]> git.karo-electronics.de Git - karo-tx-linux.git/blob - mm/migrate.c
RDMA/qib,hfi1: Fix MR reference count leak on write with immediate
[karo-tx-linux.git] / mm / migrate.c
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/sched/mm.h>
44
45 #include <asm/tlbflush.h>
46
47 #define CREATE_TRACE_POINTS
48 #include <trace/events/migrate.h>
49
50 #include "internal.h"
51
52 /*
53  * migrate_prep() needs to be called before we start compiling a list of pages
54  * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
55  * undesirable, use migrate_prep_local()
56  */
57 int migrate_prep(void)
58 {
59         /*
60          * Clear the LRU lists so pages can be isolated.
61          * Note that pages may be moved off the LRU after we have
62          * drained them. Those pages will fail to migrate like other
63          * pages that may be busy.
64          */
65         lru_add_drain_all();
66
67         return 0;
68 }
69
70 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
71 int migrate_prep_local(void)
72 {
73         lru_add_drain();
74
75         return 0;
76 }
77
78 int isolate_movable_page(struct page *page, isolate_mode_t mode)
79 {
80         struct address_space *mapping;
81
82         /*
83          * Avoid burning cycles with pages that are yet under __free_pages(),
84          * or just got freed under us.
85          *
86          * In case we 'win' a race for a movable page being freed under us and
87          * raise its refcount preventing __free_pages() from doing its job
88          * the put_page() at the end of this block will take care of
89          * release this page, thus avoiding a nasty leakage.
90          */
91         if (unlikely(!get_page_unless_zero(page)))
92                 goto out;
93
94         /*
95          * Check PageMovable before holding a PG_lock because page's owner
96          * assumes anybody doesn't touch PG_lock of newly allocated page
97          * so unconditionally grapping the lock ruins page's owner side.
98          */
99         if (unlikely(!__PageMovable(page)))
100                 goto out_putpage;
101         /*
102          * As movable pages are not isolated from LRU lists, concurrent
103          * compaction threads can race against page migration functions
104          * as well as race against the releasing a page.
105          *
106          * In order to avoid having an already isolated movable page
107          * being (wrongly) re-isolated while it is under migration,
108          * or to avoid attempting to isolate pages being released,
109          * lets be sure we have the page lock
110          * before proceeding with the movable page isolation steps.
111          */
112         if (unlikely(!trylock_page(page)))
113                 goto out_putpage;
114
115         if (!PageMovable(page) || PageIsolated(page))
116                 goto out_no_isolated;
117
118         mapping = page_mapping(page);
119         VM_BUG_ON_PAGE(!mapping, page);
120
121         if (!mapping->a_ops->isolate_page(page, mode))
122                 goto out_no_isolated;
123
124         /* Driver shouldn't use PG_isolated bit of page->flags */
125         WARN_ON_ONCE(PageIsolated(page));
126         __SetPageIsolated(page);
127         unlock_page(page);
128
129         return 0;
130
131 out_no_isolated:
132         unlock_page(page);
133 out_putpage:
134         put_page(page);
135 out:
136         return -EBUSY;
137 }
138
139 /* It should be called on page which is PG_movable */
140 void putback_movable_page(struct page *page)
141 {
142         struct address_space *mapping;
143
144         VM_BUG_ON_PAGE(!PageLocked(page), page);
145         VM_BUG_ON_PAGE(!PageMovable(page), page);
146         VM_BUG_ON_PAGE(!PageIsolated(page), page);
147
148         mapping = page_mapping(page);
149         mapping->a_ops->putback_page(page);
150         __ClearPageIsolated(page);
151 }
152
153 /*
154  * Put previously isolated pages back onto the appropriate lists
155  * from where they were once taken off for compaction/migration.
156  *
157  * This function shall be used whenever the isolated pageset has been
158  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
159  * and isolate_huge_page().
160  */
161 void putback_movable_pages(struct list_head *l)
162 {
163         struct page *page;
164         struct page *page2;
165
166         list_for_each_entry_safe(page, page2, l, lru) {
167                 if (unlikely(PageHuge(page))) {
168                         putback_active_hugepage(page);
169                         continue;
170                 }
171                 list_del(&page->lru);
172                 /*
173                  * We isolated non-lru movable page so here we can use
174                  * __PageMovable because LRU page's mapping cannot have
175                  * PAGE_MAPPING_MOVABLE.
176                  */
177                 if (unlikely(__PageMovable(page))) {
178                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
179                         lock_page(page);
180                         if (PageMovable(page))
181                                 putback_movable_page(page);
182                         else
183                                 __ClearPageIsolated(page);
184                         unlock_page(page);
185                         put_page(page);
186                 } else {
187                         dec_node_page_state(page, NR_ISOLATED_ANON +
188                                         page_is_file_cache(page));
189                         putback_lru_page(page);
190                 }
191         }
192 }
193
194 /*
195  * Restore a potential migration pte to a working pte entry
196  */
197 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
198                                  unsigned long addr, void *old)
199 {
200         struct page_vma_mapped_walk pvmw = {
201                 .page = old,
202                 .vma = vma,
203                 .address = addr,
204                 .flags = PVMW_SYNC | PVMW_MIGRATION,
205         };
206         struct page *new;
207         pte_t pte;
208         swp_entry_t entry;
209
210         VM_BUG_ON_PAGE(PageTail(page), page);
211         while (page_vma_mapped_walk(&pvmw)) {
212                 if (PageKsm(page))
213                         new = page;
214                 else
215                         new = page - pvmw.page->index +
216                                 linear_page_index(vma, pvmw.address);
217
218                 get_page(new);
219                 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
220                 if (pte_swp_soft_dirty(*pvmw.pte))
221                         pte = pte_mksoft_dirty(pte);
222
223                 /*
224                  * Recheck VMA as permissions can change since migration started
225                  */
226                 entry = pte_to_swp_entry(*pvmw.pte);
227                 if (is_write_migration_entry(entry))
228                         pte = maybe_mkwrite(pte, vma);
229
230 #ifdef CONFIG_HUGETLB_PAGE
231                 if (PageHuge(new)) {
232                         pte = pte_mkhuge(pte);
233                         pte = arch_make_huge_pte(pte, vma, new, 0);
234                 }
235 #endif
236                 flush_dcache_page(new);
237                 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
238
239                 if (PageHuge(new)) {
240                         if (PageAnon(new))
241                                 hugepage_add_anon_rmap(new, vma, pvmw.address);
242                         else
243                                 page_dup_rmap(new, true);
244                 } else if (PageAnon(new))
245                         page_add_anon_rmap(new, vma, pvmw.address, false);
246                 else
247                         page_add_file_rmap(new, false);
248
249                 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
250                         mlock_vma_page(new);
251
252                 /* No need to invalidate - it was non-present before */
253                 update_mmu_cache(vma, pvmw.address, pvmw.pte);
254         }
255
256         return true;
257 }
258
259 /*
260  * Get rid of all migration entries and replace them by
261  * references to the indicated page.
262  */
263 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
264 {
265         struct rmap_walk_control rwc = {
266                 .rmap_one = remove_migration_pte,
267                 .arg = old,
268         };
269
270         if (locked)
271                 rmap_walk_locked(new, &rwc);
272         else
273                 rmap_walk(new, &rwc);
274 }
275
276 /*
277  * Something used the pte of a page under migration. We need to
278  * get to the page and wait until migration is finished.
279  * When we return from this function the fault will be retried.
280  */
281 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
282                                 spinlock_t *ptl)
283 {
284         pte_t pte;
285         swp_entry_t entry;
286         struct page *page;
287
288         spin_lock(ptl);
289         pte = *ptep;
290         if (!is_swap_pte(pte))
291                 goto out;
292
293         entry = pte_to_swp_entry(pte);
294         if (!is_migration_entry(entry))
295                 goto out;
296
297         page = migration_entry_to_page(entry);
298
299         /*
300          * Once radix-tree replacement of page migration started, page_count
301          * *must* be zero. And, we don't want to call wait_on_page_locked()
302          * against a page without get_page().
303          * So, we use get_page_unless_zero(), here. Even failed, page fault
304          * will occur again.
305          */
306         if (!get_page_unless_zero(page))
307                 goto out;
308         pte_unmap_unlock(ptep, ptl);
309         wait_on_page_locked(page);
310         put_page(page);
311         return;
312 out:
313         pte_unmap_unlock(ptep, ptl);
314 }
315
316 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
317                                 unsigned long address)
318 {
319         spinlock_t *ptl = pte_lockptr(mm, pmd);
320         pte_t *ptep = pte_offset_map(pmd, address);
321         __migration_entry_wait(mm, ptep, ptl);
322 }
323
324 void migration_entry_wait_huge(struct vm_area_struct *vma,
325                 struct mm_struct *mm, pte_t *pte)
326 {
327         spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
328         __migration_entry_wait(mm, pte, ptl);
329 }
330
331 #ifdef CONFIG_BLOCK
332 /* Returns true if all buffers are successfully locked */
333 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
334                                                         enum migrate_mode mode)
335 {
336         struct buffer_head *bh = head;
337
338         /* Simple case, sync compaction */
339         if (mode != MIGRATE_ASYNC) {
340                 do {
341                         get_bh(bh);
342                         lock_buffer(bh);
343                         bh = bh->b_this_page;
344
345                 } while (bh != head);
346
347                 return true;
348         }
349
350         /* async case, we cannot block on lock_buffer so use trylock_buffer */
351         do {
352                 get_bh(bh);
353                 if (!trylock_buffer(bh)) {
354                         /*
355                          * We failed to lock the buffer and cannot stall in
356                          * async migration. Release the taken locks
357                          */
358                         struct buffer_head *failed_bh = bh;
359                         put_bh(failed_bh);
360                         bh = head;
361                         while (bh != failed_bh) {
362                                 unlock_buffer(bh);
363                                 put_bh(bh);
364                                 bh = bh->b_this_page;
365                         }
366                         return false;
367                 }
368
369                 bh = bh->b_this_page;
370         } while (bh != head);
371         return true;
372 }
373 #else
374 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
375                                                         enum migrate_mode mode)
376 {
377         return true;
378 }
379 #endif /* CONFIG_BLOCK */
380
381 /*
382  * Replace the page in the mapping.
383  *
384  * The number of remaining references must be:
385  * 1 for anonymous pages without a mapping
386  * 2 for pages with a mapping
387  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
388  */
389 int migrate_page_move_mapping(struct address_space *mapping,
390                 struct page *newpage, struct page *page,
391                 struct buffer_head *head, enum migrate_mode mode,
392                 int extra_count)
393 {
394         struct zone *oldzone, *newzone;
395         int dirty;
396         int expected_count = 1 + extra_count;
397         void **pslot;
398
399         if (!mapping) {
400                 /* Anonymous page without mapping */
401                 if (page_count(page) != expected_count)
402                         return -EAGAIN;
403
404                 /* No turning back from here */
405                 newpage->index = page->index;
406                 newpage->mapping = page->mapping;
407                 if (PageSwapBacked(page))
408                         __SetPageSwapBacked(newpage);
409
410                 return MIGRATEPAGE_SUCCESS;
411         }
412
413         oldzone = page_zone(page);
414         newzone = page_zone(newpage);
415
416         spin_lock_irq(&mapping->tree_lock);
417
418         pslot = radix_tree_lookup_slot(&mapping->page_tree,
419                                         page_index(page));
420
421         expected_count += 1 + page_has_private(page);
422         if (page_count(page) != expected_count ||
423                 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
424                 spin_unlock_irq(&mapping->tree_lock);
425                 return -EAGAIN;
426         }
427
428         if (!page_ref_freeze(page, expected_count)) {
429                 spin_unlock_irq(&mapping->tree_lock);
430                 return -EAGAIN;
431         }
432
433         /*
434          * In the async migration case of moving a page with buffers, lock the
435          * buffers using trylock before the mapping is moved. If the mapping
436          * was moved, we later failed to lock the buffers and could not move
437          * the mapping back due to an elevated page count, we would have to
438          * block waiting on other references to be dropped.
439          */
440         if (mode == MIGRATE_ASYNC && head &&
441                         !buffer_migrate_lock_buffers(head, mode)) {
442                 page_ref_unfreeze(page, expected_count);
443                 spin_unlock_irq(&mapping->tree_lock);
444                 return -EAGAIN;
445         }
446
447         /*
448          * Now we know that no one else is looking at the page:
449          * no turning back from here.
450          */
451         newpage->index = page->index;
452         newpage->mapping = page->mapping;
453         get_page(newpage);      /* add cache reference */
454         if (PageSwapBacked(page)) {
455                 __SetPageSwapBacked(newpage);
456                 if (PageSwapCache(page)) {
457                         SetPageSwapCache(newpage);
458                         set_page_private(newpage, page_private(page));
459                 }
460         } else {
461                 VM_BUG_ON_PAGE(PageSwapCache(page), page);
462         }
463
464         /* Move dirty while page refs frozen and newpage not yet exposed */
465         dirty = PageDirty(page);
466         if (dirty) {
467                 ClearPageDirty(page);
468                 SetPageDirty(newpage);
469         }
470
471         radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
472
473         /*
474          * Drop cache reference from old page by unfreezing
475          * to one less reference.
476          * We know this isn't the last reference.
477          */
478         page_ref_unfreeze(page, expected_count - 1);
479
480         spin_unlock(&mapping->tree_lock);
481         /* Leave irq disabled to prevent preemption while updating stats */
482
483         /*
484          * If moved to a different zone then also account
485          * the page for that zone. Other VM counters will be
486          * taken care of when we establish references to the
487          * new page and drop references to the old page.
488          *
489          * Note that anonymous pages are accounted for
490          * via NR_FILE_PAGES and NR_ANON_MAPPED if they
491          * are mapped to swap space.
492          */
493         if (newzone != oldzone) {
494                 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
495                 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
496                 if (PageSwapBacked(page) && !PageSwapCache(page)) {
497                         __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
498                         __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
499                 }
500                 if (dirty && mapping_cap_account_dirty(mapping)) {
501                         __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
502                         __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
503                         __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
504                         __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
505                 }
506         }
507         local_irq_enable();
508
509         return MIGRATEPAGE_SUCCESS;
510 }
511 EXPORT_SYMBOL(migrate_page_move_mapping);
512
513 /*
514  * The expected number of remaining references is the same as that
515  * of migrate_page_move_mapping().
516  */
517 int migrate_huge_page_move_mapping(struct address_space *mapping,
518                                    struct page *newpage, struct page *page)
519 {
520         int expected_count;
521         void **pslot;
522
523         spin_lock_irq(&mapping->tree_lock);
524
525         pslot = radix_tree_lookup_slot(&mapping->page_tree,
526                                         page_index(page));
527
528         expected_count = 2 + page_has_private(page);
529         if (page_count(page) != expected_count ||
530                 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
531                 spin_unlock_irq(&mapping->tree_lock);
532                 return -EAGAIN;
533         }
534
535         if (!page_ref_freeze(page, expected_count)) {
536                 spin_unlock_irq(&mapping->tree_lock);
537                 return -EAGAIN;
538         }
539
540         newpage->index = page->index;
541         newpage->mapping = page->mapping;
542
543         get_page(newpage);
544
545         radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
546
547         page_ref_unfreeze(page, expected_count - 1);
548
549         spin_unlock_irq(&mapping->tree_lock);
550
551         return MIGRATEPAGE_SUCCESS;
552 }
553
554 /*
555  * Gigantic pages are so large that we do not guarantee that page++ pointer
556  * arithmetic will work across the entire page.  We need something more
557  * specialized.
558  */
559 static void __copy_gigantic_page(struct page *dst, struct page *src,
560                                 int nr_pages)
561 {
562         int i;
563         struct page *dst_base = dst;
564         struct page *src_base = src;
565
566         for (i = 0; i < nr_pages; ) {
567                 cond_resched();
568                 copy_highpage(dst, src);
569
570                 i++;
571                 dst = mem_map_next(dst, dst_base, i);
572                 src = mem_map_next(src, src_base, i);
573         }
574 }
575
576 static void copy_huge_page(struct page *dst, struct page *src)
577 {
578         int i;
579         int nr_pages;
580
581         if (PageHuge(src)) {
582                 /* hugetlbfs page */
583                 struct hstate *h = page_hstate(src);
584                 nr_pages = pages_per_huge_page(h);
585
586                 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
587                         __copy_gigantic_page(dst, src, nr_pages);
588                         return;
589                 }
590         } else {
591                 /* thp page */
592                 BUG_ON(!PageTransHuge(src));
593                 nr_pages = hpage_nr_pages(src);
594         }
595
596         for (i = 0; i < nr_pages; i++) {
597                 cond_resched();
598                 copy_highpage(dst + i, src + i);
599         }
600 }
601
602 /*
603  * Copy the page to its new location
604  */
605 void migrate_page_copy(struct page *newpage, struct page *page)
606 {
607         int cpupid;
608
609         if (PageHuge(page) || PageTransHuge(page))
610                 copy_huge_page(newpage, page);
611         else
612                 copy_highpage(newpage, page);
613
614         if (PageError(page))
615                 SetPageError(newpage);
616         if (PageReferenced(page))
617                 SetPageReferenced(newpage);
618         if (PageUptodate(page))
619                 SetPageUptodate(newpage);
620         if (TestClearPageActive(page)) {
621                 VM_BUG_ON_PAGE(PageUnevictable(page), page);
622                 SetPageActive(newpage);
623         } else if (TestClearPageUnevictable(page))
624                 SetPageUnevictable(newpage);
625         if (PageChecked(page))
626                 SetPageChecked(newpage);
627         if (PageMappedToDisk(page))
628                 SetPageMappedToDisk(newpage);
629
630         /* Move dirty on pages not done by migrate_page_move_mapping() */
631         if (PageDirty(page))
632                 SetPageDirty(newpage);
633
634         if (page_is_young(page))
635                 set_page_young(newpage);
636         if (page_is_idle(page))
637                 set_page_idle(newpage);
638
639         /*
640          * Copy NUMA information to the new page, to prevent over-eager
641          * future migrations of this same page.
642          */
643         cpupid = page_cpupid_xchg_last(page, -1);
644         page_cpupid_xchg_last(newpage, cpupid);
645
646         ksm_migrate_page(newpage, page);
647         /*
648          * Please do not reorder this without considering how mm/ksm.c's
649          * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
650          */
651         if (PageSwapCache(page))
652                 ClearPageSwapCache(page);
653         ClearPagePrivate(page);
654         set_page_private(page, 0);
655
656         /*
657          * If any waiters have accumulated on the new page then
658          * wake them up.
659          */
660         if (PageWriteback(newpage))
661                 end_page_writeback(newpage);
662
663         copy_page_owner(page, newpage);
664
665         mem_cgroup_migrate(page, newpage);
666 }
667 EXPORT_SYMBOL(migrate_page_copy);
668
669 /************************************************************
670  *                    Migration functions
671  ***********************************************************/
672
673 /*
674  * Common logic to directly migrate a single LRU page suitable for
675  * pages that do not use PagePrivate/PagePrivate2.
676  *
677  * Pages are locked upon entry and exit.
678  */
679 int migrate_page(struct address_space *mapping,
680                 struct page *newpage, struct page *page,
681                 enum migrate_mode mode)
682 {
683         int rc;
684
685         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
686
687         rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
688
689         if (rc != MIGRATEPAGE_SUCCESS)
690                 return rc;
691
692         migrate_page_copy(newpage, page);
693         return MIGRATEPAGE_SUCCESS;
694 }
695 EXPORT_SYMBOL(migrate_page);
696
697 #ifdef CONFIG_BLOCK
698 /*
699  * Migration function for pages with buffers. This function can only be used
700  * if the underlying filesystem guarantees that no other references to "page"
701  * exist.
702  */
703 int buffer_migrate_page(struct address_space *mapping,
704                 struct page *newpage, struct page *page, enum migrate_mode mode)
705 {
706         struct buffer_head *bh, *head;
707         int rc;
708
709         if (!page_has_buffers(page))
710                 return migrate_page(mapping, newpage, page, mode);
711
712         head = page_buffers(page);
713
714         rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
715
716         if (rc != MIGRATEPAGE_SUCCESS)
717                 return rc;
718
719         /*
720          * In the async case, migrate_page_move_mapping locked the buffers
721          * with an IRQ-safe spinlock held. In the sync case, the buffers
722          * need to be locked now
723          */
724         if (mode != MIGRATE_ASYNC)
725                 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
726
727         ClearPagePrivate(page);
728         set_page_private(newpage, page_private(page));
729         set_page_private(page, 0);
730         put_page(page);
731         get_page(newpage);
732
733         bh = head;
734         do {
735                 set_bh_page(bh, newpage, bh_offset(bh));
736                 bh = bh->b_this_page;
737
738         } while (bh != head);
739
740         SetPagePrivate(newpage);
741
742         migrate_page_copy(newpage, page);
743
744         bh = head;
745         do {
746                 unlock_buffer(bh);
747                 put_bh(bh);
748                 bh = bh->b_this_page;
749
750         } while (bh != head);
751
752         return MIGRATEPAGE_SUCCESS;
753 }
754 EXPORT_SYMBOL(buffer_migrate_page);
755 #endif
756
757 /*
758  * Writeback a page to clean the dirty state
759  */
760 static int writeout(struct address_space *mapping, struct page *page)
761 {
762         struct writeback_control wbc = {
763                 .sync_mode = WB_SYNC_NONE,
764                 .nr_to_write = 1,
765                 .range_start = 0,
766                 .range_end = LLONG_MAX,
767                 .for_reclaim = 1
768         };
769         int rc;
770
771         if (!mapping->a_ops->writepage)
772                 /* No write method for the address space */
773                 return -EINVAL;
774
775         if (!clear_page_dirty_for_io(page))
776                 /* Someone else already triggered a write */
777                 return -EAGAIN;
778
779         /*
780          * A dirty page may imply that the underlying filesystem has
781          * the page on some queue. So the page must be clean for
782          * migration. Writeout may mean we loose the lock and the
783          * page state is no longer what we checked for earlier.
784          * At this point we know that the migration attempt cannot
785          * be successful.
786          */
787         remove_migration_ptes(page, page, false);
788
789         rc = mapping->a_ops->writepage(page, &wbc);
790
791         if (rc != AOP_WRITEPAGE_ACTIVATE)
792                 /* unlocked. Relock */
793                 lock_page(page);
794
795         return (rc < 0) ? -EIO : -EAGAIN;
796 }
797
798 /*
799  * Default handling if a filesystem does not provide a migration function.
800  */
801 static int fallback_migrate_page(struct address_space *mapping,
802         struct page *newpage, struct page *page, enum migrate_mode mode)
803 {
804         if (PageDirty(page)) {
805                 /* Only writeback pages in full synchronous migration */
806                 if (mode != MIGRATE_SYNC)
807                         return -EBUSY;
808                 return writeout(mapping, page);
809         }
810
811         /*
812          * Buffers may be managed in a filesystem specific way.
813          * We must have no buffers or drop them.
814          */
815         if (page_has_private(page) &&
816             !try_to_release_page(page, GFP_KERNEL))
817                 return -EAGAIN;
818
819         return migrate_page(mapping, newpage, page, mode);
820 }
821
822 /*
823  * Move a page to a newly allocated page
824  * The page is locked and all ptes have been successfully removed.
825  *
826  * The new page will have replaced the old page if this function
827  * is successful.
828  *
829  * Return value:
830  *   < 0 - error code
831  *  MIGRATEPAGE_SUCCESS - success
832  */
833 static int move_to_new_page(struct page *newpage, struct page *page,
834                                 enum migrate_mode mode)
835 {
836         struct address_space *mapping;
837         int rc = -EAGAIN;
838         bool is_lru = !__PageMovable(page);
839
840         VM_BUG_ON_PAGE(!PageLocked(page), page);
841         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
842
843         mapping = page_mapping(page);
844
845         if (likely(is_lru)) {
846                 if (!mapping)
847                         rc = migrate_page(mapping, newpage, page, mode);
848                 else if (mapping->a_ops->migratepage)
849                         /*
850                          * Most pages have a mapping and most filesystems
851                          * provide a migratepage callback. Anonymous pages
852                          * are part of swap space which also has its own
853                          * migratepage callback. This is the most common path
854                          * for page migration.
855                          */
856                         rc = mapping->a_ops->migratepage(mapping, newpage,
857                                                         page, mode);
858                 else
859                         rc = fallback_migrate_page(mapping, newpage,
860                                                         page, mode);
861         } else {
862                 /*
863                  * In case of non-lru page, it could be released after
864                  * isolation step. In that case, we shouldn't try migration.
865                  */
866                 VM_BUG_ON_PAGE(!PageIsolated(page), page);
867                 if (!PageMovable(page)) {
868                         rc = MIGRATEPAGE_SUCCESS;
869                         __ClearPageIsolated(page);
870                         goto out;
871                 }
872
873                 rc = mapping->a_ops->migratepage(mapping, newpage,
874                                                 page, mode);
875                 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
876                         !PageIsolated(page));
877         }
878
879         /*
880          * When successful, old pagecache page->mapping must be cleared before
881          * page is freed; but stats require that PageAnon be left as PageAnon.
882          */
883         if (rc == MIGRATEPAGE_SUCCESS) {
884                 if (__PageMovable(page)) {
885                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
886
887                         /*
888                          * We clear PG_movable under page_lock so any compactor
889                          * cannot try to migrate this page.
890                          */
891                         __ClearPageIsolated(page);
892                 }
893
894                 /*
895                  * Anonymous and movable page->mapping will be cleard by
896                  * free_pages_prepare so don't reset it here for keeping
897                  * the type to work PageAnon, for example.
898                  */
899                 if (!PageMappingFlags(page))
900                         page->mapping = NULL;
901         }
902 out:
903         return rc;
904 }
905
906 static int __unmap_and_move(struct page *page, struct page *newpage,
907                                 int force, enum migrate_mode mode)
908 {
909         int rc = -EAGAIN;
910         int page_was_mapped = 0;
911         struct anon_vma *anon_vma = NULL;
912         bool is_lru = !__PageMovable(page);
913
914         if (!trylock_page(page)) {
915                 if (!force || mode == MIGRATE_ASYNC)
916                         goto out;
917
918                 /*
919                  * It's not safe for direct compaction to call lock_page.
920                  * For example, during page readahead pages are added locked
921                  * to the LRU. Later, when the IO completes the pages are
922                  * marked uptodate and unlocked. However, the queueing
923                  * could be merging multiple pages for one bio (e.g.
924                  * mpage_readpages). If an allocation happens for the
925                  * second or third page, the process can end up locking
926                  * the same page twice and deadlocking. Rather than
927                  * trying to be clever about what pages can be locked,
928                  * avoid the use of lock_page for direct compaction
929                  * altogether.
930                  */
931                 if (current->flags & PF_MEMALLOC)
932                         goto out;
933
934                 lock_page(page);
935         }
936
937         if (PageWriteback(page)) {
938                 /*
939                  * Only in the case of a full synchronous migration is it
940                  * necessary to wait for PageWriteback. In the async case,
941                  * the retry loop is too short and in the sync-light case,
942                  * the overhead of stalling is too much
943                  */
944                 if (mode != MIGRATE_SYNC) {
945                         rc = -EBUSY;
946                         goto out_unlock;
947                 }
948                 if (!force)
949                         goto out_unlock;
950                 wait_on_page_writeback(page);
951         }
952
953         /*
954          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
955          * we cannot notice that anon_vma is freed while we migrates a page.
956          * This get_anon_vma() delays freeing anon_vma pointer until the end
957          * of migration. File cache pages are no problem because of page_lock()
958          * File Caches may use write_page() or lock_page() in migration, then,
959          * just care Anon page here.
960          *
961          * Only page_get_anon_vma() understands the subtleties of
962          * getting a hold on an anon_vma from outside one of its mms.
963          * But if we cannot get anon_vma, then we won't need it anyway,
964          * because that implies that the anon page is no longer mapped
965          * (and cannot be remapped so long as we hold the page lock).
966          */
967         if (PageAnon(page) && !PageKsm(page))
968                 anon_vma = page_get_anon_vma(page);
969
970         /*
971          * Block others from accessing the new page when we get around to
972          * establishing additional references. We are usually the only one
973          * holding a reference to newpage at this point. We used to have a BUG
974          * here if trylock_page(newpage) fails, but would like to allow for
975          * cases where there might be a race with the previous use of newpage.
976          * This is much like races on refcount of oldpage: just don't BUG().
977          */
978         if (unlikely(!trylock_page(newpage)))
979                 goto out_unlock;
980
981         if (unlikely(!is_lru)) {
982                 rc = move_to_new_page(newpage, page, mode);
983                 goto out_unlock_both;
984         }
985
986         /*
987          * Corner case handling:
988          * 1. When a new swap-cache page is read into, it is added to the LRU
989          * and treated as swapcache but it has no rmap yet.
990          * Calling try_to_unmap() against a page->mapping==NULL page will
991          * trigger a BUG.  So handle it here.
992          * 2. An orphaned page (see truncate_complete_page) might have
993          * fs-private metadata. The page can be picked up due to memory
994          * offlining.  Everywhere else except page reclaim, the page is
995          * invisible to the vm, so the page can not be migrated.  So try to
996          * free the metadata, so the page can be freed.
997          */
998         if (!page->mapping) {
999                 VM_BUG_ON_PAGE(PageAnon(page), page);
1000                 if (page_has_private(page)) {
1001                         try_to_free_buffers(page);
1002                         goto out_unlock_both;
1003                 }
1004         } else if (page_mapped(page)) {
1005                 /* Establish migration ptes */
1006                 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1007                                 page);
1008                 try_to_unmap(page,
1009                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1010                 page_was_mapped = 1;
1011         }
1012
1013         if (!page_mapped(page))
1014                 rc = move_to_new_page(newpage, page, mode);
1015
1016         if (page_was_mapped)
1017                 remove_migration_ptes(page,
1018                         rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1019
1020 out_unlock_both:
1021         unlock_page(newpage);
1022 out_unlock:
1023         /* Drop an anon_vma reference if we took one */
1024         if (anon_vma)
1025                 put_anon_vma(anon_vma);
1026         unlock_page(page);
1027 out:
1028         /*
1029          * If migration is successful, decrease refcount of the newpage
1030          * which will not free the page because new page owner increased
1031          * refcounter. As well, if it is LRU page, add the page to LRU
1032          * list in here.
1033          */
1034         if (rc == MIGRATEPAGE_SUCCESS) {
1035                 if (unlikely(__PageMovable(newpage)))
1036                         put_page(newpage);
1037                 else
1038                         putback_lru_page(newpage);
1039         }
1040
1041         return rc;
1042 }
1043
1044 /*
1045  * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move().  Work
1046  * around it.
1047  */
1048 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1049 #define ICE_noinline noinline
1050 #else
1051 #define ICE_noinline
1052 #endif
1053
1054 /*
1055  * Obtain the lock on page, remove all ptes and migrate the page
1056  * to the newly allocated page in newpage.
1057  */
1058 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1059                                    free_page_t put_new_page,
1060                                    unsigned long private, struct page *page,
1061                                    int force, enum migrate_mode mode,
1062                                    enum migrate_reason reason)
1063 {
1064         int rc = MIGRATEPAGE_SUCCESS;
1065         int *result = NULL;
1066         struct page *newpage;
1067
1068         newpage = get_new_page(page, private, &result);
1069         if (!newpage)
1070                 return -ENOMEM;
1071
1072         if (page_count(page) == 1) {
1073                 /* page was freed from under us. So we are done. */
1074                 ClearPageActive(page);
1075                 ClearPageUnevictable(page);
1076                 if (unlikely(__PageMovable(page))) {
1077                         lock_page(page);
1078                         if (!PageMovable(page))
1079                                 __ClearPageIsolated(page);
1080                         unlock_page(page);
1081                 }
1082                 if (put_new_page)
1083                         put_new_page(newpage, private);
1084                 else
1085                         put_page(newpage);
1086                 goto out;
1087         }
1088
1089         if (unlikely(PageTransHuge(page))) {
1090                 lock_page(page);
1091                 rc = split_huge_page(page);
1092                 unlock_page(page);
1093                 if (rc)
1094                         goto out;
1095         }
1096
1097         rc = __unmap_and_move(page, newpage, force, mode);
1098         if (rc == MIGRATEPAGE_SUCCESS)
1099                 set_page_owner_migrate_reason(newpage, reason);
1100
1101 out:
1102         if (rc != -EAGAIN) {
1103                 /*
1104                  * A page that has been migrated has all references
1105                  * removed and will be freed. A page that has not been
1106                  * migrated will have kepts its references and be
1107                  * restored.
1108                  */
1109                 list_del(&page->lru);
1110
1111                 /*
1112                  * Compaction can migrate also non-LRU pages which are
1113                  * not accounted to NR_ISOLATED_*. They can be recognized
1114                  * as __PageMovable
1115                  */
1116                 if (likely(!__PageMovable(page)))
1117                         dec_node_page_state(page, NR_ISOLATED_ANON +
1118                                         page_is_file_cache(page));
1119         }
1120
1121         /*
1122          * If migration is successful, releases reference grabbed during
1123          * isolation. Otherwise, restore the page to right list unless
1124          * we want to retry.
1125          */
1126         if (rc == MIGRATEPAGE_SUCCESS) {
1127                 put_page(page);
1128                 if (reason == MR_MEMORY_FAILURE) {
1129                         /*
1130                          * Set PG_HWPoison on just freed page
1131                          * intentionally. Although it's rather weird,
1132                          * it's how HWPoison flag works at the moment.
1133                          */
1134                         if (!test_set_page_hwpoison(page))
1135                                 num_poisoned_pages_inc();
1136                 }
1137         } else {
1138                 if (rc != -EAGAIN) {
1139                         if (likely(!__PageMovable(page))) {
1140                                 putback_lru_page(page);
1141                                 goto put_new;
1142                         }
1143
1144                         lock_page(page);
1145                         if (PageMovable(page))
1146                                 putback_movable_page(page);
1147                         else
1148                                 __ClearPageIsolated(page);
1149                         unlock_page(page);
1150                         put_page(page);
1151                 }
1152 put_new:
1153                 if (put_new_page)
1154                         put_new_page(newpage, private);
1155                 else
1156                         put_page(newpage);
1157         }
1158
1159         if (result) {
1160                 if (rc)
1161                         *result = rc;
1162                 else
1163                         *result = page_to_nid(newpage);
1164         }
1165         return rc;
1166 }
1167
1168 /*
1169  * Counterpart of unmap_and_move_page() for hugepage migration.
1170  *
1171  * This function doesn't wait the completion of hugepage I/O
1172  * because there is no race between I/O and migration for hugepage.
1173  * Note that currently hugepage I/O occurs only in direct I/O
1174  * where no lock is held and PG_writeback is irrelevant,
1175  * and writeback status of all subpages are counted in the reference
1176  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1177  * under direct I/O, the reference of the head page is 512 and a bit more.)
1178  * This means that when we try to migrate hugepage whose subpages are
1179  * doing direct I/O, some references remain after try_to_unmap() and
1180  * hugepage migration fails without data corruption.
1181  *
1182  * There is also no race when direct I/O is issued on the page under migration,
1183  * because then pte is replaced with migration swap entry and direct I/O code
1184  * will wait in the page fault for migration to complete.
1185  */
1186 static int unmap_and_move_huge_page(new_page_t get_new_page,
1187                                 free_page_t put_new_page, unsigned long private,
1188                                 struct page *hpage, int force,
1189                                 enum migrate_mode mode, int reason)
1190 {
1191         int rc = -EAGAIN;
1192         int *result = NULL;
1193         int page_was_mapped = 0;
1194         struct page *new_hpage;
1195         struct anon_vma *anon_vma = NULL;
1196
1197         /*
1198          * Movability of hugepages depends on architectures and hugepage size.
1199          * This check is necessary because some callers of hugepage migration
1200          * like soft offline and memory hotremove don't walk through page
1201          * tables or check whether the hugepage is pmd-based or not before
1202          * kicking migration.
1203          */
1204         if (!hugepage_migration_supported(page_hstate(hpage))) {
1205                 putback_active_hugepage(hpage);
1206                 return -ENOSYS;
1207         }
1208
1209         new_hpage = get_new_page(hpage, private, &result);
1210         if (!new_hpage)
1211                 return -ENOMEM;
1212
1213         if (!trylock_page(hpage)) {
1214                 if (!force || mode != MIGRATE_SYNC)
1215                         goto out;
1216                 lock_page(hpage);
1217         }
1218
1219         if (PageAnon(hpage))
1220                 anon_vma = page_get_anon_vma(hpage);
1221
1222         if (unlikely(!trylock_page(new_hpage)))
1223                 goto put_anon;
1224
1225         if (page_mapped(hpage)) {
1226                 try_to_unmap(hpage,
1227                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1228                 page_was_mapped = 1;
1229         }
1230
1231         if (!page_mapped(hpage))
1232                 rc = move_to_new_page(new_hpage, hpage, mode);
1233
1234         if (page_was_mapped)
1235                 remove_migration_ptes(hpage,
1236                         rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1237
1238         unlock_page(new_hpage);
1239
1240 put_anon:
1241         if (anon_vma)
1242                 put_anon_vma(anon_vma);
1243
1244         if (rc == MIGRATEPAGE_SUCCESS) {
1245                 hugetlb_cgroup_migrate(hpage, new_hpage);
1246                 put_new_page = NULL;
1247                 set_page_owner_migrate_reason(new_hpage, reason);
1248         }
1249
1250         unlock_page(hpage);
1251 out:
1252         if (rc != -EAGAIN)
1253                 putback_active_hugepage(hpage);
1254
1255         /*
1256          * If migration was not successful and there's a freeing callback, use
1257          * it.  Otherwise, put_page() will drop the reference grabbed during
1258          * isolation.
1259          */
1260         if (put_new_page)
1261                 put_new_page(new_hpage, private);
1262         else
1263                 putback_active_hugepage(new_hpage);
1264
1265         if (result) {
1266                 if (rc)
1267                         *result = rc;
1268                 else
1269                         *result = page_to_nid(new_hpage);
1270         }
1271         return rc;
1272 }
1273
1274 /*
1275  * migrate_pages - migrate the pages specified in a list, to the free pages
1276  *                 supplied as the target for the page migration
1277  *
1278  * @from:               The list of pages to be migrated.
1279  * @get_new_page:       The function used to allocate free pages to be used
1280  *                      as the target of the page migration.
1281  * @put_new_page:       The function used to free target pages if migration
1282  *                      fails, or NULL if no special handling is necessary.
1283  * @private:            Private data to be passed on to get_new_page()
1284  * @mode:               The migration mode that specifies the constraints for
1285  *                      page migration, if any.
1286  * @reason:             The reason for page migration.
1287  *
1288  * The function returns after 10 attempts or if no pages are movable any more
1289  * because the list has become empty or no retryable pages exist any more.
1290  * The caller should call putback_movable_pages() to return pages to the LRU
1291  * or free list only if ret != 0.
1292  *
1293  * Returns the number of pages that were not migrated, or an error code.
1294  */
1295 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1296                 free_page_t put_new_page, unsigned long private,
1297                 enum migrate_mode mode, int reason)
1298 {
1299         int retry = 1;
1300         int nr_failed = 0;
1301         int nr_succeeded = 0;
1302         int pass = 0;
1303         struct page *page;
1304         struct page *page2;
1305         int swapwrite = current->flags & PF_SWAPWRITE;
1306         int rc;
1307
1308         if (!swapwrite)
1309                 current->flags |= PF_SWAPWRITE;
1310
1311         for(pass = 0; pass < 10 && retry; pass++) {
1312                 retry = 0;
1313
1314                 list_for_each_entry_safe(page, page2, from, lru) {
1315                         cond_resched();
1316
1317                         if (PageHuge(page))
1318                                 rc = unmap_and_move_huge_page(get_new_page,
1319                                                 put_new_page, private, page,
1320                                                 pass > 2, mode, reason);
1321                         else
1322                                 rc = unmap_and_move(get_new_page, put_new_page,
1323                                                 private, page, pass > 2, mode,
1324                                                 reason);
1325
1326                         switch(rc) {
1327                         case -ENOMEM:
1328                                 nr_failed++;
1329                                 goto out;
1330                         case -EAGAIN:
1331                                 retry++;
1332                                 break;
1333                         case MIGRATEPAGE_SUCCESS:
1334                                 nr_succeeded++;
1335                                 break;
1336                         default:
1337                                 /*
1338                                  * Permanent failure (-EBUSY, -ENOSYS, etc.):
1339                                  * unlike -EAGAIN case, the failed page is
1340                                  * removed from migration page list and not
1341                                  * retried in the next outer loop.
1342                                  */
1343                                 nr_failed++;
1344                                 break;
1345                         }
1346                 }
1347         }
1348         nr_failed += retry;
1349         rc = nr_failed;
1350 out:
1351         if (nr_succeeded)
1352                 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1353         if (nr_failed)
1354                 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1355         trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1356
1357         if (!swapwrite)
1358                 current->flags &= ~PF_SWAPWRITE;
1359
1360         return rc;
1361 }
1362
1363 #ifdef CONFIG_NUMA
1364 /*
1365  * Move a list of individual pages
1366  */
1367 struct page_to_node {
1368         unsigned long addr;
1369         struct page *page;
1370         int node;
1371         int status;
1372 };
1373
1374 static struct page *new_page_node(struct page *p, unsigned long private,
1375                 int **result)
1376 {
1377         struct page_to_node *pm = (struct page_to_node *)private;
1378
1379         while (pm->node != MAX_NUMNODES && pm->page != p)
1380                 pm++;
1381
1382         if (pm->node == MAX_NUMNODES)
1383                 return NULL;
1384
1385         *result = &pm->status;
1386
1387         if (PageHuge(p))
1388                 return alloc_huge_page_node(page_hstate(compound_head(p)),
1389                                         pm->node);
1390         else
1391                 return __alloc_pages_node(pm->node,
1392                                 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1393 }
1394
1395 /*
1396  * Move a set of pages as indicated in the pm array. The addr
1397  * field must be set to the virtual address of the page to be moved
1398  * and the node number must contain a valid target node.
1399  * The pm array ends with node = MAX_NUMNODES.
1400  */
1401 static int do_move_page_to_node_array(struct mm_struct *mm,
1402                                       struct page_to_node *pm,
1403                                       int migrate_all)
1404 {
1405         int err;
1406         struct page_to_node *pp;
1407         LIST_HEAD(pagelist);
1408
1409         down_read(&mm->mmap_sem);
1410
1411         /*
1412          * Build a list of pages to migrate
1413          */
1414         for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1415                 struct vm_area_struct *vma;
1416                 struct page *page;
1417
1418                 err = -EFAULT;
1419                 vma = find_vma(mm, pp->addr);
1420                 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1421                         goto set_status;
1422
1423                 /* FOLL_DUMP to ignore special (like zero) pages */
1424                 page = follow_page(vma, pp->addr,
1425                                 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1426
1427                 err = PTR_ERR(page);
1428                 if (IS_ERR(page))
1429                         goto set_status;
1430
1431                 err = -ENOENT;
1432                 if (!page)
1433                         goto set_status;
1434
1435                 pp->page = page;
1436                 err = page_to_nid(page);
1437
1438                 if (err == pp->node)
1439                         /*
1440                          * Node already in the right place
1441                          */
1442                         goto put_and_set;
1443
1444                 err = -EACCES;
1445                 if (page_mapcount(page) > 1 &&
1446                                 !migrate_all)
1447                         goto put_and_set;
1448
1449                 if (PageHuge(page)) {
1450                         if (PageHead(page))
1451                                 isolate_huge_page(page, &pagelist);
1452                         goto put_and_set;
1453                 }
1454
1455                 err = isolate_lru_page(page);
1456                 if (!err) {
1457                         list_add_tail(&page->lru, &pagelist);
1458                         inc_node_page_state(page, NR_ISOLATED_ANON +
1459                                             page_is_file_cache(page));
1460                 }
1461 put_and_set:
1462                 /*
1463                  * Either remove the duplicate refcount from
1464                  * isolate_lru_page() or drop the page ref if it was
1465                  * not isolated.
1466                  */
1467                 put_page(page);
1468 set_status:
1469                 pp->status = err;
1470         }
1471
1472         err = 0;
1473         if (!list_empty(&pagelist)) {
1474                 err = migrate_pages(&pagelist, new_page_node, NULL,
1475                                 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1476                 if (err)
1477                         putback_movable_pages(&pagelist);
1478         }
1479
1480         up_read(&mm->mmap_sem);
1481         return err;
1482 }
1483
1484 /*
1485  * Migrate an array of page address onto an array of nodes and fill
1486  * the corresponding array of status.
1487  */
1488 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1489                          unsigned long nr_pages,
1490                          const void __user * __user *pages,
1491                          const int __user *nodes,
1492                          int __user *status, int flags)
1493 {
1494         struct page_to_node *pm;
1495         unsigned long chunk_nr_pages;
1496         unsigned long chunk_start;
1497         int err;
1498
1499         err = -ENOMEM;
1500         pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1501         if (!pm)
1502                 goto out;
1503
1504         migrate_prep();
1505
1506         /*
1507          * Store a chunk of page_to_node array in a page,
1508          * but keep the last one as a marker
1509          */
1510         chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1511
1512         for (chunk_start = 0;
1513              chunk_start < nr_pages;
1514              chunk_start += chunk_nr_pages) {
1515                 int j;
1516
1517                 if (chunk_start + chunk_nr_pages > nr_pages)
1518                         chunk_nr_pages = nr_pages - chunk_start;
1519
1520                 /* fill the chunk pm with addrs and nodes from user-space */
1521                 for (j = 0; j < chunk_nr_pages; j++) {
1522                         const void __user *p;
1523                         int node;
1524
1525                         err = -EFAULT;
1526                         if (get_user(p, pages + j + chunk_start))
1527                                 goto out_pm;
1528                         pm[j].addr = (unsigned long) p;
1529
1530                         if (get_user(node, nodes + j + chunk_start))
1531                                 goto out_pm;
1532
1533                         err = -ENODEV;
1534                         if (node < 0 || node >= MAX_NUMNODES)
1535                                 goto out_pm;
1536
1537                         if (!node_state(node, N_MEMORY))
1538                                 goto out_pm;
1539
1540                         err = -EACCES;
1541                         if (!node_isset(node, task_nodes))
1542                                 goto out_pm;
1543
1544                         pm[j].node = node;
1545                 }
1546
1547                 /* End marker for this chunk */
1548                 pm[chunk_nr_pages].node = MAX_NUMNODES;
1549
1550                 /* Migrate this chunk */
1551                 err = do_move_page_to_node_array(mm, pm,
1552                                                  flags & MPOL_MF_MOVE_ALL);
1553                 if (err < 0)
1554                         goto out_pm;
1555
1556                 /* Return status information */
1557                 for (j = 0; j < chunk_nr_pages; j++)
1558                         if (put_user(pm[j].status, status + j + chunk_start)) {
1559                                 err = -EFAULT;
1560                                 goto out_pm;
1561                         }
1562         }
1563         err = 0;
1564
1565 out_pm:
1566         free_page((unsigned long)pm);
1567 out:
1568         return err;
1569 }
1570
1571 /*
1572  * Determine the nodes of an array of pages and store it in an array of status.
1573  */
1574 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1575                                 const void __user **pages, int *status)
1576 {
1577         unsigned long i;
1578
1579         down_read(&mm->mmap_sem);
1580
1581         for (i = 0; i < nr_pages; i++) {
1582                 unsigned long addr = (unsigned long)(*pages);
1583                 struct vm_area_struct *vma;
1584                 struct page *page;
1585                 int err = -EFAULT;
1586
1587                 vma = find_vma(mm, addr);
1588                 if (!vma || addr < vma->vm_start)
1589                         goto set_status;
1590
1591                 /* FOLL_DUMP to ignore special (like zero) pages */
1592                 page = follow_page(vma, addr, FOLL_DUMP);
1593
1594                 err = PTR_ERR(page);
1595                 if (IS_ERR(page))
1596                         goto set_status;
1597
1598                 err = page ? page_to_nid(page) : -ENOENT;
1599 set_status:
1600                 *status = err;
1601
1602                 pages++;
1603                 status++;
1604         }
1605
1606         up_read(&mm->mmap_sem);
1607 }
1608
1609 /*
1610  * Determine the nodes of a user array of pages and store it in
1611  * a user array of status.
1612  */
1613 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1614                          const void __user * __user *pages,
1615                          int __user *status)
1616 {
1617 #define DO_PAGES_STAT_CHUNK_NR 16
1618         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1619         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1620
1621         while (nr_pages) {
1622                 unsigned long chunk_nr;
1623
1624                 chunk_nr = nr_pages;
1625                 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1626                         chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1627
1628                 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1629                         break;
1630
1631                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1632
1633                 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1634                         break;
1635
1636                 pages += chunk_nr;
1637                 status += chunk_nr;
1638                 nr_pages -= chunk_nr;
1639         }
1640         return nr_pages ? -EFAULT : 0;
1641 }
1642
1643 /*
1644  * Move a list of pages in the address space of the currently executing
1645  * process.
1646  */
1647 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1648                 const void __user * __user *, pages,
1649                 const int __user *, nodes,
1650                 int __user *, status, int, flags)
1651 {
1652         const struct cred *cred = current_cred(), *tcred;
1653         struct task_struct *task;
1654         struct mm_struct *mm;
1655         int err;
1656         nodemask_t task_nodes;
1657
1658         /* Check flags */
1659         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1660                 return -EINVAL;
1661
1662         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1663                 return -EPERM;
1664
1665         /* Find the mm_struct */
1666         rcu_read_lock();
1667         task = pid ? find_task_by_vpid(pid) : current;
1668         if (!task) {
1669                 rcu_read_unlock();
1670                 return -ESRCH;
1671         }
1672         get_task_struct(task);
1673
1674         /*
1675          * Check if this process has the right to modify the specified
1676          * process. The right exists if the process has administrative
1677          * capabilities, superuser privileges or the same
1678          * userid as the target process.
1679          */
1680         tcred = __task_cred(task);
1681         if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1682             !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1683             !capable(CAP_SYS_NICE)) {
1684                 rcu_read_unlock();
1685                 err = -EPERM;
1686                 goto out;
1687         }
1688         rcu_read_unlock();
1689
1690         err = security_task_movememory(task);
1691         if (err)
1692                 goto out;
1693
1694         task_nodes = cpuset_mems_allowed(task);
1695         mm = get_task_mm(task);
1696         put_task_struct(task);
1697
1698         if (!mm)
1699                 return -EINVAL;
1700
1701         if (nodes)
1702                 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1703                                     nodes, status, flags);
1704         else
1705                 err = do_pages_stat(mm, nr_pages, pages, status);
1706
1707         mmput(mm);
1708         return err;
1709
1710 out:
1711         put_task_struct(task);
1712         return err;
1713 }
1714
1715 #ifdef CONFIG_NUMA_BALANCING
1716 /*
1717  * Returns true if this is a safe migration target node for misplaced NUMA
1718  * pages. Currently it only checks the watermarks which crude
1719  */
1720 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1721                                    unsigned long nr_migrate_pages)
1722 {
1723         int z;
1724
1725         for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1726                 struct zone *zone = pgdat->node_zones + z;
1727
1728                 if (!populated_zone(zone))
1729                         continue;
1730
1731                 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1732                 if (!zone_watermark_ok(zone, 0,
1733                                        high_wmark_pages(zone) +
1734                                        nr_migrate_pages,
1735                                        0, 0))
1736                         continue;
1737                 return true;
1738         }
1739         return false;
1740 }
1741
1742 static struct page *alloc_misplaced_dst_page(struct page *page,
1743                                            unsigned long data,
1744                                            int **result)
1745 {
1746         int nid = (int) data;
1747         struct page *newpage;
1748
1749         newpage = __alloc_pages_node(nid,
1750                                          (GFP_HIGHUSER_MOVABLE |
1751                                           __GFP_THISNODE | __GFP_NOMEMALLOC |
1752                                           __GFP_NORETRY | __GFP_NOWARN) &
1753                                          ~__GFP_RECLAIM, 0);
1754
1755         return newpage;
1756 }
1757
1758 /*
1759  * page migration rate limiting control.
1760  * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1761  * window of time. Default here says do not migrate more than 1280M per second.
1762  */
1763 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1764 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1765
1766 /* Returns true if the node is migrate rate-limited after the update */
1767 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1768                                         unsigned long nr_pages)
1769 {
1770         /*
1771          * Rate-limit the amount of data that is being migrated to a node.
1772          * Optimal placement is no good if the memory bus is saturated and
1773          * all the time is being spent migrating!
1774          */
1775         if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1776                 spin_lock(&pgdat->numabalancing_migrate_lock);
1777                 pgdat->numabalancing_migrate_nr_pages = 0;
1778                 pgdat->numabalancing_migrate_next_window = jiffies +
1779                         msecs_to_jiffies(migrate_interval_millisecs);
1780                 spin_unlock(&pgdat->numabalancing_migrate_lock);
1781         }
1782         if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1783                 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1784                                                                 nr_pages);
1785                 return true;
1786         }
1787
1788         /*
1789          * This is an unlocked non-atomic update so errors are possible.
1790          * The consequences are failing to migrate when we potentiall should
1791          * have which is not severe enough to warrant locking. If it is ever
1792          * a problem, it can be converted to a per-cpu counter.
1793          */
1794         pgdat->numabalancing_migrate_nr_pages += nr_pages;
1795         return false;
1796 }
1797
1798 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1799 {
1800         int page_lru;
1801
1802         VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1803
1804         /* Avoid migrating to a node that is nearly full */
1805         if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1806                 return 0;
1807
1808         if (isolate_lru_page(page))
1809                 return 0;
1810
1811         /*
1812          * migrate_misplaced_transhuge_page() skips page migration's usual
1813          * check on page_count(), so we must do it here, now that the page
1814          * has been isolated: a GUP pin, or any other pin, prevents migration.
1815          * The expected page count is 3: 1 for page's mapcount and 1 for the
1816          * caller's pin and 1 for the reference taken by isolate_lru_page().
1817          */
1818         if (PageTransHuge(page) && page_count(page) != 3) {
1819                 putback_lru_page(page);
1820                 return 0;
1821         }
1822
1823         page_lru = page_is_file_cache(page);
1824         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1825                                 hpage_nr_pages(page));
1826
1827         /*
1828          * Isolating the page has taken another reference, so the
1829          * caller's reference can be safely dropped without the page
1830          * disappearing underneath us during migration.
1831          */
1832         put_page(page);
1833         return 1;
1834 }
1835
1836 bool pmd_trans_migrating(pmd_t pmd)
1837 {
1838         struct page *page = pmd_page(pmd);
1839         return PageLocked(page);
1840 }
1841
1842 /*
1843  * Attempt to migrate a misplaced page to the specified destination
1844  * node. Caller is expected to have an elevated reference count on
1845  * the page that will be dropped by this function before returning.
1846  */
1847 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1848                            int node)
1849 {
1850         pg_data_t *pgdat = NODE_DATA(node);
1851         int isolated;
1852         int nr_remaining;
1853         LIST_HEAD(migratepages);
1854
1855         /*
1856          * Don't migrate file pages that are mapped in multiple processes
1857          * with execute permissions as they are probably shared libraries.
1858          */
1859         if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1860             (vma->vm_flags & VM_EXEC))
1861                 goto out;
1862
1863         /*
1864          * Rate-limit the amount of data that is being migrated to a node.
1865          * Optimal placement is no good if the memory bus is saturated and
1866          * all the time is being spent migrating!
1867          */
1868         if (numamigrate_update_ratelimit(pgdat, 1))
1869                 goto out;
1870
1871         isolated = numamigrate_isolate_page(pgdat, page);
1872         if (!isolated)
1873                 goto out;
1874
1875         list_add(&page->lru, &migratepages);
1876         nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1877                                      NULL, node, MIGRATE_ASYNC,
1878                                      MR_NUMA_MISPLACED);
1879         if (nr_remaining) {
1880                 if (!list_empty(&migratepages)) {
1881                         list_del(&page->lru);
1882                         dec_node_page_state(page, NR_ISOLATED_ANON +
1883                                         page_is_file_cache(page));
1884                         putback_lru_page(page);
1885                 }
1886                 isolated = 0;
1887         } else
1888                 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1889         BUG_ON(!list_empty(&migratepages));
1890         return isolated;
1891
1892 out:
1893         put_page(page);
1894         return 0;
1895 }
1896 #endif /* CONFIG_NUMA_BALANCING */
1897
1898 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1899 /*
1900  * Migrates a THP to a given target node. page must be locked and is unlocked
1901  * before returning.
1902  */
1903 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1904                                 struct vm_area_struct *vma,
1905                                 pmd_t *pmd, pmd_t entry,
1906                                 unsigned long address,
1907                                 struct page *page, int node)
1908 {
1909         spinlock_t *ptl;
1910         pg_data_t *pgdat = NODE_DATA(node);
1911         int isolated = 0;
1912         struct page *new_page = NULL;
1913         int page_lru = page_is_file_cache(page);
1914         unsigned long mmun_start = address & HPAGE_PMD_MASK;
1915         unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1916         pmd_t orig_entry;
1917
1918         /*
1919          * Rate-limit the amount of data that is being migrated to a node.
1920          * Optimal placement is no good if the memory bus is saturated and
1921          * all the time is being spent migrating!
1922          */
1923         if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1924                 goto out_dropref;
1925
1926         new_page = alloc_pages_node(node,
1927                 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1928                 HPAGE_PMD_ORDER);
1929         if (!new_page)
1930                 goto out_fail;
1931         prep_transhuge_page(new_page);
1932
1933         isolated = numamigrate_isolate_page(pgdat, page);
1934         if (!isolated) {
1935                 put_page(new_page);
1936                 goto out_fail;
1937         }
1938         /*
1939          * We are not sure a pending tlb flush here is for a huge page
1940          * mapping or not. Hence use the tlb range variant
1941          */
1942         if (mm_tlb_flush_pending(mm))
1943                 flush_tlb_range(vma, mmun_start, mmun_end);
1944
1945         /* Prepare a page as a migration target */
1946         __SetPageLocked(new_page);
1947         if (PageSwapBacked(page))
1948                 __SetPageSwapBacked(new_page);
1949
1950         /* anon mapping, we can simply copy page->mapping to the new page: */
1951         new_page->mapping = page->mapping;
1952         new_page->index = page->index;
1953         migrate_page_copy(new_page, page);
1954         WARN_ON(PageLRU(new_page));
1955
1956         /* Recheck the target PMD */
1957         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1958         ptl = pmd_lock(mm, pmd);
1959         if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1960 fail_putback:
1961                 spin_unlock(ptl);
1962                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1963
1964                 /* Reverse changes made by migrate_page_copy() */
1965                 if (TestClearPageActive(new_page))
1966                         SetPageActive(page);
1967                 if (TestClearPageUnevictable(new_page))
1968                         SetPageUnevictable(page);
1969
1970                 unlock_page(new_page);
1971                 put_page(new_page);             /* Free it */
1972
1973                 /* Retake the callers reference and putback on LRU */
1974                 get_page(page);
1975                 putback_lru_page(page);
1976                 mod_node_page_state(page_pgdat(page),
1977                          NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1978
1979                 goto out_unlock;
1980         }
1981
1982         orig_entry = *pmd;
1983         entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1984         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1985
1986         /*
1987          * Clear the old entry under pagetable lock and establish the new PTE.
1988          * Any parallel GUP will either observe the old page blocking on the
1989          * page lock, block on the page table lock or observe the new page.
1990          * The SetPageUptodate on the new page and page_add_new_anon_rmap
1991          * guarantee the copy is visible before the pagetable update.
1992          */
1993         flush_cache_range(vma, mmun_start, mmun_end);
1994         page_add_anon_rmap(new_page, vma, mmun_start, true);
1995         pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1996         set_pmd_at(mm, mmun_start, pmd, entry);
1997         update_mmu_cache_pmd(vma, address, &entry);
1998
1999         if (page_count(page) != 2) {
2000                 set_pmd_at(mm, mmun_start, pmd, orig_entry);
2001                 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
2002                 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2003                 update_mmu_cache_pmd(vma, address, &entry);
2004                 page_remove_rmap(new_page, true);
2005                 goto fail_putback;
2006         }
2007
2008         mlock_migrate_page(new_page, page);
2009         page_remove_rmap(page, true);
2010         set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2011
2012         spin_unlock(ptl);
2013         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2014
2015         /* Take an "isolate" reference and put new page on the LRU. */
2016         get_page(new_page);
2017         putback_lru_page(new_page);
2018
2019         unlock_page(new_page);
2020         unlock_page(page);
2021         put_page(page);                 /* Drop the rmap reference */
2022         put_page(page);                 /* Drop the LRU isolation reference */
2023
2024         count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2025         count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2026
2027         mod_node_page_state(page_pgdat(page),
2028                         NR_ISOLATED_ANON + page_lru,
2029                         -HPAGE_PMD_NR);
2030         return isolated;
2031
2032 out_fail:
2033         count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2034 out_dropref:
2035         ptl = pmd_lock(mm, pmd);
2036         if (pmd_same(*pmd, entry)) {
2037                 entry = pmd_modify(entry, vma->vm_page_prot);
2038                 set_pmd_at(mm, mmun_start, pmd, entry);
2039                 update_mmu_cache_pmd(vma, address, &entry);
2040         }
2041         spin_unlock(ptl);
2042
2043 out_unlock:
2044         unlock_page(page);
2045         put_page(page);
2046         return 0;
2047 }
2048 #endif /* CONFIG_NUMA_BALANCING */
2049
2050 #endif /* CONFIG_NUMA */