isolated non-lru movable page so here we can use
181	* __PageMovable because LRU page's mapping cannot have
182	* PAGE_MAPPING_MOVABLE.
183	*/
184	if (unlikely(__PageMovable(page))) {
185	VM_BUG_ON_PAGE(!PageIsolated(page), page);
186	lock_page(page);
187	if (PageMovable(page))
188	putback_movable_page(page);
189	else
190	__ClearPageIsolated(page);
191	unlock_page(page);
192	put_page(page);
193	} else {
194	dec_node_page_state(page, NR_ISOLATED_ANON +
195	page_is_file_cache(page));
196	putback_lru_page(page);
197	}
198	}
199	}
200
201	/*
202	* Restore a potential migration pte to a working pte entry
203	*/
204	static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
205	unsigned long addr, void *old)
206	{
207	struct mm_struct *mm = vma->vm_mm;
208	swp_entry_t entry;
209	pmd_t *pmd;
210	pte_t *ptep, pte;
211	spinlock_t *ptl;
212
213	if (unlikely(PageHuge(new))) {
214	ptep = huge_pte_offset(mm, addr);
215	if (!ptep)
216	goto out;
217	ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
218	} else {
219	pmd = mm_find_pmd(mm, addr);
220	if (!pmd)
221	goto out;
222
223	ptep = pte_offset_map(pmd, addr);
224
225	/*
226	* Peek to check is_swap_pte() before taking ptlock? No, we
227	* can race mremap's move_ptes(), which skips anon_vma lock.
228	*/
229
230	ptl = pte_lockptr(mm, pmd);
231	}
232
233	spin_lock(ptl);
234	pte = *ptep;
235	if (!is_swap_pte(pte))
236	goto unlock;
237
238	entry = pte_to_swp_entry(pte);
239
240	if (!is_migration_entry(entry) ||
241	migration_entry_to_page(entry) != old)
242	goto unlock;
243
244	get_page(new);
245	pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
246	if (pte_swp_soft_dirty(*ptep))
247	pte = pte_mksoft_dirty(pte);
248
249	/* Recheck VMA as permissions can change since migration started */
250	if (is_write_migration_entry(entry))
251	pte = maybe_mkwrite(pte, vma);
252
253	#ifdef CONFIG_HUGETLB_PAGE
254	if (PageHuge(new)) {
255	pte = pte_mkhuge(pte);
256	pte = arch_make_huge_pte(pte, vma, new, 0);
257	}
258	#endif
259	flush_dcache_page(new);
260	set_pte_at(mm, addr, ptep, pte);
261
262	if (PageHuge(new)) {
263	if (PageAnon(new))
264	hugepage_add_anon_rmap(new, vma, addr);
265	else
266	page_dup_rmap(new, true);
267	} else if (PageAnon(new))
268	page_add_anon_rmap(new, vma, addr, false);
269	else
270	page_add_file_rmap(new, false);
271
272	if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
273	mlock_vma_page(new);
274
275	/* No need to invalidate - it was non-present before */
276	update_mmu_cache(vma, addr, ptep);
277	unlock:
278	pte_unmap_unlock(ptep, ptl);
279	out:
280	return SWAP_AGAIN;
281	}
282
283	/*
284	* Get rid of all migration entries and replace them by
285	* references to the indicated page.
286	*/
287	void remove_migration_ptes(struct page *old, struct page *new, bool locked)
288	{
289	struct rmap_walk_control rwc = {
290	.rmap_one = remove_migration_pte,
291	.arg = old,
292	};
293
294	if (locked)
295	rmap_walk_locked(new, &rwc);
296	else
297	rmap_walk(new, &rwc);
298	}
299
300	/*
301	* Something used the pte of a page under migration. We need to
302	* get to the page and wait until migration is finished.
303	* When we return from this function the fault will be retried.
304	*/
305	void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
306	spinlock_t *ptl)
307	{
308	pte_t pte;
309	swp_entry_t entry;
310	struct page *page;
311	#ifdef CONFIG_AMLOGIC_CMA
312	bool need_wait = 0;
313	#endif
314
315	spin_lock(ptl);
316	pte = *ptep;
317	if (!is_swap_pte(pte))
318	goto out;
319
320	entry = pte_to_swp_entry(pte);
321	if (!is_migration_entry(entry))
322	goto out;
323
324	page = migration_entry_to_page(entry);
325	#ifdef CONFIG_AMLOGIC_CMA
326	/* This page is under cma allocating, do not increase it ref */
327	if (PageCmaAllocating(page)) {
328	pr_debug("%s, Page:%lx, flags:%lx, m:%d, c:%d, map:%p\n",
329	__func__, page_to_pfn(page), page->flags,
330	page_mapcount(page), page_count(page),
331	page->mapping);
332	need_wait = 1;
333	goto out;
334	}
335	#endif
336
337	/*
338	* Once radix-tree replacement of page migration started, page_count
339	* *must* be zero. And, we don't want to call wait_on_page_locked()
340	* against a page without get_page().
341	* So, we use get_page_unless_zero(), here. Even failed, page fault
342	* will occur again.
343	*/
344	if (!get_page_unless_zero(page))
345	goto out;
346	pte_unmap_unlock(ptep, ptl);
347	wait_on_page_locked(page);
348	put_page(page);
349	return;
350	out:
351	pte_unmap_unlock(ptep, ptl);
352	#ifdef CONFIG_AMLOGIC_CMA
353	if (need_wait)
354	usleep_range(1000, 1100);
355	#endif
356	}
357
358	void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
359	unsigned long address)
360	{
361	spinlock_t *ptl = pte_lockptr(mm, pmd);
362	pte_t *ptep = pte_offset_map(pmd, address);
363	__migration_entry_wait(mm, ptep, ptl);
364	}
365
366	void migration_entry_wait_huge(struct vm_area_struct *vma,
367	struct mm_struct *mm, pte_t *pte)
368	{
369	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
370	__migration_entry_wait(mm, pte, ptl);
371	}
372
373	#ifdef CONFIG_BLOCK
374	/* Returns true if all buffers are successfully locked */
375	static bool buffer_migrate_lock_buffers(struct buffer_head *head,
376	enum migrate_mode mode)
377	{
378	struct buffer_head *bh = head;
379
380	/* Simple case, sync compaction */
381	if (mode != MIGRATE_ASYNC) {
382	do {
383	get_bh(bh);
384	lock_buffer(bh);
385	bh = bh->b_this_page;
386
387	} while (bh != head);
388
389	return true;
390	}
391
392	/* async case, we cannot block on lock_buffer so use trylock_buffer */
393	do {
394	get_bh(bh);
395	if (!trylock_buffer(bh)) {
396	/*
397	* We failed to lock the buffer and cannot stall in
398	* async migration. Release the taken locks
399	*/
400	struct buffer_head *failed_bh = bh;
401	put_bh(failed_bh);
402	bh = head;
403	while (bh != failed_bh) {
404	unlock_buffer(bh);
405	put_bh(bh);
406	bh = bh->b_this_page;
407	}
408	return false;
409	}
410
411	bh = bh->b_this_page;
412	} while (bh != head);
413	return true;
414	}
415	#else
416	static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
417	enum migrate_mode mode)
418	{
419	return true;
420	}
421	#endif /* CONFIG_BLOCK */
422
423	/*
424	* Replace the page in the mapping.
425	*
426	* The number of remaining references must be:
427	* 1 for anonymous pages without a mapping
428	* 2 for pages with a mapping
429	* 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
430	*/
431	int migrate_page_move_mapping(struct address_space *mapping,
432	struct page *newpage, struct page *page,
433	struct buffer_head *head, enum migrate_mode mode,
434	int extra_count)
435	{
436	struct zone *oldzone, *newzone;
437	int dirty;
438	int expected_count = 1 + extra_count;
439	void **pslot;
440
441	if (!mapping) {
442	/* Anonymous page without mapping */
443	#ifdef CONFIG_AMLOGIC_CMA
444	if (page_count(page) != expected_count)
445	cma_debug(2, page, " anon page cnt miss match, e:%d\n",
446	expected_count);
447	#endif
448	if (page_count(page) != expected_count)
449	return -EAGAIN;
450
451	/* No turning back from here */
452	newpage->index = page->index;
453	newpage->mapping = page->mapping;
454	if (PageSwapBacked(page))
455	__SetPageSwapBacked(newpage);
456
457	return MIGRATEPAGE_SUCCESS;
458	}
459
460	oldzone = page_zone(page);
461	newzone = page_zone(newpage);
462
463	spin_lock_irq(&mapping->tree_lock);
464
465	pslot = radix_tree_lookup_slot(&mapping->page_tree,
466	page_index(page));
467
468	expected_count += 1 + page_has_private(page);
469	if (page_count(page) != expected_count ||
470	radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
471	spin_unlock_irq(&mapping->tree_lock);
472	#ifdef CONFIG_AMLOGIC_CMA
473	cma_debug(2, page, " file page cnt miss match, e:%d, p:%d\n",
474	expected_count, page_has_private(page));
475	#endif
476	return -EAGAIN;
477	}
478
479	if (!page_ref_freeze(page, expected_count)) {
480	spin_unlock_irq(&mapping->tree_lock);
481	return -EAGAIN;
482	}
483
484	/*
485	* In the async migration case of moving a page with buffers, lock the
486	* buffers using trylock before the mapping is moved. If the mapping
487	* was moved, we later failed to lock the buffers and could not move
488	* the mapping back due to an elevated page count, we would have to
489	* block waiting on other references to be dropped.
490	*/
491	if (mode == MIGRATE_ASYNC && head &&
492	!buffer_migrate_lock_buffers(head, mode)) {
493	page_ref_unfreeze(page, expected_count);
494	spin_unlock_irq(&mapping->tree_lock);
495	return -EAGAIN;
496	}
497
498	/*
499	* Now we know that no one else is looking at the page:
500	* no turning back from here.
501	*/
502	newpage->index = page->index;
503	newpage->mapping = page->mapping;
504	if (PageSwapBacked(page))
505	__SetPageSwapBacked(newpage);
506
507	get_page(newpage); /* add cache reference */
508	if (PageSwapCache(page)) {
509	SetPageSwapCache(newpage);
510	set_page_private(newpage, page_private(page));
511	}
512
513	/* Move dirty while page refs frozen and newpage not yet exposed */
514	dirty = PageDirty(page);
515	if (dirty) {
516	ClearPageDirty(page);
517	SetPageDirty(newpage);
518	}
519
520	radix_tree_replace_slot(pslot, newpage);
521
522	/*
523	* Drop cache reference from old page by unfreezing
524	* to one less reference.
525	* We know this isn't the last reference.
526	*/
527	page_ref_unfreeze(page, expected_count - 1);
528
529	spin_unlock(&mapping->tree_lock);
530	/* Leave irq disabled to prevent preemption while updating stats */
531
532	/*
533	* If moved to a different zone then also account
534	* the page for that zone. Other VM counters will be
535	* taken care of when we establish references to the
536	* new page and drop references to the old page.
537	*
538	* Note that anonymous pages are accounted for
539	* via NR_FILE_PAGES and NR_ANON_MAPPED if they
540	* are mapped to swap space.
541	*/
542	if (newzone != oldzone) {
543	__dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
544	__inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
545	if (PageSwapBacked(page) && !PageSwapCache(page)) {
546	__dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
547	__inc_node_state(newzone->zone_pgdat, NR_SHMEM);
548	}
549	if (dirty && mapping_cap_account_dirty(mapping)) {
550	__dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
551	__dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
552	__inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
553	__inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
554	}
555	}
556	local_irq_enable();
557
558	return MIGRATEPAGE_SUCCESS;
559	}
560	EXPORT_SYMBOL(migrate_page_move_mapping);
561
562	/*
563	* The expected number of remaining references is the same as that
564	* of migrate_page_move_mapping().
565	*/
566	int migrate_huge_page_move_mapping(struct address_space *mapping,
567	struct page *newpage, struct page *page)
568	{
569	int expected_count;
570	void **pslot;
571
572	spin_lock_irq(&mapping->tree_lock);
573
574	pslot = radix_tree_lookup_slot(&mapping->page_tree,
575	page_index(page));
576
577	expected_count = 2 + page_has_private(page);
578	if (page_count(page) != expected_count ||
579	radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
580	spin_unlock_irq(&mapping->tree_lock);
581	return -EAGAIN;
582	}
583
584	if (!page_ref_freeze(page, expected_count)) {
585	spin_unlock_irq(&mapping->tree_lock);
586	return -EAGAIN;
587	}
588
589	newpage->index = page->index;
590	newpage->mapping = page->mapping;
591
592	get_page(newpage);
593
594	radix_tree_replace_slot(pslot, newpage);
595
596	page_ref_unfreeze(page, expected_count - 1);
597
598	spin_unlock_irq(&mapping->tree_lock);
599
600	return MIGRATEPAGE_SUCCESS;
601	}
602
603	/*
604	* Gigantic pages are so large that we do not guarantee that page++ pointer
605	* arithmetic will work across the entire page. We need something more
606	* specialized.
607	*/
608	static void __copy_gigantic_page(struct page *dst, struct page *src,
609	int nr_pages)
610	{
611	int i;
612	struct page *dst_base = dst;
613	struct page *src_base = src;
614
615	for (i = 0; i < nr_pages; ) {
616	cond_resched();
617	copy_highpage(dst, src);
618
619	i++;
620	dst = mem_map_next(dst, dst_base, i);
621	src = mem_map_next(src, src_base, i);
622	}
623	}
624
625	static void copy_huge_page(struct page *dst, struct page *src)
626	{
627	int i;
628	int nr_pages;
629
630	if (PageHuge(src)) {
631	/* hugetlbfs page */
632	struct hstate *h = page_hstate(src);
633	nr_pages = pages_per_huge_page(h);
634
635	if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
636	__copy_gigantic_page(dst, src, nr_pages);
637	return;
638	}
639	} else {
640	/* thp page */
641	BUG_ON(!PageTransHuge(src));
642	nr_pages = hpage_nr_pages(src);
643	}
644
645	for (i = 0; i < nr_pages; i++) {
646	cond_resched();
647	copy_highpage(dst + i, src + i);
648	}
649	}
650
651	/*
652	* Copy the page to its new location
653	*/
654	void migrate_page_copy(struct page *newpage, struct page *page)
655	{
656	int cpupid;
657
658	if (PageHuge(page) || PageTransHuge(page))
659	copy_huge_page(newpage, page);
660	else
661	copy_highpage(newpage, page);
662
663	if (PageError(page))
664	SetPageError(newpage);
665	if (PageReferenced(page))
666	SetPageReferenced(newpage);
667	if (PageUptodate(page))
668	SetPageUptodate(newpage);
669	if (TestClearPageActive(page)) {
670	VM_BUG_ON_PAGE(PageUnevictable(page), page);
671	SetPageActive(newpage);
672	} else if (TestClearPageUnevictable(page))
673	SetPageUnevictable(newpage);
674	if (PageWorkingset(page))
675	SetPageWorkingset(newpage);
676	if (PageChecked(page))
677	SetPageChecked(newpage);
678	if (PageMappedToDisk(page))
679	SetPageMappedToDisk(newpage);
680
681	/* Move dirty on pages not done by migrate_page_move_mapping() */
682	if (PageDirty(page))
683	SetPageDirty(newpage);
684
685	if (page_is_young(page))
686	set_page_young(newpage);
687	if (page_is_idle(page))
688	set_page_idle(newpage);
689
690	/*
691	* Copy NUMA information to the new page, to prevent over-eager
692	* future migrations of this same page.
693	*/
694	cpupid = page_cpupid_xchg_last(page, -1);
695	page_cpupid_xchg_last(newpage, cpupid);
696
697	ksm_migrate_page(newpage, page);
698	/*
699	* Please do not reorder this without considering how mm/ksm.c's
700	* get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
701	*/
702	if (PageSwapCache(page))
703	ClearPageSwapCache(page);
704	ClearPagePrivate(page);
705	set_page_private(page, 0);
706
707	/*
708	* If any waiters have accumulated on the new page then
709	* wake them up.
710	*/
711	if (PageWriteback(newpage))
712	end_page_writeback(newpage);
713
714	copy_page_owner(page, newpage);
715
716	mem_cgroup_migrate(page, newpage);
717	}
718	EXPORT_SYMBOL(migrate_page_copy);
719
720	/************************************************************
721	* Migration functions
722	***********************************************************/
723
724	/*
725	* Common logic to directly migrate a single LRU page suitable for
726	* pages that do not use PagePrivate/PagePrivate2.
727	*
728	* Pages are locked upon entry and exit.
729	*/
730	int migrate_page(struct address_space *mapping,
731	struct page *newpage, struct page *page,
732	enum migrate_mode mode)
733	{
734	int rc;
735
736	BUG_ON(PageWriteback(page)); /* Writeback must be complete */
737
738	rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
739
740	if (rc != MIGRATEPAGE_SUCCESS)
741	return rc;
742
743	migrate_page_copy(newpage, page);
744	return MIGRATEPAGE_SUCCESS;
745	}
746	EXPORT_SYMBOL(migrate_page);
747
748	#ifdef CONFIG_BLOCK
749	/*
750	* Migration function for pages with buffers. This function can only be used
751	* if the underlying filesystem guarantees that no other references to "page"
752	* exist.
753	*/
754	int buffer_migrate_page(struct address_space *mapping,
755	struct page *newpage, struct page *page, enum migrate_mode mode)
756	{
757	struct buffer_head *bh, *head;
758	int rc;
759
760	if (!page_has_buffers(page))
761	return migrate_page(mapping, newpage, page, mode);
762
763	head = page_buffers(page);
764
765	rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
766
767	if (rc != MIGRATEPAGE_SUCCESS)
768	return rc;
769
770	/*
771	* In the async case, migrate_page_move_mapping locked the buffers
772	* with an IRQ-safe spinlock held. In the sync case, the buffers
773	* need to be locked now
774	*/
775	if (mode != MIGRATE_ASYNC)
776	BUG_ON(!buffer_migrate_lock_buffers(head, mode));
777
778	ClearPagePrivate(page);
779	set_page_private(newpage, page_private(page));
780	set_page_private(page, 0);
781	put_page(page);
782	get_page(newpage);
783
784	bh = head;
785	do {
786	set_bh_page(bh, newpage, bh_offset(bh));
787	bh = bh->b_this_page;
788
789	} while (bh != head);
790
791	SetPagePrivate(newpage);
792
793	migrate_page_copy(newpage, page);
794
795	bh = head;
796	do {
797	unlock_buffer(bh);
798	put_bh(bh);
799	bh = bh->b_this_page;
800
801	} while (bh != head);
802
803	return MIGRATEPAGE_SUCCESS;
804	}
805	EXPORT_SYMBOL(buffer_migrate_page);
806	#endif
807
808	/*
809	* Writeback a page to clean the dirty state
810	*/
811	static int writeout(struct address_space *mapping, struct page *page)
812	{
813	struct writeback_control wbc = {
814	.sync_mode = WB_SYNC_NONE,
815	.nr_to_write = 1,
816	.range_start = 0,
817	.range_end = LLONG_MAX,
818	.for_reclaim = 1
819	};
820	int rc;
821
822	if (!mapping->a_ops->writepage)
823	/* No write method for the address space */
824	return -EINVAL;
825
826	if (!clear_page_dirty_for_io(page))
827	/* Someone else already triggered a write */
828	return -EAGAIN;
829
830	/*
831	* A dirty page may imply that the underlying filesystem has
832	* the page on some queue. So the page must be clean for
833	* migration. Writeout may mean we loose the lock and the
834	* page state is no longer what we checked for earlier.
835	* At this point we know that the migration attempt cannot
836	* be successful.
837	*/
838	remove_migration_ptes(page, page, false);
839
840	rc = mapping->a_ops->writepage(page, &wbc);
841
842	if (rc != AOP_WRITEPAGE_ACTIVATE)
843	/* unlocked. Relock */
844	lock_page(page);
845
846	return (rc < 0) ? -EIO : -EAGAIN;
847	}
848
849	/*
850	* Default handling if a filesystem does not provide a migration function.
851	*/
852	static int fallback_migrate_page(struct address_space *mapping,
853	struct page *newpage, struct page *page, enum migrate_mode mode)
854	{
855	if (PageDirty(page)) {
856	/* Only writeback pages in full synchronous migration */
857	if (mode != MIGRATE_SYNC)
858	return -EBUSY;
859	return writeout(mapping, page);
860	}
861
862	/*
863	* Buffers may be managed in a filesystem specific way.
864	* We must have no buffers or drop them.
865	*/
866	if (page_has_private(page) &&
867	!try_to_release_page(page, GFP_KERNEL))
868	return -EAGAIN;
869
870	return migrate_page(mapping, newpage, page, mode);
871	}
872
873	/*
874	* Move a page to a newly allocated page
875	* The page is locked and all ptes have been successfully removed.
876	*
877	* The new page will have replaced the old page if this function
878	* is successful.
879	*
880	* Return value:
881	* < 0 - error code
882	* MIGRATEPAGE_SUCCESS - success
883	*/
884	static int move_to_new_page(struct page *newpage, struct page *page,
885	enum migrate_mode mode)
886	{
887	struct address_space *mapping;
888	int rc = -EAGAIN;
889	bool is_lru = !__PageMovable(page);
890
891	VM_BUG_ON_PAGE(!PageLocked(page), page);
892	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
893
894	mapping = page_mapping(page);
895
896	if (likely(is_lru)) {
897	if (!mapping)
898	rc = migrate_page(mapping, newpage, page, mode);
899	else if (mapping->a_ops->migratepage)
900	/*
901	* Most pages have a mapping and most filesystems
902	* provide a migratepage callback. Anonymous pages
903	* are part of swap space which also has its own
904	* migratepage callback. This is the most common path
905	* for page migration.
906	*/
907	rc = mapping->a_ops->migratepage(mapping, newpage,
908	page, mode);
909	else
910	rc = fallback_migrate_page(mapping, newpage,
911	page, mode);
912	} else {
913	/*
914	* In case of non-lru page, it could be released after
915	* isolation step. In that case, we shouldn't try migration.
916	*/
917	VM_BUG_ON_PAGE(!PageIsolated(page), page);
918	if (!PageMovable(page)) {
919	rc = MIGRATEPAGE_SUCCESS;
920	__ClearPageIsolated(page);
921	goto out;
922	}
923
924	rc = mapping->a_ops->migratepage(mapping, newpage,
925	page, mode);
926	WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
927	!PageIsolated(page));
928	}
929
930	/*
931	* When successful, old pagecache page->mapping must be cleared before
932	* page is freed; but stats require that PageAnon be left as PageAnon.
933	*/
934	if (rc == MIGRATEPAGE_SUCCESS) {
935	if (__PageMovable(page)) {
936	VM_BUG_ON_PAGE(!PageIsolated(page), page);
937
938	/*
939	* We clear PG_movable under page_lock so any compactor
940	* cannot try to migrate this page.
941	*/
942	__ClearPageIsolated(page);
943	}
944
945	/*
946	* Anonymous and movable page->mapping will be cleard by
947	* free_pages_prepare so don't reset it here for keeping
948	* the type to work PageAnon, for example.
949	*/
950	if (!PageMappingFlags(page))
951	page->mapping = NULL;
952	}
953	out:
954	return rc;
955	}
956
957	static int __unmap_and_move(struct page *page, struct page *newpage,
958	int force, enum migrate_mode mode)
959	{
960	int rc = -EAGAIN;
961	int page_was_mapped = 0;
962	struct anon_vma *anon_vma = NULL;
963	bool is_lru = !__PageMovable(page);
964
965	if (!trylock_page(page)) {
966	if (!force || mode == MIGRATE_ASYNC)
967	goto out;
968
969	/*
970	* It's not safe for direct compaction to call lock_page.
971	* For example, during page readahead pages are added locked
972	* to the LRU. Later, when the IO completes the pages are
973	* marked uptodate and unlocked. However, the queueing
974	* could be merging multiple pages for one bio (e.g.
975	* mpage_readpages). If an allocation happens for the
976	* second or third page, the process can end up locking
977	* the same page twice and deadlocking. Rather than
978	* trying to be clever about what pages can be locked,
979	* avoid the use of lock_page for direct compaction
980	* altogether.
981	*/
982	if (current->flags & PF_MEMALLOC)
983	goto out;
984
985	lock_page(page);
986	}
987
988	if (PageWriteback(page)) {
989	/*
990	* Only in the case of a full synchronous migration is it
991	* necessary to wait for PageWriteback. In the async case,
992	* the retry loop is too short and in the sync-light case,
993	* the overhead of stalling is too much
994	*/
995	if (mode != MIGRATE_SYNC) {
996	rc = -EBUSY;
997	goto out_unlock;
998	}
999	if (!force)
1000	goto out_unlock;
1001	wait_on_page_writeback(page);
1002	}
1003
1004	/*
1005	* By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1006	* we cannot notice that anon_vma is freed while we migrates a page.
1007	* This get_anon_vma() delays freeing anon_vma pointer until the end
1008	* of migration. File cache pages are no problem because of page_lock()
1009	* File Caches may use write_page() or lock_page() in migration, then,
1010	* just care Anon page here.
1011	*
1012	* Only page_get_anon_vma() understands the subtleties of
1013	* getting a hold on an anon_vma from outside one of its mms.
1014	* But if we cannot get anon_vma, then we won't need it anyway,
1015	* because that implies that the anon page is no longer mapped
1016	* (and cannot be remapped so long as we hold the page lock).
1017	*/
1018	if (PageAnon(page) && !PageKsm(page))
1019	anon_vma = page_get_anon_vma(page);
1020
1021	/*
1022	* Block others from accessing the new page when we get around to
1023	* establishing additional references. We are usually the only one
1024	* holding a reference to newpage at this point. We used to have a BUG
1025	* here if trylock_page(newpage) fails, but would like to allow for
1026	* cases where there might be a race with the previous use of newpage.
1027	* This is much like races on refcount of oldpage: just don't BUG().
1028	*/
1029	if (unlikely(!trylock_page(newpage)))
1030	goto out_unlock;
1031
1032	if (unlikely(!is_lru)) {
1033	rc = move_to_new_page(newpage, page, mode);
1034	goto out_unlock_both;
1035	}
1036
1037	/*
1038	* Corner case handling:
1039	* 1. When a new swap-cache page is read into, it is added to the LRU
1040	* and treated as swapcache but it has no rmap yet.
1041	* Calling try_to_unmap() against a page->mapping==NULL page will
1042	* trigger a BUG. So handle it here.
1043	* 2. An orphaned page (see truncate_complete_page) might have
1044	* fs-private metadata. The page can be picked up due to memory
1045	* offlining. Everywhere else except page reclaim, the page is
1046	* invisible to the vm, so the page can not be migrated. So try to
1047	* free the metadata, so the page can be freed.
1048	*/
1049	if (!page->mapping) {
1050	VM_BUG_ON_PAGE(PageAnon(page), page);
1051	if (page_has_private(page)) {
1052	try_to_free_buffers(page);
1053	goto out_unlock_both;
1054	}
1055	} else if (page_mapped(page)) {
1056	#ifdef CONFIG_AMLOGIC_CMA
1057	int ret;
1058
1059	ret = try_to_unmap(page,
1060	TTU_MIGRATION | TTU_IGNORE_MLOCK |
1061	TTU_IGNORE_ACCESS);
1062	if (ret != SWAP_SUCCESS)
1063	cma_debug(2, page, " unmap failed:%d\n", ret);
1064	#else
1065	/* Establish migration ptes */
1066	VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1067	page);
1068	try_to_unmap(page,
1069	TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1070	#endif
1071	page_was_mapped = 1;
1072	}
1073
1074	if (!page_mapped(page))
1075	rc = move_to_new_page(newpage, page, mode);
1076
1077	if (page_was_mapped)
1078	remove_migration_ptes(page,
1079	rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1080
1081	out_unlock_both:
1082	unlock_page(newpage);
1083	out_unlock:
1084	/* Drop an anon_vma reference if we took one */
1085	if (anon_vma)
1086	put_anon_vma(anon_vma);
1087	unlock_page(page);
1088	out:
1089	/*
1090	* If migration is successful, decrease refcount of the newpage
1091	* which will not free the page because new page owner increased
1092	* refcounter. As well, if it is LRU page, add the page to LRU
1093	* list in here. Use the old state of the isolated source page to
1094	* determine if we migrated a LRU page. newpage was already unlocked
1095	* and possibly modified by its owner - don't rely on the page
1096	* state.
1097	*/
1098	if (rc == MIGRATEPAGE_SUCCESS) {
1099	if (unlikely(!is_lru))
1100	put_page(newpage);
1101	else
1102	putback_lru_page(newpage);
1103	}
1104
1105	return rc;
1106	}
1107
1108	/*
1109	* gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1110	* around it.
1111	*/
1112	#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1113	#define ICE_noinline noinline
1114	#else
1115	#define ICE_noinline
1116	#endif
1117
1118	/*
1119	* Obtain the lock on page, remove all ptes and migrate the page
1120	* to the newly allocated page in newpage.
1121	*/
1122	static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1123	free_page_t put_new_page,
1124	unsigned long private, struct page *page,
1125	int force, enum migrate_mode mode,
1126	enum migrate_reason reason)
1127	{
1128	int rc = MIGRATEPAGE_SUCCESS;
1129	int *result = NULL;
1130	struct page *newpage;
1131
1132	newpage = get_new_page(page, private, &result);
1133	if (!newpage)
1134	return -ENOMEM;
1135
1136	if (page_count(page) == 1) {
1137	/* page was freed from under us. So we are done. */
1138	ClearPageActive(page);
1139	ClearPageUnevictable(page);
1140	if (unlikely(__PageMovable(page))) {
1141	lock_page(page);
1142	if (!PageMovable(page))
1143	__ClearPageIsolated(page);
1144	unlock_page(page);
1145	}
1146	if (put_new_page)
1147	put_new_page(newpage, private);
1148	else
1149	put_page(newpage);
1150	goto out;
1151	}
1152
1153	if (unlikely(PageTransHuge(page))) {
1154	lock_page(page);
1155	rc = split_huge_page(page);
1156	unlock_page(page);
1157	if (rc)
1158	goto out;
1159	}
1160
1161	rc = __unmap_and_move(page, newpage, force, mode);
1162	if (rc == MIGRATEPAGE_SUCCESS)
1163	set_page_owner_migrate_reason(newpage, reason);
1164
1165	out:
1166	if (rc != -EAGAIN) {
1167	/*
1168	* A page that has been migrated has all references
1169	* removed and will be freed. A page that has not been
1170	* migrated will have kepts its references and be
1171	* restored.
1172	*/
1173	list_del(&page->lru);
1174
1175	/*
1176	* Compaction can migrate also non-LRU pages which are
1177	* not accounted to NR_ISOLATED_*. They can be recognized
1178	* as __PageMovable
1179	*/
1180	if (likely(!__PageMovable(page)))
1181	dec_node_page_state(page, NR_ISOLATED_ANON +
1182	page_is_file_cache(page));
1183	}
1184
1185	/*
1186	* If migration is successful, releases reference grabbed during
1187	* isolation. Otherwise, restore the page to right list unless
1188	* we want to retry.
1189	*/
1190	if (rc == MIGRATEPAGE_SUCCESS) {
1191	put_page(page);
1192	if (reason == MR_MEMORY_FAILURE) {
1193	/*
1194	* Set PG_HWPoison on just freed page
1195	* intentionally. Although it's rather weird,
1196	* it's how HWPoison flag works at the moment.
1197	*/
1198	if (!test_set_page_hwpoison(page))
1199	num_poisoned_pages_inc();
1200	}
1201	} else {
1202	if (rc != -EAGAIN) {
1203	if (likely(!__PageMovable(page))) {
1204	putback_lru_page(page);
1205	goto put_new;
1206	}
1207
1208	lock_page(page);
1209	if (PageMovable(page))
1210	putback_movable_page(page);
1211	else
1212	__ClearPageIsolated(page);
1213	unlock_page(page);
1214	put_page(page);
1215	}
1216	put_new:
1217	if (put_new_page)
1218	put_new_page(newpage, private);
1219	else
1220	put_page(newpage);
1221	}
1222
1223	if (result) {
1224	if (rc)
1225	*result = rc;
1226	else
1227	*result = page_to_nid(newpage);
1228	}
1229	#ifdef CONFIG_AMLOGIC_CMA
1230	if (reason == MR_CMA && rc == MIGRATEPAGE_SUCCESS)
1231	ClearPageCmaAllocating(page);
1232	#endif
1233	return rc;
1234	}
1235
1236	/*
1237	* Counterpart of unmap_and_move_page() for hugepage migration.
1238	*
1239	* This function doesn't wait the completion of hugepage I/O
1240	* because there is no race between I/O and migration for hugepage.
1241	* Note that currently hugepage I/O occurs only in direct I/O
1242	* where no lock is held and PG_writeback is irrelevant,
1243	* and writeback status of all subpages are counted in the reference
1244	* count of the head page (i.e. if all subpages of a 2MB hugepage are
1245	* under direct I/O, the reference of the head page is 512 and a bit more.)
1246	* This means that when we try to migrate hugepage whose subpages are
1247	* doing direct I/O, some references remain after try_to_unmap() and
1248	* hugepage migration fails without data corruption.
1249	*
1250	* There is also no race when direct I/O is issued on the page under migration,
1251	* because then pte is replaced with migration swap entry and direct I/O code
1252	* will wait in the page fault for migration to complete.
1253	*/
1254	static int unmap_and_move_huge_page(new_page_t get_new_page,
1255	free_page_t put_new_page, unsigned long private,
1256	struct page *hpage, int force,
1257	enum migrate_mode mode, int reason)
1258	{
1259	int rc = -EAGAIN;
1260	int *result = NULL;
1261	int page_was_mapped = 0;
1262	struct page *new_hpage;
1263	struct anon_vma *anon_vma = NULL;
1264
1265	/*
1266	* Movability of hugepages depends on architectures and hugepage size.
1267	* This check is necessary because some callers of hugepage migration
1268	* like soft offline and memory hotremove don't walk through page
1269	* tables or check whether the hugepage is pmd-based or not before
1270	* kicking migration.
1271	*/
1272	if (!hugepage_migration_supported(page_hstate(hpage))) {
1273	putback_active_hugepage(hpage);
1274	return -ENOSYS;
1275	}
1276
1277	new_hpage = get_new_page(hpage, private, &result);
1278	if (!new_hpage)
1279	return -ENOMEM;
1280
1281	if (!trylock_page(hpage)) {
1282	if (!force || mode != MIGRATE_SYNC)
1283	goto out;
1284	lock_page(hpage);
1285	}
1286
1287	/*
1288	* Check for pages which are in the process of being freed. Without
1289	* page_mapping() set, hugetlbfs specific move page routine will not
1290	* be called and we could leak usage counts for subpools.
1291	*/
1292	if (page_private(hpage) && !page_mapping(hpage)) {
1293	rc = -EBUSY;
1294	goto out_unlock;
1295	}
1296
1297	if (PageAnon(hpage))
1298	anon_vma = page_get_anon_vma(hpage);
1299
1300	if (unlikely(!trylock_page(new_hpage)))
1301	goto put_anon;
1302
1303	if (page_mapped(hpage)) {
1304	try_to_unmap(hpage,
1305	TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1306	page_was_mapped = 1;
1307	}
1308
1309	if (!page_mapped(hpage))
1310	rc = move_to_new_page(new_hpage, hpage, mode);
1311
1312	if (page_was_mapped)
1313	remove_migration_ptes(hpage,
1314	rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1315
1316	unlock_page(new_hpage);
1317
1318	put_anon:
1319	if (anon_vma)
1320	put_anon_vma(anon_vma);
1321
1322	if (rc == MIGRATEPAGE_SUCCESS) {
1323	hugetlb_cgroup_migrate(hpage, new_hpage);
1324	put_new_page = NULL;
1325	set_page_owner_migrate_reason(new_hpage, reason);
1326	}
1327
1328	out_unlock:
1329	unlock_page(hpage);
1330	out:
1331	if (rc != -EAGAIN)
1332	putback_active_hugepage(hpage);
1333
1334	/*
1335	* If migration was not successful and there's a freeing callback, use
1336	* it. Otherwise, put_page() will drop the reference grabbed during
1337	* isolation.
1338	*/
1339	if (put_new_page)
1340	put_new_page(new_hpage, private);
1341	else
1342	putback_active_hugepage(new_hpage);
1343
1344	if (result) {
1345	if (rc)
1346	*result = rc;
1347	else
1348	*result = page_to_nid(new_hpage);
1349	}
1350	return rc;
1351	}
1352
1353	/*
1354	* migrate_pages - migrate the pages specified in a list, to the free pages
1355	* supplied as the target for the page migration
1356	*
1357	* @from: The list of pages to be migrated.
1358	* @get_new_page: The function used to allocate free pages to be used
1359	* as the target of the page migration.
1360	* @put_new_page: The function used to free target pages if migration
1361	* fails, or NULL if no special handling is necessary.
1362	* @private: Private data to be passed on to get_new_page()
1363	* @mode: The migration mode that specifies the constraints for
1364	* page migration, if any.
1365	* @reason: The reason for page migration.
1366	*
1367	* The function returns after 10 attempts or if no pages are movable any more
1368	* because the list has become empty or no retryable pages exist any more.
1369	* The caller should call putback_movable_pages() to return pages to the LRU
1370	* or free list only if ret != 0.
1371	*
1372	* Returns the number of pages that were not migrated, or an error code.
1373	*/
1374	int migrate_pages(struct list_head *from, new_page_t get_new_page,
1375	free_page_t put_new_page, unsigned long private,
1376	enum migrate_mode mode, int reason)
1377	{
1378	int retry = 1;
1379	int nr_failed = 0;
1380	int nr_succeeded = 0;
1381	int pass = 0;
1382	struct page *page;
1383	struct page *page2;
1384	int swapwrite = current->flags & PF_SWAPWRITE;
1385	int rc;
1386
1387	if (!swapwrite)
1388	current->flags |= PF_SWAPWRITE;
1389
1390	for(pass = 0; pass < 10 && retry; pass++) {
1391	retry = 0;
1392
1393	list_for_each_entry_safe(page, page2, from, lru) {
1394	cond_resched();
1395
1396	if (PageHuge(page))
1397	rc = unmap_and_move_huge_page(get_new_page,
1398	put_new_page, private, page,
1399	pass > 2, mode, reason);
1400	else
1401	rc = unmap_and_move(get_new_page, put_new_page,
1402	private, page, pass > 2, mode,
1403	reason);
1404
1405	switch(rc) {
1406	case -ENOMEM:
1407	nr_failed++;
1408	#ifdef CONFIG_AMLOGIC_CMA
1409	cma_debug(2, page, " NO MEM\n");
1410	#endif
1411	goto out;
1412	case -EAGAIN:
1413	retry++;
1414	break;
1415	case MIGRATEPAGE_SUCCESS:
1416	nr_succeeded++;
1417	break;
1418	default:
1419	/*
1420	* Permanent failure (-EBUSY, -ENOSYS, etc.):
1421	* unlike -EAGAIN case, the failed page is
1422	* removed from migration page list and not
1423	* retried in the next outer loop.
1424	*/
1425	nr_failed++;
1426	#ifdef CONFIG_AMLOGIC_CMA
1427	cma_debug(2, page, " failed:%d\n", rc);
1428	#endif
1429	break;
1430	}
1431	}
1432	}
1433	nr_failed += retry;
1434	rc = nr_failed;
1435	out:
1436	if (nr_succeeded)
1437	count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1438	if (nr_failed)
1439	count_vm_events(PGMIGRATE_FAIL, nr_failed);
1440	trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1441
1442	if (!swapwrite)
1443	current->flags &= ~PF_SWAPWRITE;
1444
1445	return rc;
1446	}
1447
1448	#ifdef CONFIG_NUMA
1449	/*
1450	* Move a list of individual pages
1451	*/
1452	struct page_to_node {
1453	unsigned long addr;
1454	struct page *page;
1455	int node;
1456	int status;
1457	};
1458
1459	static struct page *new_page_node(struct page *p, unsigned long private,
1460	int **result)
1461	{
1462	struct page_to_node *pm = (struct page_to_node *)private;
1463
1464	while (pm->node != MAX_NUMNODES && pm->page != p)
1465	pm++;
1466
1467	if (pm->node == MAX_NUMNODES)
1468	return NULL;
1469
1470	*result = &pm->status;
1471
1472	if (PageHuge(p))
1473	return alloc_huge_page_node(page_hstate(compound_head(p)),
1474	pm->node);
1475	else
1476	return __alloc_pages_node(pm->node,
1477	GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1478	}
1479
1480	/*
1481	* Move a set of pages as indicated in the pm array. The addr
1482	* field must be set to the virtual address of the page to be moved
1483	* and the node number must contain a valid target node.
1484	* The pm array ends with node = MAX_NUMNODES.
1485	*/
1486	static int do_move_page_to_node_array(struct mm_struct *mm,
1487	struct page_to_node *pm,
1488	int migrate_all)
1489	{
1490	int err;
1491	struct page_to_node *pp;
1492	LIST_HEAD(pagelist);
1493
1494	down_read(&mm->mmap_sem);
1495
1496	/*
1497	* Build a list of pages to migrate
1498	*/
1499	for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1500	struct vm_area_struct *vma;
1501	struct page *page;
1502
1503	err = -EFAULT;
1504	vma = find_vma(mm, pp->addr);
1505	if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1506	goto set_status;
1507
1508	/* FOLL_DUMP to ignore special (like zero) pages */
1509	page = follow_page(vma, pp->addr,
1510	FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1511
1512	err = PTR_ERR(page);
1513	if (IS_ERR(page))
1514	goto set_status;
1515
1516	err = -ENOENT;
1517	if (!page)
1518	goto set_status;
1519
1520	pp->page = page;
1521	err = page_to_nid(page);
1522
1523	if (err == pp->node)
1524	/*
1525	* Node already in the right place
1526	*/
1527	goto put_and_set;
1528
1529	err = -EACCES;
1530	if (page_mapcount(page) > 1 &&
1531	!migrate_all)
1532	goto put_and_set;
1533
1534	if (PageHuge(page)) {
1535	if (PageHead(page))
1536	isolate_huge_page(page, &pagelist);
1537	goto put_and_set;
1538	}
1539
1540	err = isolate_lru_page(page);
1541	if (!err) {
1542	list_add_tail(&page->lru, &pagelist);
1543	inc_node_page_state(page, NR_ISOLATED_ANON +
1544	page_is_file_cache(page));
1545	}
1546	put_and_set:
1547	/*
1548	* Either remove the duplicate refcount from
1549	* isolate_lru_page() or drop the page ref if it was
1550	* not isolated.
1551	*/
1552	put_page(page);
1553	set_status:
1554	pp->status = err;
1555	}
1556
1557	err = 0;
1558	if (!list_empty(&pagelist)) {
1559	err = migrate_pages(&pagelist, new_page_node, NULL,
1560	(unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1561	if (err)
1562	putback_movable_pages(&pagelist);
1563	}
1564
1565	up_read(&mm->mmap_sem);
1566	return err;
1567	}
1568
1569	/*
1570	* Migrate an array of page address onto an array of nodes and fill
1571	* the corresponding array of status.
1572	*/
1573	static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1574	unsigned long nr_pages,
1575	const void __user * __user *pages,
1576	const int __user *nodes,
1577	int __user *status, int flags)
1578	{
1579	struct page_to_node *pm;
1580	unsigned long chunk_nr_pages;
1581	unsigned long chunk_start;
1582	int err;
1583
1584	err = -ENOMEM;
1585	pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1586	if (!pm)
1587	goto out;
1588
1589	migrate_prep();
1590
1591	/*
1592	* Store a chunk of page_to_node array in a page,
1593	* but keep the last one as a marker
1594	*/
1595	chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1596
1597	for (chunk_start = 0;
1598	chunk_start < nr_pages;
1599	chunk_start += chunk_nr_pages) {
1600	int j;
1601
1602	if (chunk_start + chunk_nr_pages > nr_pages)
1603	chunk_nr_pages = nr_pages - chunk_start;
1604
1605	/* fill the chunk pm with addrs and nodes from user-space */
1606	for (j = 0; j < chunk_nr_pages; j++) {
1607	const void __user *p;
1608	int node;
1609
1610	err = -EFAULT;
1611	if (get_user(p, pages + j + chunk_start))
1612	goto out_pm;
1613	pm[j].addr = (unsigned long) p;
1614
1615	if (get_user(node, nodes + j + chunk_start))
1616	goto out_pm;
1617
1618	err = -ENODEV;
1619	if (node < 0 || node >= MAX_NUMNODES)
1620	goto out_pm;
1621
1622	if (!node_state(node, N_MEMORY))
1623	goto out_pm;
1624
1625	err = -EACCES;
1626	if (!node_isset(node, task_nodes))
1627	goto out_pm;
1628
1629	pm[j].node = node;
1630	}
1631
1632	/* End marker for this chunk */
1633	pm[chunk_nr_pages].node = MAX_NUMNODES;
1634
1635	/* Migrate this chunk */
1636	err = do_move_page_to_node_array(mm, pm,
1637	flags & MPOL_MF_MOVE_ALL);
1638	if (err < 0)
1639	goto out_pm;
1640
1641	/* Return status information */
1642	for (j = 0; j < chunk_nr_pages; j++)
1643	if (put_user(pm[j].status, status + j + chunk_start)) {
1644	err = -EFAULT;
1645	goto out_pm;
1646	}
1647	}
1648	err = 0;
1649
1650	out_pm:
1651	free_page((unsigned long)pm);
1652	out:
1653	return err;
1654	}
1655
1656	/*
1657	* Determine the nodes of an array of pages and store it in an array of status.
1658	*/
1659	static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1660	const void __user **pages, int *status)
1661	{
1662	unsigned long i;
1663
1664	down_read(&mm->mmap_sem);
1665
1666	for (i = 0; i < nr_pages; i++) {
1667	unsigned long addr = (unsigned long)(*pages);
1668	struct vm_area_struct *vma;
1669	struct page *page;
1670	int err = -EFAULT;
1671
1672	vma = find_vma(mm, addr);
1673	if (!vma || addr < vma->vm_start)
1674	goto set_status;
1675
1676	/* FOLL_DUMP to ignore special (like zero) pages */
1677	page = follow_page(vma, addr, FOLL_DUMP);
1678
1679	err = PTR_ERR(page);
1680	if (IS_ERR(page))
1681	goto set_status;
1682
1683	err = page ? page_to_nid(page) : -ENOENT;
1684	set_status:
1685	*status = err;
1686
1687	pages++;
1688	status++;
1689	}
1690
1691	up_read(&mm->mmap_sem);
1692	}
1693
1694	/*
1695	* Determine the nodes of a user array of pages and store it in
1696	* a user array of status.
1697	*/
1698	static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1699	const void __user * __user *pages,
1700	int __user *status)
1701	{
1702	#define DO_PAGES_STAT_CHUNK_NR 16
1703	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1704	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1705
1706	while (nr_pages) {
1707	unsigned long chunk_nr;
1708
1709	chunk_nr = nr_pages;
1710	if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1711	chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1712
1713	if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1714	break;
1715
1716	do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1717
1718	if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1719	break;
1720
1721	pages += chunk_nr;
1722	status += chunk_nr;
1723	nr_pages -= chunk_nr;
1724	}
1725	return nr_pages ? -EFAULT : 0;
1726	}
1727
1728	/*
1729	* Move a list of pages in the address space of the currently executing
1730	* process.
1731	*/
1732	SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1733	const void __user * __user *, pages,
1734	const int __user *, nodes,
1735	int __user *, status, int, flags)
1736	{
1737	struct task_struct *task;
1738	struct mm_struct *mm;
1739	int err;
1740	nodemask_t task_nodes;
1741
1742	/* Check flags */
1743	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1744	return -EINVAL;
1745
1746	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1747	return -EPERM;
1748
1749	/* Find the mm_struct */
1750	rcu_read_lock();
1751	task = pid ? find_task_by_vpid(pid) : current;
1752	if (!task) {
1753	rcu_read_unlock();
1754	return -ESRCH;
1755	}
1756	get_task_struct(task);
1757
1758	/*
1759	* Check if this process has the right to modify the specified
1760	* process. Use the regular "ptrace_may_access()" checks.
1761	*/
1762	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1763	rcu_read_unlock();
1764	err = -EPERM;
1765	goto out;
1766	}
1767	rcu_read_unlock();
1768
1769	err = security_task_movememory(task);
1770	if (err)
1771	goto out;
1772
1773	task_nodes = cpuset_mems_allowed(task);
1774	mm = get_task_mm(task);
1775	put_task_struct(task);
1776
1777	if (!mm)
1778	return -EINVAL;
1779
1780	if (nodes)
1781	err = do_pages_move(mm, task_nodes, nr_pages, pages,
1782	nodes, status, flags);
1783	else
1784	err = do_pages_stat(mm, nr_pages, pages, status);
1785
1786	mmput(mm);
1787	return err;
1788
1789	out:
1790	put_task_struct(task);
1791	return err;
1792	}
1793
1794	#ifdef CONFIG_NUMA_BALANCING
1795	/*
1796	* Returns true if this is a safe migration target node for misplaced NUMA
1797	* pages. Currently it only checks the watermarks which crude
1798	*/
1799	static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1800	unsigned long nr_migrate_pages)
1801	{
1802	int z;
1803
1804	if (!pgdat_reclaimable(pgdat))
1805	return false;
1806
1807	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1808	struct zone *zone = pgdat->node_zones + z;
1809
1810	if (!populated_zone(zone))
1811	continue;
1812
1813	/* Avoid waking kswapd by allocating pages_to_migrate pages. */
1814	if (!zone_watermark_ok(zone, 0,
1815	high_wmark_pages(zone) +
1816	nr_migrate_pages,
1817	0, 0))
1818	continue;
1819	return true;
1820	}
1821	return false;
1822	}
1823
1824	static struct page *alloc_misplaced_dst_page(struct page *page,
1825	unsigned long data,
1826	int **result)
1827	{
1828	int nid = (int) data;
1829	struct page *newpage;
1830
1831	newpage = __alloc_pages_node(nid,
1832	(GFP_HIGHUSER_MOVABLE |
1833	__GFP_THISNODE | __GFP_NOMEMALLOC |
1834	__GFP_NORETRY | __GFP_NOWARN) &
1835	~__GFP_RECLAIM, 0);
1836
1837	return newpage;
1838	}
1839
1840	/*
1841	* page migration rate limiting control.
1842	* Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1843	* window of time. Default here says do not migrate more than 1280M per second.
1844	*/
1845	static unsigned int migrate_interval_millisecs __read_mostly = 100;
1846	static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1847
1848	/* Returns true if the node is migrate rate-limited after the update */
1849	static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1850	unsigned long nr_pages)
1851	{
1852	/*
1853	* Rate-limit the amount of data that is being migrated to a node.
1854	* Optimal placement is no good if the memory bus is saturated and
1855	* all the time is being spent migrating!
1856	*/
1857	if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1858	spin_lock(&pgdat->numabalancing_migrate_lock);
1859	pgdat->numabalancing_migrate_nr_pages = 0;
1860	pgdat->numabalancing_migrate_next_window = jiffies +
1861	msecs_to_jiffies(migrate_interval_millisecs);
1862	spin_unlock(&pgdat->numabalancing_migrate_lock);
1863	}
1864	if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1865	trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1866	nr_pages);
1867	return true;
1868	}
1869
1870	/*
1871	* This is an unlocked non-atomic update so errors are possible.
1872	* The consequences are failing to migrate when we potentiall should
1873	* have which is not severe enough to warrant locking. If it is ever
1874	* a problem, it can be converted to a per-cpu counter.
1875	*/
1876	pgdat->numabalancing_migrate_nr_pages += nr_pages;
1877	return false;
1878	}
1879
1880	static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1881	{
1882	int page_lru;
1883
1884	VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1885
1886	/* Avoid migrating to a node that is nearly full */
1887	if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1888	return 0;
1889
1890	if (isolate_lru_page(page))
1891	return 0;
1892
1893	/*
1894	* migrate_misplaced_transhuge_page() skips page migration's usual
1895	* check on page_count(), so we must do it here, now that the page
1896	* has been isolated: a GUP pin, or any other pin, prevents migration.
1897	* The expected page count is 3: 1 for page's mapcount and 1 for the
1898	* caller's pin and 1 for the reference taken by isolate_lru_page().
1899	*/
1900	if (PageTransHuge(page) && page_count(page) != 3) {
1901	putback_lru_page(page);
1902	return 0;
1903	}
1904
1905	page_lru = page_is_file_cache(page);
1906	mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1907	hpage_nr_pages(page));
1908
1909	/*
1910	* Isolating the page has taken another reference, so the
1911	* caller's reference can be safely dropped without the page
1912	* disappearing underneath us during migration.
1913	*/
1914	put_page(page);
1915	return 1;
1916	}
1917
1918	bool pmd_trans_migrating(pmd_t pmd)
1919	{
1920	struct page *page = pmd_page(pmd);
1921	return PageLocked(page);
1922	}
1923
1924	/*
1925	* Attempt to migrate a misplaced page to the specified destination
1926	* node. Caller is expected to have an elevated reference count on
1927	* the page that will be dropped by this function before returning.
1928	*/
1929	int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1930	int node)
1931	{
1932	pg_data_t *pgdat = NODE_DATA(node);
1933	int isolated;
1934	int nr_remaining;
1935	LIST_HEAD(migratepages);
1936
1937	/*
1938	* Don't migrate file pages that are mapped in multiple processes
1939	* with execute permissions as they are probably shared libraries.
1940	*/
1941	if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1942	(vma->vm_flags & VM_EXEC))
1943	goto out;
1944
1945	/*
1946	* Rate-limit the amount of data that is being migrated to a node.
1947	* Optimal placement is no good if the memory bus is saturated and
1948	* all the time is being spent migrating!
1949	*/
1950	if (numamigrate_update_ratelimit(pgdat, 1))
1951	goto out;
1952
1953	isolated = numamigrate_isolate_page(pgdat, page);
1954	if (!isolated)
1955	goto out;
1956
1957	list_add(&page->lru, &migratepages);
1958	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1959	NULL, node, MIGRATE_ASYNC,
1960	MR_NUMA_MISPLACED);
1961	if (nr_remaining) {
1962	if (!list_empty(&migratepages)) {
1963	list_del(&page->lru);
1964	dec_node_page_state(page, NR_ISOLATED_ANON +
1965	page_is_file_cache(page));
1966	putback_lru_page(page);
1967	}
1968	isolated = 0;
1969	} else
1970	count_vm_numa_event(NUMA_PAGE_MIGRATE);
1971	BUG_ON(!list_empty(&migratepages));
1972	return isolated;
1973
1974	out:
1975	put_page(page);
1976	return 0;
1977	}
1978	#endif /* CONFIG_NUMA_BALANCING */
1979
1980	#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1981	/*
1982	* Migrates a THP to a given target node. page must be locked and is unlocked
1983	* before returning.
1984	*/
1985	int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1986	struct vm_area_struct *vma,
1987	pmd_t *pmd, pmd_t entry,
1988	unsigned long address,
1989	struct page *page, int node)
1990	{
1991	spinlock_t *ptl;
1992	pg_data_t *pgdat = NODE_DATA(node);
1993	int isolated = 0;
1994	struct page *new_page = NULL;
1995	int page_lru = page_is_file_cache(page);
1996	unsigned long mmun_start = address & HPAGE_PMD_MASK;
1997	unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1998	pmd_t orig_entry;
1999
2000	/*
2001	* Rate-limit the amount of data that is being migrated to a node.
2002	* Optimal placement is no good if the memory bus is saturated and
2003	* all the time is being spent migrating!
2004	*/
2005	if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
2006	goto out_dropref;
2007
2008	new_page = alloc_pages_node(node,
2009	(GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2010	HPAGE_PMD_ORDER);
2011	if (!new_page)
2012	goto out_fail;
2013	prep_transhuge_page(new_page);
2014
2015	isolated = numamigrate_isolate_page(pgdat, page);
2016	if (!isolated) {
2017	put_page(new_page);
2018	goto out_fail;
2019	}
2020	/*
2021	* We are not sure a pending tlb flush here is for a huge page
2022	* mapping or not. Hence use the tlb range variant
2023	*/
2024	if (mm_tlb_flush_pending(mm))
2025	flush_tlb_range(vma, mmun_start, mmun_end);
2026
2027	/* Prepare a page as a migration target */
2028	__SetPageLocked(new_page);
2029	__SetPageSwapBacked(new_page);
2030
2031	/* anon mapping, we can simply copy page->mapping to the new page: */
2032	new_page->mapping = page->mapping;
2033	new_page->index = page->index;
2034	migrate_page_copy(new_page, page);
2035	WARN_ON(PageLRU(new_page));
2036
2037	/* Recheck the target PMD */
2038	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2039	ptl = pmd_lock(mm, pmd);
2040	if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
2041	fail_putback:
2042	spin_unlock(ptl);
2043	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2044
2045	/* Reverse changes made by migrate_page_copy() */
2046	if (TestClearPageActive(new_page))
2047	SetPageActive(page);
2048	if (TestClearPageUnevictable(new_page))
2049	SetPageUnevictable(page);
2050
2051	unlock_page(new_page);
2052	put_page(new_page); /* Free it */
2053
2054	/* Retake the callers reference and putback on LRU */
2055	get_page(page);
2056	putback_lru_page(page);
2057	mod_node_page_state(page_pgdat(page),
2058	NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2059
2060	goto out_unlock;
2061	}
2062
2063	orig_entry = *pmd;
2064	entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2065	entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2066
2067	/*
2068	* Clear the old entry under pagetable lock and establish the new PTE.
2069	* Any parallel GUP will either observe the old page blocking on the
2070	* page lock, block on the page table lock or observe the new page.
2071	* The SetPageUptodate on the new page and page_add_new_anon_rmap
2072	* guarantee the copy is visible before the pagetable update.
2073	*/
2074	flush_cache_range(vma, mmun_start, mmun_end);
2075	page_add_anon_rmap(new_page, vma, mmun_start, true);
2076	pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2077	set_pmd_at(mm, mmun_start, pmd, entry);
2078	update_mmu_cache_pmd(vma, address, &entry);
2079
2080	if (page_count(page) != 2) {
2081	set_pmd_at(mm, mmun_start, pmd, orig_entry);
2082	flush_pmd_tlb_range(vma, mmun_start, mmun_end);
2083	mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2084	update_mmu_cache_pmd(vma, address, &entry);
2085	page_remove_rmap(new_page, true);
2086	goto fail_putback;
2087	}
2088
2089	mlock_migrate_page(new_page, page);
2090	page_remove_rmap(page, true);
2091	set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2092
2093	spin_unlock(ptl);
2094	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2095
2096	/* Take an "isolate" reference and put new page on the LRU. */
2097	get_page(new_page);
2098	putback_lru_page(new_page);
2099
2100	unlock_page(new_page);
2101	unlock_page(page);
2102	put_page(page); /* Drop the rmap reference */
2103	put_page(page); /* Drop the LRU isolation reference */
2104
2105	count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2106	count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2107
2108	mod_node_page_state(page_pgdat(page),
2109	NR_ISOLATED_ANON + page_lru,
2110	-HPAGE_PMD_NR);
2111	return isolated;
2112
2113	out_fail:
2114	count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2115	out_dropref:
2116	ptl = pmd_lock(mm, pmd);
2117	if (pmd_same(*pmd, entry)) {
2118	entry = pmd_modify(entry, vma->vm_page_prot);
2119	set_pmd_at(mm, mmun_start, pmd, entry);
2120	update_mmu_cache_pmd(vma, address, &entry);
2121	}
2122	spin_unlock(ptl);
2123
2124	out_unlock:
2125	unlock_page(page);
2126	put_page(page);
2127	return 0;
2128	}
2129	#endif /* CONFIG_NUMA_BALANCING */
2130
2131	#endif /* CONFIG_NUMA */
2132