Loading...
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/pagewalk.h>
42#include <linux/pfn_t.h>
43#include <linux/memremap.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/balloon_compaction.h>
46#include <linux/mmu_notifier.h>
47#include <linux/page_idle.h>
48#include <linux/page_owner.h>
49#include <linux/sched/mm.h>
50#include <linux/ptrace.h>
51#include <linux/oom.h>
52
53#include <asm/tlbflush.h>
54
55#define CREATE_TRACE_POINTS
56#include <trace/events/migrate.h>
57
58#include "internal.h"
59
60int isolate_movable_page(struct page *page, isolate_mode_t mode)
61{
62 struct address_space *mapping;
63
64 /*
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
67 *
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
72 */
73 if (unlikely(!get_page_unless_zero(page)))
74 goto out;
75
76 /*
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
80 */
81 if (unlikely(!__PageMovable(page)))
82 goto out_putpage;
83 /*
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
87 *
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
93 */
94 if (unlikely(!trylock_page(page)))
95 goto out_putpage;
96
97 if (!PageMovable(page) || PageIsolated(page))
98 goto out_no_isolated;
99
100 mapping = page_mapping(page);
101 VM_BUG_ON_PAGE(!mapping, page);
102
103 if (!mapping->a_ops->isolate_page(page, mode))
104 goto out_no_isolated;
105
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 __SetPageIsolated(page);
109 unlock_page(page);
110
111 return 0;
112
113out_no_isolated:
114 unlock_page(page);
115out_putpage:
116 put_page(page);
117out:
118 return -EBUSY;
119}
120
121static void putback_movable_page(struct page *page)
122{
123 struct address_space *mapping;
124
125 mapping = page_mapping(page);
126 mapping->a_ops->putback_page(page);
127 __ClearPageIsolated(page);
128}
129
130/*
131 * Put previously isolated pages back onto the appropriate lists
132 * from where they were once taken off for compaction/migration.
133 *
134 * This function shall be used whenever the isolated pageset has been
135 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136 * and isolate_huge_page().
137 */
138void putback_movable_pages(struct list_head *l)
139{
140 struct page *page;
141 struct page *page2;
142
143 list_for_each_entry_safe(page, page2, l, lru) {
144 if (unlikely(PageHuge(page))) {
145 putback_active_hugepage(page);
146 continue;
147 }
148 list_del(&page->lru);
149 /*
150 * We isolated non-lru movable page so here we can use
151 * __PageMovable because LRU page's mapping cannot have
152 * PAGE_MAPPING_MOVABLE.
153 */
154 if (unlikely(__PageMovable(page))) {
155 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 lock_page(page);
157 if (PageMovable(page))
158 putback_movable_page(page);
159 else
160 __ClearPageIsolated(page);
161 unlock_page(page);
162 put_page(page);
163 } else {
164 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
165 page_is_file_lru(page), -thp_nr_pages(page));
166 putback_lru_page(page);
167 }
168 }
169}
170
171/*
172 * Restore a potential migration pte to a working pte entry
173 */
174static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
175 unsigned long addr, void *old)
176{
177 struct page_vma_mapped_walk pvmw = {
178 .page = old,
179 .vma = vma,
180 .address = addr,
181 .flags = PVMW_SYNC | PVMW_MIGRATION,
182 };
183 struct page *new;
184 pte_t pte;
185 swp_entry_t entry;
186
187 VM_BUG_ON_PAGE(PageTail(page), page);
188 while (page_vma_mapped_walk(&pvmw)) {
189 if (PageKsm(page))
190 new = page;
191 else
192 new = page - pvmw.page->index +
193 linear_page_index(vma, pvmw.address);
194
195#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
196 /* PMD-mapped THP migration entry */
197 if (!pvmw.pte) {
198 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
199 remove_migration_pmd(&pvmw, new);
200 continue;
201 }
202#endif
203
204 get_page(new);
205 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
206 if (pte_swp_soft_dirty(*pvmw.pte))
207 pte = pte_mksoft_dirty(pte);
208
209 /*
210 * Recheck VMA as permissions can change since migration started
211 */
212 entry = pte_to_swp_entry(*pvmw.pte);
213 if (is_writable_migration_entry(entry))
214 pte = maybe_mkwrite(pte, vma);
215 else if (pte_swp_uffd_wp(*pvmw.pte))
216 pte = pte_mkuffd_wp(pte);
217
218 if (unlikely(is_device_private_page(new))) {
219 if (pte_write(pte))
220 entry = make_writable_device_private_entry(
221 page_to_pfn(new));
222 else
223 entry = make_readable_device_private_entry(
224 page_to_pfn(new));
225 pte = swp_entry_to_pte(entry);
226 if (pte_swp_soft_dirty(*pvmw.pte))
227 pte = pte_swp_mksoft_dirty(pte);
228 if (pte_swp_uffd_wp(*pvmw.pte))
229 pte = pte_swp_mkuffd_wp(pte);
230 }
231
232#ifdef CONFIG_HUGETLB_PAGE
233 if (PageHuge(new)) {
234 unsigned int shift = huge_page_shift(hstate_vma(vma));
235
236 pte = pte_mkhuge(pte);
237 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
238 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
239 if (PageAnon(new))
240 hugepage_add_anon_rmap(new, vma, pvmw.address);
241 else
242 page_dup_rmap(new, true);
243 } else
244#endif
245 {
246 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
247
248 if (PageAnon(new))
249 page_add_anon_rmap(new, vma, pvmw.address, false);
250 else
251 page_add_file_rmap(new, false);
252 }
253 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
254 mlock_vma_page(new);
255
256 if (PageTransHuge(page) && PageMlocked(page))
257 clear_page_mlock(page);
258
259 /* No need to invalidate - it was non-present before */
260 update_mmu_cache(vma, pvmw.address, pvmw.pte);
261 }
262
263 return true;
264}
265
266/*
267 * Get rid of all migration entries and replace them by
268 * references to the indicated page.
269 */
270void remove_migration_ptes(struct page *old, struct page *new, bool locked)
271{
272 struct rmap_walk_control rwc = {
273 .rmap_one = remove_migration_pte,
274 .arg = old,
275 };
276
277 if (locked)
278 rmap_walk_locked(new, &rwc);
279 else
280 rmap_walk(new, &rwc);
281}
282
283/*
284 * Something used the pte of a page under migration. We need to
285 * get to the page and wait until migration is finished.
286 * When we return from this function the fault will be retried.
287 */
288void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
289 spinlock_t *ptl)
290{
291 pte_t pte;
292 swp_entry_t entry;
293 struct page *page;
294
295 spin_lock(ptl);
296 pte = *ptep;
297 if (!is_swap_pte(pte))
298 goto out;
299
300 entry = pte_to_swp_entry(pte);
301 if (!is_migration_entry(entry))
302 goto out;
303
304 page = pfn_swap_entry_to_page(entry);
305 page = compound_head(page);
306
307 /*
308 * Once page cache replacement of page migration started, page_count
309 * is zero; but we must not call put_and_wait_on_page_locked() without
310 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
311 */
312 if (!get_page_unless_zero(page))
313 goto out;
314 pte_unmap_unlock(ptep, ptl);
315 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
316 return;
317out:
318 pte_unmap_unlock(ptep, ptl);
319}
320
321void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
322 unsigned long address)
323{
324 spinlock_t *ptl = pte_lockptr(mm, pmd);
325 pte_t *ptep = pte_offset_map(pmd, address);
326 __migration_entry_wait(mm, ptep, ptl);
327}
328
329void migration_entry_wait_huge(struct vm_area_struct *vma,
330 struct mm_struct *mm, pte_t *pte)
331{
332 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
333 __migration_entry_wait(mm, pte, ptl);
334}
335
336#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
337void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
338{
339 spinlock_t *ptl;
340 struct page *page;
341
342 ptl = pmd_lock(mm, pmd);
343 if (!is_pmd_migration_entry(*pmd))
344 goto unlock;
345 page = pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd));
346 if (!get_page_unless_zero(page))
347 goto unlock;
348 spin_unlock(ptl);
349 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
350 return;
351unlock:
352 spin_unlock(ptl);
353}
354#endif
355
356static int expected_page_refs(struct address_space *mapping, struct page *page)
357{
358 int expected_count = 1;
359
360 /*
361 * Device private pages have an extra refcount as they are
362 * ZONE_DEVICE pages.
363 */
364 expected_count += is_device_private_page(page);
365 if (mapping)
366 expected_count += thp_nr_pages(page) + page_has_private(page);
367
368 return expected_count;
369}
370
371/*
372 * Replace the page in the mapping.
373 *
374 * The number of remaining references must be:
375 * 1 for anonymous pages without a mapping
376 * 2 for pages with a mapping
377 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
378 */
379int migrate_page_move_mapping(struct address_space *mapping,
380 struct page *newpage, struct page *page, int extra_count)
381{
382 XA_STATE(xas, &mapping->i_pages, page_index(page));
383 struct zone *oldzone, *newzone;
384 int dirty;
385 int expected_count = expected_page_refs(mapping, page) + extra_count;
386 int nr = thp_nr_pages(page);
387
388 if (!mapping) {
389 /* Anonymous page without mapping */
390 if (page_count(page) != expected_count)
391 return -EAGAIN;
392
393 /* No turning back from here */
394 newpage->index = page->index;
395 newpage->mapping = page->mapping;
396 if (PageSwapBacked(page))
397 __SetPageSwapBacked(newpage);
398
399 return MIGRATEPAGE_SUCCESS;
400 }
401
402 oldzone = page_zone(page);
403 newzone = page_zone(newpage);
404
405 xas_lock_irq(&xas);
406 if (page_count(page) != expected_count || xas_load(&xas) != page) {
407 xas_unlock_irq(&xas);
408 return -EAGAIN;
409 }
410
411 if (!page_ref_freeze(page, expected_count)) {
412 xas_unlock_irq(&xas);
413 return -EAGAIN;
414 }
415
416 /*
417 * Now we know that no one else is looking at the page:
418 * no turning back from here.
419 */
420 newpage->index = page->index;
421 newpage->mapping = page->mapping;
422 page_ref_add(newpage, nr); /* add cache reference */
423 if (PageSwapBacked(page)) {
424 __SetPageSwapBacked(newpage);
425 if (PageSwapCache(page)) {
426 SetPageSwapCache(newpage);
427 set_page_private(newpage, page_private(page));
428 }
429 } else {
430 VM_BUG_ON_PAGE(PageSwapCache(page), page);
431 }
432
433 /* Move dirty while page refs frozen and newpage not yet exposed */
434 dirty = PageDirty(page);
435 if (dirty) {
436 ClearPageDirty(page);
437 SetPageDirty(newpage);
438 }
439
440 xas_store(&xas, newpage);
441 if (PageTransHuge(page)) {
442 int i;
443
444 for (i = 1; i < nr; i++) {
445 xas_next(&xas);
446 xas_store(&xas, newpage);
447 }
448 }
449
450 /*
451 * Drop cache reference from old page by unfreezing
452 * to one less reference.
453 * We know this isn't the last reference.
454 */
455 page_ref_unfreeze(page, expected_count - nr);
456
457 xas_unlock(&xas);
458 /* Leave irq disabled to prevent preemption while updating stats */
459
460 /*
461 * If moved to a different zone then also account
462 * the page for that zone. Other VM counters will be
463 * taken care of when we establish references to the
464 * new page and drop references to the old page.
465 *
466 * Note that anonymous pages are accounted for
467 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
468 * are mapped to swap space.
469 */
470 if (newzone != oldzone) {
471 struct lruvec *old_lruvec, *new_lruvec;
472 struct mem_cgroup *memcg;
473
474 memcg = page_memcg(page);
475 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
476 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
477
478 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
479 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
480 if (PageSwapBacked(page) && !PageSwapCache(page)) {
481 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
482 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
483 }
484#ifdef CONFIG_SWAP
485 if (PageSwapCache(page)) {
486 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
487 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
488 }
489#endif
490 if (dirty && mapping_can_writeback(mapping)) {
491 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
492 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
493 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
494 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
495 }
496 }
497 local_irq_enable();
498
499 return MIGRATEPAGE_SUCCESS;
500}
501EXPORT_SYMBOL(migrate_page_move_mapping);
502
503/*
504 * The expected number of remaining references is the same as that
505 * of migrate_page_move_mapping().
506 */
507int migrate_huge_page_move_mapping(struct address_space *mapping,
508 struct page *newpage, struct page *page)
509{
510 XA_STATE(xas, &mapping->i_pages, page_index(page));
511 int expected_count;
512
513 xas_lock_irq(&xas);
514 expected_count = 2 + page_has_private(page);
515 if (page_count(page) != expected_count || xas_load(&xas) != page) {
516 xas_unlock_irq(&xas);
517 return -EAGAIN;
518 }
519
520 if (!page_ref_freeze(page, expected_count)) {
521 xas_unlock_irq(&xas);
522 return -EAGAIN;
523 }
524
525 newpage->index = page->index;
526 newpage->mapping = page->mapping;
527
528 get_page(newpage);
529
530 xas_store(&xas, newpage);
531
532 page_ref_unfreeze(page, expected_count - 1);
533
534 xas_unlock_irq(&xas);
535
536 return MIGRATEPAGE_SUCCESS;
537}
538
539/*
540 * Copy the page to its new location
541 */
542void migrate_page_states(struct page *newpage, struct page *page)
543{
544 int cpupid;
545
546 if (PageError(page))
547 SetPageError(newpage);
548 if (PageReferenced(page))
549 SetPageReferenced(newpage);
550 if (PageUptodate(page))
551 SetPageUptodate(newpage);
552 if (TestClearPageActive(page)) {
553 VM_BUG_ON_PAGE(PageUnevictable(page), page);
554 SetPageActive(newpage);
555 } else if (TestClearPageUnevictable(page))
556 SetPageUnevictable(newpage);
557 if (PageWorkingset(page))
558 SetPageWorkingset(newpage);
559 if (PageChecked(page))
560 SetPageChecked(newpage);
561 if (PageMappedToDisk(page))
562 SetPageMappedToDisk(newpage);
563
564 /* Move dirty on pages not done by migrate_page_move_mapping() */
565 if (PageDirty(page))
566 SetPageDirty(newpage);
567
568 if (page_is_young(page))
569 set_page_young(newpage);
570 if (page_is_idle(page))
571 set_page_idle(newpage);
572
573 /*
574 * Copy NUMA information to the new page, to prevent over-eager
575 * future migrations of this same page.
576 */
577 cpupid = page_cpupid_xchg_last(page, -1);
578 page_cpupid_xchg_last(newpage, cpupid);
579
580 ksm_migrate_page(newpage, page);
581 /*
582 * Please do not reorder this without considering how mm/ksm.c's
583 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
584 */
585 if (PageSwapCache(page))
586 ClearPageSwapCache(page);
587 ClearPagePrivate(page);
588
589 /* page->private contains hugetlb specific flags */
590 if (!PageHuge(page))
591 set_page_private(page, 0);
592
593 /*
594 * If any waiters have accumulated on the new page then
595 * wake them up.
596 */
597 if (PageWriteback(newpage))
598 end_page_writeback(newpage);
599
600 /*
601 * PG_readahead shares the same bit with PG_reclaim. The above
602 * end_page_writeback() may clear PG_readahead mistakenly, so set the
603 * bit after that.
604 */
605 if (PageReadahead(page))
606 SetPageReadahead(newpage);
607
608 copy_page_owner(page, newpage);
609
610 if (!PageHuge(page))
611 mem_cgroup_migrate(page, newpage);
612}
613EXPORT_SYMBOL(migrate_page_states);
614
615void migrate_page_copy(struct page *newpage, struct page *page)
616{
617 if (PageHuge(page) || PageTransHuge(page))
618 copy_huge_page(newpage, page);
619 else
620 copy_highpage(newpage, page);
621
622 migrate_page_states(newpage, page);
623}
624EXPORT_SYMBOL(migrate_page_copy);
625
626/************************************************************
627 * Migration functions
628 ***********************************************************/
629
630/*
631 * Common logic to directly migrate a single LRU page suitable for
632 * pages that do not use PagePrivate/PagePrivate2.
633 *
634 * Pages are locked upon entry and exit.
635 */
636int migrate_page(struct address_space *mapping,
637 struct page *newpage, struct page *page,
638 enum migrate_mode mode)
639{
640 int rc;
641
642 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
643
644 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
645
646 if (rc != MIGRATEPAGE_SUCCESS)
647 return rc;
648
649 if (mode != MIGRATE_SYNC_NO_COPY)
650 migrate_page_copy(newpage, page);
651 else
652 migrate_page_states(newpage, page);
653 return MIGRATEPAGE_SUCCESS;
654}
655EXPORT_SYMBOL(migrate_page);
656
657#ifdef CONFIG_BLOCK
658/* Returns true if all buffers are successfully locked */
659static bool buffer_migrate_lock_buffers(struct buffer_head *head,
660 enum migrate_mode mode)
661{
662 struct buffer_head *bh = head;
663
664 /* Simple case, sync compaction */
665 if (mode != MIGRATE_ASYNC) {
666 do {
667 lock_buffer(bh);
668 bh = bh->b_this_page;
669
670 } while (bh != head);
671
672 return true;
673 }
674
675 /* async case, we cannot block on lock_buffer so use trylock_buffer */
676 do {
677 if (!trylock_buffer(bh)) {
678 /*
679 * We failed to lock the buffer and cannot stall in
680 * async migration. Release the taken locks
681 */
682 struct buffer_head *failed_bh = bh;
683 bh = head;
684 while (bh != failed_bh) {
685 unlock_buffer(bh);
686 bh = bh->b_this_page;
687 }
688 return false;
689 }
690
691 bh = bh->b_this_page;
692 } while (bh != head);
693 return true;
694}
695
696static int __buffer_migrate_page(struct address_space *mapping,
697 struct page *newpage, struct page *page, enum migrate_mode mode,
698 bool check_refs)
699{
700 struct buffer_head *bh, *head;
701 int rc;
702 int expected_count;
703
704 if (!page_has_buffers(page))
705 return migrate_page(mapping, newpage, page, mode);
706
707 /* Check whether page does not have extra refs before we do more work */
708 expected_count = expected_page_refs(mapping, page);
709 if (page_count(page) != expected_count)
710 return -EAGAIN;
711
712 head = page_buffers(page);
713 if (!buffer_migrate_lock_buffers(head, mode))
714 return -EAGAIN;
715
716 if (check_refs) {
717 bool busy;
718 bool invalidated = false;
719
720recheck_buffers:
721 busy = false;
722 spin_lock(&mapping->private_lock);
723 bh = head;
724 do {
725 if (atomic_read(&bh->b_count)) {
726 busy = true;
727 break;
728 }
729 bh = bh->b_this_page;
730 } while (bh != head);
731 if (busy) {
732 if (invalidated) {
733 rc = -EAGAIN;
734 goto unlock_buffers;
735 }
736 spin_unlock(&mapping->private_lock);
737 invalidate_bh_lrus();
738 invalidated = true;
739 goto recheck_buffers;
740 }
741 }
742
743 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
744 if (rc != MIGRATEPAGE_SUCCESS)
745 goto unlock_buffers;
746
747 attach_page_private(newpage, detach_page_private(page));
748
749 bh = head;
750 do {
751 set_bh_page(bh, newpage, bh_offset(bh));
752 bh = bh->b_this_page;
753
754 } while (bh != head);
755
756 if (mode != MIGRATE_SYNC_NO_COPY)
757 migrate_page_copy(newpage, page);
758 else
759 migrate_page_states(newpage, page);
760
761 rc = MIGRATEPAGE_SUCCESS;
762unlock_buffers:
763 if (check_refs)
764 spin_unlock(&mapping->private_lock);
765 bh = head;
766 do {
767 unlock_buffer(bh);
768 bh = bh->b_this_page;
769
770 } while (bh != head);
771
772 return rc;
773}
774
775/*
776 * Migration function for pages with buffers. This function can only be used
777 * if the underlying filesystem guarantees that no other references to "page"
778 * exist. For example attached buffer heads are accessed only under page lock.
779 */
780int buffer_migrate_page(struct address_space *mapping,
781 struct page *newpage, struct page *page, enum migrate_mode mode)
782{
783 return __buffer_migrate_page(mapping, newpage, page, mode, false);
784}
785EXPORT_SYMBOL(buffer_migrate_page);
786
787/*
788 * Same as above except that this variant is more careful and checks that there
789 * are also no buffer head references. This function is the right one for
790 * mappings where buffer heads are directly looked up and referenced (such as
791 * block device mappings).
792 */
793int buffer_migrate_page_norefs(struct address_space *mapping,
794 struct page *newpage, struct page *page, enum migrate_mode mode)
795{
796 return __buffer_migrate_page(mapping, newpage, page, mode, true);
797}
798#endif
799
800/*
801 * Writeback a page to clean the dirty state
802 */
803static int writeout(struct address_space *mapping, struct page *page)
804{
805 struct writeback_control wbc = {
806 .sync_mode = WB_SYNC_NONE,
807 .nr_to_write = 1,
808 .range_start = 0,
809 .range_end = LLONG_MAX,
810 .for_reclaim = 1
811 };
812 int rc;
813
814 if (!mapping->a_ops->writepage)
815 /* No write method for the address space */
816 return -EINVAL;
817
818 if (!clear_page_dirty_for_io(page))
819 /* Someone else already triggered a write */
820 return -EAGAIN;
821
822 /*
823 * A dirty page may imply that the underlying filesystem has
824 * the page on some queue. So the page must be clean for
825 * migration. Writeout may mean we loose the lock and the
826 * page state is no longer what we checked for earlier.
827 * At this point we know that the migration attempt cannot
828 * be successful.
829 */
830 remove_migration_ptes(page, page, false);
831
832 rc = mapping->a_ops->writepage(page, &wbc);
833
834 if (rc != AOP_WRITEPAGE_ACTIVATE)
835 /* unlocked. Relock */
836 lock_page(page);
837
838 return (rc < 0) ? -EIO : -EAGAIN;
839}
840
841/*
842 * Default handling if a filesystem does not provide a migration function.
843 */
844static int fallback_migrate_page(struct address_space *mapping,
845 struct page *newpage, struct page *page, enum migrate_mode mode)
846{
847 if (PageDirty(page)) {
848 /* Only writeback pages in full synchronous migration */
849 switch (mode) {
850 case MIGRATE_SYNC:
851 case MIGRATE_SYNC_NO_COPY:
852 break;
853 default:
854 return -EBUSY;
855 }
856 return writeout(mapping, page);
857 }
858
859 /*
860 * Buffers may be managed in a filesystem specific way.
861 * We must have no buffers or drop them.
862 */
863 if (page_has_private(page) &&
864 !try_to_release_page(page, GFP_KERNEL))
865 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
866
867 return migrate_page(mapping, newpage, page, mode);
868}
869
870/*
871 * Move a page to a newly allocated page
872 * The page is locked and all ptes have been successfully removed.
873 *
874 * The new page will have replaced the old page if this function
875 * is successful.
876 *
877 * Return value:
878 * < 0 - error code
879 * MIGRATEPAGE_SUCCESS - success
880 */
881static int move_to_new_page(struct page *newpage, struct page *page,
882 enum migrate_mode mode)
883{
884 struct address_space *mapping;
885 int rc = -EAGAIN;
886 bool is_lru = !__PageMovable(page);
887
888 VM_BUG_ON_PAGE(!PageLocked(page), page);
889 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
890
891 mapping = page_mapping(page);
892
893 if (likely(is_lru)) {
894 if (!mapping)
895 rc = migrate_page(mapping, newpage, page, mode);
896 else if (mapping->a_ops->migratepage)
897 /*
898 * Most pages have a mapping and most filesystems
899 * provide a migratepage callback. Anonymous pages
900 * are part of swap space which also has its own
901 * migratepage callback. This is the most common path
902 * for page migration.
903 */
904 rc = mapping->a_ops->migratepage(mapping, newpage,
905 page, mode);
906 else
907 rc = fallback_migrate_page(mapping, newpage,
908 page, mode);
909 } else {
910 /*
911 * In case of non-lru page, it could be released after
912 * isolation step. In that case, we shouldn't try migration.
913 */
914 VM_BUG_ON_PAGE(!PageIsolated(page), page);
915 if (!PageMovable(page)) {
916 rc = MIGRATEPAGE_SUCCESS;
917 __ClearPageIsolated(page);
918 goto out;
919 }
920
921 rc = mapping->a_ops->migratepage(mapping, newpage,
922 page, mode);
923 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
924 !PageIsolated(page));
925 }
926
927 /*
928 * When successful, old pagecache page->mapping must be cleared before
929 * page is freed; but stats require that PageAnon be left as PageAnon.
930 */
931 if (rc == MIGRATEPAGE_SUCCESS) {
932 if (__PageMovable(page)) {
933 VM_BUG_ON_PAGE(!PageIsolated(page), page);
934
935 /*
936 * We clear PG_movable under page_lock so any compactor
937 * cannot try to migrate this page.
938 */
939 __ClearPageIsolated(page);
940 }
941
942 /*
943 * Anonymous and movable page->mapping will be cleared by
944 * free_pages_prepare so don't reset it here for keeping
945 * the type to work PageAnon, for example.
946 */
947 if (!PageMappingFlags(page))
948 page->mapping = NULL;
949
950 if (likely(!is_zone_device_page(newpage)))
951 flush_dcache_page(newpage);
952
953 }
954out:
955 return rc;
956}
957
958static int __unmap_and_move(struct page *page, struct page *newpage,
959 int force, enum migrate_mode mode)
960{
961 int rc = -EAGAIN;
962 int page_was_mapped = 0;
963 struct anon_vma *anon_vma = NULL;
964 bool is_lru = !__PageMovable(page);
965
966 if (!trylock_page(page)) {
967 if (!force || mode == MIGRATE_ASYNC)
968 goto out;
969
970 /*
971 * It's not safe for direct compaction to call lock_page.
972 * For example, during page readahead pages are added locked
973 * to the LRU. Later, when the IO completes the pages are
974 * marked uptodate and unlocked. However, the queueing
975 * could be merging multiple pages for one bio (e.g.
976 * mpage_readahead). If an allocation happens for the
977 * second or third page, the process can end up locking
978 * the same page twice and deadlocking. Rather than
979 * trying to be clever about what pages can be locked,
980 * avoid the use of lock_page for direct compaction
981 * altogether.
982 */
983 if (current->flags & PF_MEMALLOC)
984 goto out;
985
986 lock_page(page);
987 }
988
989 if (PageWriteback(page)) {
990 /*
991 * Only in the case of a full synchronous migration is it
992 * necessary to wait for PageWriteback. In the async case,
993 * the retry loop is too short and in the sync-light case,
994 * the overhead of stalling is too much
995 */
996 switch (mode) {
997 case MIGRATE_SYNC:
998 case MIGRATE_SYNC_NO_COPY:
999 break;
1000 default:
1001 rc = -EBUSY;
1002 goto out_unlock;
1003 }
1004 if (!force)
1005 goto out_unlock;
1006 wait_on_page_writeback(page);
1007 }
1008
1009 /*
1010 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1011 * we cannot notice that anon_vma is freed while we migrates a page.
1012 * This get_anon_vma() delays freeing anon_vma pointer until the end
1013 * of migration. File cache pages are no problem because of page_lock()
1014 * File Caches may use write_page() or lock_page() in migration, then,
1015 * just care Anon page here.
1016 *
1017 * Only page_get_anon_vma() understands the subtleties of
1018 * getting a hold on an anon_vma from outside one of its mms.
1019 * But if we cannot get anon_vma, then we won't need it anyway,
1020 * because that implies that the anon page is no longer mapped
1021 * (and cannot be remapped so long as we hold the page lock).
1022 */
1023 if (PageAnon(page) && !PageKsm(page))
1024 anon_vma = page_get_anon_vma(page);
1025
1026 /*
1027 * Block others from accessing the new page when we get around to
1028 * establishing additional references. We are usually the only one
1029 * holding a reference to newpage at this point. We used to have a BUG
1030 * here if trylock_page(newpage) fails, but would like to allow for
1031 * cases where there might be a race with the previous use of newpage.
1032 * This is much like races on refcount of oldpage: just don't BUG().
1033 */
1034 if (unlikely(!trylock_page(newpage)))
1035 goto out_unlock;
1036
1037 if (unlikely(!is_lru)) {
1038 rc = move_to_new_page(newpage, page, mode);
1039 goto out_unlock_both;
1040 }
1041
1042 /*
1043 * Corner case handling:
1044 * 1. When a new swap-cache page is read into, it is added to the LRU
1045 * and treated as swapcache but it has no rmap yet.
1046 * Calling try_to_unmap() against a page->mapping==NULL page will
1047 * trigger a BUG. So handle it here.
1048 * 2. An orphaned page (see truncate_cleanup_page) might have
1049 * fs-private metadata. The page can be picked up due to memory
1050 * offlining. Everywhere else except page reclaim, the page is
1051 * invisible to the vm, so the page can not be migrated. So try to
1052 * free the metadata, so the page can be freed.
1053 */
1054 if (!page->mapping) {
1055 VM_BUG_ON_PAGE(PageAnon(page), page);
1056 if (page_has_private(page)) {
1057 try_to_free_buffers(page);
1058 goto out_unlock_both;
1059 }
1060 } else if (page_mapped(page)) {
1061 /* Establish migration ptes */
1062 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1063 page);
1064 try_to_migrate(page, 0);
1065 page_was_mapped = 1;
1066 }
1067
1068 if (!page_mapped(page))
1069 rc = move_to_new_page(newpage, page, mode);
1070
1071 if (page_was_mapped)
1072 remove_migration_ptes(page,
1073 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1074
1075out_unlock_both:
1076 unlock_page(newpage);
1077out_unlock:
1078 /* Drop an anon_vma reference if we took one */
1079 if (anon_vma)
1080 put_anon_vma(anon_vma);
1081 unlock_page(page);
1082out:
1083 /*
1084 * If migration is successful, decrease refcount of the newpage
1085 * which will not free the page because new page owner increased
1086 * refcounter. As well, if it is LRU page, add the page to LRU
1087 * list in here. Use the old state of the isolated source page to
1088 * determine if we migrated a LRU page. newpage was already unlocked
1089 * and possibly modified by its owner - don't rely on the page
1090 * state.
1091 */
1092 if (rc == MIGRATEPAGE_SUCCESS) {
1093 if (unlikely(!is_lru))
1094 put_page(newpage);
1095 else
1096 putback_lru_page(newpage);
1097 }
1098
1099 return rc;
1100}
1101
1102/*
1103 * Obtain the lock on page, remove all ptes and migrate the page
1104 * to the newly allocated page in newpage.
1105 */
1106static int unmap_and_move(new_page_t get_new_page,
1107 free_page_t put_new_page,
1108 unsigned long private, struct page *page,
1109 int force, enum migrate_mode mode,
1110 enum migrate_reason reason,
1111 struct list_head *ret)
1112{
1113 int rc = MIGRATEPAGE_SUCCESS;
1114 struct page *newpage = NULL;
1115
1116 if (!thp_migration_supported() && PageTransHuge(page))
1117 return -ENOSYS;
1118
1119 if (page_count(page) == 1) {
1120 /* page was freed from under us. So we are done. */
1121 ClearPageActive(page);
1122 ClearPageUnevictable(page);
1123 if (unlikely(__PageMovable(page))) {
1124 lock_page(page);
1125 if (!PageMovable(page))
1126 __ClearPageIsolated(page);
1127 unlock_page(page);
1128 }
1129 goto out;
1130 }
1131
1132 newpage = get_new_page(page, private);
1133 if (!newpage)
1134 return -ENOMEM;
1135
1136 rc = __unmap_and_move(page, newpage, force, mode);
1137 if (rc == MIGRATEPAGE_SUCCESS)
1138 set_page_owner_migrate_reason(newpage, reason);
1139
1140out:
1141 if (rc != -EAGAIN) {
1142 /*
1143 * A page that has been migrated has all references
1144 * removed and will be freed. A page that has not been
1145 * migrated will have kept its references and be restored.
1146 */
1147 list_del(&page->lru);
1148 }
1149
1150 /*
1151 * If migration is successful, releases reference grabbed during
1152 * isolation. Otherwise, restore the page to right list unless
1153 * we want to retry.
1154 */
1155 if (rc == MIGRATEPAGE_SUCCESS) {
1156 /*
1157 * Compaction can migrate also non-LRU pages which are
1158 * not accounted to NR_ISOLATED_*. They can be recognized
1159 * as __PageMovable
1160 */
1161 if (likely(!__PageMovable(page)))
1162 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1163 page_is_file_lru(page), -thp_nr_pages(page));
1164
1165 if (reason != MR_MEMORY_FAILURE)
1166 /*
1167 * We release the page in page_handle_poison.
1168 */
1169 put_page(page);
1170 } else {
1171 if (rc != -EAGAIN)
1172 list_add_tail(&page->lru, ret);
1173
1174 if (put_new_page)
1175 put_new_page(newpage, private);
1176 else
1177 put_page(newpage);
1178 }
1179
1180 return rc;
1181}
1182
1183/*
1184 * Counterpart of unmap_and_move_page() for hugepage migration.
1185 *
1186 * This function doesn't wait the completion of hugepage I/O
1187 * because there is no race between I/O and migration for hugepage.
1188 * Note that currently hugepage I/O occurs only in direct I/O
1189 * where no lock is held and PG_writeback is irrelevant,
1190 * and writeback status of all subpages are counted in the reference
1191 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1192 * under direct I/O, the reference of the head page is 512 and a bit more.)
1193 * This means that when we try to migrate hugepage whose subpages are
1194 * doing direct I/O, some references remain after try_to_unmap() and
1195 * hugepage migration fails without data corruption.
1196 *
1197 * There is also no race when direct I/O is issued on the page under migration,
1198 * because then pte is replaced with migration swap entry and direct I/O code
1199 * will wait in the page fault for migration to complete.
1200 */
1201static int unmap_and_move_huge_page(new_page_t get_new_page,
1202 free_page_t put_new_page, unsigned long private,
1203 struct page *hpage, int force,
1204 enum migrate_mode mode, int reason,
1205 struct list_head *ret)
1206{
1207 int rc = -EAGAIN;
1208 int page_was_mapped = 0;
1209 struct page *new_hpage;
1210 struct anon_vma *anon_vma = NULL;
1211 struct address_space *mapping = NULL;
1212
1213 /*
1214 * Migratability of hugepages depends on architectures and their size.
1215 * This check is necessary because some callers of hugepage migration
1216 * like soft offline and memory hotremove don't walk through page
1217 * tables or check whether the hugepage is pmd-based or not before
1218 * kicking migration.
1219 */
1220 if (!hugepage_migration_supported(page_hstate(hpage))) {
1221 list_move_tail(&hpage->lru, ret);
1222 return -ENOSYS;
1223 }
1224
1225 if (page_count(hpage) == 1) {
1226 /* page was freed from under us. So we are done. */
1227 putback_active_hugepage(hpage);
1228 return MIGRATEPAGE_SUCCESS;
1229 }
1230
1231 new_hpage = get_new_page(hpage, private);
1232 if (!new_hpage)
1233 return -ENOMEM;
1234
1235 if (!trylock_page(hpage)) {
1236 if (!force)
1237 goto out;
1238 switch (mode) {
1239 case MIGRATE_SYNC:
1240 case MIGRATE_SYNC_NO_COPY:
1241 break;
1242 default:
1243 goto out;
1244 }
1245 lock_page(hpage);
1246 }
1247
1248 /*
1249 * Check for pages which are in the process of being freed. Without
1250 * page_mapping() set, hugetlbfs specific move page routine will not
1251 * be called and we could leak usage counts for subpools.
1252 */
1253 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1254 rc = -EBUSY;
1255 goto out_unlock;
1256 }
1257
1258 if (PageAnon(hpage))
1259 anon_vma = page_get_anon_vma(hpage);
1260
1261 if (unlikely(!trylock_page(new_hpage)))
1262 goto put_anon;
1263
1264 if (page_mapped(hpage)) {
1265 bool mapping_locked = false;
1266 enum ttu_flags ttu = 0;
1267
1268 if (!PageAnon(hpage)) {
1269 /*
1270 * In shared mappings, try_to_unmap could potentially
1271 * call huge_pmd_unshare. Because of this, take
1272 * semaphore in write mode here and set TTU_RMAP_LOCKED
1273 * to let lower levels know we have taken the lock.
1274 */
1275 mapping = hugetlb_page_mapping_lock_write(hpage);
1276 if (unlikely(!mapping))
1277 goto unlock_put_anon;
1278
1279 mapping_locked = true;
1280 ttu |= TTU_RMAP_LOCKED;
1281 }
1282
1283 try_to_migrate(hpage, ttu);
1284 page_was_mapped = 1;
1285
1286 if (mapping_locked)
1287 i_mmap_unlock_write(mapping);
1288 }
1289
1290 if (!page_mapped(hpage))
1291 rc = move_to_new_page(new_hpage, hpage, mode);
1292
1293 if (page_was_mapped)
1294 remove_migration_ptes(hpage,
1295 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1296
1297unlock_put_anon:
1298 unlock_page(new_hpage);
1299
1300put_anon:
1301 if (anon_vma)
1302 put_anon_vma(anon_vma);
1303
1304 if (rc == MIGRATEPAGE_SUCCESS) {
1305 move_hugetlb_state(hpage, new_hpage, reason);
1306 put_new_page = NULL;
1307 }
1308
1309out_unlock:
1310 unlock_page(hpage);
1311out:
1312 if (rc == MIGRATEPAGE_SUCCESS)
1313 putback_active_hugepage(hpage);
1314 else if (rc != -EAGAIN)
1315 list_move_tail(&hpage->lru, ret);
1316
1317 /*
1318 * If migration was not successful and there's a freeing callback, use
1319 * it. Otherwise, put_page() will drop the reference grabbed during
1320 * isolation.
1321 */
1322 if (put_new_page)
1323 put_new_page(new_hpage, private);
1324 else
1325 putback_active_hugepage(new_hpage);
1326
1327 return rc;
1328}
1329
1330static inline int try_split_thp(struct page *page, struct page **page2,
1331 struct list_head *from)
1332{
1333 int rc = 0;
1334
1335 lock_page(page);
1336 rc = split_huge_page_to_list(page, from);
1337 unlock_page(page);
1338 if (!rc)
1339 list_safe_reset_next(page, *page2, lru);
1340
1341 return rc;
1342}
1343
1344/*
1345 * migrate_pages - migrate the pages specified in a list, to the free pages
1346 * supplied as the target for the page migration
1347 *
1348 * @from: The list of pages to be migrated.
1349 * @get_new_page: The function used to allocate free pages to be used
1350 * as the target of the page migration.
1351 * @put_new_page: The function used to free target pages if migration
1352 * fails, or NULL if no special handling is necessary.
1353 * @private: Private data to be passed on to get_new_page()
1354 * @mode: The migration mode that specifies the constraints for
1355 * page migration, if any.
1356 * @reason: The reason for page migration.
1357 *
1358 * The function returns after 10 attempts or if no pages are movable any more
1359 * because the list has become empty or no retryable pages exist any more.
1360 * It is caller's responsibility to call putback_movable_pages() to return pages
1361 * to the LRU or free list only if ret != 0.
1362 *
1363 * Returns the number of pages that were not migrated, or an error code.
1364 */
1365int migrate_pages(struct list_head *from, new_page_t get_new_page,
1366 free_page_t put_new_page, unsigned long private,
1367 enum migrate_mode mode, int reason)
1368{
1369 int retry = 1;
1370 int thp_retry = 1;
1371 int nr_failed = 0;
1372 int nr_succeeded = 0;
1373 int nr_thp_succeeded = 0;
1374 int nr_thp_failed = 0;
1375 int nr_thp_split = 0;
1376 int pass = 0;
1377 bool is_thp = false;
1378 struct page *page;
1379 struct page *page2;
1380 int swapwrite = current->flags & PF_SWAPWRITE;
1381 int rc, nr_subpages;
1382 LIST_HEAD(ret_pages);
1383 bool nosplit = (reason == MR_NUMA_MISPLACED);
1384
1385 trace_mm_migrate_pages_start(mode, reason);
1386
1387 if (!swapwrite)
1388 current->flags |= PF_SWAPWRITE;
1389
1390 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1391 retry = 0;
1392 thp_retry = 0;
1393
1394 list_for_each_entry_safe(page, page2, from, lru) {
1395retry:
1396 /*
1397 * THP statistics is based on the source huge page.
1398 * Capture required information that might get lost
1399 * during migration.
1400 */
1401 is_thp = PageTransHuge(page) && !PageHuge(page);
1402 nr_subpages = thp_nr_pages(page);
1403 cond_resched();
1404
1405 if (PageHuge(page))
1406 rc = unmap_and_move_huge_page(get_new_page,
1407 put_new_page, private, page,
1408 pass > 2, mode, reason,
1409 &ret_pages);
1410 else
1411 rc = unmap_and_move(get_new_page, put_new_page,
1412 private, page, pass > 2, mode,
1413 reason, &ret_pages);
1414 /*
1415 * The rules are:
1416 * Success: non hugetlb page will be freed, hugetlb
1417 * page will be put back
1418 * -EAGAIN: stay on the from list
1419 * -ENOMEM: stay on the from list
1420 * Other errno: put on ret_pages list then splice to
1421 * from list
1422 */
1423 switch(rc) {
1424 /*
1425 * THP migration might be unsupported or the
1426 * allocation could've failed so we should
1427 * retry on the same page with the THP split
1428 * to base pages.
1429 *
1430 * Head page is retried immediately and tail
1431 * pages are added to the tail of the list so
1432 * we encounter them after the rest of the list
1433 * is processed.
1434 */
1435 case -ENOSYS:
1436 /* THP migration is unsupported */
1437 if (is_thp) {
1438 if (!try_split_thp(page, &page2, from)) {
1439 nr_thp_split++;
1440 goto retry;
1441 }
1442
1443 nr_thp_failed++;
1444 nr_failed += nr_subpages;
1445 break;
1446 }
1447
1448 /* Hugetlb migration is unsupported */
1449 nr_failed++;
1450 break;
1451 case -ENOMEM:
1452 /*
1453 * When memory is low, don't bother to try to migrate
1454 * other pages, just exit.
1455 * THP NUMA faulting doesn't split THP to retry.
1456 */
1457 if (is_thp && !nosplit) {
1458 if (!try_split_thp(page, &page2, from)) {
1459 nr_thp_split++;
1460 goto retry;
1461 }
1462
1463 nr_thp_failed++;
1464 nr_failed += nr_subpages;
1465 goto out;
1466 }
1467 nr_failed++;
1468 goto out;
1469 case -EAGAIN:
1470 if (is_thp) {
1471 thp_retry++;
1472 break;
1473 }
1474 retry++;
1475 break;
1476 case MIGRATEPAGE_SUCCESS:
1477 if (is_thp) {
1478 nr_thp_succeeded++;
1479 nr_succeeded += nr_subpages;
1480 break;
1481 }
1482 nr_succeeded++;
1483 break;
1484 default:
1485 /*
1486 * Permanent failure (-EBUSY, etc.):
1487 * unlike -EAGAIN case, the failed page is
1488 * removed from migration page list and not
1489 * retried in the next outer loop.
1490 */
1491 if (is_thp) {
1492 nr_thp_failed++;
1493 nr_failed += nr_subpages;
1494 break;
1495 }
1496 nr_failed++;
1497 break;
1498 }
1499 }
1500 }
1501 nr_failed += retry + thp_retry;
1502 nr_thp_failed += thp_retry;
1503 rc = nr_failed;
1504out:
1505 /*
1506 * Put the permanent failure page back to migration list, they
1507 * will be put back to the right list by the caller.
1508 */
1509 list_splice(&ret_pages, from);
1510
1511 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1512 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1513 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1514 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1515 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1516 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1517 nr_thp_failed, nr_thp_split, mode, reason);
1518
1519 if (!swapwrite)
1520 current->flags &= ~PF_SWAPWRITE;
1521
1522 return rc;
1523}
1524
1525struct page *alloc_migration_target(struct page *page, unsigned long private)
1526{
1527 struct migration_target_control *mtc;
1528 gfp_t gfp_mask;
1529 unsigned int order = 0;
1530 struct page *new_page = NULL;
1531 int nid;
1532 int zidx;
1533
1534 mtc = (struct migration_target_control *)private;
1535 gfp_mask = mtc->gfp_mask;
1536 nid = mtc->nid;
1537 if (nid == NUMA_NO_NODE)
1538 nid = page_to_nid(page);
1539
1540 if (PageHuge(page)) {
1541 struct hstate *h = page_hstate(compound_head(page));
1542
1543 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1544 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1545 }
1546
1547 if (PageTransHuge(page)) {
1548 /*
1549 * clear __GFP_RECLAIM to make the migration callback
1550 * consistent with regular THP allocations.
1551 */
1552 gfp_mask &= ~__GFP_RECLAIM;
1553 gfp_mask |= GFP_TRANSHUGE;
1554 order = HPAGE_PMD_ORDER;
1555 }
1556 zidx = zone_idx(page_zone(page));
1557 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1558 gfp_mask |= __GFP_HIGHMEM;
1559
1560 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1561
1562 if (new_page && PageTransHuge(new_page))
1563 prep_transhuge_page(new_page);
1564
1565 return new_page;
1566}
1567
1568#ifdef CONFIG_NUMA
1569
1570static int store_status(int __user *status, int start, int value, int nr)
1571{
1572 while (nr-- > 0) {
1573 if (put_user(value, status + start))
1574 return -EFAULT;
1575 start++;
1576 }
1577
1578 return 0;
1579}
1580
1581static int do_move_pages_to_node(struct mm_struct *mm,
1582 struct list_head *pagelist, int node)
1583{
1584 int err;
1585 struct migration_target_control mtc = {
1586 .nid = node,
1587 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1588 };
1589
1590 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1591 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1592 if (err)
1593 putback_movable_pages(pagelist);
1594 return err;
1595}
1596
1597/*
1598 * Resolves the given address to a struct page, isolates it from the LRU and
1599 * puts it to the given pagelist.
1600 * Returns:
1601 * errno - if the page cannot be found/isolated
1602 * 0 - when it doesn't have to be migrated because it is already on the
1603 * target node
1604 * 1 - when it has been queued
1605 */
1606static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1607 int node, struct list_head *pagelist, bool migrate_all)
1608{
1609 struct vm_area_struct *vma;
1610 struct page *page;
1611 unsigned int follflags;
1612 int err;
1613
1614 mmap_read_lock(mm);
1615 err = -EFAULT;
1616 vma = find_vma(mm, addr);
1617 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1618 goto out;
1619
1620 /* FOLL_DUMP to ignore special (like zero) pages */
1621 follflags = FOLL_GET | FOLL_DUMP;
1622 page = follow_page(vma, addr, follflags);
1623
1624 err = PTR_ERR(page);
1625 if (IS_ERR(page))
1626 goto out;
1627
1628 err = -ENOENT;
1629 if (!page)
1630 goto out;
1631
1632 err = 0;
1633 if (page_to_nid(page) == node)
1634 goto out_putpage;
1635
1636 err = -EACCES;
1637 if (page_mapcount(page) > 1 && !migrate_all)
1638 goto out_putpage;
1639
1640 if (PageHuge(page)) {
1641 if (PageHead(page)) {
1642 isolate_huge_page(page, pagelist);
1643 err = 1;
1644 }
1645 } else {
1646 struct page *head;
1647
1648 head = compound_head(page);
1649 err = isolate_lru_page(head);
1650 if (err)
1651 goto out_putpage;
1652
1653 err = 1;
1654 list_add_tail(&head->lru, pagelist);
1655 mod_node_page_state(page_pgdat(head),
1656 NR_ISOLATED_ANON + page_is_file_lru(head),
1657 thp_nr_pages(head));
1658 }
1659out_putpage:
1660 /*
1661 * Either remove the duplicate refcount from
1662 * isolate_lru_page() or drop the page ref if it was
1663 * not isolated.
1664 */
1665 put_page(page);
1666out:
1667 mmap_read_unlock(mm);
1668 return err;
1669}
1670
1671static int move_pages_and_store_status(struct mm_struct *mm, int node,
1672 struct list_head *pagelist, int __user *status,
1673 int start, int i, unsigned long nr_pages)
1674{
1675 int err;
1676
1677 if (list_empty(pagelist))
1678 return 0;
1679
1680 err = do_move_pages_to_node(mm, pagelist, node);
1681 if (err) {
1682 /*
1683 * Positive err means the number of failed
1684 * pages to migrate. Since we are going to
1685 * abort and return the number of non-migrated
1686 * pages, so need to include the rest of the
1687 * nr_pages that have not been attempted as
1688 * well.
1689 */
1690 if (err > 0)
1691 err += nr_pages - i - 1;
1692 return err;
1693 }
1694 return store_status(status, start, node, i - start);
1695}
1696
1697/*
1698 * Migrate an array of page address onto an array of nodes and fill
1699 * the corresponding array of status.
1700 */
1701static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1702 unsigned long nr_pages,
1703 const void __user * __user *pages,
1704 const int __user *nodes,
1705 int __user *status, int flags)
1706{
1707 int current_node = NUMA_NO_NODE;
1708 LIST_HEAD(pagelist);
1709 int start, i;
1710 int err = 0, err1;
1711
1712 lru_cache_disable();
1713
1714 for (i = start = 0; i < nr_pages; i++) {
1715 const void __user *p;
1716 unsigned long addr;
1717 int node;
1718
1719 err = -EFAULT;
1720 if (get_user(p, pages + i))
1721 goto out_flush;
1722 if (get_user(node, nodes + i))
1723 goto out_flush;
1724 addr = (unsigned long)untagged_addr(p);
1725
1726 err = -ENODEV;
1727 if (node < 0 || node >= MAX_NUMNODES)
1728 goto out_flush;
1729 if (!node_state(node, N_MEMORY))
1730 goto out_flush;
1731
1732 err = -EACCES;
1733 if (!node_isset(node, task_nodes))
1734 goto out_flush;
1735
1736 if (current_node == NUMA_NO_NODE) {
1737 current_node = node;
1738 start = i;
1739 } else if (node != current_node) {
1740 err = move_pages_and_store_status(mm, current_node,
1741 &pagelist, status, start, i, nr_pages);
1742 if (err)
1743 goto out;
1744 start = i;
1745 current_node = node;
1746 }
1747
1748 /*
1749 * Errors in the page lookup or isolation are not fatal and we simply
1750 * report them via status
1751 */
1752 err = add_page_for_migration(mm, addr, current_node,
1753 &pagelist, flags & MPOL_MF_MOVE_ALL);
1754
1755 if (err > 0) {
1756 /* The page is successfully queued for migration */
1757 continue;
1758 }
1759
1760 /*
1761 * If the page is already on the target node (!err), store the
1762 * node, otherwise, store the err.
1763 */
1764 err = store_status(status, i, err ? : current_node, 1);
1765 if (err)
1766 goto out_flush;
1767
1768 err = move_pages_and_store_status(mm, current_node, &pagelist,
1769 status, start, i, nr_pages);
1770 if (err)
1771 goto out;
1772 current_node = NUMA_NO_NODE;
1773 }
1774out_flush:
1775 /* Make sure we do not overwrite the existing error */
1776 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1777 status, start, i, nr_pages);
1778 if (err >= 0)
1779 err = err1;
1780out:
1781 lru_cache_enable();
1782 return err;
1783}
1784
1785/*
1786 * Determine the nodes of an array of pages and store it in an array of status.
1787 */
1788static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1789 const void __user **pages, int *status)
1790{
1791 unsigned long i;
1792
1793 mmap_read_lock(mm);
1794
1795 for (i = 0; i < nr_pages; i++) {
1796 unsigned long addr = (unsigned long)(*pages);
1797 struct vm_area_struct *vma;
1798 struct page *page;
1799 int err = -EFAULT;
1800
1801 vma = vma_lookup(mm, addr);
1802 if (!vma)
1803 goto set_status;
1804
1805 /* FOLL_DUMP to ignore special (like zero) pages */
1806 page = follow_page(vma, addr, FOLL_DUMP);
1807
1808 err = PTR_ERR(page);
1809 if (IS_ERR(page))
1810 goto set_status;
1811
1812 err = page ? page_to_nid(page) : -ENOENT;
1813set_status:
1814 *status = err;
1815
1816 pages++;
1817 status++;
1818 }
1819
1820 mmap_read_unlock(mm);
1821}
1822
1823/*
1824 * Determine the nodes of a user array of pages and store it in
1825 * a user array of status.
1826 */
1827static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1828 const void __user * __user *pages,
1829 int __user *status)
1830{
1831#define DO_PAGES_STAT_CHUNK_NR 16
1832 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1833 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1834
1835 while (nr_pages) {
1836 unsigned long chunk_nr;
1837
1838 chunk_nr = nr_pages;
1839 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1840 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1841
1842 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1843 break;
1844
1845 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1846
1847 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1848 break;
1849
1850 pages += chunk_nr;
1851 status += chunk_nr;
1852 nr_pages -= chunk_nr;
1853 }
1854 return nr_pages ? -EFAULT : 0;
1855}
1856
1857static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1858{
1859 struct task_struct *task;
1860 struct mm_struct *mm;
1861
1862 /*
1863 * There is no need to check if current process has the right to modify
1864 * the specified process when they are same.
1865 */
1866 if (!pid) {
1867 mmget(current->mm);
1868 *mem_nodes = cpuset_mems_allowed(current);
1869 return current->mm;
1870 }
1871
1872 /* Find the mm_struct */
1873 rcu_read_lock();
1874 task = find_task_by_vpid(pid);
1875 if (!task) {
1876 rcu_read_unlock();
1877 return ERR_PTR(-ESRCH);
1878 }
1879 get_task_struct(task);
1880
1881 /*
1882 * Check if this process has the right to modify the specified
1883 * process. Use the regular "ptrace_may_access()" checks.
1884 */
1885 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1886 rcu_read_unlock();
1887 mm = ERR_PTR(-EPERM);
1888 goto out;
1889 }
1890 rcu_read_unlock();
1891
1892 mm = ERR_PTR(security_task_movememory(task));
1893 if (IS_ERR(mm))
1894 goto out;
1895 *mem_nodes = cpuset_mems_allowed(task);
1896 mm = get_task_mm(task);
1897out:
1898 put_task_struct(task);
1899 if (!mm)
1900 mm = ERR_PTR(-EINVAL);
1901 return mm;
1902}
1903
1904/*
1905 * Move a list of pages in the address space of the currently executing
1906 * process.
1907 */
1908static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1909 const void __user * __user *pages,
1910 const int __user *nodes,
1911 int __user *status, int flags)
1912{
1913 struct mm_struct *mm;
1914 int err;
1915 nodemask_t task_nodes;
1916
1917 /* Check flags */
1918 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1919 return -EINVAL;
1920
1921 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1922 return -EPERM;
1923
1924 mm = find_mm_struct(pid, &task_nodes);
1925 if (IS_ERR(mm))
1926 return PTR_ERR(mm);
1927
1928 if (nodes)
1929 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1930 nodes, status, flags);
1931 else
1932 err = do_pages_stat(mm, nr_pages, pages, status);
1933
1934 mmput(mm);
1935 return err;
1936}
1937
1938SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1939 const void __user * __user *, pages,
1940 const int __user *, nodes,
1941 int __user *, status, int, flags)
1942{
1943 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1944}
1945
1946#ifdef CONFIG_COMPAT
1947COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1948 compat_uptr_t __user *, pages32,
1949 const int __user *, nodes,
1950 int __user *, status,
1951 int, flags)
1952{
1953 const void __user * __user *pages;
1954 int i;
1955
1956 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1957 for (i = 0; i < nr_pages; i++) {
1958 compat_uptr_t p;
1959
1960 if (get_user(p, pages32 + i) ||
1961 put_user(compat_ptr(p), pages + i))
1962 return -EFAULT;
1963 }
1964 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1965}
1966#endif /* CONFIG_COMPAT */
1967
1968#ifdef CONFIG_NUMA_BALANCING
1969/*
1970 * Returns true if this is a safe migration target node for misplaced NUMA
1971 * pages. Currently it only checks the watermarks which crude
1972 */
1973static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1974 unsigned long nr_migrate_pages)
1975{
1976 int z;
1977
1978 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1979 struct zone *zone = pgdat->node_zones + z;
1980
1981 if (!populated_zone(zone))
1982 continue;
1983
1984 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1985 if (!zone_watermark_ok(zone, 0,
1986 high_wmark_pages(zone) +
1987 nr_migrate_pages,
1988 ZONE_MOVABLE, 0))
1989 continue;
1990 return true;
1991 }
1992 return false;
1993}
1994
1995static struct page *alloc_misplaced_dst_page(struct page *page,
1996 unsigned long data)
1997{
1998 int nid = (int) data;
1999 struct page *newpage;
2000
2001 newpage = __alloc_pages_node(nid,
2002 (GFP_HIGHUSER_MOVABLE |
2003 __GFP_THISNODE | __GFP_NOMEMALLOC |
2004 __GFP_NORETRY | __GFP_NOWARN) &
2005 ~__GFP_RECLAIM, 0);
2006
2007 return newpage;
2008}
2009
2010static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2011 unsigned long data)
2012{
2013 int nid = (int) data;
2014 struct page *newpage;
2015
2016 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2017 HPAGE_PMD_ORDER);
2018 if (!newpage)
2019 goto out;
2020
2021 prep_transhuge_page(newpage);
2022
2023out:
2024 return newpage;
2025}
2026
2027static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2028{
2029 int page_lru;
2030
2031 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2032
2033 /* Do not migrate THP mapped by multiple processes */
2034 if (PageTransHuge(page) && total_mapcount(page) > 1)
2035 return 0;
2036
2037 /* Avoid migrating to a node that is nearly full */
2038 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2039 return 0;
2040
2041 if (isolate_lru_page(page))
2042 return 0;
2043
2044 page_lru = page_is_file_lru(page);
2045 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2046 thp_nr_pages(page));
2047
2048 /*
2049 * Isolating the page has taken another reference, so the
2050 * caller's reference can be safely dropped without the page
2051 * disappearing underneath us during migration.
2052 */
2053 put_page(page);
2054 return 1;
2055}
2056
2057/*
2058 * Attempt to migrate a misplaced page to the specified destination
2059 * node. Caller is expected to have an elevated reference count on
2060 * the page that will be dropped by this function before returning.
2061 */
2062int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2063 int node)
2064{
2065 pg_data_t *pgdat = NODE_DATA(node);
2066 int isolated;
2067 int nr_remaining;
2068 LIST_HEAD(migratepages);
2069 new_page_t *new;
2070 bool compound;
2071 int nr_pages = thp_nr_pages(page);
2072
2073 /*
2074 * PTE mapped THP or HugeTLB page can't reach here so the page could
2075 * be either base page or THP. And it must be head page if it is
2076 * THP.
2077 */
2078 compound = PageTransHuge(page);
2079
2080 if (compound)
2081 new = alloc_misplaced_dst_page_thp;
2082 else
2083 new = alloc_misplaced_dst_page;
2084
2085 /*
2086 * Don't migrate file pages that are mapped in multiple processes
2087 * with execute permissions as they are probably shared libraries.
2088 */
2089 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2090 (vma->vm_flags & VM_EXEC))
2091 goto out;
2092
2093 /*
2094 * Also do not migrate dirty pages as not all filesystems can move
2095 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2096 */
2097 if (page_is_file_lru(page) && PageDirty(page))
2098 goto out;
2099
2100 isolated = numamigrate_isolate_page(pgdat, page);
2101 if (!isolated)
2102 goto out;
2103
2104 list_add(&page->lru, &migratepages);
2105 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2106 MIGRATE_ASYNC, MR_NUMA_MISPLACED);
2107 if (nr_remaining) {
2108 if (!list_empty(&migratepages)) {
2109 list_del(&page->lru);
2110 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2111 page_is_file_lru(page), -nr_pages);
2112 putback_lru_page(page);
2113 }
2114 isolated = 0;
2115 } else
2116 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages);
2117 BUG_ON(!list_empty(&migratepages));
2118 return isolated;
2119
2120out:
2121 put_page(page);
2122 return 0;
2123}
2124#endif /* CONFIG_NUMA_BALANCING */
2125#endif /* CONFIG_NUMA */
2126
2127#ifdef CONFIG_DEVICE_PRIVATE
2128static int migrate_vma_collect_skip(unsigned long start,
2129 unsigned long end,
2130 struct mm_walk *walk)
2131{
2132 struct migrate_vma *migrate = walk->private;
2133 unsigned long addr;
2134
2135 for (addr = start; addr < end; addr += PAGE_SIZE) {
2136 migrate->dst[migrate->npages] = 0;
2137 migrate->src[migrate->npages++] = 0;
2138 }
2139
2140 return 0;
2141}
2142
2143static int migrate_vma_collect_hole(unsigned long start,
2144 unsigned long end,
2145 __always_unused int depth,
2146 struct mm_walk *walk)
2147{
2148 struct migrate_vma *migrate = walk->private;
2149 unsigned long addr;
2150
2151 /* Only allow populating anonymous memory. */
2152 if (!vma_is_anonymous(walk->vma))
2153 return migrate_vma_collect_skip(start, end, walk);
2154
2155 for (addr = start; addr < end; addr += PAGE_SIZE) {
2156 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2157 migrate->dst[migrate->npages] = 0;
2158 migrate->npages++;
2159 migrate->cpages++;
2160 }
2161
2162 return 0;
2163}
2164
2165static int migrate_vma_collect_pmd(pmd_t *pmdp,
2166 unsigned long start,
2167 unsigned long end,
2168 struct mm_walk *walk)
2169{
2170 struct migrate_vma *migrate = walk->private;
2171 struct vm_area_struct *vma = walk->vma;
2172 struct mm_struct *mm = vma->vm_mm;
2173 unsigned long addr = start, unmapped = 0;
2174 spinlock_t *ptl;
2175 pte_t *ptep;
2176
2177again:
2178 if (pmd_none(*pmdp))
2179 return migrate_vma_collect_hole(start, end, -1, walk);
2180
2181 if (pmd_trans_huge(*pmdp)) {
2182 struct page *page;
2183
2184 ptl = pmd_lock(mm, pmdp);
2185 if (unlikely(!pmd_trans_huge(*pmdp))) {
2186 spin_unlock(ptl);
2187 goto again;
2188 }
2189
2190 page = pmd_page(*pmdp);
2191 if (is_huge_zero_page(page)) {
2192 spin_unlock(ptl);
2193 split_huge_pmd(vma, pmdp, addr);
2194 if (pmd_trans_unstable(pmdp))
2195 return migrate_vma_collect_skip(start, end,
2196 walk);
2197 } else {
2198 int ret;
2199
2200 get_page(page);
2201 spin_unlock(ptl);
2202 if (unlikely(!trylock_page(page)))
2203 return migrate_vma_collect_skip(start, end,
2204 walk);
2205 ret = split_huge_page(page);
2206 unlock_page(page);
2207 put_page(page);
2208 if (ret)
2209 return migrate_vma_collect_skip(start, end,
2210 walk);
2211 if (pmd_none(*pmdp))
2212 return migrate_vma_collect_hole(start, end, -1,
2213 walk);
2214 }
2215 }
2216
2217 if (unlikely(pmd_bad(*pmdp)))
2218 return migrate_vma_collect_skip(start, end, walk);
2219
2220 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2221 arch_enter_lazy_mmu_mode();
2222
2223 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2224 unsigned long mpfn = 0, pfn;
2225 struct page *page;
2226 swp_entry_t entry;
2227 pte_t pte;
2228
2229 pte = *ptep;
2230
2231 if (pte_none(pte)) {
2232 if (vma_is_anonymous(vma)) {
2233 mpfn = MIGRATE_PFN_MIGRATE;
2234 migrate->cpages++;
2235 }
2236 goto next;
2237 }
2238
2239 if (!pte_present(pte)) {
2240 /*
2241 * Only care about unaddressable device page special
2242 * page table entry. Other special swap entries are not
2243 * migratable, and we ignore regular swapped page.
2244 */
2245 entry = pte_to_swp_entry(pte);
2246 if (!is_device_private_entry(entry))
2247 goto next;
2248
2249 page = pfn_swap_entry_to_page(entry);
2250 if (!(migrate->flags &
2251 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2252 page->pgmap->owner != migrate->pgmap_owner)
2253 goto next;
2254
2255 mpfn = migrate_pfn(page_to_pfn(page)) |
2256 MIGRATE_PFN_MIGRATE;
2257 if (is_writable_device_private_entry(entry))
2258 mpfn |= MIGRATE_PFN_WRITE;
2259 } else {
2260 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2261 goto next;
2262 pfn = pte_pfn(pte);
2263 if (is_zero_pfn(pfn)) {
2264 mpfn = MIGRATE_PFN_MIGRATE;
2265 migrate->cpages++;
2266 goto next;
2267 }
2268 page = vm_normal_page(migrate->vma, addr, pte);
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 = 0;
2276 goto next;
2277 }
2278
2279 /*
2280 * By getting a reference on the page we pin it and that blocks
2281 * any kind of migration. Side effect is that it "freezes" the
2282 * pte.
2283 *
2284 * We drop this reference after isolating the page from the lru
2285 * for non device page (device page are not on the lru and thus
2286 * can't be dropped from it).
2287 */
2288 get_page(page);
2289 migrate->cpages++;
2290
2291 /*
2292 * Optimize for the common case where page is only mapped once
2293 * in one process. If we can lock the page, then we can safely
2294 * set up a special migration page table entry now.
2295 */
2296 if (trylock_page(page)) {
2297 pte_t swp_pte;
2298
2299 mpfn |= MIGRATE_PFN_LOCKED;
2300 ptep_get_and_clear(mm, addr, ptep);
2301
2302 /* Setup special migration page table entry */
2303 if (mpfn & MIGRATE_PFN_WRITE)
2304 entry = make_writable_migration_entry(
2305 page_to_pfn(page));
2306 else
2307 entry = make_readable_migration_entry(
2308 page_to_pfn(page));
2309 swp_pte = swp_entry_to_pte(entry);
2310 if (pte_present(pte)) {
2311 if (pte_soft_dirty(pte))
2312 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2313 if (pte_uffd_wp(pte))
2314 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2315 } else {
2316 if (pte_swp_soft_dirty(pte))
2317 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2318 if (pte_swp_uffd_wp(pte))
2319 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2320 }
2321 set_pte_at(mm, addr, ptep, swp_pte);
2322
2323 /*
2324 * This is like regular unmap: we remove the rmap and
2325 * drop page refcount. Page won't be freed, as we took
2326 * a reference just above.
2327 */
2328 page_remove_rmap(page, false);
2329 put_page(page);
2330
2331 if (pte_present(pte))
2332 unmapped++;
2333 }
2334
2335next:
2336 migrate->dst[migrate->npages] = 0;
2337 migrate->src[migrate->npages++] = mpfn;
2338 }
2339 arch_leave_lazy_mmu_mode();
2340 pte_unmap_unlock(ptep - 1, ptl);
2341
2342 /* Only flush the TLB if we actually modified any entries */
2343 if (unmapped)
2344 flush_tlb_range(walk->vma, start, end);
2345
2346 return 0;
2347}
2348
2349static const struct mm_walk_ops migrate_vma_walk_ops = {
2350 .pmd_entry = migrate_vma_collect_pmd,
2351 .pte_hole = migrate_vma_collect_hole,
2352};
2353
2354/*
2355 * migrate_vma_collect() - collect pages over a range of virtual addresses
2356 * @migrate: migrate struct containing all migration information
2357 *
2358 * This will walk the CPU page table. For each virtual address backed by a
2359 * valid page, it updates the src array and takes a reference on the page, in
2360 * order to pin the page until we lock it and unmap it.
2361 */
2362static void migrate_vma_collect(struct migrate_vma *migrate)
2363{
2364 struct mmu_notifier_range range;
2365
2366 /*
2367 * Note that the pgmap_owner is passed to the mmu notifier callback so
2368 * that the registered device driver can skip invalidating device
2369 * private page mappings that won't be migrated.
2370 */
2371 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
2372 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
2373 migrate->pgmap_owner);
2374 mmu_notifier_invalidate_range_start(&range);
2375
2376 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2377 &migrate_vma_walk_ops, migrate);
2378
2379 mmu_notifier_invalidate_range_end(&range);
2380 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2381}
2382
2383/*
2384 * migrate_vma_check_page() - check if page is pinned or not
2385 * @page: struct page to check
2386 *
2387 * Pinned pages cannot be migrated. This is the same test as in
2388 * migrate_page_move_mapping(), except that here we allow migration of a
2389 * ZONE_DEVICE page.
2390 */
2391static bool migrate_vma_check_page(struct page *page)
2392{
2393 /*
2394 * One extra ref because caller holds an extra reference, either from
2395 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2396 * a device page.
2397 */
2398 int extra = 1;
2399
2400 /*
2401 * FIXME support THP (transparent huge page), it is bit more complex to
2402 * check them than regular pages, because they can be mapped with a pmd
2403 * or with a pte (split pte mapping).
2404 */
2405 if (PageCompound(page))
2406 return false;
2407
2408 /* Page from ZONE_DEVICE have one extra reference */
2409 if (is_zone_device_page(page)) {
2410 /*
2411 * Private page can never be pin as they have no valid pte and
2412 * GUP will fail for those. Yet if there is a pending migration
2413 * a thread might try to wait on the pte migration entry and
2414 * will bump the page reference count. Sadly there is no way to
2415 * differentiate a regular pin from migration wait. Hence to
2416 * avoid 2 racing thread trying to migrate back to CPU to enter
2417 * infinite loop (one stopping migration because the other is
2418 * waiting on pte migration entry). We always return true here.
2419 *
2420 * FIXME proper solution is to rework migration_entry_wait() so
2421 * it does not need to take a reference on page.
2422 */
2423 return is_device_private_page(page);
2424 }
2425
2426 /* For file back page */
2427 if (page_mapping(page))
2428 extra += 1 + page_has_private(page);
2429
2430 if ((page_count(page) - extra) > page_mapcount(page))
2431 return false;
2432
2433 return true;
2434}
2435
2436/*
2437 * migrate_vma_prepare() - lock pages and isolate them from the lru
2438 * @migrate: migrate struct containing all migration information
2439 *
2440 * This locks pages that have been collected by migrate_vma_collect(). Once each
2441 * page is locked it is isolated from the lru (for non-device pages). Finally,
2442 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2443 * migrated by concurrent kernel threads.
2444 */
2445static void migrate_vma_prepare(struct migrate_vma *migrate)
2446{
2447 const unsigned long npages = migrate->npages;
2448 const unsigned long start = migrate->start;
2449 unsigned long addr, i, restore = 0;
2450 bool allow_drain = true;
2451
2452 lru_add_drain();
2453
2454 for (i = 0; (i < npages) && migrate->cpages; i++) {
2455 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2456 bool remap = true;
2457
2458 if (!page)
2459 continue;
2460
2461 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2462 /*
2463 * Because we are migrating several pages there can be
2464 * a deadlock between 2 concurrent migration where each
2465 * are waiting on each other page lock.
2466 *
2467 * Make migrate_vma() a best effort thing and backoff
2468 * for any page we can not lock right away.
2469 */
2470 if (!trylock_page(page)) {
2471 migrate->src[i] = 0;
2472 migrate->cpages--;
2473 put_page(page);
2474 continue;
2475 }
2476 remap = false;
2477 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2478 }
2479
2480 /* ZONE_DEVICE pages are not on LRU */
2481 if (!is_zone_device_page(page)) {
2482 if (!PageLRU(page) && allow_drain) {
2483 /* Drain CPU's pagevec */
2484 lru_add_drain_all();
2485 allow_drain = false;
2486 }
2487
2488 if (isolate_lru_page(page)) {
2489 if (remap) {
2490 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2491 migrate->cpages--;
2492 restore++;
2493 } else {
2494 migrate->src[i] = 0;
2495 unlock_page(page);
2496 migrate->cpages--;
2497 put_page(page);
2498 }
2499 continue;
2500 }
2501
2502 /* Drop the reference we took in collect */
2503 put_page(page);
2504 }
2505
2506 if (!migrate_vma_check_page(page)) {
2507 if (remap) {
2508 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2509 migrate->cpages--;
2510 restore++;
2511
2512 if (!is_zone_device_page(page)) {
2513 get_page(page);
2514 putback_lru_page(page);
2515 }
2516 } else {
2517 migrate->src[i] = 0;
2518 unlock_page(page);
2519 migrate->cpages--;
2520
2521 if (!is_zone_device_page(page))
2522 putback_lru_page(page);
2523 else
2524 put_page(page);
2525 }
2526 }
2527 }
2528
2529 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2530 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2531
2532 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2533 continue;
2534
2535 remove_migration_pte(page, migrate->vma, addr, page);
2536
2537 migrate->src[i] = 0;
2538 unlock_page(page);
2539 put_page(page);
2540 restore--;
2541 }
2542}
2543
2544/*
2545 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2546 * @migrate: migrate struct containing all migration information
2547 *
2548 * Replace page mapping (CPU page table pte) with a special migration pte entry
2549 * and check again if it has been pinned. Pinned pages are restored because we
2550 * cannot migrate them.
2551 *
2552 * This is the last step before we call the device driver callback to allocate
2553 * destination memory and copy contents of original page over to new page.
2554 */
2555static void migrate_vma_unmap(struct migrate_vma *migrate)
2556{
2557 const unsigned long npages = migrate->npages;
2558 const unsigned long start = migrate->start;
2559 unsigned long addr, i, restore = 0;
2560
2561 for (i = 0; i < npages; i++) {
2562 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2563
2564 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2565 continue;
2566
2567 if (page_mapped(page)) {
2568 try_to_migrate(page, 0);
2569 if (page_mapped(page))
2570 goto restore;
2571 }
2572
2573 if (migrate_vma_check_page(page))
2574 continue;
2575
2576restore:
2577 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2578 migrate->cpages--;
2579 restore++;
2580 }
2581
2582 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2583 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2584
2585 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2586 continue;
2587
2588 remove_migration_ptes(page, page, false);
2589
2590 migrate->src[i] = 0;
2591 unlock_page(page);
2592 restore--;
2593
2594 if (is_zone_device_page(page))
2595 put_page(page);
2596 else
2597 putback_lru_page(page);
2598 }
2599}
2600
2601/**
2602 * migrate_vma_setup() - prepare to migrate a range of memory
2603 * @args: contains the vma, start, and pfns arrays for the migration
2604 *
2605 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2606 * without an error.
2607 *
2608 * Prepare to migrate a range of memory virtual address range by collecting all
2609 * the pages backing each virtual address in the range, saving them inside the
2610 * src array. Then lock those pages and unmap them. Once the pages are locked
2611 * and unmapped, check whether each page is pinned or not. Pages that aren't
2612 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2613 * corresponding src array entry. Then restores any pages that are pinned, by
2614 * remapping and unlocking those pages.
2615 *
2616 * The caller should then allocate destination memory and copy source memory to
2617 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2618 * flag set). Once these are allocated and copied, the caller must update each
2619 * corresponding entry in the dst array with the pfn value of the destination
2620 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2621 * (destination pages must have their struct pages locked, via lock_page()).
2622 *
2623 * Note that the caller does not have to migrate all the pages that are marked
2624 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2625 * device memory to system memory. If the caller cannot migrate a device page
2626 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2627 * consequences for the userspace process, so it must be avoided if at all
2628 * possible.
2629 *
2630 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2631 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2632 * allowing the caller to allocate device memory for those unbacked virtual
2633 * addresses. For this the caller simply has to allocate device memory and
2634 * properly set the destination entry like for regular migration. Note that
2635 * this can still fail, and thus inside the device driver you must check if the
2636 * migration was successful for those entries after calling migrate_vma_pages(),
2637 * just like for regular migration.
2638 *
2639 * After that, the callers must call migrate_vma_pages() to go over each entry
2640 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2641 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2642 * then migrate_vma_pages() to migrate struct page information from the source
2643 * struct page to the destination struct page. If it fails to migrate the
2644 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2645 * src array.
2646 *
2647 * At this point all successfully migrated pages have an entry in the src
2648 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2649 * array entry with MIGRATE_PFN_VALID flag set.
2650 *
2651 * Once migrate_vma_pages() returns the caller may inspect which pages were
2652 * successfully migrated, and which were not. Successfully migrated pages will
2653 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2654 *
2655 * It is safe to update device page table after migrate_vma_pages() because
2656 * both destination and source page are still locked, and the mmap_lock is held
2657 * in read mode (hence no one can unmap the range being migrated).
2658 *
2659 * Once the caller is done cleaning up things and updating its page table (if it
2660 * chose to do so, this is not an obligation) it finally calls
2661 * migrate_vma_finalize() to update the CPU page table to point to new pages
2662 * for successfully migrated pages or otherwise restore the CPU page table to
2663 * point to the original source pages.
2664 */
2665int migrate_vma_setup(struct migrate_vma *args)
2666{
2667 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2668
2669 args->start &= PAGE_MASK;
2670 args->end &= PAGE_MASK;
2671 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2672 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2673 return -EINVAL;
2674 if (nr_pages <= 0)
2675 return -EINVAL;
2676 if (args->start < args->vma->vm_start ||
2677 args->start >= args->vma->vm_end)
2678 return -EINVAL;
2679 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2680 return -EINVAL;
2681 if (!args->src || !args->dst)
2682 return -EINVAL;
2683
2684 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2685 args->cpages = 0;
2686 args->npages = 0;
2687
2688 migrate_vma_collect(args);
2689
2690 if (args->cpages)
2691 migrate_vma_prepare(args);
2692 if (args->cpages)
2693 migrate_vma_unmap(args);
2694
2695 /*
2696 * At this point pages are locked and unmapped, and thus they have
2697 * stable content and can safely be copied to destination memory that
2698 * is allocated by the drivers.
2699 */
2700 return 0;
2701
2702}
2703EXPORT_SYMBOL(migrate_vma_setup);
2704
2705/*
2706 * This code closely matches the code in:
2707 * __handle_mm_fault()
2708 * handle_pte_fault()
2709 * do_anonymous_page()
2710 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2711 * private page.
2712 */
2713static void migrate_vma_insert_page(struct migrate_vma *migrate,
2714 unsigned long addr,
2715 struct page *page,
2716 unsigned long *src)
2717{
2718 struct vm_area_struct *vma = migrate->vma;
2719 struct mm_struct *mm = vma->vm_mm;
2720 bool flush = false;
2721 spinlock_t *ptl;
2722 pte_t entry;
2723 pgd_t *pgdp;
2724 p4d_t *p4dp;
2725 pud_t *pudp;
2726 pmd_t *pmdp;
2727 pte_t *ptep;
2728
2729 /* Only allow populating anonymous memory */
2730 if (!vma_is_anonymous(vma))
2731 goto abort;
2732
2733 pgdp = pgd_offset(mm, addr);
2734 p4dp = p4d_alloc(mm, pgdp, addr);
2735 if (!p4dp)
2736 goto abort;
2737 pudp = pud_alloc(mm, p4dp, addr);
2738 if (!pudp)
2739 goto abort;
2740 pmdp = pmd_alloc(mm, pudp, addr);
2741 if (!pmdp)
2742 goto abort;
2743
2744 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2745 goto abort;
2746
2747 /*
2748 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2749 * pte_offset_map() on pmds where a huge pmd might be created
2750 * from a different thread.
2751 *
2752 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2753 * parallel threads are excluded by other means.
2754 *
2755 * Here we only have mmap_read_lock(mm).
2756 */
2757 if (pte_alloc(mm, pmdp))
2758 goto abort;
2759
2760 /* See the comment in pte_alloc_one_map() */
2761 if (unlikely(pmd_trans_unstable(pmdp)))
2762 goto abort;
2763
2764 if (unlikely(anon_vma_prepare(vma)))
2765 goto abort;
2766 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2767 goto abort;
2768
2769 /*
2770 * The memory barrier inside __SetPageUptodate makes sure that
2771 * preceding stores to the page contents become visible before
2772 * the set_pte_at() write.
2773 */
2774 __SetPageUptodate(page);
2775
2776 if (is_zone_device_page(page)) {
2777 if (is_device_private_page(page)) {
2778 swp_entry_t swp_entry;
2779
2780 if (vma->vm_flags & VM_WRITE)
2781 swp_entry = make_writable_device_private_entry(
2782 page_to_pfn(page));
2783 else
2784 swp_entry = make_readable_device_private_entry(
2785 page_to_pfn(page));
2786 entry = swp_entry_to_pte(swp_entry);
2787 } else {
2788 /*
2789 * For now we only support migrating to un-addressable
2790 * device memory.
2791 */
2792 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2793 goto abort;
2794 }
2795 } else {
2796 entry = mk_pte(page, vma->vm_page_prot);
2797 if (vma->vm_flags & VM_WRITE)
2798 entry = pte_mkwrite(pte_mkdirty(entry));
2799 }
2800
2801 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2802
2803 if (check_stable_address_space(mm))
2804 goto unlock_abort;
2805
2806 if (pte_present(*ptep)) {
2807 unsigned long pfn = pte_pfn(*ptep);
2808
2809 if (!is_zero_pfn(pfn))
2810 goto unlock_abort;
2811 flush = true;
2812 } else if (!pte_none(*ptep))
2813 goto unlock_abort;
2814
2815 /*
2816 * Check for userfaultfd but do not deliver the fault. Instead,
2817 * just back off.
2818 */
2819 if (userfaultfd_missing(vma))
2820 goto unlock_abort;
2821
2822 inc_mm_counter(mm, MM_ANONPAGES);
2823 page_add_new_anon_rmap(page, vma, addr, false);
2824 if (!is_zone_device_page(page))
2825 lru_cache_add_inactive_or_unevictable(page, vma);
2826 get_page(page);
2827
2828 if (flush) {
2829 flush_cache_page(vma, addr, pte_pfn(*ptep));
2830 ptep_clear_flush_notify(vma, addr, ptep);
2831 set_pte_at_notify(mm, addr, ptep, entry);
2832 update_mmu_cache(vma, addr, ptep);
2833 } else {
2834 /* No need to invalidate - it was non-present before */
2835 set_pte_at(mm, addr, ptep, entry);
2836 update_mmu_cache(vma, addr, ptep);
2837 }
2838
2839 pte_unmap_unlock(ptep, ptl);
2840 *src = MIGRATE_PFN_MIGRATE;
2841 return;
2842
2843unlock_abort:
2844 pte_unmap_unlock(ptep, ptl);
2845abort:
2846 *src &= ~MIGRATE_PFN_MIGRATE;
2847}
2848
2849/**
2850 * migrate_vma_pages() - migrate meta-data from src page to dst page
2851 * @migrate: migrate struct containing all migration information
2852 *
2853 * This migrates struct page meta-data from source struct page to destination
2854 * struct page. This effectively finishes the migration from source page to the
2855 * destination page.
2856 */
2857void migrate_vma_pages(struct migrate_vma *migrate)
2858{
2859 const unsigned long npages = migrate->npages;
2860 const unsigned long start = migrate->start;
2861 struct mmu_notifier_range range;
2862 unsigned long addr, i;
2863 bool notified = false;
2864
2865 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2866 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2867 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2868 struct address_space *mapping;
2869 int r;
2870
2871 if (!newpage) {
2872 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2873 continue;
2874 }
2875
2876 if (!page) {
2877 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2878 continue;
2879 if (!notified) {
2880 notified = true;
2881
2882 mmu_notifier_range_init_owner(&range,
2883 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
2884 migrate->vma->vm_mm, addr, migrate->end,
2885 migrate->pgmap_owner);
2886 mmu_notifier_invalidate_range_start(&range);
2887 }
2888 migrate_vma_insert_page(migrate, addr, newpage,
2889 &migrate->src[i]);
2890 continue;
2891 }
2892
2893 mapping = page_mapping(page);
2894
2895 if (is_zone_device_page(newpage)) {
2896 if (is_device_private_page(newpage)) {
2897 /*
2898 * For now only support private anonymous when
2899 * migrating to un-addressable device memory.
2900 */
2901 if (mapping) {
2902 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2903 continue;
2904 }
2905 } else {
2906 /*
2907 * Other types of ZONE_DEVICE page are not
2908 * supported.
2909 */
2910 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2911 continue;
2912 }
2913 }
2914
2915 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2916 if (r != MIGRATEPAGE_SUCCESS)
2917 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2918 }
2919
2920 /*
2921 * No need to double call mmu_notifier->invalidate_range() callback as
2922 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2923 * did already call it.
2924 */
2925 if (notified)
2926 mmu_notifier_invalidate_range_only_end(&range);
2927}
2928EXPORT_SYMBOL(migrate_vma_pages);
2929
2930/**
2931 * migrate_vma_finalize() - restore CPU page table entry
2932 * @migrate: migrate struct containing all migration information
2933 *
2934 * This replaces the special migration pte entry with either a mapping to the
2935 * new page if migration was successful for that page, or to the original page
2936 * otherwise.
2937 *
2938 * This also unlocks the pages and puts them back on the lru, or drops the extra
2939 * refcount, for device pages.
2940 */
2941void migrate_vma_finalize(struct migrate_vma *migrate)
2942{
2943 const unsigned long npages = migrate->npages;
2944 unsigned long i;
2945
2946 for (i = 0; i < npages; i++) {
2947 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2948 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2949
2950 if (!page) {
2951 if (newpage) {
2952 unlock_page(newpage);
2953 put_page(newpage);
2954 }
2955 continue;
2956 }
2957
2958 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2959 if (newpage) {
2960 unlock_page(newpage);
2961 put_page(newpage);
2962 }
2963 newpage = page;
2964 }
2965
2966 remove_migration_ptes(page, newpage, false);
2967 unlock_page(page);
2968
2969 if (is_zone_device_page(page))
2970 put_page(page);
2971 else
2972 putback_lru_page(page);
2973
2974 if (newpage != page) {
2975 unlock_page(newpage);
2976 if (is_zone_device_page(newpage))
2977 put_page(newpage);
2978 else
2979 putback_lru_page(newpage);
2980 }
2981 }
2982}
2983EXPORT_SYMBOL(migrate_vma_finalize);
2984#endif /* CONFIG_DEVICE_PRIVATE */
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter
13 */
14
15#include <linux/migrate.h>
16#include <linux/export.h>
17#include <linux/swap.h>
18#include <linux/swapops.h>
19#include <linux/pagemap.h>
20#include <linux/buffer_head.h>
21#include <linux/mm_inline.h>
22#include <linux/nsproxy.h>
23#include <linux/pagevec.h>
24#include <linux/ksm.h>
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
29#include <linux/writeback.h>
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
32#include <linux/security.h>
33#include <linux/memcontrol.h>
34#include <linux/syscalls.h>
35#include <linux/hugetlb.h>
36#include <linux/gfp.h>
37
38#include <asm/tlbflush.h>
39
40#include "internal.h"
41
42/*
43 * migrate_prep() needs to be called before we start compiling a list of pages
44 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
45 * undesirable, use migrate_prep_local()
46 */
47int migrate_prep(void)
48{
49 /*
50 * Clear the LRU lists so pages can be isolated.
51 * Note that pages may be moved off the LRU after we have
52 * drained them. Those pages will fail to migrate like other
53 * pages that may be busy.
54 */
55 lru_add_drain_all();
56
57 return 0;
58}
59
60/* Do the necessary work of migrate_prep but not if it involves other CPUs */
61int migrate_prep_local(void)
62{
63 lru_add_drain();
64
65 return 0;
66}
67
68/*
69 * Add isolated pages on the list back to the LRU under page lock
70 * to avoid leaking evictable pages back onto unevictable list.
71 */
72void putback_lru_pages(struct list_head *l)
73{
74 struct page *page;
75 struct page *page2;
76
77 list_for_each_entry_safe(page, page2, l, lru) {
78 list_del(&page->lru);
79 dec_zone_page_state(page, NR_ISOLATED_ANON +
80 page_is_file_cache(page));
81 putback_lru_page(page);
82 }
83}
84
85/*
86 * Restore a potential migration pte to a working pte entry
87 */
88static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
89 unsigned long addr, void *old)
90{
91 struct mm_struct *mm = vma->vm_mm;
92 swp_entry_t entry;
93 pgd_t *pgd;
94 pud_t *pud;
95 pmd_t *pmd;
96 pte_t *ptep, pte;
97 spinlock_t *ptl;
98
99 if (unlikely(PageHuge(new))) {
100 ptep = huge_pte_offset(mm, addr);
101 if (!ptep)
102 goto out;
103 ptl = &mm->page_table_lock;
104 } else {
105 pgd = pgd_offset(mm, addr);
106 if (!pgd_present(*pgd))
107 goto out;
108
109 pud = pud_offset(pgd, addr);
110 if (!pud_present(*pud))
111 goto out;
112
113 pmd = pmd_offset(pud, addr);
114 if (pmd_trans_huge(*pmd))
115 goto out;
116 if (!pmd_present(*pmd))
117 goto out;
118
119 ptep = pte_offset_map(pmd, addr);
120
121 /*
122 * Peek to check is_swap_pte() before taking ptlock? No, we
123 * can race mremap's move_ptes(), which skips anon_vma lock.
124 */
125
126 ptl = pte_lockptr(mm, pmd);
127 }
128
129 spin_lock(ptl);
130 pte = *ptep;
131 if (!is_swap_pte(pte))
132 goto unlock;
133
134 entry = pte_to_swp_entry(pte);
135
136 if (!is_migration_entry(entry) ||
137 migration_entry_to_page(entry) != old)
138 goto unlock;
139
140 get_page(new);
141 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
142 if (is_write_migration_entry(entry))
143 pte = pte_mkwrite(pte);
144#ifdef CONFIG_HUGETLB_PAGE
145 if (PageHuge(new))
146 pte = pte_mkhuge(pte);
147#endif
148 flush_cache_page(vma, addr, pte_pfn(pte));
149 set_pte_at(mm, addr, ptep, pte);
150
151 if (PageHuge(new)) {
152 if (PageAnon(new))
153 hugepage_add_anon_rmap(new, vma, addr);
154 else
155 page_dup_rmap(new);
156 } else if (PageAnon(new))
157 page_add_anon_rmap(new, vma, addr);
158 else
159 page_add_file_rmap(new);
160
161 /* No need to invalidate - it was non-present before */
162 update_mmu_cache(vma, addr, ptep);
163unlock:
164 pte_unmap_unlock(ptep, ptl);
165out:
166 return SWAP_AGAIN;
167}
168
169/*
170 * Get rid of all migration entries and replace them by
171 * references to the indicated page.
172 */
173static void remove_migration_ptes(struct page *old, struct page *new)
174{
175 rmap_walk(new, remove_migration_pte, old);
176}
177
178/*
179 * Something used the pte of a page under migration. We need to
180 * get to the page and wait until migration is finished.
181 * When we return from this function the fault will be retried.
182 */
183void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
184 unsigned long address)
185{
186 pte_t *ptep, pte;
187 spinlock_t *ptl;
188 swp_entry_t entry;
189 struct page *page;
190
191 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
192 pte = *ptep;
193 if (!is_swap_pte(pte))
194 goto out;
195
196 entry = pte_to_swp_entry(pte);
197 if (!is_migration_entry(entry))
198 goto out;
199
200 page = migration_entry_to_page(entry);
201
202 /*
203 * Once radix-tree replacement of page migration started, page_count
204 * *must* be zero. And, we don't want to call wait_on_page_locked()
205 * against a page without get_page().
206 * So, we use get_page_unless_zero(), here. Even failed, page fault
207 * will occur again.
208 */
209 if (!get_page_unless_zero(page))
210 goto out;
211 pte_unmap_unlock(ptep, ptl);
212 wait_on_page_locked(page);
213 put_page(page);
214 return;
215out:
216 pte_unmap_unlock(ptep, ptl);
217}
218
219#ifdef CONFIG_BLOCK
220/* Returns true if all buffers are successfully locked */
221static bool buffer_migrate_lock_buffers(struct buffer_head *head,
222 enum migrate_mode mode)
223{
224 struct buffer_head *bh = head;
225
226 /* Simple case, sync compaction */
227 if (mode != MIGRATE_ASYNC) {
228 do {
229 get_bh(bh);
230 lock_buffer(bh);
231 bh = bh->b_this_page;
232
233 } while (bh != head);
234
235 return true;
236 }
237
238 /* async case, we cannot block on lock_buffer so use trylock_buffer */
239 do {
240 get_bh(bh);
241 if (!trylock_buffer(bh)) {
242 /*
243 * We failed to lock the buffer and cannot stall in
244 * async migration. Release the taken locks
245 */
246 struct buffer_head *failed_bh = bh;
247 put_bh(failed_bh);
248 bh = head;
249 while (bh != failed_bh) {
250 unlock_buffer(bh);
251 put_bh(bh);
252 bh = bh->b_this_page;
253 }
254 return false;
255 }
256
257 bh = bh->b_this_page;
258 } while (bh != head);
259 return true;
260}
261#else
262static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
263 enum migrate_mode mode)
264{
265 return true;
266}
267#endif /* CONFIG_BLOCK */
268
269/*
270 * Replace the page in the mapping.
271 *
272 * The number of remaining references must be:
273 * 1 for anonymous pages without a mapping
274 * 2 for pages with a mapping
275 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
276 */
277static int migrate_page_move_mapping(struct address_space *mapping,
278 struct page *newpage, struct page *page,
279 struct buffer_head *head, enum migrate_mode mode)
280{
281 int expected_count;
282 void **pslot;
283
284 if (!mapping) {
285 /* Anonymous page without mapping */
286 if (page_count(page) != 1)
287 return -EAGAIN;
288 return 0;
289 }
290
291 spin_lock_irq(&mapping->tree_lock);
292
293 pslot = radix_tree_lookup_slot(&mapping->page_tree,
294 page_index(page));
295
296 expected_count = 2 + page_has_private(page);
297 if (page_count(page) != expected_count ||
298 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
299 spin_unlock_irq(&mapping->tree_lock);
300 return -EAGAIN;
301 }
302
303 if (!page_freeze_refs(page, expected_count)) {
304 spin_unlock_irq(&mapping->tree_lock);
305 return -EAGAIN;
306 }
307
308 /*
309 * In the async migration case of moving a page with buffers, lock the
310 * buffers using trylock before the mapping is moved. If the mapping
311 * was moved, we later failed to lock the buffers and could not move
312 * the mapping back due to an elevated page count, we would have to
313 * block waiting on other references to be dropped.
314 */
315 if (mode == MIGRATE_ASYNC && head &&
316 !buffer_migrate_lock_buffers(head, mode)) {
317 page_unfreeze_refs(page, expected_count);
318 spin_unlock_irq(&mapping->tree_lock);
319 return -EAGAIN;
320 }
321
322 /*
323 * Now we know that no one else is looking at the page.
324 */
325 get_page(newpage); /* add cache reference */
326 if (PageSwapCache(page)) {
327 SetPageSwapCache(newpage);
328 set_page_private(newpage, page_private(page));
329 }
330
331 radix_tree_replace_slot(pslot, newpage);
332
333 /*
334 * Drop cache reference from old page by unfreezing
335 * to one less reference.
336 * We know this isn't the last reference.
337 */
338 page_unfreeze_refs(page, expected_count - 1);
339
340 /*
341 * If moved to a different zone then also account
342 * the page for that zone. Other VM counters will be
343 * taken care of when we establish references to the
344 * new page and drop references to the old page.
345 *
346 * Note that anonymous pages are accounted for
347 * via NR_FILE_PAGES and NR_ANON_PAGES if they
348 * are mapped to swap space.
349 */
350 __dec_zone_page_state(page, NR_FILE_PAGES);
351 __inc_zone_page_state(newpage, NR_FILE_PAGES);
352 if (!PageSwapCache(page) && PageSwapBacked(page)) {
353 __dec_zone_page_state(page, NR_SHMEM);
354 __inc_zone_page_state(newpage, NR_SHMEM);
355 }
356 spin_unlock_irq(&mapping->tree_lock);
357
358 return 0;
359}
360
361/*
362 * The expected number of remaining references is the same as that
363 * of migrate_page_move_mapping().
364 */
365int migrate_huge_page_move_mapping(struct address_space *mapping,
366 struct page *newpage, struct page *page)
367{
368 int expected_count;
369 void **pslot;
370
371 if (!mapping) {
372 if (page_count(page) != 1)
373 return -EAGAIN;
374 return 0;
375 }
376
377 spin_lock_irq(&mapping->tree_lock);
378
379 pslot = radix_tree_lookup_slot(&mapping->page_tree,
380 page_index(page));
381
382 expected_count = 2 + page_has_private(page);
383 if (page_count(page) != expected_count ||
384 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
385 spin_unlock_irq(&mapping->tree_lock);
386 return -EAGAIN;
387 }
388
389 if (!page_freeze_refs(page, expected_count)) {
390 spin_unlock_irq(&mapping->tree_lock);
391 return -EAGAIN;
392 }
393
394 get_page(newpage);
395
396 radix_tree_replace_slot(pslot, newpage);
397
398 page_unfreeze_refs(page, expected_count - 1);
399
400 spin_unlock_irq(&mapping->tree_lock);
401 return 0;
402}
403
404/*
405 * Copy the page to its new location
406 */
407void migrate_page_copy(struct page *newpage, struct page *page)
408{
409 if (PageHuge(page))
410 copy_huge_page(newpage, page);
411 else
412 copy_highpage(newpage, page);
413
414 if (PageError(page))
415 SetPageError(newpage);
416 if (PageReferenced(page))
417 SetPageReferenced(newpage);
418 if (PageUptodate(page))
419 SetPageUptodate(newpage);
420 if (TestClearPageActive(page)) {
421 VM_BUG_ON(PageUnevictable(page));
422 SetPageActive(newpage);
423 } else if (TestClearPageUnevictable(page))
424 SetPageUnevictable(newpage);
425 if (PageChecked(page))
426 SetPageChecked(newpage);
427 if (PageMappedToDisk(page))
428 SetPageMappedToDisk(newpage);
429
430 if (PageDirty(page)) {
431 clear_page_dirty_for_io(page);
432 /*
433 * Want to mark the page and the radix tree as dirty, and
434 * redo the accounting that clear_page_dirty_for_io undid,
435 * but we can't use set_page_dirty because that function
436 * is actually a signal that all of the page has become dirty.
437 * Whereas only part of our page may be dirty.
438 */
439 if (PageSwapBacked(page))
440 SetPageDirty(newpage);
441 else
442 __set_page_dirty_nobuffers(newpage);
443 }
444
445 mlock_migrate_page(newpage, page);
446 ksm_migrate_page(newpage, page);
447
448 ClearPageSwapCache(page);
449 ClearPagePrivate(page);
450 set_page_private(page, 0);
451
452 /*
453 * If any waiters have accumulated on the new page then
454 * wake them up.
455 */
456 if (PageWriteback(newpage))
457 end_page_writeback(newpage);
458}
459
460/************************************************************
461 * Migration functions
462 ***********************************************************/
463
464/* Always fail migration. Used for mappings that are not movable */
465int fail_migrate_page(struct address_space *mapping,
466 struct page *newpage, struct page *page)
467{
468 return -EIO;
469}
470EXPORT_SYMBOL(fail_migrate_page);
471
472/*
473 * Common logic to directly migrate a single page suitable for
474 * pages that do not use PagePrivate/PagePrivate2.
475 *
476 * Pages are locked upon entry and exit.
477 */
478int migrate_page(struct address_space *mapping,
479 struct page *newpage, struct page *page,
480 enum migrate_mode mode)
481{
482 int rc;
483
484 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
485
486 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
487
488 if (rc)
489 return rc;
490
491 migrate_page_copy(newpage, page);
492 return 0;
493}
494EXPORT_SYMBOL(migrate_page);
495
496#ifdef CONFIG_BLOCK
497/*
498 * Migration function for pages with buffers. This function can only be used
499 * if the underlying filesystem guarantees that no other references to "page"
500 * exist.
501 */
502int buffer_migrate_page(struct address_space *mapping,
503 struct page *newpage, struct page *page, enum migrate_mode mode)
504{
505 struct buffer_head *bh, *head;
506 int rc;
507
508 if (!page_has_buffers(page))
509 return migrate_page(mapping, newpage, page, mode);
510
511 head = page_buffers(page);
512
513 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
514
515 if (rc)
516 return rc;
517
518 /*
519 * In the async case, migrate_page_move_mapping locked the buffers
520 * with an IRQ-safe spinlock held. In the sync case, the buffers
521 * need to be locked now
522 */
523 if (mode != MIGRATE_ASYNC)
524 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
525
526 ClearPagePrivate(page);
527 set_page_private(newpage, page_private(page));
528 set_page_private(page, 0);
529 put_page(page);
530 get_page(newpage);
531
532 bh = head;
533 do {
534 set_bh_page(bh, newpage, bh_offset(bh));
535 bh = bh->b_this_page;
536
537 } while (bh != head);
538
539 SetPagePrivate(newpage);
540
541 migrate_page_copy(newpage, page);
542
543 bh = head;
544 do {
545 unlock_buffer(bh);
546 put_bh(bh);
547 bh = bh->b_this_page;
548
549 } while (bh != head);
550
551 return 0;
552}
553EXPORT_SYMBOL(buffer_migrate_page);
554#endif
555
556/*
557 * Writeback a page to clean the dirty state
558 */
559static int writeout(struct address_space *mapping, struct page *page)
560{
561 struct writeback_control wbc = {
562 .sync_mode = WB_SYNC_NONE,
563 .nr_to_write = 1,
564 .range_start = 0,
565 .range_end = LLONG_MAX,
566 .for_reclaim = 1
567 };
568 int rc;
569
570 if (!mapping->a_ops->writepage)
571 /* No write method for the address space */
572 return -EINVAL;
573
574 if (!clear_page_dirty_for_io(page))
575 /* Someone else already triggered a write */
576 return -EAGAIN;
577
578 /*
579 * A dirty page may imply that the underlying filesystem has
580 * the page on some queue. So the page must be clean for
581 * migration. Writeout may mean we loose the lock and the
582 * page state is no longer what we checked for earlier.
583 * At this point we know that the migration attempt cannot
584 * be successful.
585 */
586 remove_migration_ptes(page, page);
587
588 rc = mapping->a_ops->writepage(page, &wbc);
589
590 if (rc != AOP_WRITEPAGE_ACTIVATE)
591 /* unlocked. Relock */
592 lock_page(page);
593
594 return (rc < 0) ? -EIO : -EAGAIN;
595}
596
597/*
598 * Default handling if a filesystem does not provide a migration function.
599 */
600static int fallback_migrate_page(struct address_space *mapping,
601 struct page *newpage, struct page *page, enum migrate_mode mode)
602{
603 if (PageDirty(page)) {
604 /* Only writeback pages in full synchronous migration */
605 if (mode != MIGRATE_SYNC)
606 return -EBUSY;
607 return writeout(mapping, page);
608 }
609
610 /*
611 * Buffers may be managed in a filesystem specific way.
612 * We must have no buffers or drop them.
613 */
614 if (page_has_private(page) &&
615 !try_to_release_page(page, GFP_KERNEL))
616 return -EAGAIN;
617
618 return migrate_page(mapping, newpage, page, mode);
619}
620
621/*
622 * Move a page to a newly allocated page
623 * The page is locked and all ptes have been successfully removed.
624 *
625 * The new page will have replaced the old page if this function
626 * is successful.
627 *
628 * Return value:
629 * < 0 - error code
630 * == 0 - success
631 */
632static int move_to_new_page(struct page *newpage, struct page *page,
633 int remap_swapcache, enum migrate_mode mode)
634{
635 struct address_space *mapping;
636 int rc;
637
638 /*
639 * Block others from accessing the page when we get around to
640 * establishing additional references. We are the only one
641 * holding a reference to the new page at this point.
642 */
643 if (!trylock_page(newpage))
644 BUG();
645
646 /* Prepare mapping for the new page.*/
647 newpage->index = page->index;
648 newpage->mapping = page->mapping;
649 if (PageSwapBacked(page))
650 SetPageSwapBacked(newpage);
651
652 mapping = page_mapping(page);
653 if (!mapping)
654 rc = migrate_page(mapping, newpage, page, mode);
655 else if (mapping->a_ops->migratepage)
656 /*
657 * Most pages have a mapping and most filesystems provide a
658 * migratepage callback. Anonymous pages are part of swap
659 * space which also has its own migratepage callback. This
660 * is the most common path for page migration.
661 */
662 rc = mapping->a_ops->migratepage(mapping,
663 newpage, page, mode);
664 else
665 rc = fallback_migrate_page(mapping, newpage, page, mode);
666
667 if (rc) {
668 newpage->mapping = NULL;
669 } else {
670 if (remap_swapcache)
671 remove_migration_ptes(page, newpage);
672 page->mapping = NULL;
673 }
674
675 unlock_page(newpage);
676
677 return rc;
678}
679
680static int __unmap_and_move(struct page *page, struct page *newpage,
681 int force, bool offlining, enum migrate_mode mode)
682{
683 int rc = -EAGAIN;
684 int remap_swapcache = 1;
685 int charge = 0;
686 struct mem_cgroup *mem;
687 struct anon_vma *anon_vma = NULL;
688
689 if (!trylock_page(page)) {
690 if (!force || mode == MIGRATE_ASYNC)
691 goto out;
692
693 /*
694 * It's not safe for direct compaction to call lock_page.
695 * For example, during page readahead pages are added locked
696 * to the LRU. Later, when the IO completes the pages are
697 * marked uptodate and unlocked. However, the queueing
698 * could be merging multiple pages for one bio (e.g.
699 * mpage_readpages). If an allocation happens for the
700 * second or third page, the process can end up locking
701 * the same page twice and deadlocking. Rather than
702 * trying to be clever about what pages can be locked,
703 * avoid the use of lock_page for direct compaction
704 * altogether.
705 */
706 if (current->flags & PF_MEMALLOC)
707 goto out;
708
709 lock_page(page);
710 }
711
712 /*
713 * Only memory hotplug's offline_pages() caller has locked out KSM,
714 * and can safely migrate a KSM page. The other cases have skipped
715 * PageKsm along with PageReserved - but it is only now when we have
716 * the page lock that we can be certain it will not go KSM beneath us
717 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
718 * its pagecount raised, but only here do we take the page lock which
719 * serializes that).
720 */
721 if (PageKsm(page) && !offlining) {
722 rc = -EBUSY;
723 goto unlock;
724 }
725
726 /* charge against new page */
727 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
728 if (charge == -ENOMEM) {
729 rc = -ENOMEM;
730 goto unlock;
731 }
732 BUG_ON(charge);
733
734 if (PageWriteback(page)) {
735 /*
736 * Only in the case of a full syncronous migration is it
737 * necessary to wait for PageWriteback. In the async case,
738 * the retry loop is too short and in the sync-light case,
739 * the overhead of stalling is too much
740 */
741 if (mode != MIGRATE_SYNC) {
742 rc = -EBUSY;
743 goto uncharge;
744 }
745 if (!force)
746 goto uncharge;
747 wait_on_page_writeback(page);
748 }
749 /*
750 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
751 * we cannot notice that anon_vma is freed while we migrates a page.
752 * This get_anon_vma() delays freeing anon_vma pointer until the end
753 * of migration. File cache pages are no problem because of page_lock()
754 * File Caches may use write_page() or lock_page() in migration, then,
755 * just care Anon page here.
756 */
757 if (PageAnon(page)) {
758 /*
759 * Only page_lock_anon_vma() understands the subtleties of
760 * getting a hold on an anon_vma from outside one of its mms.
761 */
762 anon_vma = page_get_anon_vma(page);
763 if (anon_vma) {
764 /*
765 * Anon page
766 */
767 } else if (PageSwapCache(page)) {
768 /*
769 * We cannot be sure that the anon_vma of an unmapped
770 * swapcache page is safe to use because we don't
771 * know in advance if the VMA that this page belonged
772 * to still exists. If the VMA and others sharing the
773 * data have been freed, then the anon_vma could
774 * already be invalid.
775 *
776 * To avoid this possibility, swapcache pages get
777 * migrated but are not remapped when migration
778 * completes
779 */
780 remap_swapcache = 0;
781 } else {
782 goto uncharge;
783 }
784 }
785
786 /*
787 * Corner case handling:
788 * 1. When a new swap-cache page is read into, it is added to the LRU
789 * and treated as swapcache but it has no rmap yet.
790 * Calling try_to_unmap() against a page->mapping==NULL page will
791 * trigger a BUG. So handle it here.
792 * 2. An orphaned page (see truncate_complete_page) might have
793 * fs-private metadata. The page can be picked up due to memory
794 * offlining. Everywhere else except page reclaim, the page is
795 * invisible to the vm, so the page can not be migrated. So try to
796 * free the metadata, so the page can be freed.
797 */
798 if (!page->mapping) {
799 VM_BUG_ON(PageAnon(page));
800 if (page_has_private(page)) {
801 try_to_free_buffers(page);
802 goto uncharge;
803 }
804 goto skip_unmap;
805 }
806
807 /* Establish migration ptes or remove ptes */
808 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
809
810skip_unmap:
811 if (!page_mapped(page))
812 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
813
814 if (rc && remap_swapcache)
815 remove_migration_ptes(page, page);
816
817 /* Drop an anon_vma reference if we took one */
818 if (anon_vma)
819 put_anon_vma(anon_vma);
820
821uncharge:
822 if (!charge)
823 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
824unlock:
825 unlock_page(page);
826out:
827 return rc;
828}
829
830/*
831 * Obtain the lock on page, remove all ptes and migrate the page
832 * to the newly allocated page in newpage.
833 */
834static int unmap_and_move(new_page_t get_new_page, unsigned long private,
835 struct page *page, int force, bool offlining,
836 enum migrate_mode mode)
837{
838 int rc = 0;
839 int *result = NULL;
840 struct page *newpage = get_new_page(page, private, &result);
841
842 if (!newpage)
843 return -ENOMEM;
844
845 if (page_count(page) == 1) {
846 /* page was freed from under us. So we are done. */
847 goto out;
848 }
849
850 if (unlikely(PageTransHuge(page)))
851 if (unlikely(split_huge_page(page)))
852 goto out;
853
854 rc = __unmap_and_move(page, newpage, force, offlining, mode);
855out:
856 if (rc != -EAGAIN) {
857 /*
858 * A page that has been migrated has all references
859 * removed and will be freed. A page that has not been
860 * migrated will have kepts its references and be
861 * restored.
862 */
863 list_del(&page->lru);
864 dec_zone_page_state(page, NR_ISOLATED_ANON +
865 page_is_file_cache(page));
866 putback_lru_page(page);
867 }
868 /*
869 * Move the new page to the LRU. If migration was not successful
870 * then this will free the page.
871 */
872 putback_lru_page(newpage);
873 if (result) {
874 if (rc)
875 *result = rc;
876 else
877 *result = page_to_nid(newpage);
878 }
879 return rc;
880}
881
882/*
883 * Counterpart of unmap_and_move_page() for hugepage migration.
884 *
885 * This function doesn't wait the completion of hugepage I/O
886 * because there is no race between I/O and migration for hugepage.
887 * Note that currently hugepage I/O occurs only in direct I/O
888 * where no lock is held and PG_writeback is irrelevant,
889 * and writeback status of all subpages are counted in the reference
890 * count of the head page (i.e. if all subpages of a 2MB hugepage are
891 * under direct I/O, the reference of the head page is 512 and a bit more.)
892 * This means that when we try to migrate hugepage whose subpages are
893 * doing direct I/O, some references remain after try_to_unmap() and
894 * hugepage migration fails without data corruption.
895 *
896 * There is also no race when direct I/O is issued on the page under migration,
897 * because then pte is replaced with migration swap entry and direct I/O code
898 * will wait in the page fault for migration to complete.
899 */
900static int unmap_and_move_huge_page(new_page_t get_new_page,
901 unsigned long private, struct page *hpage,
902 int force, bool offlining,
903 enum migrate_mode mode)
904{
905 int rc = 0;
906 int *result = NULL;
907 struct page *new_hpage = get_new_page(hpage, private, &result);
908 struct anon_vma *anon_vma = NULL;
909
910 if (!new_hpage)
911 return -ENOMEM;
912
913 rc = -EAGAIN;
914
915 if (!trylock_page(hpage)) {
916 if (!force || mode != MIGRATE_SYNC)
917 goto out;
918 lock_page(hpage);
919 }
920
921 if (PageAnon(hpage))
922 anon_vma = page_get_anon_vma(hpage);
923
924 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
925
926 if (!page_mapped(hpage))
927 rc = move_to_new_page(new_hpage, hpage, 1, mode);
928
929 if (rc)
930 remove_migration_ptes(hpage, hpage);
931
932 if (anon_vma)
933 put_anon_vma(anon_vma);
934 unlock_page(hpage);
935
936out:
937 if (rc != -EAGAIN) {
938 list_del(&hpage->lru);
939 put_page(hpage);
940 }
941
942 put_page(new_hpage);
943
944 if (result) {
945 if (rc)
946 *result = rc;
947 else
948 *result = page_to_nid(new_hpage);
949 }
950 return rc;
951}
952
953/*
954 * migrate_pages
955 *
956 * The function takes one list of pages to migrate and a function
957 * that determines from the page to be migrated and the private data
958 * the target of the move and allocates the page.
959 *
960 * The function returns after 10 attempts or if no pages
961 * are movable anymore because to has become empty
962 * or no retryable pages exist anymore.
963 * Caller should call putback_lru_pages to return pages to the LRU
964 * or free list only if ret != 0.
965 *
966 * Return: Number of pages not migrated or error code.
967 */
968int migrate_pages(struct list_head *from,
969 new_page_t get_new_page, unsigned long private, bool offlining,
970 enum migrate_mode mode)
971{
972 int retry = 1;
973 int nr_failed = 0;
974 int pass = 0;
975 struct page *page;
976 struct page *page2;
977 int swapwrite = current->flags & PF_SWAPWRITE;
978 int rc;
979
980 if (!swapwrite)
981 current->flags |= PF_SWAPWRITE;
982
983 for(pass = 0; pass < 10 && retry; pass++) {
984 retry = 0;
985
986 list_for_each_entry_safe(page, page2, from, lru) {
987 cond_resched();
988
989 rc = unmap_and_move(get_new_page, private,
990 page, pass > 2, offlining,
991 mode);
992
993 switch(rc) {
994 case -ENOMEM:
995 goto out;
996 case -EAGAIN:
997 retry++;
998 break;
999 case 0:
1000 break;
1001 default:
1002 /* Permanent failure */
1003 nr_failed++;
1004 break;
1005 }
1006 }
1007 }
1008 rc = 0;
1009out:
1010 if (!swapwrite)
1011 current->flags &= ~PF_SWAPWRITE;
1012
1013 if (rc)
1014 return rc;
1015
1016 return nr_failed + retry;
1017}
1018
1019int migrate_huge_pages(struct list_head *from,
1020 new_page_t get_new_page, unsigned long private, bool offlining,
1021 enum migrate_mode mode)
1022{
1023 int retry = 1;
1024 int nr_failed = 0;
1025 int pass = 0;
1026 struct page *page;
1027 struct page *page2;
1028 int rc;
1029
1030 for (pass = 0; pass < 10 && retry; pass++) {
1031 retry = 0;
1032
1033 list_for_each_entry_safe(page, page2, from, lru) {
1034 cond_resched();
1035
1036 rc = unmap_and_move_huge_page(get_new_page,
1037 private, page, pass > 2, offlining,
1038 mode);
1039
1040 switch(rc) {
1041 case -ENOMEM:
1042 goto out;
1043 case -EAGAIN:
1044 retry++;
1045 break;
1046 case 0:
1047 break;
1048 default:
1049 /* Permanent failure */
1050 nr_failed++;
1051 break;
1052 }
1053 }
1054 }
1055 rc = 0;
1056out:
1057 if (rc)
1058 return rc;
1059
1060 return nr_failed + retry;
1061}
1062
1063#ifdef CONFIG_NUMA
1064/*
1065 * Move a list of individual pages
1066 */
1067struct page_to_node {
1068 unsigned long addr;
1069 struct page *page;
1070 int node;
1071 int status;
1072};
1073
1074static struct page *new_page_node(struct page *p, unsigned long private,
1075 int **result)
1076{
1077 struct page_to_node *pm = (struct page_to_node *)private;
1078
1079 while (pm->node != MAX_NUMNODES && pm->page != p)
1080 pm++;
1081
1082 if (pm->node == MAX_NUMNODES)
1083 return NULL;
1084
1085 *result = &pm->status;
1086
1087 return alloc_pages_exact_node(pm->node,
1088 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1089}
1090
1091/*
1092 * Move a set of pages as indicated in the pm array. The addr
1093 * field must be set to the virtual address of the page to be moved
1094 * and the node number must contain a valid target node.
1095 * The pm array ends with node = MAX_NUMNODES.
1096 */
1097static int do_move_page_to_node_array(struct mm_struct *mm,
1098 struct page_to_node *pm,
1099 int migrate_all)
1100{
1101 int err;
1102 struct page_to_node *pp;
1103 LIST_HEAD(pagelist);
1104
1105 down_read(&mm->mmap_sem);
1106
1107 /*
1108 * Build a list of pages to migrate
1109 */
1110 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1111 struct vm_area_struct *vma;
1112 struct page *page;
1113
1114 err = -EFAULT;
1115 vma = find_vma(mm, pp->addr);
1116 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1117 goto set_status;
1118
1119 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1120
1121 err = PTR_ERR(page);
1122 if (IS_ERR(page))
1123 goto set_status;
1124
1125 err = -ENOENT;
1126 if (!page)
1127 goto set_status;
1128
1129 /* Use PageReserved to check for zero page */
1130 if (PageReserved(page) || PageKsm(page))
1131 goto put_and_set;
1132
1133 pp->page = page;
1134 err = page_to_nid(page);
1135
1136 if (err == pp->node)
1137 /*
1138 * Node already in the right place
1139 */
1140 goto put_and_set;
1141
1142 err = -EACCES;
1143 if (page_mapcount(page) > 1 &&
1144 !migrate_all)
1145 goto put_and_set;
1146
1147 err = isolate_lru_page(page);
1148 if (!err) {
1149 list_add_tail(&page->lru, &pagelist);
1150 inc_zone_page_state(page, NR_ISOLATED_ANON +
1151 page_is_file_cache(page));
1152 }
1153put_and_set:
1154 /*
1155 * Either remove the duplicate refcount from
1156 * isolate_lru_page() or drop the page ref if it was
1157 * not isolated.
1158 */
1159 put_page(page);
1160set_status:
1161 pp->status = err;
1162 }
1163
1164 err = 0;
1165 if (!list_empty(&pagelist)) {
1166 err = migrate_pages(&pagelist, new_page_node,
1167 (unsigned long)pm, 0, MIGRATE_SYNC);
1168 if (err)
1169 putback_lru_pages(&pagelist);
1170 }
1171
1172 up_read(&mm->mmap_sem);
1173 return err;
1174}
1175
1176/*
1177 * Migrate an array of page address onto an array of nodes and fill
1178 * the corresponding array of status.
1179 */
1180static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1181 unsigned long nr_pages,
1182 const void __user * __user *pages,
1183 const int __user *nodes,
1184 int __user *status, int flags)
1185{
1186 struct page_to_node *pm;
1187 unsigned long chunk_nr_pages;
1188 unsigned long chunk_start;
1189 int err;
1190
1191 err = -ENOMEM;
1192 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1193 if (!pm)
1194 goto out;
1195
1196 migrate_prep();
1197
1198 /*
1199 * Store a chunk of page_to_node array in a page,
1200 * but keep the last one as a marker
1201 */
1202 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1203
1204 for (chunk_start = 0;
1205 chunk_start < nr_pages;
1206 chunk_start += chunk_nr_pages) {
1207 int j;
1208
1209 if (chunk_start + chunk_nr_pages > nr_pages)
1210 chunk_nr_pages = nr_pages - chunk_start;
1211
1212 /* fill the chunk pm with addrs and nodes from user-space */
1213 for (j = 0; j < chunk_nr_pages; j++) {
1214 const void __user *p;
1215 int node;
1216
1217 err = -EFAULT;
1218 if (get_user(p, pages + j + chunk_start))
1219 goto out_pm;
1220 pm[j].addr = (unsigned long) p;
1221
1222 if (get_user(node, nodes + j + chunk_start))
1223 goto out_pm;
1224
1225 err = -ENODEV;
1226 if (node < 0 || node >= MAX_NUMNODES)
1227 goto out_pm;
1228
1229 if (!node_state(node, N_HIGH_MEMORY))
1230 goto out_pm;
1231
1232 err = -EACCES;
1233 if (!node_isset(node, task_nodes))
1234 goto out_pm;
1235
1236 pm[j].node = node;
1237 }
1238
1239 /* End marker for this chunk */
1240 pm[chunk_nr_pages].node = MAX_NUMNODES;
1241
1242 /* Migrate this chunk */
1243 err = do_move_page_to_node_array(mm, pm,
1244 flags & MPOL_MF_MOVE_ALL);
1245 if (err < 0)
1246 goto out_pm;
1247
1248 /* Return status information */
1249 for (j = 0; j < chunk_nr_pages; j++)
1250 if (put_user(pm[j].status, status + j + chunk_start)) {
1251 err = -EFAULT;
1252 goto out_pm;
1253 }
1254 }
1255 err = 0;
1256
1257out_pm:
1258 free_page((unsigned long)pm);
1259out:
1260 return err;
1261}
1262
1263/*
1264 * Determine the nodes of an array of pages and store it in an array of status.
1265 */
1266static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1267 const void __user **pages, int *status)
1268{
1269 unsigned long i;
1270
1271 down_read(&mm->mmap_sem);
1272
1273 for (i = 0; i < nr_pages; i++) {
1274 unsigned long addr = (unsigned long)(*pages);
1275 struct vm_area_struct *vma;
1276 struct page *page;
1277 int err = -EFAULT;
1278
1279 vma = find_vma(mm, addr);
1280 if (!vma || addr < vma->vm_start)
1281 goto set_status;
1282
1283 page = follow_page(vma, addr, 0);
1284
1285 err = PTR_ERR(page);
1286 if (IS_ERR(page))
1287 goto set_status;
1288
1289 err = -ENOENT;
1290 /* Use PageReserved to check for zero page */
1291 if (!page || PageReserved(page) || PageKsm(page))
1292 goto set_status;
1293
1294 err = page_to_nid(page);
1295set_status:
1296 *status = err;
1297
1298 pages++;
1299 status++;
1300 }
1301
1302 up_read(&mm->mmap_sem);
1303}
1304
1305/*
1306 * Determine the nodes of a user array of pages and store it in
1307 * a user array of status.
1308 */
1309static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1310 const void __user * __user *pages,
1311 int __user *status)
1312{
1313#define DO_PAGES_STAT_CHUNK_NR 16
1314 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1315 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1316
1317 while (nr_pages) {
1318 unsigned long chunk_nr;
1319
1320 chunk_nr = nr_pages;
1321 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1322 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1323
1324 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1325 break;
1326
1327 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1328
1329 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1330 break;
1331
1332 pages += chunk_nr;
1333 status += chunk_nr;
1334 nr_pages -= chunk_nr;
1335 }
1336 return nr_pages ? -EFAULT : 0;
1337}
1338
1339/*
1340 * Move a list of pages in the address space of the currently executing
1341 * process.
1342 */
1343SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1344 const void __user * __user *, pages,
1345 const int __user *, nodes,
1346 int __user *, status, int, flags)
1347{
1348 const struct cred *cred = current_cred(), *tcred;
1349 struct task_struct *task;
1350 struct mm_struct *mm;
1351 int err;
1352 nodemask_t task_nodes;
1353
1354 /* Check flags */
1355 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1356 return -EINVAL;
1357
1358 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1359 return -EPERM;
1360
1361 /* Find the mm_struct */
1362 rcu_read_lock();
1363 task = pid ? find_task_by_vpid(pid) : current;
1364 if (!task) {
1365 rcu_read_unlock();
1366 return -ESRCH;
1367 }
1368 get_task_struct(task);
1369
1370 /*
1371 * Check if this process has the right to modify the specified
1372 * process. The right exists if the process has administrative
1373 * capabilities, superuser privileges or the same
1374 * userid as the target process.
1375 */
1376 tcred = __task_cred(task);
1377 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1378 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1379 !capable(CAP_SYS_NICE)) {
1380 rcu_read_unlock();
1381 err = -EPERM;
1382 goto out;
1383 }
1384 rcu_read_unlock();
1385
1386 err = security_task_movememory(task);
1387 if (err)
1388 goto out;
1389
1390 task_nodes = cpuset_mems_allowed(task);
1391 mm = get_task_mm(task);
1392 put_task_struct(task);
1393
1394 if (!mm)
1395 return -EINVAL;
1396
1397 if (nodes)
1398 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1399 nodes, status, flags);
1400 else
1401 err = do_pages_stat(mm, nr_pages, pages, status);
1402
1403 mmput(mm);
1404 return err;
1405
1406out:
1407 put_task_struct(task);
1408 return err;
1409}
1410
1411/*
1412 * Call migration functions in the vma_ops that may prepare
1413 * memory in a vm for migration. migration functions may perform
1414 * the migration for vmas that do not have an underlying page struct.
1415 */
1416int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1417 const nodemask_t *from, unsigned long flags)
1418{
1419 struct vm_area_struct *vma;
1420 int err = 0;
1421
1422 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1423 if (vma->vm_ops && vma->vm_ops->migrate) {
1424 err = vma->vm_ops->migrate(vma, to, from, flags);
1425 if (err)
1426 break;
1427 }
1428 }
1429 return err;
1430}
1431#endif