Merge branch 'akpm' (patches from Andrew)
[muen/linux.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 grabbing 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                         }
252                 }
253
254 #ifdef CONFIG_HUGETLB_PAGE
255                 if (PageHuge(new)) {
256                         pte = pte_mkhuge(pte);
257                         pte = arch_make_huge_pte(pte, vma, new, 0);
258                         set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
259                         if (PageAnon(new))
260                                 hugepage_add_anon_rmap(new, vma, pvmw.address);
261                         else
262                                 page_dup_rmap(new, true);
263                 } else
264 #endif
265                 {
266                         set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
267
268                         if (PageAnon(new))
269                                 page_add_anon_rmap(new, vma, pvmw.address, false);
270                         else
271                                 page_add_file_rmap(new, false);
272                 }
273                 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
274                         mlock_vma_page(new);
275
276                 if (PageTransHuge(page) && PageMlocked(page))
277                         clear_page_mlock(page);
278
279                 /* No need to invalidate - it was non-present before */
280                 update_mmu_cache(vma, pvmw.address, pvmw.pte);
281         }
282
283         return true;
284 }
285
286 /*
287  * Get rid of all migration entries and replace them by
288  * references to the indicated page.
289  */
290 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
291 {
292         struct rmap_walk_control rwc = {
293                 .rmap_one = remove_migration_pte,
294                 .arg = old,
295         };
296
297         if (locked)
298                 rmap_walk_locked(new, &rwc);
299         else
300                 rmap_walk(new, &rwc);
301 }
302
303 /*
304  * Something used the pte of a page under migration. We need to
305  * get to the page and wait until migration is finished.
306  * When we return from this function the fault will be retried.
307  */
308 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
309                                 spinlock_t *ptl)
310 {
311         pte_t pte;
312         swp_entry_t entry;
313         struct page *page;
314
315         spin_lock(ptl);
316         pte = *ptep;
317         if (!is_swap_pte(pte))
318                 goto out;
319
320         entry = pte_to_swp_entry(pte);
321         if (!is_migration_entry(entry))
322                 goto out;
323
324         page = migration_entry_to_page(entry);
325
326         /*
327          * Once page cache replacement of page migration started, page_count
328          * is zero; but we must not call put_and_wait_on_page_locked() without
329          * a ref. Use get_page_unless_zero(), and just fault again if it fails.
330          */
331         if (!get_page_unless_zero(page))
332                 goto out;
333         pte_unmap_unlock(ptep, ptl);
334         put_and_wait_on_page_locked(page);
335         return;
336 out:
337         pte_unmap_unlock(ptep, ptl);
338 }
339
340 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
341                                 unsigned long address)
342 {
343         spinlock_t *ptl = pte_lockptr(mm, pmd);
344         pte_t *ptep = pte_offset_map(pmd, address);
345         __migration_entry_wait(mm, ptep, ptl);
346 }
347
348 void migration_entry_wait_huge(struct vm_area_struct *vma,
349                 struct mm_struct *mm, pte_t *pte)
350 {
351         spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
352         __migration_entry_wait(mm, pte, ptl);
353 }
354
355 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
356 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
357 {
358         spinlock_t *ptl;
359         struct page *page;
360
361         ptl = pmd_lock(mm, pmd);
362         if (!is_pmd_migration_entry(*pmd))
363                 goto unlock;
364         page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
365         if (!get_page_unless_zero(page))
366                 goto unlock;
367         spin_unlock(ptl);
368         put_and_wait_on_page_locked(page);
369         return;
370 unlock:
371         spin_unlock(ptl);
372 }
373 #endif
374
375 static int expected_page_refs(struct address_space *mapping, struct page *page)
376 {
377         int expected_count = 1;
378
379         /*
380          * Device public or private pages have an extra refcount as they are
381          * ZONE_DEVICE pages.
382          */
383         expected_count += is_device_private_page(page);
384         expected_count += is_device_public_page(page);
385         if (mapping)
386                 expected_count += hpage_nr_pages(page) + page_has_private(page);
387
388         return expected_count;
389 }
390
391 /*
392  * Replace the page in the mapping.
393  *
394  * The number of remaining references must be:
395  * 1 for anonymous pages without a mapping
396  * 2 for pages with a mapping
397  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
398  */
399 int migrate_page_move_mapping(struct address_space *mapping,
400                 struct page *newpage, struct page *page, enum migrate_mode mode,
401                 int extra_count)
402 {
403         XA_STATE(xas, &mapping->i_pages, page_index(page));
404         struct zone *oldzone, *newzone;
405         int dirty;
406         int expected_count = expected_page_refs(mapping, page) + extra_count;
407
408         if (!mapping) {
409                 /* Anonymous page without mapping */
410                 if (page_count(page) != expected_count)
411                         return -EAGAIN;
412
413                 /* No turning back from here */
414                 newpage->index = page->index;
415                 newpage->mapping = page->mapping;
416                 if (PageSwapBacked(page))
417                         __SetPageSwapBacked(newpage);
418
419                 return MIGRATEPAGE_SUCCESS;
420         }
421
422         oldzone = page_zone(page);
423         newzone = page_zone(newpage);
424
425         xas_lock_irq(&xas);
426         if (page_count(page) != expected_count || xas_load(&xas) != page) {
427                 xas_unlock_irq(&xas);
428                 return -EAGAIN;
429         }
430
431         if (!page_ref_freeze(page, expected_count)) {
432                 xas_unlock_irq(&xas);
433                 return -EAGAIN;
434         }
435
436         /*
437          * Now we know that no one else is looking at the page:
438          * no turning back from here.
439          */
440         newpage->index = page->index;
441         newpage->mapping = page->mapping;
442         page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
443         if (PageSwapBacked(page)) {
444                 __SetPageSwapBacked(newpage);
445                 if (PageSwapCache(page)) {
446                         SetPageSwapCache(newpage);
447                         set_page_private(newpage, page_private(page));
448                 }
449         } else {
450                 VM_BUG_ON_PAGE(PageSwapCache(page), page);
451         }
452
453         /* Move dirty while page refs frozen and newpage not yet exposed */
454         dirty = PageDirty(page);
455         if (dirty) {
456                 ClearPageDirty(page);
457                 SetPageDirty(newpage);
458         }
459
460         xas_store(&xas, newpage);
461         if (PageTransHuge(page)) {
462                 int i;
463
464                 for (i = 1; i < HPAGE_PMD_NR; i++) {
465                         xas_next(&xas);
466                         xas_store(&xas, newpage + i);
467                 }
468         }
469
470         /*
471          * Drop cache reference from old page by unfreezing
472          * to one less reference.
473          * We know this isn't the last reference.
474          */
475         page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
476
477         xas_unlock(&xas);
478         /* Leave irq disabled to prevent preemption while updating stats */
479
480         /*
481          * If moved to a different zone then also account
482          * the page for that zone. Other VM counters will be
483          * taken care of when we establish references to the
484          * new page and drop references to the old page.
485          *
486          * Note that anonymous pages are accounted for
487          * via NR_FILE_PAGES and NR_ANON_MAPPED if they
488          * are mapped to swap space.
489          */
490         if (newzone != oldzone) {
491                 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
492                 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
493                 if (PageSwapBacked(page) && !PageSwapCache(page)) {
494                         __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
495                         __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
496                 }
497                 if (dirty && mapping_cap_account_dirty(mapping)) {
498                         __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
499                         __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
500                         __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
501                         __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
502                 }
503         }
504         local_irq_enable();
505
506         return MIGRATEPAGE_SUCCESS;
507 }
508 EXPORT_SYMBOL(migrate_page_move_mapping);
509
510 /*
511  * The expected number of remaining references is the same as that
512  * of migrate_page_move_mapping().
513  */
514 int migrate_huge_page_move_mapping(struct address_space *mapping,
515                                    struct page *newpage, struct page *page)
516 {
517         XA_STATE(xas, &mapping->i_pages, page_index(page));
518         int expected_count;
519
520         xas_lock_irq(&xas);
521         expected_count = 2 + page_has_private(page);
522         if (page_count(page) != expected_count || xas_load(&xas) != page) {
523                 xas_unlock_irq(&xas);
524                 return -EAGAIN;
525         }
526
527         if (!page_ref_freeze(page, expected_count)) {
528                 xas_unlock_irq(&xas);
529                 return -EAGAIN;
530         }
531
532         newpage->index = page->index;
533         newpage->mapping = page->mapping;
534
535         get_page(newpage);
536
537         xas_store(&xas, newpage);
538
539         page_ref_unfreeze(page, expected_count - 1);
540
541         xas_unlock_irq(&xas);
542
543         return MIGRATEPAGE_SUCCESS;
544 }
545
546 /*
547  * Gigantic pages are so large that we do not guarantee that page++ pointer
548  * arithmetic will work across the entire page.  We need something more
549  * specialized.
550  */
551 static void __copy_gigantic_page(struct page *dst, struct page *src,
552                                 int nr_pages)
553 {
554         int i;
555         struct page *dst_base = dst;
556         struct page *src_base = src;
557
558         for (i = 0; i < nr_pages; ) {
559                 cond_resched();
560                 copy_highpage(dst, src);
561
562                 i++;
563                 dst = mem_map_next(dst, dst_base, i);
564                 src = mem_map_next(src, src_base, i);
565         }
566 }
567
568 static void copy_huge_page(struct page *dst, struct page *src)
569 {
570         int i;
571         int nr_pages;
572
573         if (PageHuge(src)) {
574                 /* hugetlbfs page */
575                 struct hstate *h = page_hstate(src);
576                 nr_pages = pages_per_huge_page(h);
577
578                 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
579                         __copy_gigantic_page(dst, src, nr_pages);
580                         return;
581                 }
582         } else {
583                 /* thp page */
584                 BUG_ON(!PageTransHuge(src));
585                 nr_pages = hpage_nr_pages(src);
586         }
587
588         for (i = 0; i < nr_pages; i++) {
589                 cond_resched();
590                 copy_highpage(dst + i, src + i);
591         }
592 }
593
594 /*
595  * Copy the page to its new location
596  */
597 void migrate_page_states(struct page *newpage, struct page *page)
598 {
599         int cpupid;
600
601         if (PageError(page))
602                 SetPageError(newpage);
603         if (PageReferenced(page))
604                 SetPageReferenced(newpage);
605         if (PageUptodate(page))
606                 SetPageUptodate(newpage);
607         if (TestClearPageActive(page)) {
608                 VM_BUG_ON_PAGE(PageUnevictable(page), page);
609                 SetPageActive(newpage);
610         } else if (TestClearPageUnevictable(page))
611                 SetPageUnevictable(newpage);
612         if (PageWorkingset(page))
613                 SetPageWorkingset(newpage);
614         if (PageChecked(page))
615                 SetPageChecked(newpage);
616         if (PageMappedToDisk(page))
617                 SetPageMappedToDisk(newpage);
618
619         /* Move dirty on pages not done by migrate_page_move_mapping() */
620         if (PageDirty(page))
621                 SetPageDirty(newpage);
622
623         if (page_is_young(page))
624                 set_page_young(newpage);
625         if (page_is_idle(page))
626                 set_page_idle(newpage);
627
628         /*
629          * Copy NUMA information to the new page, to prevent over-eager
630          * future migrations of this same page.
631          */
632         cpupid = page_cpupid_xchg_last(page, -1);
633         page_cpupid_xchg_last(newpage, cpupid);
634
635         ksm_migrate_page(newpage, page);
636         /*
637          * Please do not reorder this without considering how mm/ksm.c's
638          * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
639          */
640         if (PageSwapCache(page))
641                 ClearPageSwapCache(page);
642         ClearPagePrivate(page);
643         set_page_private(page, 0);
644
645         /*
646          * If any waiters have accumulated on the new page then
647          * wake them up.
648          */
649         if (PageWriteback(newpage))
650                 end_page_writeback(newpage);
651
652         copy_page_owner(page, newpage);
653
654         mem_cgroup_migrate(page, newpage);
655 }
656 EXPORT_SYMBOL(migrate_page_states);
657
658 void migrate_page_copy(struct page *newpage, struct page *page)
659 {
660         if (PageHuge(page) || PageTransHuge(page))
661                 copy_huge_page(newpage, page);
662         else
663                 copy_highpage(newpage, page);
664
665         migrate_page_states(newpage, page);
666 }
667 EXPORT_SYMBOL(migrate_page_copy);
668
669 /************************************************************
670  *                    Migration functions
671  ***********************************************************/
672
673 /*
674  * Common logic to directly migrate a single LRU page suitable for
675  * pages that do not use PagePrivate/PagePrivate2.
676  *
677  * Pages are locked upon entry and exit.
678  */
679 int migrate_page(struct address_space *mapping,
680                 struct page *newpage, struct page *page,
681                 enum migrate_mode mode)
682 {
683         int rc;
684
685         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
686
687         rc = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
688
689         if (rc != MIGRATEPAGE_SUCCESS)
690                 return rc;
691
692         if (mode != MIGRATE_SYNC_NO_COPY)
693                 migrate_page_copy(newpage, page);
694         else
695                 migrate_page_states(newpage, page);
696         return MIGRATEPAGE_SUCCESS;
697 }
698 EXPORT_SYMBOL(migrate_page);
699
700 #ifdef CONFIG_BLOCK
701 /* Returns true if all buffers are successfully locked */
702 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
703                                                         enum migrate_mode mode)
704 {
705         struct buffer_head *bh = head;
706
707         /* Simple case, sync compaction */
708         if (mode != MIGRATE_ASYNC) {
709                 do {
710                         lock_buffer(bh);
711                         bh = bh->b_this_page;
712
713                 } while (bh != head);
714
715                 return true;
716         }
717
718         /* async case, we cannot block on lock_buffer so use trylock_buffer */
719         do {
720                 if (!trylock_buffer(bh)) {
721                         /*
722                          * We failed to lock the buffer and cannot stall in
723                          * async migration. Release the taken locks
724                          */
725                         struct buffer_head *failed_bh = bh;
726                         bh = head;
727                         while (bh != failed_bh) {
728                                 unlock_buffer(bh);
729                                 bh = bh->b_this_page;
730                         }
731                         return false;
732                 }
733
734                 bh = bh->b_this_page;
735         } while (bh != head);
736         return true;
737 }
738
739 static int __buffer_migrate_page(struct address_space *mapping,
740                 struct page *newpage, struct page *page, enum migrate_mode mode,
741                 bool check_refs)
742 {
743         struct buffer_head *bh, *head;
744         int rc;
745         int expected_count;
746
747         if (!page_has_buffers(page))
748                 return migrate_page(mapping, newpage, page, mode);
749
750         /* Check whether page does not have extra refs before we do more work */
751         expected_count = expected_page_refs(mapping, page);
752         if (page_count(page) != expected_count)
753                 return -EAGAIN;
754
755         head = page_buffers(page);
756         if (!buffer_migrate_lock_buffers(head, mode))
757                 return -EAGAIN;
758
759         if (check_refs) {
760                 bool busy;
761                 bool invalidated = false;
762
763 recheck_buffers:
764                 busy = false;
765                 spin_lock(&mapping->private_lock);
766                 bh = head;
767                 do {
768                         if (atomic_read(&bh->b_count)) {
769                                 busy = true;
770                                 break;
771                         }
772                         bh = bh->b_this_page;
773                 } while (bh != head);
774                 spin_unlock(&mapping->private_lock);
775                 if (busy) {
776                         if (invalidated) {
777                                 rc = -EAGAIN;
778                                 goto unlock_buffers;
779                         }
780                         invalidate_bh_lrus();
781                         invalidated = true;
782                         goto recheck_buffers;
783                 }
784         }
785
786         rc = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
787         if (rc != MIGRATEPAGE_SUCCESS)
788                 goto unlock_buffers;
789
790         ClearPagePrivate(page);
791         set_page_private(newpage, page_private(page));
792         set_page_private(page, 0);
793         put_page(page);
794         get_page(newpage);
795
796         bh = head;
797         do {
798                 set_bh_page(bh, newpage, bh_offset(bh));
799                 bh = bh->b_this_page;
800
801         } while (bh != head);
802
803         SetPagePrivate(newpage);
804
805         if (mode != MIGRATE_SYNC_NO_COPY)
806                 migrate_page_copy(newpage, page);
807         else
808                 migrate_page_states(newpage, page);
809
810         rc = MIGRATEPAGE_SUCCESS;
811 unlock_buffers:
812         bh = head;
813         do {
814                 unlock_buffer(bh);
815                 bh = bh->b_this_page;
816
817         } while (bh != head);
818
819         return rc;
820 }
821
822 /*
823  * Migration function for pages with buffers. This function can only be used
824  * if the underlying filesystem guarantees that no other references to "page"
825  * exist. For example attached buffer heads are accessed only under page lock.
826  */
827 int buffer_migrate_page(struct address_space *mapping,
828                 struct page *newpage, struct page *page, enum migrate_mode mode)
829 {
830         return __buffer_migrate_page(mapping, newpage, page, mode, false);
831 }
832 EXPORT_SYMBOL(buffer_migrate_page);
833
834 /*
835  * Same as above except that this variant is more careful and checks that there
836  * are also no buffer head references. This function is the right one for
837  * mappings where buffer heads are directly looked up and referenced (such as
838  * block device mappings).
839  */
840 int buffer_migrate_page_norefs(struct address_space *mapping,
841                 struct page *newpage, struct page *page, enum migrate_mode mode)
842 {
843         return __buffer_migrate_page(mapping, newpage, page, mode, true);
844 }
845 #endif
846
847 /*
848  * Writeback a page to clean the dirty state
849  */
850 static int writeout(struct address_space *mapping, struct page *page)
851 {
852         struct writeback_control wbc = {
853                 .sync_mode = WB_SYNC_NONE,
854                 .nr_to_write = 1,
855                 .range_start = 0,
856                 .range_end = LLONG_MAX,
857                 .for_reclaim = 1
858         };
859         int rc;
860
861         if (!mapping->a_ops->writepage)
862                 /* No write method for the address space */
863                 return -EINVAL;
864
865         if (!clear_page_dirty_for_io(page))
866                 /* Someone else already triggered a write */
867                 return -EAGAIN;
868
869         /*
870          * A dirty page may imply that the underlying filesystem has
871          * the page on some queue. So the page must be clean for
872          * migration. Writeout may mean we loose the lock and the
873          * page state is no longer what we checked for earlier.
874          * At this point we know that the migration attempt cannot
875          * be successful.
876          */
877         remove_migration_ptes(page, page, false);
878
879         rc = mapping->a_ops->writepage(page, &wbc);
880
881         if (rc != AOP_WRITEPAGE_ACTIVATE)
882                 /* unlocked. Relock */
883                 lock_page(page);
884
885         return (rc < 0) ? -EIO : -EAGAIN;
886 }
887
888 /*
889  * Default handling if a filesystem does not provide a migration function.
890  */
891 static int fallback_migrate_page(struct address_space *mapping,
892         struct page *newpage, struct page *page, enum migrate_mode mode)
893 {
894         if (PageDirty(page)) {
895                 /* Only writeback pages in full synchronous migration */
896                 switch (mode) {
897                 case MIGRATE_SYNC:
898                 case MIGRATE_SYNC_NO_COPY:
899                         break;
900                 default:
901                         return -EBUSY;
902                 }
903                 return writeout(mapping, page);
904         }
905
906         /*
907          * Buffers may be managed in a filesystem specific way.
908          * We must have no buffers or drop them.
909          */
910         if (page_has_private(page) &&
911             !try_to_release_page(page, GFP_KERNEL))
912                 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
913
914         return migrate_page(mapping, newpage, page, mode);
915 }
916
917 /*
918  * Move a page to a newly allocated page
919  * The page is locked and all ptes have been successfully removed.
920  *
921  * The new page will have replaced the old page if this function
922  * is successful.
923  *
924  * Return value:
925  *   < 0 - error code
926  *  MIGRATEPAGE_SUCCESS - success
927  */
928 static int move_to_new_page(struct page *newpage, struct page *page,
929                                 enum migrate_mode mode)
930 {
931         struct address_space *mapping;
932         int rc = -EAGAIN;
933         bool is_lru = !__PageMovable(page);
934
935         VM_BUG_ON_PAGE(!PageLocked(page), page);
936         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
937
938         mapping = page_mapping(page);
939
940         if (likely(is_lru)) {
941                 if (!mapping)
942                         rc = migrate_page(mapping, newpage, page, mode);
943                 else if (mapping->a_ops->migratepage)
944                         /*
945                          * Most pages have a mapping and most filesystems
946                          * provide a migratepage callback. Anonymous pages
947                          * are part of swap space which also has its own
948                          * migratepage callback. This is the most common path
949                          * for page migration.
950                          */
951                         rc = mapping->a_ops->migratepage(mapping, newpage,
952                                                         page, mode);
953                 else
954                         rc = fallback_migrate_page(mapping, newpage,
955                                                         page, mode);
956         } else {
957                 /*
958                  * In case of non-lru page, it could be released after
959                  * isolation step. In that case, we shouldn't try migration.
960                  */
961                 VM_BUG_ON_PAGE(!PageIsolated(page), page);
962                 if (!PageMovable(page)) {
963                         rc = MIGRATEPAGE_SUCCESS;
964                         __ClearPageIsolated(page);
965                         goto out;
966                 }
967
968                 rc = mapping->a_ops->migratepage(mapping, newpage,
969                                                 page, mode);
970                 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
971                         !PageIsolated(page));
972         }
973
974         /*
975          * When successful, old pagecache page->mapping must be cleared before
976          * page is freed; but stats require that PageAnon be left as PageAnon.
977          */
978         if (rc == MIGRATEPAGE_SUCCESS) {
979                 if (__PageMovable(page)) {
980                         VM_BUG_ON_PAGE(!PageIsolated(page), page);
981
982                         /*
983                          * We clear PG_movable under page_lock so any compactor
984                          * cannot try to migrate this page.
985                          */
986                         __ClearPageIsolated(page);
987                 }
988
989                 /*
990                  * Anonymous and movable page->mapping will be cleard by
991                  * free_pages_prepare so don't reset it here for keeping
992                  * the type to work PageAnon, for example.
993                  */
994                 if (!PageMappingFlags(page))
995                         page->mapping = NULL;
996
997                 if (unlikely(is_zone_device_page(newpage))) {
998                         if (is_device_public_page(newpage))
999                                 flush_dcache_page(newpage);
1000                 } else
1001                         flush_dcache_page(newpage);
1002
1003         }
1004 out:
1005         return rc;
1006 }
1007
1008 static int __unmap_and_move(struct page *page, struct page *newpage,
1009                                 int force, enum migrate_mode mode)
1010 {
1011         int rc = -EAGAIN;
1012         int page_was_mapped = 0;
1013         struct anon_vma *anon_vma = NULL;
1014         bool is_lru = !__PageMovable(page);
1015
1016         if (!trylock_page(page)) {
1017                 if (!force || mode == MIGRATE_ASYNC)
1018                         goto out;
1019
1020                 /*
1021                  * It's not safe for direct compaction to call lock_page.
1022                  * For example, during page readahead pages are added locked
1023                  * to the LRU. Later, when the IO completes the pages are
1024                  * marked uptodate and unlocked. However, the queueing
1025                  * could be merging multiple pages for one bio (e.g.
1026                  * mpage_readpages). If an allocation happens for the
1027                  * second or third page, the process can end up locking
1028                  * the same page twice and deadlocking. Rather than
1029                  * trying to be clever about what pages can be locked,
1030                  * avoid the use of lock_page for direct compaction
1031                  * altogether.
1032                  */
1033                 if (current->flags & PF_MEMALLOC)
1034                         goto out;
1035
1036                 lock_page(page);
1037         }
1038
1039         if (PageWriteback(page)) {
1040                 /*
1041                  * Only in the case of a full synchronous migration is it
1042                  * necessary to wait for PageWriteback. In the async case,
1043                  * the retry loop is too short and in the sync-light case,
1044                  * the overhead of stalling is too much
1045                  */
1046                 switch (mode) {
1047                 case MIGRATE_SYNC:
1048                 case MIGRATE_SYNC_NO_COPY:
1049                         break;
1050                 default:
1051                         rc = -EBUSY;
1052                         goto out_unlock;
1053                 }
1054                 if (!force)
1055                         goto out_unlock;
1056                 wait_on_page_writeback(page);
1057         }
1058
1059         /*
1060          * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1061          * we cannot notice that anon_vma is freed while we migrates a page.
1062          * This get_anon_vma() delays freeing anon_vma pointer until the end
1063          * of migration. File cache pages are no problem because of page_lock()
1064          * File Caches may use write_page() or lock_page() in migration, then,
1065          * just care Anon page here.
1066          *
1067          * Only page_get_anon_vma() understands the subtleties of
1068          * getting a hold on an anon_vma from outside one of its mms.
1069          * But if we cannot get anon_vma, then we won't need it anyway,
1070          * because that implies that the anon page is no longer mapped
1071          * (and cannot be remapped so long as we hold the page lock).
1072          */
1073         if (PageAnon(page) && !PageKsm(page))
1074                 anon_vma = page_get_anon_vma(page);
1075
1076         /*
1077          * Block others from accessing the new page when we get around to
1078          * establishing additional references. We are usually the only one
1079          * holding a reference to newpage at this point. We used to have a BUG
1080          * here if trylock_page(newpage) fails, but would like to allow for
1081          * cases where there might be a race with the previous use of newpage.
1082          * This is much like races on refcount of oldpage: just don't BUG().
1083          */
1084         if (unlikely(!trylock_page(newpage)))
1085                 goto out_unlock;
1086
1087         if (unlikely(!is_lru)) {
1088                 rc = move_to_new_page(newpage, page, mode);
1089                 goto out_unlock_both;
1090         }
1091
1092         /*
1093          * Corner case handling:
1094          * 1. When a new swap-cache page is read into, it is added to the LRU
1095          * and treated as swapcache but it has no rmap yet.
1096          * Calling try_to_unmap() against a page->mapping==NULL page will
1097          * trigger a BUG.  So handle it here.
1098          * 2. An orphaned page (see truncate_complete_page) might have
1099          * fs-private metadata. The page can be picked up due to memory
1100          * offlining.  Everywhere else except page reclaim, the page is
1101          * invisible to the vm, so the page can not be migrated.  So try to
1102          * free the metadata, so the page can be freed.
1103          */
1104         if (!page->mapping) {
1105                 VM_BUG_ON_PAGE(PageAnon(page), page);
1106                 if (page_has_private(page)) {
1107                         try_to_free_buffers(page);
1108                         goto out_unlock_both;
1109                 }
1110         } else if (page_mapped(page)) {
1111                 /* Establish migration ptes */
1112                 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1113                                 page);
1114                 try_to_unmap(page,
1115                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1116                 page_was_mapped = 1;
1117         }
1118
1119         if (!page_mapped(page))
1120                 rc = move_to_new_page(newpage, page, mode);
1121
1122         if (page_was_mapped)
1123                 remove_migration_ptes(page,
1124                         rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1125
1126 out_unlock_both:
1127         unlock_page(newpage);
1128 out_unlock:
1129         /* Drop an anon_vma reference if we took one */
1130         if (anon_vma)
1131                 put_anon_vma(anon_vma);
1132         unlock_page(page);
1133 out:
1134         /*
1135          * If migration is successful, decrease refcount of the newpage
1136          * which will not free the page because new page owner increased
1137          * refcounter. As well, if it is LRU page, add the page to LRU
1138          * list in here. Use the old state of the isolated source page to
1139          * determine if we migrated a LRU page. newpage was already unlocked
1140          * and possibly modified by its owner - don't rely on the page
1141          * state.
1142          */
1143         if (rc == MIGRATEPAGE_SUCCESS) {
1144                 if (unlikely(!is_lru))
1145                         put_page(newpage);
1146                 else
1147                         putback_lru_page(newpage);
1148         }
1149
1150         return rc;
1151 }
1152
1153 /*
1154  * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move().  Work
1155  * around it.
1156  */
1157 #if defined(CONFIG_ARM) && \
1158         defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1159 #define ICE_noinline noinline
1160 #else
1161 #define ICE_noinline
1162 #endif
1163
1164 /*
1165  * Obtain the lock on page, remove all ptes and migrate the page
1166  * to the newly allocated page in newpage.
1167  */
1168 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1169                                    free_page_t put_new_page,
1170                                    unsigned long private, struct page *page,
1171                                    int force, enum migrate_mode mode,
1172                                    enum migrate_reason reason)
1173 {
1174         int rc = MIGRATEPAGE_SUCCESS;
1175         struct page *newpage;
1176
1177         if (!thp_migration_supported() && PageTransHuge(page))
1178                 return -ENOMEM;
1179
1180         newpage = get_new_page(page, private);
1181         if (!newpage)
1182                 return -ENOMEM;
1183
1184         if (page_count(page) == 1) {
1185                 /* page was freed from under us. So we are done. */
1186                 ClearPageActive(page);
1187                 ClearPageUnevictable(page);
1188                 if (unlikely(__PageMovable(page))) {
1189                         lock_page(page);
1190                         if (!PageMovable(page))
1191                                 __ClearPageIsolated(page);
1192                         unlock_page(page);
1193                 }
1194                 if (put_new_page)
1195                         put_new_page(newpage, private);
1196                 else
1197                         put_page(newpage);
1198                 goto out;
1199         }
1200
1201         rc = __unmap_and_move(page, newpage, force, mode);
1202         if (rc == MIGRATEPAGE_SUCCESS)
1203                 set_page_owner_migrate_reason(newpage, reason);
1204
1205 out:
1206         if (rc != -EAGAIN) {
1207                 /*
1208                  * A page that has been migrated has all references
1209                  * removed and will be freed. A page that has not been
1210                  * migrated will have kepts its references and be
1211                  * restored.
1212                  */
1213                 list_del(&page->lru);
1214
1215                 /*
1216                  * Compaction can migrate also non-LRU pages which are
1217                  * not accounted to NR_ISOLATED_*. They can be recognized
1218                  * as __PageMovable
1219                  */
1220                 if (likely(!__PageMovable(page)))
1221                         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1222                                         page_is_file_cache(page), -hpage_nr_pages(page));
1223         }
1224
1225         /*
1226          * If migration is successful, releases reference grabbed during
1227          * isolation. Otherwise, restore the page to right list unless
1228          * we want to retry.
1229          */
1230         if (rc == MIGRATEPAGE_SUCCESS) {
1231                 put_page(page);
1232                 if (reason == MR_MEMORY_FAILURE) {
1233                         /*
1234                          * Set PG_HWPoison on just freed page
1235                          * intentionally. Although it's rather weird,
1236                          * it's how HWPoison flag works at the moment.
1237                          */
1238                         if (set_hwpoison_free_buddy_page(page))
1239                                 num_poisoned_pages_inc();
1240                 }
1241         } else {
1242                 if (rc != -EAGAIN) {
1243                         if (likely(!__PageMovable(page))) {
1244                                 putback_lru_page(page);
1245                                 goto put_new;
1246                         }
1247
1248                         lock_page(page);
1249                         if (PageMovable(page))
1250                                 putback_movable_page(page);
1251                         else
1252                                 __ClearPageIsolated(page);
1253                         unlock_page(page);
1254                         put_page(page);
1255                 }
1256 put_new:
1257                 if (put_new_page)
1258                         put_new_page(newpage, private);
1259                 else
1260                         put_page(newpage);
1261         }
1262
1263         return rc;
1264 }
1265
1266 /*
1267  * Counterpart of unmap_and_move_page() for hugepage migration.
1268  *
1269  * This function doesn't wait the completion of hugepage I/O
1270  * because there is no race between I/O and migration for hugepage.
1271  * Note that currently hugepage I/O occurs only in direct I/O
1272  * where no lock is held and PG_writeback is irrelevant,
1273  * and writeback status of all subpages are counted in the reference
1274  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1275  * under direct I/O, the reference of the head page is 512 and a bit more.)
1276  * This means that when we try to migrate hugepage whose subpages are
1277  * doing direct I/O, some references remain after try_to_unmap() and
1278  * hugepage migration fails without data corruption.
1279  *
1280  * There is also no race when direct I/O is issued on the page under migration,
1281  * because then pte is replaced with migration swap entry and direct I/O code
1282  * will wait in the page fault for migration to complete.
1283  */
1284 static int unmap_and_move_huge_page(new_page_t get_new_page,
1285                                 free_page_t put_new_page, unsigned long private,
1286                                 struct page *hpage, int force,
1287                                 enum migrate_mode mode, int reason)
1288 {
1289         int rc = -EAGAIN;
1290         int page_was_mapped = 0;
1291         struct page *new_hpage;
1292         struct anon_vma *anon_vma = NULL;
1293
1294         /*
1295          * Migratability of hugepages depends on architectures and their size.
1296          * This check is necessary because some callers of hugepage migration
1297          * like soft offline and memory hotremove don't walk through page
1298          * tables or check whether the hugepage is pmd-based or not before
1299          * kicking migration.
1300          */
1301         if (!hugepage_migration_supported(page_hstate(hpage))) {
1302                 putback_active_hugepage(hpage);
1303                 return -ENOSYS;
1304         }
1305
1306         new_hpage = get_new_page(hpage, private);
1307         if (!new_hpage)
1308                 return -ENOMEM;
1309
1310         if (!trylock_page(hpage)) {
1311                 if (!force)
1312                         goto out;
1313                 switch (mode) {
1314                 case MIGRATE_SYNC:
1315                 case MIGRATE_SYNC_NO_COPY:
1316                         break;
1317                 default:
1318                         goto out;
1319                 }
1320                 lock_page(hpage);
1321         }
1322
1323         /*
1324          * Check for pages which are in the process of being freed.  Without
1325          * page_mapping() set, hugetlbfs specific move page routine will not
1326          * be called and we could leak usage counts for subpools.
1327          */
1328         if (page_private(hpage) && !page_mapping(hpage)) {
1329                 rc = -EBUSY;
1330                 goto out_unlock;
1331         }
1332
1333         if (PageAnon(hpage))
1334                 anon_vma = page_get_anon_vma(hpage);
1335
1336         if (unlikely(!trylock_page(new_hpage)))
1337                 goto put_anon;
1338
1339         if (page_mapped(hpage)) {
1340                 try_to_unmap(hpage,
1341                         TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1342                 page_was_mapped = 1;
1343         }
1344
1345         if (!page_mapped(hpage))
1346                 rc = move_to_new_page(new_hpage, hpage, mode);
1347
1348         if (page_was_mapped)
1349                 remove_migration_ptes(hpage,
1350                         rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1351
1352         unlock_page(new_hpage);
1353
1354 put_anon:
1355         if (anon_vma)
1356                 put_anon_vma(anon_vma);
1357
1358         if (rc == MIGRATEPAGE_SUCCESS) {
1359                 move_hugetlb_state(hpage, new_hpage, reason);
1360                 put_new_page = NULL;
1361         }
1362
1363 out_unlock:
1364         unlock_page(hpage);
1365 out:
1366         if (rc != -EAGAIN)
1367                 putback_active_hugepage(hpage);
1368
1369         /*
1370          * If migration was not successful and there's a freeing callback, use
1371          * it.  Otherwise, put_page() will drop the reference grabbed during
1372          * isolation.
1373          */
1374         if (put_new_page)
1375                 put_new_page(new_hpage, private);
1376         else
1377                 putback_active_hugepage(new_hpage);
1378
1379         return rc;
1380 }
1381
1382 /*
1383  * migrate_pages - migrate the pages specified in a list, to the free pages
1384  *                 supplied as the target for the page migration
1385  *
1386  * @from:               The list of pages to be migrated.
1387  * @get_new_page:       The function used to allocate free pages to be used
1388  *                      as the target of the page migration.
1389  * @put_new_page:       The function used to free target pages if migration
1390  *                      fails, or NULL if no special handling is necessary.
1391  * @private:            Private data to be passed on to get_new_page()
1392  * @mode:               The migration mode that specifies the constraints for
1393  *                      page migration, if any.
1394  * @reason:             The reason for page migration.
1395  *
1396  * The function returns after 10 attempts or if no pages are movable any more
1397  * because the list has become empty or no retryable pages exist any more.
1398  * The caller should call putback_movable_pages() to return pages to the LRU
1399  * or free list only if ret != 0.
1400  *
1401  * Returns the number of pages that were not migrated, or an error code.
1402  */
1403 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1404                 free_page_t put_new_page, unsigned long private,
1405                 enum migrate_mode mode, int reason)
1406 {
1407         int retry = 1;
1408         int nr_failed = 0;
1409         int nr_succeeded = 0;
1410         int pass = 0;
1411         struct page *page;
1412         struct page *page2;
1413         int swapwrite = current->flags & PF_SWAPWRITE;
1414         int rc;
1415
1416         if (!swapwrite)
1417                 current->flags |= PF_SWAPWRITE;
1418
1419         for(pass = 0; pass < 10 && retry; pass++) {
1420                 retry = 0;
1421
1422                 list_for_each_entry_safe(page, page2, from, lru) {
1423 retry:
1424                         cond_resched();
1425
1426                         if (PageHuge(page))
1427                                 rc = unmap_and_move_huge_page(get_new_page,
1428                                                 put_new_page, private, page,
1429                                                 pass > 2, mode, reason);
1430                         else
1431                                 rc = unmap_and_move(get_new_page, put_new_page,
1432                                                 private, page, pass > 2, mode,
1433                                                 reason);
1434
1435                         switch(rc) {
1436                         case -ENOMEM:
1437                                 /*
1438                                  * THP migration might be unsupported or the
1439                                  * allocation could've failed so we should
1440                                  * retry on the same page with the THP split
1441                                  * to base pages.
1442                                  *
1443                                  * Head page is retried immediately and tail
1444                                  * pages are added to the tail of the list so
1445                                  * we encounter them after the rest of the list
1446                                  * is processed.
1447                                  */
1448                                 if (PageTransHuge(page) && !PageHuge(page)) {
1449                                         lock_page(page);
1450                                         rc = split_huge_page_to_list(page, from);
1451                                         unlock_page(page);
1452                                         if (!rc) {
1453                                                 list_safe_reset_next(page, page2, lru);
1454                                                 goto retry;
1455                                         }
1456                                 }
1457                                 nr_failed++;
1458                                 goto out;
1459                         case -EAGAIN:
1460                                 retry++;
1461                                 break;
1462                         case MIGRATEPAGE_SUCCESS:
1463                                 nr_succeeded++;
1464                                 break;
1465                         default:
1466                                 /*
1467                                  * Permanent failure (-EBUSY, -ENOSYS, etc.):
1468                                  * unlike -EAGAIN case, the failed page is
1469                                  * removed from migration page list and not
1470                                  * retried in the next outer loop.
1471                                  */
1472                                 nr_failed++;
1473                                 break;
1474                         }
1475                 }
1476         }
1477         nr_failed += retry;
1478         rc = nr_failed;
1479 out:
1480         if (nr_succeeded)
1481                 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1482         if (nr_failed)
1483                 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1484         trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1485
1486         if (!swapwrite)
1487                 current->flags &= ~PF_SWAPWRITE;
1488
1489         return rc;
1490 }
1491
1492 #ifdef CONFIG_NUMA
1493
1494 static int store_status(int __user *status, int start, int value, int nr)
1495 {
1496         while (nr-- > 0) {
1497                 if (put_user(value, status + start))
1498                         return -EFAULT;
1499                 start++;
1500         }
1501
1502         return 0;
1503 }
1504
1505 static int do_move_pages_to_node(struct mm_struct *mm,
1506                 struct list_head *pagelist, int node)
1507 {
1508         int err;
1509
1510         if (list_empty(pagelist))
1511                 return 0;
1512
1513         err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1514                         MIGRATE_SYNC, MR_SYSCALL);
1515         if (err)
1516                 putback_movable_pages(pagelist);
1517         return err;
1518 }
1519
1520 /*
1521  * Resolves the given address to a struct page, isolates it from the LRU and
1522  * puts it to the given pagelist.
1523  * Returns -errno if the page cannot be found/isolated or 0 when it has been
1524  * queued or the page doesn't need to be migrated because it is already on
1525  * the target node
1526  */
1527 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1528                 int node, struct list_head *pagelist, bool migrate_all)
1529 {
1530         struct vm_area_struct *vma;
1531         struct page *page;
1532         unsigned int follflags;
1533         int err;
1534
1535         down_read(&mm->mmap_sem);
1536         err = -EFAULT;
1537         vma = find_vma(mm, addr);
1538         if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1539                 goto out;
1540
1541         /* FOLL_DUMP to ignore special (like zero) pages */
1542         follflags = FOLL_GET | FOLL_DUMP;
1543         page = follow_page(vma, addr, follflags);
1544
1545         err = PTR_ERR(page);
1546         if (IS_ERR(page))
1547                 goto out;
1548
1549         err = -ENOENT;
1550         if (!page)
1551                 goto out;
1552
1553         err = 0;
1554         if (page_to_nid(page) == node)
1555                 goto out_putpage;
1556
1557         err = -EACCES;
1558         if (page_mapcount(page) > 1 && !migrate_all)
1559                 goto out_putpage;
1560
1561         if (PageHuge(page)) {
1562                 if (PageHead(page)) {
1563                         isolate_huge_page(page, pagelist);
1564                         err = 0;
1565                 }
1566         } else {
1567                 struct page *head;
1568
1569                 head = compound_head(page);
1570                 err = isolate_lru_page(head);
1571                 if (err)
1572                         goto out_putpage;
1573
1574                 err = 0;
1575                 list_add_tail(&head->lru, pagelist);
1576                 mod_node_page_state(page_pgdat(head),
1577                         NR_ISOLATED_ANON + page_is_file_cache(head),
1578                         hpage_nr_pages(head));
1579         }
1580 out_putpage:
1581         /*
1582          * Either remove the duplicate refcount from
1583          * isolate_lru_page() or drop the page ref if it was
1584          * not isolated.
1585          */
1586         put_page(page);
1587 out:
1588         up_read(&mm->mmap_sem);
1589         return err;
1590 }
1591
1592 /*
1593  * Migrate an array of page address onto an array of nodes and fill
1594  * the corresponding array of status.
1595  */
1596 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1597                          unsigned long nr_pages,
1598                          const void __user * __user *pages,
1599                          const int __user *nodes,
1600                          int __user *status, int flags)
1601 {
1602         int current_node = NUMA_NO_NODE;
1603         LIST_HEAD(pagelist);
1604         int start, i;
1605         int err = 0, err1;
1606
1607         migrate_prep();
1608
1609         for (i = start = 0; i < nr_pages; i++) {
1610                 const void __user *p;
1611                 unsigned long addr;
1612                 int node;
1613
1614                 err = -EFAULT;
1615                 if (get_user(p, pages + i))
1616                         goto out_flush;
1617                 if (get_user(node, nodes + i))
1618                         goto out_flush;
1619                 addr = (unsigned long)p;
1620
1621                 err = -ENODEV;
1622                 if (node < 0 || node >= MAX_NUMNODES)
1623                         goto out_flush;
1624                 if (!node_state(node, N_MEMORY))
1625                         goto out_flush;
1626
1627                 err = -EACCES;
1628                 if (!node_isset(node, task_nodes))
1629                         goto out_flush;
1630
1631                 if (current_node == NUMA_NO_NODE) {
1632                         current_node = node;
1633                         start = i;
1634                 } else if (node != current_node) {
1635                         err = do_move_pages_to_node(mm, &pagelist, current_node);
1636                         if (err)
1637                                 goto out;
1638                         err = store_status(status, start, current_node, i - start);
1639                         if (err)
1640                                 goto out;
1641                         start = i;
1642                         current_node = node;
1643                 }
1644
1645                 /*
1646                  * Errors in the page lookup or isolation are not fatal and we simply
1647                  * report them via status
1648                  */
1649                 err = add_page_for_migration(mm, addr, current_node,
1650                                 &pagelist, flags & MPOL_MF_MOVE_ALL);
1651                 if (!err)
1652                         continue;
1653
1654                 err = store_status(status, i, err, 1);
1655                 if (err)
1656                         goto out_flush;
1657
1658                 err = do_move_pages_to_node(mm, &pagelist, current_node);
1659                 if (err)
1660                         goto out;
1661                 if (i > start) {
1662                         err = store_status(status, start, current_node, i - start);
1663                         if (err)
1664                                 goto out;
1665                 }
1666                 current_node = NUMA_NO_NODE;
1667         }
1668 out_flush:
1669         if (list_empty(&pagelist))
1670                 return err;
1671
1672         /* Make sure we do not overwrite the existing error */
1673         err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1674         if (!err1)
1675                 err1 = store_status(status, start, current_node, i - start);
1676         if (!err)
1677                 err = err1;
1678 out:
1679         return err;
1680 }
1681
1682 /*
1683  * Determine the nodes of an array of pages and store it in an array of status.
1684  */
1685 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1686                                 const void __user **pages, int *status)
1687 {
1688         unsigned long i;
1689
1690         down_read(&mm->mmap_sem);
1691
1692         for (i = 0; i < nr_pages; i++) {
1693                 unsigned long addr = (unsigned long)(*pages);
1694                 struct vm_area_struct *vma;
1695                 struct page *page;
1696                 int err = -EFAULT;
1697
1698                 vma = find_vma(mm, addr);
1699                 if (!vma || addr < vma->vm_start)
1700                         goto set_status;
1701
1702                 /* FOLL_DUMP to ignore special (like zero) pages */
1703                 page = follow_page(vma, addr, FOLL_DUMP);
1704
1705                 err = PTR_ERR(page);
1706                 if (IS_ERR(page))
1707                         goto set_status;
1708
1709                 err = page ? page_to_nid(page) : -ENOENT;
1710 set_status:
1711                 *status = err;
1712
1713                 pages++;
1714                 status++;
1715         }
1716
1717         up_read(&mm->mmap_sem);
1718 }
1719
1720 /*
1721  * Determine the nodes of a user array of pages and store it in
1722  * a user array of status.
1723  */
1724 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1725                          const void __user * __user *pages,
1726                          int __user *status)
1727 {
1728 #define DO_PAGES_STAT_CHUNK_NR 16
1729         const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1730         int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1731
1732         while (nr_pages) {
1733                 unsigned long chunk_nr;
1734
1735                 chunk_nr = nr_pages;
1736                 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1737                         chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1738
1739                 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1740                         break;
1741
1742                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1743
1744                 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1745                         break;
1746
1747                 pages += chunk_nr;
1748                 status += chunk_nr;
1749                 nr_pages -= chunk_nr;
1750         }
1751         return nr_pages ? -EFAULT : 0;
1752 }
1753
1754 /*
1755  * Move a list of pages in the address space of the currently executing
1756  * process.
1757  */
1758 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1759                              const void __user * __user *pages,
1760                              const int __user *nodes,
1761                              int __user *status, int flags)
1762 {
1763         struct task_struct *task;
1764         struct mm_struct *mm;
1765         int err;
1766         nodemask_t task_nodes;
1767
1768         /* Check flags */
1769         if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1770                 return -EINVAL;
1771
1772         if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1773                 return -EPERM;
1774
1775         /* Find the mm_struct */
1776         rcu_read_lock();
1777         task = pid ? find_task_by_vpid(pid) : current;
1778         if (!task) {
1779                 rcu_read_unlock();
1780                 return -ESRCH;
1781         }
1782         get_task_struct(task);
1783
1784         /*
1785          * Check if this process has the right to modify the specified
1786          * process. Use the regular "ptrace_may_access()" checks.
1787          */
1788         if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1789                 rcu_read_unlock();
1790                 err = -EPERM;
1791                 goto out;
1792         }
1793         rcu_read_unlock();
1794
1795         err = security_task_movememory(task);
1796         if (err)
1797                 goto out;
1798
1799         task_nodes = cpuset_mems_allowed(task);
1800         mm = get_task_mm(task);
1801         put_task_struct(task);
1802
1803         if (!mm)
1804                 return -EINVAL;
1805
1806         if (nodes)
1807                 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1808                                     nodes, status, flags);
1809         else
1810                 err = do_pages_stat(mm, nr_pages, pages, status);
1811
1812         mmput(mm);
1813         return err;
1814
1815 out:
1816         put_task_struct(task);
1817         return err;
1818 }
1819
1820 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1821                 const void __user * __user *, pages,
1822                 const int __user *, nodes,
1823                 int __user *, status, int, flags)
1824 {
1825         return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1826 }
1827
1828 #ifdef CONFIG_COMPAT
1829 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1830                        compat_uptr_t __user *, pages32,
1831                        const int __user *, nodes,
1832                        int __user *, status,
1833                        int, flags)
1834 {
1835         const void __user * __user *pages;
1836         int i;
1837
1838         pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1839         for (i = 0; i < nr_pages; i++) {
1840                 compat_uptr_t p;
1841
1842                 if (get_user(p, pages32 + i) ||
1843                         put_user(compat_ptr(p), pages + i))
1844                         return -EFAULT;
1845         }
1846         return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1847 }
1848 #endif /* CONFIG_COMPAT */
1849
1850 #ifdef CONFIG_NUMA_BALANCING
1851 /*
1852  * Returns true if this is a safe migration target node for misplaced NUMA
1853  * pages. Currently it only checks the watermarks which crude
1854  */
1855 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1856                                    unsigned long nr_migrate_pages)
1857 {
1858         int z;
1859
1860         for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1861                 struct zone *zone = pgdat->node_zones + z;
1862
1863                 if (!populated_zone(zone))
1864                         continue;
1865
1866                 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1867                 if (!zone_watermark_ok(zone, 0,
1868                                        high_wmark_pages(zone) +
1869                                        nr_migrate_pages,
1870                                        0, 0))
1871                         continue;
1872                 return true;
1873         }
1874         return false;
1875 }
1876
1877 static struct page *alloc_misplaced_dst_page(struct page *page,
1878                                            unsigned long data)
1879 {
1880         int nid = (int) data;
1881         struct page *newpage;
1882
1883         newpage = __alloc_pages_node(nid,
1884                                          (GFP_HIGHUSER_MOVABLE |
1885                                           __GFP_THISNODE | __GFP_NOMEMALLOC |
1886                                           __GFP_NORETRY | __GFP_NOWARN) &
1887                                          ~__GFP_RECLAIM, 0);
1888
1889         return newpage;
1890 }
1891
1892 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1893 {
1894         int page_lru;
1895
1896         VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1897
1898         /* Avoid migrating to a node that is nearly full */
1899         if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1900                 return 0;
1901
1902         if (isolate_lru_page(page))
1903                 return 0;
1904
1905         /*
1906          * migrate_misplaced_transhuge_page() skips page migration's usual
1907          * check on page_count(), so we must do it here, now that the page
1908          * has been isolated: a GUP pin, or any other pin, prevents migration.
1909          * The expected page count is 3: 1 for page's mapcount and 1 for the
1910          * caller's pin and 1 for the reference taken by isolate_lru_page().
1911          */
1912         if (PageTransHuge(page) && page_count(page) != 3) {
1913                 putback_lru_page(page);
1914                 return 0;
1915         }
1916
1917         page_lru = page_is_file_cache(page);
1918         mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1919                                 hpage_nr_pages(page));
1920
1921         /*
1922          * Isolating the page has taken another reference, so the
1923          * caller's reference can be safely dropped without the page
1924          * disappearing underneath us during migration.
1925          */
1926         put_page(page);
1927         return 1;
1928 }
1929
1930 bool pmd_trans_migrating(pmd_t pmd)
1931 {
1932         struct page *page = pmd_page(pmd);
1933         return PageLocked(page);
1934 }
1935
1936 /*
1937  * Attempt to migrate a misplaced page to the specified destination
1938  * node. Caller is expected to have an elevated reference count on
1939  * the page that will be dropped by this function before returning.
1940  */
1941 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1942                            int node)
1943 {
1944         pg_data_t *pgdat = NODE_DATA(node);
1945         int isolated;
1946         int nr_remaining;
1947         LIST_HEAD(migratepages);
1948
1949         /*
1950          * Don't migrate file pages that are mapped in multiple processes
1951          * with execute permissions as they are probably shared libraries.
1952          */
1953         if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1954             (vma->vm_flags & VM_EXEC))
1955                 goto out;
1956
1957         /*
1958          * Also do not migrate dirty pages as not all filesystems can move
1959          * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1960          */
1961         if (page_is_file_cache(page) && PageDirty(page))
1962                 goto out;
1963
1964         isolated = numamigrate_isolate_page(pgdat, page);
1965         if (!isolated)
1966                 goto out;
1967
1968         list_add(&page->lru, &migratepages);
1969         nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1970                                      NULL, node, MIGRATE_ASYNC,
1971                                      MR_NUMA_MISPLACED);
1972         if (nr_remaining) {
1973                 if (!list_empty(&migratepages)) {
1974                         list_del(&page->lru);
1975                         dec_node_page_state(page, NR_ISOLATED_ANON +
1976                                         page_is_file_cache(page));
1977                         putback_lru_page(page);
1978                 }
1979                 isolated = 0;
1980         } else
1981                 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1982         BUG_ON(!list_empty(&migratepages));
1983         return isolated;
1984
1985 out:
1986         put_page(page);
1987         return 0;
1988 }
1989 #endif /* CONFIG_NUMA_BALANCING */
1990
1991 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1992 /*
1993  * Migrates a THP to a given target node. page must be locked and is unlocked
1994  * before returning.
1995  */
1996 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1997                                 struct vm_area_struct *vma,
1998                                 pmd_t *pmd, pmd_t entry,
1999                                 unsigned long address,
2000                                 struct page *page, int node)
2001 {
2002         spinlock_t *ptl;
2003         pg_data_t *pgdat = NODE_DATA(node);
2004         int isolated = 0;
2005         struct page *new_page = NULL;
2006         int page_lru = page_is_file_cache(page);
2007         unsigned long start = address & HPAGE_PMD_MASK;
2008
2009         new_page = alloc_pages_node(node,
2010                 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2011                 HPAGE_PMD_ORDER);
2012         if (!new_page)
2013                 goto out_fail;
2014         prep_transhuge_page(new_page);
2015
2016         isolated = numamigrate_isolate_page(pgdat, page);
2017         if (!isolated) {
2018                 put_page(new_page);
2019                 goto out_fail;
2020         }
2021
2022         /* Prepare a page as a migration target */
2023         __SetPageLocked(new_page);
2024         if (PageSwapBacked(page))
2025                 __SetPageSwapBacked(new_page);
2026
2027         /* anon mapping, we can simply copy page->mapping to the new page: */
2028         new_page->mapping = page->mapping;
2029         new_page->index = page->index;
2030         /* flush the cache before copying using the kernel virtual address */
2031         flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2032         migrate_page_copy(new_page, page);
2033         WARN_ON(PageLRU(new_page));
2034
2035         /* Recheck the target PMD */
2036         ptl = pmd_lock(mm, pmd);
2037         if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2038                 spin_unlock(ptl);
2039
2040                 /* Reverse changes made by migrate_page_copy() */
2041                 if (TestClearPageActive(new_page))
2042                         SetPageActive(page);
2043                 if (TestClearPageUnevictable(new_page))
2044                         SetPageUnevictable(page);
2045
2046                 unlock_page(new_page);
2047                 put_page(new_page);             /* Free it */
2048
2049                 /* Retake the callers reference and putback on LRU */
2050                 get_page(page);
2051                 putback_lru_page(page);
2052                 mod_node_page_state(page_pgdat(page),
2053                          NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2054
2055                 goto out_unlock;
2056         }
2057
2058         entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2059         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2060
2061         /*
2062          * Overwrite the old entry under pagetable lock and establish
2063          * the new PTE. Any parallel GUP will either observe the old
2064          * page blocking on the page lock, block on the page table
2065          * lock or observe the new page. The SetPageUptodate on the
2066          * new page and page_add_new_anon_rmap guarantee the copy is
2067          * visible before the pagetable update.
2068          */
2069         page_add_anon_rmap(new_page, vma, start, true);
2070         /*
2071          * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2072          * has already been flushed globally.  So no TLB can be currently
2073          * caching this non present pmd mapping.  There's no need to clear the
2074          * pmd before doing set_pmd_at(), nor to flush the TLB after
2075          * set_pmd_at().  Clearing the pmd here would introduce a race
2076          * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2077          * mmap_sem for reading.  If the pmd is set to NULL at any given time,
2078          * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2079          * pmd.
2080          */
2081         set_pmd_at(mm, start, pmd, entry);
2082         update_mmu_cache_pmd(vma, address, &entry);
2083
2084         page_ref_unfreeze(page, 2);
2085         mlock_migrate_page(new_page, page);
2086         page_remove_rmap(page, true);
2087         set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2088
2089         spin_unlock(ptl);
2090
2091         /* Take an "isolate" reference and put new page on the LRU. */
2092         get_page(new_page);
2093         putback_lru_page(new_page);
2094
2095         unlock_page(new_page);
2096         unlock_page(page);
2097         put_page(page);                 /* Drop the rmap reference */
2098         put_page(page);                 /* Drop the LRU isolation reference */
2099
2100         count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2101         count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2102
2103         mod_node_page_state(page_pgdat(page),
2104                         NR_ISOLATED_ANON + page_lru,
2105                         -HPAGE_PMD_NR);
2106         return isolated;
2107
2108 out_fail:
2109         count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2110         ptl = pmd_lock(mm, pmd);
2111         if (pmd_same(*pmd, entry)) {
2112                 entry = pmd_modify(entry, vma->vm_page_prot);
2113                 set_pmd_at(mm, start, pmd, entry);
2114                 update_mmu_cache_pmd(vma, address, &entry);
2115         }
2116         spin_unlock(ptl);
2117
2118 out_unlock:
2119         unlock_page(page);
2120         put_page(page);
2121         return 0;
2122 }
2123 #endif /* CONFIG_NUMA_BALANCING */
2124
2125 #endif /* CONFIG_NUMA */
2126
2127 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2128 struct migrate_vma {
2129         struct vm_area_struct   *vma;
2130         unsigned long           *dst;
2131         unsigned long           *src;
2132         unsigned long           cpages;
2133         unsigned long           npages;
2134         unsigned long           start;
2135         unsigned long           end;
2136 };
2137
2138 static int migrate_vma_collect_hole(unsigned long start,
2139                                     unsigned long end,
2140                                     struct mm_walk *walk)
2141 {
2142         struct migrate_vma *migrate = walk->private;
2143         unsigned long addr;
2144
2145         for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2146                 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2147                 migrate->dst[migrate->npages] = 0;
2148                 migrate->npages++;
2149                 migrate->cpages++;
2150         }
2151
2152         return 0;
2153 }
2154
2155 static int migrate_vma_collect_skip(unsigned long start,
2156                                     unsigned long end,
2157                                     struct mm_walk *walk)
2158 {
2159         struct migrate_vma *migrate = walk->private;
2160         unsigned long addr;
2161
2162         for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2163                 migrate->dst[migrate->npages] = 0;
2164                 migrate->src[migrate->npages++] = 0;
2165         }
2166
2167         return 0;
2168 }
2169
2170 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2171                                    unsigned long start,
2172                                    unsigned long end,
2173                                    struct mm_walk *walk)
2174 {
2175         struct migrate_vma *migrate = walk->private;
2176         struct vm_area_struct *vma = walk->vma;
2177         struct mm_struct *mm = vma->vm_mm;
2178         unsigned long addr = start, unmapped = 0;
2179         spinlock_t *ptl;
2180         pte_t *ptep;
2181
2182 again:
2183         if (pmd_none(*pmdp))
2184                 return migrate_vma_collect_hole(start, end, walk);
2185
2186         if (pmd_trans_huge(*pmdp)) {
2187                 struct page *page;
2188
2189                 ptl = pmd_lock(mm, pmdp);
2190                 if (unlikely(!pmd_trans_huge(*pmdp))) {
2191                         spin_unlock(ptl);
2192                         goto again;
2193                 }
2194
2195                 page = pmd_page(*pmdp);
2196                 if (is_huge_zero_page(page)) {
2197                         spin_unlock(ptl);
2198                         split_huge_pmd(vma, pmdp, addr);
2199                         if (pmd_trans_unstable(pmdp))
2200                                 return migrate_vma_collect_skip(start, end,
2201                                                                 walk);
2202                 } else {
2203                         int ret;
2204
2205                         get_page(page);
2206                         spin_unlock(ptl);
2207                         if (unlikely(!trylock_page(page)))
2208                                 return migrate_vma_collect_skip(start, end,
2209                                                                 walk);
2210                         ret = split_huge_page(page);
2211                         unlock_page(page);
2212                         put_page(page);
2213                         if (ret)
2214                                 return migrate_vma_collect_skip(start, end,
2215                                                                 walk);
2216                         if (pmd_none(*pmdp))
2217                                 return migrate_vma_collect_hole(start, end,
2218                                                                 walk);
2219                 }
2220         }
2221
2222         if (unlikely(pmd_bad(*pmdp)))
2223                 return migrate_vma_collect_skip(start, end, walk);
2224
2225         ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2226         arch_enter_lazy_mmu_mode();
2227
2228         for (; addr < end; addr += PAGE_SIZE, ptep++) {
2229                 unsigned long mpfn, pfn;
2230                 struct page *page;
2231                 swp_entry_t entry;
2232                 pte_t pte;
2233
2234                 pte = *ptep;
2235                 pfn = pte_pfn(pte);
2236
2237                 if (pte_none(pte)) {
2238                         mpfn = MIGRATE_PFN_MIGRATE;
2239                         migrate->cpages++;
2240                         pfn = 0;
2241                         goto next;
2242                 }
2243
2244                 if (!pte_present(pte)) {
2245                         mpfn = pfn = 0;
2246
2247                         /*
2248                          * Only care about unaddressable device page special
2249                          * page table entry. Other special swap entries are not
2250                          * migratable, and we ignore regular swapped page.
2251                          */
2252                         entry = pte_to_swp_entry(pte);
2253                         if (!is_device_private_entry(entry))
2254                                 goto next;
2255
2256                         page = device_private_entry_to_page(entry);
2257                         mpfn = migrate_pfn(page_to_pfn(page))|
2258                                 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2259                         if (is_write_device_private_entry(entry))
2260                                 mpfn |= MIGRATE_PFN_WRITE;
2261                 } else {
2262                         if (is_zero_pfn(pfn)) {
2263                                 mpfn = MIGRATE_PFN_MIGRATE;
2264                                 migrate->cpages++;
2265                                 pfn = 0;
2266                                 goto next;
2267                         }
2268                         page = _vm_normal_page(migrate->vma, addr, pte, true);
2269                         mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2270                         mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2271                 }
2272
2273                 /* FIXME support THP */
2274                 if (!page || !page->mapping || PageTransCompound(page)) {
2275                         mpfn = pfn = 0;
2276                         goto next;
2277                 }
2278                 pfn = page_to_pfn(page);
2279
2280                 /*
2281                  * By getting a reference on the page we pin it and that blocks
2282                  * any kind of migration. Side effect is that it "freezes" the
2283                  * pte.
2284                  *
2285                  * We drop this reference after isolating the page from the lru
2286                  * for non device page (device page are not on the lru and thus
2287                  * can't be dropped from it).
2288                  */
2289                 get_page(page);
2290                 migrate->cpages++;
2291
2292                 /*
2293                  * Optimize for the common case where page is only mapped once
2294                  * in one process. If we can lock the page, then we can safely
2295                  * set up a special migration page table entry now.
2296                  */
2297                 if (trylock_page(page)) {
2298                         pte_t swp_pte;
2299
2300                         mpfn |= MIGRATE_PFN_LOCKED;
2301                         ptep_get_and_clear(mm, addr, ptep);
2302
2303                         /* Setup special migration page table entry */
2304                         entry = make_migration_entry(page, mpfn &
2305                                                      MIGRATE_PFN_WRITE);
2306                         swp_pte = swp_entry_to_pte(entry);
2307                         if (pte_soft_dirty(pte))
2308                                 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2309                         set_pte_at(mm, addr, ptep, swp_pte);
2310
2311                         /*
2312                          * This is like regular unmap: we remove the rmap and
2313                          * drop page refcount. Page won't be freed, as we took
2314                          * a reference just above.
2315                          */
2316                         page_remove_rmap(page, false);
2317                         put_page(page);
2318
2319                         if (pte_present(pte))
2320                                 unmapped++;
2321                 }
2322
2323 next:
2324                 migrate->dst[migrate->npages] = 0;
2325                 migrate->src[migrate->npages++] = mpfn;
2326         }
2327         arch_leave_lazy_mmu_mode();
2328         pte_unmap_unlock(ptep - 1, ptl);
2329
2330         /* Only flush the TLB if we actually modified any entries */
2331         if (unmapped)
2332                 flush_tlb_range(walk->vma, start, end);
2333
2334         return 0;
2335 }
2336
2337 /*
2338  * migrate_vma_collect() - collect pages over a range of virtual addresses
2339  * @migrate: migrate struct containing all migration information
2340  *
2341  * This will walk the CPU page table. For each virtual address backed by a
2342  * valid page, it updates the src array and takes a reference on the page, in
2343  * order to pin the page until we lock it and unmap it.
2344  */
2345 static void migrate_vma_collect(struct migrate_vma *migrate)
2346 {
2347         struct mmu_notifier_range range;
2348         struct mm_walk mm_walk;
2349
2350         mm_walk.pmd_entry = migrate_vma_collect_pmd;
2351         mm_walk.pte_entry = NULL;
2352         mm_walk.pte_hole = migrate_vma_collect_hole;
2353         mm_walk.hugetlb_entry = NULL;
2354         mm_walk.test_walk = NULL;
2355         mm_walk.vma = migrate->vma;
2356         mm_walk.mm = migrate->vma->vm_mm;
2357         mm_walk.private = migrate;
2358
2359         mmu_notifier_range_init(&range, mm_walk.mm, migrate->start,
2360                                 migrate->end);
2361         mmu_notifier_invalidate_range_start(&range);
2362         walk_page_range(migrate->start, migrate->end, &mm_walk);
2363         mmu_notifier_invalidate_range_end(&range);
2364
2365         migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2366 }
2367
2368 /*
2369  * migrate_vma_check_page() - check if page is pinned or not
2370  * @page: struct page to check
2371  *
2372  * Pinned pages cannot be migrated. This is the same test as in
2373  * migrate_page_move_mapping(), except that here we allow migration of a
2374  * ZONE_DEVICE page.
2375  */
2376 static bool migrate_vma_check_page(struct page *page)
2377 {
2378         /*
2379          * One extra ref because caller holds an extra reference, either from
2380          * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2381          * a device page.
2382          */
2383         int extra = 1;
2384
2385         /*
2386          * FIXME support THP (transparent huge page), it is bit more complex to
2387          * check them than regular pages, because they can be mapped with a pmd
2388          * or with a pte (split pte mapping).
2389          */
2390         if (PageCompound(page))
2391                 return false;
2392
2393         /* Page from ZONE_DEVICE have one extra reference */
2394         if (is_zone_device_page(page)) {
2395                 /*
2396                  * Private page can never be pin as they have no valid pte and
2397                  * GUP will fail for those. Yet if there is a pending migration
2398                  * a thread might try to wait on the pte migration entry and
2399                  * will bump the page reference count. Sadly there is no way to
2400                  * differentiate a regular pin from migration wait. Hence to
2401                  * avoid 2 racing thread trying to migrate back to CPU to enter
2402                  * infinite loop (one stoping migration because the other is
2403                  * waiting on pte migration entry). We always return true here.
2404                  *
2405                  * FIXME proper solution is to rework migration_entry_wait() so
2406                  * it does not need to take a reference on page.
2407                  */
2408                 if (is_device_private_page(page))
2409                         return true;
2410
2411                 /*
2412                  * Only allow device public page to be migrated and account for
2413                  * the extra reference count imply by ZONE_DEVICE pages.
2414                  */
2415                 if (!is_device_public_page(page))
2416                         return false;
2417                 extra++;
2418         }
2419
2420         /* For file back page */
2421         if (page_mapping(page))
2422                 extra += 1 + page_has_private(page);
2423
2424         if ((page_count(page) - extra) > page_mapcount(page))
2425                 return false;
2426
2427         return true;
2428 }
2429
2430 /*
2431  * migrate_vma_prepare() - lock pages and isolate them from the lru
2432  * @migrate: migrate struct containing all migration information
2433  *
2434  * This locks pages that have been collected by migrate_vma_collect(). Once each
2435  * page is locked it is isolated from the lru (for non-device pages). Finally,
2436  * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2437  * migrated by concurrent kernel threads.
2438  */
2439 static void migrate_vma_prepare(struct migrate_vma *migrate)
2440 {
2441         const unsigned long npages = migrate->npages;
2442         const unsigned long start = migrate->start;
2443         unsigned long addr, i, restore = 0;
2444         bool allow_drain = true;
2445
2446         lru_add_drain();
2447
2448         for (i = 0; (i < npages) && migrate->cpages; i++) {
2449                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2450                 bool remap = true;
2451
2452                 if (!page)
2453                         continue;
2454
2455                 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2456                         /*
2457                          * Because we are migrating several pages there can be
2458                          * a deadlock between 2 concurrent migration where each
2459                          * are waiting on each other page lock.
2460                          *
2461                          * Make migrate_vma() a best effort thing and backoff
2462                          * for any page we can not lock right away.
2463                          */
2464                         if (!trylock_page(page)) {
2465                                 migrate->src[i] = 0;
2466                                 migrate->cpages--;
2467                                 put_page(page);
2468                                 continue;
2469                         }
2470                         remap = false;
2471                         migrate->src[i] |= MIGRATE_PFN_LOCKED;
2472                 }
2473
2474                 /* ZONE_DEVICE pages are not on LRU */
2475                 if (!is_zone_device_page(page)) {
2476                         if (!PageLRU(page) && allow_drain) {
2477                                 /* Drain CPU's pagevec */
2478                                 lru_add_drain_all();
2479                                 allow_drain = false;
2480                         }
2481
2482                         if (isolate_lru_page(page)) {
2483                                 if (remap) {
2484                                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2485                                         migrate->cpages--;
2486                                         restore++;
2487                                 } else {
2488                                         migrate->src[i] = 0;
2489                                         unlock_page(page);
2490                                         migrate->cpages--;
2491                                         put_page(page);
2492                                 }
2493                                 continue;
2494                         }
2495
2496                         /* Drop the reference we took in collect */
2497                         put_page(page);
2498                 }
2499
2500                 if (!migrate_vma_check_page(page)) {
2501                         if (remap) {
2502                                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2503                                 migrate->cpages--;
2504                                 restore++;
2505
2506                                 if (!is_zone_device_page(page)) {
2507                                         get_page(page);
2508                                         putback_lru_page(page);
2509                                 }
2510                         } else {
2511                                 migrate->src[i] = 0;
2512                                 unlock_page(page);
2513                                 migrate->cpages--;
2514
2515                                 if (!is_zone_device_page(page))
2516                                         putback_lru_page(page);
2517                                 else
2518                                         put_page(page);
2519                         }
2520                 }
2521         }
2522
2523         for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2524                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2525
2526                 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2527                         continue;
2528
2529                 remove_migration_pte(page, migrate->vma, addr, page);
2530
2531                 migrate->src[i] = 0;
2532                 unlock_page(page);
2533                 put_page(page);
2534                 restore--;
2535         }
2536 }
2537
2538 /*
2539  * migrate_vma_unmap() - replace page mapping with special migration pte entry
2540  * @migrate: migrate struct containing all migration information
2541  *
2542  * Replace page mapping (CPU page table pte) with a special migration pte entry
2543  * and check again if it has been pinned. Pinned pages are restored because we
2544  * cannot migrate them.
2545  *
2546  * This is the last step before we call the device driver callback to allocate
2547  * destination memory and copy contents of original page over to new page.
2548  */
2549 static void migrate_vma_unmap(struct migrate_vma *migrate)
2550 {
2551         int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2552         const unsigned long npages = migrate->npages;
2553         const unsigned long start = migrate->start;
2554         unsigned long addr, i, restore = 0;
2555
2556         for (i = 0; i < npages; 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                 if (page_mapped(page)) {
2563                         try_to_unmap(page, flags);
2564                         if (page_mapped(page))
2565                                 goto restore;
2566                 }
2567
2568                 if (migrate_vma_check_page(page))
2569                         continue;
2570
2571 restore:
2572                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2573                 migrate->cpages--;
2574                 restore++;
2575         }
2576
2577         for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2578                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2579
2580                 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2581                         continue;
2582
2583                 remove_migration_ptes(page, page, false);
2584
2585                 migrate->src[i] = 0;
2586                 unlock_page(page);
2587                 restore--;
2588
2589                 if (is_zone_device_page(page))
2590                         put_page(page);
2591                 else
2592                         putback_lru_page(page);
2593         }
2594 }
2595
2596 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2597                                     unsigned long addr,
2598                                     struct page *page,
2599                                     unsigned long *src,
2600                                     unsigned long *dst)
2601 {
2602         struct vm_area_struct *vma = migrate->vma;
2603         struct mm_struct *mm = vma->vm_mm;
2604         struct mem_cgroup *memcg;
2605         bool flush = false;
2606         spinlock_t *ptl;
2607         pte_t entry;
2608         pgd_t *pgdp;
2609         p4d_t *p4dp;
2610         pud_t *pudp;
2611         pmd_t *pmdp;
2612         pte_t *ptep;
2613
2614         /* Only allow populating anonymous memory */
2615         if (!vma_is_anonymous(vma))
2616                 goto abort;
2617
2618         pgdp = pgd_offset(mm, addr);
2619         p4dp = p4d_alloc(mm, pgdp, addr);
2620         if (!p4dp)
2621                 goto abort;
2622         pudp = pud_alloc(mm, p4dp, addr);
2623         if (!pudp)
2624                 goto abort;
2625         pmdp = pmd_alloc(mm, pudp, addr);
2626         if (!pmdp)
2627                 goto abort;
2628
2629         if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2630                 goto abort;
2631
2632         /*
2633          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2634          * pte_offset_map() on pmds where a huge pmd might be created
2635          * from a different thread.
2636          *
2637          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2638          * parallel threads are excluded by other means.
2639          *
2640          * Here we only have down_read(mmap_sem).
2641          */
2642         if (pte_alloc(mm, pmdp))
2643                 goto abort;
2644
2645         /* See the comment in pte_alloc_one_map() */
2646         if (unlikely(pmd_trans_unstable(pmdp)))
2647                 goto abort;
2648
2649         if (unlikely(anon_vma_prepare(vma)))
2650                 goto abort;
2651         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2652                 goto abort;
2653
2654         /*
2655          * The memory barrier inside __SetPageUptodate makes sure that
2656          * preceding stores to the page contents become visible before
2657          * the set_pte_at() write.
2658          */
2659         __SetPageUptodate(page);
2660
2661         if (is_zone_device_page(page)) {
2662                 if (is_device_private_page(page)) {
2663                         swp_entry_t swp_entry;
2664
2665                         swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2666                         entry = swp_entry_to_pte(swp_entry);
2667                 } else if (is_device_public_page(page)) {
2668                         entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2669                         if (vma->vm_flags & VM_WRITE)
2670                                 entry = pte_mkwrite(pte_mkdirty(entry));
2671                         entry = pte_mkdevmap(entry);
2672                 }
2673         } else {
2674                 entry = mk_pte(page, vma->vm_page_prot);
2675                 if (vma->vm_flags & VM_WRITE)
2676                         entry = pte_mkwrite(pte_mkdirty(entry));
2677         }
2678
2679         ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2680
2681         if (pte_present(*ptep)) {
2682                 unsigned long pfn = pte_pfn(*ptep);
2683
2684                 if (!is_zero_pfn(pfn)) {
2685                         pte_unmap_unlock(ptep, ptl);
2686                         mem_cgroup_cancel_charge(page, memcg, false);
2687                         goto abort;
2688                 }
2689                 flush = true;
2690         } else if (!pte_none(*ptep)) {
2691                 pte_unmap_unlock(ptep, ptl);
2692                 mem_cgroup_cancel_charge(page, memcg, false);
2693                 goto abort;
2694         }
2695
2696         /*
2697          * Check for usefaultfd but do not deliver the fault. Instead,
2698          * just back off.
2699          */
2700         if (userfaultfd_missing(vma)) {
2701                 pte_unmap_unlock(ptep, ptl);
2702                 mem_cgroup_cancel_charge(page, memcg, false);
2703                 goto abort;
2704         }
2705
2706         inc_mm_counter(mm, MM_ANONPAGES);
2707         page_add_new_anon_rmap(page, vma, addr, false);
2708         mem_cgroup_commit_charge(page, memcg, false, false);
2709         if (!is_zone_device_page(page))
2710                 lru_cache_add_active_or_unevictable(page, vma);
2711         get_page(page);
2712
2713         if (flush) {
2714                 flush_cache_page(vma, addr, pte_pfn(*ptep));
2715                 ptep_clear_flush_notify(vma, addr, ptep);
2716                 set_pte_at_notify(mm, addr, ptep, entry);
2717                 update_mmu_cache(vma, addr, ptep);
2718         } else {
2719                 /* No need to invalidate - it was non-present before */
2720                 set_pte_at(mm, addr, ptep, entry);
2721                 update_mmu_cache(vma, addr, ptep);
2722         }
2723
2724         pte_unmap_unlock(ptep, ptl);
2725         *src = MIGRATE_PFN_MIGRATE;
2726         return;
2727
2728 abort:
2729         *src &= ~MIGRATE_PFN_MIGRATE;
2730 }
2731
2732 /*
2733  * migrate_vma_pages() - migrate meta-data from src page to dst page
2734  * @migrate: migrate struct containing all migration information
2735  *
2736  * This migrates struct page meta-data from source struct page to destination
2737  * struct page. This effectively finishes the migration from source page to the
2738  * destination page.
2739  */
2740 static void migrate_vma_pages(struct migrate_vma *migrate)
2741 {
2742         const unsigned long npages = migrate->npages;
2743         const unsigned long start = migrate->start;
2744         struct mmu_notifier_range range;
2745         unsigned long addr, i;
2746         bool notified = false;
2747
2748         for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2749                 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2750                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2751                 struct address_space *mapping;
2752                 int r;
2753
2754                 if (!newpage) {
2755                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2756                         continue;
2757                 }
2758
2759                 if (!page) {
2760                         if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2761                                 continue;
2762                         }
2763                         if (!notified) {
2764                                 notified = true;
2765
2766                                 mmu_notifier_range_init(&range,
2767                                                         migrate->vma->vm_mm,
2768                                                         addr, migrate->end);
2769                                 mmu_notifier_invalidate_range_start(&range);
2770                         }
2771                         migrate_vma_insert_page(migrate, addr, newpage,
2772                                                 &migrate->src[i],
2773                                                 &migrate->dst[i]);
2774                         continue;
2775                 }
2776
2777                 mapping = page_mapping(page);
2778
2779                 if (is_zone_device_page(newpage)) {
2780                         if (is_device_private_page(newpage)) {
2781                                 /*
2782                                  * For now only support private anonymous when
2783                                  * migrating to un-addressable device memory.
2784                                  */
2785                                 if (mapping) {
2786                                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2787                                         continue;
2788                                 }
2789                         } else if (!is_device_public_page(newpage)) {
2790                                 /*
2791                                  * Other types of ZONE_DEVICE page are not
2792                                  * supported.
2793                                  */
2794                                 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2795                                 continue;
2796                         }
2797                 }
2798
2799                 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2800                 if (r != MIGRATEPAGE_SUCCESS)
2801                         migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2802         }
2803
2804         /*
2805          * No need to double call mmu_notifier->invalidate_range() callback as
2806          * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2807          * did already call it.
2808          */
2809         if (notified)
2810                 mmu_notifier_invalidate_range_only_end(&range);
2811 }
2812
2813 /*
2814  * migrate_vma_finalize() - restore CPU page table entry
2815  * @migrate: migrate struct containing all migration information
2816  *
2817  * This replaces the special migration pte entry with either a mapping to the
2818  * new page if migration was successful for that page, or to the original page
2819  * otherwise.
2820  *
2821  * This also unlocks the pages and puts them back on the lru, or drops the extra
2822  * refcount, for device pages.
2823  */
2824 static void migrate_vma_finalize(struct migrate_vma *migrate)
2825 {
2826         const unsigned long npages = migrate->npages;
2827         unsigned long i;
2828
2829         for (i = 0; i < npages; i++) {
2830                 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2831                 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2832
2833                 if (!page) {
2834                         if (newpage) {
2835                                 unlock_page(newpage);
2836                                 put_page(newpage);
2837                         }
2838                         continue;
2839                 }
2840
2841                 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2842                         if (newpage) {
2843                                 unlock_page(newpage);
2844                                 put_page(newpage);
2845                         }
2846                         newpage = page;
2847                 }
2848
2849                 remove_migration_ptes(page, newpage, false);
2850                 unlock_page(page);
2851                 migrate->cpages--;
2852
2853                 if (is_zone_device_page(page))
2854                         put_page(page);
2855                 else
2856                         putback_lru_page(page);
2857
2858                 if (newpage != page) {
2859                         unlock_page(newpage);
2860                         if (is_zone_device_page(newpage))
2861                                 put_page(newpage);
2862                         else
2863                                 putback_lru_page(newpage);
2864                 }
2865         }
2866 }
2867
2868 /*
2869  * migrate_vma() - migrate a range of memory inside vma
2870  *
2871  * @ops: migration callback for allocating destination memory and copying
2872  * @vma: virtual memory area containing the range to be migrated
2873  * @start: start address of the range to migrate (inclusive)
2874  * @end: end address of the range to migrate (exclusive)
2875  * @src: array of hmm_pfn_t containing source pfns
2876  * @dst: array of hmm_pfn_t containing destination pfns
2877  * @private: pointer passed back to each of the callback
2878  * Returns: 0 on success, error code otherwise
2879  *
2880  * This function tries to migrate a range of memory virtual address range, using
2881  * callbacks to allocate and copy memory from source to destination. First it
2882  * collects all the pages backing each virtual address in the range, saving this
2883  * inside the src array. Then it locks those pages and unmaps them. Once the pages
2884  * are locked and unmapped, it checks whether each page is pinned or not. Pages
2885  * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2886  * in the corresponding src array entry. It then restores any pages that are
2887  * pinned, by remapping and unlocking those pages.
2888  *
2889  * At this point it calls the alloc_and_copy() callback. For documentation on
2890  * what is expected from that callback, see struct migrate_vma_ops comments in
2891  * include/linux/migrate.h
2892  *
2893  * After the alloc_and_copy() callback, this function goes over each entry in
2894  * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2895  * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2896  * then the function tries to migrate struct page information from the source
2897  * struct page to the destination struct page. If it fails to migrate the struct
2898  * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2899  * array.
2900  *
2901  * At this point all successfully migrated pages have an entry in the src
2902  * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2903  * array entry with MIGRATE_PFN_VALID flag set.
2904  *
2905  * It then calls the finalize_and_map() callback. See comments for "struct
2906  * migrate_vma_ops", in include/linux/migrate.h for details about
2907  * finalize_and_map() behavior.
2908  *
2909  * After the finalize_and_map() callback, for successfully migrated pages, this
2910  * function updates the CPU page table to point to new pages, otherwise it
2911  * restores the CPU page table to point to the original source pages.
2912  *
2913  * Function returns 0 after the above steps, even if no pages were migrated
2914  * (The function only returns an error if any of the arguments are invalid.)
2915  *
2916  * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2917  * unsigned long entries.
2918  */
2919 int migrate_vma(const struct migrate_vma_ops *ops,
2920                 struct vm_area_struct *vma,
2921                 unsigned long start,
2922                 unsigned long end,
2923                 unsigned long *src,
2924                 unsigned long *dst,
2925                 void *private)
2926 {
2927         struct migrate_vma migrate;
2928
2929         /* Sanity check the arguments */
2930         start &= PAGE_MASK;
2931         end &= PAGE_MASK;
2932         if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2933                         vma_is_dax(vma))
2934                 return -EINVAL;
2935         if (start < vma->vm_start || start >= vma->vm_end)
2936                 return -EINVAL;
2937         if (end <= vma->vm_start || end > vma->vm_end)
2938                 return -EINVAL;
2939         if (!ops || !src || !dst || start >= end)
2940                 return -EINVAL;
2941
2942         memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2943         migrate.src = src;
2944         migrate.dst = dst;
2945         migrate.start = start;
2946         migrate.npages = 0;
2947         migrate.cpages = 0;
2948         migrate.end = end;
2949         migrate.vma = vma;
2950
2951         /* Collect, and try to unmap source pages */
2952         migrate_vma_collect(&migrate);
2953         if (!migrate.cpages)
2954                 return 0;
2955
2956         /* Lock and isolate page */
2957         migrate_vma_prepare(&migrate);
2958         if (!migrate.cpages)
2959                 return 0;
2960
2961         /* Unmap pages */
2962         migrate_vma_unmap(&migrate);
2963         if (!migrate.cpages)
2964                 return 0;
2965
2966         /*
2967          * At this point pages are locked and unmapped, and thus they have
2968          * stable content and can safely be copied to destination memory that
2969          * is allocated by the callback.
2970          *
2971          * Note that migration can fail in migrate_vma_struct_page() for each
2972          * individual page.
2973          */
2974         ops->alloc_and_copy(vma, src, dst, start, end, private);
2975
2976         /* This does the real migration of struct page */
2977         migrate_vma_pages(&migrate);
2978
2979         ops->finalize_and_map(vma, src, dst, start, end, private);
2980
2981         /* Unlock and remap pages */
2982         migrate_vma_finalize(&migrate);
2983
2984         return 0;
2985 }
2986 EXPORT_SYMBOL(migrate_vma);
2987 #endif /* defined(MIGRATE_VMA_HELPER) */