mm: thp: fix MADV_DONTNEED vs migrate_misplaced_transhuge_page race condition
[linux-2.6-block.git] / mm / migrate.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Memory Migration functionality - linux/mm/migrate.c
4  *
5  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6  *
7  * Page migration was first developed in the context of the memory hotplug
8  * project. The main authors of the migration code are:
9  *
10  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11  * Hirokazu Takahashi <taka@valinux.co.jp>
12  * Dave Hansen <haveblue@us.ibm.com>
13  * Christoph Lameter
14  */
15
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/page_idle.h>
47 #include <linux/page_owner.h>
48 #include <linux/sched/mm.h>
49 #include <linux/ptrace.h>
50
51 #include <asm/tlbflush.h>
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/migrate.h>
55
56 #include "internal.h"
57
58 /*
59  * migrate_prep() needs to be called before we start compiling a list of pages
60  * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61  * undesirable, use migrate_prep_local()
62  */
63 int migrate_prep(void)
64 {
65         /*
66          * Clear the LRU lists so pages can be isolated.
67          * Note that pages may be moved off the LRU after we have
68          * drained them. Those pages will fail to migrate like other
69          * pages that may be busy.
70          */
71         lru_add_drain_all();
72
73         return 0;
74 }
75
76 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 int migrate_prep_local(void)
78 {
79         lru_add_drain();
80
81         return 0;
82 }
83
84 int isolate_movable_page(struct page *page, isolate_mode_t mode)
85 {
86         struct address_space *mapping;
87
88         /*
89          * Avoid burning cycles with pages that are yet under __free_pages(),
90          * or just got freed under us.
91          *
92          * In case we 'win' a race for a movable page being freed under us and
93          * raise its refcount preventing __free_pages() from doing its job
94          * the put_page() at the end of this block will take care of
95          * release this page, thus avoiding a nasty leakage.
96          */
97         if (unlikely(!get_page_unless_zero(page)))
98                 goto out;
99
100         /*
101          * Check PageMovable before holding a PG_lock because page's owner
102          * assumes anybody doesn't touch PG_lock of newly allocated page
103          * so unconditionally grapping the lock ruins page's owner side.
104          */
105         if (unlikely(!__PageMovable(page)))
106                 goto out_putpage;
107         /*
108          * As movable pages are not isolated from LRU lists, concurrent
109          * compaction threads can race against page migration functions
110          * as well as race against the releasing a page.
111          *
112          * In order to avoid having an already isolated movable page
113          * being (wrongly) re-isolated while it is under migration,
114          * or to avoid attempting to isolate pages being released,
115          * lets be sure we have the page lock
116          * before proceeding with the movable page isolation steps.
117          */
118         if (unlikely(!trylock_page(page)))
119                 goto out_putpage;
120
121         if (!PageMovable(page) || PageIsolated(page))
122                 goto out_no_isolated;
123
124         mapping = page_mapping(page);
125         VM_BUG_ON_PAGE(!mapping, page);
126
127         if (!mapping->a_ops->isolate_page(page, mode))
128                 goto out_no_isolated;
129
130         /* Driver shouldn't use PG_isolated bit of page->flags */
131         WARN_ON_ONCE(PageIsolated(page));
132         __SetPageIsolated(page);
133         unlock_page(page);
134
135         return 0;
136
137 out_no_isolated:
138         unlock_page(page);
139 out_putpage:
140         put_page(page);
141 out:
142         return -EBUSY;
143 }
144
145 /* It should be called on page which is PG_movable */
146 void putback_movable_page(struct page *page)
147 {
148         struct address_space *mapping;
149
150         VM_BUG_ON_PAGE(!PageLocked(page), page);
151         VM_BUG_ON_PAGE(!PageMovable(page), page);
152         VM_BUG_ON_PAGE(!PageIsolated(page), page);
153
154         mapping = page_mapping(page);
155         mapping->a_ops->putback_page(page);
156         __ClearPageIsolated(page);
157 }
158
159 /*
160  * Put previously isolated pages back onto the appropriate lists
161  * from where they were once taken off for compaction/migration.
162  *
163  * This function shall be used whenever the isolated pageset has been
164  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165  * and isolate_huge_page().
166  */
167 void putback_movable_pages(struct list_head *l)
168 {
169         struct page *page;
170         struct page *page2;
171
172         list_for_each_entry_safe(page, page2, l, lru) {
173                 if (unlikely(PageHuge(page))) {
174                         putback_active_hugepage(page);
175                         continue;
176                 }
177                 list_del(&page->lru);
178                 /*
179                  * We isolated non-lru movable page so here we can use
180                  * __PageMovable because LRU page's mapping cannot have
181                  * PAGE_MAPPING_MOVABLE.
182                  */
183                 if (unlikely(__PageMovable(page))) {
184                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
185                         lock_page(page);
186                         if (PageMovable(page))
187                                 putback_movable_page(page);
188                         else
189                                 __ClearPageIsolated(page);
190                         unlock_page(page);
191                         put_page(page);
192                 } else {
193                         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194                                         page_is_file_cache(page), -hpage_nr_pages(page));
195                         putback_lru_page(page);
196                 }
197         }
198 }
199
200 /*
201  * Restore a potential migration pte to a working pte entry
202  */
203 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
204                                  unsigned long addr, void *old)
205 {
206         struct page_vma_mapped_walk pvmw = {
207                 .page = old,
208                 .vma = vma,
209                 .address = addr,
210                 .flags = PVMW_SYNC | PVMW_MIGRATION,
211         };
212         struct page *new;
213         pte_t pte;
214         swp_entry_t entry;
215
216         VM_BUG_ON_PAGE(PageTail(page), page);
217         while (page_vma_mapped_walk(&pvmw)) {
218                 if (PageKsm(page))
219                         new = page;
220                 else
221                         new = page - pvmw.page->index +
222                                 linear_page_index(vma, pvmw.address);
223
224 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225                 /* PMD-mapped THP migration entry */
226                 if (!pvmw.pte) {
227                         VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228                         remove_migration_pmd(&pvmw, new);
229                         continue;
230                 }
231 #endif
232
233                 get_page(new);
234                 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235                 if (pte_swp_soft_dirty(*pvmw.pte))
236                         pte = pte_mksoft_dirty(pte);
237
238                 /*
239                  * Recheck VMA as permissions can change since migration started
240                  */
241                 entry = pte_to_swp_entry(*pvmw.pte);
242                 if (is_write_migration_entry(entry))
243                         pte = maybe_mkwrite(pte, vma);
244
245                 if (unlikely(is_zone_device_page(new))) {
246                         if (is_device_private_page(new)) {
247                                 entry = make_device_private_entry(new, pte_write(pte));
248                                 pte = swp_entry_to_pte(entry);
249                         } else if (is_device_public_page(new)) {
250                                 pte = pte_mkdevmap(pte);
251                                 flush_dcache_page(new);
252                         }
253                 } else
254                         flush_dcache_page(new);
255
256 #ifdef CONFIG_HUGETLB_PAGE
257                 if (PageHuge(new)) {
258                         pte = pte_mkhuge(pte);
259                         pte = arch_make_huge_pte(pte, vma, new, 0);
260                         set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
261                         if (PageAnon(new))
262                                 hugepage_add_anon_rmap(new, vma, pvmw.address);
263                         else
264                                 page_dup_rmap(new, true);
265                 } else
266 #endif
267                 {
268                         set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269
270                         if (PageAnon(new))
271                                 page_add_anon_rmap(new, vma, pvmw.address, false);
272                         else
273                                 page_add_file_rmap(new, false);
274                 }
275                 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276                         mlock_vma_page(new);
277
278                 if (PageTransHuge(page) && PageMlocked(page))
279                         clear_page_mlock(page);
280
281                 /* No need to invalidate - it was non-present before */
282                 update_mmu_cache(vma, pvmw.address, pvmw.pte);
283         }
284
285         return true;
286 }
287
288 /*
289  * Get rid of all migration entries and replace them by
290  * references to the indicated page.
291  */
292 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
293 {
294         struct rmap_walk_control rwc = {
295                 .rmap_one = remove_migration_pte,
296                 .arg = old,
297         };
298
299         if (locked)
300                 rmap_walk_locked(new, &rwc);
301         else
302                 rmap_walk(new, &rwc);
303 }
304
305 /*
306  * Something used the pte of a page under migration. We need to
307  * get to the page and wait until migration is finished.
308  * When we return from this function the fault will be retried.
309  */
310 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
311                                 spinlock_t *ptl)
312 {
313         pte_t pte;
314         swp_entry_t entry;
315         struct page *page;
316
317         spin_lock(ptl);
318         pte = *ptep;
319         if (!is_swap_pte(pte))
320                 goto out;
321
322         entry = pte_to_swp_entry(pte);
323         if (!is_migration_entry(entry))
324                 goto out;
325
326         page = migration_entry_to_page(entry);
327
328         /*
329          * Once radix-tree replacement of page migration started, page_count
330          * *must* be zero. And, we don't want to call wait_on_page_locked()
331          * against a page without get_page().
332          * So, we use get_page_unless_zero(), here. Even failed, page fault
333          * will occur again.
334          */
335         if (!get_page_unless_zero(page))
336                 goto out;
337         pte_unmap_unlock(ptep, ptl);
338         wait_on_page_locked(page);
339         put_page(page);
340         return;
341 out:
342         pte_unmap_unlock(ptep, ptl);
343 }
344
345 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
346                                 unsigned long address)
347 {
348         spinlock_t *ptl = pte_lockptr(mm, pmd);
349         pte_t *ptep = pte_offset_map(pmd, address);
350         __migration_entry_wait(mm, ptep, ptl);
351 }
352
353 void migration_entry_wait_huge(struct vm_area_struct *vma,
354                 struct mm_struct *mm, pte_t *pte)
355 {
356         spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
357         __migration_entry_wait(mm, pte, ptl);
358 }
359
360 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
361 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
362 {
363         spinlock_t *ptl;
364         struct page *page;
365
366         ptl = pmd_lock(mm, pmd);
367         if (!is_pmd_migration_entry(*pmd))
368                 goto unlock;
369         page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
370         if (!get_page_unless_zero(page))
371                 goto unlock;
372         spin_unlock(ptl);
373         wait_on_page_locked(page);
374         put_page(page);
375         return;
376 unlock:
377         spin_unlock(ptl);
378 }
379 #endif
380
381 #ifdef CONFIG_BLOCK
382 /* Returns true if all buffers are successfully locked */
383 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
384                                                         enum migrate_mode mode)
385 {
386         struct buffer_head *bh = head;
387
388         /* Simple case, sync compaction */
389         if (mode != MIGRATE_ASYNC) {
390                 do {
391                         get_bh(bh);
392                         lock_buffer(bh);
393                         bh = bh->b_this_page;
394
395                 } while (bh != head);
396
397                 return true;
398         }
399
400         /* async case, we cannot block on lock_buffer so use trylock_buffer */
401         do {
402                 get_bh(bh);
403                 if (!trylock_buffer(bh)) {
404                         /*
405                          * We failed to lock the buffer and cannot stall in
406                          * async migration. Release the taken locks
407                          */
408                         struct buffer_head *failed_bh = bh;
409                         put_bh(failed_bh);
410                         bh = head;
411                         while (bh != failed_bh) {
412                                 unlock_buffer(bh);
413                                 put_bh(bh);
414                                 bh = bh->b_this_page;
415                         }
416                         return false;
417                 }
418
419                 bh = bh->b_this_page;
420         } while (bh != head);
421         return true;
422 }
423 #else
424 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
425                                                         enum migrate_mode mode)
426 {
427         return true;
428 }
429 #endif /* CONFIG_BLOCK */
430
431 /*
432  * Replace the page in the mapping.
433  *
434  * The number of remaining references must be:
435  * 1 for anonymous pages without a mapping
436  * 2 for pages with a mapping
437  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
438  */
439 int migrate_page_move_mapping(struct address_space *mapping,
440                 struct page *newpage, struct page *page,
441                 struct buffer_head *head, enum migrate_mode mode,
442                 int extra_count)
443 {
444         struct zone *oldzone, *newzone;
445         int dirty;
446         int expected_count = 1 + extra_count;
447         void **pslot;
448
449         /*
450          * Device public or private pages have an extra refcount as they are
451          * ZONE_DEVICE pages.
452          */
453         expected_count += is_device_private_page(page);
454         expected_count += is_device_public_page(page);
455
456         if (!mapping) {
457                 /* Anonymous page without mapping */
458                 if (page_count(page) != expected_count)
459                         return -EAGAIN;
460
461                 /* No turning back from here */
462                 newpage->index = page->index;
463                 newpage->mapping = page->mapping;
464                 if (PageSwapBacked(page))
465                         __SetPageSwapBacked(newpage);
466
467                 return MIGRATEPAGE_SUCCESS;
468         }
469
470         oldzone = page_zone(page);
471         newzone = page_zone(newpage);
472
473         xa_lock_irq(&mapping->i_pages);
474
475         pslot = radix_tree_lookup_slot(&mapping->i_pages,
476                                         page_index(page));
477
478         expected_count += hpage_nr_pages(page) + page_has_private(page);
479         if (page_count(page) != expected_count ||
480                 radix_tree_deref_slot_protected(pslot,
481                                         &mapping->i_pages.xa_lock) != page) {
482                 xa_unlock_irq(&mapping->i_pages);
483                 return -EAGAIN;
484         }
485
486         if (!page_ref_freeze(page, expected_count)) {
487                 xa_unlock_irq(&mapping->i_pages);
488                 return -EAGAIN;
489         }
490
491         /*
492          * In the async migration case of moving a page with buffers, lock the
493          * buffers using trylock before the mapping is moved. If the mapping
494          * was moved, we later failed to lock the buffers and could not move
495          * the mapping back due to an elevated page count, we would have to
496          * block waiting on other references to be dropped.
497          */
498         if (mode == MIGRATE_ASYNC && head &&
499                         !buffer_migrate_lock_buffers(head, mode)) {
500                 page_ref_unfreeze(page, expected_count);
501                 xa_unlock_irq(&mapping->i_pages);
502                 return -EAGAIN;
503         }
504
505         /*
506          * Now we know that no one else is looking at the page:
507          * no turning back from here.
508          */
509         newpage->index = page->index;
510         newpage->mapping = page->mapping;
511         page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
512         if (PageSwapBacked(page)) {
513                 __SetPageSwapBacked(newpage);
514                 if (PageSwapCache(page)) {
515                         SetPageSwapCache(newpage);
516                         set_page_private(newpage, page_private(page));
517                 }
518         } else {
519                 VM_BUG_ON_PAGE(PageSwapCache(page), page);
520         }
521
522         /* Move dirty while page refs frozen and newpage not yet exposed */
523         dirty = PageDirty(page);
524         if (dirty) {
525                 ClearPageDirty(page);
526                 SetPageDirty(newpage);
527         }
528
529         radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
530         if (PageTransHuge(page)) {
531                 int i;
532                 int index = page_index(page);
533
534                 for (i = 1; i < HPAGE_PMD_NR; i++) {
535                         pslot = radix_tree_lookup_slot(&mapping->i_pages,
536                                                        index + i);
537                         radix_tree_replace_slot(&mapping->i_pages, pslot,
538                                                 newpage + i);
539                 }
540         }
541
542         /*
543          * Drop cache reference from old page by unfreezing
544          * to one less reference.
545          * We know this isn't the last reference.
546          */
547         page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
548
549         xa_unlock(&mapping->i_pages);
550         /* Leave irq disabled to prevent preemption while updating stats */
551
552         /*
553          * If moved to a different zone then also account
554          * the page for that zone. Other VM counters will be
555          * taken care of when we establish references to the
556          * new page and drop references to the old page.
557          *
558          * Note that anonymous pages are accounted for
559          * via NR_FILE_PAGES and NR_ANON_MAPPED if they
560          * are mapped to swap space.
561          */
562         if (newzone != oldzone) {
563                 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
564                 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
565                 if (PageSwapBacked(page) && !PageSwapCache(page)) {
566                         __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
567                         __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
568                 }
569                 if (dirty && mapping_cap_account_dirty(mapping)) {
570                         __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
571                         __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
572                         __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
573                         __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
574                 }
575         }
576         local_irq_enable();
577
578         return MIGRATEPAGE_SUCCESS;
579 }
580 EXPORT_SYMBOL(migrate_page_move_mapping);
581
582 /*
583  * The expected number of remaining references is the same as that
584  * of migrate_page_move_mapping().
585  */
586 int migrate_huge_page_move_mapping(struct address_space *mapping,
587                                    struct page *newpage, struct page *page)
588 {
589         int expected_count;
590         void **pslot;
591
592         xa_lock_irq(&mapping->i_pages);
593
594         pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
595
596         expected_count = 2 + page_has_private(page);
597         if (page_count(page) != expected_count ||
598                 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
599                 xa_unlock_irq(&mapping->i_pages);
600                 return -EAGAIN;
601         }
602
603         if (!page_ref_freeze(page, expected_count)) {
604                 xa_unlock_irq(&mapping->i_pages);
605                 return -EAGAIN;
606         }
607
608         newpage->index = page->index;
609         newpage->mapping = page->mapping;
610
611         get_page(newpage);
612
613         radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
614
615         page_ref_unfreeze(page, expected_count - 1);
616
617         xa_unlock_irq(&mapping->i_pages);
618
619         return MIGRATEPAGE_SUCCESS;
620 }
621
622 /*
623  * Gigantic pages are so large that we do not guarantee that page++ pointer
624  * arithmetic will work across the entire page.  We need something more
625  * specialized.
626  */
627 static void __copy_gigantic_page(struct page *dst, struct page *src,
628                                 int nr_pages)
629 {
630         int i;
631         struct page *dst_base = dst;
632         struct page *src_base = src;
633
634         for (i = 0; i < nr_pages; ) {
635                 cond_resched();
636                 copy_highpage(dst, src);
637
638                 i++;
639                 dst = mem_map_next(dst, dst_base, i);
640                 src = mem_map_next(src, src_base, i);
641         }
642 }
643
644 static void copy_huge_page(struct page *dst, struct page *src)
645 {
646         int i;
647         int nr_pages;
648
649         if (PageHuge(src)) {
650                 /* hugetlbfs page */
651                 struct hstate *h = page_hstate(src);
652                 nr_pages = pages_per_huge_page(h);
653
654                 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
655                         __copy_gigantic_page(dst, src, nr_pages);
656                         return;
657                 }
658         } else {
659                 /* thp page */
660                 BUG_ON(!PageTransHuge(src));
661                 nr_pages = hpage_nr_pages(src);
662         }
663
664         for (i = 0; i < nr_pages; i++) {
665                 cond_resched();
666                 copy_highpage(dst + i, src + i);
667         }
668 }
669
670 /*
671  * Copy the page to its new location
672  */
673 void migrate_page_states(struct page *newpage, struct page *page)
674 {
675         int cpupid;
676
677         if (PageError(page))
678                 SetPageError(newpage);
679         if (PageReferenced(page))
680                 SetPageReferenced(newpage);
681         if (PageUptodate(page))
682                 SetPageUptodate(newpage);
683         if (TestClearPageActive(page)) {
684                 VM_BUG_ON_PAGE(PageUnevictable(page), page);
685                 SetPageActive(newpage);
686         } else if (TestClearPageUnevictable(page))
687                 SetPageUnevictable(newpage);
688         if (PageWorkingset(page))
689                 SetPageWorkingset(newpage);
690         if (PageChecked(page))
691                 SetPageChecked(newpage);
692         if (PageMappedToDisk(page))
693                 SetPageMappedToDisk(newpage);
694
695         /* Move dirty on pages not done by migrate_page_move_mapping() */
696         if (PageDirty(page))
697                 SetPageDirty(newpage);
698
699         if (page_is_young(page))
700                 set_page_young(newpage);
701         if (page_is_idle(page))
702                 set_page_idle(newpage);
703
704         /*
705          * Copy NUMA information to the new page, to prevent over-eager
706          * future migrations of this same page.
707          */
708         cpupid = page_cpupid_xchg_last(page, -1);
709         page_cpupid_xchg_last(newpage, cpupid);
710
711         ksm_migrate_page(newpage, page);
712         /*
713          * Please do not reorder this without considering how mm/ksm.c's
714          * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
715          */
716         if (PageSwapCache(page))
717                 ClearPageSwapCache(page);
718         ClearPagePrivate(page);
719         set_page_private(page, 0);
720
721         /*
722          * If any waiters have accumulated on the new page then
723          * wake them up.
724          */
725         if (PageWriteback(newpage))
726                 end_page_writeback(newpage);
727
728         copy_page_owner(page, newpage);
729
730         mem_cgroup_migrate(page, newpage);
731 }
732 EXPORT_SYMBOL(migrate_page_states);
733
734 void migrate_page_copy(struct page *newpage, struct page *page)
735 {
736         if (PageHuge(page) || PageTransHuge(page))
737                 copy_huge_page(newpage, page);
738         else
739                 copy_highpage(newpage, page);
740
741         migrate_page_states(newpage, page);
742 }
743 EXPORT_SYMBOL(migrate_page_copy);
744
745 /************************************************************
746  *                    Migration functions
747  ***********************************************************/
748
749 /*
750  * Common logic to directly migrate a single LRU page suitable for
751  * pages that do not use PagePrivate/PagePrivate2.
752  *
753  * Pages are locked upon entry and exit.
754  */
755 int migrate_page(struct address_space *mapping,
756                 struct page *newpage, struct page *page,
757                 enum migrate_mode mode)
758 {
759         int rc;
760
761         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
762
763         rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
764
765         if (rc != MIGRATEPAGE_SUCCESS)
766                 return rc;
767
768         if (mode != MIGRATE_SYNC_NO_COPY)
769                 migrate_page_copy(newpage, page);
770         else
771                 migrate_page_states(newpage, page);
772         return MIGRATEPAGE_SUCCESS;
773 }
774 EXPORT_SYMBOL(migrate_page);
775
776 #ifdef CONFIG_BLOCK
777 /*
778  * Migration function for pages with buffers. This function can only be used
779  * if the underlying filesystem guarantees that no other references to "page"
780  * exist.
781  */
782 int buffer_migrate_page(struct address_space *mapping,
783                 struct page *newpage, struct page *page, enum migrate_mode mode)
784 {
785         struct buffer_head *bh, *head;
786         int rc;
787
788         if (!page_has_buffers(page))
789                 return migrate_page(mapping, newpage, page, mode);
790
791         head = page_buffers(page);
792
793         rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
794
795         if (rc != MIGRATEPAGE_SUCCESS)
796                 return rc;
797
798         /*
799          * In the async case, migrate_page_move_mapping locked the buffers
800          * with an IRQ-safe spinlock held. In the sync case, the buffers
801          * need to be locked now
802          */
803         if (mode != MIGRATE_ASYNC)
804                 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
805
806         ClearPagePrivate(page);
807         set_page_private(newpage, page_private(page));
808         set_page_private(page, 0);
809         put_page(page);
810         get_page(newpage);
811
812         bh = head;
813         do {
814                 set_bh_page(bh, newpage, bh_offset(bh));
815                 bh = bh->b_this_page;
816
817         } while (bh != head);
818
819         SetPagePrivate(newpage);
820
821         if (mode != MIGRATE_SYNC_NO_COPY)
822                 migrate_page_copy(newpage, page);
823         else
824                 migrate_page_states(newpage, page);
825
826         bh = head;
827         do {
828                 unlock_buffer(bh);
829                 put_bh(bh);
830                 bh = bh->b_this_page;
831
832         } while (bh != head);
833
834         return MIGRATEPAGE_SUCCESS;
835 }
836 EXPORT_SYMBOL(buffer_migrate_page);
837 #endif
838
839 /*
840  * Writeback a page to clean the dirty state
841  */
842 static int writeout(struct address_space *mapping, struct page *page)
843 {
844         struct writeback_control wbc = {
845                 .sync_mode = WB_SYNC_NONE,
846                 .nr_to_write = 1,
847                 .range_start = 0,
848                 .range_end = LLONG_MAX,
849                 .for_reclaim = 1
850         };
851         int rc;
852
853         if (!mapping->a_ops->writepage)
854                 /* No write method for the address space */
855                 return -EINVAL;
856
857         if (!clear_page_dirty_for_io(page))
858                 /* Someone else already triggered a write */
859                 return -EAGAIN;
860
861         /*
862          * A dirty page may imply that the underlying filesystem has
863          * the page on some queue. So the page must be clean for
864          * migration. Writeout may mean we loose the lock and the
865          * page state is no longer what we checked for earlier.
866          * At this point we know that the migration attempt cannot
867          * be successful.
868          */
869         remove_migration_ptes(page, page, false);
870
871         rc = mapping->a_ops->writepage(page, &wbc);
872
873         if (rc != AOP_WRITEPAGE_ACTIVATE)
874                 /* unlocked. Relock */
875                 lock_page(page);
876
877         return (rc < 0) ? -EIO : -EAGAIN;
878 }
879
880 /*
881  * Default handling if a filesystem does not provide a migration function.
882  */
883 static int fallback_migrate_page(struct address_space *mapping,
884         struct page *newpage, struct page *page, enum migrate_mode mode)
885 {
886         if (PageDirty(page)) {
887                 /* Only writeback pages in full synchronous migration */
888                 switch (mode) {
889                 case MIGRATE_SYNC:
890                 case MIGRATE_SYNC_NO_COPY:
891                         break;
892                 default:
893                         return -EBUSY;
894                 }
895                 return writeout(mapping, page);
896         }
897
898         /*
899          * Buffers may be managed in a filesystem specific way.
900          * We must have no buffers or drop them.
901          */
902         if (page_has_private(page) &&
903             !try_to_release_page(page, GFP_KERNEL))
904                 return -EAGAIN;
905
906         return migrate_page(mapping, newpage, page, mode);
907 }
908
909 /*
910  * Move a page to a newly allocated page
911  * The page is locked and all ptes have been successfully removed.
912  *
913  * The new page will have replaced the old page if this function
914  * is successful.
915  *
916  * Return value:
917  *   < 0 - error code
918  *  MIGRATEPAGE_SUCCESS - success
919  */
920 static int move_to_new_page(struct page *newpage, struct page *page,
921                                 enum migrate_mode mode)
922 {
923         struct address_space *mapping;
924         int rc = -EAGAIN;
925         bool is_lru = !__PageMovable(page);
926
927         VM_BUG_ON_PAGE(!PageLocked(page), page);
928         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
929
930         mapping = page_mapping(page);
931
932         if (likely(is_lru)) {
933                 if (!mapping)
934                         rc = migrate_page(mapping, newpage, page, mode);
935                 else if (mapping->a_ops->migratepage)
936                         /*
937                          * Most pages have a mapping and most filesystems
938                          * provide a migratepage callback. Anonymous pages
939                          * are part of swap space which also has its own
940                          * migratepage callback. This is the most common path
941                          * for page migration.
942                          */
943                         rc = mapping->a_ops->migratepage(mapping, newpage,
944                                                         page, mode);
945                 else
946                         rc = fallback_migrate_page(mapping, newpage,
947                                                         page, mode);
948         } else {
949                 /*
950                  * In case of non-lru page, it could be released after
951                  * isolation step. In that case, we shouldn't try migration.
952                  */
953                 VM_BUG_ON_PAGE(!PageIsolated(page), page);
954                 if (!PageMovable(page)) {
955                         rc = MIGRATEPAGE_SUCCESS;
956                         __ClearPageIsolated(page);
957                         goto out;
958                 }
959
960                 rc = mapping->a_ops->migratepage(mapping, newpage,
961                                                 page, mode);
962                 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
963                         !PageIsolated(page));
964         }
965
966         /*
967          * When successful, old pagecache page->mapping must be cleared before
968          * page is freed; but stats require that PageAnon be left as PageAnon.
969          */
970         if (rc == MIGRATEPAGE_SUCCESS) {
971                 if (__PageMovable(page)) {
972                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
973
974                         /*
975                          * We clear PG_movable under page_lock so any compactor
976                          * cannot try to migrate this page.
977                          */
978                         __ClearPageIsolated(page);
979                 }
980
981                 /*
982                  * Anonymous and movable page->mapping will be cleard by
983                  * free_pages_prepare so don't reset it here for keeping
984                  * the type to work PageAnon, for example.
985                  */
986                 if (!PageMappingFlags(page))
987                         page->mapping = NULL;
988         }
989 out:
990         return rc;
991 }
992
993 static int __unmap_and_move(struct page *page, struct page *newpage,
994                                 int force, enum migrate_mode mode)
995 {
996         int rc = -EAGAIN;
997         int page_was_mapped = 0;
998         struct anon_vma *anon_vma = NULL;
999         bool is_lru = !__PageMovable(page);
1000
1001         if (!trylock_page(page)) {
1002                 if (!force || mode == MIGRATE_ASYNC)
1003                         goto out;
1004
1005                 /*
1006                  * It's not safe for direct compaction to call lock_page.
1007                  * For example, during page readahead pages are added locked
1008                  * to the LRU. Later, when the IO completes the pages are
1009                  * marked uptodate and unlocked. However, the queueing
1010                  * could be merging multiple pages for one bio (e.g.
1011                  * mpage_readpages). If an allocation happens for the
1012                  * second or third page, the process can end up locking
1013                  * the same page twice and deadlocking. Rather than
1014                  * trying to be clever about what pages can be locked,
1015                  * avoid the use of lock_page for direct compaction
1016                  * altogether.
1017                  */
1018                 if (current->flags & PF_MEMALLOC)
1019                         goto out;
1020
1021                 lock_page(page);
1022         }
1023
1024         if (PageWriteback(page)) {
1025                 /*
1026                  * Only in the case of a full synchronous migration is it
1027                  * necessary to wait for PageWriteback. In the async case,
1028                  * the retry loop is too short and in the sync-light case,
1029                  * the overhead of stalling is too much
1030                  */
1031                 switch (mode) {
1032                 case MIGRATE_SYNC:
1033                 case MIGRATE_SYNC_NO_COPY:
1034                         break;
1035                 default:
1036                         rc = -EBUSY;
1037                         goto out_unlock;
1038                 }
1039                 if (!force)
1040                         goto out_unlock;
1041                 wait_on_page_writeback(page);
1042         }
1043
1044         /*
1045          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1046          * we cannot notice that anon_vma is freed while we migrates a page.
1047          * This get_anon_vma() delays freeing anon_vma pointer until the end
1048          * of migration. File cache pages are no problem because of page_lock()
1049          * File Caches may use write_page() or lock_page() in migration, then,
1050          * just care Anon page here.
1051          *
1052          * Only page_get_anon_vma() understands the subtleties of
1053          * getting a hold on an anon_vma from outside one of its mms.
1054          * But if we cannot get anon_vma, then we won't need it anyway,
1055          * because that implies that the anon page is no longer mapped
1056          * (and cannot be remapped so long as we hold the page lock).
1057          */
1058         if (PageAnon(page) && !PageKsm(page))
1059                 anon_vma = page_get_anon_vma(page);
1060
1061         /*
1062          * Block others from accessing the new page when we get around to
1063          * establishing additional references. We are usually the only one
1064          * holding a reference to newpage at this point. We used to have a BUG
1065          * here if trylock_page(newpage) fails, but would like to allow for
1066          * cases where there might be a race with the previous use of newpage.
1067          * This is much like races on refcount of oldpage: just don't BUG().
1068          */
1069         if (unlikely(!trylock_page(newpage)))
1070                 goto out_unlock;
1071
1072         if (unlikely(!is_lru)) {
1073                 rc = move_to_new_page(newpage, page, mode);
1074                 goto out_unlock_both;
1075         }
1076
1077         /*
1078          * Corner case handling:
1079          * 1. When a new swap-cache page is read into, it is added to the LRU
1080          * and treated as swapcache but it has no rmap yet.
1081          * Calling try_to_unmap() against a page->mapping==NULL page will
1082          * trigger a BUG.  So handle it here.
1083          * 2. An orphaned page (see truncate_complete_page) might have
1084          * fs-private metadata. The page can be picked up due to memory
1085          * offlining.  Everywhere else except page reclaim, the page is
1086          * invisible to the vm, so the page can not be migrated.  So try to
1087          * free the metadata, so the page can be freed.
1088          */
1089         if (!page->mapping) {
1090                 VM_BUG_ON_PAGE(PageAnon(page), page);
1091                 if (page_has_private(page)) {
1092                         try_to_free_buffers(page);
1093                         goto out_unlock_both;
1094                 }
1095         } else if (page_mapped(page)) {
1096                 /* Establish migration ptes */
1097                 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1098                                 page);
1099                 try_to_unmap(page,
1100                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1101                 page_was_mapped = 1;
1102         }
1103
1104         if (!page_mapped(page))
1105                 rc = move_to_new_page(newpage, page, mode);
1106
1107         if (page_was_mapped)
1108                 remove_migration_ptes(page,
1109                         rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1110
1111 out_unlock_both:
1112         unlock_page(newpage);
1113 out_unlock:
1114         /* Drop an anon_vma reference if we took one */
1115         if (anon_vma)
1116                 put_anon_vma(anon_vma);
1117         unlock_page(page);
1118 out:
1119         /*
1120          * If migration is successful, decrease refcount of the newpage
1121          * which will not free the page because new page owner increased
1122          * refcounter. As well, if it is LRU page, add the page to LRU
1123          * list in here.
1124          */
1125         if (rc == MIGRATEPAGE_SUCCESS) {
1126                 if (unlikely(__PageMovable(newpage)))
1127                         put_page(newpage);
1128                 else
1129                         putback_lru_page(newpage);
1130         }
1131
1132         return rc;
1133 }
1134
1135 /*
1136  * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move().  Work
1137  * around it.
1138  */
1139 #if defined(CONFIG_ARM) && \
1140         defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1141 #define ICE_noinline noinline
1142 #else
1143 #define ICE_noinline
1144 #endif
1145
1146 /*
1147  * Obtain the lock on page, remove all ptes and migrate the page
1148  * to the newly allocated page in newpage.
1149  */
1150 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1151                                    free_page_t put_new_page,
1152                                    unsigned long private, struct page *page,
1153                                    int force, enum migrate_mode mode,
1154                                    enum migrate_reason reason)
1155 {
1156         int rc = MIGRATEPAGE_SUCCESS;
1157         struct page *newpage;
1158
1159         if (!thp_migration_supported() && PageTransHuge(page))
1160                 return -ENOMEM;
1161
1162         newpage = get_new_page(page, private);
1163         if (!newpage)
1164                 return -ENOMEM;
1165
1166         if (page_count(page) == 1) {
1167                 /* page was freed from under us. So we are done. */
1168                 ClearPageActive(page);
1169                 ClearPageUnevictable(page);
1170                 if (unlikely(__PageMovable(page))) {
1171                         lock_page(page);
1172                         if (!PageMovable(page))
1173                                 __ClearPageIsolated(page);
1174                         unlock_page(page);
1175                 }
1176                 if (put_new_page)
1177                         put_new_page(newpage, private);
1178                 else
1179                         put_page(newpage);
1180                 goto out;
1181         }
1182
1183         rc = __unmap_and_move(page, newpage, force, mode);
1184         if (rc == MIGRATEPAGE_SUCCESS)
1185                 set_page_owner_migrate_reason(newpage, reason);
1186
1187 out:
1188         if (rc != -EAGAIN) {
1189                 /*
1190                  * A page that has been migrated has all references
1191                  * removed and will be freed. A page that has not been
1192                  * migrated will have kepts its references and be
1193                  * restored.
1194                  */
1195                 list_del(&page->lru);
1196
1197                 /*
1198                  * Compaction can migrate also non-LRU pages which are
1199                  * not accounted to NR_ISOLATED_*. They can be recognized
1200                  * as __PageMovable
1201                  */
1202                 if (likely(!__PageMovable(page)))
1203                         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1204                                         page_is_file_cache(page), -hpage_nr_pages(page));
1205         }
1206
1207         /*
1208          * If migration is successful, releases reference grabbed during
1209          * isolation. Otherwise, restore the page to right list unless
1210          * we want to retry.
1211          */
1212         if (rc == MIGRATEPAGE_SUCCESS) {
1213                 put_page(page);
1214                 if (reason == MR_MEMORY_FAILURE) {
1215                         /*
1216                          * Set PG_HWPoison on just freed page
1217                          * intentionally. Although it's rather weird,
1218                          * it's how HWPoison flag works at the moment.
1219                          */
1220                         if (set_hwpoison_free_buddy_page(page))
1221                                 num_poisoned_pages_inc();
1222                 }
1223         } else {
1224                 if (rc != -EAGAIN) {
1225                         if (likely(!__PageMovable(page))) {
1226                                 putback_lru_page(page);
1227                                 goto put_new;
1228                         }
1229
1230                         lock_page(page);
1231                         if (PageMovable(page))
1232                                 putback_movable_page(page);
1233                         else
1234                                 __ClearPageIsolated(page);
1235                         unlock_page(page);
1236                         put_page(page);
1237                 }
1238 put_new:
1239                 if (put_new_page)
1240                         put_new_page(newpage, private);
1241                 else
1242                         put_page(newpage);
1243         }
1244
1245         return rc;
1246 }
1247
1248 /*
1249  * Counterpart of unmap_and_move_page() for hugepage migration.
1250  *
1251  * This function doesn't wait the completion of hugepage I/O
1252  * because there is no race between I/O and migration for hugepage.
1253  * Note that currently hugepage I/O occurs only in direct I/O
1254  * where no lock is held and PG_writeback is irrelevant,
1255  * and writeback status of all subpages are counted in the reference
1256  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1257  * under direct I/O, the reference of the head page is 512 and a bit more.)
1258  * This means that when we try to migrate hugepage whose subpages are
1259  * doing direct I/O, some references remain after try_to_unmap() and
1260  * hugepage migration fails without data corruption.
1261  *
1262  * There is also no race when direct I/O is issued on the page under migration,
1263  * because then pte is replaced with migration swap entry and direct I/O code
1264  * will wait in the page fault for migration to complete.
1265  */
1266 static int unmap_and_move_huge_page(new_page_t get_new_page,
1267                                 free_page_t put_new_page, unsigned long private,
1268                                 struct page *hpage, int force,
1269                                 enum migrate_mode mode, int reason)
1270 {
1271         int rc = -EAGAIN;
1272         int page_was_mapped = 0;
1273         struct page *new_hpage;
1274         struct anon_vma *anon_vma = NULL;
1275
1276         /*
1277          * Movability of hugepages depends on architectures and hugepage size.
1278          * This check is necessary because some callers of hugepage migration
1279          * like soft offline and memory hotremove don't walk through page
1280          * tables or check whether the hugepage is pmd-based or not before
1281          * kicking migration.
1282          */
1283         if (!hugepage_migration_supported(page_hstate(hpage))) {
1284                 putback_active_hugepage(hpage);
1285                 return -ENOSYS;
1286         }
1287
1288         new_hpage = get_new_page(hpage, private);
1289         if (!new_hpage)
1290                 return -ENOMEM;
1291
1292         if (!trylock_page(hpage)) {
1293                 if (!force)
1294                         goto out;
1295                 switch (mode) {
1296                 case MIGRATE_SYNC:
1297                 case MIGRATE_SYNC_NO_COPY:
1298                         break;
1299                 default:
1300                         goto out;
1301                 }
1302                 lock_page(hpage);
1303         }
1304
1305         if (PageAnon(hpage))
1306                 anon_vma = page_get_anon_vma(hpage);
1307
1308         if (unlikely(!trylock_page(new_hpage)))
1309                 goto put_anon;
1310
1311         if (page_mapped(hpage)) {
1312                 try_to_unmap(hpage,
1313                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1314                 page_was_mapped = 1;
1315         }
1316
1317         if (!page_mapped(hpage))
1318                 rc = move_to_new_page(new_hpage, hpage, mode);
1319
1320         if (page_was_mapped)
1321                 remove_migration_ptes(hpage,
1322                         rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1323
1324         unlock_page(new_hpage);
1325
1326 put_anon:
1327         if (anon_vma)
1328                 put_anon_vma(anon_vma);
1329
1330         if (rc == MIGRATEPAGE_SUCCESS) {
1331                 move_hugetlb_state(hpage, new_hpage, reason);
1332                 put_new_page = NULL;
1333         }
1334
1335         unlock_page(hpage);
1336 out:
1337         if (rc != -EAGAIN)
1338                 putback_active_hugepage(hpage);
1339
1340         /*
1341          * If migration was not successful and there's a freeing callback, use
1342          * it.  Otherwise, put_page() will drop the reference grabbed during
1343          * isolation.
1344          */
1345         if (put_new_page)
1346                 put_new_page(new_hpage, private);
1347         else
1348                 putback_active_hugepage(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 retry:
1395                         cond_resched();
1396
1397                         if (PageHuge(page))
1398                                 rc = unmap_and_move_huge_page(get_new_page,
1399                                                 put_new_page, private, page,
1400                                                 pass > 2, mode, reason);
1401                         else
1402                                 rc = unmap_and_move(get_new_page, put_new_page,
1403                                                 private, page, pass > 2, mode,
1404                                                 reason);
1405
1406                         switch(rc) {
1407                         case -ENOMEM:
1408                                 /*
1409                                  * THP migration might be unsupported or the
1410                                  * allocation could've failed so we should
1411                                  * retry on the same page with the THP split
1412                                  * to base pages.
1413                                  *
1414                                  * Head page is retried immediately and tail
1415                                  * pages are added to the tail of the list so
1416                                  * we encounter them after the rest of the list
1417                                  * is processed.
1418                                  */
1419                                 if (PageTransHuge(page) && !PageHuge(page)) {
1420                                         lock_page(page);
1421                                         rc = split_huge_page_to_list(page, from);
1422                                         unlock_page(page);
1423                                         if (!rc) {
1424                                                 list_safe_reset_next(page, page2, lru);
1425                                                 goto retry;
1426                                         }
1427                                 }
1428                                 nr_failed++;
1429                                 goto out;
1430                         case -EAGAIN:
1431                                 retry++;
1432                                 break;
1433                         case MIGRATEPAGE_SUCCESS:
1434                                 nr_succeeded++;
1435                                 break;
1436                         default:
1437                                 /*
1438                                  * Permanent failure (-EBUSY, -ENOSYS, etc.):
1439                                  * unlike -EAGAIN case, the failed page is
1440                                  * removed from migration page list and not
1441                                  * retried in the next outer loop.
1442                                  */
1443                                 nr_failed++;
1444                                 break;
1445                         }
1446                 }
1447         }
1448         nr_failed += retry;
1449         rc = nr_failed;
1450 out:
1451         if (nr_succeeded)
1452                 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1453         if (nr_failed)
1454                 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1455         trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1456
1457         if (!swapwrite)
1458                 current->flags &= ~PF_SWAPWRITE;
1459
1460         return rc;
1461 }
1462
1463 #ifdef CONFIG_NUMA
1464
1465 static int store_status(int __user *status, int start, int value, int nr)
1466 {
1467         while (nr-- > 0) {
1468                 if (put_user(value, status + start))
1469                         return -EFAULT;
1470                 start++;
1471         }
1472
1473         return 0;
1474 }
1475
1476 static int do_move_pages_to_node(struct mm_struct *mm,
1477                 struct list_head *pagelist, int node)
1478 {
1479         int err;
1480
1481         if (list_empty(pagelist))
1482                 return 0;
1483
1484         err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1485                         MIGRATE_SYNC, MR_SYSCALL);
1486         if (err)
1487                 putback_movable_pages(pagelist);
1488         return err;
1489 }
1490
1491 /*
1492  * Resolves the given address to a struct page, isolates it from the LRU and
1493  * puts it to the given pagelist.
1494  * Returns -errno if the page cannot be found/isolated or 0 when it has been
1495  * queued or the page doesn't need to be migrated because it is already on
1496  * the target node
1497  */
1498 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1499                 int node, struct list_head *pagelist, bool migrate_all)
1500 {
1501         struct vm_area_struct *vma;
1502         struct page *page;
1503         unsigned int follflags;
1504         int err;
1505
1506         down_read(&mm->mmap_sem);
1507         err = -EFAULT;
1508         vma = find_vma(mm, addr);
1509         if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1510                 goto out;
1511
1512         /* FOLL_DUMP to ignore special (like zero) pages */
1513         follflags = FOLL_GET | FOLL_DUMP;
1514         page = follow_page(vma, addr, follflags);
1515
1516         err = PTR_ERR(page);
1517         if (IS_ERR(page))
1518                 goto out;
1519
1520         err = -ENOENT;
1521         if (!page)
1522                 goto out;
1523
1524         err = 0;
1525         if (page_to_nid(page) == node)
1526                 goto out_putpage;
1527
1528         err = -EACCES;
1529         if (page_mapcount(page) > 1 && !migrate_all)
1530                 goto out_putpage;
1531
1532         if (PageHuge(page)) {
1533                 if (PageHead(page)) {
1534                         isolate_huge_page(page, pagelist);
1535                         err = 0;
1536                 }
1537         } else {
1538                 struct page *head;
1539
1540                 head = compound_head(page);
1541                 err = isolate_lru_page(head);
1542                 if (err)
1543                         goto out_putpage;
1544
1545                 err = 0;
1546                 list_add_tail(&head->lru, pagelist);
1547                 mod_node_page_state(page_pgdat(head),
1548                         NR_ISOLATED_ANON + page_is_file_cache(head),
1549                         hpage_nr_pages(head));
1550         }
1551 out_putpage:
1552         /*
1553          * Either remove the duplicate refcount from
1554          * isolate_lru_page() or drop the page ref if it was
1555          * not isolated.
1556          */
1557         put_page(page);
1558 out:
1559         up_read(&mm->mmap_sem);
1560         return err;
1561 }
1562
1563 /*
1564  * Migrate an array of page address onto an array of nodes and fill
1565  * the corresponding array of status.
1566  */
1567 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1568                          unsigned long nr_pages,
1569                          const void __user * __user *pages,
1570                          const int __user *nodes,
1571                          int __user *status, int flags)
1572 {
1573         int current_node = NUMA_NO_NODE;
1574         LIST_HEAD(pagelist);
1575         int start, i;
1576         int err = 0, err1;
1577
1578         migrate_prep();
1579
1580         for (i = start = 0; i < nr_pages; i++) {
1581                 const void __user *p;
1582                 unsigned long addr;
1583                 int node;
1584
1585                 err = -EFAULT;
1586                 if (get_user(p, pages + i))
1587                         goto out_flush;
1588                 if (get_user(node, nodes + i))
1589                         goto out_flush;
1590                 addr = (unsigned long)p;
1591
1592                 err = -ENODEV;
1593                 if (node < 0 || node >= MAX_NUMNODES)
1594                         goto out_flush;
1595                 if (!node_state(node, N_MEMORY))
1596                         goto out_flush;
1597
1598                 err = -EACCES;
1599                 if (!node_isset(node, task_nodes))
1600                         goto out_flush;
1601
1602                 if (current_node == NUMA_NO_NODE) {
1603                         current_node = node;
1604                         start = i;
1605                 } else if (node != current_node) {
1606                         err = do_move_pages_to_node(mm, &pagelist, current_node);
1607                         if (err)
1608                                 goto out;
1609                         err = store_status(status, start, current_node, i - start);
1610                         if (err)
1611                                 goto out;
1612                         start = i;
1613                         current_node = node;
1614                 }
1615
1616                 /*
1617                  * Errors in the page lookup or isolation are not fatal and we simply
1618                  * report them via status
1619                  */
1620                 err = add_page_for_migration(mm, addr, current_node,
1621                                 &pagelist, flags & MPOL_MF_MOVE_ALL);
1622                 if (!err)
1623                         continue;
1624
1625                 err = store_status(status, i, err, 1);
1626                 if (err)
1627                         goto out_flush;
1628
1629                 err = do_move_pages_to_node(mm, &pagelist, current_node);
1630                 if (err)
1631                         goto out;
1632                 if (i > start) {
1633                         err = store_status(status, start, current_node, i - start);
1634                         if (err)
1635                                 goto out;
1636                 }
1637                 current_node = NUMA_NO_NODE;
1638         }
1639 out_flush:
1640         if (list_empty(&pagelist))
1641                 return err;
1642
1643         /* Make sure we do not overwrite the existing error */
1644         err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1645         if (!err1)
1646                 err1 = store_status(status, start, current_node, i - start);
1647         if (!err)
1648                 err = err1;
1649 out:
1650         return err;
1651 }
1652
1653 /*
1654  * Determine the nodes of an array of pages and store it in an array of status.
1655  */
1656 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1657                                 const void __user **pages, int *status)
1658 {
1659         unsigned long i;
1660
1661         down_read(&mm->mmap_sem);
1662
1663         for (i = 0; i < nr_pages; i++) {
1664                 unsigned long addr = (unsigned long)(*pages);
1665                 struct vm_area_struct *vma;
1666                 struct page *page;
1667                 int err = -EFAULT;
1668
1669                 vma = find_vma(mm, addr);
1670                 if (!vma || addr < vma->vm_start)
1671                         goto set_status;
1672
1673                 /* FOLL_DUMP to ignore special (like zero) pages */
1674                 page = follow_page(vma, addr, FOLL_DUMP);
1675
1676                 err = PTR_ERR(page);
1677                 if (IS_ERR(page))
1678                         goto set_status;
1679
1680                 err = page ? page_to_nid(page) : -ENOENT;
1681 set_status:
1682                 *status = err;
1683
1684                 pages++;
1685                 status++;
1686         }
1687
1688         up_read(&mm->mmap_sem);
1689 }
1690
1691 /*
1692  * Determine the nodes of a user array of pages and store it in
1693  * a user array of status.
1694  */
1695 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1696                          const void __user * __user *pages,
1697                          int __user *status)
1698 {
1699 #define DO_PAGES_STAT_CHUNK_NR 16
1700         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1701         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1702
1703         while (nr_pages) {
1704                 unsigned long chunk_nr;
1705
1706                 chunk_nr = nr_pages;
1707                 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1708                         chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1709
1710                 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1711                         break;
1712
1713                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1714
1715                 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1716                         break;
1717
1718                 pages += chunk_nr;
1719                 status += chunk_nr;
1720                 nr_pages -= chunk_nr;
1721         }
1722         return nr_pages ? -EFAULT : 0;
1723 }
1724
1725 /*
1726  * Move a list of pages in the address space of the currently executing
1727  * process.
1728  */
1729 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1730                              const void __user * __user *pages,
1731                              const int __user *nodes,
1732                              int __user *status, int flags)
1733 {
1734         struct task_struct *task;
1735         struct mm_struct *mm;
1736         int err;
1737         nodemask_t task_nodes;
1738
1739         /* Check flags */
1740         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1741                 return -EINVAL;
1742
1743         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1744                 return -EPERM;
1745
1746         /* Find the mm_struct */
1747         rcu_read_lock();
1748         task = pid ? find_task_by_vpid(pid) : current;
1749         if (!task) {
1750                 rcu_read_unlock();
1751                 return -ESRCH;
1752         }
1753         get_task_struct(task);
1754
1755         /*
1756          * Check if this process has the right to modify the specified
1757          * process. Use the regular "ptrace_may_access()" checks.
1758          */
1759         if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1760                 rcu_read_unlock();
1761                 err = -EPERM;
1762                 goto out;
1763         }
1764         rcu_read_unlock();
1765
1766         err = security_task_movememory(task);
1767         if (err)
1768                 goto out;
1769
1770         task_nodes = cpuset_mems_allowed(task);
1771         mm = get_task_mm(task);
1772         put_task_struct(task);
1773
1774         if (!mm)
1775                 return -EINVAL;
1776
1777         if (nodes)
1778                 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1779                                     nodes, status, flags);
1780         else
1781                 err = do_pages_stat(mm, nr_pages, pages, status);
1782
1783         mmput(mm);
1784         return err;
1785
1786 out:
1787         put_task_struct(task);
1788         return err;
1789 }
1790
1791 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1792                 const void __user * __user *, pages,
1793                 const int __user *, nodes,
1794                 int __user *, status, int, flags)
1795 {
1796         return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1797 }
1798
1799 #ifdef CONFIG_COMPAT
1800 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1801                        compat_uptr_t __user *, pages32,
1802                        const int __user *, nodes,
1803                        int __user *, status,
1804                        int, flags)
1805 {
1806         const void __user * __user *pages;
1807         int i;
1808
1809         pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1810         for (i = 0; i < nr_pages; i++) {
1811                 compat_uptr_t p;
1812
1813                 if (get_user(p, pages32 + i) ||
1814                         put_user(compat_ptr(p), pages + i))
1815                         return -EFAULT;
1816         }
1817         return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1818 }
1819 #endif /* CONFIG_COMPAT */
1820
1821 #ifdef CONFIG_NUMA_BALANCING
1822 /*
1823  * Returns true if this is a safe migration target node for misplaced NUMA
1824  * pages. Currently it only checks the watermarks which crude
1825  */
1826 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1827                                    unsigned long nr_migrate_pages)
1828 {
1829         int z;
1830
1831         for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1832                 struct zone *zone = pgdat->node_zones + z;
1833
1834                 if (!populated_zone(zone))
1835                         continue;
1836
1837                 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1838                 if (!zone_watermark_ok(zone, 0,
1839                                        high_wmark_pages(zone) +
1840                                        nr_migrate_pages,
1841                                        0, 0))
1842                         continue;
1843                 return true;
1844         }
1845         return false;
1846 }
1847
1848 static struct page *alloc_misplaced_dst_page(struct page *page,
1849                                            unsigned long data)
1850 {
1851         int nid = (int) data;
1852         struct page *newpage;
1853
1854         newpage = __alloc_pages_node(nid,
1855                                          (GFP_HIGHUSER_MOVABLE |
1856                                           __GFP_THISNODE | __GFP_NOMEMALLOC |
1857                                           __GFP_NORETRY | __GFP_NOWARN) &
1858                                          ~__GFP_RECLAIM, 0);
1859
1860         return newpage;
1861 }
1862
1863 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1864 {
1865         int page_lru;
1866
1867         VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1868
1869         /* Avoid migrating to a node that is nearly full */
1870         if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1871                 return 0;
1872
1873         if (isolate_lru_page(page))
1874                 return 0;
1875
1876         /*
1877          * migrate_misplaced_transhuge_page() skips page migration's usual
1878          * check on page_count(), so we must do it here, now that the page
1879          * has been isolated: a GUP pin, or any other pin, prevents migration.
1880          * The expected page count is 3: 1 for page's mapcount and 1 for the
1881          * caller's pin and 1 for the reference taken by isolate_lru_page().
1882          */
1883         if (PageTransHuge(page) && page_count(page) != 3) {
1884                 putback_lru_page(page);
1885                 return 0;
1886         }
1887
1888         page_lru = page_is_file_cache(page);
1889         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1890                                 hpage_nr_pages(page));
1891
1892         /*
1893          * Isolating the page has taken another reference, so the
1894          * caller's reference can be safely dropped without the page
1895          * disappearing underneath us during migration.
1896          */
1897         put_page(page);
1898         return 1;
1899 }
1900
1901 bool pmd_trans_migrating(pmd_t pmd)
1902 {
1903         struct page *page = pmd_page(pmd);
1904         return PageLocked(page);
1905 }
1906
1907 /*
1908  * Attempt to migrate a misplaced page to the specified destination
1909  * node. Caller is expected to have an elevated reference count on
1910  * the page that will be dropped by this function before returning.
1911  */
1912 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1913                            int node)
1914 {
1915         pg_data_t *pgdat = NODE_DATA(node);
1916         int isolated;
1917         int nr_remaining;
1918         LIST_HEAD(migratepages);
1919
1920         /*
1921          * Don't migrate file pages that are mapped in multiple processes
1922          * with execute permissions as they are probably shared libraries.
1923          */
1924         if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1925             (vma->vm_flags & VM_EXEC))
1926                 goto out;
1927
1928         /*
1929          * Also do not migrate dirty pages as not all filesystems can move
1930          * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1931          */
1932         if (page_is_file_cache(page) && PageDirty(page))
1933                 goto out;
1934
1935         isolated = numamigrate_isolate_page(pgdat, page);
1936         if (!isolated)
1937                 goto out;
1938
1939         list_add(&page->lru, &migratepages);
1940         nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1941                                      NULL, node, MIGRATE_ASYNC,
1942                                      MR_NUMA_MISPLACED);
1943         if (nr_remaining) {
1944                 if (!list_empty(&migratepages)) {
1945                         list_del(&page->lru);
1946                         dec_node_page_state(page, NR_ISOLATED_ANON +
1947                                         page_is_file_cache(page));
1948                         putback_lru_page(page);
1949                 }
1950                 isolated = 0;
1951         } else
1952                 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1953         BUG_ON(!list_empty(&migratepages));
1954         return isolated;
1955
1956 out:
1957         put_page(page);
1958         return 0;
1959 }
1960 #endif /* CONFIG_NUMA_BALANCING */
1961
1962 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1963 /*
1964  * Migrates a THP to a given target node. page must be locked and is unlocked
1965  * before returning.
1966  */
1967 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1968                                 struct vm_area_struct *vma,
1969                                 pmd_t *pmd, pmd_t entry,
1970                                 unsigned long address,
1971                                 struct page *page, int node)
1972 {
1973         spinlock_t *ptl;
1974         pg_data_t *pgdat = NODE_DATA(node);
1975         int isolated = 0;
1976         struct page *new_page = NULL;
1977         int page_lru = page_is_file_cache(page);
1978         unsigned long mmun_start = address & HPAGE_PMD_MASK;
1979         unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1980
1981         new_page = alloc_pages_node(node,
1982                 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1983                 HPAGE_PMD_ORDER);
1984         if (!new_page)
1985                 goto out_fail;
1986         prep_transhuge_page(new_page);
1987
1988         isolated = numamigrate_isolate_page(pgdat, page);
1989         if (!isolated) {
1990                 put_page(new_page);
1991                 goto out_fail;
1992         }
1993
1994         /* Prepare a page as a migration target */
1995         __SetPageLocked(new_page);
1996         if (PageSwapBacked(page))
1997                 __SetPageSwapBacked(new_page);
1998
1999         /* anon mapping, we can simply copy page->mapping to the new page: */
2000         new_page->mapping = page->mapping;
2001         new_page->index = page->index;
2002         migrate_page_copy(new_page, page);
2003         WARN_ON(PageLRU(new_page));
2004
2005         /* Recheck the target PMD */
2006         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2007         ptl = pmd_lock(mm, pmd);
2008         if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2009                 spin_unlock(ptl);
2010                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2011
2012                 /* Reverse changes made by migrate_page_copy() */
2013                 if (TestClearPageActive(new_page))
2014                         SetPageActive(page);
2015                 if (TestClearPageUnevictable(new_page))
2016                         SetPageUnevictable(page);
2017
2018                 unlock_page(new_page);
2019                 put_page(new_page);             /* Free it */
2020
2021                 /* Retake the callers reference and putback on LRU */
2022                 get_page(page);
2023                 putback_lru_page(page);
2024                 mod_node_page_state(page_pgdat(page),
2025                          NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2026
2027                 goto out_unlock;
2028         }
2029
2030         entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2031         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2032
2033         /*
2034          * Overwrite the old entry under pagetable lock and establish
2035          * the new PTE. Any parallel GUP will either observe the old
2036          * page blocking on the page lock, block on the page table
2037          * lock or observe the new page. The SetPageUptodate on the
2038          * new page and page_add_new_anon_rmap guarantee the copy is
2039          * visible before the pagetable update.
2040          */
2041         flush_cache_range(vma, mmun_start, mmun_end);
2042         page_add_anon_rmap(new_page, vma, mmun_start, true);
2043         /*
2044          * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2045          * has already been flushed globally.  So no TLB can be currently
2046          * caching this non present pmd mapping.  There's no need to clear the
2047          * pmd before doing set_pmd_at(), nor to flush the TLB after
2048          * set_pmd_at().  Clearing the pmd here would introduce a race
2049          * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2050          * mmap_sem for reading.  If the pmd is set to NULL at any given time,
2051          * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2052          * pmd.
2053          */
2054         set_pmd_at(mm, mmun_start, pmd, entry);
2055         update_mmu_cache_pmd(vma, address, &entry);
2056
2057         page_ref_unfreeze(page, 2);
2058         mlock_migrate_page(new_page, page);
2059         page_remove_rmap(page, true);
2060         set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2061
2062         spin_unlock(ptl);
2063         /*
2064          * No need to double call mmu_notifier->invalidate_range() callback as
2065          * the above pmdp_huge_clear_flush_notify() did already call it.
2066          */
2067         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2068
2069         /* Take an "isolate" reference and put new page on the LRU. */
2070         get_page(new_page);
2071         putback_lru_page(new_page);
2072
2073         unlock_page(new_page);
2074         unlock_page(page);
2075         put_page(page);                 /* Drop the rmap reference */
2076         put_page(page);                 /* Drop the LRU isolation reference */
2077
2078         count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2079         count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2080
2081         mod_node_page_state(page_pgdat(page),
2082                         NR_ISOLATED_ANON + page_lru,
2083                         -HPAGE_PMD_NR);
2084         return isolated;
2085
2086 out_fail:
2087         count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2088         ptl = pmd_lock(mm, pmd);
2089         if (pmd_same(*pmd, entry)) {
2090                 entry = pmd_modify(entry, vma->vm_page_prot);
2091                 set_pmd_at(mm, mmun_start, pmd, entry);
2092                 update_mmu_cache_pmd(vma, address, &entry);
2093         }
2094         spin_unlock(ptl);
2095
2096 out_unlock:
2097         unlock_page(page);
2098         put_page(page);
2099         return 0;
2100 }
2101 #endif /* CONFIG_NUMA_BALANCING */
2102
2103 #endif /* CONFIG_NUMA */
2104
2105 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2106 struct migrate_vma {
2107         struct vm_area_struct   *vma;
2108         unsigned long           *dst;
2109         unsigned long           *src;
2110         unsigned long           cpages;
2111         unsigned long           npages;
2112         unsigned long           start;
2113         unsigned long           end;
2114 };
2115
2116 static int migrate_vma_collect_hole(unsigned long start,
2117                                     unsigned long end,
2118                                     struct mm_walk *walk)
2119 {
2120         struct migrate_vma *migrate = walk->private;
2121         unsigned long addr;
2122
2123         for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2124                 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2125                 migrate->dst[migrate->npages] = 0;
2126                 migrate->npages++;
2127                 migrate->cpages++;
2128         }
2129
2130         return 0;
2131 }
2132
2133 static int migrate_vma_collect_skip(unsigned long start,
2134                                     unsigned long end,
2135                                     struct mm_walk *walk)
2136 {
2137         struct migrate_vma *migrate = walk->private;
2138         unsigned long addr;
2139
2140         for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2141                 migrate->dst[migrate->npages] = 0;
2142                 migrate->src[migrate->npages++] = 0;
2143         }
2144
2145         return 0;
2146 }
2147
2148 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2149                                    unsigned long start,
2150                                    unsigned long end,
2151                                    struct mm_walk *walk)
2152 {
2153         struct migrate_vma *migrate = walk->private;
2154         struct vm_area_struct *vma = walk->vma;
2155         struct mm_struct *mm = vma->vm_mm;
2156         unsigned long addr = start, unmapped = 0;
2157         spinlock_t *ptl;
2158         pte_t *ptep;
2159
2160 again:
2161         if (pmd_none(*pmdp))
2162                 return migrate_vma_collect_hole(start, end, walk);
2163
2164         if (pmd_trans_huge(*pmdp)) {
2165                 struct page *page;
2166
2167                 ptl = pmd_lock(mm, pmdp);
2168                 if (unlikely(!pmd_trans_huge(*pmdp))) {
2169                         spin_unlock(ptl);
2170                         goto again;
2171                 }
2172
2173                 page = pmd_page(*pmdp);
2174                 if (is_huge_zero_page(page)) {
2175                         spin_unlock(ptl);
2176                         split_huge_pmd(vma, pmdp, addr);
2177                         if (pmd_trans_unstable(pmdp))
2178                                 return migrate_vma_collect_skip(start, end,
2179                                                                 walk);
2180                 } else {
2181                         int ret;
2182
2183                         get_page(page);
2184                         spin_unlock(ptl);
2185                         if (unlikely(!trylock_page(page)))
2186                                 return migrate_vma_collect_skip(start, end,
2187                                                                 walk);
2188                         ret = split_huge_page(page);
2189                         unlock_page(page);
2190                         put_page(page);
2191                         if (ret)
2192                                 return migrate_vma_collect_skip(start, end,
2193                                                                 walk);
2194                         if (pmd_none(*pmdp))
2195                                 return migrate_vma_collect_hole(start, end,
2196                                                                 walk);
2197                 }
2198         }
2199
2200         if (unlikely(pmd_bad(*pmdp)))
2201                 return migrate_vma_collect_skip(start, end, walk);
2202
2203         ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2204         arch_enter_lazy_mmu_mode();
2205
2206         for (; addr < end; addr += PAGE_SIZE, ptep++) {
2207                 unsigned long mpfn, pfn;
2208                 struct page *page;
2209                 swp_entry_t entry;
2210                 pte_t pte;
2211
2212                 pte = *ptep;
2213                 pfn = pte_pfn(pte);
2214
2215                 if (pte_none(pte)) {
2216                         mpfn = MIGRATE_PFN_MIGRATE;
2217                         migrate->cpages++;
2218                         pfn = 0;
2219                         goto next;
2220                 }
2221
2222                 if (!pte_present(pte)) {
2223                         mpfn = pfn = 0;
2224
2225                         /*
2226                          * Only care about unaddressable device page special
2227                          * page table entry. Other special swap entries are not
2228                          * migratable, and we ignore regular swapped page.
2229                          */
2230                         entry = pte_to_swp_entry(pte);
2231                         if (!is_device_private_entry(entry))
2232                                 goto next;
2233
2234                         page = device_private_entry_to_page(entry);
2235                         mpfn = migrate_pfn(page_to_pfn(page))|
2236                                 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2237                         if (is_write_device_private_entry(entry))
2238                                 mpfn |= MIGRATE_PFN_WRITE;
2239                 } else {
2240                         if (is_zero_pfn(pfn)) {
2241                                 mpfn = MIGRATE_PFN_MIGRATE;
2242                                 migrate->cpages++;
2243                                 pfn = 0;
2244                                 goto next;
2245                         }
2246                         page = _vm_normal_page(migrate->vma, addr, pte, true);
2247                         mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2248                         mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2249                 }
2250
2251                 /* FIXME support THP */
2252                 if (!page || !page->mapping || PageTransCompound(page)) {
2253                         mpfn = pfn = 0;
2254                         goto next;
2255                 }
2256                 pfn = page_to_pfn(page);
2257
2258                 /*
2259                  * By getting a reference on the page we pin it and that blocks
2260                  * any kind of migration. Side effect is that it "freezes" the
2261                  * pte.
2262                  *
2263                  * We drop this reference after isolating the page from the lru
2264                  * for non device page (device page are not on the lru and thus
2265                  * can't be dropped from it).
2266                  */
2267                 get_page(page);
2268                 migrate->cpages++;
2269
2270                 /*
2271                  * Optimize for the common case where page is only mapped once
2272                  * in one process. If we can lock the page, then we can safely
2273                  * set up a special migration page table entry now.
2274                  */
2275                 if (trylock_page(page)) {
2276                         pte_t swp_pte;
2277
2278                         mpfn |= MIGRATE_PFN_LOCKED;
2279                         ptep_get_and_clear(mm, addr, ptep);
2280
2281                         /* Setup special migration page table entry */
2282                         entry = make_migration_entry(page, mpfn &
2283                                                      MIGRATE_PFN_WRITE);
2284                         swp_pte = swp_entry_to_pte(entry);
2285                         if (pte_soft_dirty(pte))
2286                                 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2287                         set_pte_at(mm, addr, ptep, swp_pte);
2288
2289                         /*
2290                          * This is like regular unmap: we remove the rmap and
2291                          * drop page refcount. Page won't be freed, as we took
2292                          * a reference just above.
2293                          */
2294                         page_remove_rmap(page, false);
2295                         put_page(page);
2296
2297                         if (pte_present(pte))
2298                                 unmapped++;
2299                 }
2300
2301 next:
2302                 migrate->dst[migrate->npages] = 0;
2303                 migrate->src[migrate->npages++] = mpfn;
2304         }
2305         arch_leave_lazy_mmu_mode();
2306         pte_unmap_unlock(ptep - 1, ptl);
2307
2308         /* Only flush the TLB if we actually modified any entries */
2309         if (unmapped)
2310                 flush_tlb_range(walk->vma, start, end);
2311
2312         return 0;
2313 }
2314
2315 /*
2316  * migrate_vma_collect() - collect pages over a range of virtual addresses
2317  * @migrate: migrate struct containing all migration information
2318  *
2319  * This will walk the CPU page table. For each virtual address backed by a
2320  * valid page, it updates the src array and takes a reference on the page, in
2321  * order to pin the page until we lock it and unmap it.
2322  */
2323 static void migrate_vma_collect(struct migrate_vma *migrate)
2324 {
2325         struct mm_walk mm_walk;
2326
2327         mm_walk.pmd_entry = migrate_vma_collect_pmd;
2328         mm_walk.pte_entry = NULL;
2329         mm_walk.pte_hole = migrate_vma_collect_hole;
2330         mm_walk.hugetlb_entry = NULL;
2331         mm_walk.test_walk = NULL;
2332         mm_walk.vma = migrate->vma;
2333         mm_walk.mm = migrate->vma->vm_mm;
2334         mm_walk.private = migrate;
2335
2336         mmu_notifier_invalidate_range_start(mm_walk.mm,
2337                                             migrate->start,
2338                                             migrate->end);
2339         walk_page_range(migrate->start, migrate->end, &mm_walk);
2340         mmu_notifier_invalidate_range_end(mm_walk.mm,
2341                                           migrate->start,
2342                                           migrate->end);
2343
2344         migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2345 }
2346
2347 /*
2348  * migrate_vma_check_page() - check if page is pinned or not
2349  * @page: struct page to check
2350  *
2351  * Pinned pages cannot be migrated. This is the same test as in
2352  * migrate_page_move_mapping(), except that here we allow migration of a
2353  * ZONE_DEVICE page.
2354  */
2355 static bool migrate_vma_check_page(struct page *page)
2356 {
2357         /*
2358          * One extra ref because caller holds an extra reference, either from
2359          * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2360          * a device page.
2361          */
2362         int extra = 1;
2363
2364         /*
2365          * FIXME support THP (transparent huge page), it is bit more complex to
2366          * check them than regular pages, because they can be mapped with a pmd
2367          * or with a pte (split pte mapping).
2368          */
2369         if (PageCompound(page))
2370                 return false;
2371
2372         /* Page from ZONE_DEVICE have one extra reference */
2373         if (is_zone_device_page(page)) {
2374                 /*
2375                  * Private page can never be pin as they have no valid pte and
2376                  * GUP will fail for those. Yet if there is a pending migration
2377                  * a thread might try to wait on the pte migration entry and
2378                  * will bump the page reference count. Sadly there is no way to
2379                  * differentiate a regular pin from migration wait. Hence to
2380                  * avoid 2 racing thread trying to migrate back to CPU to enter
2381                  * infinite loop (one stoping migration because the other is
2382                  * waiting on pte migration entry). We always return true here.
2383                  *
2384                  * FIXME proper solution is to rework migration_entry_wait() so
2385                  * it does not need to take a reference on page.
2386                  */
2387                 if (is_device_private_page(page))
2388                         return true;
2389
2390                 /*
2391                  * Only allow device public page to be migrated and account for
2392                  * the extra reference count imply by ZONE_DEVICE pages.
2393                  */
2394                 if (!is_device_public_page(page))
2395                         return false;
2396                 extra++;
2397         }
2398
2399         /* For file back page */
2400         if (page_mapping(page))
2401                 extra += 1 + page_has_private(page);
2402
2403         if ((page_count(page) - extra) > page_mapcount(page))
2404                 return false;
2405
2406         return true;
2407 }
2408
2409 /*
2410  * migrate_vma_prepare() - lock pages and isolate them from the lru
2411  * @migrate: migrate struct containing all migration information
2412  *
2413  * This locks pages that have been collected by migrate_vma_collect(). Once each
2414  * page is locked it is isolated from the lru (for non-device pages). Finally,
2415  * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2416  * migrated by concurrent kernel threads.
2417  */
2418 static void migrate_vma_prepare(struct migrate_vma *migrate)
2419 {
2420         const unsigned long npages = migrate->npages;
2421         const unsigned long start = migrate->start;
2422         unsigned long addr, i, restore = 0;
2423         bool allow_drain = true;
2424
2425         lru_add_drain();
2426
2427         for (i = 0; (i < npages) && migrate->cpages; i++) {
2428                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2429                 bool remap = true;
2430
2431                 if (!page)
2432                         continue;
2433
2434                 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2435                         /*
2436                          * Because we are migrating several pages there can be
2437                          * a deadlock between 2 concurrent migration where each
2438                          * are waiting on each other page lock.
2439                          *
2440                          * Make migrate_vma() a best effort thing and backoff
2441                          * for any page we can not lock right away.
2442                          */
2443                         if (!trylock_page(page)) {
2444                                 migrate->src[i] = 0;
2445                                 migrate->cpages--;
2446                                 put_page(page);
2447                                 continue;
2448                         }
2449                         remap = false;
2450                         migrate->src[i] |= MIGRATE_PFN_LOCKED;
2451                 }
2452
2453                 /* ZONE_DEVICE pages are not on LRU */
2454                 if (!is_zone_device_page(page)) {
2455                         if (!PageLRU(page) && allow_drain) {
2456                                 /* Drain CPU's pagevec */
2457                                 lru_add_drain_all();
2458                                 allow_drain = false;
2459                         }
2460
2461                         if (isolate_lru_page(page)) {
2462                                 if (remap) {
2463                                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2464                                         migrate->cpages--;
2465                                         restore++;
2466                                 } else {
2467                                         migrate->src[i] = 0;
2468                                         unlock_page(page);
2469                                         migrate->cpages--;
2470                                         put_page(page);
2471                                 }
2472                                 continue;
2473                         }
2474
2475                         /* Drop the reference we took in collect */
2476                         put_page(page);
2477                 }
2478
2479                 if (!migrate_vma_check_page(page)) {
2480                         if (remap) {
2481                                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2482                                 migrate->cpages--;
2483                                 restore++;
2484
2485                                 if (!is_zone_device_page(page)) {
2486                                         get_page(page);
2487                                         putback_lru_page(page);
2488                                 }
2489                         } else {
2490                                 migrate->src[i] = 0;
2491                                 unlock_page(page);
2492                                 migrate->cpages--;
2493
2494                                 if (!is_zone_device_page(page))
2495                                         putback_lru_page(page);
2496                                 else
2497                                         put_page(page);
2498                         }
2499                 }
2500         }
2501
2502         for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2503                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2504
2505                 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2506                         continue;
2507
2508                 remove_migration_pte(page, migrate->vma, addr, page);
2509
2510                 migrate->src[i] = 0;
2511                 unlock_page(page);
2512                 put_page(page);
2513                 restore--;
2514         }
2515 }
2516
2517 /*
2518  * migrate_vma_unmap() - replace page mapping with special migration pte entry
2519  * @migrate: migrate struct containing all migration information
2520  *
2521  * Replace page mapping (CPU page table pte) with a special migration pte entry
2522  * and check again if it has been pinned. Pinned pages are restored because we
2523  * cannot migrate them.
2524  *
2525  * This is the last step before we call the device driver callback to allocate
2526  * destination memory and copy contents of original page over to new page.
2527  */
2528 static void migrate_vma_unmap(struct migrate_vma *migrate)
2529 {
2530         int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2531         const unsigned long npages = migrate->npages;
2532         const unsigned long start = migrate->start;
2533         unsigned long addr, i, restore = 0;
2534
2535         for (i = 0; i < npages; i++) {
2536                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2537
2538                 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2539                         continue;
2540
2541                 if (page_mapped(page)) {
2542                         try_to_unmap(page, flags);
2543                         if (page_mapped(page))
2544                                 goto restore;
2545                 }
2546
2547                 if (migrate_vma_check_page(page))
2548                         continue;
2549
2550 restore:
2551                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2552                 migrate->cpages--;
2553                 restore++;
2554         }
2555
2556         for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2557                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2558
2559                 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2560                         continue;
2561
2562                 remove_migration_ptes(page, page, false);
2563
2564                 migrate->src[i] = 0;
2565                 unlock_page(page);
2566                 restore--;
2567
2568                 if (is_zone_device_page(page))
2569                         put_page(page);
2570                 else
2571                         putback_lru_page(page);
2572         }
2573 }
2574
2575 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2576                                     unsigned long addr,
2577                                     struct page *page,
2578                                     unsigned long *src,
2579                                     unsigned long *dst)
2580 {
2581         struct vm_area_struct *vma = migrate->vma;
2582         struct mm_struct *mm = vma->vm_mm;
2583         struct mem_cgroup *memcg;
2584         bool flush = false;
2585         spinlock_t *ptl;
2586         pte_t entry;
2587         pgd_t *pgdp;
2588         p4d_t *p4dp;
2589         pud_t *pudp;
2590         pmd_t *pmdp;
2591         pte_t *ptep;
2592
2593         /* Only allow populating anonymous memory */
2594         if (!vma_is_anonymous(vma))
2595                 goto abort;
2596
2597         pgdp = pgd_offset(mm, addr);
2598         p4dp = p4d_alloc(mm, pgdp, addr);
2599         if (!p4dp)
2600                 goto abort;
2601         pudp = pud_alloc(mm, p4dp, addr);
2602         if (!pudp)
2603                 goto abort;
2604         pmdp = pmd_alloc(mm, pudp, addr);
2605         if (!pmdp)
2606                 goto abort;
2607
2608         if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2609                 goto abort;
2610
2611         /*
2612          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2613          * pte_offset_map() on pmds where a huge pmd might be created
2614          * from a different thread.
2615          *
2616          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2617          * parallel threads are excluded by other means.
2618          *
2619          * Here we only have down_read(mmap_sem).
2620          */
2621         if (pte_alloc(mm, pmdp, addr))
2622                 goto abort;
2623
2624         /* See the comment in pte_alloc_one_map() */
2625         if (unlikely(pmd_trans_unstable(pmdp)))
2626                 goto abort;
2627
2628         if (unlikely(anon_vma_prepare(vma)))
2629                 goto abort;
2630         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2631                 goto abort;
2632
2633         /*
2634          * The memory barrier inside __SetPageUptodate makes sure that
2635          * preceding stores to the page contents become visible before
2636          * the set_pte_at() write.
2637          */
2638         __SetPageUptodate(page);
2639
2640         if (is_zone_device_page(page)) {
2641                 if (is_device_private_page(page)) {
2642                         swp_entry_t swp_entry;
2643
2644                         swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2645                         entry = swp_entry_to_pte(swp_entry);
2646                 } else if (is_device_public_page(page)) {
2647                         entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2648                         if (vma->vm_flags & VM_WRITE)
2649                                 entry = pte_mkwrite(pte_mkdirty(entry));
2650                         entry = pte_mkdevmap(entry);
2651                 }
2652         } else {
2653                 entry = mk_pte(page, vma->vm_page_prot);
2654                 if (vma->vm_flags & VM_WRITE)
2655                         entry = pte_mkwrite(pte_mkdirty(entry));
2656         }
2657
2658         ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2659
2660         if (pte_present(*ptep)) {
2661                 unsigned long pfn = pte_pfn(*ptep);
2662
2663                 if (!is_zero_pfn(pfn)) {
2664                         pte_unmap_unlock(ptep, ptl);
2665                         mem_cgroup_cancel_charge(page, memcg, false);
2666                         goto abort;
2667                 }
2668                 flush = true;
2669         } else if (!pte_none(*ptep)) {
2670                 pte_unmap_unlock(ptep, ptl);
2671                 mem_cgroup_cancel_charge(page, memcg, false);
2672                 goto abort;
2673         }
2674
2675         /*
2676          * Check for usefaultfd but do not deliver the fault. Instead,
2677          * just back off.
2678          */
2679         if (userfaultfd_missing(vma)) {
2680                 pte_unmap_unlock(ptep, ptl);
2681                 mem_cgroup_cancel_charge(page, memcg, false);
2682                 goto abort;
2683         }
2684
2685         inc_mm_counter(mm, MM_ANONPAGES);
2686         page_add_new_anon_rmap(page, vma, addr, false);
2687         mem_cgroup_commit_charge(page, memcg, false, false);
2688         if (!is_zone_device_page(page))
2689                 lru_cache_add_active_or_unevictable(page, vma);
2690         get_page(page);
2691
2692         if (flush) {
2693                 flush_cache_page(vma, addr, pte_pfn(*ptep));
2694                 ptep_clear_flush_notify(vma, addr, ptep);
2695                 set_pte_at_notify(mm, addr, ptep, entry);
2696                 update_mmu_cache(vma, addr, ptep);
2697         } else {
2698                 /* No need to invalidate - it was non-present before */
2699                 set_pte_at(mm, addr, ptep, entry);
2700                 update_mmu_cache(vma, addr, ptep);
2701         }
2702
2703         pte_unmap_unlock(ptep, ptl);
2704         *src = MIGRATE_PFN_MIGRATE;
2705         return;
2706
2707 abort:
2708         *src &= ~MIGRATE_PFN_MIGRATE;
2709 }
2710
2711 /*
2712  * migrate_vma_pages() - migrate meta-data from src page to dst page
2713  * @migrate: migrate struct containing all migration information
2714  *
2715  * This migrates struct page meta-data from source struct page to destination
2716  * struct page. This effectively finishes the migration from source page to the
2717  * destination page.
2718  */
2719 static void migrate_vma_pages(struct migrate_vma *migrate)
2720 {
2721         const unsigned long npages = migrate->npages;
2722         const unsigned long start = migrate->start;
2723         struct vm_area_struct *vma = migrate->vma;
2724         struct mm_struct *mm = vma->vm_mm;
2725         unsigned long addr, i, mmu_start;
2726         bool notified = false;
2727
2728         for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2729                 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2730                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2731                 struct address_space *mapping;
2732                 int r;
2733
2734                 if (!newpage) {
2735                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2736                         continue;
2737                 }
2738
2739                 if (!page) {
2740                         if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2741                                 continue;
2742                         }
2743                         if (!notified) {
2744                                 mmu_start = addr;
2745                                 notified = true;
2746                                 mmu_notifier_invalidate_range_start(mm,
2747                                                                 mmu_start,
2748                                                                 migrate->end);
2749                         }
2750                         migrate_vma_insert_page(migrate, addr, newpage,
2751                                                 &migrate->src[i],
2752                                                 &migrate->dst[i]);
2753                         continue;
2754                 }
2755
2756                 mapping = page_mapping(page);
2757
2758                 if (is_zone_device_page(newpage)) {
2759                         if (is_device_private_page(newpage)) {
2760                                 /*
2761                                  * For now only support private anonymous when
2762                                  * migrating to un-addressable device memory.
2763                                  */
2764                                 if (mapping) {
2765                                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2766                                         continue;
2767                                 }
2768                         } else if (!is_device_public_page(newpage)) {
2769                                 /*
2770                                  * Other types of ZONE_DEVICE page are not
2771                                  * supported.
2772                                  */
2773                                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2774                                 continue;
2775                         }
2776                 }
2777
2778                 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2779                 if (r != MIGRATEPAGE_SUCCESS)
2780                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2781         }
2782
2783         /*
2784          * No need to double call mmu_notifier->invalidate_range() callback as
2785          * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2786          * did already call it.
2787          */
2788         if (notified)
2789                 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2790                                                        migrate->end);
2791 }
2792
2793 /*
2794  * migrate_vma_finalize() - restore CPU page table entry
2795  * @migrate: migrate struct containing all migration information
2796  *
2797  * This replaces the special migration pte entry with either a mapping to the
2798  * new page if migration was successful for that page, or to the original page
2799  * otherwise.
2800  *
2801  * This also unlocks the pages and puts them back on the lru, or drops the extra
2802  * refcount, for device pages.
2803  */
2804 static void migrate_vma_finalize(struct migrate_vma *migrate)
2805 {
2806         const unsigned long npages = migrate->npages;
2807         unsigned long i;
2808
2809         for (i = 0; i < npages; i++) {
2810                 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2811                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2812
2813                 if (!page) {
2814                         if (newpage) {
2815                                 unlock_page(newpage);
2816                                 put_page(newpage);
2817                         }
2818                         continue;
2819                 }
2820
2821                 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2822                         if (newpage) {
2823                                 unlock_page(newpage);
2824                                 put_page(newpage);
2825                         }
2826                         newpage = page;
2827                 }
2828
2829                 remove_migration_ptes(page, newpage, false);
2830                 unlock_page(page);
2831                 migrate->cpages--;
2832
2833                 if (is_zone_device_page(page))
2834                         put_page(page);
2835                 else
2836                         putback_lru_page(page);
2837
2838                 if (newpage != page) {
2839                         unlock_page(newpage);
2840                         if (is_zone_device_page(newpage))
2841                                 put_page(newpage);
2842                         else
2843                                 putback_lru_page(newpage);
2844                 }
2845         }
2846 }
2847
2848 /*
2849  * migrate_vma() - migrate a range of memory inside vma
2850  *
2851  * @ops: migration callback for allocating destination memory and copying
2852  * @vma: virtual memory area containing the range to be migrated
2853  * @start: start address of the range to migrate (inclusive)
2854  * @end: end address of the range to migrate (exclusive)
2855  * @src: array of hmm_pfn_t containing source pfns
2856  * @dst: array of hmm_pfn_t containing destination pfns
2857  * @private: pointer passed back to each of the callback
2858  * Returns: 0 on success, error code otherwise
2859  *
2860  * This function tries to migrate a range of memory virtual address range, using
2861  * callbacks to allocate and copy memory from source to destination. First it
2862  * collects all the pages backing each virtual address in the range, saving this
2863  * inside the src array. Then it locks those pages and unmaps them. Once the pages
2864  * are locked and unmapped, it checks whether each page is pinned or not. Pages
2865  * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2866  * in the corresponding src array entry. It then restores any pages that are
2867  * pinned, by remapping and unlocking those pages.
2868  *
2869  * At this point it calls the alloc_and_copy() callback. For documentation on
2870  * what is expected from that callback, see struct migrate_vma_ops comments in
2871  * include/linux/migrate.h
2872  *
2873  * After the alloc_and_copy() callback, this function goes over each entry in
2874  * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2875  * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2876  * then the function tries to migrate struct page information from the source
2877  * struct page to the destination struct page. If it fails to migrate the struct
2878  * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2879  * array.
2880  *
2881  * At this point all successfully migrated pages have an entry in the src
2882  * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2883  * array entry with MIGRATE_PFN_VALID flag set.
2884  *
2885  * It then calls the finalize_and_map() callback. See comments for "struct
2886  * migrate_vma_ops", in include/linux/migrate.h for details about
2887  * finalize_and_map() behavior.
2888  *
2889  * After the finalize_and_map() callback, for successfully migrated pages, this
2890  * function updates the CPU page table to point to new pages, otherwise it
2891  * restores the CPU page table to point to the original source pages.
2892  *
2893  * Function returns 0 after the above steps, even if no pages were migrated
2894  * (The function only returns an error if any of the arguments are invalid.)
2895  *
2896  * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2897  * unsigned long entries.
2898  */
2899 int migrate_vma(const struct migrate_vma_ops *ops,
2900                 struct vm_area_struct *vma,
2901                 unsigned long start,
2902                 unsigned long end,
2903                 unsigned long *src,
2904                 unsigned long *dst,
2905                 void *private)
2906 {
2907         struct migrate_vma migrate;
2908
2909         /* Sanity check the arguments */
2910         start &= PAGE_MASK;
2911         end &= PAGE_MASK;
2912         if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2913                         vma_is_dax(vma))
2914                 return -EINVAL;
2915         if (start < vma->vm_start || start >= vma->vm_end)
2916                 return -EINVAL;
2917         if (end <= vma->vm_start || end > vma->vm_end)
2918                 return -EINVAL;
2919         if (!ops || !src || !dst || start >= end)
2920                 return -EINVAL;
2921
2922         memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2923         migrate.src = src;
2924         migrate.dst = dst;
2925         migrate.start = start;
2926         migrate.npages = 0;
2927         migrate.cpages = 0;
2928         migrate.end = end;
2929         migrate.vma = vma;
2930
2931         /* Collect, and try to unmap source pages */
2932         migrate_vma_collect(&migrate);
2933         if (!migrate.cpages)
2934                 return 0;
2935
2936         /* Lock and isolate page */
2937         migrate_vma_prepare(&migrate);
2938         if (!migrate.cpages)
2939                 return 0;
2940
2941         /* Unmap pages */
2942         migrate_vma_unmap(&migrate);
2943         if (!migrate.cpages)
2944                 return 0;
2945
2946         /*
2947          * At this point pages are locked and unmapped, and thus they have
2948          * stable content and can safely be copied to destination memory that
2949          * is allocated by the callback.
2950          *
2951          * Note that migration can fail in migrate_vma_struct_page() for each
2952          * individual page.
2953          */
2954         ops->alloc_and_copy(vma, src, dst, start, end, private);
2955
2956         /* This does the real migration of struct page */
2957         migrate_vma_pages(&migrate);
2958
2959         ops->finalize_and_map(vma, src, dst, start, end, private);
2960
2961         /* Unlock and remap pages */
2962         migrate_vma_finalize(&migrate);
2963
2964         return 0;
2965 }
2966 EXPORT_SYMBOL(migrate_vma);
2967 #endif /* defined(MIGRATE_VMA_HELPER) */