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// 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/ksm.h>
24#include <linux/rmap.h>
25#include <linux/topology.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
28#include <linux/writeback.h>
29#include <linux/mempolicy.h>
30#include <linux/vmalloc.h>
31#include <linux/security.h>
32#include <linux/backing-dev.h>
33#include <linux/compaction.h>
34#include <linux/syscalls.h>
35#include <linux/compat.h>
36#include <linux/hugetlb.h>
37#include <linux/gfp.h>
38#include <linux/pfn_t.h>
39#include <linux/page_idle.h>
40#include <linux/page_owner.h>
41#include <linux/sched/mm.h>
42#include <linux/ptrace.h>
43#include <linux/memory.h>
44#include <linux/sched/sysctl.h>
45#include <linux/memory-tiers.h>
46#include <linux/pagewalk.h>
47
48#include <asm/tlbflush.h>
49
50#include <trace/events/migrate.h>
51
52#include "internal.h"
53
54bool isolate_movable_page(struct page *page, isolate_mode_t mode)
55{
56 struct folio *folio = folio_get_nontail_page(page);
57 const struct movable_operations *mops;
58
59 /*
60 * Avoid burning cycles with pages that are yet under __free_pages(),
61 * or just got freed under us.
62 *
63 * In case we 'win' a race for a movable page being freed under us and
64 * raise its refcount preventing __free_pages() from doing its job
65 * the put_page() at the end of this block will take care of
66 * release this page, thus avoiding a nasty leakage.
67 */
68 if (!folio)
69 goto out;
70
71 if (unlikely(folio_test_slab(folio)))
72 goto out_putfolio;
73 /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
74 smp_rmb();
75 /*
76 * Check movable flag before taking the page lock because
77 * we use non-atomic bitops on newly allocated page flags so
78 * unconditionally grabbing the lock ruins page's owner side.
79 */
80 if (unlikely(!__folio_test_movable(folio)))
81 goto out_putfolio;
82 /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
83 smp_rmb();
84 if (unlikely(folio_test_slab(folio)))
85 goto out_putfolio;
86
87 /*
88 * As movable pages are not isolated from LRU lists, concurrent
89 * compaction threads can race against page migration functions
90 * as well as race against the releasing a page.
91 *
92 * In order to avoid having an already isolated movable page
93 * being (wrongly) re-isolated while it is under migration,
94 * or to avoid attempting to isolate pages being released,
95 * lets be sure we have the page lock
96 * before proceeding with the movable page isolation steps.
97 */
98 if (unlikely(!folio_trylock(folio)))
99 goto out_putfolio;
100
101 if (!folio_test_movable(folio) || folio_test_isolated(folio))
102 goto out_no_isolated;
103
104 mops = folio_movable_ops(folio);
105 VM_BUG_ON_FOLIO(!mops, folio);
106
107 if (!mops->isolate_page(&folio->page, mode))
108 goto out_no_isolated;
109
110 /* Driver shouldn't use the isolated flag */
111 WARN_ON_ONCE(folio_test_isolated(folio));
112 folio_set_isolated(folio);
113 folio_unlock(folio);
114
115 return true;
116
117out_no_isolated:
118 folio_unlock(folio);
119out_putfolio:
120 folio_put(folio);
121out:
122 return false;
123}
124
125static void putback_movable_folio(struct folio *folio)
126{
127 const struct movable_operations *mops = folio_movable_ops(folio);
128
129 mops->putback_page(&folio->page);
130 folio_clear_isolated(folio);
131}
132
133/*
134 * Put previously isolated pages back onto the appropriate lists
135 * from where they were once taken off for compaction/migration.
136 *
137 * This function shall be used whenever the isolated pageset has been
138 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
139 * and isolate_hugetlb().
140 */
141void putback_movable_pages(struct list_head *l)
142{
143 struct folio *folio;
144 struct folio *folio2;
145
146 list_for_each_entry_safe(folio, folio2, l, lru) {
147 if (unlikely(folio_test_hugetlb(folio))) {
148 folio_putback_active_hugetlb(folio);
149 continue;
150 }
151 list_del(&folio->lru);
152 /*
153 * We isolated non-lru movable folio so here we can use
154 * __folio_test_movable because LRU folio's mapping cannot
155 * have PAGE_MAPPING_MOVABLE.
156 */
157 if (unlikely(__folio_test_movable(folio))) {
158 VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
159 folio_lock(folio);
160 if (folio_test_movable(folio))
161 putback_movable_folio(folio);
162 else
163 folio_clear_isolated(folio);
164 folio_unlock(folio);
165 folio_put(folio);
166 } else {
167 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
168 folio_is_file_lru(folio), -folio_nr_pages(folio));
169 folio_putback_lru(folio);
170 }
171 }
172}
173
174/* Must be called with an elevated refcount on the non-hugetlb folio */
175bool isolate_folio_to_list(struct folio *folio, struct list_head *list)
176{
177 bool isolated, lru;
178
179 if (folio_test_hugetlb(folio))
180 return isolate_hugetlb(folio, list);
181
182 lru = !__folio_test_movable(folio);
183 if (lru)
184 isolated = folio_isolate_lru(folio);
185 else
186 isolated = isolate_movable_page(&folio->page,
187 ISOLATE_UNEVICTABLE);
188
189 if (!isolated)
190 return false;
191
192 list_add(&folio->lru, list);
193 if (lru)
194 node_stat_add_folio(folio, NR_ISOLATED_ANON +
195 folio_is_file_lru(folio));
196
197 return true;
198}
199
200static bool try_to_map_unused_to_zeropage(struct page_vma_mapped_walk *pvmw,
201 struct folio *folio,
202 unsigned long idx)
203{
204 struct page *page = folio_page(folio, idx);
205 bool contains_data;
206 pte_t newpte;
207 void *addr;
208
209 if (PageCompound(page))
210 return false;
211 VM_BUG_ON_PAGE(!PageAnon(page), page);
212 VM_BUG_ON_PAGE(!PageLocked(page), page);
213 VM_BUG_ON_PAGE(pte_present(*pvmw->pte), page);
214
215 if (folio_test_mlocked(folio) || (pvmw->vma->vm_flags & VM_LOCKED) ||
216 mm_forbids_zeropage(pvmw->vma->vm_mm))
217 return false;
218
219 /*
220 * The pmd entry mapping the old thp was flushed and the pte mapping
221 * this subpage has been non present. If the subpage is only zero-filled
222 * then map it to the shared zeropage.
223 */
224 addr = kmap_local_page(page);
225 contains_data = memchr_inv(addr, 0, PAGE_SIZE);
226 kunmap_local(addr);
227
228 if (contains_data)
229 return false;
230
231 newpte = pte_mkspecial(pfn_pte(my_zero_pfn(pvmw->address),
232 pvmw->vma->vm_page_prot));
233 set_pte_at(pvmw->vma->vm_mm, pvmw->address, pvmw->pte, newpte);
234
235 dec_mm_counter(pvmw->vma->vm_mm, mm_counter(folio));
236 return true;
237}
238
239struct rmap_walk_arg {
240 struct folio *folio;
241 bool map_unused_to_zeropage;
242};
243
244/*
245 * Restore a potential migration pte to a working pte entry
246 */
247static bool remove_migration_pte(struct folio *folio,
248 struct vm_area_struct *vma, unsigned long addr, void *arg)
249{
250 struct rmap_walk_arg *rmap_walk_arg = arg;
251 DEFINE_FOLIO_VMA_WALK(pvmw, rmap_walk_arg->folio, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
252
253 while (page_vma_mapped_walk(&pvmw)) {
254 rmap_t rmap_flags = RMAP_NONE;
255 pte_t old_pte;
256 pte_t pte;
257 swp_entry_t entry;
258 struct page *new;
259 unsigned long idx = 0;
260
261 /* pgoff is invalid for ksm pages, but they are never large */
262 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
263 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
264 new = folio_page(folio, idx);
265
266#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
267 /* PMD-mapped THP migration entry */
268 if (!pvmw.pte) {
269 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
270 !folio_test_pmd_mappable(folio), folio);
271 remove_migration_pmd(&pvmw, new);
272 continue;
273 }
274#endif
275 if (rmap_walk_arg->map_unused_to_zeropage &&
276 try_to_map_unused_to_zeropage(&pvmw, folio, idx))
277 continue;
278
279 folio_get(folio);
280 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
281 old_pte = ptep_get(pvmw.pte);
282
283 entry = pte_to_swp_entry(old_pte);
284 if (!is_migration_entry_young(entry))
285 pte = pte_mkold(pte);
286 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
287 pte = pte_mkdirty(pte);
288 if (pte_swp_soft_dirty(old_pte))
289 pte = pte_mksoft_dirty(pte);
290 else
291 pte = pte_clear_soft_dirty(pte);
292
293 if (is_writable_migration_entry(entry))
294 pte = pte_mkwrite(pte, vma);
295 else if (pte_swp_uffd_wp(old_pte))
296 pte = pte_mkuffd_wp(pte);
297
298 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
299 rmap_flags |= RMAP_EXCLUSIVE;
300
301 if (unlikely(is_device_private_page(new))) {
302 if (pte_write(pte))
303 entry = make_writable_device_private_entry(
304 page_to_pfn(new));
305 else
306 entry = make_readable_device_private_entry(
307 page_to_pfn(new));
308 pte = swp_entry_to_pte(entry);
309 if (pte_swp_soft_dirty(old_pte))
310 pte = pte_swp_mksoft_dirty(pte);
311 if (pte_swp_uffd_wp(old_pte))
312 pte = pte_swp_mkuffd_wp(pte);
313 }
314
315#ifdef CONFIG_HUGETLB_PAGE
316 if (folio_test_hugetlb(folio)) {
317 struct hstate *h = hstate_vma(vma);
318 unsigned int shift = huge_page_shift(h);
319 unsigned long psize = huge_page_size(h);
320
321 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
322 if (folio_test_anon(folio))
323 hugetlb_add_anon_rmap(folio, vma, pvmw.address,
324 rmap_flags);
325 else
326 hugetlb_add_file_rmap(folio);
327 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
328 psize);
329 } else
330#endif
331 {
332 if (folio_test_anon(folio))
333 folio_add_anon_rmap_pte(folio, new, vma,
334 pvmw.address, rmap_flags);
335 else
336 folio_add_file_rmap_pte(folio, new, vma);
337 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
338 }
339 if (vma->vm_flags & VM_LOCKED)
340 mlock_drain_local();
341
342 trace_remove_migration_pte(pvmw.address, pte_val(pte),
343 compound_order(new));
344
345 /* No need to invalidate - it was non-present before */
346 update_mmu_cache(vma, pvmw.address, pvmw.pte);
347 }
348
349 return true;
350}
351
352/*
353 * Get rid of all migration entries and replace them by
354 * references to the indicated page.
355 */
356void remove_migration_ptes(struct folio *src, struct folio *dst, int flags)
357{
358 struct rmap_walk_arg rmap_walk_arg = {
359 .folio = src,
360 .map_unused_to_zeropage = flags & RMP_USE_SHARED_ZEROPAGE,
361 };
362
363 struct rmap_walk_control rwc = {
364 .rmap_one = remove_migration_pte,
365 .arg = &rmap_walk_arg,
366 };
367
368 VM_BUG_ON_FOLIO((flags & RMP_USE_SHARED_ZEROPAGE) && (src != dst), src);
369
370 if (flags & RMP_LOCKED)
371 rmap_walk_locked(dst, &rwc);
372 else
373 rmap_walk(dst, &rwc);
374}
375
376/*
377 * Something used the pte of a page under migration. We need to
378 * get to the page and wait until migration is finished.
379 * When we return from this function the fault will be retried.
380 */
381void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
382 unsigned long address)
383{
384 spinlock_t *ptl;
385 pte_t *ptep;
386 pte_t pte;
387 swp_entry_t entry;
388
389 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
390 if (!ptep)
391 return;
392
393 pte = ptep_get(ptep);
394 pte_unmap(ptep);
395
396 if (!is_swap_pte(pte))
397 goto out;
398
399 entry = pte_to_swp_entry(pte);
400 if (!is_migration_entry(entry))
401 goto out;
402
403 migration_entry_wait_on_locked(entry, ptl);
404 return;
405out:
406 spin_unlock(ptl);
407}
408
409#ifdef CONFIG_HUGETLB_PAGE
410/*
411 * The vma read lock must be held upon entry. Holding that lock prevents either
412 * the pte or the ptl from being freed.
413 *
414 * This function will release the vma lock before returning.
415 */
416void migration_entry_wait_huge(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep)
417{
418 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
419 pte_t pte;
420
421 hugetlb_vma_assert_locked(vma);
422 spin_lock(ptl);
423 pte = huge_ptep_get(vma->vm_mm, addr, ptep);
424
425 if (unlikely(!is_hugetlb_entry_migration(pte))) {
426 spin_unlock(ptl);
427 hugetlb_vma_unlock_read(vma);
428 } else {
429 /*
430 * If migration entry existed, safe to release vma lock
431 * here because the pgtable page won't be freed without the
432 * pgtable lock released. See comment right above pgtable
433 * lock release in migration_entry_wait_on_locked().
434 */
435 hugetlb_vma_unlock_read(vma);
436 migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
437 }
438}
439#endif
440
441#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
442void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
443{
444 spinlock_t *ptl;
445
446 ptl = pmd_lock(mm, pmd);
447 if (!is_pmd_migration_entry(*pmd))
448 goto unlock;
449 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
450 return;
451unlock:
452 spin_unlock(ptl);
453}
454#endif
455
456static int folio_expected_refs(struct address_space *mapping,
457 struct folio *folio)
458{
459 int refs = 1;
460 if (!mapping)
461 return refs;
462
463 refs += folio_nr_pages(folio);
464 if (folio_test_private(folio))
465 refs++;
466
467 return refs;
468}
469
470/*
471 * Replace the folio in the mapping.
472 *
473 * The number of remaining references must be:
474 * 1 for anonymous folios without a mapping
475 * 2 for folios with a mapping
476 * 3 for folios with a mapping and the private flag set.
477 */
478static int __folio_migrate_mapping(struct address_space *mapping,
479 struct folio *newfolio, struct folio *folio, int expected_count)
480{
481 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
482 struct zone *oldzone, *newzone;
483 int dirty;
484 long nr = folio_nr_pages(folio);
485 long entries, i;
486
487 if (!mapping) {
488 /* Take off deferred split queue while frozen and memcg set */
489 if (folio_test_large(folio) &&
490 folio_test_large_rmappable(folio)) {
491 if (!folio_ref_freeze(folio, expected_count))
492 return -EAGAIN;
493 folio_unqueue_deferred_split(folio);
494 folio_ref_unfreeze(folio, expected_count);
495 }
496
497 /* No turning back from here */
498 newfolio->index = folio->index;
499 newfolio->mapping = folio->mapping;
500 if (folio_test_anon(folio) && folio_test_large(folio))
501 mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON, 1);
502 if (folio_test_swapbacked(folio))
503 __folio_set_swapbacked(newfolio);
504
505 return MIGRATEPAGE_SUCCESS;
506 }
507
508 oldzone = folio_zone(folio);
509 newzone = folio_zone(newfolio);
510
511 xas_lock_irq(&xas);
512 if (!folio_ref_freeze(folio, expected_count)) {
513 xas_unlock_irq(&xas);
514 return -EAGAIN;
515 }
516
517 /* Take off deferred split queue while frozen and memcg set */
518 folio_unqueue_deferred_split(folio);
519
520 /*
521 * Now we know that no one else is looking at the folio:
522 * no turning back from here.
523 */
524 newfolio->index = folio->index;
525 newfolio->mapping = folio->mapping;
526 if (folio_test_anon(folio) && folio_test_large(folio))
527 mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON, 1);
528 folio_ref_add(newfolio, nr); /* add cache reference */
529 if (folio_test_swapbacked(folio)) {
530 __folio_set_swapbacked(newfolio);
531 if (folio_test_swapcache(folio)) {
532 folio_set_swapcache(newfolio);
533 newfolio->private = folio_get_private(folio);
534 }
535 entries = nr;
536 } else {
537 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
538 entries = 1;
539 }
540
541 /* Move dirty while folio refs frozen and newfolio not yet exposed */
542 dirty = folio_test_dirty(folio);
543 if (dirty) {
544 folio_clear_dirty(folio);
545 folio_set_dirty(newfolio);
546 }
547
548 /* Swap cache still stores N entries instead of a high-order entry */
549 for (i = 0; i < entries; i++) {
550 xas_store(&xas, newfolio);
551 xas_next(&xas);
552 }
553
554 /*
555 * Drop cache reference from old folio by unfreezing
556 * to one less reference.
557 * We know this isn't the last reference.
558 */
559 folio_ref_unfreeze(folio, expected_count - nr);
560
561 xas_unlock(&xas);
562 /* Leave irq disabled to prevent preemption while updating stats */
563
564 /*
565 * If moved to a different zone then also account
566 * the folio for that zone. Other VM counters will be
567 * taken care of when we establish references to the
568 * new folio and drop references to the old folio.
569 *
570 * Note that anonymous folios are accounted for
571 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
572 * are mapped to swap space.
573 */
574 if (newzone != oldzone) {
575 struct lruvec *old_lruvec, *new_lruvec;
576 struct mem_cgroup *memcg;
577
578 memcg = folio_memcg(folio);
579 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
580 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
581
582 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
583 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
584 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
585 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
586 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
587
588 if (folio_test_pmd_mappable(folio)) {
589 __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
590 __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
591 }
592 }
593#ifdef CONFIG_SWAP
594 if (folio_test_swapcache(folio)) {
595 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
596 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
597 }
598#endif
599 if (dirty && mapping_can_writeback(mapping)) {
600 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
601 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
602 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
603 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
604 }
605 }
606 local_irq_enable();
607
608 return MIGRATEPAGE_SUCCESS;
609}
610
611int folio_migrate_mapping(struct address_space *mapping,
612 struct folio *newfolio, struct folio *folio, int extra_count)
613{
614 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
615
616 if (folio_ref_count(folio) != expected_count)
617 return -EAGAIN;
618
619 return __folio_migrate_mapping(mapping, newfolio, folio, expected_count);
620}
621EXPORT_SYMBOL(folio_migrate_mapping);
622
623/*
624 * The expected number of remaining references is the same as that
625 * of folio_migrate_mapping().
626 */
627int migrate_huge_page_move_mapping(struct address_space *mapping,
628 struct folio *dst, struct folio *src)
629{
630 XA_STATE(xas, &mapping->i_pages, folio_index(src));
631 int rc, expected_count = folio_expected_refs(mapping, src);
632
633 if (folio_ref_count(src) != expected_count)
634 return -EAGAIN;
635
636 rc = folio_mc_copy(dst, src);
637 if (unlikely(rc))
638 return rc;
639
640 xas_lock_irq(&xas);
641 if (!folio_ref_freeze(src, expected_count)) {
642 xas_unlock_irq(&xas);
643 return -EAGAIN;
644 }
645
646 dst->index = src->index;
647 dst->mapping = src->mapping;
648
649 folio_ref_add(dst, folio_nr_pages(dst));
650
651 xas_store(&xas, dst);
652
653 folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
654
655 xas_unlock_irq(&xas);
656
657 return MIGRATEPAGE_SUCCESS;
658}
659
660/*
661 * Copy the flags and some other ancillary information
662 */
663void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
664{
665 int cpupid;
666
667 if (folio_test_referenced(folio))
668 folio_set_referenced(newfolio);
669 if (folio_test_uptodate(folio))
670 folio_mark_uptodate(newfolio);
671 if (folio_test_clear_active(folio)) {
672 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
673 folio_set_active(newfolio);
674 } else if (folio_test_clear_unevictable(folio))
675 folio_set_unevictable(newfolio);
676 if (folio_test_workingset(folio))
677 folio_set_workingset(newfolio);
678 if (folio_test_checked(folio))
679 folio_set_checked(newfolio);
680 /*
681 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
682 * migration entries. We can still have PG_anon_exclusive set on an
683 * effectively unmapped and unreferenced first sub-pages of an
684 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
685 */
686 if (folio_test_mappedtodisk(folio))
687 folio_set_mappedtodisk(newfolio);
688
689 /* Move dirty on pages not done by folio_migrate_mapping() */
690 if (folio_test_dirty(folio))
691 folio_set_dirty(newfolio);
692
693 if (folio_test_young(folio))
694 folio_set_young(newfolio);
695 if (folio_test_idle(folio))
696 folio_set_idle(newfolio);
697
698 folio_migrate_refs(newfolio, folio);
699 /*
700 * Copy NUMA information to the new page, to prevent over-eager
701 * future migrations of this same page.
702 */
703 cpupid = folio_xchg_last_cpupid(folio, -1);
704 /*
705 * For memory tiering mode, when migrate between slow and fast
706 * memory node, reset cpupid, because that is used to record
707 * page access time in slow memory node.
708 */
709 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
710 bool f_toptier = node_is_toptier(folio_nid(folio));
711 bool t_toptier = node_is_toptier(folio_nid(newfolio));
712
713 if (f_toptier != t_toptier)
714 cpupid = -1;
715 }
716 folio_xchg_last_cpupid(newfolio, cpupid);
717
718 folio_migrate_ksm(newfolio, folio);
719 /*
720 * Please do not reorder this without considering how mm/ksm.c's
721 * ksm_get_folio() depends upon ksm_migrate_page() and the
722 * swapcache flag.
723 */
724 if (folio_test_swapcache(folio))
725 folio_clear_swapcache(folio);
726 folio_clear_private(folio);
727
728 /* page->private contains hugetlb specific flags */
729 if (!folio_test_hugetlb(folio))
730 folio->private = NULL;
731
732 /*
733 * If any waiters have accumulated on the new page then
734 * wake them up.
735 */
736 if (folio_test_writeback(newfolio))
737 folio_end_writeback(newfolio);
738
739 /*
740 * PG_readahead shares the same bit with PG_reclaim. The above
741 * end_page_writeback() may clear PG_readahead mistakenly, so set the
742 * bit after that.
743 */
744 if (folio_test_readahead(folio))
745 folio_set_readahead(newfolio);
746
747 folio_copy_owner(newfolio, folio);
748 pgalloc_tag_swap(newfolio, folio);
749
750 mem_cgroup_migrate(folio, newfolio);
751}
752EXPORT_SYMBOL(folio_migrate_flags);
753
754/************************************************************
755 * Migration functions
756 ***********************************************************/
757
758static int __migrate_folio(struct address_space *mapping, struct folio *dst,
759 struct folio *src, void *src_private,
760 enum migrate_mode mode)
761{
762 int rc, expected_count = folio_expected_refs(mapping, src);
763
764 /* Check whether src does not have extra refs before we do more work */
765 if (folio_ref_count(src) != expected_count)
766 return -EAGAIN;
767
768 rc = folio_mc_copy(dst, src);
769 if (unlikely(rc))
770 return rc;
771
772 rc = __folio_migrate_mapping(mapping, dst, src, expected_count);
773 if (rc != MIGRATEPAGE_SUCCESS)
774 return rc;
775
776 if (src_private)
777 folio_attach_private(dst, folio_detach_private(src));
778
779 folio_migrate_flags(dst, src);
780 return MIGRATEPAGE_SUCCESS;
781}
782
783/**
784 * migrate_folio() - Simple folio migration.
785 * @mapping: The address_space containing the folio.
786 * @dst: The folio to migrate the data to.
787 * @src: The folio containing the current data.
788 * @mode: How to migrate the page.
789 *
790 * Common logic to directly migrate a single LRU folio suitable for
791 * folios that do not have private data.
792 *
793 * Folios are locked upon entry and exit.
794 */
795int migrate_folio(struct address_space *mapping, struct folio *dst,
796 struct folio *src, enum migrate_mode mode)
797{
798 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
799 return __migrate_folio(mapping, dst, src, NULL, mode);
800}
801EXPORT_SYMBOL(migrate_folio);
802
803#ifdef CONFIG_BUFFER_HEAD
804/* Returns true if all buffers are successfully locked */
805static bool buffer_migrate_lock_buffers(struct buffer_head *head,
806 enum migrate_mode mode)
807{
808 struct buffer_head *bh = head;
809 struct buffer_head *failed_bh;
810
811 do {
812 if (!trylock_buffer(bh)) {
813 if (mode == MIGRATE_ASYNC)
814 goto unlock;
815 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
816 goto unlock;
817 lock_buffer(bh);
818 }
819
820 bh = bh->b_this_page;
821 } while (bh != head);
822
823 return true;
824
825unlock:
826 /* We failed to lock the buffer and cannot stall. */
827 failed_bh = bh;
828 bh = head;
829 while (bh != failed_bh) {
830 unlock_buffer(bh);
831 bh = bh->b_this_page;
832 }
833
834 return false;
835}
836
837static int __buffer_migrate_folio(struct address_space *mapping,
838 struct folio *dst, struct folio *src, enum migrate_mode mode,
839 bool check_refs)
840{
841 struct buffer_head *bh, *head;
842 int rc;
843 int expected_count;
844
845 head = folio_buffers(src);
846 if (!head)
847 return migrate_folio(mapping, dst, src, mode);
848
849 /* Check whether page does not have extra refs before we do more work */
850 expected_count = folio_expected_refs(mapping, src);
851 if (folio_ref_count(src) != expected_count)
852 return -EAGAIN;
853
854 if (!buffer_migrate_lock_buffers(head, mode))
855 return -EAGAIN;
856
857 if (check_refs) {
858 bool busy;
859 bool invalidated = false;
860
861recheck_buffers:
862 busy = false;
863 spin_lock(&mapping->i_private_lock);
864 bh = head;
865 do {
866 if (atomic_read(&bh->b_count)) {
867 busy = true;
868 break;
869 }
870 bh = bh->b_this_page;
871 } while (bh != head);
872 if (busy) {
873 if (invalidated) {
874 rc = -EAGAIN;
875 goto unlock_buffers;
876 }
877 spin_unlock(&mapping->i_private_lock);
878 invalidate_bh_lrus();
879 invalidated = true;
880 goto recheck_buffers;
881 }
882 }
883
884 rc = filemap_migrate_folio(mapping, dst, src, mode);
885 if (rc != MIGRATEPAGE_SUCCESS)
886 goto unlock_buffers;
887
888 bh = head;
889 do {
890 folio_set_bh(bh, dst, bh_offset(bh));
891 bh = bh->b_this_page;
892 } while (bh != head);
893
894unlock_buffers:
895 if (check_refs)
896 spin_unlock(&mapping->i_private_lock);
897 bh = head;
898 do {
899 unlock_buffer(bh);
900 bh = bh->b_this_page;
901 } while (bh != head);
902
903 return rc;
904}
905
906/**
907 * buffer_migrate_folio() - Migration function for folios with buffers.
908 * @mapping: The address space containing @src.
909 * @dst: The folio to migrate to.
910 * @src: The folio to migrate from.
911 * @mode: How to migrate the folio.
912 *
913 * This function can only be used if the underlying filesystem guarantees
914 * that no other references to @src exist. For example attached buffer
915 * heads are accessed only under the folio lock. If your filesystem cannot
916 * provide this guarantee, buffer_migrate_folio_norefs() may be more
917 * appropriate.
918 *
919 * Return: 0 on success or a negative errno on failure.
920 */
921int buffer_migrate_folio(struct address_space *mapping,
922 struct folio *dst, struct folio *src, enum migrate_mode mode)
923{
924 return __buffer_migrate_folio(mapping, dst, src, mode, false);
925}
926EXPORT_SYMBOL(buffer_migrate_folio);
927
928/**
929 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
930 * @mapping: The address space containing @src.
931 * @dst: The folio to migrate to.
932 * @src: The folio to migrate from.
933 * @mode: How to migrate the folio.
934 *
935 * Like buffer_migrate_folio() except that this variant is more careful
936 * and checks that there are also no buffer head references. This function
937 * is the right one for mappings where buffer heads are directly looked
938 * up and referenced (such as block device mappings).
939 *
940 * Return: 0 on success or a negative errno on failure.
941 */
942int buffer_migrate_folio_norefs(struct address_space *mapping,
943 struct folio *dst, struct folio *src, enum migrate_mode mode)
944{
945 return __buffer_migrate_folio(mapping, dst, src, mode, true);
946}
947EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
948#endif /* CONFIG_BUFFER_HEAD */
949
950int filemap_migrate_folio(struct address_space *mapping,
951 struct folio *dst, struct folio *src, enum migrate_mode mode)
952{
953 return __migrate_folio(mapping, dst, src, folio_get_private(src), mode);
954}
955EXPORT_SYMBOL_GPL(filemap_migrate_folio);
956
957/*
958 * Writeback a folio to clean the dirty state
959 */
960static int writeout(struct address_space *mapping, struct folio *folio)
961{
962 struct writeback_control wbc = {
963 .sync_mode = WB_SYNC_NONE,
964 .nr_to_write = 1,
965 .range_start = 0,
966 .range_end = LLONG_MAX,
967 .for_reclaim = 1
968 };
969 int rc;
970
971 if (!mapping->a_ops->writepage)
972 /* No write method for the address space */
973 return -EINVAL;
974
975 if (!folio_clear_dirty_for_io(folio))
976 /* Someone else already triggered a write */
977 return -EAGAIN;
978
979 /*
980 * A dirty folio may imply that the underlying filesystem has
981 * the folio on some queue. So the folio must be clean for
982 * migration. Writeout may mean we lose the lock and the
983 * folio state is no longer what we checked for earlier.
984 * At this point we know that the migration attempt cannot
985 * be successful.
986 */
987 remove_migration_ptes(folio, folio, 0);
988
989 rc = mapping->a_ops->writepage(&folio->page, &wbc);
990
991 if (rc != AOP_WRITEPAGE_ACTIVATE)
992 /* unlocked. Relock */
993 folio_lock(folio);
994
995 return (rc < 0) ? -EIO : -EAGAIN;
996}
997
998/*
999 * Default handling if a filesystem does not provide a migration function.
1000 */
1001static int fallback_migrate_folio(struct address_space *mapping,
1002 struct folio *dst, struct folio *src, enum migrate_mode mode)
1003{
1004 if (folio_test_dirty(src)) {
1005 /* Only writeback folios in full synchronous migration */
1006 switch (mode) {
1007 case MIGRATE_SYNC:
1008 break;
1009 default:
1010 return -EBUSY;
1011 }
1012 return writeout(mapping, src);
1013 }
1014
1015 /*
1016 * Buffers may be managed in a filesystem specific way.
1017 * We must have no buffers or drop them.
1018 */
1019 if (!filemap_release_folio(src, GFP_KERNEL))
1020 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
1021
1022 return migrate_folio(mapping, dst, src, mode);
1023}
1024
1025/*
1026 * Move a page to a newly allocated page
1027 * The page is locked and all ptes have been successfully removed.
1028 *
1029 * The new page will have replaced the old page if this function
1030 * is successful.
1031 *
1032 * Return value:
1033 * < 0 - error code
1034 * MIGRATEPAGE_SUCCESS - success
1035 */
1036static int move_to_new_folio(struct folio *dst, struct folio *src,
1037 enum migrate_mode mode)
1038{
1039 int rc = -EAGAIN;
1040 bool is_lru = !__folio_test_movable(src);
1041
1042 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
1043 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
1044
1045 if (likely(is_lru)) {
1046 struct address_space *mapping = folio_mapping(src);
1047
1048 if (!mapping)
1049 rc = migrate_folio(mapping, dst, src, mode);
1050 else if (mapping_inaccessible(mapping))
1051 rc = -EOPNOTSUPP;
1052 else if (mapping->a_ops->migrate_folio)
1053 /*
1054 * Most folios have a mapping and most filesystems
1055 * provide a migrate_folio callback. Anonymous folios
1056 * are part of swap space which also has its own
1057 * migrate_folio callback. This is the most common path
1058 * for page migration.
1059 */
1060 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
1061 mode);
1062 else
1063 rc = fallback_migrate_folio(mapping, dst, src, mode);
1064 } else {
1065 const struct movable_operations *mops;
1066
1067 /*
1068 * In case of non-lru page, it could be released after
1069 * isolation step. In that case, we shouldn't try migration.
1070 */
1071 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1072 if (!folio_test_movable(src)) {
1073 rc = MIGRATEPAGE_SUCCESS;
1074 folio_clear_isolated(src);
1075 goto out;
1076 }
1077
1078 mops = folio_movable_ops(src);
1079 rc = mops->migrate_page(&dst->page, &src->page, mode);
1080 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
1081 !folio_test_isolated(src));
1082 }
1083
1084 /*
1085 * When successful, old pagecache src->mapping must be cleared before
1086 * src is freed; but stats require that PageAnon be left as PageAnon.
1087 */
1088 if (rc == MIGRATEPAGE_SUCCESS) {
1089 if (__folio_test_movable(src)) {
1090 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1091
1092 /*
1093 * We clear PG_movable under page_lock so any compactor
1094 * cannot try to migrate this page.
1095 */
1096 folio_clear_isolated(src);
1097 }
1098
1099 /*
1100 * Anonymous and movable src->mapping will be cleared by
1101 * free_pages_prepare so don't reset it here for keeping
1102 * the type to work PageAnon, for example.
1103 */
1104 if (!folio_mapping_flags(src))
1105 src->mapping = NULL;
1106
1107 if (likely(!folio_is_zone_device(dst)))
1108 flush_dcache_folio(dst);
1109 }
1110out:
1111 return rc;
1112}
1113
1114/*
1115 * To record some information during migration, we use unused private
1116 * field of struct folio of the newly allocated destination folio.
1117 * This is safe because nobody is using it except us.
1118 */
1119enum {
1120 PAGE_WAS_MAPPED = BIT(0),
1121 PAGE_WAS_MLOCKED = BIT(1),
1122 PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1123};
1124
1125static void __migrate_folio_record(struct folio *dst,
1126 int old_page_state,
1127 struct anon_vma *anon_vma)
1128{
1129 dst->private = (void *)anon_vma + old_page_state;
1130}
1131
1132static void __migrate_folio_extract(struct folio *dst,
1133 int *old_page_state,
1134 struct anon_vma **anon_vmap)
1135{
1136 unsigned long private = (unsigned long)dst->private;
1137
1138 *anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1139 *old_page_state = private & PAGE_OLD_STATES;
1140 dst->private = NULL;
1141}
1142
1143/* Restore the source folio to the original state upon failure */
1144static void migrate_folio_undo_src(struct folio *src,
1145 int page_was_mapped,
1146 struct anon_vma *anon_vma,
1147 bool locked,
1148 struct list_head *ret)
1149{
1150 if (page_was_mapped)
1151 remove_migration_ptes(src, src, 0);
1152 /* Drop an anon_vma reference if we took one */
1153 if (anon_vma)
1154 put_anon_vma(anon_vma);
1155 if (locked)
1156 folio_unlock(src);
1157 if (ret)
1158 list_move_tail(&src->lru, ret);
1159}
1160
1161/* Restore the destination folio to the original state upon failure */
1162static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1163 free_folio_t put_new_folio, unsigned long private)
1164{
1165 if (locked)
1166 folio_unlock(dst);
1167 if (put_new_folio)
1168 put_new_folio(dst, private);
1169 else
1170 folio_put(dst);
1171}
1172
1173/* Cleanup src folio upon migration success */
1174static void migrate_folio_done(struct folio *src,
1175 enum migrate_reason reason)
1176{
1177 /*
1178 * Compaction can migrate also non-LRU pages which are
1179 * not accounted to NR_ISOLATED_*. They can be recognized
1180 * as __folio_test_movable
1181 */
1182 if (likely(!__folio_test_movable(src)) && reason != MR_DEMOTION)
1183 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1184 folio_is_file_lru(src), -folio_nr_pages(src));
1185
1186 if (reason != MR_MEMORY_FAILURE)
1187 /* We release the page in page_handle_poison. */
1188 folio_put(src);
1189}
1190
1191/* Obtain the lock on page, remove all ptes. */
1192static int migrate_folio_unmap(new_folio_t get_new_folio,
1193 free_folio_t put_new_folio, unsigned long private,
1194 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1195 enum migrate_reason reason, struct list_head *ret)
1196{
1197 struct folio *dst;
1198 int rc = -EAGAIN;
1199 int old_page_state = 0;
1200 struct anon_vma *anon_vma = NULL;
1201 bool is_lru = data_race(!__folio_test_movable(src));
1202 bool locked = false;
1203 bool dst_locked = false;
1204
1205 if (folio_ref_count(src) == 1) {
1206 /* Folio was freed from under us. So we are done. */
1207 folio_clear_active(src);
1208 folio_clear_unevictable(src);
1209 /* free_pages_prepare() will clear PG_isolated. */
1210 list_del(&src->lru);
1211 migrate_folio_done(src, reason);
1212 return MIGRATEPAGE_SUCCESS;
1213 }
1214
1215 dst = get_new_folio(src, private);
1216 if (!dst)
1217 return -ENOMEM;
1218 *dstp = dst;
1219
1220 dst->private = NULL;
1221
1222 if (!folio_trylock(src)) {
1223 if (mode == MIGRATE_ASYNC)
1224 goto out;
1225
1226 /*
1227 * It's not safe for direct compaction to call lock_page.
1228 * For example, during page readahead pages are added locked
1229 * to the LRU. Later, when the IO completes the pages are
1230 * marked uptodate and unlocked. However, the queueing
1231 * could be merging multiple pages for one bio (e.g.
1232 * mpage_readahead). If an allocation happens for the
1233 * second or third page, the process can end up locking
1234 * the same page twice and deadlocking. Rather than
1235 * trying to be clever about what pages can be locked,
1236 * avoid the use of lock_page for direct compaction
1237 * altogether.
1238 */
1239 if (current->flags & PF_MEMALLOC)
1240 goto out;
1241
1242 /*
1243 * In "light" mode, we can wait for transient locks (eg
1244 * inserting a page into the page table), but it's not
1245 * worth waiting for I/O.
1246 */
1247 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1248 goto out;
1249
1250 folio_lock(src);
1251 }
1252 locked = true;
1253 if (folio_test_mlocked(src))
1254 old_page_state |= PAGE_WAS_MLOCKED;
1255
1256 if (folio_test_writeback(src)) {
1257 /*
1258 * Only in the case of a full synchronous migration is it
1259 * necessary to wait for PageWriteback. In the async case,
1260 * the retry loop is too short and in the sync-light case,
1261 * the overhead of stalling is too much
1262 */
1263 switch (mode) {
1264 case MIGRATE_SYNC:
1265 break;
1266 default:
1267 rc = -EBUSY;
1268 goto out;
1269 }
1270 folio_wait_writeback(src);
1271 }
1272
1273 /*
1274 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1275 * we cannot notice that anon_vma is freed while we migrate a page.
1276 * This get_anon_vma() delays freeing anon_vma pointer until the end
1277 * of migration. File cache pages are no problem because of page_lock()
1278 * File Caches may use write_page() or lock_page() in migration, then,
1279 * just care Anon page here.
1280 *
1281 * Only folio_get_anon_vma() understands the subtleties of
1282 * getting a hold on an anon_vma from outside one of its mms.
1283 * But if we cannot get anon_vma, then we won't need it anyway,
1284 * because that implies that the anon page is no longer mapped
1285 * (and cannot be remapped so long as we hold the page lock).
1286 */
1287 if (folio_test_anon(src) && !folio_test_ksm(src))
1288 anon_vma = folio_get_anon_vma(src);
1289
1290 /*
1291 * Block others from accessing the new page when we get around to
1292 * establishing additional references. We are usually the only one
1293 * holding a reference to dst at this point. We used to have a BUG
1294 * here if folio_trylock(dst) fails, but would like to allow for
1295 * cases where there might be a race with the previous use of dst.
1296 * This is much like races on refcount of oldpage: just don't BUG().
1297 */
1298 if (unlikely(!folio_trylock(dst)))
1299 goto out;
1300 dst_locked = true;
1301
1302 if (unlikely(!is_lru)) {
1303 __migrate_folio_record(dst, old_page_state, anon_vma);
1304 return MIGRATEPAGE_UNMAP;
1305 }
1306
1307 /*
1308 * Corner case handling:
1309 * 1. When a new swap-cache page is read into, it is added to the LRU
1310 * and treated as swapcache but it has no rmap yet.
1311 * Calling try_to_unmap() against a src->mapping==NULL page will
1312 * trigger a BUG. So handle it here.
1313 * 2. An orphaned page (see truncate_cleanup_page) might have
1314 * fs-private metadata. The page can be picked up due to memory
1315 * offlining. Everywhere else except page reclaim, the page is
1316 * invisible to the vm, so the page can not be migrated. So try to
1317 * free the metadata, so the page can be freed.
1318 */
1319 if (!src->mapping) {
1320 if (folio_test_private(src)) {
1321 try_to_free_buffers(src);
1322 goto out;
1323 }
1324 } else if (folio_mapped(src)) {
1325 /* Establish migration ptes */
1326 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1327 !folio_test_ksm(src) && !anon_vma, src);
1328 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1329 old_page_state |= PAGE_WAS_MAPPED;
1330 }
1331
1332 if (!folio_mapped(src)) {
1333 __migrate_folio_record(dst, old_page_state, anon_vma);
1334 return MIGRATEPAGE_UNMAP;
1335 }
1336
1337out:
1338 /*
1339 * A folio that has not been unmapped will be restored to
1340 * right list unless we want to retry.
1341 */
1342 if (rc == -EAGAIN)
1343 ret = NULL;
1344
1345 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1346 anon_vma, locked, ret);
1347 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1348
1349 return rc;
1350}
1351
1352/* Migrate the folio to the newly allocated folio in dst. */
1353static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1354 struct folio *src, struct folio *dst,
1355 enum migrate_mode mode, enum migrate_reason reason,
1356 struct list_head *ret)
1357{
1358 int rc;
1359 int old_page_state = 0;
1360 struct anon_vma *anon_vma = NULL;
1361 bool is_lru = !__folio_test_movable(src);
1362 struct list_head *prev;
1363
1364 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1365 prev = dst->lru.prev;
1366 list_del(&dst->lru);
1367
1368 rc = move_to_new_folio(dst, src, mode);
1369 if (rc)
1370 goto out;
1371
1372 if (unlikely(!is_lru))
1373 goto out_unlock_both;
1374
1375 /*
1376 * When successful, push dst to LRU immediately: so that if it
1377 * turns out to be an mlocked page, remove_migration_ptes() will
1378 * automatically build up the correct dst->mlock_count for it.
1379 *
1380 * We would like to do something similar for the old page, when
1381 * unsuccessful, and other cases when a page has been temporarily
1382 * isolated from the unevictable LRU: but this case is the easiest.
1383 */
1384 folio_add_lru(dst);
1385 if (old_page_state & PAGE_WAS_MLOCKED)
1386 lru_add_drain();
1387
1388 if (old_page_state & PAGE_WAS_MAPPED)
1389 remove_migration_ptes(src, dst, 0);
1390
1391out_unlock_both:
1392 folio_unlock(dst);
1393 set_page_owner_migrate_reason(&dst->page, reason);
1394 /*
1395 * If migration is successful, decrease refcount of dst,
1396 * which will not free the page because new page owner increased
1397 * refcounter.
1398 */
1399 folio_put(dst);
1400
1401 /*
1402 * A folio that has been migrated has all references removed
1403 * and will be freed.
1404 */
1405 list_del(&src->lru);
1406 /* Drop an anon_vma reference if we took one */
1407 if (anon_vma)
1408 put_anon_vma(anon_vma);
1409 folio_unlock(src);
1410 migrate_folio_done(src, reason);
1411
1412 return rc;
1413out:
1414 /*
1415 * A folio that has not been migrated will be restored to
1416 * right list unless we want to retry.
1417 */
1418 if (rc == -EAGAIN) {
1419 list_add(&dst->lru, prev);
1420 __migrate_folio_record(dst, old_page_state, anon_vma);
1421 return rc;
1422 }
1423
1424 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1425 anon_vma, true, ret);
1426 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1427
1428 return rc;
1429}
1430
1431/*
1432 * Counterpart of unmap_and_move_page() for hugepage migration.
1433 *
1434 * This function doesn't wait the completion of hugepage I/O
1435 * because there is no race between I/O and migration for hugepage.
1436 * Note that currently hugepage I/O occurs only in direct I/O
1437 * where no lock is held and PG_writeback is irrelevant,
1438 * and writeback status of all subpages are counted in the reference
1439 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1440 * under direct I/O, the reference of the head page is 512 and a bit more.)
1441 * This means that when we try to migrate hugepage whose subpages are
1442 * doing direct I/O, some references remain after try_to_unmap() and
1443 * hugepage migration fails without data corruption.
1444 *
1445 * There is also no race when direct I/O is issued on the page under migration,
1446 * because then pte is replaced with migration swap entry and direct I/O code
1447 * will wait in the page fault for migration to complete.
1448 */
1449static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1450 free_folio_t put_new_folio, unsigned long private,
1451 struct folio *src, int force, enum migrate_mode mode,
1452 int reason, struct list_head *ret)
1453{
1454 struct folio *dst;
1455 int rc = -EAGAIN;
1456 int page_was_mapped = 0;
1457 struct anon_vma *anon_vma = NULL;
1458 struct address_space *mapping = NULL;
1459
1460 if (folio_ref_count(src) == 1) {
1461 /* page was freed from under us. So we are done. */
1462 folio_putback_active_hugetlb(src);
1463 return MIGRATEPAGE_SUCCESS;
1464 }
1465
1466 dst = get_new_folio(src, private);
1467 if (!dst)
1468 return -ENOMEM;
1469
1470 if (!folio_trylock(src)) {
1471 if (!force)
1472 goto out;
1473 switch (mode) {
1474 case MIGRATE_SYNC:
1475 break;
1476 default:
1477 goto out;
1478 }
1479 folio_lock(src);
1480 }
1481
1482 /*
1483 * Check for pages which are in the process of being freed. Without
1484 * folio_mapping() set, hugetlbfs specific move page routine will not
1485 * be called and we could leak usage counts for subpools.
1486 */
1487 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1488 rc = -EBUSY;
1489 goto out_unlock;
1490 }
1491
1492 if (folio_test_anon(src))
1493 anon_vma = folio_get_anon_vma(src);
1494
1495 if (unlikely(!folio_trylock(dst)))
1496 goto put_anon;
1497
1498 if (folio_mapped(src)) {
1499 enum ttu_flags ttu = 0;
1500
1501 if (!folio_test_anon(src)) {
1502 /*
1503 * In shared mappings, try_to_unmap could potentially
1504 * call huge_pmd_unshare. Because of this, take
1505 * semaphore in write mode here and set TTU_RMAP_LOCKED
1506 * to let lower levels know we have taken the lock.
1507 */
1508 mapping = hugetlb_folio_mapping_lock_write(src);
1509 if (unlikely(!mapping))
1510 goto unlock_put_anon;
1511
1512 ttu = TTU_RMAP_LOCKED;
1513 }
1514
1515 try_to_migrate(src, ttu);
1516 page_was_mapped = 1;
1517
1518 if (ttu & TTU_RMAP_LOCKED)
1519 i_mmap_unlock_write(mapping);
1520 }
1521
1522 if (!folio_mapped(src))
1523 rc = move_to_new_folio(dst, src, mode);
1524
1525 if (page_was_mapped)
1526 remove_migration_ptes(src,
1527 rc == MIGRATEPAGE_SUCCESS ? dst : src, 0);
1528
1529unlock_put_anon:
1530 folio_unlock(dst);
1531
1532put_anon:
1533 if (anon_vma)
1534 put_anon_vma(anon_vma);
1535
1536 if (rc == MIGRATEPAGE_SUCCESS) {
1537 move_hugetlb_state(src, dst, reason);
1538 put_new_folio = NULL;
1539 }
1540
1541out_unlock:
1542 folio_unlock(src);
1543out:
1544 if (rc == MIGRATEPAGE_SUCCESS)
1545 folio_putback_active_hugetlb(src);
1546 else if (rc != -EAGAIN)
1547 list_move_tail(&src->lru, ret);
1548
1549 /*
1550 * If migration was not successful and there's a freeing callback, use
1551 * it. Otherwise, put_page() will drop the reference grabbed during
1552 * isolation.
1553 */
1554 if (put_new_folio)
1555 put_new_folio(dst, private);
1556 else
1557 folio_putback_active_hugetlb(dst);
1558
1559 return rc;
1560}
1561
1562static inline int try_split_folio(struct folio *folio, struct list_head *split_folios,
1563 enum migrate_mode mode)
1564{
1565 int rc;
1566
1567 if (mode == MIGRATE_ASYNC) {
1568 if (!folio_trylock(folio))
1569 return -EAGAIN;
1570 } else {
1571 folio_lock(folio);
1572 }
1573 rc = split_folio_to_list(folio, split_folios);
1574 folio_unlock(folio);
1575 if (!rc)
1576 list_move_tail(&folio->lru, split_folios);
1577
1578 return rc;
1579}
1580
1581#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1582#define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1583#else
1584#define NR_MAX_BATCHED_MIGRATION 512
1585#endif
1586#define NR_MAX_MIGRATE_PAGES_RETRY 10
1587#define NR_MAX_MIGRATE_ASYNC_RETRY 3
1588#define NR_MAX_MIGRATE_SYNC_RETRY \
1589 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1590
1591struct migrate_pages_stats {
1592 int nr_succeeded; /* Normal and large folios migrated successfully, in
1593 units of base pages */
1594 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1595 units of base pages. Untried folios aren't counted */
1596 int nr_thp_succeeded; /* THP migrated successfully */
1597 int nr_thp_failed; /* THP failed to be migrated */
1598 int nr_thp_split; /* THP split before migrating */
1599 int nr_split; /* Large folio (include THP) split before migrating */
1600};
1601
1602/*
1603 * Returns the number of hugetlb folios that were not migrated, or an error code
1604 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1605 * any more because the list has become empty or no retryable hugetlb folios
1606 * exist any more. It is caller's responsibility to call putback_movable_pages()
1607 * only if ret != 0.
1608 */
1609static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1610 free_folio_t put_new_folio, unsigned long private,
1611 enum migrate_mode mode, int reason,
1612 struct migrate_pages_stats *stats,
1613 struct list_head *ret_folios)
1614{
1615 int retry = 1;
1616 int nr_failed = 0;
1617 int nr_retry_pages = 0;
1618 int pass = 0;
1619 struct folio *folio, *folio2;
1620 int rc, nr_pages;
1621
1622 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1623 retry = 0;
1624 nr_retry_pages = 0;
1625
1626 list_for_each_entry_safe(folio, folio2, from, lru) {
1627 if (!folio_test_hugetlb(folio))
1628 continue;
1629
1630 nr_pages = folio_nr_pages(folio);
1631
1632 cond_resched();
1633
1634 /*
1635 * Migratability of hugepages depends on architectures and
1636 * their size. This check is necessary because some callers
1637 * of hugepage migration like soft offline and memory
1638 * hotremove don't walk through page tables or check whether
1639 * the hugepage is pmd-based or not before kicking migration.
1640 */
1641 if (!hugepage_migration_supported(folio_hstate(folio))) {
1642 nr_failed++;
1643 stats->nr_failed_pages += nr_pages;
1644 list_move_tail(&folio->lru, ret_folios);
1645 continue;
1646 }
1647
1648 rc = unmap_and_move_huge_page(get_new_folio,
1649 put_new_folio, private,
1650 folio, pass > 2, mode,
1651 reason, ret_folios);
1652 /*
1653 * The rules are:
1654 * Success: hugetlb folio will be put back
1655 * -EAGAIN: stay on the from list
1656 * -ENOMEM: stay on the from list
1657 * Other errno: put on ret_folios list
1658 */
1659 switch(rc) {
1660 case -ENOMEM:
1661 /*
1662 * When memory is low, don't bother to try to migrate
1663 * other folios, just exit.
1664 */
1665 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1666 return -ENOMEM;
1667 case -EAGAIN:
1668 retry++;
1669 nr_retry_pages += nr_pages;
1670 break;
1671 case MIGRATEPAGE_SUCCESS:
1672 stats->nr_succeeded += nr_pages;
1673 break;
1674 default:
1675 /*
1676 * Permanent failure (-EBUSY, etc.):
1677 * unlike -EAGAIN case, the failed folio is
1678 * removed from migration folio list and not
1679 * retried in the next outer loop.
1680 */
1681 nr_failed++;
1682 stats->nr_failed_pages += nr_pages;
1683 break;
1684 }
1685 }
1686 }
1687 /*
1688 * nr_failed is number of hugetlb folios failed to be migrated. After
1689 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1690 * folios as failed.
1691 */
1692 nr_failed += retry;
1693 stats->nr_failed_pages += nr_retry_pages;
1694
1695 return nr_failed;
1696}
1697
1698/*
1699 * migrate_pages_batch() first unmaps folios in the from list as many as
1700 * possible, then move the unmapped folios.
1701 *
1702 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1703 * lock or bit when we have locked more than one folio. Which may cause
1704 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1705 * length of the from list must be <= 1.
1706 */
1707static int migrate_pages_batch(struct list_head *from,
1708 new_folio_t get_new_folio, free_folio_t put_new_folio,
1709 unsigned long private, enum migrate_mode mode, int reason,
1710 struct list_head *ret_folios, struct list_head *split_folios,
1711 struct migrate_pages_stats *stats, int nr_pass)
1712{
1713 int retry = 1;
1714 int thp_retry = 1;
1715 int nr_failed = 0;
1716 int nr_retry_pages = 0;
1717 int pass = 0;
1718 bool is_thp = false;
1719 bool is_large = false;
1720 struct folio *folio, *folio2, *dst = NULL, *dst2;
1721 int rc, rc_saved = 0, nr_pages;
1722 LIST_HEAD(unmap_folios);
1723 LIST_HEAD(dst_folios);
1724 bool nosplit = (reason == MR_NUMA_MISPLACED);
1725
1726 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1727 !list_empty(from) && !list_is_singular(from));
1728
1729 for (pass = 0; pass < nr_pass && retry; pass++) {
1730 retry = 0;
1731 thp_retry = 0;
1732 nr_retry_pages = 0;
1733
1734 list_for_each_entry_safe(folio, folio2, from, lru) {
1735 is_large = folio_test_large(folio);
1736 is_thp = folio_test_pmd_mappable(folio);
1737 nr_pages = folio_nr_pages(folio);
1738
1739 cond_resched();
1740
1741 /*
1742 * The rare folio on the deferred split list should
1743 * be split now. It should not count as a failure:
1744 * but increment nr_failed because, without doing so,
1745 * migrate_pages() may report success with (split but
1746 * unmigrated) pages still on its fromlist; whereas it
1747 * always reports success when its fromlist is empty.
1748 * stats->nr_thp_failed should be increased too,
1749 * otherwise stats inconsistency will happen when
1750 * migrate_pages_batch is called via migrate_pages()
1751 * with MIGRATE_SYNC and MIGRATE_ASYNC.
1752 *
1753 * Only check it without removing it from the list.
1754 * Since the folio can be on deferred_split_scan()
1755 * local list and removing it can cause the local list
1756 * corruption. Folio split process below can handle it
1757 * with the help of folio_ref_freeze().
1758 *
1759 * nr_pages > 2 is needed to avoid checking order-1
1760 * page cache folios. They exist, in contrast to
1761 * non-existent order-1 anonymous folios, and do not
1762 * use _deferred_list.
1763 */
1764 if (nr_pages > 2 &&
1765 !list_empty(&folio->_deferred_list) &&
1766 folio_test_partially_mapped(folio)) {
1767 if (!try_split_folio(folio, split_folios, mode)) {
1768 nr_failed++;
1769 stats->nr_thp_failed += is_thp;
1770 stats->nr_thp_split += is_thp;
1771 stats->nr_split++;
1772 continue;
1773 }
1774 }
1775
1776 /*
1777 * Large folio migration might be unsupported or
1778 * the allocation might be failed so we should retry
1779 * on the same folio with the large folio split
1780 * to normal folios.
1781 *
1782 * Split folios are put in split_folios, and
1783 * we will migrate them after the rest of the
1784 * list is processed.
1785 */
1786 if (!thp_migration_supported() && is_thp) {
1787 nr_failed++;
1788 stats->nr_thp_failed++;
1789 if (!try_split_folio(folio, split_folios, mode)) {
1790 stats->nr_thp_split++;
1791 stats->nr_split++;
1792 continue;
1793 }
1794 stats->nr_failed_pages += nr_pages;
1795 list_move_tail(&folio->lru, ret_folios);
1796 continue;
1797 }
1798
1799 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1800 private, folio, &dst, mode, reason,
1801 ret_folios);
1802 /*
1803 * The rules are:
1804 * Success: folio will be freed
1805 * Unmap: folio will be put on unmap_folios list,
1806 * dst folio put on dst_folios list
1807 * -EAGAIN: stay on the from list
1808 * -ENOMEM: stay on the from list
1809 * Other errno: put on ret_folios list
1810 */
1811 switch(rc) {
1812 case -ENOMEM:
1813 /*
1814 * When memory is low, don't bother to try to migrate
1815 * other folios, move unmapped folios, then exit.
1816 */
1817 nr_failed++;
1818 stats->nr_thp_failed += is_thp;
1819 /* Large folio NUMA faulting doesn't split to retry. */
1820 if (is_large && !nosplit) {
1821 int ret = try_split_folio(folio, split_folios, mode);
1822
1823 if (!ret) {
1824 stats->nr_thp_split += is_thp;
1825 stats->nr_split++;
1826 break;
1827 } else if (reason == MR_LONGTERM_PIN &&
1828 ret == -EAGAIN) {
1829 /*
1830 * Try again to split large folio to
1831 * mitigate the failure of longterm pinning.
1832 */
1833 retry++;
1834 thp_retry += is_thp;
1835 nr_retry_pages += nr_pages;
1836 /* Undo duplicated failure counting. */
1837 nr_failed--;
1838 stats->nr_thp_failed -= is_thp;
1839 break;
1840 }
1841 }
1842
1843 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1844 /* nr_failed isn't updated for not used */
1845 stats->nr_thp_failed += thp_retry;
1846 rc_saved = rc;
1847 if (list_empty(&unmap_folios))
1848 goto out;
1849 else
1850 goto move;
1851 case -EAGAIN:
1852 retry++;
1853 thp_retry += is_thp;
1854 nr_retry_pages += nr_pages;
1855 break;
1856 case MIGRATEPAGE_SUCCESS:
1857 stats->nr_succeeded += nr_pages;
1858 stats->nr_thp_succeeded += is_thp;
1859 break;
1860 case MIGRATEPAGE_UNMAP:
1861 list_move_tail(&folio->lru, &unmap_folios);
1862 list_add_tail(&dst->lru, &dst_folios);
1863 break;
1864 default:
1865 /*
1866 * Permanent failure (-EBUSY, etc.):
1867 * unlike -EAGAIN case, the failed folio is
1868 * removed from migration folio list and not
1869 * retried in the next outer loop.
1870 */
1871 nr_failed++;
1872 stats->nr_thp_failed += is_thp;
1873 stats->nr_failed_pages += nr_pages;
1874 break;
1875 }
1876 }
1877 }
1878 nr_failed += retry;
1879 stats->nr_thp_failed += thp_retry;
1880 stats->nr_failed_pages += nr_retry_pages;
1881move:
1882 /* Flush TLBs for all unmapped folios */
1883 try_to_unmap_flush();
1884
1885 retry = 1;
1886 for (pass = 0; pass < nr_pass && retry; pass++) {
1887 retry = 0;
1888 thp_retry = 0;
1889 nr_retry_pages = 0;
1890
1891 dst = list_first_entry(&dst_folios, struct folio, lru);
1892 dst2 = list_next_entry(dst, lru);
1893 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1894 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1895 nr_pages = folio_nr_pages(folio);
1896
1897 cond_resched();
1898
1899 rc = migrate_folio_move(put_new_folio, private,
1900 folio, dst, mode,
1901 reason, ret_folios);
1902 /*
1903 * The rules are:
1904 * Success: folio will be freed
1905 * -EAGAIN: stay on the unmap_folios list
1906 * Other errno: put on ret_folios list
1907 */
1908 switch(rc) {
1909 case -EAGAIN:
1910 retry++;
1911 thp_retry += is_thp;
1912 nr_retry_pages += nr_pages;
1913 break;
1914 case MIGRATEPAGE_SUCCESS:
1915 stats->nr_succeeded += nr_pages;
1916 stats->nr_thp_succeeded += is_thp;
1917 break;
1918 default:
1919 nr_failed++;
1920 stats->nr_thp_failed += is_thp;
1921 stats->nr_failed_pages += nr_pages;
1922 break;
1923 }
1924 dst = dst2;
1925 dst2 = list_next_entry(dst, lru);
1926 }
1927 }
1928 nr_failed += retry;
1929 stats->nr_thp_failed += thp_retry;
1930 stats->nr_failed_pages += nr_retry_pages;
1931
1932 rc = rc_saved ? : nr_failed;
1933out:
1934 /* Cleanup remaining folios */
1935 dst = list_first_entry(&dst_folios, struct folio, lru);
1936 dst2 = list_next_entry(dst, lru);
1937 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1938 int old_page_state = 0;
1939 struct anon_vma *anon_vma = NULL;
1940
1941 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1942 migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1943 anon_vma, true, ret_folios);
1944 list_del(&dst->lru);
1945 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1946 dst = dst2;
1947 dst2 = list_next_entry(dst, lru);
1948 }
1949
1950 return rc;
1951}
1952
1953static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1954 free_folio_t put_new_folio, unsigned long private,
1955 enum migrate_mode mode, int reason,
1956 struct list_head *ret_folios, struct list_head *split_folios,
1957 struct migrate_pages_stats *stats)
1958{
1959 int rc, nr_failed = 0;
1960 LIST_HEAD(folios);
1961 struct migrate_pages_stats astats;
1962
1963 memset(&astats, 0, sizeof(astats));
1964 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1965 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1966 reason, &folios, split_folios, &astats,
1967 NR_MAX_MIGRATE_ASYNC_RETRY);
1968 stats->nr_succeeded += astats.nr_succeeded;
1969 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1970 stats->nr_thp_split += astats.nr_thp_split;
1971 stats->nr_split += astats.nr_split;
1972 if (rc < 0) {
1973 stats->nr_failed_pages += astats.nr_failed_pages;
1974 stats->nr_thp_failed += astats.nr_thp_failed;
1975 list_splice_tail(&folios, ret_folios);
1976 return rc;
1977 }
1978 stats->nr_thp_failed += astats.nr_thp_split;
1979 /*
1980 * Do not count rc, as pages will be retried below.
1981 * Count nr_split only, since it includes nr_thp_split.
1982 */
1983 nr_failed += astats.nr_split;
1984 /*
1985 * Fall back to migrate all failed folios one by one synchronously. All
1986 * failed folios except split THPs will be retried, so their failure
1987 * isn't counted
1988 */
1989 list_splice_tail_init(&folios, from);
1990 while (!list_empty(from)) {
1991 list_move(from->next, &folios);
1992 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1993 private, mode, reason, ret_folios,
1994 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1995 list_splice_tail_init(&folios, ret_folios);
1996 if (rc < 0)
1997 return rc;
1998 nr_failed += rc;
1999 }
2000
2001 return nr_failed;
2002}
2003
2004/*
2005 * migrate_pages - migrate the folios specified in a list, to the free folios
2006 * supplied as the target for the page migration
2007 *
2008 * @from: The list of folios to be migrated.
2009 * @get_new_folio: The function used to allocate free folios to be used
2010 * as the target of the folio migration.
2011 * @put_new_folio: The function used to free target folios if migration
2012 * fails, or NULL if no special handling is necessary.
2013 * @private: Private data to be passed on to get_new_folio()
2014 * @mode: The migration mode that specifies the constraints for
2015 * folio migration, if any.
2016 * @reason: The reason for folio migration.
2017 * @ret_succeeded: Set to the number of folios migrated successfully if
2018 * the caller passes a non-NULL pointer.
2019 *
2020 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
2021 * are movable any more because the list has become empty or no retryable folios
2022 * exist any more. It is caller's responsibility to call putback_movable_pages()
2023 * only if ret != 0.
2024 *
2025 * Returns the number of {normal folio, large folio, hugetlb} that were not
2026 * migrated, or an error code. The number of large folio splits will be
2027 * considered as the number of non-migrated large folio, no matter how many
2028 * split folios of the large folio are migrated successfully.
2029 */
2030int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
2031 free_folio_t put_new_folio, unsigned long private,
2032 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
2033{
2034 int rc, rc_gather;
2035 int nr_pages;
2036 struct folio *folio, *folio2;
2037 LIST_HEAD(folios);
2038 LIST_HEAD(ret_folios);
2039 LIST_HEAD(split_folios);
2040 struct migrate_pages_stats stats;
2041
2042 trace_mm_migrate_pages_start(mode, reason);
2043
2044 memset(&stats, 0, sizeof(stats));
2045
2046 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
2047 mode, reason, &stats, &ret_folios);
2048 if (rc_gather < 0)
2049 goto out;
2050
2051again:
2052 nr_pages = 0;
2053 list_for_each_entry_safe(folio, folio2, from, lru) {
2054 /* Retried hugetlb folios will be kept in list */
2055 if (folio_test_hugetlb(folio)) {
2056 list_move_tail(&folio->lru, &ret_folios);
2057 continue;
2058 }
2059
2060 nr_pages += folio_nr_pages(folio);
2061 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
2062 break;
2063 }
2064 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
2065 list_cut_before(&folios, from, &folio2->lru);
2066 else
2067 list_splice_init(from, &folios);
2068 if (mode == MIGRATE_ASYNC)
2069 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
2070 private, mode, reason, &ret_folios,
2071 &split_folios, &stats,
2072 NR_MAX_MIGRATE_PAGES_RETRY);
2073 else
2074 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
2075 private, mode, reason, &ret_folios,
2076 &split_folios, &stats);
2077 list_splice_tail_init(&folios, &ret_folios);
2078 if (rc < 0) {
2079 rc_gather = rc;
2080 list_splice_tail(&split_folios, &ret_folios);
2081 goto out;
2082 }
2083 if (!list_empty(&split_folios)) {
2084 /*
2085 * Failure isn't counted since all split folios of a large folio
2086 * is counted as 1 failure already. And, we only try to migrate
2087 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
2088 */
2089 migrate_pages_batch(&split_folios, get_new_folio,
2090 put_new_folio, private, MIGRATE_ASYNC, reason,
2091 &ret_folios, NULL, &stats, 1);
2092 list_splice_tail_init(&split_folios, &ret_folios);
2093 }
2094 rc_gather += rc;
2095 if (!list_empty(from))
2096 goto again;
2097out:
2098 /*
2099 * Put the permanent failure folio back to migration list, they
2100 * will be put back to the right list by the caller.
2101 */
2102 list_splice(&ret_folios, from);
2103
2104 /*
2105 * Return 0 in case all split folios of fail-to-migrate large folios
2106 * are migrated successfully.
2107 */
2108 if (list_empty(from))
2109 rc_gather = 0;
2110
2111 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
2112 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
2113 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
2114 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
2115 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
2116 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
2117 stats.nr_thp_succeeded, stats.nr_thp_failed,
2118 stats.nr_thp_split, stats.nr_split, mode,
2119 reason);
2120
2121 if (ret_succeeded)
2122 *ret_succeeded = stats.nr_succeeded;
2123
2124 return rc_gather;
2125}
2126
2127struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2128{
2129 struct migration_target_control *mtc;
2130 gfp_t gfp_mask;
2131 unsigned int order = 0;
2132 int nid;
2133 int zidx;
2134
2135 mtc = (struct migration_target_control *)private;
2136 gfp_mask = mtc->gfp_mask;
2137 nid = mtc->nid;
2138 if (nid == NUMA_NO_NODE)
2139 nid = folio_nid(src);
2140
2141 if (folio_test_hugetlb(src)) {
2142 struct hstate *h = folio_hstate(src);
2143
2144 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2145 return alloc_hugetlb_folio_nodemask(h, nid,
2146 mtc->nmask, gfp_mask,
2147 htlb_allow_alloc_fallback(mtc->reason));
2148 }
2149
2150 if (folio_test_large(src)) {
2151 /*
2152 * clear __GFP_RECLAIM to make the migration callback
2153 * consistent with regular THP allocations.
2154 */
2155 gfp_mask &= ~__GFP_RECLAIM;
2156 gfp_mask |= GFP_TRANSHUGE;
2157 order = folio_order(src);
2158 }
2159 zidx = zone_idx(folio_zone(src));
2160 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2161 gfp_mask |= __GFP_HIGHMEM;
2162
2163 return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2164}
2165
2166#ifdef CONFIG_NUMA
2167
2168static int store_status(int __user *status, int start, int value, int nr)
2169{
2170 while (nr-- > 0) {
2171 if (put_user(value, status + start))
2172 return -EFAULT;
2173 start++;
2174 }
2175
2176 return 0;
2177}
2178
2179static int do_move_pages_to_node(struct list_head *pagelist, int node)
2180{
2181 int err;
2182 struct migration_target_control mtc = {
2183 .nid = node,
2184 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2185 .reason = MR_SYSCALL,
2186 };
2187
2188 err = migrate_pages(pagelist, alloc_migration_target, NULL,
2189 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2190 if (err)
2191 putback_movable_pages(pagelist);
2192 return err;
2193}
2194
2195static int __add_folio_for_migration(struct folio *folio, int node,
2196 struct list_head *pagelist, bool migrate_all)
2197{
2198 if (is_zero_folio(folio) || is_huge_zero_folio(folio))
2199 return -EFAULT;
2200
2201 if (folio_is_zone_device(folio))
2202 return -ENOENT;
2203
2204 if (folio_nid(folio) == node)
2205 return 0;
2206
2207 if (folio_likely_mapped_shared(folio) && !migrate_all)
2208 return -EACCES;
2209
2210 if (folio_test_hugetlb(folio)) {
2211 if (isolate_hugetlb(folio, pagelist))
2212 return 1;
2213 } else if (folio_isolate_lru(folio)) {
2214 list_add_tail(&folio->lru, pagelist);
2215 node_stat_mod_folio(folio,
2216 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2217 folio_nr_pages(folio));
2218 return 1;
2219 }
2220 return -EBUSY;
2221}
2222
2223/*
2224 * Resolves the given address to a struct folio, isolates it from the LRU and
2225 * puts it to the given pagelist.
2226 * Returns:
2227 * errno - if the folio cannot be found/isolated
2228 * 0 - when it doesn't have to be migrated because it is already on the
2229 * target node
2230 * 1 - when it has been queued
2231 */
2232static int add_folio_for_migration(struct mm_struct *mm, const void __user *p,
2233 int node, struct list_head *pagelist, bool migrate_all)
2234{
2235 struct vm_area_struct *vma;
2236 struct folio_walk fw;
2237 struct folio *folio;
2238 unsigned long addr;
2239 int err = -EFAULT;
2240
2241 mmap_read_lock(mm);
2242 addr = (unsigned long)untagged_addr_remote(mm, p);
2243
2244 vma = vma_lookup(mm, addr);
2245 if (vma && vma_migratable(vma)) {
2246 folio = folio_walk_start(&fw, vma, addr, FW_ZEROPAGE);
2247 if (folio) {
2248 err = __add_folio_for_migration(folio, node, pagelist,
2249 migrate_all);
2250 folio_walk_end(&fw, vma);
2251 } else {
2252 err = -ENOENT;
2253 }
2254 }
2255 mmap_read_unlock(mm);
2256 return err;
2257}
2258
2259static int move_pages_and_store_status(int node,
2260 struct list_head *pagelist, int __user *status,
2261 int start, int i, unsigned long nr_pages)
2262{
2263 int err;
2264
2265 if (list_empty(pagelist))
2266 return 0;
2267
2268 err = do_move_pages_to_node(pagelist, node);
2269 if (err) {
2270 /*
2271 * Positive err means the number of failed
2272 * pages to migrate. Since we are going to
2273 * abort and return the number of non-migrated
2274 * pages, so need to include the rest of the
2275 * nr_pages that have not been attempted as
2276 * well.
2277 */
2278 if (err > 0)
2279 err += nr_pages - i;
2280 return err;
2281 }
2282 return store_status(status, start, node, i - start);
2283}
2284
2285/*
2286 * Migrate an array of page address onto an array of nodes and fill
2287 * the corresponding array of status.
2288 */
2289static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2290 unsigned long nr_pages,
2291 const void __user * __user *pages,
2292 const int __user *nodes,
2293 int __user *status, int flags)
2294{
2295 compat_uptr_t __user *compat_pages = (void __user *)pages;
2296 int current_node = NUMA_NO_NODE;
2297 LIST_HEAD(pagelist);
2298 int start, i;
2299 int err = 0, err1;
2300
2301 lru_cache_disable();
2302
2303 for (i = start = 0; i < nr_pages; i++) {
2304 const void __user *p;
2305 int node;
2306
2307 err = -EFAULT;
2308 if (in_compat_syscall()) {
2309 compat_uptr_t cp;
2310
2311 if (get_user(cp, compat_pages + i))
2312 goto out_flush;
2313
2314 p = compat_ptr(cp);
2315 } else {
2316 if (get_user(p, pages + i))
2317 goto out_flush;
2318 }
2319 if (get_user(node, nodes + i))
2320 goto out_flush;
2321
2322 err = -ENODEV;
2323 if (node < 0 || node >= MAX_NUMNODES)
2324 goto out_flush;
2325 if (!node_state(node, N_MEMORY))
2326 goto out_flush;
2327
2328 err = -EACCES;
2329 if (!node_isset(node, task_nodes))
2330 goto out_flush;
2331
2332 if (current_node == NUMA_NO_NODE) {
2333 current_node = node;
2334 start = i;
2335 } else if (node != current_node) {
2336 err = move_pages_and_store_status(current_node,
2337 &pagelist, status, start, i, nr_pages);
2338 if (err)
2339 goto out;
2340 start = i;
2341 current_node = node;
2342 }
2343
2344 /*
2345 * Errors in the page lookup or isolation are not fatal and we simply
2346 * report them via status
2347 */
2348 err = add_folio_for_migration(mm, p, current_node, &pagelist,
2349 flags & MPOL_MF_MOVE_ALL);
2350
2351 if (err > 0) {
2352 /* The page is successfully queued for migration */
2353 continue;
2354 }
2355
2356 /*
2357 * The move_pages() man page does not have an -EEXIST choice, so
2358 * use -EFAULT instead.
2359 */
2360 if (err == -EEXIST)
2361 err = -EFAULT;
2362
2363 /*
2364 * If the page is already on the target node (!err), store the
2365 * node, otherwise, store the err.
2366 */
2367 err = store_status(status, i, err ? : current_node, 1);
2368 if (err)
2369 goto out_flush;
2370
2371 err = move_pages_and_store_status(current_node, &pagelist,
2372 status, start, i, nr_pages);
2373 if (err) {
2374 /* We have accounted for page i */
2375 if (err > 0)
2376 err--;
2377 goto out;
2378 }
2379 current_node = NUMA_NO_NODE;
2380 }
2381out_flush:
2382 /* Make sure we do not overwrite the existing error */
2383 err1 = move_pages_and_store_status(current_node, &pagelist,
2384 status, start, i, nr_pages);
2385 if (err >= 0)
2386 err = err1;
2387out:
2388 lru_cache_enable();
2389 return err;
2390}
2391
2392/*
2393 * Determine the nodes of an array of pages and store it in an array of status.
2394 */
2395static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2396 const void __user **pages, int *status)
2397{
2398 unsigned long i;
2399
2400 mmap_read_lock(mm);
2401
2402 for (i = 0; i < nr_pages; i++) {
2403 unsigned long addr = (unsigned long)(*pages);
2404 struct vm_area_struct *vma;
2405 struct folio_walk fw;
2406 struct folio *folio;
2407 int err = -EFAULT;
2408
2409 vma = vma_lookup(mm, addr);
2410 if (!vma)
2411 goto set_status;
2412
2413 folio = folio_walk_start(&fw, vma, addr, FW_ZEROPAGE);
2414 if (folio) {
2415 if (is_zero_folio(folio) || is_huge_zero_folio(folio))
2416 err = -EFAULT;
2417 else if (folio_is_zone_device(folio))
2418 err = -ENOENT;
2419 else
2420 err = folio_nid(folio);
2421 folio_walk_end(&fw, vma);
2422 } else {
2423 err = -ENOENT;
2424 }
2425set_status:
2426 *status = err;
2427
2428 pages++;
2429 status++;
2430 }
2431
2432 mmap_read_unlock(mm);
2433}
2434
2435static int get_compat_pages_array(const void __user *chunk_pages[],
2436 const void __user * __user *pages,
2437 unsigned long chunk_nr)
2438{
2439 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2440 compat_uptr_t p;
2441 int i;
2442
2443 for (i = 0; i < chunk_nr; i++) {
2444 if (get_user(p, pages32 + i))
2445 return -EFAULT;
2446 chunk_pages[i] = compat_ptr(p);
2447 }
2448
2449 return 0;
2450}
2451
2452/*
2453 * Determine the nodes of a user array of pages and store it in
2454 * a user array of status.
2455 */
2456static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2457 const void __user * __user *pages,
2458 int __user *status)
2459{
2460#define DO_PAGES_STAT_CHUNK_NR 16UL
2461 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2462 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2463
2464 while (nr_pages) {
2465 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2466
2467 if (in_compat_syscall()) {
2468 if (get_compat_pages_array(chunk_pages, pages,
2469 chunk_nr))
2470 break;
2471 } else {
2472 if (copy_from_user(chunk_pages, pages,
2473 chunk_nr * sizeof(*chunk_pages)))
2474 break;
2475 }
2476
2477 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2478
2479 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2480 break;
2481
2482 pages += chunk_nr;
2483 status += chunk_nr;
2484 nr_pages -= chunk_nr;
2485 }
2486 return nr_pages ? -EFAULT : 0;
2487}
2488
2489static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2490{
2491 struct task_struct *task;
2492 struct mm_struct *mm;
2493
2494 /*
2495 * There is no need to check if current process has the right to modify
2496 * the specified process when they are same.
2497 */
2498 if (!pid) {
2499 mmget(current->mm);
2500 *mem_nodes = cpuset_mems_allowed(current);
2501 return current->mm;
2502 }
2503
2504 task = find_get_task_by_vpid(pid);
2505 if (!task) {
2506 return ERR_PTR(-ESRCH);
2507 }
2508
2509 /*
2510 * Check if this process has the right to modify the specified
2511 * process. Use the regular "ptrace_may_access()" checks.
2512 */
2513 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2514 mm = ERR_PTR(-EPERM);
2515 goto out;
2516 }
2517
2518 mm = ERR_PTR(security_task_movememory(task));
2519 if (IS_ERR(mm))
2520 goto out;
2521 *mem_nodes = cpuset_mems_allowed(task);
2522 mm = get_task_mm(task);
2523out:
2524 put_task_struct(task);
2525 if (!mm)
2526 mm = ERR_PTR(-EINVAL);
2527 return mm;
2528}
2529
2530/*
2531 * Move a list of pages in the address space of the currently executing
2532 * process.
2533 */
2534static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2535 const void __user * __user *pages,
2536 const int __user *nodes,
2537 int __user *status, int flags)
2538{
2539 struct mm_struct *mm;
2540 int err;
2541 nodemask_t task_nodes;
2542
2543 /* Check flags */
2544 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2545 return -EINVAL;
2546
2547 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2548 return -EPERM;
2549
2550 mm = find_mm_struct(pid, &task_nodes);
2551 if (IS_ERR(mm))
2552 return PTR_ERR(mm);
2553
2554 if (nodes)
2555 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2556 nodes, status, flags);
2557 else
2558 err = do_pages_stat(mm, nr_pages, pages, status);
2559
2560 mmput(mm);
2561 return err;
2562}
2563
2564SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2565 const void __user * __user *, pages,
2566 const int __user *, nodes,
2567 int __user *, status, int, flags)
2568{
2569 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2570}
2571
2572#ifdef CONFIG_NUMA_BALANCING
2573/*
2574 * Returns true if this is a safe migration target node for misplaced NUMA
2575 * pages. Currently it only checks the watermarks which is crude.
2576 */
2577static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2578 unsigned long nr_migrate_pages)
2579{
2580 int z;
2581
2582 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2583 struct zone *zone = pgdat->node_zones + z;
2584
2585 if (!managed_zone(zone))
2586 continue;
2587
2588 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2589 if (!zone_watermark_ok(zone, 0,
2590 high_wmark_pages(zone) +
2591 nr_migrate_pages,
2592 ZONE_MOVABLE, ALLOC_CMA))
2593 continue;
2594 return true;
2595 }
2596 return false;
2597}
2598
2599static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2600 unsigned long data)
2601{
2602 int nid = (int) data;
2603 int order = folio_order(src);
2604 gfp_t gfp = __GFP_THISNODE;
2605
2606 if (order > 0)
2607 gfp |= GFP_TRANSHUGE_LIGHT;
2608 else {
2609 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2610 __GFP_NOWARN;
2611 gfp &= ~__GFP_RECLAIM;
2612 }
2613 return __folio_alloc_node(gfp, order, nid);
2614}
2615
2616/*
2617 * Prepare for calling migrate_misplaced_folio() by isolating the folio if
2618 * permitted. Must be called with the PTL still held.
2619 */
2620int migrate_misplaced_folio_prepare(struct folio *folio,
2621 struct vm_area_struct *vma, int node)
2622{
2623 int nr_pages = folio_nr_pages(folio);
2624 pg_data_t *pgdat = NODE_DATA(node);
2625
2626 if (folio_is_file_lru(folio)) {
2627 /*
2628 * Do not migrate file folios that are mapped in multiple
2629 * processes with execute permissions as they are probably
2630 * shared libraries.
2631 *
2632 * See folio_likely_mapped_shared() on possible imprecision
2633 * when we cannot easily detect if a folio is shared.
2634 */
2635 if ((vma->vm_flags & VM_EXEC) &&
2636 folio_likely_mapped_shared(folio))
2637 return -EACCES;
2638
2639 /*
2640 * Do not migrate dirty folios as not all filesystems can move
2641 * dirty folios in MIGRATE_ASYNC mode which is a waste of
2642 * cycles.
2643 */
2644 if (folio_test_dirty(folio))
2645 return -EAGAIN;
2646 }
2647
2648 /* Avoid migrating to a node that is nearly full */
2649 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2650 int z;
2651
2652 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2653 return -EAGAIN;
2654 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2655 if (managed_zone(pgdat->node_zones + z))
2656 break;
2657 }
2658
2659 /*
2660 * If there are no managed zones, it should not proceed
2661 * further.
2662 */
2663 if (z < 0)
2664 return -EAGAIN;
2665
2666 wakeup_kswapd(pgdat->node_zones + z, 0,
2667 folio_order(folio), ZONE_MOVABLE);
2668 return -EAGAIN;
2669 }
2670
2671 if (!folio_isolate_lru(folio))
2672 return -EAGAIN;
2673
2674 node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2675 nr_pages);
2676 return 0;
2677}
2678
2679/*
2680 * Attempt to migrate a misplaced folio to the specified destination
2681 * node. Caller is expected to have isolated the folio by calling
2682 * migrate_misplaced_folio_prepare(), which will result in an
2683 * elevated reference count on the folio. This function will un-isolate the
2684 * folio, dereferencing the folio before returning.
2685 */
2686int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2687 int node)
2688{
2689 pg_data_t *pgdat = NODE_DATA(node);
2690 int nr_remaining;
2691 unsigned int nr_succeeded;
2692 LIST_HEAD(migratepages);
2693 struct mem_cgroup *memcg = get_mem_cgroup_from_folio(folio);
2694 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2695
2696 list_add(&folio->lru, &migratepages);
2697 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2698 NULL, node, MIGRATE_ASYNC,
2699 MR_NUMA_MISPLACED, &nr_succeeded);
2700 if (nr_remaining && !list_empty(&migratepages))
2701 putback_movable_pages(&migratepages);
2702 if (nr_succeeded) {
2703 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2704 count_memcg_events(memcg, NUMA_PAGE_MIGRATE, nr_succeeded);
2705 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
2706 && !node_is_toptier(folio_nid(folio))
2707 && node_is_toptier(node))
2708 mod_lruvec_state(lruvec, PGPROMOTE_SUCCESS, nr_succeeded);
2709 }
2710 mem_cgroup_put(memcg);
2711 BUG_ON(!list_empty(&migratepages));
2712 return nr_remaining ? -EAGAIN : 0;
2713}
2714#endif /* CONFIG_NUMA_BALANCING */
2715#endif /* CONFIG_NUMA */