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/pfn_t.h>
42#include <linux/memremap.h>
43#include <linux/userfaultfd_k.h>
44#include <linux/balloon_compaction.h>
45#include <linux/mmu_notifier.h>
46#include <linux/page_idle.h>
47#include <linux/page_owner.h>
48#include <linux/sched/mm.h>
49#include <linux/ptrace.h>
50
51#include <asm/tlbflush.h>
52
53#define CREATE_TRACE_POINTS
54#include <trace/events/migrate.h>
55
56#include "internal.h"
57
58/*
59 * migrate_prep() needs to be called before we start compiling a list of pages
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
62 */
63int migrate_prep(void)
64{
65 /*
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
70 */
71 lru_add_drain_all();
72
73 return 0;
74}
75
76/* Do the necessary work of migrate_prep but not if it involves other CPUs */
77int migrate_prep_local(void)
78{
79 lru_add_drain();
80
81 return 0;
82}
83
84int isolate_movable_page(struct page *page, isolate_mode_t mode)
85{
86 struct address_space *mapping;
87
88 /*
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
91 *
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
96 */
97 if (unlikely(!get_page_unless_zero(page)))
98 goto out;
99
100 /*
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
104 */
105 if (unlikely(!__PageMovable(page)))
106 goto out_putpage;
107 /*
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
111 *
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
117 */
118 if (unlikely(!trylock_page(page)))
119 goto out_putpage;
120
121 if (!PageMovable(page) || PageIsolated(page))
122 goto out_no_isolated;
123
124 mapping = page_mapping(page);
125 VM_BUG_ON_PAGE(!mapping, page);
126
127 if (!mapping->a_ops->isolate_page(page, mode))
128 goto out_no_isolated;
129
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page));
132 __SetPageIsolated(page);
133 unlock_page(page);
134
135 return 0;
136
137out_no_isolated:
138 unlock_page(page);
139out_putpage:
140 put_page(page);
141out:
142 return -EBUSY;
143}
144
145/* It should be called on page which is PG_movable */
146void putback_movable_page(struct page *page)
147{
148 struct address_space *mapping;
149
150 VM_BUG_ON_PAGE(!PageLocked(page), page);
151 VM_BUG_ON_PAGE(!PageMovable(page), page);
152 VM_BUG_ON_PAGE(!PageIsolated(page), page);
153
154 mapping = page_mapping(page);
155 mapping->a_ops->putback_page(page);
156 __ClearPageIsolated(page);
157}
158
159/*
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
162 *
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
166 */
167void putback_movable_pages(struct list_head *l)
168{
169 struct page *page;
170 struct page *page2;
171
172 list_for_each_entry_safe(page, page2, l, lru) {
173 if (unlikely(PageHuge(page))) {
174 putback_active_hugepage(page);
175 continue;
176 }
177 list_del(&page->lru);
178 /*
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
182 */
183 if (unlikely(__PageMovable(page))) {
184 VM_BUG_ON_PAGE(!PageIsolated(page), page);
185 lock_page(page);
186 if (PageMovable(page))
187 putback_movable_page(page);
188 else
189 __ClearPageIsolated(page);
190 unlock_page(page);
191 put_page(page);
192 } else {
193 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194 page_is_file_cache(page), -hpage_nr_pages(page));
195 putback_lru_page(page);
196 }
197 }
198}
199
200/*
201 * Restore a potential migration pte to a working pte entry
202 */
203static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
204 unsigned long addr, void *old)
205{
206 struct page_vma_mapped_walk pvmw = {
207 .page = old,
208 .vma = vma,
209 .address = addr,
210 .flags = PVMW_SYNC | PVMW_MIGRATION,
211 };
212 struct page *new;
213 pte_t pte;
214 swp_entry_t entry;
215
216 VM_BUG_ON_PAGE(PageTail(page), page);
217 while (page_vma_mapped_walk(&pvmw)) {
218 if (PageKsm(page))
219 new = page;
220 else
221 new = page - pvmw.page->index +
222 linear_page_index(vma, pvmw.address);
223
224#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
226 if (!pvmw.pte) {
227 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228 remove_migration_pmd(&pvmw, new);
229 continue;
230 }
231#endif
232
233 get_page(new);
234 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235 if (pte_swp_soft_dirty(*pvmw.pte))
236 pte = pte_mksoft_dirty(pte);
237
238 /*
239 * Recheck VMA as permissions can change since migration started
240 */
241 entry = pte_to_swp_entry(*pvmw.pte);
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
244
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry = make_device_private_entry(new, pte_write(pte));
248 pte = swp_entry_to_pte(entry);
249 } else if (is_device_public_page(new)) {
250 pte = pte_mkdevmap(pte);
251 flush_dcache_page(new);
252 }
253 } else
254 flush_dcache_page(new);
255
256#ifdef CONFIG_HUGETLB_PAGE
257 if (PageHuge(new)) {
258 pte = pte_mkhuge(pte);
259 pte = arch_make_huge_pte(pte, vma, new, 0);
260 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
261 if (PageAnon(new))
262 hugepage_add_anon_rmap(new, vma, pvmw.address);
263 else
264 page_dup_rmap(new, true);
265 } else
266#endif
267 {
268 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269
270 if (PageAnon(new))
271 page_add_anon_rmap(new, vma, pvmw.address, false);
272 else
273 page_add_file_rmap(new, false);
274 }
275 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 mlock_vma_page(new);
277
278 /* No need to invalidate - it was non-present before */
279 update_mmu_cache(vma, pvmw.address, pvmw.pte);
280 }
281
282 return true;
283}
284
285/*
286 * Get rid of all migration entries and replace them by
287 * references to the indicated page.
288 */
289void remove_migration_ptes(struct page *old, struct page *new, bool locked)
290{
291 struct rmap_walk_control rwc = {
292 .rmap_one = remove_migration_pte,
293 .arg = old,
294 };
295
296 if (locked)
297 rmap_walk_locked(new, &rwc);
298 else
299 rmap_walk(new, &rwc);
300}
301
302/*
303 * Something used the pte of a page under migration. We need to
304 * get to the page and wait until migration is finished.
305 * When we return from this function the fault will be retried.
306 */
307void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
308 spinlock_t *ptl)
309{
310 pte_t pte;
311 swp_entry_t entry;
312 struct page *page;
313
314 spin_lock(ptl);
315 pte = *ptep;
316 if (!is_swap_pte(pte))
317 goto out;
318
319 entry = pte_to_swp_entry(pte);
320 if (!is_migration_entry(entry))
321 goto out;
322
323 page = migration_entry_to_page(entry);
324
325 /*
326 * Once radix-tree replacement of page migration started, page_count
327 * *must* be zero. And, we don't want to call wait_on_page_locked()
328 * against a page without get_page().
329 * So, we use get_page_unless_zero(), here. Even failed, page fault
330 * will occur again.
331 */
332 if (!get_page_unless_zero(page))
333 goto out;
334 pte_unmap_unlock(ptep, ptl);
335 wait_on_page_locked(page);
336 put_page(page);
337 return;
338out:
339 pte_unmap_unlock(ptep, ptl);
340}
341
342void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
343 unsigned long address)
344{
345 spinlock_t *ptl = pte_lockptr(mm, pmd);
346 pte_t *ptep = pte_offset_map(pmd, address);
347 __migration_entry_wait(mm, ptep, ptl);
348}
349
350void migration_entry_wait_huge(struct vm_area_struct *vma,
351 struct mm_struct *mm, pte_t *pte)
352{
353 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
354 __migration_entry_wait(mm, pte, ptl);
355}
356
357#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
358void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
359{
360 spinlock_t *ptl;
361 struct page *page;
362
363 ptl = pmd_lock(mm, pmd);
364 if (!is_pmd_migration_entry(*pmd))
365 goto unlock;
366 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
367 if (!get_page_unless_zero(page))
368 goto unlock;
369 spin_unlock(ptl);
370 wait_on_page_locked(page);
371 put_page(page);
372 return;
373unlock:
374 spin_unlock(ptl);
375}
376#endif
377
378#ifdef CONFIG_BLOCK
379/* Returns true if all buffers are successfully locked */
380static bool buffer_migrate_lock_buffers(struct buffer_head *head,
381 enum migrate_mode mode)
382{
383 struct buffer_head *bh = head;
384
385 /* Simple case, sync compaction */
386 if (mode != MIGRATE_ASYNC) {
387 do {
388 get_bh(bh);
389 lock_buffer(bh);
390 bh = bh->b_this_page;
391
392 } while (bh != head);
393
394 return true;
395 }
396
397 /* async case, we cannot block on lock_buffer so use trylock_buffer */
398 do {
399 get_bh(bh);
400 if (!trylock_buffer(bh)) {
401 /*
402 * We failed to lock the buffer and cannot stall in
403 * async migration. Release the taken locks
404 */
405 struct buffer_head *failed_bh = bh;
406 put_bh(failed_bh);
407 bh = head;
408 while (bh != failed_bh) {
409 unlock_buffer(bh);
410 put_bh(bh);
411 bh = bh->b_this_page;
412 }
413 return false;
414 }
415
416 bh = bh->b_this_page;
417 } while (bh != head);
418 return true;
419}
420#else
421static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
422 enum migrate_mode mode)
423{
424 return true;
425}
426#endif /* CONFIG_BLOCK */
427
428/*
429 * Replace the page in the mapping.
430 *
431 * The number of remaining references must be:
432 * 1 for anonymous pages without a mapping
433 * 2 for pages with a mapping
434 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
435 */
436int migrate_page_move_mapping(struct address_space *mapping,
437 struct page *newpage, struct page *page,
438 struct buffer_head *head, enum migrate_mode mode,
439 int extra_count)
440{
441 struct zone *oldzone, *newzone;
442 int dirty;
443 int expected_count = 1 + extra_count;
444 void **pslot;
445
446 /*
447 * Device public or private pages have an extra refcount as they are
448 * ZONE_DEVICE pages.
449 */
450 expected_count += is_device_private_page(page);
451 expected_count += is_device_public_page(page);
452
453 if (!mapping) {
454 /* Anonymous page without mapping */
455 if (page_count(page) != expected_count)
456 return -EAGAIN;
457
458 /* No turning back from here */
459 newpage->index = page->index;
460 newpage->mapping = page->mapping;
461 if (PageSwapBacked(page))
462 __SetPageSwapBacked(newpage);
463
464 return MIGRATEPAGE_SUCCESS;
465 }
466
467 oldzone = page_zone(page);
468 newzone = page_zone(newpage);
469
470 xa_lock_irq(&mapping->i_pages);
471
472 pslot = radix_tree_lookup_slot(&mapping->i_pages,
473 page_index(page));
474
475 expected_count += hpage_nr_pages(page) + page_has_private(page);
476 if (page_count(page) != expected_count ||
477 radix_tree_deref_slot_protected(pslot,
478 &mapping->i_pages.xa_lock) != page) {
479 xa_unlock_irq(&mapping->i_pages);
480 return -EAGAIN;
481 }
482
483 if (!page_ref_freeze(page, expected_count)) {
484 xa_unlock_irq(&mapping->i_pages);
485 return -EAGAIN;
486 }
487
488 /*
489 * In the async migration case of moving a page with buffers, lock the
490 * buffers using trylock before the mapping is moved. If the mapping
491 * was moved, we later failed to lock the buffers and could not move
492 * the mapping back due to an elevated page count, we would have to
493 * block waiting on other references to be dropped.
494 */
495 if (mode == MIGRATE_ASYNC && head &&
496 !buffer_migrate_lock_buffers(head, mode)) {
497 page_ref_unfreeze(page, expected_count);
498 xa_unlock_irq(&mapping->i_pages);
499 return -EAGAIN;
500 }
501
502 /*
503 * Now we know that no one else is looking at the page:
504 * no turning back from here.
505 */
506 newpage->index = page->index;
507 newpage->mapping = page->mapping;
508 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
509 if (PageSwapBacked(page)) {
510 __SetPageSwapBacked(newpage);
511 if (PageSwapCache(page)) {
512 SetPageSwapCache(newpage);
513 set_page_private(newpage, page_private(page));
514 }
515 } else {
516 VM_BUG_ON_PAGE(PageSwapCache(page), page);
517 }
518
519 /* Move dirty while page refs frozen and newpage not yet exposed */
520 dirty = PageDirty(page);
521 if (dirty) {
522 ClearPageDirty(page);
523 SetPageDirty(newpage);
524 }
525
526 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
527 if (PageTransHuge(page)) {
528 int i;
529 int index = page_index(page);
530
531 for (i = 1; i < HPAGE_PMD_NR; i++) {
532 pslot = radix_tree_lookup_slot(&mapping->i_pages,
533 index + i);
534 radix_tree_replace_slot(&mapping->i_pages, pslot,
535 newpage + i);
536 }
537 }
538
539 /*
540 * Drop cache reference from old page by unfreezing
541 * to one less reference.
542 * We know this isn't the last reference.
543 */
544 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
545
546 xa_unlock(&mapping->i_pages);
547 /* Leave irq disabled to prevent preemption while updating stats */
548
549 /*
550 * If moved to a different zone then also account
551 * the page for that zone. Other VM counters will be
552 * taken care of when we establish references to the
553 * new page and drop references to the old page.
554 *
555 * Note that anonymous pages are accounted for
556 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
557 * are mapped to swap space.
558 */
559 if (newzone != oldzone) {
560 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
561 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
562 if (PageSwapBacked(page) && !PageSwapCache(page)) {
563 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
564 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
565 }
566 if (dirty && mapping_cap_account_dirty(mapping)) {
567 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
568 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
569 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
570 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
571 }
572 }
573 local_irq_enable();
574
575 return MIGRATEPAGE_SUCCESS;
576}
577EXPORT_SYMBOL(migrate_page_move_mapping);
578
579/*
580 * The expected number of remaining references is the same as that
581 * of migrate_page_move_mapping().
582 */
583int migrate_huge_page_move_mapping(struct address_space *mapping,
584 struct page *newpage, struct page *page)
585{
586 int expected_count;
587 void **pslot;
588
589 xa_lock_irq(&mapping->i_pages);
590
591 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
592
593 expected_count = 2 + page_has_private(page);
594 if (page_count(page) != expected_count ||
595 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
596 xa_unlock_irq(&mapping->i_pages);
597 return -EAGAIN;
598 }
599
600 if (!page_ref_freeze(page, expected_count)) {
601 xa_unlock_irq(&mapping->i_pages);
602 return -EAGAIN;
603 }
604
605 newpage->index = page->index;
606 newpage->mapping = page->mapping;
607
608 get_page(newpage);
609
610 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
611
612 page_ref_unfreeze(page, expected_count - 1);
613
614 xa_unlock_irq(&mapping->i_pages);
615
616 return MIGRATEPAGE_SUCCESS;
617}
618
619/*
620 * Gigantic pages are so large that we do not guarantee that page++ pointer
621 * arithmetic will work across the entire page. We need something more
622 * specialized.
623 */
624static void __copy_gigantic_page(struct page *dst, struct page *src,
625 int nr_pages)
626{
627 int i;
628 struct page *dst_base = dst;
629 struct page *src_base = src;
630
631 for (i = 0; i < nr_pages; ) {
632 cond_resched();
633 copy_highpage(dst, src);
634
635 i++;
636 dst = mem_map_next(dst, dst_base, i);
637 src = mem_map_next(src, src_base, i);
638 }
639}
640
641static void copy_huge_page(struct page *dst, struct page *src)
642{
643 int i;
644 int nr_pages;
645
646 if (PageHuge(src)) {
647 /* hugetlbfs page */
648 struct hstate *h = page_hstate(src);
649 nr_pages = pages_per_huge_page(h);
650
651 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
652 __copy_gigantic_page(dst, src, nr_pages);
653 return;
654 }
655 } else {
656 /* thp page */
657 BUG_ON(!PageTransHuge(src));
658 nr_pages = hpage_nr_pages(src);
659 }
660
661 for (i = 0; i < nr_pages; i++) {
662 cond_resched();
663 copy_highpage(dst + i, src + i);
664 }
665}
666
667/*
668 * Copy the page to its new location
669 */
670void migrate_page_states(struct page *newpage, struct page *page)
671{
672 int cpupid;
673
674 if (PageError(page))
675 SetPageError(newpage);
676 if (PageReferenced(page))
677 SetPageReferenced(newpage);
678 if (PageUptodate(page))
679 SetPageUptodate(newpage);
680 if (TestClearPageActive(page)) {
681 VM_BUG_ON_PAGE(PageUnevictable(page), page);
682 SetPageActive(newpage);
683 } else if (TestClearPageUnevictable(page))
684 SetPageUnevictable(newpage);
685 if (PageChecked(page))
686 SetPageChecked(newpage);
687 if (PageMappedToDisk(page))
688 SetPageMappedToDisk(newpage);
689
690 /* Move dirty on pages not done by migrate_page_move_mapping() */
691 if (PageDirty(page))
692 SetPageDirty(newpage);
693
694 if (page_is_young(page))
695 set_page_young(newpage);
696 if (page_is_idle(page))
697 set_page_idle(newpage);
698
699 /*
700 * Copy NUMA information to the new page, to prevent over-eager
701 * future migrations of this same page.
702 */
703 cpupid = page_cpupid_xchg_last(page, -1);
704 page_cpupid_xchg_last(newpage, cpupid);
705
706 ksm_migrate_page(newpage, page);
707 /*
708 * Please do not reorder this without considering how mm/ksm.c's
709 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
710 */
711 if (PageSwapCache(page))
712 ClearPageSwapCache(page);
713 ClearPagePrivate(page);
714 set_page_private(page, 0);
715
716 /*
717 * If any waiters have accumulated on the new page then
718 * wake them up.
719 */
720 if (PageWriteback(newpage))
721 end_page_writeback(newpage);
722
723 copy_page_owner(page, newpage);
724
725 mem_cgroup_migrate(page, newpage);
726}
727EXPORT_SYMBOL(migrate_page_states);
728
729void migrate_page_copy(struct page *newpage, struct page *page)
730{
731 if (PageHuge(page) || PageTransHuge(page))
732 copy_huge_page(newpage, page);
733 else
734 copy_highpage(newpage, page);
735
736 migrate_page_states(newpage, page);
737}
738EXPORT_SYMBOL(migrate_page_copy);
739
740/************************************************************
741 * Migration functions
742 ***********************************************************/
743
744/*
745 * Common logic to directly migrate a single LRU page suitable for
746 * pages that do not use PagePrivate/PagePrivate2.
747 *
748 * Pages are locked upon entry and exit.
749 */
750int migrate_page(struct address_space *mapping,
751 struct page *newpage, struct page *page,
752 enum migrate_mode mode)
753{
754 int rc;
755
756 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
757
758 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
759
760 if (rc != MIGRATEPAGE_SUCCESS)
761 return rc;
762
763 if (mode != MIGRATE_SYNC_NO_COPY)
764 migrate_page_copy(newpage, page);
765 else
766 migrate_page_states(newpage, page);
767 return MIGRATEPAGE_SUCCESS;
768}
769EXPORT_SYMBOL(migrate_page);
770
771#ifdef CONFIG_BLOCK
772/*
773 * Migration function for pages with buffers. This function can only be used
774 * if the underlying filesystem guarantees that no other references to "page"
775 * exist.
776 */
777int buffer_migrate_page(struct address_space *mapping,
778 struct page *newpage, struct page *page, enum migrate_mode mode)
779{
780 struct buffer_head *bh, *head;
781 int rc;
782
783 if (!page_has_buffers(page))
784 return migrate_page(mapping, newpage, page, mode);
785
786 head = page_buffers(page);
787
788 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
789
790 if (rc != MIGRATEPAGE_SUCCESS)
791 return rc;
792
793 /*
794 * In the async case, migrate_page_move_mapping locked the buffers
795 * with an IRQ-safe spinlock held. In the sync case, the buffers
796 * need to be locked now
797 */
798 if (mode != MIGRATE_ASYNC)
799 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
800
801 ClearPagePrivate(page);
802 set_page_private(newpage, page_private(page));
803 set_page_private(page, 0);
804 put_page(page);
805 get_page(newpage);
806
807 bh = head;
808 do {
809 set_bh_page(bh, newpage, bh_offset(bh));
810 bh = bh->b_this_page;
811
812 } while (bh != head);
813
814 SetPagePrivate(newpage);
815
816 if (mode != MIGRATE_SYNC_NO_COPY)
817 migrate_page_copy(newpage, page);
818 else
819 migrate_page_states(newpage, page);
820
821 bh = head;
822 do {
823 unlock_buffer(bh);
824 put_bh(bh);
825 bh = bh->b_this_page;
826
827 } while (bh != head);
828
829 return MIGRATEPAGE_SUCCESS;
830}
831EXPORT_SYMBOL(buffer_migrate_page);
832#endif
833
834/*
835 * Writeback a page to clean the dirty state
836 */
837static int writeout(struct address_space *mapping, struct page *page)
838{
839 struct writeback_control wbc = {
840 .sync_mode = WB_SYNC_NONE,
841 .nr_to_write = 1,
842 .range_start = 0,
843 .range_end = LLONG_MAX,
844 .for_reclaim = 1
845 };
846 int rc;
847
848 if (!mapping->a_ops->writepage)
849 /* No write method for the address space */
850 return -EINVAL;
851
852 if (!clear_page_dirty_for_io(page))
853 /* Someone else already triggered a write */
854 return -EAGAIN;
855
856 /*
857 * A dirty page may imply that the underlying filesystem has
858 * the page on some queue. So the page must be clean for
859 * migration. Writeout may mean we loose the lock and the
860 * page state is no longer what we checked for earlier.
861 * At this point we know that the migration attempt cannot
862 * be successful.
863 */
864 remove_migration_ptes(page, page, false);
865
866 rc = mapping->a_ops->writepage(page, &wbc);
867
868 if (rc != AOP_WRITEPAGE_ACTIVATE)
869 /* unlocked. Relock */
870 lock_page(page);
871
872 return (rc < 0) ? -EIO : -EAGAIN;
873}
874
875/*
876 * Default handling if a filesystem does not provide a migration function.
877 */
878static int fallback_migrate_page(struct address_space *mapping,
879 struct page *newpage, struct page *page, enum migrate_mode mode)
880{
881 if (PageDirty(page)) {
882 /* Only writeback pages in full synchronous migration */
883 switch (mode) {
884 case MIGRATE_SYNC:
885 case MIGRATE_SYNC_NO_COPY:
886 break;
887 default:
888 return -EBUSY;
889 }
890 return writeout(mapping, page);
891 }
892
893 /*
894 * Buffers may be managed in a filesystem specific way.
895 * We must have no buffers or drop them.
896 */
897 if (page_has_private(page) &&
898 !try_to_release_page(page, GFP_KERNEL))
899 return -EAGAIN;
900
901 return migrate_page(mapping, newpage, page, mode);
902}
903
904/*
905 * Move a page to a newly allocated page
906 * The page is locked and all ptes have been successfully removed.
907 *
908 * The new page will have replaced the old page if this function
909 * is successful.
910 *
911 * Return value:
912 * < 0 - error code
913 * MIGRATEPAGE_SUCCESS - success
914 */
915static int move_to_new_page(struct page *newpage, struct page *page,
916 enum migrate_mode mode)
917{
918 struct address_space *mapping;
919 int rc = -EAGAIN;
920 bool is_lru = !__PageMovable(page);
921
922 VM_BUG_ON_PAGE(!PageLocked(page), page);
923 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
924
925 mapping = page_mapping(page);
926
927 if (likely(is_lru)) {
928 if (!mapping)
929 rc = migrate_page(mapping, newpage, page, mode);
930 else if (mapping->a_ops->migratepage)
931 /*
932 * Most pages have a mapping and most filesystems
933 * provide a migratepage callback. Anonymous pages
934 * are part of swap space which also has its own
935 * migratepage callback. This is the most common path
936 * for page migration.
937 */
938 rc = mapping->a_ops->migratepage(mapping, newpage,
939 page, mode);
940 else
941 rc = fallback_migrate_page(mapping, newpage,
942 page, mode);
943 } else {
944 /*
945 * In case of non-lru page, it could be released after
946 * isolation step. In that case, we shouldn't try migration.
947 */
948 VM_BUG_ON_PAGE(!PageIsolated(page), page);
949 if (!PageMovable(page)) {
950 rc = MIGRATEPAGE_SUCCESS;
951 __ClearPageIsolated(page);
952 goto out;
953 }
954
955 rc = mapping->a_ops->migratepage(mapping, newpage,
956 page, mode);
957 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
958 !PageIsolated(page));
959 }
960
961 /*
962 * When successful, old pagecache page->mapping must be cleared before
963 * page is freed; but stats require that PageAnon be left as PageAnon.
964 */
965 if (rc == MIGRATEPAGE_SUCCESS) {
966 if (__PageMovable(page)) {
967 VM_BUG_ON_PAGE(!PageIsolated(page), page);
968
969 /*
970 * We clear PG_movable under page_lock so any compactor
971 * cannot try to migrate this page.
972 */
973 __ClearPageIsolated(page);
974 }
975
976 /*
977 * Anonymous and movable page->mapping will be cleard by
978 * free_pages_prepare so don't reset it here for keeping
979 * the type to work PageAnon, for example.
980 */
981 if (!PageMappingFlags(page))
982 page->mapping = NULL;
983 }
984out:
985 return rc;
986}
987
988static int __unmap_and_move(struct page *page, struct page *newpage,
989 int force, enum migrate_mode mode)
990{
991 int rc = -EAGAIN;
992 int page_was_mapped = 0;
993 struct anon_vma *anon_vma = NULL;
994 bool is_lru = !__PageMovable(page);
995
996 if (!trylock_page(page)) {
997 if (!force || mode == MIGRATE_ASYNC)
998 goto out;
999
1000 /*
1001 * It's not safe for direct compaction to call lock_page.
1002 * For example, during page readahead pages are added locked
1003 * to the LRU. Later, when the IO completes the pages are
1004 * marked uptodate and unlocked. However, the queueing
1005 * could be merging multiple pages for one bio (e.g.
1006 * mpage_readpages). If an allocation happens for the
1007 * second or third page, the process can end up locking
1008 * the same page twice and deadlocking. Rather than
1009 * trying to be clever about what pages can be locked,
1010 * avoid the use of lock_page for direct compaction
1011 * altogether.
1012 */
1013 if (current->flags & PF_MEMALLOC)
1014 goto out;
1015
1016 lock_page(page);
1017 }
1018
1019 if (PageWriteback(page)) {
1020 /*
1021 * Only in the case of a full synchronous migration is it
1022 * necessary to wait for PageWriteback. In the async case,
1023 * the retry loop is too short and in the sync-light case,
1024 * the overhead of stalling is too much
1025 */
1026 switch (mode) {
1027 case MIGRATE_SYNC:
1028 case MIGRATE_SYNC_NO_COPY:
1029 break;
1030 default:
1031 rc = -EBUSY;
1032 goto out_unlock;
1033 }
1034 if (!force)
1035 goto out_unlock;
1036 wait_on_page_writeback(page);
1037 }
1038
1039 /*
1040 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1041 * we cannot notice that anon_vma is freed while we migrates a page.
1042 * This get_anon_vma() delays freeing anon_vma pointer until the end
1043 * of migration. File cache pages are no problem because of page_lock()
1044 * File Caches may use write_page() or lock_page() in migration, then,
1045 * just care Anon page here.
1046 *
1047 * Only page_get_anon_vma() understands the subtleties of
1048 * getting a hold on an anon_vma from outside one of its mms.
1049 * But if we cannot get anon_vma, then we won't need it anyway,
1050 * because that implies that the anon page is no longer mapped
1051 * (and cannot be remapped so long as we hold the page lock).
1052 */
1053 if (PageAnon(page) && !PageKsm(page))
1054 anon_vma = page_get_anon_vma(page);
1055
1056 /*
1057 * Block others from accessing the new page when we get around to
1058 * establishing additional references. We are usually the only one
1059 * holding a reference to newpage at this point. We used to have a BUG
1060 * here if trylock_page(newpage) fails, but would like to allow for
1061 * cases where there might be a race with the previous use of newpage.
1062 * This is much like races on refcount of oldpage: just don't BUG().
1063 */
1064 if (unlikely(!trylock_page(newpage)))
1065 goto out_unlock;
1066
1067 if (unlikely(!is_lru)) {
1068 rc = move_to_new_page(newpage, page, mode);
1069 goto out_unlock_both;
1070 }
1071
1072 /*
1073 * Corner case handling:
1074 * 1. When a new swap-cache page is read into, it is added to the LRU
1075 * and treated as swapcache but it has no rmap yet.
1076 * Calling try_to_unmap() against a page->mapping==NULL page will
1077 * trigger a BUG. So handle it here.
1078 * 2. An orphaned page (see truncate_complete_page) might have
1079 * fs-private metadata. The page can be picked up due to memory
1080 * offlining. Everywhere else except page reclaim, the page is
1081 * invisible to the vm, so the page can not be migrated. So try to
1082 * free the metadata, so the page can be freed.
1083 */
1084 if (!page->mapping) {
1085 VM_BUG_ON_PAGE(PageAnon(page), page);
1086 if (page_has_private(page)) {
1087 try_to_free_buffers(page);
1088 goto out_unlock_both;
1089 }
1090 } else if (page_mapped(page)) {
1091 /* Establish migration ptes */
1092 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1093 page);
1094 try_to_unmap(page,
1095 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1096 page_was_mapped = 1;
1097 }
1098
1099 if (!page_mapped(page))
1100 rc = move_to_new_page(newpage, page, mode);
1101
1102 if (page_was_mapped)
1103 remove_migration_ptes(page,
1104 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1105
1106out_unlock_both:
1107 unlock_page(newpage);
1108out_unlock:
1109 /* Drop an anon_vma reference if we took one */
1110 if (anon_vma)
1111 put_anon_vma(anon_vma);
1112 unlock_page(page);
1113out:
1114 /*
1115 * If migration is successful, decrease refcount of the newpage
1116 * which will not free the page because new page owner increased
1117 * refcounter. As well, if it is LRU page, add the page to LRU
1118 * list in here.
1119 */
1120 if (rc == MIGRATEPAGE_SUCCESS) {
1121 if (unlikely(__PageMovable(newpage)))
1122 put_page(newpage);
1123 else
1124 putback_lru_page(newpage);
1125 }
1126
1127 return rc;
1128}
1129
1130/*
1131 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1132 * around it.
1133 */
1134#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1135#define ICE_noinline noinline
1136#else
1137#define ICE_noinline
1138#endif
1139
1140/*
1141 * Obtain the lock on page, remove all ptes and migrate the page
1142 * to the newly allocated page in newpage.
1143 */
1144static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1145 free_page_t put_new_page,
1146 unsigned long private, struct page *page,
1147 int force, enum migrate_mode mode,
1148 enum migrate_reason reason)
1149{
1150 int rc = MIGRATEPAGE_SUCCESS;
1151 struct page *newpage;
1152
1153 if (!thp_migration_supported() && PageTransHuge(page))
1154 return -ENOMEM;
1155
1156 newpage = get_new_page(page, private);
1157 if (!newpage)
1158 return -ENOMEM;
1159
1160 if (page_count(page) == 1) {
1161 /* page was freed from under us. So we are done. */
1162 ClearPageActive(page);
1163 ClearPageUnevictable(page);
1164 if (unlikely(__PageMovable(page))) {
1165 lock_page(page);
1166 if (!PageMovable(page))
1167 __ClearPageIsolated(page);
1168 unlock_page(page);
1169 }
1170 if (put_new_page)
1171 put_new_page(newpage, private);
1172 else
1173 put_page(newpage);
1174 goto out;
1175 }
1176
1177 rc = __unmap_and_move(page, newpage, force, mode);
1178 if (rc == MIGRATEPAGE_SUCCESS)
1179 set_page_owner_migrate_reason(newpage, reason);
1180
1181out:
1182 if (rc != -EAGAIN) {
1183 /*
1184 * A page that has been migrated has all references
1185 * removed and will be freed. A page that has not been
1186 * migrated will have kepts its references and be
1187 * restored.
1188 */
1189 list_del(&page->lru);
1190
1191 /*
1192 * Compaction can migrate also non-LRU pages which are
1193 * not accounted to NR_ISOLATED_*. They can be recognized
1194 * as __PageMovable
1195 */
1196 if (likely(!__PageMovable(page)))
1197 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1198 page_is_file_cache(page), -hpage_nr_pages(page));
1199 }
1200
1201 /*
1202 * If migration is successful, releases reference grabbed during
1203 * isolation. Otherwise, restore the page to right list unless
1204 * we want to retry.
1205 */
1206 if (rc == MIGRATEPAGE_SUCCESS) {
1207 put_page(page);
1208 if (reason == MR_MEMORY_FAILURE) {
1209 /*
1210 * Set PG_HWPoison on just freed page
1211 * intentionally. Although it's rather weird,
1212 * it's how HWPoison flag works at the moment.
1213 */
1214 if (!test_set_page_hwpoison(page))
1215 num_poisoned_pages_inc();
1216 }
1217 } else {
1218 if (rc != -EAGAIN) {
1219 if (likely(!__PageMovable(page))) {
1220 putback_lru_page(page);
1221 goto put_new;
1222 }
1223
1224 lock_page(page);
1225 if (PageMovable(page))
1226 putback_movable_page(page);
1227 else
1228 __ClearPageIsolated(page);
1229 unlock_page(page);
1230 put_page(page);
1231 }
1232put_new:
1233 if (put_new_page)
1234 put_new_page(newpage, private);
1235 else
1236 put_page(newpage);
1237 }
1238
1239 return rc;
1240}
1241
1242/*
1243 * Counterpart of unmap_and_move_page() for hugepage migration.
1244 *
1245 * This function doesn't wait the completion of hugepage I/O
1246 * because there is no race between I/O and migration for hugepage.
1247 * Note that currently hugepage I/O occurs only in direct I/O
1248 * where no lock is held and PG_writeback is irrelevant,
1249 * and writeback status of all subpages are counted in the reference
1250 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1251 * under direct I/O, the reference of the head page is 512 and a bit more.)
1252 * This means that when we try to migrate hugepage whose subpages are
1253 * doing direct I/O, some references remain after try_to_unmap() and
1254 * hugepage migration fails without data corruption.
1255 *
1256 * There is also no race when direct I/O is issued on the page under migration,
1257 * because then pte is replaced with migration swap entry and direct I/O code
1258 * will wait in the page fault for migration to complete.
1259 */
1260static int unmap_and_move_huge_page(new_page_t get_new_page,
1261 free_page_t put_new_page, unsigned long private,
1262 struct page *hpage, int force,
1263 enum migrate_mode mode, int reason)
1264{
1265 int rc = -EAGAIN;
1266 int page_was_mapped = 0;
1267 struct page *new_hpage;
1268 struct anon_vma *anon_vma = NULL;
1269
1270 /*
1271 * Movability of hugepages depends on architectures and hugepage size.
1272 * This check is necessary because some callers of hugepage migration
1273 * like soft offline and memory hotremove don't walk through page
1274 * tables or check whether the hugepage is pmd-based or not before
1275 * kicking migration.
1276 */
1277 if (!hugepage_migration_supported(page_hstate(hpage))) {
1278 putback_active_hugepage(hpage);
1279 return -ENOSYS;
1280 }
1281
1282 new_hpage = get_new_page(hpage, private);
1283 if (!new_hpage)
1284 return -ENOMEM;
1285
1286 if (!trylock_page(hpage)) {
1287 if (!force)
1288 goto out;
1289 switch (mode) {
1290 case MIGRATE_SYNC:
1291 case MIGRATE_SYNC_NO_COPY:
1292 break;
1293 default:
1294 goto out;
1295 }
1296 lock_page(hpage);
1297 }
1298
1299 if (PageAnon(hpage))
1300 anon_vma = page_get_anon_vma(hpage);
1301
1302 if (unlikely(!trylock_page(new_hpage)))
1303 goto put_anon;
1304
1305 if (page_mapped(hpage)) {
1306 try_to_unmap(hpage,
1307 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1308 page_was_mapped = 1;
1309 }
1310
1311 if (!page_mapped(hpage))
1312 rc = move_to_new_page(new_hpage, hpage, mode);
1313
1314 if (page_was_mapped)
1315 remove_migration_ptes(hpage,
1316 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1317
1318 unlock_page(new_hpage);
1319
1320put_anon:
1321 if (anon_vma)
1322 put_anon_vma(anon_vma);
1323
1324 if (rc == MIGRATEPAGE_SUCCESS) {
1325 move_hugetlb_state(hpage, new_hpage, reason);
1326 put_new_page = NULL;
1327 }
1328
1329 unlock_page(hpage);
1330out:
1331 if (rc != -EAGAIN)
1332 putback_active_hugepage(hpage);
1333 if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage))
1334 num_poisoned_pages_inc();
1335
1336 /*
1337 * If migration was not successful and there's a freeing callback, use
1338 * it. Otherwise, put_page() will drop the reference grabbed during
1339 * isolation.
1340 */
1341 if (put_new_page)
1342 put_new_page(new_hpage, private);
1343 else
1344 putback_active_hugepage(new_hpage);
1345
1346 return rc;
1347}
1348
1349/*
1350 * migrate_pages - migrate the pages specified in a list, to the free pages
1351 * supplied as the target for the page migration
1352 *
1353 * @from: The list of pages to be migrated.
1354 * @get_new_page: The function used to allocate free pages to be used
1355 * as the target of the page migration.
1356 * @put_new_page: The function used to free target pages if migration
1357 * fails, or NULL if no special handling is necessary.
1358 * @private: Private data to be passed on to get_new_page()
1359 * @mode: The migration mode that specifies the constraints for
1360 * page migration, if any.
1361 * @reason: The reason for page migration.
1362 *
1363 * The function returns after 10 attempts or if no pages are movable any more
1364 * because the list has become empty or no retryable pages exist any more.
1365 * The caller should call putback_movable_pages() to return pages to the LRU
1366 * or free list only if ret != 0.
1367 *
1368 * Returns the number of pages that were not migrated, or an error code.
1369 */
1370int migrate_pages(struct list_head *from, new_page_t get_new_page,
1371 free_page_t put_new_page, unsigned long private,
1372 enum migrate_mode mode, int reason)
1373{
1374 int retry = 1;
1375 int nr_failed = 0;
1376 int nr_succeeded = 0;
1377 int pass = 0;
1378 struct page *page;
1379 struct page *page2;
1380 int swapwrite = current->flags & PF_SWAPWRITE;
1381 int rc;
1382
1383 if (!swapwrite)
1384 current->flags |= PF_SWAPWRITE;
1385
1386 for(pass = 0; pass < 10 && retry; pass++) {
1387 retry = 0;
1388
1389 list_for_each_entry_safe(page, page2, from, lru) {
1390retry:
1391 cond_resched();
1392
1393 if (PageHuge(page))
1394 rc = unmap_and_move_huge_page(get_new_page,
1395 put_new_page, private, page,
1396 pass > 2, mode, reason);
1397 else
1398 rc = unmap_and_move(get_new_page, put_new_page,
1399 private, page, pass > 2, mode,
1400 reason);
1401
1402 switch(rc) {
1403 case -ENOMEM:
1404 /*
1405 * THP migration might be unsupported or the
1406 * allocation could've failed so we should
1407 * retry on the same page with the THP split
1408 * to base pages.
1409 *
1410 * Head page is retried immediately and tail
1411 * pages are added to the tail of the list so
1412 * we encounter them after the rest of the list
1413 * is processed.
1414 */
1415 if (PageTransHuge(page)) {
1416 lock_page(page);
1417 rc = split_huge_page_to_list(page, from);
1418 unlock_page(page);
1419 if (!rc) {
1420 list_safe_reset_next(page, page2, lru);
1421 goto retry;
1422 }
1423 }
1424 nr_failed++;
1425 goto out;
1426 case -EAGAIN:
1427 retry++;
1428 break;
1429 case MIGRATEPAGE_SUCCESS:
1430 nr_succeeded++;
1431 break;
1432 default:
1433 /*
1434 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1435 * unlike -EAGAIN case, the failed page is
1436 * removed from migration page list and not
1437 * retried in the next outer loop.
1438 */
1439 nr_failed++;
1440 break;
1441 }
1442 }
1443 }
1444 nr_failed += retry;
1445 rc = nr_failed;
1446out:
1447 if (nr_succeeded)
1448 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1449 if (nr_failed)
1450 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1451 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1452
1453 if (!swapwrite)
1454 current->flags &= ~PF_SWAPWRITE;
1455
1456 return rc;
1457}
1458
1459#ifdef CONFIG_NUMA
1460
1461static int store_status(int __user *status, int start, int value, int nr)
1462{
1463 while (nr-- > 0) {
1464 if (put_user(value, status + start))
1465 return -EFAULT;
1466 start++;
1467 }
1468
1469 return 0;
1470}
1471
1472static int do_move_pages_to_node(struct mm_struct *mm,
1473 struct list_head *pagelist, int node)
1474{
1475 int err;
1476
1477 if (list_empty(pagelist))
1478 return 0;
1479
1480 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1481 MIGRATE_SYNC, MR_SYSCALL);
1482 if (err)
1483 putback_movable_pages(pagelist);
1484 return err;
1485}
1486
1487/*
1488 * Resolves the given address to a struct page, isolates it from the LRU and
1489 * puts it to the given pagelist.
1490 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1491 * queued or the page doesn't need to be migrated because it is already on
1492 * the target node
1493 */
1494static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1495 int node, struct list_head *pagelist, bool migrate_all)
1496{
1497 struct vm_area_struct *vma;
1498 struct page *page;
1499 unsigned int follflags;
1500 int err;
1501
1502 down_read(&mm->mmap_sem);
1503 err = -EFAULT;
1504 vma = find_vma(mm, addr);
1505 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1506 goto out;
1507
1508 /* FOLL_DUMP to ignore special (like zero) pages */
1509 follflags = FOLL_GET | FOLL_DUMP;
1510 page = follow_page(vma, addr, follflags);
1511
1512 err = PTR_ERR(page);
1513 if (IS_ERR(page))
1514 goto out;
1515
1516 err = -ENOENT;
1517 if (!page)
1518 goto out;
1519
1520 err = 0;
1521 if (page_to_nid(page) == node)
1522 goto out_putpage;
1523
1524 err = -EACCES;
1525 if (page_mapcount(page) > 1 && !migrate_all)
1526 goto out_putpage;
1527
1528 if (PageHuge(page)) {
1529 if (PageHead(page)) {
1530 isolate_huge_page(page, pagelist);
1531 err = 0;
1532 }
1533 } else {
1534 struct page *head;
1535
1536 head = compound_head(page);
1537 err = isolate_lru_page(head);
1538 if (err)
1539 goto out_putpage;
1540
1541 err = 0;
1542 list_add_tail(&head->lru, pagelist);
1543 mod_node_page_state(page_pgdat(head),
1544 NR_ISOLATED_ANON + page_is_file_cache(head),
1545 hpage_nr_pages(head));
1546 }
1547out_putpage:
1548 /*
1549 * Either remove the duplicate refcount from
1550 * isolate_lru_page() or drop the page ref if it was
1551 * not isolated.
1552 */
1553 put_page(page);
1554out:
1555 up_read(&mm->mmap_sem);
1556 return err;
1557}
1558
1559/*
1560 * Migrate an array of page address onto an array of nodes and fill
1561 * the corresponding array of status.
1562 */
1563static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1564 unsigned long nr_pages,
1565 const void __user * __user *pages,
1566 const int __user *nodes,
1567 int __user *status, int flags)
1568{
1569 int current_node = NUMA_NO_NODE;
1570 LIST_HEAD(pagelist);
1571 int start, i;
1572 int err = 0, err1;
1573
1574 migrate_prep();
1575
1576 for (i = start = 0; i < nr_pages; i++) {
1577 const void __user *p;
1578 unsigned long addr;
1579 int node;
1580
1581 err = -EFAULT;
1582 if (get_user(p, pages + i))
1583 goto out_flush;
1584 if (get_user(node, nodes + i))
1585 goto out_flush;
1586 addr = (unsigned long)p;
1587
1588 err = -ENODEV;
1589 if (node < 0 || node >= MAX_NUMNODES)
1590 goto out_flush;
1591 if (!node_state(node, N_MEMORY))
1592 goto out_flush;
1593
1594 err = -EACCES;
1595 if (!node_isset(node, task_nodes))
1596 goto out_flush;
1597
1598 if (current_node == NUMA_NO_NODE) {
1599 current_node = node;
1600 start = i;
1601 } else if (node != current_node) {
1602 err = do_move_pages_to_node(mm, &pagelist, current_node);
1603 if (err)
1604 goto out;
1605 err = store_status(status, start, current_node, i - start);
1606 if (err)
1607 goto out;
1608 start = i;
1609 current_node = node;
1610 }
1611
1612 /*
1613 * Errors in the page lookup or isolation are not fatal and we simply
1614 * report them via status
1615 */
1616 err = add_page_for_migration(mm, addr, current_node,
1617 &pagelist, flags & MPOL_MF_MOVE_ALL);
1618 if (!err)
1619 continue;
1620
1621 err = store_status(status, i, err, 1);
1622 if (err)
1623 goto out_flush;
1624
1625 err = do_move_pages_to_node(mm, &pagelist, current_node);
1626 if (err)
1627 goto out;
1628 if (i > start) {
1629 err = store_status(status, start, current_node, i - start);
1630 if (err)
1631 goto out;
1632 }
1633 current_node = NUMA_NO_NODE;
1634 }
1635out_flush:
1636 if (list_empty(&pagelist))
1637 return err;
1638
1639 /* Make sure we do not overwrite the existing error */
1640 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1641 if (!err1)
1642 err1 = store_status(status, start, current_node, i - start);
1643 if (!err)
1644 err = err1;
1645out:
1646 return err;
1647}
1648
1649/*
1650 * Determine the nodes of an array of pages and store it in an array of status.
1651 */
1652static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1653 const void __user **pages, int *status)
1654{
1655 unsigned long i;
1656
1657 down_read(&mm->mmap_sem);
1658
1659 for (i = 0; i < nr_pages; i++) {
1660 unsigned long addr = (unsigned long)(*pages);
1661 struct vm_area_struct *vma;
1662 struct page *page;
1663 int err = -EFAULT;
1664
1665 vma = find_vma(mm, addr);
1666 if (!vma || addr < vma->vm_start)
1667 goto set_status;
1668
1669 /* FOLL_DUMP to ignore special (like zero) pages */
1670 page = follow_page(vma, addr, FOLL_DUMP);
1671
1672 err = PTR_ERR(page);
1673 if (IS_ERR(page))
1674 goto set_status;
1675
1676 err = page ? page_to_nid(page) : -ENOENT;
1677set_status:
1678 *status = err;
1679
1680 pages++;
1681 status++;
1682 }
1683
1684 up_read(&mm->mmap_sem);
1685}
1686
1687/*
1688 * Determine the nodes of a user array of pages and store it in
1689 * a user array of status.
1690 */
1691static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1692 const void __user * __user *pages,
1693 int __user *status)
1694{
1695#define DO_PAGES_STAT_CHUNK_NR 16
1696 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1697 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1698
1699 while (nr_pages) {
1700 unsigned long chunk_nr;
1701
1702 chunk_nr = nr_pages;
1703 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1704 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1705
1706 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1707 break;
1708
1709 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1710
1711 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1712 break;
1713
1714 pages += chunk_nr;
1715 status += chunk_nr;
1716 nr_pages -= chunk_nr;
1717 }
1718 return nr_pages ? -EFAULT : 0;
1719}
1720
1721/*
1722 * Move a list of pages in the address space of the currently executing
1723 * process.
1724 */
1725static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1726 const void __user * __user *pages,
1727 const int __user *nodes,
1728 int __user *status, int flags)
1729{
1730 struct task_struct *task;
1731 struct mm_struct *mm;
1732 int err;
1733 nodemask_t task_nodes;
1734
1735 /* Check flags */
1736 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1737 return -EINVAL;
1738
1739 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1740 return -EPERM;
1741
1742 /* Find the mm_struct */
1743 rcu_read_lock();
1744 task = pid ? find_task_by_vpid(pid) : current;
1745 if (!task) {
1746 rcu_read_unlock();
1747 return -ESRCH;
1748 }
1749 get_task_struct(task);
1750
1751 /*
1752 * Check if this process has the right to modify the specified
1753 * process. Use the regular "ptrace_may_access()" checks.
1754 */
1755 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1756 rcu_read_unlock();
1757 err = -EPERM;
1758 goto out;
1759 }
1760 rcu_read_unlock();
1761
1762 err = security_task_movememory(task);
1763 if (err)
1764 goto out;
1765
1766 task_nodes = cpuset_mems_allowed(task);
1767 mm = get_task_mm(task);
1768 put_task_struct(task);
1769
1770 if (!mm)
1771 return -EINVAL;
1772
1773 if (nodes)
1774 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1775 nodes, status, flags);
1776 else
1777 err = do_pages_stat(mm, nr_pages, pages, status);
1778
1779 mmput(mm);
1780 return err;
1781
1782out:
1783 put_task_struct(task);
1784 return err;
1785}
1786
1787SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1788 const void __user * __user *, pages,
1789 const int __user *, nodes,
1790 int __user *, status, int, flags)
1791{
1792 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1793}
1794
1795#ifdef CONFIG_COMPAT
1796COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1797 compat_uptr_t __user *, pages32,
1798 const int __user *, nodes,
1799 int __user *, status,
1800 int, flags)
1801{
1802 const void __user * __user *pages;
1803 int i;
1804
1805 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1806 for (i = 0; i < nr_pages; i++) {
1807 compat_uptr_t p;
1808
1809 if (get_user(p, pages32 + i) ||
1810 put_user(compat_ptr(p), pages + i))
1811 return -EFAULT;
1812 }
1813 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1814}
1815#endif /* CONFIG_COMPAT */
1816
1817#ifdef CONFIG_NUMA_BALANCING
1818/*
1819 * Returns true if this is a safe migration target node for misplaced NUMA
1820 * pages. Currently it only checks the watermarks which crude
1821 */
1822static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1823 unsigned long nr_migrate_pages)
1824{
1825 int z;
1826
1827 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1828 struct zone *zone = pgdat->node_zones + z;
1829
1830 if (!populated_zone(zone))
1831 continue;
1832
1833 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1834 if (!zone_watermark_ok(zone, 0,
1835 high_wmark_pages(zone) +
1836 nr_migrate_pages,
1837 0, 0))
1838 continue;
1839 return true;
1840 }
1841 return false;
1842}
1843
1844static struct page *alloc_misplaced_dst_page(struct page *page,
1845 unsigned long data)
1846{
1847 int nid = (int) data;
1848 struct page *newpage;
1849
1850 newpage = __alloc_pages_node(nid,
1851 (GFP_HIGHUSER_MOVABLE |
1852 __GFP_THISNODE | __GFP_NOMEMALLOC |
1853 __GFP_NORETRY | __GFP_NOWARN) &
1854 ~__GFP_RECLAIM, 0);
1855
1856 return newpage;
1857}
1858
1859/*
1860 * page migration rate limiting control.
1861 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1862 * window of time. Default here says do not migrate more than 1280M per second.
1863 */
1864static unsigned int migrate_interval_millisecs __read_mostly = 100;
1865static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1866
1867/* Returns true if the node is migrate rate-limited after the update */
1868static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1869 unsigned long nr_pages)
1870{
1871 /*
1872 * Rate-limit the amount of data that is being migrated to a node.
1873 * Optimal placement is no good if the memory bus is saturated and
1874 * all the time is being spent migrating!
1875 */
1876 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1877 spin_lock(&pgdat->numabalancing_migrate_lock);
1878 pgdat->numabalancing_migrate_nr_pages = 0;
1879 pgdat->numabalancing_migrate_next_window = jiffies +
1880 msecs_to_jiffies(migrate_interval_millisecs);
1881 spin_unlock(&pgdat->numabalancing_migrate_lock);
1882 }
1883 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1884 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1885 nr_pages);
1886 return true;
1887 }
1888
1889 /*
1890 * This is an unlocked non-atomic update so errors are possible.
1891 * The consequences are failing to migrate when we potentiall should
1892 * have which is not severe enough to warrant locking. If it is ever
1893 * a problem, it can be converted to a per-cpu counter.
1894 */
1895 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1896 return false;
1897}
1898
1899static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1900{
1901 int page_lru;
1902
1903 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1904
1905 /* Avoid migrating to a node that is nearly full */
1906 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1907 return 0;
1908
1909 if (isolate_lru_page(page))
1910 return 0;
1911
1912 /*
1913 * migrate_misplaced_transhuge_page() skips page migration's usual
1914 * check on page_count(), so we must do it here, now that the page
1915 * has been isolated: a GUP pin, or any other pin, prevents migration.
1916 * The expected page count is 3: 1 for page's mapcount and 1 for the
1917 * caller's pin and 1 for the reference taken by isolate_lru_page().
1918 */
1919 if (PageTransHuge(page) && page_count(page) != 3) {
1920 putback_lru_page(page);
1921 return 0;
1922 }
1923
1924 page_lru = page_is_file_cache(page);
1925 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1926 hpage_nr_pages(page));
1927
1928 /*
1929 * Isolating the page has taken another reference, so the
1930 * caller's reference can be safely dropped without the page
1931 * disappearing underneath us during migration.
1932 */
1933 put_page(page);
1934 return 1;
1935}
1936
1937bool pmd_trans_migrating(pmd_t pmd)
1938{
1939 struct page *page = pmd_page(pmd);
1940 return PageLocked(page);
1941}
1942
1943/*
1944 * Attempt to migrate a misplaced page to the specified destination
1945 * node. Caller is expected to have an elevated reference count on
1946 * the page that will be dropped by this function before returning.
1947 */
1948int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1949 int node)
1950{
1951 pg_data_t *pgdat = NODE_DATA(node);
1952 int isolated;
1953 int nr_remaining;
1954 LIST_HEAD(migratepages);
1955
1956 /*
1957 * Don't migrate file pages that are mapped in multiple processes
1958 * with execute permissions as they are probably shared libraries.
1959 */
1960 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1961 (vma->vm_flags & VM_EXEC))
1962 goto out;
1963
1964 /*
1965 * Also do not migrate dirty pages as not all filesystems can move
1966 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1967 */
1968 if (page_is_file_cache(page) && PageDirty(page))
1969 goto out;
1970
1971 /*
1972 * Rate-limit the amount of data that is being migrated to a node.
1973 * Optimal placement is no good if the memory bus is saturated and
1974 * all the time is being spent migrating!
1975 */
1976 if (numamigrate_update_ratelimit(pgdat, 1))
1977 goto out;
1978
1979 isolated = numamigrate_isolate_page(pgdat, page);
1980 if (!isolated)
1981 goto out;
1982
1983 list_add(&page->lru, &migratepages);
1984 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1985 NULL, node, MIGRATE_ASYNC,
1986 MR_NUMA_MISPLACED);
1987 if (nr_remaining) {
1988 if (!list_empty(&migratepages)) {
1989 list_del(&page->lru);
1990 dec_node_page_state(page, NR_ISOLATED_ANON +
1991 page_is_file_cache(page));
1992 putback_lru_page(page);
1993 }
1994 isolated = 0;
1995 } else
1996 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1997 BUG_ON(!list_empty(&migratepages));
1998 return isolated;
1999
2000out:
2001 put_page(page);
2002 return 0;
2003}
2004#endif /* CONFIG_NUMA_BALANCING */
2005
2006#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2007/*
2008 * Migrates a THP to a given target node. page must be locked and is unlocked
2009 * before returning.
2010 */
2011int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2012 struct vm_area_struct *vma,
2013 pmd_t *pmd, pmd_t entry,
2014 unsigned long address,
2015 struct page *page, int node)
2016{
2017 spinlock_t *ptl;
2018 pg_data_t *pgdat = NODE_DATA(node);
2019 int isolated = 0;
2020 struct page *new_page = NULL;
2021 int page_lru = page_is_file_cache(page);
2022 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2023 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2024
2025 /*
2026 * Rate-limit the amount of data that is being migrated to a node.
2027 * Optimal placement is no good if the memory bus is saturated and
2028 * all the time is being spent migrating!
2029 */
2030 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
2031 goto out_dropref;
2032
2033 new_page = alloc_pages_node(node,
2034 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2035 HPAGE_PMD_ORDER);
2036 if (!new_page)
2037 goto out_fail;
2038 prep_transhuge_page(new_page);
2039
2040 isolated = numamigrate_isolate_page(pgdat, page);
2041 if (!isolated) {
2042 put_page(new_page);
2043 goto out_fail;
2044 }
2045
2046 /* Prepare a page as a migration target */
2047 __SetPageLocked(new_page);
2048 if (PageSwapBacked(page))
2049 __SetPageSwapBacked(new_page);
2050
2051 /* anon mapping, we can simply copy page->mapping to the new page: */
2052 new_page->mapping = page->mapping;
2053 new_page->index = page->index;
2054 migrate_page_copy(new_page, page);
2055 WARN_ON(PageLRU(new_page));
2056
2057 /* Recheck the target PMD */
2058 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2059 ptl = pmd_lock(mm, pmd);
2060 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2061 spin_unlock(ptl);
2062 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2063
2064 /* Reverse changes made by migrate_page_copy() */
2065 if (TestClearPageActive(new_page))
2066 SetPageActive(page);
2067 if (TestClearPageUnevictable(new_page))
2068 SetPageUnevictable(page);
2069
2070 unlock_page(new_page);
2071 put_page(new_page); /* Free it */
2072
2073 /* Retake the callers reference and putback on LRU */
2074 get_page(page);
2075 putback_lru_page(page);
2076 mod_node_page_state(page_pgdat(page),
2077 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2078
2079 goto out_unlock;
2080 }
2081
2082 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2083 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2084
2085 /*
2086 * Clear the old entry under pagetable lock and establish the new PTE.
2087 * Any parallel GUP will either observe the old page blocking on the
2088 * page lock, block on the page table lock or observe the new page.
2089 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2090 * guarantee the copy is visible before the pagetable update.
2091 */
2092 flush_cache_range(vma, mmun_start, mmun_end);
2093 page_add_anon_rmap(new_page, vma, mmun_start, true);
2094 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2095 set_pmd_at(mm, mmun_start, pmd, entry);
2096 update_mmu_cache_pmd(vma, address, &entry);
2097
2098 page_ref_unfreeze(page, 2);
2099 mlock_migrate_page(new_page, page);
2100 page_remove_rmap(page, true);
2101 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2102
2103 spin_unlock(ptl);
2104 /*
2105 * No need to double call mmu_notifier->invalidate_range() callback as
2106 * the above pmdp_huge_clear_flush_notify() did already call it.
2107 */
2108 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2109
2110 /* Take an "isolate" reference and put new page on the LRU. */
2111 get_page(new_page);
2112 putback_lru_page(new_page);
2113
2114 unlock_page(new_page);
2115 unlock_page(page);
2116 put_page(page); /* Drop the rmap reference */
2117 put_page(page); /* Drop the LRU isolation reference */
2118
2119 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2120 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2121
2122 mod_node_page_state(page_pgdat(page),
2123 NR_ISOLATED_ANON + page_lru,
2124 -HPAGE_PMD_NR);
2125 return isolated;
2126
2127out_fail:
2128 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2129out_dropref:
2130 ptl = pmd_lock(mm, pmd);
2131 if (pmd_same(*pmd, entry)) {
2132 entry = pmd_modify(entry, vma->vm_page_prot);
2133 set_pmd_at(mm, mmun_start, pmd, entry);
2134 update_mmu_cache_pmd(vma, address, &entry);
2135 }
2136 spin_unlock(ptl);
2137
2138out_unlock:
2139 unlock_page(page);
2140 put_page(page);
2141 return 0;
2142}
2143#endif /* CONFIG_NUMA_BALANCING */
2144
2145#endif /* CONFIG_NUMA */
2146
2147#if defined(CONFIG_MIGRATE_VMA_HELPER)
2148struct migrate_vma {
2149 struct vm_area_struct *vma;
2150 unsigned long *dst;
2151 unsigned long *src;
2152 unsigned long cpages;
2153 unsigned long npages;
2154 unsigned long start;
2155 unsigned long end;
2156};
2157
2158static int migrate_vma_collect_hole(unsigned long start,
2159 unsigned long end,
2160 struct mm_walk *walk)
2161{
2162 struct migrate_vma *migrate = walk->private;
2163 unsigned long addr;
2164
2165 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2166 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2167 migrate->dst[migrate->npages] = 0;
2168 migrate->npages++;
2169 migrate->cpages++;
2170 }
2171
2172 return 0;
2173}
2174
2175static int migrate_vma_collect_skip(unsigned long start,
2176 unsigned long end,
2177 struct mm_walk *walk)
2178{
2179 struct migrate_vma *migrate = walk->private;
2180 unsigned long addr;
2181
2182 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2183 migrate->dst[migrate->npages] = 0;
2184 migrate->src[migrate->npages++] = 0;
2185 }
2186
2187 return 0;
2188}
2189
2190static int migrate_vma_collect_pmd(pmd_t *pmdp,
2191 unsigned long start,
2192 unsigned long end,
2193 struct mm_walk *walk)
2194{
2195 struct migrate_vma *migrate = walk->private;
2196 struct vm_area_struct *vma = walk->vma;
2197 struct mm_struct *mm = vma->vm_mm;
2198 unsigned long addr = start, unmapped = 0;
2199 spinlock_t *ptl;
2200 pte_t *ptep;
2201
2202again:
2203 if (pmd_none(*pmdp))
2204 return migrate_vma_collect_hole(start, end, walk);
2205
2206 if (pmd_trans_huge(*pmdp)) {
2207 struct page *page;
2208
2209 ptl = pmd_lock(mm, pmdp);
2210 if (unlikely(!pmd_trans_huge(*pmdp))) {
2211 spin_unlock(ptl);
2212 goto again;
2213 }
2214
2215 page = pmd_page(*pmdp);
2216 if (is_huge_zero_page(page)) {
2217 spin_unlock(ptl);
2218 split_huge_pmd(vma, pmdp, addr);
2219 if (pmd_trans_unstable(pmdp))
2220 return migrate_vma_collect_skip(start, end,
2221 walk);
2222 } else {
2223 int ret;
2224
2225 get_page(page);
2226 spin_unlock(ptl);
2227 if (unlikely(!trylock_page(page)))
2228 return migrate_vma_collect_skip(start, end,
2229 walk);
2230 ret = split_huge_page(page);
2231 unlock_page(page);
2232 put_page(page);
2233 if (ret)
2234 return migrate_vma_collect_skip(start, end,
2235 walk);
2236 if (pmd_none(*pmdp))
2237 return migrate_vma_collect_hole(start, end,
2238 walk);
2239 }
2240 }
2241
2242 if (unlikely(pmd_bad(*pmdp)))
2243 return migrate_vma_collect_skip(start, end, walk);
2244
2245 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2246 arch_enter_lazy_mmu_mode();
2247
2248 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2249 unsigned long mpfn, pfn;
2250 struct page *page;
2251 swp_entry_t entry;
2252 pte_t pte;
2253
2254 pte = *ptep;
2255 pfn = pte_pfn(pte);
2256
2257 if (pte_none(pte)) {
2258 mpfn = MIGRATE_PFN_MIGRATE;
2259 migrate->cpages++;
2260 pfn = 0;
2261 goto next;
2262 }
2263
2264 if (!pte_present(pte)) {
2265 mpfn = pfn = 0;
2266
2267 /*
2268 * Only care about unaddressable device page special
2269 * page table entry. Other special swap entries are not
2270 * migratable, and we ignore regular swapped page.
2271 */
2272 entry = pte_to_swp_entry(pte);
2273 if (!is_device_private_entry(entry))
2274 goto next;
2275
2276 page = device_private_entry_to_page(entry);
2277 mpfn = migrate_pfn(page_to_pfn(page))|
2278 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2279 if (is_write_device_private_entry(entry))
2280 mpfn |= MIGRATE_PFN_WRITE;
2281 } else {
2282 if (is_zero_pfn(pfn)) {
2283 mpfn = MIGRATE_PFN_MIGRATE;
2284 migrate->cpages++;
2285 pfn = 0;
2286 goto next;
2287 }
2288 page = _vm_normal_page(migrate->vma, addr, pte, true);
2289 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2290 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2291 }
2292
2293 /* FIXME support THP */
2294 if (!page || !page->mapping || PageTransCompound(page)) {
2295 mpfn = pfn = 0;
2296 goto next;
2297 }
2298 pfn = page_to_pfn(page);
2299
2300 /*
2301 * By getting a reference on the page we pin it and that blocks
2302 * any kind of migration. Side effect is that it "freezes" the
2303 * pte.
2304 *
2305 * We drop this reference after isolating the page from the lru
2306 * for non device page (device page are not on the lru and thus
2307 * can't be dropped from it).
2308 */
2309 get_page(page);
2310 migrate->cpages++;
2311
2312 /*
2313 * Optimize for the common case where page is only mapped once
2314 * in one process. If we can lock the page, then we can safely
2315 * set up a special migration page table entry now.
2316 */
2317 if (trylock_page(page)) {
2318 pte_t swp_pte;
2319
2320 mpfn |= MIGRATE_PFN_LOCKED;
2321 ptep_get_and_clear(mm, addr, ptep);
2322
2323 /* Setup special migration page table entry */
2324 entry = make_migration_entry(page, mpfn &
2325 MIGRATE_PFN_WRITE);
2326 swp_pte = swp_entry_to_pte(entry);
2327 if (pte_soft_dirty(pte))
2328 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2329 set_pte_at(mm, addr, ptep, swp_pte);
2330
2331 /*
2332 * This is like regular unmap: we remove the rmap and
2333 * drop page refcount. Page won't be freed, as we took
2334 * a reference just above.
2335 */
2336 page_remove_rmap(page, false);
2337 put_page(page);
2338
2339 if (pte_present(pte))
2340 unmapped++;
2341 }
2342
2343next:
2344 migrate->dst[migrate->npages] = 0;
2345 migrate->src[migrate->npages++] = mpfn;
2346 }
2347 arch_leave_lazy_mmu_mode();
2348 pte_unmap_unlock(ptep - 1, ptl);
2349
2350 /* Only flush the TLB if we actually modified any entries */
2351 if (unmapped)
2352 flush_tlb_range(walk->vma, start, end);
2353
2354 return 0;
2355}
2356
2357/*
2358 * migrate_vma_collect() - collect pages over a range of virtual addresses
2359 * @migrate: migrate struct containing all migration information
2360 *
2361 * This will walk the CPU page table. For each virtual address backed by a
2362 * valid page, it updates the src array and takes a reference on the page, in
2363 * order to pin the page until we lock it and unmap it.
2364 */
2365static void migrate_vma_collect(struct migrate_vma *migrate)
2366{
2367 struct mm_walk mm_walk;
2368
2369 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2370 mm_walk.pte_entry = NULL;
2371 mm_walk.pte_hole = migrate_vma_collect_hole;
2372 mm_walk.hugetlb_entry = NULL;
2373 mm_walk.test_walk = NULL;
2374 mm_walk.vma = migrate->vma;
2375 mm_walk.mm = migrate->vma->vm_mm;
2376 mm_walk.private = migrate;
2377
2378 mmu_notifier_invalidate_range_start(mm_walk.mm,
2379 migrate->start,
2380 migrate->end);
2381 walk_page_range(migrate->start, migrate->end, &mm_walk);
2382 mmu_notifier_invalidate_range_end(mm_walk.mm,
2383 migrate->start,
2384 migrate->end);
2385
2386 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2387}
2388
2389/*
2390 * migrate_vma_check_page() - check if page is pinned or not
2391 * @page: struct page to check
2392 *
2393 * Pinned pages cannot be migrated. This is the same test as in
2394 * migrate_page_move_mapping(), except that here we allow migration of a
2395 * ZONE_DEVICE page.
2396 */
2397static bool migrate_vma_check_page(struct page *page)
2398{
2399 /*
2400 * One extra ref because caller holds an extra reference, either from
2401 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2402 * a device page.
2403 */
2404 int extra = 1;
2405
2406 /*
2407 * FIXME support THP (transparent huge page), it is bit more complex to
2408 * check them than regular pages, because they can be mapped with a pmd
2409 * or with a pte (split pte mapping).
2410 */
2411 if (PageCompound(page))
2412 return false;
2413
2414 /* Page from ZONE_DEVICE have one extra reference */
2415 if (is_zone_device_page(page)) {
2416 /*
2417 * Private page can never be pin as they have no valid pte and
2418 * GUP will fail for those. Yet if there is a pending migration
2419 * a thread might try to wait on the pte migration entry and
2420 * will bump the page reference count. Sadly there is no way to
2421 * differentiate a regular pin from migration wait. Hence to
2422 * avoid 2 racing thread trying to migrate back to CPU to enter
2423 * infinite loop (one stoping migration because the other is
2424 * waiting on pte migration entry). We always return true here.
2425 *
2426 * FIXME proper solution is to rework migration_entry_wait() so
2427 * it does not need to take a reference on page.
2428 */
2429 if (is_device_private_page(page))
2430 return true;
2431
2432 /*
2433 * Only allow device public page to be migrated and account for
2434 * the extra reference count imply by ZONE_DEVICE pages.
2435 */
2436 if (!is_device_public_page(page))
2437 return false;
2438 extra++;
2439 }
2440
2441 /* For file back page */
2442 if (page_mapping(page))
2443 extra += 1 + page_has_private(page);
2444
2445 if ((page_count(page) - extra) > page_mapcount(page))
2446 return false;
2447
2448 return true;
2449}
2450
2451/*
2452 * migrate_vma_prepare() - lock pages and isolate them from the lru
2453 * @migrate: migrate struct containing all migration information
2454 *
2455 * This locks pages that have been collected by migrate_vma_collect(). Once each
2456 * page is locked it is isolated from the lru (for non-device pages). Finally,
2457 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2458 * migrated by concurrent kernel threads.
2459 */
2460static void migrate_vma_prepare(struct migrate_vma *migrate)
2461{
2462 const unsigned long npages = migrate->npages;
2463 const unsigned long start = migrate->start;
2464 unsigned long addr, i, restore = 0;
2465 bool allow_drain = true;
2466
2467 lru_add_drain();
2468
2469 for (i = 0; (i < npages) && migrate->cpages; i++) {
2470 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2471 bool remap = true;
2472
2473 if (!page)
2474 continue;
2475
2476 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2477 /*
2478 * Because we are migrating several pages there can be
2479 * a deadlock between 2 concurrent migration where each
2480 * are waiting on each other page lock.
2481 *
2482 * Make migrate_vma() a best effort thing and backoff
2483 * for any page we can not lock right away.
2484 */
2485 if (!trylock_page(page)) {
2486 migrate->src[i] = 0;
2487 migrate->cpages--;
2488 put_page(page);
2489 continue;
2490 }
2491 remap = false;
2492 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2493 }
2494
2495 /* ZONE_DEVICE pages are not on LRU */
2496 if (!is_zone_device_page(page)) {
2497 if (!PageLRU(page) && allow_drain) {
2498 /* Drain CPU's pagevec */
2499 lru_add_drain_all();
2500 allow_drain = false;
2501 }
2502
2503 if (isolate_lru_page(page)) {
2504 if (remap) {
2505 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2506 migrate->cpages--;
2507 restore++;
2508 } else {
2509 migrate->src[i] = 0;
2510 unlock_page(page);
2511 migrate->cpages--;
2512 put_page(page);
2513 }
2514 continue;
2515 }
2516
2517 /* Drop the reference we took in collect */
2518 put_page(page);
2519 }
2520
2521 if (!migrate_vma_check_page(page)) {
2522 if (remap) {
2523 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2524 migrate->cpages--;
2525 restore++;
2526
2527 if (!is_zone_device_page(page)) {
2528 get_page(page);
2529 putback_lru_page(page);
2530 }
2531 } else {
2532 migrate->src[i] = 0;
2533 unlock_page(page);
2534 migrate->cpages--;
2535
2536 if (!is_zone_device_page(page))
2537 putback_lru_page(page);
2538 else
2539 put_page(page);
2540 }
2541 }
2542 }
2543
2544 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2545 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2546
2547 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2548 continue;
2549
2550 remove_migration_pte(page, migrate->vma, addr, page);
2551
2552 migrate->src[i] = 0;
2553 unlock_page(page);
2554 put_page(page);
2555 restore--;
2556 }
2557}
2558
2559/*
2560 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2561 * @migrate: migrate struct containing all migration information
2562 *
2563 * Replace page mapping (CPU page table pte) with a special migration pte entry
2564 * and check again if it has been pinned. Pinned pages are restored because we
2565 * cannot migrate them.
2566 *
2567 * This is the last step before we call the device driver callback to allocate
2568 * destination memory and copy contents of original page over to new page.
2569 */
2570static void migrate_vma_unmap(struct migrate_vma *migrate)
2571{
2572 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2573 const unsigned long npages = migrate->npages;
2574 const unsigned long start = migrate->start;
2575 unsigned long addr, i, restore = 0;
2576
2577 for (i = 0; i < npages; i++) {
2578 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2579
2580 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2581 continue;
2582
2583 if (page_mapped(page)) {
2584 try_to_unmap(page, flags);
2585 if (page_mapped(page))
2586 goto restore;
2587 }
2588
2589 if (migrate_vma_check_page(page))
2590 continue;
2591
2592restore:
2593 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2594 migrate->cpages--;
2595 restore++;
2596 }
2597
2598 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2599 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2600
2601 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2602 continue;
2603
2604 remove_migration_ptes(page, page, false);
2605
2606 migrate->src[i] = 0;
2607 unlock_page(page);
2608 restore--;
2609
2610 if (is_zone_device_page(page))
2611 put_page(page);
2612 else
2613 putback_lru_page(page);
2614 }
2615}
2616
2617static void migrate_vma_insert_page(struct migrate_vma *migrate,
2618 unsigned long addr,
2619 struct page *page,
2620 unsigned long *src,
2621 unsigned long *dst)
2622{
2623 struct vm_area_struct *vma = migrate->vma;
2624 struct mm_struct *mm = vma->vm_mm;
2625 struct mem_cgroup *memcg;
2626 bool flush = false;
2627 spinlock_t *ptl;
2628 pte_t entry;
2629 pgd_t *pgdp;
2630 p4d_t *p4dp;
2631 pud_t *pudp;
2632 pmd_t *pmdp;
2633 pte_t *ptep;
2634
2635 /* Only allow populating anonymous memory */
2636 if (!vma_is_anonymous(vma))
2637 goto abort;
2638
2639 pgdp = pgd_offset(mm, addr);
2640 p4dp = p4d_alloc(mm, pgdp, addr);
2641 if (!p4dp)
2642 goto abort;
2643 pudp = pud_alloc(mm, p4dp, addr);
2644 if (!pudp)
2645 goto abort;
2646 pmdp = pmd_alloc(mm, pudp, addr);
2647 if (!pmdp)
2648 goto abort;
2649
2650 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2651 goto abort;
2652
2653 /*
2654 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2655 * pte_offset_map() on pmds where a huge pmd might be created
2656 * from a different thread.
2657 *
2658 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2659 * parallel threads are excluded by other means.
2660 *
2661 * Here we only have down_read(mmap_sem).
2662 */
2663 if (pte_alloc(mm, pmdp, addr))
2664 goto abort;
2665
2666 /* See the comment in pte_alloc_one_map() */
2667 if (unlikely(pmd_trans_unstable(pmdp)))
2668 goto abort;
2669
2670 if (unlikely(anon_vma_prepare(vma)))
2671 goto abort;
2672 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2673 goto abort;
2674
2675 /*
2676 * The memory barrier inside __SetPageUptodate makes sure that
2677 * preceding stores to the page contents become visible before
2678 * the set_pte_at() write.
2679 */
2680 __SetPageUptodate(page);
2681
2682 if (is_zone_device_page(page)) {
2683 if (is_device_private_page(page)) {
2684 swp_entry_t swp_entry;
2685
2686 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2687 entry = swp_entry_to_pte(swp_entry);
2688 } else if (is_device_public_page(page)) {
2689 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2690 if (vma->vm_flags & VM_WRITE)
2691 entry = pte_mkwrite(pte_mkdirty(entry));
2692 entry = pte_mkdevmap(entry);
2693 }
2694 } else {
2695 entry = mk_pte(page, vma->vm_page_prot);
2696 if (vma->vm_flags & VM_WRITE)
2697 entry = pte_mkwrite(pte_mkdirty(entry));
2698 }
2699
2700 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2701
2702 if (pte_present(*ptep)) {
2703 unsigned long pfn = pte_pfn(*ptep);
2704
2705 if (!is_zero_pfn(pfn)) {
2706 pte_unmap_unlock(ptep, ptl);
2707 mem_cgroup_cancel_charge(page, memcg, false);
2708 goto abort;
2709 }
2710 flush = true;
2711 } else if (!pte_none(*ptep)) {
2712 pte_unmap_unlock(ptep, ptl);
2713 mem_cgroup_cancel_charge(page, memcg, false);
2714 goto abort;
2715 }
2716
2717 /*
2718 * Check for usefaultfd but do not deliver the fault. Instead,
2719 * just back off.
2720 */
2721 if (userfaultfd_missing(vma)) {
2722 pte_unmap_unlock(ptep, ptl);
2723 mem_cgroup_cancel_charge(page, memcg, false);
2724 goto abort;
2725 }
2726
2727 inc_mm_counter(mm, MM_ANONPAGES);
2728 page_add_new_anon_rmap(page, vma, addr, false);
2729 mem_cgroup_commit_charge(page, memcg, false, false);
2730 if (!is_zone_device_page(page))
2731 lru_cache_add_active_or_unevictable(page, vma);
2732 get_page(page);
2733
2734 if (flush) {
2735 flush_cache_page(vma, addr, pte_pfn(*ptep));
2736 ptep_clear_flush_notify(vma, addr, ptep);
2737 set_pte_at_notify(mm, addr, ptep, entry);
2738 update_mmu_cache(vma, addr, ptep);
2739 } else {
2740 /* No need to invalidate - it was non-present before */
2741 set_pte_at(mm, addr, ptep, entry);
2742 update_mmu_cache(vma, addr, ptep);
2743 }
2744
2745 pte_unmap_unlock(ptep, ptl);
2746 *src = MIGRATE_PFN_MIGRATE;
2747 return;
2748
2749abort:
2750 *src &= ~MIGRATE_PFN_MIGRATE;
2751}
2752
2753/*
2754 * migrate_vma_pages() - migrate meta-data from src page to dst page
2755 * @migrate: migrate struct containing all migration information
2756 *
2757 * This migrates struct page meta-data from source struct page to destination
2758 * struct page. This effectively finishes the migration from source page to the
2759 * destination page.
2760 */
2761static void migrate_vma_pages(struct migrate_vma *migrate)
2762{
2763 const unsigned long npages = migrate->npages;
2764 const unsigned long start = migrate->start;
2765 struct vm_area_struct *vma = migrate->vma;
2766 struct mm_struct *mm = vma->vm_mm;
2767 unsigned long addr, i, mmu_start;
2768 bool notified = false;
2769
2770 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2771 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2772 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2773 struct address_space *mapping;
2774 int r;
2775
2776 if (!newpage) {
2777 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2778 continue;
2779 }
2780
2781 if (!page) {
2782 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2783 continue;
2784 }
2785 if (!notified) {
2786 mmu_start = addr;
2787 notified = true;
2788 mmu_notifier_invalidate_range_start(mm,
2789 mmu_start,
2790 migrate->end);
2791 }
2792 migrate_vma_insert_page(migrate, addr, newpage,
2793 &migrate->src[i],
2794 &migrate->dst[i]);
2795 continue;
2796 }
2797
2798 mapping = page_mapping(page);
2799
2800 if (is_zone_device_page(newpage)) {
2801 if (is_device_private_page(newpage)) {
2802 /*
2803 * For now only support private anonymous when
2804 * migrating to un-addressable device memory.
2805 */
2806 if (mapping) {
2807 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2808 continue;
2809 }
2810 } else if (!is_device_public_page(newpage)) {
2811 /*
2812 * Other types of ZONE_DEVICE page are not
2813 * supported.
2814 */
2815 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2816 continue;
2817 }
2818 }
2819
2820 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2821 if (r != MIGRATEPAGE_SUCCESS)
2822 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2823 }
2824
2825 /*
2826 * No need to double call mmu_notifier->invalidate_range() callback as
2827 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2828 * did already call it.
2829 */
2830 if (notified)
2831 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2832 migrate->end);
2833}
2834
2835/*
2836 * migrate_vma_finalize() - restore CPU page table entry
2837 * @migrate: migrate struct containing all migration information
2838 *
2839 * This replaces the special migration pte entry with either a mapping to the
2840 * new page if migration was successful for that page, or to the original page
2841 * otherwise.
2842 *
2843 * This also unlocks the pages and puts them back on the lru, or drops the extra
2844 * refcount, for device pages.
2845 */
2846static void migrate_vma_finalize(struct migrate_vma *migrate)
2847{
2848 const unsigned long npages = migrate->npages;
2849 unsigned long i;
2850
2851 for (i = 0; i < npages; i++) {
2852 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2853 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2854
2855 if (!page) {
2856 if (newpage) {
2857 unlock_page(newpage);
2858 put_page(newpage);
2859 }
2860 continue;
2861 }
2862
2863 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2864 if (newpage) {
2865 unlock_page(newpage);
2866 put_page(newpage);
2867 }
2868 newpage = page;
2869 }
2870
2871 remove_migration_ptes(page, newpage, false);
2872 unlock_page(page);
2873 migrate->cpages--;
2874
2875 if (is_zone_device_page(page))
2876 put_page(page);
2877 else
2878 putback_lru_page(page);
2879
2880 if (newpage != page) {
2881 unlock_page(newpage);
2882 if (is_zone_device_page(newpage))
2883 put_page(newpage);
2884 else
2885 putback_lru_page(newpage);
2886 }
2887 }
2888}
2889
2890/*
2891 * migrate_vma() - migrate a range of memory inside vma
2892 *
2893 * @ops: migration callback for allocating destination memory and copying
2894 * @vma: virtual memory area containing the range to be migrated
2895 * @start: start address of the range to migrate (inclusive)
2896 * @end: end address of the range to migrate (exclusive)
2897 * @src: array of hmm_pfn_t containing source pfns
2898 * @dst: array of hmm_pfn_t containing destination pfns
2899 * @private: pointer passed back to each of the callback
2900 * Returns: 0 on success, error code otherwise
2901 *
2902 * This function tries to migrate a range of memory virtual address range, using
2903 * callbacks to allocate and copy memory from source to destination. First it
2904 * collects all the pages backing each virtual address in the range, saving this
2905 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2906 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2907 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2908 * in the corresponding src array entry. It then restores any pages that are
2909 * pinned, by remapping and unlocking those pages.
2910 *
2911 * At this point it calls the alloc_and_copy() callback. For documentation on
2912 * what is expected from that callback, see struct migrate_vma_ops comments in
2913 * include/linux/migrate.h
2914 *
2915 * After the alloc_and_copy() callback, this function goes over each entry in
2916 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2917 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2918 * then the function tries to migrate struct page information from the source
2919 * struct page to the destination struct page. If it fails to migrate the struct
2920 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2921 * array.
2922 *
2923 * At this point all successfully migrated pages have an entry in the src
2924 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2925 * array entry with MIGRATE_PFN_VALID flag set.
2926 *
2927 * It then calls the finalize_and_map() callback. See comments for "struct
2928 * migrate_vma_ops", in include/linux/migrate.h for details about
2929 * finalize_and_map() behavior.
2930 *
2931 * After the finalize_and_map() callback, for successfully migrated pages, this
2932 * function updates the CPU page table to point to new pages, otherwise it
2933 * restores the CPU page table to point to the original source pages.
2934 *
2935 * Function returns 0 after the above steps, even if no pages were migrated
2936 * (The function only returns an error if any of the arguments are invalid.)
2937 *
2938 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2939 * unsigned long entries.
2940 */
2941int migrate_vma(const struct migrate_vma_ops *ops,
2942 struct vm_area_struct *vma,
2943 unsigned long start,
2944 unsigned long end,
2945 unsigned long *src,
2946 unsigned long *dst,
2947 void *private)
2948{
2949 struct migrate_vma migrate;
2950
2951 /* Sanity check the arguments */
2952 start &= PAGE_MASK;
2953 end &= PAGE_MASK;
2954 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL))
2955 return -EINVAL;
2956 if (start < vma->vm_start || start >= vma->vm_end)
2957 return -EINVAL;
2958 if (end <= vma->vm_start || end > vma->vm_end)
2959 return -EINVAL;
2960 if (!ops || !src || !dst || start >= end)
2961 return -EINVAL;
2962
2963 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2964 migrate.src = src;
2965 migrate.dst = dst;
2966 migrate.start = start;
2967 migrate.npages = 0;
2968 migrate.cpages = 0;
2969 migrate.end = end;
2970 migrate.vma = vma;
2971
2972 /* Collect, and try to unmap source pages */
2973 migrate_vma_collect(&migrate);
2974 if (!migrate.cpages)
2975 return 0;
2976
2977 /* Lock and isolate page */
2978 migrate_vma_prepare(&migrate);
2979 if (!migrate.cpages)
2980 return 0;
2981
2982 /* Unmap pages */
2983 migrate_vma_unmap(&migrate);
2984 if (!migrate.cpages)
2985 return 0;
2986
2987 /*
2988 * At this point pages are locked and unmapped, and thus they have
2989 * stable content and can safely be copied to destination memory that
2990 * is allocated by the callback.
2991 *
2992 * Note that migration can fail in migrate_vma_struct_page() for each
2993 * individual page.
2994 */
2995 ops->alloc_and_copy(vma, src, dst, start, end, private);
2996
2997 /* This does the real migration of struct page */
2998 migrate_vma_pages(&migrate);
2999
3000 ops->finalize_and_map(vma, src, dst, start, end, private);
3001
3002 /* Unlock and remap pages */
3003 migrate_vma_finalize(&migrate);
3004
3005 return 0;
3006}
3007EXPORT_SYMBOL(migrate_vma);
3008#endif /* defined(MIGRATE_VMA_HELPER) */
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/ksm.h>
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
29#include <linux/writeback.h>
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
32#include <linux/security.h>
33#include <linux/backing-dev.h>
34#include <linux/compaction.h>
35#include <linux/syscalls.h>
36#include <linux/compat.h>
37#include <linux/hugetlb.h>
38#include <linux/hugetlb_cgroup.h>
39#include <linux/gfp.h>
40#include <linux/pfn_t.h>
41#include <linux/memremap.h>
42#include <linux/userfaultfd_k.h>
43#include <linux/balloon_compaction.h>
44#include <linux/page_idle.h>
45#include <linux/page_owner.h>
46#include <linux/sched/mm.h>
47#include <linux/ptrace.h>
48#include <linux/oom.h>
49#include <linux/memory.h>
50#include <linux/random.h>
51#include <linux/sched/sysctl.h>
52#include <linux/memory-tiers.h>
53
54#include <asm/tlbflush.h>
55
56#include <trace/events/migrate.h>
57
58#include "internal.h"
59
60bool isolate_movable_page(struct page *page, isolate_mode_t mode)
61{
62 struct folio *folio = folio_get_nontail_page(page);
63 const struct movable_operations *mops;
64
65 /*
66 * Avoid burning cycles with pages that are yet under __free_pages(),
67 * or just got freed under us.
68 *
69 * In case we 'win' a race for a movable page being freed under us and
70 * raise its refcount preventing __free_pages() from doing its job
71 * the put_page() at the end of this block will take care of
72 * release this page, thus avoiding a nasty leakage.
73 */
74 if (!folio)
75 goto out;
76
77 if (unlikely(folio_test_slab(folio)))
78 goto out_putfolio;
79 /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
80 smp_rmb();
81 /*
82 * Check movable flag before taking the page lock because
83 * we use non-atomic bitops on newly allocated page flags so
84 * unconditionally grabbing the lock ruins page's owner side.
85 */
86 if (unlikely(!__folio_test_movable(folio)))
87 goto out_putfolio;
88 /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
89 smp_rmb();
90 if (unlikely(folio_test_slab(folio)))
91 goto out_putfolio;
92
93 /*
94 * As movable pages are not isolated from LRU lists, concurrent
95 * compaction threads can race against page migration functions
96 * as well as race against the releasing a page.
97 *
98 * In order to avoid having an already isolated movable page
99 * being (wrongly) re-isolated while it is under migration,
100 * or to avoid attempting to isolate pages being released,
101 * lets be sure we have the page lock
102 * before proceeding with the movable page isolation steps.
103 */
104 if (unlikely(!folio_trylock(folio)))
105 goto out_putfolio;
106
107 if (!folio_test_movable(folio) || folio_test_isolated(folio))
108 goto out_no_isolated;
109
110 mops = folio_movable_ops(folio);
111 VM_BUG_ON_FOLIO(!mops, folio);
112
113 if (!mops->isolate_page(&folio->page, mode))
114 goto out_no_isolated;
115
116 /* Driver shouldn't use PG_isolated bit of page->flags */
117 WARN_ON_ONCE(folio_test_isolated(folio));
118 folio_set_isolated(folio);
119 folio_unlock(folio);
120
121 return true;
122
123out_no_isolated:
124 folio_unlock(folio);
125out_putfolio:
126 folio_put(folio);
127out:
128 return false;
129}
130
131static void putback_movable_folio(struct folio *folio)
132{
133 const struct movable_operations *mops = folio_movable_ops(folio);
134
135 mops->putback_page(&folio->page);
136 folio_clear_isolated(folio);
137}
138
139/*
140 * Put previously isolated pages back onto the appropriate lists
141 * from where they were once taken off for compaction/migration.
142 *
143 * This function shall be used whenever the isolated pageset has been
144 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
145 * and isolate_hugetlb().
146 */
147void putback_movable_pages(struct list_head *l)
148{
149 struct folio *folio;
150 struct folio *folio2;
151
152 list_for_each_entry_safe(folio, folio2, l, lru) {
153 if (unlikely(folio_test_hugetlb(folio))) {
154 folio_putback_active_hugetlb(folio);
155 continue;
156 }
157 list_del(&folio->lru);
158 /*
159 * We isolated non-lru movable folio so here we can use
160 * __folio_test_movable because LRU folio's mapping cannot
161 * have PAGE_MAPPING_MOVABLE.
162 */
163 if (unlikely(__folio_test_movable(folio))) {
164 VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
165 folio_lock(folio);
166 if (folio_test_movable(folio))
167 putback_movable_folio(folio);
168 else
169 folio_clear_isolated(folio);
170 folio_unlock(folio);
171 folio_put(folio);
172 } else {
173 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
174 folio_is_file_lru(folio), -folio_nr_pages(folio));
175 folio_putback_lru(folio);
176 }
177 }
178}
179
180/*
181 * Restore a potential migration pte to a working pte entry
182 */
183static bool remove_migration_pte(struct folio *folio,
184 struct vm_area_struct *vma, unsigned long addr, void *old)
185{
186 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
187
188 while (page_vma_mapped_walk(&pvmw)) {
189 rmap_t rmap_flags = RMAP_NONE;
190 pte_t old_pte;
191 pte_t pte;
192 swp_entry_t entry;
193 struct page *new;
194 unsigned long idx = 0;
195
196 /* pgoff is invalid for ksm pages, but they are never large */
197 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
198 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
199 new = folio_page(folio, idx);
200
201#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
202 /* PMD-mapped THP migration entry */
203 if (!pvmw.pte) {
204 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
205 !folio_test_pmd_mappable(folio), folio);
206 remove_migration_pmd(&pvmw, new);
207 continue;
208 }
209#endif
210
211 folio_get(folio);
212 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
213 old_pte = ptep_get(pvmw.pte);
214
215 entry = pte_to_swp_entry(old_pte);
216 if (!is_migration_entry_young(entry))
217 pte = pte_mkold(pte);
218 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
219 pte = pte_mkdirty(pte);
220 if (pte_swp_soft_dirty(old_pte))
221 pte = pte_mksoft_dirty(pte);
222 else
223 pte = pte_clear_soft_dirty(pte);
224
225 if (is_writable_migration_entry(entry))
226 pte = pte_mkwrite(pte, vma);
227 else if (pte_swp_uffd_wp(old_pte))
228 pte = pte_mkuffd_wp(pte);
229
230 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
231 rmap_flags |= RMAP_EXCLUSIVE;
232
233 if (unlikely(is_device_private_page(new))) {
234 if (pte_write(pte))
235 entry = make_writable_device_private_entry(
236 page_to_pfn(new));
237 else
238 entry = make_readable_device_private_entry(
239 page_to_pfn(new));
240 pte = swp_entry_to_pte(entry);
241 if (pte_swp_soft_dirty(old_pte))
242 pte = pte_swp_mksoft_dirty(pte);
243 if (pte_swp_uffd_wp(old_pte))
244 pte = pte_swp_mkuffd_wp(pte);
245 }
246
247#ifdef CONFIG_HUGETLB_PAGE
248 if (folio_test_hugetlb(folio)) {
249 struct hstate *h = hstate_vma(vma);
250 unsigned int shift = huge_page_shift(h);
251 unsigned long psize = huge_page_size(h);
252
253 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
254 if (folio_test_anon(folio))
255 hugetlb_add_anon_rmap(folio, vma, pvmw.address,
256 rmap_flags);
257 else
258 hugetlb_add_file_rmap(folio);
259 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
260 psize);
261 } else
262#endif
263 {
264 if (folio_test_anon(folio))
265 folio_add_anon_rmap_pte(folio, new, vma,
266 pvmw.address, rmap_flags);
267 else
268 folio_add_file_rmap_pte(folio, new, vma);
269 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
270 }
271 if (vma->vm_flags & VM_LOCKED)
272 mlock_drain_local();
273
274 trace_remove_migration_pte(pvmw.address, pte_val(pte),
275 compound_order(new));
276
277 /* No need to invalidate - it was non-present before */
278 update_mmu_cache(vma, pvmw.address, pvmw.pte);
279 }
280
281 return true;
282}
283
284/*
285 * Get rid of all migration entries and replace them by
286 * references to the indicated page.
287 */
288void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
289{
290 struct rmap_walk_control rwc = {
291 .rmap_one = remove_migration_pte,
292 .arg = src,
293 };
294
295 if (locked)
296 rmap_walk_locked(dst, &rwc);
297 else
298 rmap_walk(dst, &rwc);
299}
300
301/*
302 * Something used the pte of a page under migration. We need to
303 * get to the page and wait until migration is finished.
304 * When we return from this function the fault will be retried.
305 */
306void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
307 unsigned long address)
308{
309 spinlock_t *ptl;
310 pte_t *ptep;
311 pte_t pte;
312 swp_entry_t entry;
313
314 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
315 if (!ptep)
316 return;
317
318 pte = ptep_get(ptep);
319 pte_unmap(ptep);
320
321 if (!is_swap_pte(pte))
322 goto out;
323
324 entry = pte_to_swp_entry(pte);
325 if (!is_migration_entry(entry))
326 goto out;
327
328 migration_entry_wait_on_locked(entry, ptl);
329 return;
330out:
331 spin_unlock(ptl);
332}
333
334#ifdef CONFIG_HUGETLB_PAGE
335/*
336 * The vma read lock must be held upon entry. Holding that lock prevents either
337 * the pte or the ptl from being freed.
338 *
339 * This function will release the vma lock before returning.
340 */
341void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
342{
343 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
344 pte_t pte;
345
346 hugetlb_vma_assert_locked(vma);
347 spin_lock(ptl);
348 pte = huge_ptep_get(ptep);
349
350 if (unlikely(!is_hugetlb_entry_migration(pte))) {
351 spin_unlock(ptl);
352 hugetlb_vma_unlock_read(vma);
353 } else {
354 /*
355 * If migration entry existed, safe to release vma lock
356 * here because the pgtable page won't be freed without the
357 * pgtable lock released. See comment right above pgtable
358 * lock release in migration_entry_wait_on_locked().
359 */
360 hugetlb_vma_unlock_read(vma);
361 migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
362 }
363}
364#endif
365
366#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
367void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
368{
369 spinlock_t *ptl;
370
371 ptl = pmd_lock(mm, pmd);
372 if (!is_pmd_migration_entry(*pmd))
373 goto unlock;
374 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
375 return;
376unlock:
377 spin_unlock(ptl);
378}
379#endif
380
381static int folio_expected_refs(struct address_space *mapping,
382 struct folio *folio)
383{
384 int refs = 1;
385 if (!mapping)
386 return refs;
387
388 refs += folio_nr_pages(folio);
389 if (folio_test_private(folio))
390 refs++;
391
392 return refs;
393}
394
395/*
396 * Replace the page in the mapping.
397 *
398 * The number of remaining references must be:
399 * 1 for anonymous pages without a mapping
400 * 2 for pages with a mapping
401 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
402 */
403int folio_migrate_mapping(struct address_space *mapping,
404 struct folio *newfolio, struct folio *folio, int extra_count)
405{
406 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
407 struct zone *oldzone, *newzone;
408 int dirty;
409 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
410 long nr = folio_nr_pages(folio);
411 long entries, i;
412
413 if (!mapping) {
414 /* Anonymous page without mapping */
415 if (folio_ref_count(folio) != expected_count)
416 return -EAGAIN;
417
418 /* No turning back from here */
419 newfolio->index = folio->index;
420 newfolio->mapping = folio->mapping;
421 if (folio_test_swapbacked(folio))
422 __folio_set_swapbacked(newfolio);
423
424 return MIGRATEPAGE_SUCCESS;
425 }
426
427 oldzone = folio_zone(folio);
428 newzone = folio_zone(newfolio);
429
430 xas_lock_irq(&xas);
431 if (!folio_ref_freeze(folio, expected_count)) {
432 xas_unlock_irq(&xas);
433 return -EAGAIN;
434 }
435
436 /*
437 * Now we know that no one else is looking at the folio:
438 * no turning back from here.
439 */
440 newfolio->index = folio->index;
441 newfolio->mapping = folio->mapping;
442 folio_ref_add(newfolio, nr); /* add cache reference */
443 if (folio_test_swapbacked(folio)) {
444 __folio_set_swapbacked(newfolio);
445 if (folio_test_swapcache(folio)) {
446 folio_set_swapcache(newfolio);
447 newfolio->private = folio_get_private(folio);
448 }
449 entries = nr;
450 } else {
451 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
452 entries = 1;
453 }
454
455 /* Move dirty while page refs frozen and newpage not yet exposed */
456 dirty = folio_test_dirty(folio);
457 if (dirty) {
458 folio_clear_dirty(folio);
459 folio_set_dirty(newfolio);
460 }
461
462 /* Swap cache still stores N entries instead of a high-order entry */
463 for (i = 0; i < entries; i++) {
464 xas_store(&xas, newfolio);
465 xas_next(&xas);
466 }
467
468 /*
469 * Drop cache reference from old page by unfreezing
470 * to one less reference.
471 * We know this isn't the last reference.
472 */
473 folio_ref_unfreeze(folio, expected_count - nr);
474
475 xas_unlock(&xas);
476 /* Leave irq disabled to prevent preemption while updating stats */
477
478 /*
479 * If moved to a different zone then also account
480 * the page for that zone. Other VM counters will be
481 * taken care of when we establish references to the
482 * new page and drop references to the old page.
483 *
484 * Note that anonymous pages are accounted for
485 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
486 * are mapped to swap space.
487 */
488 if (newzone != oldzone) {
489 struct lruvec *old_lruvec, *new_lruvec;
490 struct mem_cgroup *memcg;
491
492 memcg = folio_memcg(folio);
493 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
494 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
495
496 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
497 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
498 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
499 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
500 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
501
502 if (folio_test_pmd_mappable(folio)) {
503 __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
504 __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
505 }
506 }
507#ifdef CONFIG_SWAP
508 if (folio_test_swapcache(folio)) {
509 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
510 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
511 }
512#endif
513 if (dirty && mapping_can_writeback(mapping)) {
514 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
515 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
516 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
517 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
518 }
519 }
520 local_irq_enable();
521
522 return MIGRATEPAGE_SUCCESS;
523}
524EXPORT_SYMBOL(folio_migrate_mapping);
525
526/*
527 * The expected number of remaining references is the same as that
528 * of folio_migrate_mapping().
529 */
530int migrate_huge_page_move_mapping(struct address_space *mapping,
531 struct folio *dst, struct folio *src)
532{
533 XA_STATE(xas, &mapping->i_pages, folio_index(src));
534 int expected_count;
535
536 xas_lock_irq(&xas);
537 expected_count = folio_expected_refs(mapping, src);
538 if (!folio_ref_freeze(src, expected_count)) {
539 xas_unlock_irq(&xas);
540 return -EAGAIN;
541 }
542
543 dst->index = src->index;
544 dst->mapping = src->mapping;
545
546 folio_ref_add(dst, folio_nr_pages(dst));
547
548 xas_store(&xas, dst);
549
550 folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
551
552 xas_unlock_irq(&xas);
553
554 return MIGRATEPAGE_SUCCESS;
555}
556
557/*
558 * Copy the flags and some other ancillary information
559 */
560void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
561{
562 int cpupid;
563
564 if (folio_test_error(folio))
565 folio_set_error(newfolio);
566 if (folio_test_referenced(folio))
567 folio_set_referenced(newfolio);
568 if (folio_test_uptodate(folio))
569 folio_mark_uptodate(newfolio);
570 if (folio_test_clear_active(folio)) {
571 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
572 folio_set_active(newfolio);
573 } else if (folio_test_clear_unevictable(folio))
574 folio_set_unevictable(newfolio);
575 if (folio_test_workingset(folio))
576 folio_set_workingset(newfolio);
577 if (folio_test_checked(folio))
578 folio_set_checked(newfolio);
579 /*
580 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
581 * migration entries. We can still have PG_anon_exclusive set on an
582 * effectively unmapped and unreferenced first sub-pages of an
583 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
584 */
585 if (folio_test_mappedtodisk(folio))
586 folio_set_mappedtodisk(newfolio);
587
588 /* Move dirty on pages not done by folio_migrate_mapping() */
589 if (folio_test_dirty(folio))
590 folio_set_dirty(newfolio);
591
592 if (folio_test_young(folio))
593 folio_set_young(newfolio);
594 if (folio_test_idle(folio))
595 folio_set_idle(newfolio);
596
597 /*
598 * Copy NUMA information to the new page, to prevent over-eager
599 * future migrations of this same page.
600 */
601 cpupid = folio_xchg_last_cpupid(folio, -1);
602 /*
603 * For memory tiering mode, when migrate between slow and fast
604 * memory node, reset cpupid, because that is used to record
605 * page access time in slow memory node.
606 */
607 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
608 bool f_toptier = node_is_toptier(folio_nid(folio));
609 bool t_toptier = node_is_toptier(folio_nid(newfolio));
610
611 if (f_toptier != t_toptier)
612 cpupid = -1;
613 }
614 folio_xchg_last_cpupid(newfolio, cpupid);
615
616 folio_migrate_ksm(newfolio, folio);
617 /*
618 * Please do not reorder this without considering how mm/ksm.c's
619 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
620 */
621 if (folio_test_swapcache(folio))
622 folio_clear_swapcache(folio);
623 folio_clear_private(folio);
624
625 /* page->private contains hugetlb specific flags */
626 if (!folio_test_hugetlb(folio))
627 folio->private = NULL;
628
629 /*
630 * If any waiters have accumulated on the new page then
631 * wake them up.
632 */
633 if (folio_test_writeback(newfolio))
634 folio_end_writeback(newfolio);
635
636 /*
637 * PG_readahead shares the same bit with PG_reclaim. The above
638 * end_page_writeback() may clear PG_readahead mistakenly, so set the
639 * bit after that.
640 */
641 if (folio_test_readahead(folio))
642 folio_set_readahead(newfolio);
643
644 folio_copy_owner(newfolio, folio);
645
646 mem_cgroup_migrate(folio, newfolio);
647}
648EXPORT_SYMBOL(folio_migrate_flags);
649
650void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
651{
652 folio_copy(newfolio, folio);
653 folio_migrate_flags(newfolio, folio);
654}
655EXPORT_SYMBOL(folio_migrate_copy);
656
657/************************************************************
658 * Migration functions
659 ***********************************************************/
660
661int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
662 struct folio *src, enum migrate_mode mode, int extra_count)
663{
664 int rc;
665
666 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
667
668 rc = folio_migrate_mapping(mapping, dst, src, extra_count);
669
670 if (rc != MIGRATEPAGE_SUCCESS)
671 return rc;
672
673 if (mode != MIGRATE_SYNC_NO_COPY)
674 folio_migrate_copy(dst, src);
675 else
676 folio_migrate_flags(dst, src);
677 return MIGRATEPAGE_SUCCESS;
678}
679
680/**
681 * migrate_folio() - Simple folio migration.
682 * @mapping: The address_space containing the folio.
683 * @dst: The folio to migrate the data to.
684 * @src: The folio containing the current data.
685 * @mode: How to migrate the page.
686 *
687 * Common logic to directly migrate a single LRU folio suitable for
688 * folios that do not use PagePrivate/PagePrivate2.
689 *
690 * Folios are locked upon entry and exit.
691 */
692int migrate_folio(struct address_space *mapping, struct folio *dst,
693 struct folio *src, enum migrate_mode mode)
694{
695 return migrate_folio_extra(mapping, dst, src, mode, 0);
696}
697EXPORT_SYMBOL(migrate_folio);
698
699#ifdef CONFIG_BUFFER_HEAD
700/* Returns true if all buffers are successfully locked */
701static bool buffer_migrate_lock_buffers(struct buffer_head *head,
702 enum migrate_mode mode)
703{
704 struct buffer_head *bh = head;
705 struct buffer_head *failed_bh;
706
707 do {
708 if (!trylock_buffer(bh)) {
709 if (mode == MIGRATE_ASYNC)
710 goto unlock;
711 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
712 goto unlock;
713 lock_buffer(bh);
714 }
715
716 bh = bh->b_this_page;
717 } while (bh != head);
718
719 return true;
720
721unlock:
722 /* We failed to lock the buffer and cannot stall. */
723 failed_bh = bh;
724 bh = head;
725 while (bh != failed_bh) {
726 unlock_buffer(bh);
727 bh = bh->b_this_page;
728 }
729
730 return false;
731}
732
733static int __buffer_migrate_folio(struct address_space *mapping,
734 struct folio *dst, struct folio *src, enum migrate_mode mode,
735 bool check_refs)
736{
737 struct buffer_head *bh, *head;
738 int rc;
739 int expected_count;
740
741 head = folio_buffers(src);
742 if (!head)
743 return migrate_folio(mapping, dst, src, mode);
744
745 /* Check whether page does not have extra refs before we do more work */
746 expected_count = folio_expected_refs(mapping, src);
747 if (folio_ref_count(src) != expected_count)
748 return -EAGAIN;
749
750 if (!buffer_migrate_lock_buffers(head, mode))
751 return -EAGAIN;
752
753 if (check_refs) {
754 bool busy;
755 bool invalidated = false;
756
757recheck_buffers:
758 busy = false;
759 spin_lock(&mapping->i_private_lock);
760 bh = head;
761 do {
762 if (atomic_read(&bh->b_count)) {
763 busy = true;
764 break;
765 }
766 bh = bh->b_this_page;
767 } while (bh != head);
768 if (busy) {
769 if (invalidated) {
770 rc = -EAGAIN;
771 goto unlock_buffers;
772 }
773 spin_unlock(&mapping->i_private_lock);
774 invalidate_bh_lrus();
775 invalidated = true;
776 goto recheck_buffers;
777 }
778 }
779
780 rc = folio_migrate_mapping(mapping, dst, src, 0);
781 if (rc != MIGRATEPAGE_SUCCESS)
782 goto unlock_buffers;
783
784 folio_attach_private(dst, folio_detach_private(src));
785
786 bh = head;
787 do {
788 folio_set_bh(bh, dst, bh_offset(bh));
789 bh = bh->b_this_page;
790 } while (bh != head);
791
792 if (mode != MIGRATE_SYNC_NO_COPY)
793 folio_migrate_copy(dst, src);
794 else
795 folio_migrate_flags(dst, src);
796
797 rc = MIGRATEPAGE_SUCCESS;
798unlock_buffers:
799 if (check_refs)
800 spin_unlock(&mapping->i_private_lock);
801 bh = head;
802 do {
803 unlock_buffer(bh);
804 bh = bh->b_this_page;
805 } while (bh != head);
806
807 return rc;
808}
809
810/**
811 * buffer_migrate_folio() - Migration function for folios with buffers.
812 * @mapping: The address space containing @src.
813 * @dst: The folio to migrate to.
814 * @src: The folio to migrate from.
815 * @mode: How to migrate the folio.
816 *
817 * This function can only be used if the underlying filesystem guarantees
818 * that no other references to @src exist. For example attached buffer
819 * heads are accessed only under the folio lock. If your filesystem cannot
820 * provide this guarantee, buffer_migrate_folio_norefs() may be more
821 * appropriate.
822 *
823 * Return: 0 on success or a negative errno on failure.
824 */
825int buffer_migrate_folio(struct address_space *mapping,
826 struct folio *dst, struct folio *src, enum migrate_mode mode)
827{
828 return __buffer_migrate_folio(mapping, dst, src, mode, false);
829}
830EXPORT_SYMBOL(buffer_migrate_folio);
831
832/**
833 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
834 * @mapping: The address space containing @src.
835 * @dst: The folio to migrate to.
836 * @src: The folio to migrate from.
837 * @mode: How to migrate the folio.
838 *
839 * Like buffer_migrate_folio() except that this variant is more careful
840 * and checks that there are also no buffer head references. This function
841 * is the right one for mappings where buffer heads are directly looked
842 * up and referenced (such as block device mappings).
843 *
844 * Return: 0 on success or a negative errno on failure.
845 */
846int buffer_migrate_folio_norefs(struct address_space *mapping,
847 struct folio *dst, struct folio *src, enum migrate_mode mode)
848{
849 return __buffer_migrate_folio(mapping, dst, src, mode, true);
850}
851EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
852#endif /* CONFIG_BUFFER_HEAD */
853
854int filemap_migrate_folio(struct address_space *mapping,
855 struct folio *dst, struct folio *src, enum migrate_mode mode)
856{
857 int ret;
858
859 ret = folio_migrate_mapping(mapping, dst, src, 0);
860 if (ret != MIGRATEPAGE_SUCCESS)
861 return ret;
862
863 if (folio_get_private(src))
864 folio_attach_private(dst, folio_detach_private(src));
865
866 if (mode != MIGRATE_SYNC_NO_COPY)
867 folio_migrate_copy(dst, src);
868 else
869 folio_migrate_flags(dst, src);
870 return MIGRATEPAGE_SUCCESS;
871}
872EXPORT_SYMBOL_GPL(filemap_migrate_folio);
873
874/*
875 * Writeback a folio to clean the dirty state
876 */
877static int writeout(struct address_space *mapping, struct folio *folio)
878{
879 struct writeback_control wbc = {
880 .sync_mode = WB_SYNC_NONE,
881 .nr_to_write = 1,
882 .range_start = 0,
883 .range_end = LLONG_MAX,
884 .for_reclaim = 1
885 };
886 int rc;
887
888 if (!mapping->a_ops->writepage)
889 /* No write method for the address space */
890 return -EINVAL;
891
892 if (!folio_clear_dirty_for_io(folio))
893 /* Someone else already triggered a write */
894 return -EAGAIN;
895
896 /*
897 * A dirty folio may imply that the underlying filesystem has
898 * the folio on some queue. So the folio must be clean for
899 * migration. Writeout may mean we lose the lock and the
900 * folio state is no longer what we checked for earlier.
901 * At this point we know that the migration attempt cannot
902 * be successful.
903 */
904 remove_migration_ptes(folio, folio, false);
905
906 rc = mapping->a_ops->writepage(&folio->page, &wbc);
907
908 if (rc != AOP_WRITEPAGE_ACTIVATE)
909 /* unlocked. Relock */
910 folio_lock(folio);
911
912 return (rc < 0) ? -EIO : -EAGAIN;
913}
914
915/*
916 * Default handling if a filesystem does not provide a migration function.
917 */
918static int fallback_migrate_folio(struct address_space *mapping,
919 struct folio *dst, struct folio *src, enum migrate_mode mode)
920{
921 if (folio_test_dirty(src)) {
922 /* Only writeback folios in full synchronous migration */
923 switch (mode) {
924 case MIGRATE_SYNC:
925 case MIGRATE_SYNC_NO_COPY:
926 break;
927 default:
928 return -EBUSY;
929 }
930 return writeout(mapping, src);
931 }
932
933 /*
934 * Buffers may be managed in a filesystem specific way.
935 * We must have no buffers or drop them.
936 */
937 if (!filemap_release_folio(src, GFP_KERNEL))
938 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
939
940 return migrate_folio(mapping, dst, src, mode);
941}
942
943/*
944 * Move a page to a newly allocated page
945 * The page is locked and all ptes have been successfully removed.
946 *
947 * The new page will have replaced the old page if this function
948 * is successful.
949 *
950 * Return value:
951 * < 0 - error code
952 * MIGRATEPAGE_SUCCESS - success
953 */
954static int move_to_new_folio(struct folio *dst, struct folio *src,
955 enum migrate_mode mode)
956{
957 int rc = -EAGAIN;
958 bool is_lru = !__folio_test_movable(src);
959
960 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
961 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
962
963 if (likely(is_lru)) {
964 struct address_space *mapping = folio_mapping(src);
965
966 if (!mapping)
967 rc = migrate_folio(mapping, dst, src, mode);
968 else if (mapping_unmovable(mapping))
969 rc = -EOPNOTSUPP;
970 else if (mapping->a_ops->migrate_folio)
971 /*
972 * Most folios have a mapping and most filesystems
973 * provide a migrate_folio callback. Anonymous folios
974 * are part of swap space which also has its own
975 * migrate_folio callback. This is the most common path
976 * for page migration.
977 */
978 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
979 mode);
980 else
981 rc = fallback_migrate_folio(mapping, dst, src, mode);
982 } else {
983 const struct movable_operations *mops;
984
985 /*
986 * In case of non-lru page, it could be released after
987 * isolation step. In that case, we shouldn't try migration.
988 */
989 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
990 if (!folio_test_movable(src)) {
991 rc = MIGRATEPAGE_SUCCESS;
992 folio_clear_isolated(src);
993 goto out;
994 }
995
996 mops = folio_movable_ops(src);
997 rc = mops->migrate_page(&dst->page, &src->page, mode);
998 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
999 !folio_test_isolated(src));
1000 }
1001
1002 /*
1003 * When successful, old pagecache src->mapping must be cleared before
1004 * src is freed; but stats require that PageAnon be left as PageAnon.
1005 */
1006 if (rc == MIGRATEPAGE_SUCCESS) {
1007 if (__folio_test_movable(src)) {
1008 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1009
1010 /*
1011 * We clear PG_movable under page_lock so any compactor
1012 * cannot try to migrate this page.
1013 */
1014 folio_clear_isolated(src);
1015 }
1016
1017 /*
1018 * Anonymous and movable src->mapping will be cleared by
1019 * free_pages_prepare so don't reset it here for keeping
1020 * the type to work PageAnon, for example.
1021 */
1022 if (!folio_mapping_flags(src))
1023 src->mapping = NULL;
1024
1025 if (likely(!folio_is_zone_device(dst)))
1026 flush_dcache_folio(dst);
1027 }
1028out:
1029 return rc;
1030}
1031
1032/*
1033 * To record some information during migration, we use unused private
1034 * field of struct folio of the newly allocated destination folio.
1035 * This is safe because nobody is using it except us.
1036 */
1037enum {
1038 PAGE_WAS_MAPPED = BIT(0),
1039 PAGE_WAS_MLOCKED = BIT(1),
1040 PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1041};
1042
1043static void __migrate_folio_record(struct folio *dst,
1044 int old_page_state,
1045 struct anon_vma *anon_vma)
1046{
1047 dst->private = (void *)anon_vma + old_page_state;
1048}
1049
1050static void __migrate_folio_extract(struct folio *dst,
1051 int *old_page_state,
1052 struct anon_vma **anon_vmap)
1053{
1054 unsigned long private = (unsigned long)dst->private;
1055
1056 *anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1057 *old_page_state = private & PAGE_OLD_STATES;
1058 dst->private = NULL;
1059}
1060
1061/* Restore the source folio to the original state upon failure */
1062static void migrate_folio_undo_src(struct folio *src,
1063 int page_was_mapped,
1064 struct anon_vma *anon_vma,
1065 bool locked,
1066 struct list_head *ret)
1067{
1068 if (page_was_mapped)
1069 remove_migration_ptes(src, src, false);
1070 /* Drop an anon_vma reference if we took one */
1071 if (anon_vma)
1072 put_anon_vma(anon_vma);
1073 if (locked)
1074 folio_unlock(src);
1075 if (ret)
1076 list_move_tail(&src->lru, ret);
1077}
1078
1079/* Restore the destination folio to the original state upon failure */
1080static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1081 free_folio_t put_new_folio, unsigned long private)
1082{
1083 if (locked)
1084 folio_unlock(dst);
1085 if (put_new_folio)
1086 put_new_folio(dst, private);
1087 else
1088 folio_put(dst);
1089}
1090
1091/* Cleanup src folio upon migration success */
1092static void migrate_folio_done(struct folio *src,
1093 enum migrate_reason reason)
1094{
1095 /*
1096 * Compaction can migrate also non-LRU pages which are
1097 * not accounted to NR_ISOLATED_*. They can be recognized
1098 * as __folio_test_movable
1099 */
1100 if (likely(!__folio_test_movable(src)))
1101 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1102 folio_is_file_lru(src), -folio_nr_pages(src));
1103
1104 if (reason != MR_MEMORY_FAILURE)
1105 /* We release the page in page_handle_poison. */
1106 folio_put(src);
1107}
1108
1109/* Obtain the lock on page, remove all ptes. */
1110static int migrate_folio_unmap(new_folio_t get_new_folio,
1111 free_folio_t put_new_folio, unsigned long private,
1112 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1113 enum migrate_reason reason, struct list_head *ret)
1114{
1115 struct folio *dst;
1116 int rc = -EAGAIN;
1117 int old_page_state = 0;
1118 struct anon_vma *anon_vma = NULL;
1119 bool is_lru = !__folio_test_movable(src);
1120 bool locked = false;
1121 bool dst_locked = false;
1122
1123 if (folio_ref_count(src) == 1) {
1124 /* Folio was freed from under us. So we are done. */
1125 folio_clear_active(src);
1126 folio_clear_unevictable(src);
1127 /* free_pages_prepare() will clear PG_isolated. */
1128 list_del(&src->lru);
1129 migrate_folio_done(src, reason);
1130 return MIGRATEPAGE_SUCCESS;
1131 }
1132
1133 dst = get_new_folio(src, private);
1134 if (!dst)
1135 return -ENOMEM;
1136 *dstp = dst;
1137
1138 dst->private = NULL;
1139
1140 if (!folio_trylock(src)) {
1141 if (mode == MIGRATE_ASYNC)
1142 goto out;
1143
1144 /*
1145 * It's not safe for direct compaction to call lock_page.
1146 * For example, during page readahead pages are added locked
1147 * to the LRU. Later, when the IO completes the pages are
1148 * marked uptodate and unlocked. However, the queueing
1149 * could be merging multiple pages for one bio (e.g.
1150 * mpage_readahead). If an allocation happens for the
1151 * second or third page, the process can end up locking
1152 * the same page twice and deadlocking. Rather than
1153 * trying to be clever about what pages can be locked,
1154 * avoid the use of lock_page for direct compaction
1155 * altogether.
1156 */
1157 if (current->flags & PF_MEMALLOC)
1158 goto out;
1159
1160 /*
1161 * In "light" mode, we can wait for transient locks (eg
1162 * inserting a page into the page table), but it's not
1163 * worth waiting for I/O.
1164 */
1165 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1166 goto out;
1167
1168 folio_lock(src);
1169 }
1170 locked = true;
1171 if (folio_test_mlocked(src))
1172 old_page_state |= PAGE_WAS_MLOCKED;
1173
1174 if (folio_test_writeback(src)) {
1175 /*
1176 * Only in the case of a full synchronous migration is it
1177 * necessary to wait for PageWriteback. In the async case,
1178 * the retry loop is too short and in the sync-light case,
1179 * the overhead of stalling is too much
1180 */
1181 switch (mode) {
1182 case MIGRATE_SYNC:
1183 case MIGRATE_SYNC_NO_COPY:
1184 break;
1185 default:
1186 rc = -EBUSY;
1187 goto out;
1188 }
1189 folio_wait_writeback(src);
1190 }
1191
1192 /*
1193 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1194 * we cannot notice that anon_vma is freed while we migrate a page.
1195 * This get_anon_vma() delays freeing anon_vma pointer until the end
1196 * of migration. File cache pages are no problem because of page_lock()
1197 * File Caches may use write_page() or lock_page() in migration, then,
1198 * just care Anon page here.
1199 *
1200 * Only folio_get_anon_vma() understands the subtleties of
1201 * getting a hold on an anon_vma from outside one of its mms.
1202 * But if we cannot get anon_vma, then we won't need it anyway,
1203 * because that implies that the anon page is no longer mapped
1204 * (and cannot be remapped so long as we hold the page lock).
1205 */
1206 if (folio_test_anon(src) && !folio_test_ksm(src))
1207 anon_vma = folio_get_anon_vma(src);
1208
1209 /*
1210 * Block others from accessing the new page when we get around to
1211 * establishing additional references. We are usually the only one
1212 * holding a reference to dst at this point. We used to have a BUG
1213 * here if folio_trylock(dst) fails, but would like to allow for
1214 * cases where there might be a race with the previous use of dst.
1215 * This is much like races on refcount of oldpage: just don't BUG().
1216 */
1217 if (unlikely(!folio_trylock(dst)))
1218 goto out;
1219 dst_locked = true;
1220
1221 if (unlikely(!is_lru)) {
1222 __migrate_folio_record(dst, old_page_state, anon_vma);
1223 return MIGRATEPAGE_UNMAP;
1224 }
1225
1226 /*
1227 * Corner case handling:
1228 * 1. When a new swap-cache page is read into, it is added to the LRU
1229 * and treated as swapcache but it has no rmap yet.
1230 * Calling try_to_unmap() against a src->mapping==NULL page will
1231 * trigger a BUG. So handle it here.
1232 * 2. An orphaned page (see truncate_cleanup_page) might have
1233 * fs-private metadata. The page can be picked up due to memory
1234 * offlining. Everywhere else except page reclaim, the page is
1235 * invisible to the vm, so the page can not be migrated. So try to
1236 * free the metadata, so the page can be freed.
1237 */
1238 if (!src->mapping) {
1239 if (folio_test_private(src)) {
1240 try_to_free_buffers(src);
1241 goto out;
1242 }
1243 } else if (folio_mapped(src)) {
1244 /* Establish migration ptes */
1245 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1246 !folio_test_ksm(src) && !anon_vma, src);
1247 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1248 old_page_state |= PAGE_WAS_MAPPED;
1249 }
1250
1251 if (!folio_mapped(src)) {
1252 __migrate_folio_record(dst, old_page_state, anon_vma);
1253 return MIGRATEPAGE_UNMAP;
1254 }
1255
1256out:
1257 /*
1258 * A folio that has not been unmapped will be restored to
1259 * right list unless we want to retry.
1260 */
1261 if (rc == -EAGAIN)
1262 ret = NULL;
1263
1264 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1265 anon_vma, locked, ret);
1266 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1267
1268 return rc;
1269}
1270
1271/* Migrate the folio to the newly allocated folio in dst. */
1272static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1273 struct folio *src, struct folio *dst,
1274 enum migrate_mode mode, enum migrate_reason reason,
1275 struct list_head *ret)
1276{
1277 int rc;
1278 int old_page_state = 0;
1279 struct anon_vma *anon_vma = NULL;
1280 bool is_lru = !__folio_test_movable(src);
1281 struct list_head *prev;
1282
1283 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1284 prev = dst->lru.prev;
1285 list_del(&dst->lru);
1286
1287 rc = move_to_new_folio(dst, src, mode);
1288 if (rc)
1289 goto out;
1290
1291 if (unlikely(!is_lru))
1292 goto out_unlock_both;
1293
1294 /*
1295 * When successful, push dst to LRU immediately: so that if it
1296 * turns out to be an mlocked page, remove_migration_ptes() will
1297 * automatically build up the correct dst->mlock_count for it.
1298 *
1299 * We would like to do something similar for the old page, when
1300 * unsuccessful, and other cases when a page has been temporarily
1301 * isolated from the unevictable LRU: but this case is the easiest.
1302 */
1303 folio_add_lru(dst);
1304 if (old_page_state & PAGE_WAS_MLOCKED)
1305 lru_add_drain();
1306
1307 if (old_page_state & PAGE_WAS_MAPPED)
1308 remove_migration_ptes(src, dst, false);
1309
1310out_unlock_both:
1311 folio_unlock(dst);
1312 set_page_owner_migrate_reason(&dst->page, reason);
1313 /*
1314 * If migration is successful, decrease refcount of dst,
1315 * which will not free the page because new page owner increased
1316 * refcounter.
1317 */
1318 folio_put(dst);
1319
1320 /*
1321 * A folio that has been migrated has all references removed
1322 * and will be freed.
1323 */
1324 list_del(&src->lru);
1325 /* Drop an anon_vma reference if we took one */
1326 if (anon_vma)
1327 put_anon_vma(anon_vma);
1328 folio_unlock(src);
1329 migrate_folio_done(src, reason);
1330
1331 return rc;
1332out:
1333 /*
1334 * A folio that has not been migrated will be restored to
1335 * right list unless we want to retry.
1336 */
1337 if (rc == -EAGAIN) {
1338 list_add(&dst->lru, prev);
1339 __migrate_folio_record(dst, old_page_state, anon_vma);
1340 return rc;
1341 }
1342
1343 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1344 anon_vma, true, ret);
1345 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1346
1347 return rc;
1348}
1349
1350/*
1351 * Counterpart of unmap_and_move_page() for hugepage migration.
1352 *
1353 * This function doesn't wait the completion of hugepage I/O
1354 * because there is no race between I/O and migration for hugepage.
1355 * Note that currently hugepage I/O occurs only in direct I/O
1356 * where no lock is held and PG_writeback is irrelevant,
1357 * and writeback status of all subpages are counted in the reference
1358 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1359 * under direct I/O, the reference of the head page is 512 and a bit more.)
1360 * This means that when we try to migrate hugepage whose subpages are
1361 * doing direct I/O, some references remain after try_to_unmap() and
1362 * hugepage migration fails without data corruption.
1363 *
1364 * There is also no race when direct I/O is issued on the page under migration,
1365 * because then pte is replaced with migration swap entry and direct I/O code
1366 * will wait in the page fault for migration to complete.
1367 */
1368static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1369 free_folio_t put_new_folio, unsigned long private,
1370 struct folio *src, int force, enum migrate_mode mode,
1371 int reason, struct list_head *ret)
1372{
1373 struct folio *dst;
1374 int rc = -EAGAIN;
1375 int page_was_mapped = 0;
1376 struct anon_vma *anon_vma = NULL;
1377 struct address_space *mapping = NULL;
1378
1379 if (folio_ref_count(src) == 1) {
1380 /* page was freed from under us. So we are done. */
1381 folio_putback_active_hugetlb(src);
1382 return MIGRATEPAGE_SUCCESS;
1383 }
1384
1385 dst = get_new_folio(src, private);
1386 if (!dst)
1387 return -ENOMEM;
1388
1389 if (!folio_trylock(src)) {
1390 if (!force)
1391 goto out;
1392 switch (mode) {
1393 case MIGRATE_SYNC:
1394 case MIGRATE_SYNC_NO_COPY:
1395 break;
1396 default:
1397 goto out;
1398 }
1399 folio_lock(src);
1400 }
1401
1402 /*
1403 * Check for pages which are in the process of being freed. Without
1404 * folio_mapping() set, hugetlbfs specific move page routine will not
1405 * be called and we could leak usage counts for subpools.
1406 */
1407 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1408 rc = -EBUSY;
1409 goto out_unlock;
1410 }
1411
1412 if (folio_test_anon(src))
1413 anon_vma = folio_get_anon_vma(src);
1414
1415 if (unlikely(!folio_trylock(dst)))
1416 goto put_anon;
1417
1418 if (folio_mapped(src)) {
1419 enum ttu_flags ttu = 0;
1420
1421 if (!folio_test_anon(src)) {
1422 /*
1423 * In shared mappings, try_to_unmap could potentially
1424 * call huge_pmd_unshare. Because of this, take
1425 * semaphore in write mode here and set TTU_RMAP_LOCKED
1426 * to let lower levels know we have taken the lock.
1427 */
1428 mapping = hugetlb_page_mapping_lock_write(&src->page);
1429 if (unlikely(!mapping))
1430 goto unlock_put_anon;
1431
1432 ttu = TTU_RMAP_LOCKED;
1433 }
1434
1435 try_to_migrate(src, ttu);
1436 page_was_mapped = 1;
1437
1438 if (ttu & TTU_RMAP_LOCKED)
1439 i_mmap_unlock_write(mapping);
1440 }
1441
1442 if (!folio_mapped(src))
1443 rc = move_to_new_folio(dst, src, mode);
1444
1445 if (page_was_mapped)
1446 remove_migration_ptes(src,
1447 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1448
1449unlock_put_anon:
1450 folio_unlock(dst);
1451
1452put_anon:
1453 if (anon_vma)
1454 put_anon_vma(anon_vma);
1455
1456 if (rc == MIGRATEPAGE_SUCCESS) {
1457 move_hugetlb_state(src, dst, reason);
1458 put_new_folio = NULL;
1459 }
1460
1461out_unlock:
1462 folio_unlock(src);
1463out:
1464 if (rc == MIGRATEPAGE_SUCCESS)
1465 folio_putback_active_hugetlb(src);
1466 else if (rc != -EAGAIN)
1467 list_move_tail(&src->lru, ret);
1468
1469 /*
1470 * If migration was not successful and there's a freeing callback, use
1471 * it. Otherwise, put_page() will drop the reference grabbed during
1472 * isolation.
1473 */
1474 if (put_new_folio)
1475 put_new_folio(dst, private);
1476 else
1477 folio_putback_active_hugetlb(dst);
1478
1479 return rc;
1480}
1481
1482static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1483{
1484 int rc;
1485
1486 folio_lock(folio);
1487 rc = split_folio_to_list(folio, split_folios);
1488 folio_unlock(folio);
1489 if (!rc)
1490 list_move_tail(&folio->lru, split_folios);
1491
1492 return rc;
1493}
1494
1495#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1496#define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1497#else
1498#define NR_MAX_BATCHED_MIGRATION 512
1499#endif
1500#define NR_MAX_MIGRATE_PAGES_RETRY 10
1501#define NR_MAX_MIGRATE_ASYNC_RETRY 3
1502#define NR_MAX_MIGRATE_SYNC_RETRY \
1503 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1504
1505struct migrate_pages_stats {
1506 int nr_succeeded; /* Normal and large folios migrated successfully, in
1507 units of base pages */
1508 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1509 units of base pages. Untried folios aren't counted */
1510 int nr_thp_succeeded; /* THP migrated successfully */
1511 int nr_thp_failed; /* THP failed to be migrated */
1512 int nr_thp_split; /* THP split before migrating */
1513 int nr_split; /* Large folio (include THP) split before migrating */
1514};
1515
1516/*
1517 * Returns the number of hugetlb folios that were not migrated, or an error code
1518 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1519 * any more because the list has become empty or no retryable hugetlb folios
1520 * exist any more. It is caller's responsibility to call putback_movable_pages()
1521 * only if ret != 0.
1522 */
1523static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1524 free_folio_t put_new_folio, unsigned long private,
1525 enum migrate_mode mode, int reason,
1526 struct migrate_pages_stats *stats,
1527 struct list_head *ret_folios)
1528{
1529 int retry = 1;
1530 int nr_failed = 0;
1531 int nr_retry_pages = 0;
1532 int pass = 0;
1533 struct folio *folio, *folio2;
1534 int rc, nr_pages;
1535
1536 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1537 retry = 0;
1538 nr_retry_pages = 0;
1539
1540 list_for_each_entry_safe(folio, folio2, from, lru) {
1541 if (!folio_test_hugetlb(folio))
1542 continue;
1543
1544 nr_pages = folio_nr_pages(folio);
1545
1546 cond_resched();
1547
1548 /*
1549 * Migratability of hugepages depends on architectures and
1550 * their size. This check is necessary because some callers
1551 * of hugepage migration like soft offline and memory
1552 * hotremove don't walk through page tables or check whether
1553 * the hugepage is pmd-based or not before kicking migration.
1554 */
1555 if (!hugepage_migration_supported(folio_hstate(folio))) {
1556 nr_failed++;
1557 stats->nr_failed_pages += nr_pages;
1558 list_move_tail(&folio->lru, ret_folios);
1559 continue;
1560 }
1561
1562 rc = unmap_and_move_huge_page(get_new_folio,
1563 put_new_folio, private,
1564 folio, pass > 2, mode,
1565 reason, ret_folios);
1566 /*
1567 * The rules are:
1568 * Success: hugetlb folio will be put back
1569 * -EAGAIN: stay on the from list
1570 * -ENOMEM: stay on the from list
1571 * Other errno: put on ret_folios list
1572 */
1573 switch(rc) {
1574 case -ENOMEM:
1575 /*
1576 * When memory is low, don't bother to try to migrate
1577 * other folios, just exit.
1578 */
1579 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1580 return -ENOMEM;
1581 case -EAGAIN:
1582 retry++;
1583 nr_retry_pages += nr_pages;
1584 break;
1585 case MIGRATEPAGE_SUCCESS:
1586 stats->nr_succeeded += nr_pages;
1587 break;
1588 default:
1589 /*
1590 * Permanent failure (-EBUSY, etc.):
1591 * unlike -EAGAIN case, the failed folio is
1592 * removed from migration folio list and not
1593 * retried in the next outer loop.
1594 */
1595 nr_failed++;
1596 stats->nr_failed_pages += nr_pages;
1597 break;
1598 }
1599 }
1600 }
1601 /*
1602 * nr_failed is number of hugetlb folios failed to be migrated. After
1603 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1604 * folios as failed.
1605 */
1606 nr_failed += retry;
1607 stats->nr_failed_pages += nr_retry_pages;
1608
1609 return nr_failed;
1610}
1611
1612/*
1613 * migrate_pages_batch() first unmaps folios in the from list as many as
1614 * possible, then move the unmapped folios.
1615 *
1616 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1617 * lock or bit when we have locked more than one folio. Which may cause
1618 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1619 * length of the from list must be <= 1.
1620 */
1621static int migrate_pages_batch(struct list_head *from,
1622 new_folio_t get_new_folio, free_folio_t put_new_folio,
1623 unsigned long private, enum migrate_mode mode, int reason,
1624 struct list_head *ret_folios, struct list_head *split_folios,
1625 struct migrate_pages_stats *stats, int nr_pass)
1626{
1627 int retry = 1;
1628 int thp_retry = 1;
1629 int nr_failed = 0;
1630 int nr_retry_pages = 0;
1631 int pass = 0;
1632 bool is_thp = false;
1633 bool is_large = false;
1634 struct folio *folio, *folio2, *dst = NULL, *dst2;
1635 int rc, rc_saved = 0, nr_pages;
1636 LIST_HEAD(unmap_folios);
1637 LIST_HEAD(dst_folios);
1638 bool nosplit = (reason == MR_NUMA_MISPLACED);
1639
1640 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1641 !list_empty(from) && !list_is_singular(from));
1642
1643 for (pass = 0; pass < nr_pass && retry; pass++) {
1644 retry = 0;
1645 thp_retry = 0;
1646 nr_retry_pages = 0;
1647
1648 list_for_each_entry_safe(folio, folio2, from, lru) {
1649 is_large = folio_test_large(folio);
1650 is_thp = is_large && folio_test_pmd_mappable(folio);
1651 nr_pages = folio_nr_pages(folio);
1652
1653 cond_resched();
1654
1655 /*
1656 * Large folio migration might be unsupported or
1657 * the allocation might be failed so we should retry
1658 * on the same folio with the large folio split
1659 * to normal folios.
1660 *
1661 * Split folios are put in split_folios, and
1662 * we will migrate them after the rest of the
1663 * list is processed.
1664 */
1665 if (!thp_migration_supported() && is_thp) {
1666 nr_failed++;
1667 stats->nr_thp_failed++;
1668 if (!try_split_folio(folio, split_folios)) {
1669 stats->nr_thp_split++;
1670 stats->nr_split++;
1671 continue;
1672 }
1673 stats->nr_failed_pages += nr_pages;
1674 list_move_tail(&folio->lru, ret_folios);
1675 continue;
1676 }
1677
1678 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1679 private, folio, &dst, mode, reason,
1680 ret_folios);
1681 /*
1682 * The rules are:
1683 * Success: folio will be freed
1684 * Unmap: folio will be put on unmap_folios list,
1685 * dst folio put on dst_folios list
1686 * -EAGAIN: stay on the from list
1687 * -ENOMEM: stay on the from list
1688 * Other errno: put on ret_folios list
1689 */
1690 switch(rc) {
1691 case -ENOMEM:
1692 /*
1693 * When memory is low, don't bother to try to migrate
1694 * other folios, move unmapped folios, then exit.
1695 */
1696 nr_failed++;
1697 stats->nr_thp_failed += is_thp;
1698 /* Large folio NUMA faulting doesn't split to retry. */
1699 if (is_large && !nosplit) {
1700 int ret = try_split_folio(folio, split_folios);
1701
1702 if (!ret) {
1703 stats->nr_thp_split += is_thp;
1704 stats->nr_split++;
1705 break;
1706 } else if (reason == MR_LONGTERM_PIN &&
1707 ret == -EAGAIN) {
1708 /*
1709 * Try again to split large folio to
1710 * mitigate the failure of longterm pinning.
1711 */
1712 retry++;
1713 thp_retry += is_thp;
1714 nr_retry_pages += nr_pages;
1715 /* Undo duplicated failure counting. */
1716 nr_failed--;
1717 stats->nr_thp_failed -= is_thp;
1718 break;
1719 }
1720 }
1721
1722 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1723 /* nr_failed isn't updated for not used */
1724 stats->nr_thp_failed += thp_retry;
1725 rc_saved = rc;
1726 if (list_empty(&unmap_folios))
1727 goto out;
1728 else
1729 goto move;
1730 case -EAGAIN:
1731 retry++;
1732 thp_retry += is_thp;
1733 nr_retry_pages += nr_pages;
1734 break;
1735 case MIGRATEPAGE_SUCCESS:
1736 stats->nr_succeeded += nr_pages;
1737 stats->nr_thp_succeeded += is_thp;
1738 break;
1739 case MIGRATEPAGE_UNMAP:
1740 list_move_tail(&folio->lru, &unmap_folios);
1741 list_add_tail(&dst->lru, &dst_folios);
1742 break;
1743 default:
1744 /*
1745 * Permanent failure (-EBUSY, etc.):
1746 * unlike -EAGAIN case, the failed folio is
1747 * removed from migration folio list and not
1748 * retried in the next outer loop.
1749 */
1750 nr_failed++;
1751 stats->nr_thp_failed += is_thp;
1752 stats->nr_failed_pages += nr_pages;
1753 break;
1754 }
1755 }
1756 }
1757 nr_failed += retry;
1758 stats->nr_thp_failed += thp_retry;
1759 stats->nr_failed_pages += nr_retry_pages;
1760move:
1761 /* Flush TLBs for all unmapped folios */
1762 try_to_unmap_flush();
1763
1764 retry = 1;
1765 for (pass = 0; pass < nr_pass && retry; pass++) {
1766 retry = 0;
1767 thp_retry = 0;
1768 nr_retry_pages = 0;
1769
1770 dst = list_first_entry(&dst_folios, struct folio, lru);
1771 dst2 = list_next_entry(dst, lru);
1772 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1773 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1774 nr_pages = folio_nr_pages(folio);
1775
1776 cond_resched();
1777
1778 rc = migrate_folio_move(put_new_folio, private,
1779 folio, dst, mode,
1780 reason, ret_folios);
1781 /*
1782 * The rules are:
1783 * Success: folio will be freed
1784 * -EAGAIN: stay on the unmap_folios list
1785 * Other errno: put on ret_folios list
1786 */
1787 switch(rc) {
1788 case -EAGAIN:
1789 retry++;
1790 thp_retry += is_thp;
1791 nr_retry_pages += nr_pages;
1792 break;
1793 case MIGRATEPAGE_SUCCESS:
1794 stats->nr_succeeded += nr_pages;
1795 stats->nr_thp_succeeded += is_thp;
1796 break;
1797 default:
1798 nr_failed++;
1799 stats->nr_thp_failed += is_thp;
1800 stats->nr_failed_pages += nr_pages;
1801 break;
1802 }
1803 dst = dst2;
1804 dst2 = list_next_entry(dst, lru);
1805 }
1806 }
1807 nr_failed += retry;
1808 stats->nr_thp_failed += thp_retry;
1809 stats->nr_failed_pages += nr_retry_pages;
1810
1811 rc = rc_saved ? : nr_failed;
1812out:
1813 /* Cleanup remaining folios */
1814 dst = list_first_entry(&dst_folios, struct folio, lru);
1815 dst2 = list_next_entry(dst, lru);
1816 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1817 int old_page_state = 0;
1818 struct anon_vma *anon_vma = NULL;
1819
1820 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1821 migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1822 anon_vma, true, ret_folios);
1823 list_del(&dst->lru);
1824 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1825 dst = dst2;
1826 dst2 = list_next_entry(dst, lru);
1827 }
1828
1829 return rc;
1830}
1831
1832static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1833 free_folio_t put_new_folio, unsigned long private,
1834 enum migrate_mode mode, int reason,
1835 struct list_head *ret_folios, struct list_head *split_folios,
1836 struct migrate_pages_stats *stats)
1837{
1838 int rc, nr_failed = 0;
1839 LIST_HEAD(folios);
1840 struct migrate_pages_stats astats;
1841
1842 memset(&astats, 0, sizeof(astats));
1843 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1844 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1845 reason, &folios, split_folios, &astats,
1846 NR_MAX_MIGRATE_ASYNC_RETRY);
1847 stats->nr_succeeded += astats.nr_succeeded;
1848 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1849 stats->nr_thp_split += astats.nr_thp_split;
1850 stats->nr_split += astats.nr_split;
1851 if (rc < 0) {
1852 stats->nr_failed_pages += astats.nr_failed_pages;
1853 stats->nr_thp_failed += astats.nr_thp_failed;
1854 list_splice_tail(&folios, ret_folios);
1855 return rc;
1856 }
1857 stats->nr_thp_failed += astats.nr_thp_split;
1858 /*
1859 * Do not count rc, as pages will be retried below.
1860 * Count nr_split only, since it includes nr_thp_split.
1861 */
1862 nr_failed += astats.nr_split;
1863 /*
1864 * Fall back to migrate all failed folios one by one synchronously. All
1865 * failed folios except split THPs will be retried, so their failure
1866 * isn't counted
1867 */
1868 list_splice_tail_init(&folios, from);
1869 while (!list_empty(from)) {
1870 list_move(from->next, &folios);
1871 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1872 private, mode, reason, ret_folios,
1873 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1874 list_splice_tail_init(&folios, ret_folios);
1875 if (rc < 0)
1876 return rc;
1877 nr_failed += rc;
1878 }
1879
1880 return nr_failed;
1881}
1882
1883/*
1884 * migrate_pages - migrate the folios specified in a list, to the free folios
1885 * supplied as the target for the page migration
1886 *
1887 * @from: The list of folios to be migrated.
1888 * @get_new_folio: The function used to allocate free folios to be used
1889 * as the target of the folio migration.
1890 * @put_new_folio: The function used to free target folios if migration
1891 * fails, or NULL if no special handling is necessary.
1892 * @private: Private data to be passed on to get_new_folio()
1893 * @mode: The migration mode that specifies the constraints for
1894 * folio migration, if any.
1895 * @reason: The reason for folio migration.
1896 * @ret_succeeded: Set to the number of folios migrated successfully if
1897 * the caller passes a non-NULL pointer.
1898 *
1899 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1900 * are movable any more because the list has become empty or no retryable folios
1901 * exist any more. It is caller's responsibility to call putback_movable_pages()
1902 * only if ret != 0.
1903 *
1904 * Returns the number of {normal folio, large folio, hugetlb} that were not
1905 * migrated, or an error code. The number of large folio splits will be
1906 * considered as the number of non-migrated large folio, no matter how many
1907 * split folios of the large folio are migrated successfully.
1908 */
1909int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1910 free_folio_t put_new_folio, unsigned long private,
1911 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1912{
1913 int rc, rc_gather;
1914 int nr_pages;
1915 struct folio *folio, *folio2;
1916 LIST_HEAD(folios);
1917 LIST_HEAD(ret_folios);
1918 LIST_HEAD(split_folios);
1919 struct migrate_pages_stats stats;
1920
1921 trace_mm_migrate_pages_start(mode, reason);
1922
1923 memset(&stats, 0, sizeof(stats));
1924
1925 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1926 mode, reason, &stats, &ret_folios);
1927 if (rc_gather < 0)
1928 goto out;
1929
1930again:
1931 nr_pages = 0;
1932 list_for_each_entry_safe(folio, folio2, from, lru) {
1933 /* Retried hugetlb folios will be kept in list */
1934 if (folio_test_hugetlb(folio)) {
1935 list_move_tail(&folio->lru, &ret_folios);
1936 continue;
1937 }
1938
1939 nr_pages += folio_nr_pages(folio);
1940 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1941 break;
1942 }
1943 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1944 list_cut_before(&folios, from, &folio2->lru);
1945 else
1946 list_splice_init(from, &folios);
1947 if (mode == MIGRATE_ASYNC)
1948 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1949 private, mode, reason, &ret_folios,
1950 &split_folios, &stats,
1951 NR_MAX_MIGRATE_PAGES_RETRY);
1952 else
1953 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1954 private, mode, reason, &ret_folios,
1955 &split_folios, &stats);
1956 list_splice_tail_init(&folios, &ret_folios);
1957 if (rc < 0) {
1958 rc_gather = rc;
1959 list_splice_tail(&split_folios, &ret_folios);
1960 goto out;
1961 }
1962 if (!list_empty(&split_folios)) {
1963 /*
1964 * Failure isn't counted since all split folios of a large folio
1965 * is counted as 1 failure already. And, we only try to migrate
1966 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1967 */
1968 migrate_pages_batch(&split_folios, get_new_folio,
1969 put_new_folio, private, MIGRATE_ASYNC, reason,
1970 &ret_folios, NULL, &stats, 1);
1971 list_splice_tail_init(&split_folios, &ret_folios);
1972 }
1973 rc_gather += rc;
1974 if (!list_empty(from))
1975 goto again;
1976out:
1977 /*
1978 * Put the permanent failure folio back to migration list, they
1979 * will be put back to the right list by the caller.
1980 */
1981 list_splice(&ret_folios, from);
1982
1983 /*
1984 * Return 0 in case all split folios of fail-to-migrate large folios
1985 * are migrated successfully.
1986 */
1987 if (list_empty(from))
1988 rc_gather = 0;
1989
1990 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1991 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1992 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1993 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1994 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1995 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1996 stats.nr_thp_succeeded, stats.nr_thp_failed,
1997 stats.nr_thp_split, stats.nr_split, mode,
1998 reason);
1999
2000 if (ret_succeeded)
2001 *ret_succeeded = stats.nr_succeeded;
2002
2003 return rc_gather;
2004}
2005
2006struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2007{
2008 struct migration_target_control *mtc;
2009 gfp_t gfp_mask;
2010 unsigned int order = 0;
2011 int nid;
2012 int zidx;
2013
2014 mtc = (struct migration_target_control *)private;
2015 gfp_mask = mtc->gfp_mask;
2016 nid = mtc->nid;
2017 if (nid == NUMA_NO_NODE)
2018 nid = folio_nid(src);
2019
2020 if (folio_test_hugetlb(src)) {
2021 struct hstate *h = folio_hstate(src);
2022
2023 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2024 return alloc_hugetlb_folio_nodemask(h, nid,
2025 mtc->nmask, gfp_mask);
2026 }
2027
2028 if (folio_test_large(src)) {
2029 /*
2030 * clear __GFP_RECLAIM to make the migration callback
2031 * consistent with regular THP allocations.
2032 */
2033 gfp_mask &= ~__GFP_RECLAIM;
2034 gfp_mask |= GFP_TRANSHUGE;
2035 order = folio_order(src);
2036 }
2037 zidx = zone_idx(folio_zone(src));
2038 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2039 gfp_mask |= __GFP_HIGHMEM;
2040
2041 return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2042}
2043
2044#ifdef CONFIG_NUMA
2045
2046static int store_status(int __user *status, int start, int value, int nr)
2047{
2048 while (nr-- > 0) {
2049 if (put_user(value, status + start))
2050 return -EFAULT;
2051 start++;
2052 }
2053
2054 return 0;
2055}
2056
2057static int do_move_pages_to_node(struct list_head *pagelist, int node)
2058{
2059 int err;
2060 struct migration_target_control mtc = {
2061 .nid = node,
2062 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2063 };
2064
2065 err = migrate_pages(pagelist, alloc_migration_target, NULL,
2066 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2067 if (err)
2068 putback_movable_pages(pagelist);
2069 return err;
2070}
2071
2072/*
2073 * Resolves the given address to a struct page, isolates it from the LRU and
2074 * puts it to the given pagelist.
2075 * Returns:
2076 * errno - if the page cannot be found/isolated
2077 * 0 - when it doesn't have to be migrated because it is already on the
2078 * target node
2079 * 1 - when it has been queued
2080 */
2081static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2082 int node, struct list_head *pagelist, bool migrate_all)
2083{
2084 struct vm_area_struct *vma;
2085 unsigned long addr;
2086 struct page *page;
2087 struct folio *folio;
2088 int err;
2089
2090 mmap_read_lock(mm);
2091 addr = (unsigned long)untagged_addr_remote(mm, p);
2092
2093 err = -EFAULT;
2094 vma = vma_lookup(mm, addr);
2095 if (!vma || !vma_migratable(vma))
2096 goto out;
2097
2098 /* FOLL_DUMP to ignore special (like zero) pages */
2099 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2100
2101 err = PTR_ERR(page);
2102 if (IS_ERR(page))
2103 goto out;
2104
2105 err = -ENOENT;
2106 if (!page)
2107 goto out;
2108
2109 folio = page_folio(page);
2110 if (folio_is_zone_device(folio))
2111 goto out_putfolio;
2112
2113 err = 0;
2114 if (folio_nid(folio) == node)
2115 goto out_putfolio;
2116
2117 err = -EACCES;
2118 if (page_mapcount(page) > 1 && !migrate_all)
2119 goto out_putfolio;
2120
2121 err = -EBUSY;
2122 if (folio_test_hugetlb(folio)) {
2123 if (isolate_hugetlb(folio, pagelist))
2124 err = 1;
2125 } else {
2126 if (!folio_isolate_lru(folio))
2127 goto out_putfolio;
2128
2129 err = 1;
2130 list_add_tail(&folio->lru, pagelist);
2131 node_stat_mod_folio(folio,
2132 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2133 folio_nr_pages(folio));
2134 }
2135out_putfolio:
2136 /*
2137 * Either remove the duplicate refcount from folio_isolate_lru()
2138 * or drop the folio ref if it was not isolated.
2139 */
2140 folio_put(folio);
2141out:
2142 mmap_read_unlock(mm);
2143 return err;
2144}
2145
2146static int move_pages_and_store_status(int node,
2147 struct list_head *pagelist, int __user *status,
2148 int start, int i, unsigned long nr_pages)
2149{
2150 int err;
2151
2152 if (list_empty(pagelist))
2153 return 0;
2154
2155 err = do_move_pages_to_node(pagelist, node);
2156 if (err) {
2157 /*
2158 * Positive err means the number of failed
2159 * pages to migrate. Since we are going to
2160 * abort and return the number of non-migrated
2161 * pages, so need to include the rest of the
2162 * nr_pages that have not been attempted as
2163 * well.
2164 */
2165 if (err > 0)
2166 err += nr_pages - i;
2167 return err;
2168 }
2169 return store_status(status, start, node, i - start);
2170}
2171
2172/*
2173 * Migrate an array of page address onto an array of nodes and fill
2174 * the corresponding array of status.
2175 */
2176static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2177 unsigned long nr_pages,
2178 const void __user * __user *pages,
2179 const int __user *nodes,
2180 int __user *status, int flags)
2181{
2182 compat_uptr_t __user *compat_pages = (void __user *)pages;
2183 int current_node = NUMA_NO_NODE;
2184 LIST_HEAD(pagelist);
2185 int start, i;
2186 int err = 0, err1;
2187
2188 lru_cache_disable();
2189
2190 for (i = start = 0; i < nr_pages; i++) {
2191 const void __user *p;
2192 int node;
2193
2194 err = -EFAULT;
2195 if (in_compat_syscall()) {
2196 compat_uptr_t cp;
2197
2198 if (get_user(cp, compat_pages + i))
2199 goto out_flush;
2200
2201 p = compat_ptr(cp);
2202 } else {
2203 if (get_user(p, pages + i))
2204 goto out_flush;
2205 }
2206 if (get_user(node, nodes + i))
2207 goto out_flush;
2208
2209 err = -ENODEV;
2210 if (node < 0 || node >= MAX_NUMNODES)
2211 goto out_flush;
2212 if (!node_state(node, N_MEMORY))
2213 goto out_flush;
2214
2215 err = -EACCES;
2216 if (!node_isset(node, task_nodes))
2217 goto out_flush;
2218
2219 if (current_node == NUMA_NO_NODE) {
2220 current_node = node;
2221 start = i;
2222 } else if (node != current_node) {
2223 err = move_pages_and_store_status(current_node,
2224 &pagelist, status, start, i, nr_pages);
2225 if (err)
2226 goto out;
2227 start = i;
2228 current_node = node;
2229 }
2230
2231 /*
2232 * Errors in the page lookup or isolation are not fatal and we simply
2233 * report them via status
2234 */
2235 err = add_page_for_migration(mm, p, current_node, &pagelist,
2236 flags & MPOL_MF_MOVE_ALL);
2237
2238 if (err > 0) {
2239 /* The page is successfully queued for migration */
2240 continue;
2241 }
2242
2243 /*
2244 * The move_pages() man page does not have an -EEXIST choice, so
2245 * use -EFAULT instead.
2246 */
2247 if (err == -EEXIST)
2248 err = -EFAULT;
2249
2250 /*
2251 * If the page is already on the target node (!err), store the
2252 * node, otherwise, store the err.
2253 */
2254 err = store_status(status, i, err ? : current_node, 1);
2255 if (err)
2256 goto out_flush;
2257
2258 err = move_pages_and_store_status(current_node, &pagelist,
2259 status, start, i, nr_pages);
2260 if (err) {
2261 /* We have accounted for page i */
2262 if (err > 0)
2263 err--;
2264 goto out;
2265 }
2266 current_node = NUMA_NO_NODE;
2267 }
2268out_flush:
2269 /* Make sure we do not overwrite the existing error */
2270 err1 = move_pages_and_store_status(current_node, &pagelist,
2271 status, start, i, nr_pages);
2272 if (err >= 0)
2273 err = err1;
2274out:
2275 lru_cache_enable();
2276 return err;
2277}
2278
2279/*
2280 * Determine the nodes of an array of pages and store it in an array of status.
2281 */
2282static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2283 const void __user **pages, int *status)
2284{
2285 unsigned long i;
2286
2287 mmap_read_lock(mm);
2288
2289 for (i = 0; i < nr_pages; i++) {
2290 unsigned long addr = (unsigned long)(*pages);
2291 struct vm_area_struct *vma;
2292 struct page *page;
2293 int err = -EFAULT;
2294
2295 vma = vma_lookup(mm, addr);
2296 if (!vma)
2297 goto set_status;
2298
2299 /* FOLL_DUMP to ignore special (like zero) pages */
2300 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2301
2302 err = PTR_ERR(page);
2303 if (IS_ERR(page))
2304 goto set_status;
2305
2306 err = -ENOENT;
2307 if (!page)
2308 goto set_status;
2309
2310 if (!is_zone_device_page(page))
2311 err = page_to_nid(page);
2312
2313 put_page(page);
2314set_status:
2315 *status = err;
2316
2317 pages++;
2318 status++;
2319 }
2320
2321 mmap_read_unlock(mm);
2322}
2323
2324static int get_compat_pages_array(const void __user *chunk_pages[],
2325 const void __user * __user *pages,
2326 unsigned long chunk_nr)
2327{
2328 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2329 compat_uptr_t p;
2330 int i;
2331
2332 for (i = 0; i < chunk_nr; i++) {
2333 if (get_user(p, pages32 + i))
2334 return -EFAULT;
2335 chunk_pages[i] = compat_ptr(p);
2336 }
2337
2338 return 0;
2339}
2340
2341/*
2342 * Determine the nodes of a user array of pages and store it in
2343 * a user array of status.
2344 */
2345static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2346 const void __user * __user *pages,
2347 int __user *status)
2348{
2349#define DO_PAGES_STAT_CHUNK_NR 16UL
2350 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2351 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2352
2353 while (nr_pages) {
2354 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2355
2356 if (in_compat_syscall()) {
2357 if (get_compat_pages_array(chunk_pages, pages,
2358 chunk_nr))
2359 break;
2360 } else {
2361 if (copy_from_user(chunk_pages, pages,
2362 chunk_nr * sizeof(*chunk_pages)))
2363 break;
2364 }
2365
2366 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2367
2368 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2369 break;
2370
2371 pages += chunk_nr;
2372 status += chunk_nr;
2373 nr_pages -= chunk_nr;
2374 }
2375 return nr_pages ? -EFAULT : 0;
2376}
2377
2378static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2379{
2380 struct task_struct *task;
2381 struct mm_struct *mm;
2382
2383 /*
2384 * There is no need to check if current process has the right to modify
2385 * the specified process when they are same.
2386 */
2387 if (!pid) {
2388 mmget(current->mm);
2389 *mem_nodes = cpuset_mems_allowed(current);
2390 return current->mm;
2391 }
2392
2393 /* Find the mm_struct */
2394 rcu_read_lock();
2395 task = find_task_by_vpid(pid);
2396 if (!task) {
2397 rcu_read_unlock();
2398 return ERR_PTR(-ESRCH);
2399 }
2400 get_task_struct(task);
2401
2402 /*
2403 * Check if this process has the right to modify the specified
2404 * process. Use the regular "ptrace_may_access()" checks.
2405 */
2406 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2407 rcu_read_unlock();
2408 mm = ERR_PTR(-EPERM);
2409 goto out;
2410 }
2411 rcu_read_unlock();
2412
2413 mm = ERR_PTR(security_task_movememory(task));
2414 if (IS_ERR(mm))
2415 goto out;
2416 *mem_nodes = cpuset_mems_allowed(task);
2417 mm = get_task_mm(task);
2418out:
2419 put_task_struct(task);
2420 if (!mm)
2421 mm = ERR_PTR(-EINVAL);
2422 return mm;
2423}
2424
2425/*
2426 * Move a list of pages in the address space of the currently executing
2427 * process.
2428 */
2429static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2430 const void __user * __user *pages,
2431 const int __user *nodes,
2432 int __user *status, int flags)
2433{
2434 struct mm_struct *mm;
2435 int err;
2436 nodemask_t task_nodes;
2437
2438 /* Check flags */
2439 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2440 return -EINVAL;
2441
2442 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2443 return -EPERM;
2444
2445 mm = find_mm_struct(pid, &task_nodes);
2446 if (IS_ERR(mm))
2447 return PTR_ERR(mm);
2448
2449 if (nodes)
2450 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2451 nodes, status, flags);
2452 else
2453 err = do_pages_stat(mm, nr_pages, pages, status);
2454
2455 mmput(mm);
2456 return err;
2457}
2458
2459SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2460 const void __user * __user *, pages,
2461 const int __user *, nodes,
2462 int __user *, status, int, flags)
2463{
2464 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2465}
2466
2467#ifdef CONFIG_NUMA_BALANCING
2468/*
2469 * Returns true if this is a safe migration target node for misplaced NUMA
2470 * pages. Currently it only checks the watermarks which is crude.
2471 */
2472static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2473 unsigned long nr_migrate_pages)
2474{
2475 int z;
2476
2477 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2478 struct zone *zone = pgdat->node_zones + z;
2479
2480 if (!managed_zone(zone))
2481 continue;
2482
2483 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2484 if (!zone_watermark_ok(zone, 0,
2485 high_wmark_pages(zone) +
2486 nr_migrate_pages,
2487 ZONE_MOVABLE, 0))
2488 continue;
2489 return true;
2490 }
2491 return false;
2492}
2493
2494static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2495 unsigned long data)
2496{
2497 int nid = (int) data;
2498 int order = folio_order(src);
2499 gfp_t gfp = __GFP_THISNODE;
2500
2501 if (order > 0)
2502 gfp |= GFP_TRANSHUGE_LIGHT;
2503 else {
2504 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2505 __GFP_NOWARN;
2506 gfp &= ~__GFP_RECLAIM;
2507 }
2508 return __folio_alloc_node(gfp, order, nid);
2509}
2510
2511static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2512{
2513 int nr_pages = folio_nr_pages(folio);
2514
2515 /* Avoid migrating to a node that is nearly full */
2516 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2517 int z;
2518
2519 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2520 return 0;
2521 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2522 if (managed_zone(pgdat->node_zones + z))
2523 break;
2524 }
2525
2526 /*
2527 * If there are no managed zones, it should not proceed
2528 * further.
2529 */
2530 if (z < 0)
2531 return 0;
2532
2533 wakeup_kswapd(pgdat->node_zones + z, 0,
2534 folio_order(folio), ZONE_MOVABLE);
2535 return 0;
2536 }
2537
2538 if (!folio_isolate_lru(folio))
2539 return 0;
2540
2541 node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2542 nr_pages);
2543
2544 /*
2545 * Isolating the folio has taken another reference, so the
2546 * caller's reference can be safely dropped without the folio
2547 * disappearing underneath us during migration.
2548 */
2549 folio_put(folio);
2550 return 1;
2551}
2552
2553/*
2554 * Attempt to migrate a misplaced folio to the specified destination
2555 * node. Caller is expected to have an elevated reference count on
2556 * the folio that will be dropped by this function before returning.
2557 */
2558int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2559 int node)
2560{
2561 pg_data_t *pgdat = NODE_DATA(node);
2562 int isolated;
2563 int nr_remaining;
2564 unsigned int nr_succeeded;
2565 LIST_HEAD(migratepages);
2566 int nr_pages = folio_nr_pages(folio);
2567
2568 /*
2569 * Don't migrate file folios that are mapped in multiple processes
2570 * with execute permissions as they are probably shared libraries.
2571 * To check if the folio is shared, ideally we want to make sure
2572 * every page is mapped to the same process. Doing that is very
2573 * expensive, so check the estimated mapcount of the folio instead.
2574 */
2575 if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2576 (vma->vm_flags & VM_EXEC))
2577 goto out;
2578
2579 /*
2580 * Also do not migrate dirty folios as not all filesystems can move
2581 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2582 */
2583 if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2584 goto out;
2585
2586 isolated = numamigrate_isolate_folio(pgdat, folio);
2587 if (!isolated)
2588 goto out;
2589
2590 list_add(&folio->lru, &migratepages);
2591 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2592 NULL, node, MIGRATE_ASYNC,
2593 MR_NUMA_MISPLACED, &nr_succeeded);
2594 if (nr_remaining) {
2595 if (!list_empty(&migratepages)) {
2596 list_del(&folio->lru);
2597 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
2598 folio_is_file_lru(folio), -nr_pages);
2599 folio_putback_lru(folio);
2600 }
2601 isolated = 0;
2602 }
2603 if (nr_succeeded) {
2604 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2605 if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
2606 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2607 nr_succeeded);
2608 }
2609 BUG_ON(!list_empty(&migratepages));
2610 return isolated;
2611
2612out:
2613 folio_put(folio);
2614 return 0;
2615}
2616#endif /* CONFIG_NUMA_BALANCING */
2617#endif /* CONFIG_NUMA */