Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory Migration functionality - linux/mm/migrate.c
4 *
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 *
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
9 *
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
13 * Christoph Lameter
14 */
15
16#include <linux/migrate.h>
17#include <linux/export.h>
18#include <linux/swap.h>
19#include <linux/swapops.h>
20#include <linux/pagemap.h>
21#include <linux/buffer_head.h>
22#include <linux/mm_inline.h>
23#include <linux/nsproxy.h>
24#include <linux/pagevec.h>
25#include <linux/ksm.h>
26#include <linux/rmap.h>
27#include <linux/topology.h>
28#include <linux/cpu.h>
29#include <linux/cpuset.h>
30#include <linux/writeback.h>
31#include <linux/mempolicy.h>
32#include <linux/vmalloc.h>
33#include <linux/security.h>
34#include <linux/backing-dev.h>
35#include <linux/compaction.h>
36#include <linux/syscalls.h>
37#include <linux/compat.h>
38#include <linux/hugetlb.h>
39#include <linux/hugetlb_cgroup.h>
40#include <linux/gfp.h>
41#include <linux/pagewalk.h>
42#include <linux/pfn_t.h>
43#include <linux/memremap.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/balloon_compaction.h>
46#include <linux/mmu_notifier.h>
47#include <linux/page_idle.h>
48#include <linux/page_owner.h>
49#include <linux/sched/mm.h>
50#include <linux/ptrace.h>
51
52#include <asm/tlbflush.h>
53
54#define CREATE_TRACE_POINTS
55#include <trace/events/migrate.h>
56
57#include "internal.h"
58
59/*
60 * migrate_prep() needs to be called before we start compiling a list of pages
61 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
62 * undesirable, use migrate_prep_local()
63 */
64int migrate_prep(void)
65{
66 /*
67 * Clear the LRU lists so pages can be isolated.
68 * Note that pages may be moved off the LRU after we have
69 * drained them. Those pages will fail to migrate like other
70 * pages that may be busy.
71 */
72 lru_add_drain_all();
73
74 return 0;
75}
76
77/* Do the necessary work of migrate_prep but not if it involves other CPUs */
78int migrate_prep_local(void)
79{
80 lru_add_drain();
81
82 return 0;
83}
84
85int isolate_movable_page(struct page *page, isolate_mode_t mode)
86{
87 struct address_space *mapping;
88
89 /*
90 * Avoid burning cycles with pages that are yet under __free_pages(),
91 * or just got freed under us.
92 *
93 * In case we 'win' a race for a movable page being freed under us and
94 * raise its refcount preventing __free_pages() from doing its job
95 * the put_page() at the end of this block will take care of
96 * release this page, thus avoiding a nasty leakage.
97 */
98 if (unlikely(!get_page_unless_zero(page)))
99 goto out;
100
101 /*
102 * Check PageMovable before holding a PG_lock because page's owner
103 * assumes anybody doesn't touch PG_lock of newly allocated page
104 * so unconditionally grabbing the lock ruins page's owner side.
105 */
106 if (unlikely(!__PageMovable(page)))
107 goto out_putpage;
108 /*
109 * As movable pages are not isolated from LRU lists, concurrent
110 * compaction threads can race against page migration functions
111 * as well as race against the releasing a page.
112 *
113 * In order to avoid having an already isolated movable page
114 * being (wrongly) re-isolated while it is under migration,
115 * or to avoid attempting to isolate pages being released,
116 * lets be sure we have the page lock
117 * before proceeding with the movable page isolation steps.
118 */
119 if (unlikely(!trylock_page(page)))
120 goto out_putpage;
121
122 if (!PageMovable(page) || PageIsolated(page))
123 goto out_no_isolated;
124
125 mapping = page_mapping(page);
126 VM_BUG_ON_PAGE(!mapping, page);
127
128 if (!mapping->a_ops->isolate_page(page, mode))
129 goto out_no_isolated;
130
131 /* Driver shouldn't use PG_isolated bit of page->flags */
132 WARN_ON_ONCE(PageIsolated(page));
133 __SetPageIsolated(page);
134 unlock_page(page);
135
136 return 0;
137
138out_no_isolated:
139 unlock_page(page);
140out_putpage:
141 put_page(page);
142out:
143 return -EBUSY;
144}
145
146/* It should be called on page which is PG_movable */
147void putback_movable_page(struct page *page)
148{
149 struct address_space *mapping;
150
151 VM_BUG_ON_PAGE(!PageLocked(page), page);
152 VM_BUG_ON_PAGE(!PageMovable(page), page);
153 VM_BUG_ON_PAGE(!PageIsolated(page), page);
154
155 mapping = page_mapping(page);
156 mapping->a_ops->putback_page(page);
157 __ClearPageIsolated(page);
158}
159
160/*
161 * Put previously isolated pages back onto the appropriate lists
162 * from where they were once taken off for compaction/migration.
163 *
164 * This function shall be used whenever the isolated pageset has been
165 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
166 * and isolate_huge_page().
167 */
168void putback_movable_pages(struct list_head *l)
169{
170 struct page *page;
171 struct page *page2;
172
173 list_for_each_entry_safe(page, page2, l, lru) {
174 if (unlikely(PageHuge(page))) {
175 putback_active_hugepage(page);
176 continue;
177 }
178 list_del(&page->lru);
179 /*
180 * We isolated non-lru movable page so here we can use
181 * __PageMovable because LRU page's mapping cannot have
182 * PAGE_MAPPING_MOVABLE.
183 */
184 if (unlikely(__PageMovable(page))) {
185 VM_BUG_ON_PAGE(!PageIsolated(page), page);
186 lock_page(page);
187 if (PageMovable(page))
188 putback_movable_page(page);
189 else
190 __ClearPageIsolated(page);
191 unlock_page(page);
192 put_page(page);
193 } else {
194 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
195 page_is_file_cache(page), -hpage_nr_pages(page));
196 putback_lru_page(page);
197 }
198 }
199}
200
201/*
202 * Restore a potential migration pte to a working pte entry
203 */
204static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
205 unsigned long addr, void *old)
206{
207 struct page_vma_mapped_walk pvmw = {
208 .page = old,
209 .vma = vma,
210 .address = addr,
211 .flags = PVMW_SYNC | PVMW_MIGRATION,
212 };
213 struct page *new;
214 pte_t pte;
215 swp_entry_t entry;
216
217 VM_BUG_ON_PAGE(PageTail(page), page);
218 while (page_vma_mapped_walk(&pvmw)) {
219 if (PageKsm(page))
220 new = page;
221 else
222 new = page - pvmw.page->index +
223 linear_page_index(vma, pvmw.address);
224
225#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
226 /* PMD-mapped THP migration entry */
227 if (!pvmw.pte) {
228 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
229 remove_migration_pmd(&pvmw, new);
230 continue;
231 }
232#endif
233
234 get_page(new);
235 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
236 if (pte_swp_soft_dirty(*pvmw.pte))
237 pte = pte_mksoft_dirty(pte);
238
239 /*
240 * Recheck VMA as permissions can change since migration started
241 */
242 entry = pte_to_swp_entry(*pvmw.pte);
243 if (is_write_migration_entry(entry))
244 pte = maybe_mkwrite(pte, vma);
245
246 if (unlikely(is_zone_device_page(new))) {
247 if (is_device_private_page(new)) {
248 entry = make_device_private_entry(new, pte_write(pte));
249 pte = swp_entry_to_pte(entry);
250 }
251 }
252
253#ifdef CONFIG_HUGETLB_PAGE
254 if (PageHuge(new)) {
255 pte = pte_mkhuge(pte);
256 pte = arch_make_huge_pte(pte, vma, new, 0);
257 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
258 if (PageAnon(new))
259 hugepage_add_anon_rmap(new, vma, pvmw.address);
260 else
261 page_dup_rmap(new, true);
262 } else
263#endif
264 {
265 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
266
267 if (PageAnon(new))
268 page_add_anon_rmap(new, vma, pvmw.address, false);
269 else
270 page_add_file_rmap(new, false);
271 }
272 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
273 mlock_vma_page(new);
274
275 if (PageTransHuge(page) && PageMlocked(page))
276 clear_page_mlock(page);
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 page cache replacement of page migration started, page_count
327 * is zero; but we must not call put_and_wait_on_page_locked() without
328 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
329 */
330 if (!get_page_unless_zero(page))
331 goto out;
332 pte_unmap_unlock(ptep, ptl);
333 put_and_wait_on_page_locked(page);
334 return;
335out:
336 pte_unmap_unlock(ptep, ptl);
337}
338
339void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
340 unsigned long address)
341{
342 spinlock_t *ptl = pte_lockptr(mm, pmd);
343 pte_t *ptep = pte_offset_map(pmd, address);
344 __migration_entry_wait(mm, ptep, ptl);
345}
346
347void migration_entry_wait_huge(struct vm_area_struct *vma,
348 struct mm_struct *mm, pte_t *pte)
349{
350 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
351 __migration_entry_wait(mm, pte, ptl);
352}
353
354#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
355void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
356{
357 spinlock_t *ptl;
358 struct page *page;
359
360 ptl = pmd_lock(mm, pmd);
361 if (!is_pmd_migration_entry(*pmd))
362 goto unlock;
363 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
364 if (!get_page_unless_zero(page))
365 goto unlock;
366 spin_unlock(ptl);
367 put_and_wait_on_page_locked(page);
368 return;
369unlock:
370 spin_unlock(ptl);
371}
372#endif
373
374static int expected_page_refs(struct address_space *mapping, struct page *page)
375{
376 int expected_count = 1;
377
378 /*
379 * Device public or private pages have an extra refcount as they are
380 * ZONE_DEVICE pages.
381 */
382 expected_count += is_device_private_page(page);
383 if (mapping)
384 expected_count += hpage_nr_pages(page) + page_has_private(page);
385
386 return expected_count;
387}
388
389/*
390 * Replace the page in the mapping.
391 *
392 * The number of remaining references must be:
393 * 1 for anonymous pages without a mapping
394 * 2 for pages with a mapping
395 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
396 */
397int migrate_page_move_mapping(struct address_space *mapping,
398 struct page *newpage, struct page *page, int extra_count)
399{
400 XA_STATE(xas, &mapping->i_pages, page_index(page));
401 struct zone *oldzone, *newzone;
402 int dirty;
403 int expected_count = expected_page_refs(mapping, page) + extra_count;
404
405 if (!mapping) {
406 /* Anonymous page without mapping */
407 if (page_count(page) != expected_count)
408 return -EAGAIN;
409
410 /* No turning back from here */
411 newpage->index = page->index;
412 newpage->mapping = page->mapping;
413 if (PageSwapBacked(page))
414 __SetPageSwapBacked(newpage);
415
416 return MIGRATEPAGE_SUCCESS;
417 }
418
419 oldzone = page_zone(page);
420 newzone = page_zone(newpage);
421
422 xas_lock_irq(&xas);
423 if (page_count(page) != expected_count || xas_load(&xas) != page) {
424 xas_unlock_irq(&xas);
425 return -EAGAIN;
426 }
427
428 if (!page_ref_freeze(page, expected_count)) {
429 xas_unlock_irq(&xas);
430 return -EAGAIN;
431 }
432
433 /*
434 * Now we know that no one else is looking at the page:
435 * no turning back from here.
436 */
437 newpage->index = page->index;
438 newpage->mapping = page->mapping;
439 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
440 if (PageSwapBacked(page)) {
441 __SetPageSwapBacked(newpage);
442 if (PageSwapCache(page)) {
443 SetPageSwapCache(newpage);
444 set_page_private(newpage, page_private(page));
445 }
446 } else {
447 VM_BUG_ON_PAGE(PageSwapCache(page), page);
448 }
449
450 /* Move dirty while page refs frozen and newpage not yet exposed */
451 dirty = PageDirty(page);
452 if (dirty) {
453 ClearPageDirty(page);
454 SetPageDirty(newpage);
455 }
456
457 xas_store(&xas, newpage);
458 if (PageTransHuge(page)) {
459 int i;
460
461 for (i = 1; i < HPAGE_PMD_NR; i++) {
462 xas_next(&xas);
463 xas_store(&xas, newpage);
464 }
465 }
466
467 /*
468 * Drop cache reference from old page by unfreezing
469 * to one less reference.
470 * We know this isn't the last reference.
471 */
472 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
473
474 xas_unlock(&xas);
475 /* Leave irq disabled to prevent preemption while updating stats */
476
477 /*
478 * If moved to a different zone then also account
479 * the page for that zone. Other VM counters will be
480 * taken care of when we establish references to the
481 * new page and drop references to the old page.
482 *
483 * Note that anonymous pages are accounted for
484 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
485 * are mapped to swap space.
486 */
487 if (newzone != oldzone) {
488 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
489 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
490 if (PageSwapBacked(page) && !PageSwapCache(page)) {
491 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
492 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
493 }
494 if (dirty && mapping_cap_account_dirty(mapping)) {
495 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
496 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
497 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
498 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
499 }
500 }
501 local_irq_enable();
502
503 return MIGRATEPAGE_SUCCESS;
504}
505EXPORT_SYMBOL(migrate_page_move_mapping);
506
507/*
508 * The expected number of remaining references is the same as that
509 * of migrate_page_move_mapping().
510 */
511int migrate_huge_page_move_mapping(struct address_space *mapping,
512 struct page *newpage, struct page *page)
513{
514 XA_STATE(xas, &mapping->i_pages, page_index(page));
515 int expected_count;
516
517 xas_lock_irq(&xas);
518 expected_count = 2 + page_has_private(page);
519 if (page_count(page) != expected_count || xas_load(&xas) != page) {
520 xas_unlock_irq(&xas);
521 return -EAGAIN;
522 }
523
524 if (!page_ref_freeze(page, expected_count)) {
525 xas_unlock_irq(&xas);
526 return -EAGAIN;
527 }
528
529 newpage->index = page->index;
530 newpage->mapping = page->mapping;
531
532 get_page(newpage);
533
534 xas_store(&xas, newpage);
535
536 page_ref_unfreeze(page, expected_count - 1);
537
538 xas_unlock_irq(&xas);
539
540 return MIGRATEPAGE_SUCCESS;
541}
542
543/*
544 * Gigantic pages are so large that we do not guarantee that page++ pointer
545 * arithmetic will work across the entire page. We need something more
546 * specialized.
547 */
548static void __copy_gigantic_page(struct page *dst, struct page *src,
549 int nr_pages)
550{
551 int i;
552 struct page *dst_base = dst;
553 struct page *src_base = src;
554
555 for (i = 0; i < nr_pages; ) {
556 cond_resched();
557 copy_highpage(dst, src);
558
559 i++;
560 dst = mem_map_next(dst, dst_base, i);
561 src = mem_map_next(src, src_base, i);
562 }
563}
564
565static void copy_huge_page(struct page *dst, struct page *src)
566{
567 int i;
568 int nr_pages;
569
570 if (PageHuge(src)) {
571 /* hugetlbfs page */
572 struct hstate *h = page_hstate(src);
573 nr_pages = pages_per_huge_page(h);
574
575 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
576 __copy_gigantic_page(dst, src, nr_pages);
577 return;
578 }
579 } else {
580 /* thp page */
581 BUG_ON(!PageTransHuge(src));
582 nr_pages = hpage_nr_pages(src);
583 }
584
585 for (i = 0; i < nr_pages; i++) {
586 cond_resched();
587 copy_highpage(dst + i, src + i);
588 }
589}
590
591/*
592 * Copy the page to its new location
593 */
594void migrate_page_states(struct page *newpage, struct page *page)
595{
596 int cpupid;
597
598 if (PageError(page))
599 SetPageError(newpage);
600 if (PageReferenced(page))
601 SetPageReferenced(newpage);
602 if (PageUptodate(page))
603 SetPageUptodate(newpage);
604 if (TestClearPageActive(page)) {
605 VM_BUG_ON_PAGE(PageUnevictable(page), page);
606 SetPageActive(newpage);
607 } else if (TestClearPageUnevictable(page))
608 SetPageUnevictable(newpage);
609 if (PageWorkingset(page))
610 SetPageWorkingset(newpage);
611 if (PageChecked(page))
612 SetPageChecked(newpage);
613 if (PageMappedToDisk(page))
614 SetPageMappedToDisk(newpage);
615
616 /* Move dirty on pages not done by migrate_page_move_mapping() */
617 if (PageDirty(page))
618 SetPageDirty(newpage);
619
620 if (page_is_young(page))
621 set_page_young(newpage);
622 if (page_is_idle(page))
623 set_page_idle(newpage);
624
625 /*
626 * Copy NUMA information to the new page, to prevent over-eager
627 * future migrations of this same page.
628 */
629 cpupid = page_cpupid_xchg_last(page, -1);
630 page_cpupid_xchg_last(newpage, cpupid);
631
632 ksm_migrate_page(newpage, page);
633 /*
634 * Please do not reorder this without considering how mm/ksm.c's
635 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
636 */
637 if (PageSwapCache(page))
638 ClearPageSwapCache(page);
639 ClearPagePrivate(page);
640 set_page_private(page, 0);
641
642 /*
643 * If any waiters have accumulated on the new page then
644 * wake them up.
645 */
646 if (PageWriteback(newpage))
647 end_page_writeback(newpage);
648
649 copy_page_owner(page, newpage);
650
651 mem_cgroup_migrate(page, newpage);
652}
653EXPORT_SYMBOL(migrate_page_states);
654
655void migrate_page_copy(struct page *newpage, struct page *page)
656{
657 if (PageHuge(page) || PageTransHuge(page))
658 copy_huge_page(newpage, page);
659 else
660 copy_highpage(newpage, page);
661
662 migrate_page_states(newpage, page);
663}
664EXPORT_SYMBOL(migrate_page_copy);
665
666/************************************************************
667 * Migration functions
668 ***********************************************************/
669
670/*
671 * Common logic to directly migrate a single LRU page suitable for
672 * pages that do not use PagePrivate/PagePrivate2.
673 *
674 * Pages are locked upon entry and exit.
675 */
676int migrate_page(struct address_space *mapping,
677 struct page *newpage, struct page *page,
678 enum migrate_mode mode)
679{
680 int rc;
681
682 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
683
684 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
685
686 if (rc != MIGRATEPAGE_SUCCESS)
687 return rc;
688
689 if (mode != MIGRATE_SYNC_NO_COPY)
690 migrate_page_copy(newpage, page);
691 else
692 migrate_page_states(newpage, page);
693 return MIGRATEPAGE_SUCCESS;
694}
695EXPORT_SYMBOL(migrate_page);
696
697#ifdef CONFIG_BLOCK
698/* Returns true if all buffers are successfully locked */
699static bool buffer_migrate_lock_buffers(struct buffer_head *head,
700 enum migrate_mode mode)
701{
702 struct buffer_head *bh = head;
703
704 /* Simple case, sync compaction */
705 if (mode != MIGRATE_ASYNC) {
706 do {
707 lock_buffer(bh);
708 bh = bh->b_this_page;
709
710 } while (bh != head);
711
712 return true;
713 }
714
715 /* async case, we cannot block on lock_buffer so use trylock_buffer */
716 do {
717 if (!trylock_buffer(bh)) {
718 /*
719 * We failed to lock the buffer and cannot stall in
720 * async migration. Release the taken locks
721 */
722 struct buffer_head *failed_bh = bh;
723 bh = head;
724 while (bh != failed_bh) {
725 unlock_buffer(bh);
726 bh = bh->b_this_page;
727 }
728 return false;
729 }
730
731 bh = bh->b_this_page;
732 } while (bh != head);
733 return true;
734}
735
736static int __buffer_migrate_page(struct address_space *mapping,
737 struct page *newpage, struct page *page, enum migrate_mode mode,
738 bool check_refs)
739{
740 struct buffer_head *bh, *head;
741 int rc;
742 int expected_count;
743
744 if (!page_has_buffers(page))
745 return migrate_page(mapping, newpage, page, mode);
746
747 /* Check whether page does not have extra refs before we do more work */
748 expected_count = expected_page_refs(mapping, page);
749 if (page_count(page) != expected_count)
750 return -EAGAIN;
751
752 head = page_buffers(page);
753 if (!buffer_migrate_lock_buffers(head, mode))
754 return -EAGAIN;
755
756 if (check_refs) {
757 bool busy;
758 bool invalidated = false;
759
760recheck_buffers:
761 busy = false;
762 spin_lock(&mapping->private_lock);
763 bh = head;
764 do {
765 if (atomic_read(&bh->b_count)) {
766 busy = true;
767 break;
768 }
769 bh = bh->b_this_page;
770 } while (bh != head);
771 if (busy) {
772 if (invalidated) {
773 rc = -EAGAIN;
774 goto unlock_buffers;
775 }
776 spin_unlock(&mapping->private_lock);
777 invalidate_bh_lrus();
778 invalidated = true;
779 goto recheck_buffers;
780 }
781 }
782
783 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
784 if (rc != MIGRATEPAGE_SUCCESS)
785 goto unlock_buffers;
786
787 ClearPagePrivate(page);
788 set_page_private(newpage, page_private(page));
789 set_page_private(page, 0);
790 put_page(page);
791 get_page(newpage);
792
793 bh = head;
794 do {
795 set_bh_page(bh, newpage, bh_offset(bh));
796 bh = bh->b_this_page;
797
798 } while (bh != head);
799
800 SetPagePrivate(newpage);
801
802 if (mode != MIGRATE_SYNC_NO_COPY)
803 migrate_page_copy(newpage, page);
804 else
805 migrate_page_states(newpage, page);
806
807 rc = MIGRATEPAGE_SUCCESS;
808unlock_buffers:
809 if (check_refs)
810 spin_unlock(&mapping->private_lock);
811 bh = head;
812 do {
813 unlock_buffer(bh);
814 bh = bh->b_this_page;
815
816 } while (bh != head);
817
818 return rc;
819}
820
821/*
822 * Migration function for pages with buffers. This function can only be used
823 * if the underlying filesystem guarantees that no other references to "page"
824 * exist. For example attached buffer heads are accessed only under page lock.
825 */
826int buffer_migrate_page(struct address_space *mapping,
827 struct page *newpage, struct page *page, enum migrate_mode mode)
828{
829 return __buffer_migrate_page(mapping, newpage, page, mode, false);
830}
831EXPORT_SYMBOL(buffer_migrate_page);
832
833/*
834 * Same as above except that this variant is more careful and checks that there
835 * are also no buffer head references. This function is the right one for
836 * mappings where buffer heads are directly looked up and referenced (such as
837 * block device mappings).
838 */
839int buffer_migrate_page_norefs(struct address_space *mapping,
840 struct page *newpage, struct page *page, enum migrate_mode mode)
841{
842 return __buffer_migrate_page(mapping, newpage, page, mode, true);
843}
844#endif
845
846/*
847 * Writeback a page to clean the dirty state
848 */
849static int writeout(struct address_space *mapping, struct page *page)
850{
851 struct writeback_control wbc = {
852 .sync_mode = WB_SYNC_NONE,
853 .nr_to_write = 1,
854 .range_start = 0,
855 .range_end = LLONG_MAX,
856 .for_reclaim = 1
857 };
858 int rc;
859
860 if (!mapping->a_ops->writepage)
861 /* No write method for the address space */
862 return -EINVAL;
863
864 if (!clear_page_dirty_for_io(page))
865 /* Someone else already triggered a write */
866 return -EAGAIN;
867
868 /*
869 * A dirty page may imply that the underlying filesystem has
870 * the page on some queue. So the page must be clean for
871 * migration. Writeout may mean we loose the lock and the
872 * page state is no longer what we checked for earlier.
873 * At this point we know that the migration attempt cannot
874 * be successful.
875 */
876 remove_migration_ptes(page, page, false);
877
878 rc = mapping->a_ops->writepage(page, &wbc);
879
880 if (rc != AOP_WRITEPAGE_ACTIVATE)
881 /* unlocked. Relock */
882 lock_page(page);
883
884 return (rc < 0) ? -EIO : -EAGAIN;
885}
886
887/*
888 * Default handling if a filesystem does not provide a migration function.
889 */
890static int fallback_migrate_page(struct address_space *mapping,
891 struct page *newpage, struct page *page, enum migrate_mode mode)
892{
893 if (PageDirty(page)) {
894 /* Only writeback pages in full synchronous migration */
895 switch (mode) {
896 case MIGRATE_SYNC:
897 case MIGRATE_SYNC_NO_COPY:
898 break;
899 default:
900 return -EBUSY;
901 }
902 return writeout(mapping, page);
903 }
904
905 /*
906 * Buffers may be managed in a filesystem specific way.
907 * We must have no buffers or drop them.
908 */
909 if (page_has_private(page) &&
910 !try_to_release_page(page, GFP_KERNEL))
911 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
912
913 return migrate_page(mapping, newpage, page, mode);
914}
915
916/*
917 * Move a page to a newly allocated page
918 * The page is locked and all ptes have been successfully removed.
919 *
920 * The new page will have replaced the old page if this function
921 * is successful.
922 *
923 * Return value:
924 * < 0 - error code
925 * MIGRATEPAGE_SUCCESS - success
926 */
927static int move_to_new_page(struct page *newpage, struct page *page,
928 enum migrate_mode mode)
929{
930 struct address_space *mapping;
931 int rc = -EAGAIN;
932 bool is_lru = !__PageMovable(page);
933
934 VM_BUG_ON_PAGE(!PageLocked(page), page);
935 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
936
937 mapping = page_mapping(page);
938
939 if (likely(is_lru)) {
940 if (!mapping)
941 rc = migrate_page(mapping, newpage, page, mode);
942 else if (mapping->a_ops->migratepage)
943 /*
944 * Most pages have a mapping and most filesystems
945 * provide a migratepage callback. Anonymous pages
946 * are part of swap space which also has its own
947 * migratepage callback. This is the most common path
948 * for page migration.
949 */
950 rc = mapping->a_ops->migratepage(mapping, newpage,
951 page, mode);
952 else
953 rc = fallback_migrate_page(mapping, newpage,
954 page, mode);
955 } else {
956 /*
957 * In case of non-lru page, it could be released after
958 * isolation step. In that case, we shouldn't try migration.
959 */
960 VM_BUG_ON_PAGE(!PageIsolated(page), page);
961 if (!PageMovable(page)) {
962 rc = MIGRATEPAGE_SUCCESS;
963 __ClearPageIsolated(page);
964 goto out;
965 }
966
967 rc = mapping->a_ops->migratepage(mapping, newpage,
968 page, mode);
969 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
970 !PageIsolated(page));
971 }
972
973 /*
974 * When successful, old pagecache page->mapping must be cleared before
975 * page is freed; but stats require that PageAnon be left as PageAnon.
976 */
977 if (rc == MIGRATEPAGE_SUCCESS) {
978 if (__PageMovable(page)) {
979 VM_BUG_ON_PAGE(!PageIsolated(page), page);
980
981 /*
982 * We clear PG_movable under page_lock so any compactor
983 * cannot try to migrate this page.
984 */
985 __ClearPageIsolated(page);
986 }
987
988 /*
989 * Anonymous and movable page->mapping will be cleard by
990 * free_pages_prepare so don't reset it here for keeping
991 * the type to work PageAnon, for example.
992 */
993 if (!PageMappingFlags(page))
994 page->mapping = NULL;
995
996 if (likely(!is_zone_device_page(newpage)))
997 flush_dcache_page(newpage);
998
999 }
1000out:
1001 return rc;
1002}
1003
1004static int __unmap_and_move(struct page *page, struct page *newpage,
1005 int force, enum migrate_mode mode)
1006{
1007 int rc = -EAGAIN;
1008 int page_was_mapped = 0;
1009 struct anon_vma *anon_vma = NULL;
1010 bool is_lru = !__PageMovable(page);
1011
1012 if (!trylock_page(page)) {
1013 if (!force || mode == MIGRATE_ASYNC)
1014 goto out;
1015
1016 /*
1017 * It's not safe for direct compaction to call lock_page.
1018 * For example, during page readahead pages are added locked
1019 * to the LRU. Later, when the IO completes the pages are
1020 * marked uptodate and unlocked. However, the queueing
1021 * could be merging multiple pages for one bio (e.g.
1022 * mpage_readpages). If an allocation happens for the
1023 * second or third page, the process can end up locking
1024 * the same page twice and deadlocking. Rather than
1025 * trying to be clever about what pages can be locked,
1026 * avoid the use of lock_page for direct compaction
1027 * altogether.
1028 */
1029 if (current->flags & PF_MEMALLOC)
1030 goto out;
1031
1032 lock_page(page);
1033 }
1034
1035 if (PageWriteback(page)) {
1036 /*
1037 * Only in the case of a full synchronous migration is it
1038 * necessary to wait for PageWriteback. In the async case,
1039 * the retry loop is too short and in the sync-light case,
1040 * the overhead of stalling is too much
1041 */
1042 switch (mode) {
1043 case MIGRATE_SYNC:
1044 case MIGRATE_SYNC_NO_COPY:
1045 break;
1046 default:
1047 rc = -EBUSY;
1048 goto out_unlock;
1049 }
1050 if (!force)
1051 goto out_unlock;
1052 wait_on_page_writeback(page);
1053 }
1054
1055 /*
1056 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1057 * we cannot notice that anon_vma is freed while we migrates a page.
1058 * This get_anon_vma() delays freeing anon_vma pointer until the end
1059 * of migration. File cache pages are no problem because of page_lock()
1060 * File Caches may use write_page() or lock_page() in migration, then,
1061 * just care Anon page here.
1062 *
1063 * Only page_get_anon_vma() understands the subtleties of
1064 * getting a hold on an anon_vma from outside one of its mms.
1065 * But if we cannot get anon_vma, then we won't need it anyway,
1066 * because that implies that the anon page is no longer mapped
1067 * (and cannot be remapped so long as we hold the page lock).
1068 */
1069 if (PageAnon(page) && !PageKsm(page))
1070 anon_vma = page_get_anon_vma(page);
1071
1072 /*
1073 * Block others from accessing the new page when we get around to
1074 * establishing additional references. We are usually the only one
1075 * holding a reference to newpage at this point. We used to have a BUG
1076 * here if trylock_page(newpage) fails, but would like to allow for
1077 * cases where there might be a race with the previous use of newpage.
1078 * This is much like races on refcount of oldpage: just don't BUG().
1079 */
1080 if (unlikely(!trylock_page(newpage)))
1081 goto out_unlock;
1082
1083 if (unlikely(!is_lru)) {
1084 rc = move_to_new_page(newpage, page, mode);
1085 goto out_unlock_both;
1086 }
1087
1088 /*
1089 * Corner case handling:
1090 * 1. When a new swap-cache page is read into, it is added to the LRU
1091 * and treated as swapcache but it has no rmap yet.
1092 * Calling try_to_unmap() against a page->mapping==NULL page will
1093 * trigger a BUG. So handle it here.
1094 * 2. An orphaned page (see truncate_complete_page) might have
1095 * fs-private metadata. The page can be picked up due to memory
1096 * offlining. Everywhere else except page reclaim, the page is
1097 * invisible to the vm, so the page can not be migrated. So try to
1098 * free the metadata, so the page can be freed.
1099 */
1100 if (!page->mapping) {
1101 VM_BUG_ON_PAGE(PageAnon(page), page);
1102 if (page_has_private(page)) {
1103 try_to_free_buffers(page);
1104 goto out_unlock_both;
1105 }
1106 } else if (page_mapped(page)) {
1107 /* Establish migration ptes */
1108 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1109 page);
1110 try_to_unmap(page,
1111 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1112 page_was_mapped = 1;
1113 }
1114
1115 if (!page_mapped(page))
1116 rc = move_to_new_page(newpage, page, mode);
1117
1118 if (page_was_mapped)
1119 remove_migration_ptes(page,
1120 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1121
1122out_unlock_both:
1123 unlock_page(newpage);
1124out_unlock:
1125 /* Drop an anon_vma reference if we took one */
1126 if (anon_vma)
1127 put_anon_vma(anon_vma);
1128 unlock_page(page);
1129out:
1130 /*
1131 * If migration is successful, decrease refcount of the newpage
1132 * which will not free the page because new page owner increased
1133 * refcounter. As well, if it is LRU page, add the page to LRU
1134 * list in here. Use the old state of the isolated source page to
1135 * determine if we migrated a LRU page. newpage was already unlocked
1136 * and possibly modified by its owner - don't rely on the page
1137 * state.
1138 */
1139 if (rc == MIGRATEPAGE_SUCCESS) {
1140 if (unlikely(!is_lru))
1141 put_page(newpage);
1142 else
1143 putback_lru_page(newpage);
1144 }
1145
1146 return rc;
1147}
1148
1149/*
1150 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1151 * around it.
1152 */
1153#if defined(CONFIG_ARM) && \
1154 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1155#define ICE_noinline noinline
1156#else
1157#define ICE_noinline
1158#endif
1159
1160/*
1161 * Obtain the lock on page, remove all ptes and migrate the page
1162 * to the newly allocated page in newpage.
1163 */
1164static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1165 free_page_t put_new_page,
1166 unsigned long private, struct page *page,
1167 int force, enum migrate_mode mode,
1168 enum migrate_reason reason)
1169{
1170 int rc = MIGRATEPAGE_SUCCESS;
1171 struct page *newpage;
1172
1173 if (!thp_migration_supported() && PageTransHuge(page))
1174 return -ENOMEM;
1175
1176 newpage = get_new_page(page, private);
1177 if (!newpage)
1178 return -ENOMEM;
1179
1180 if (page_count(page) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page);
1183 ClearPageUnevictable(page);
1184 if (unlikely(__PageMovable(page))) {
1185 lock_page(page);
1186 if (!PageMovable(page))
1187 __ClearPageIsolated(page);
1188 unlock_page(page);
1189 }
1190 if (put_new_page)
1191 put_new_page(newpage, private);
1192 else
1193 put_page(newpage);
1194 goto out;
1195 }
1196
1197 rc = __unmap_and_move(page, newpage, force, mode);
1198 if (rc == MIGRATEPAGE_SUCCESS)
1199 set_page_owner_migrate_reason(newpage, reason);
1200
1201out:
1202 if (rc != -EAGAIN) {
1203 /*
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kepts its references and be
1207 * restored.
1208 */
1209 list_del(&page->lru);
1210
1211 /*
1212 * Compaction can migrate also non-LRU pages which are
1213 * not accounted to NR_ISOLATED_*. They can be recognized
1214 * as __PageMovable
1215 */
1216 if (likely(!__PageMovable(page)))
1217 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1218 page_is_file_cache(page), -hpage_nr_pages(page));
1219 }
1220
1221 /*
1222 * If migration is successful, releases reference grabbed during
1223 * isolation. Otherwise, restore the page to right list unless
1224 * we want to retry.
1225 */
1226 if (rc == MIGRATEPAGE_SUCCESS) {
1227 put_page(page);
1228 if (reason == MR_MEMORY_FAILURE) {
1229 /*
1230 * Set PG_HWPoison on just freed page
1231 * intentionally. Although it's rather weird,
1232 * it's how HWPoison flag works at the moment.
1233 */
1234 if (set_hwpoison_free_buddy_page(page))
1235 num_poisoned_pages_inc();
1236 }
1237 } else {
1238 if (rc != -EAGAIN) {
1239 if (likely(!__PageMovable(page))) {
1240 putback_lru_page(page);
1241 goto put_new;
1242 }
1243
1244 lock_page(page);
1245 if (PageMovable(page))
1246 putback_movable_page(page);
1247 else
1248 __ClearPageIsolated(page);
1249 unlock_page(page);
1250 put_page(page);
1251 }
1252put_new:
1253 if (put_new_page)
1254 put_new_page(newpage, private);
1255 else
1256 put_page(newpage);
1257 }
1258
1259 return rc;
1260}
1261
1262/*
1263 * Counterpart of unmap_and_move_page() for hugepage migration.
1264 *
1265 * This function doesn't wait the completion of hugepage I/O
1266 * because there is no race between I/O and migration for hugepage.
1267 * Note that currently hugepage I/O occurs only in direct I/O
1268 * where no lock is held and PG_writeback is irrelevant,
1269 * and writeback status of all subpages are counted in the reference
1270 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1271 * under direct I/O, the reference of the head page is 512 and a bit more.)
1272 * This means that when we try to migrate hugepage whose subpages are
1273 * doing direct I/O, some references remain after try_to_unmap() and
1274 * hugepage migration fails without data corruption.
1275 *
1276 * There is also no race when direct I/O is issued on the page under migration,
1277 * because then pte is replaced with migration swap entry and direct I/O code
1278 * will wait in the page fault for migration to complete.
1279 */
1280static int unmap_and_move_huge_page(new_page_t get_new_page,
1281 free_page_t put_new_page, unsigned long private,
1282 struct page *hpage, int force,
1283 enum migrate_mode mode, int reason)
1284{
1285 int rc = -EAGAIN;
1286 int page_was_mapped = 0;
1287 struct page *new_hpage;
1288 struct anon_vma *anon_vma = NULL;
1289
1290 /*
1291 * Migratability of hugepages depends on architectures and their size.
1292 * This check is necessary because some callers of hugepage migration
1293 * like soft offline and memory hotremove don't walk through page
1294 * tables or check whether the hugepage is pmd-based or not before
1295 * kicking migration.
1296 */
1297 if (!hugepage_migration_supported(page_hstate(hpage))) {
1298 putback_active_hugepage(hpage);
1299 return -ENOSYS;
1300 }
1301
1302 new_hpage = get_new_page(hpage, private);
1303 if (!new_hpage)
1304 return -ENOMEM;
1305
1306 if (!trylock_page(hpage)) {
1307 if (!force)
1308 goto out;
1309 switch (mode) {
1310 case MIGRATE_SYNC:
1311 case MIGRATE_SYNC_NO_COPY:
1312 break;
1313 default:
1314 goto out;
1315 }
1316 lock_page(hpage);
1317 }
1318
1319 /*
1320 * Check for pages which are in the process of being freed. Without
1321 * page_mapping() set, hugetlbfs specific move page routine will not
1322 * be called and we could leak usage counts for subpools.
1323 */
1324 if (page_private(hpage) && !page_mapping(hpage)) {
1325 rc = -EBUSY;
1326 goto out_unlock;
1327 }
1328
1329 if (PageAnon(hpage))
1330 anon_vma = page_get_anon_vma(hpage);
1331
1332 if (unlikely(!trylock_page(new_hpage)))
1333 goto put_anon;
1334
1335 if (page_mapped(hpage)) {
1336 try_to_unmap(hpage,
1337 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1338 page_was_mapped = 1;
1339 }
1340
1341 if (!page_mapped(hpage))
1342 rc = move_to_new_page(new_hpage, hpage, mode);
1343
1344 if (page_was_mapped)
1345 remove_migration_ptes(hpage,
1346 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1347
1348 unlock_page(new_hpage);
1349
1350put_anon:
1351 if (anon_vma)
1352 put_anon_vma(anon_vma);
1353
1354 if (rc == MIGRATEPAGE_SUCCESS) {
1355 move_hugetlb_state(hpage, new_hpage, reason);
1356 put_new_page = NULL;
1357 }
1358
1359out_unlock:
1360 unlock_page(hpage);
1361out:
1362 if (rc != -EAGAIN)
1363 putback_active_hugepage(hpage);
1364
1365 /*
1366 * If migration was not successful and there's a freeing callback, use
1367 * it. Otherwise, put_page() will drop the reference grabbed during
1368 * isolation.
1369 */
1370 if (put_new_page)
1371 put_new_page(new_hpage, private);
1372 else
1373 putback_active_hugepage(new_hpage);
1374
1375 return rc;
1376}
1377
1378/*
1379 * migrate_pages - migrate the pages specified in a list, to the free pages
1380 * supplied as the target for the page migration
1381 *
1382 * @from: The list of pages to be migrated.
1383 * @get_new_page: The function used to allocate free pages to be used
1384 * as the target of the page migration.
1385 * @put_new_page: The function used to free target pages if migration
1386 * fails, or NULL if no special handling is necessary.
1387 * @private: Private data to be passed on to get_new_page()
1388 * @mode: The migration mode that specifies the constraints for
1389 * page migration, if any.
1390 * @reason: The reason for page migration.
1391 *
1392 * The function returns after 10 attempts or if no pages are movable any more
1393 * because the list has become empty or no retryable pages exist any more.
1394 * The caller should call putback_movable_pages() to return pages to the LRU
1395 * or free list only if ret != 0.
1396 *
1397 * Returns the number of pages that were not migrated, or an error code.
1398 */
1399int migrate_pages(struct list_head *from, new_page_t get_new_page,
1400 free_page_t put_new_page, unsigned long private,
1401 enum migrate_mode mode, int reason)
1402{
1403 int retry = 1;
1404 int nr_failed = 0;
1405 int nr_succeeded = 0;
1406 int pass = 0;
1407 struct page *page;
1408 struct page *page2;
1409 int swapwrite = current->flags & PF_SWAPWRITE;
1410 int rc;
1411
1412 if (!swapwrite)
1413 current->flags |= PF_SWAPWRITE;
1414
1415 for(pass = 0; pass < 10 && retry; pass++) {
1416 retry = 0;
1417
1418 list_for_each_entry_safe(page, page2, from, lru) {
1419retry:
1420 cond_resched();
1421
1422 if (PageHuge(page))
1423 rc = unmap_and_move_huge_page(get_new_page,
1424 put_new_page, private, page,
1425 pass > 2, mode, reason);
1426 else
1427 rc = unmap_and_move(get_new_page, put_new_page,
1428 private, page, pass > 2, mode,
1429 reason);
1430
1431 switch(rc) {
1432 case -ENOMEM:
1433 /*
1434 * THP migration might be unsupported or the
1435 * allocation could've failed so we should
1436 * retry on the same page with the THP split
1437 * to base pages.
1438 *
1439 * Head page is retried immediately and tail
1440 * pages are added to the tail of the list so
1441 * we encounter them after the rest of the list
1442 * is processed.
1443 */
1444 if (PageTransHuge(page) && !PageHuge(page)) {
1445 lock_page(page);
1446 rc = split_huge_page_to_list(page, from);
1447 unlock_page(page);
1448 if (!rc) {
1449 list_safe_reset_next(page, page2, lru);
1450 goto retry;
1451 }
1452 }
1453 nr_failed++;
1454 goto out;
1455 case -EAGAIN:
1456 retry++;
1457 break;
1458 case MIGRATEPAGE_SUCCESS:
1459 nr_succeeded++;
1460 break;
1461 default:
1462 /*
1463 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1464 * unlike -EAGAIN case, the failed page is
1465 * removed from migration page list and not
1466 * retried in the next outer loop.
1467 */
1468 nr_failed++;
1469 break;
1470 }
1471 }
1472 }
1473 nr_failed += retry;
1474 rc = nr_failed;
1475out:
1476 if (nr_succeeded)
1477 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1478 if (nr_failed)
1479 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1480 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1481
1482 if (!swapwrite)
1483 current->flags &= ~PF_SWAPWRITE;
1484
1485 return rc;
1486}
1487
1488#ifdef CONFIG_NUMA
1489
1490static int store_status(int __user *status, int start, int value, int nr)
1491{
1492 while (nr-- > 0) {
1493 if (put_user(value, status + start))
1494 return -EFAULT;
1495 start++;
1496 }
1497
1498 return 0;
1499}
1500
1501static int do_move_pages_to_node(struct mm_struct *mm,
1502 struct list_head *pagelist, int node)
1503{
1504 int err;
1505
1506 if (list_empty(pagelist))
1507 return 0;
1508
1509 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1510 MIGRATE_SYNC, MR_SYSCALL);
1511 if (err)
1512 putback_movable_pages(pagelist);
1513 return err;
1514}
1515
1516/*
1517 * Resolves the given address to a struct page, isolates it from the LRU and
1518 * puts it to the given pagelist.
1519 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1520 * queued or the page doesn't need to be migrated because it is already on
1521 * the target node
1522 */
1523static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1524 int node, struct list_head *pagelist, bool migrate_all)
1525{
1526 struct vm_area_struct *vma;
1527 struct page *page;
1528 unsigned int follflags;
1529 int err;
1530
1531 down_read(&mm->mmap_sem);
1532 err = -EFAULT;
1533 vma = find_vma(mm, addr);
1534 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1535 goto out;
1536
1537 /* FOLL_DUMP to ignore special (like zero) pages */
1538 follflags = FOLL_GET | FOLL_DUMP;
1539 page = follow_page(vma, addr, follflags);
1540
1541 err = PTR_ERR(page);
1542 if (IS_ERR(page))
1543 goto out;
1544
1545 err = -ENOENT;
1546 if (!page)
1547 goto out;
1548
1549 err = 0;
1550 if (page_to_nid(page) == node)
1551 goto out_putpage;
1552
1553 err = -EACCES;
1554 if (page_mapcount(page) > 1 && !migrate_all)
1555 goto out_putpage;
1556
1557 if (PageHuge(page)) {
1558 if (PageHead(page)) {
1559 isolate_huge_page(page, pagelist);
1560 err = 0;
1561 }
1562 } else {
1563 struct page *head;
1564
1565 head = compound_head(page);
1566 err = isolate_lru_page(head);
1567 if (err)
1568 goto out_putpage;
1569
1570 err = 0;
1571 list_add_tail(&head->lru, pagelist);
1572 mod_node_page_state(page_pgdat(head),
1573 NR_ISOLATED_ANON + page_is_file_cache(head),
1574 hpage_nr_pages(head));
1575 }
1576out_putpage:
1577 /*
1578 * Either remove the duplicate refcount from
1579 * isolate_lru_page() or drop the page ref if it was
1580 * not isolated.
1581 */
1582 put_page(page);
1583out:
1584 up_read(&mm->mmap_sem);
1585 return err;
1586}
1587
1588/*
1589 * Migrate an array of page address onto an array of nodes and fill
1590 * the corresponding array of status.
1591 */
1592static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1593 unsigned long nr_pages,
1594 const void __user * __user *pages,
1595 const int __user *nodes,
1596 int __user *status, int flags)
1597{
1598 int current_node = NUMA_NO_NODE;
1599 LIST_HEAD(pagelist);
1600 int start, i;
1601 int err = 0, err1;
1602
1603 migrate_prep();
1604
1605 for (i = start = 0; i < nr_pages; i++) {
1606 const void __user *p;
1607 unsigned long addr;
1608 int node;
1609
1610 err = -EFAULT;
1611 if (get_user(p, pages + i))
1612 goto out_flush;
1613 if (get_user(node, nodes + i))
1614 goto out_flush;
1615 addr = (unsigned long)untagged_addr(p);
1616
1617 err = -ENODEV;
1618 if (node < 0 || node >= MAX_NUMNODES)
1619 goto out_flush;
1620 if (!node_state(node, N_MEMORY))
1621 goto out_flush;
1622
1623 err = -EACCES;
1624 if (!node_isset(node, task_nodes))
1625 goto out_flush;
1626
1627 if (current_node == NUMA_NO_NODE) {
1628 current_node = node;
1629 start = i;
1630 } else if (node != current_node) {
1631 err = do_move_pages_to_node(mm, &pagelist, current_node);
1632 if (err)
1633 goto out;
1634 err = store_status(status, start, current_node, i - start);
1635 if (err)
1636 goto out;
1637 start = i;
1638 current_node = node;
1639 }
1640
1641 /*
1642 * Errors in the page lookup or isolation are not fatal and we simply
1643 * report them via status
1644 */
1645 err = add_page_for_migration(mm, addr, current_node,
1646 &pagelist, flags & MPOL_MF_MOVE_ALL);
1647 if (!err)
1648 continue;
1649
1650 err = store_status(status, i, err, 1);
1651 if (err)
1652 goto out_flush;
1653
1654 err = do_move_pages_to_node(mm, &pagelist, current_node);
1655 if (err)
1656 goto out;
1657 if (i > start) {
1658 err = store_status(status, start, current_node, i - start);
1659 if (err)
1660 goto out;
1661 }
1662 current_node = NUMA_NO_NODE;
1663 }
1664out_flush:
1665 if (list_empty(&pagelist))
1666 return err;
1667
1668 /* Make sure we do not overwrite the existing error */
1669 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1670 if (!err1)
1671 err1 = store_status(status, start, current_node, i - start);
1672 if (!err)
1673 err = err1;
1674out:
1675 return err;
1676}
1677
1678/*
1679 * Determine the nodes of an array of pages and store it in an array of status.
1680 */
1681static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1682 const void __user **pages, int *status)
1683{
1684 unsigned long i;
1685
1686 down_read(&mm->mmap_sem);
1687
1688 for (i = 0; i < nr_pages; i++) {
1689 unsigned long addr = (unsigned long)(*pages);
1690 struct vm_area_struct *vma;
1691 struct page *page;
1692 int err = -EFAULT;
1693
1694 vma = find_vma(mm, addr);
1695 if (!vma || addr < vma->vm_start)
1696 goto set_status;
1697
1698 /* FOLL_DUMP to ignore special (like zero) pages */
1699 page = follow_page(vma, addr, FOLL_DUMP);
1700
1701 err = PTR_ERR(page);
1702 if (IS_ERR(page))
1703 goto set_status;
1704
1705 err = page ? page_to_nid(page) : -ENOENT;
1706set_status:
1707 *status = err;
1708
1709 pages++;
1710 status++;
1711 }
1712
1713 up_read(&mm->mmap_sem);
1714}
1715
1716/*
1717 * Determine the nodes of a user array of pages and store it in
1718 * a user array of status.
1719 */
1720static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1721 const void __user * __user *pages,
1722 int __user *status)
1723{
1724#define DO_PAGES_STAT_CHUNK_NR 16
1725 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1726 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1727
1728 while (nr_pages) {
1729 unsigned long chunk_nr;
1730
1731 chunk_nr = nr_pages;
1732 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1733 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1734
1735 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1736 break;
1737
1738 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1739
1740 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1741 break;
1742
1743 pages += chunk_nr;
1744 status += chunk_nr;
1745 nr_pages -= chunk_nr;
1746 }
1747 return nr_pages ? -EFAULT : 0;
1748}
1749
1750/*
1751 * Move a list of pages in the address space of the currently executing
1752 * process.
1753 */
1754static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1755 const void __user * __user *pages,
1756 const int __user *nodes,
1757 int __user *status, int flags)
1758{
1759 struct task_struct *task;
1760 struct mm_struct *mm;
1761 int err;
1762 nodemask_t task_nodes;
1763
1764 /* Check flags */
1765 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1766 return -EINVAL;
1767
1768 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1769 return -EPERM;
1770
1771 /* Find the mm_struct */
1772 rcu_read_lock();
1773 task = pid ? find_task_by_vpid(pid) : current;
1774 if (!task) {
1775 rcu_read_unlock();
1776 return -ESRCH;
1777 }
1778 get_task_struct(task);
1779
1780 /*
1781 * Check if this process has the right to modify the specified
1782 * process. Use the regular "ptrace_may_access()" checks.
1783 */
1784 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1785 rcu_read_unlock();
1786 err = -EPERM;
1787 goto out;
1788 }
1789 rcu_read_unlock();
1790
1791 err = security_task_movememory(task);
1792 if (err)
1793 goto out;
1794
1795 task_nodes = cpuset_mems_allowed(task);
1796 mm = get_task_mm(task);
1797 put_task_struct(task);
1798
1799 if (!mm)
1800 return -EINVAL;
1801
1802 if (nodes)
1803 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1804 nodes, status, flags);
1805 else
1806 err = do_pages_stat(mm, nr_pages, pages, status);
1807
1808 mmput(mm);
1809 return err;
1810
1811out:
1812 put_task_struct(task);
1813 return err;
1814}
1815
1816SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1817 const void __user * __user *, pages,
1818 const int __user *, nodes,
1819 int __user *, status, int, flags)
1820{
1821 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1822}
1823
1824#ifdef CONFIG_COMPAT
1825COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1826 compat_uptr_t __user *, pages32,
1827 const int __user *, nodes,
1828 int __user *, status,
1829 int, flags)
1830{
1831 const void __user * __user *pages;
1832 int i;
1833
1834 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1835 for (i = 0; i < nr_pages; i++) {
1836 compat_uptr_t p;
1837
1838 if (get_user(p, pages32 + i) ||
1839 put_user(compat_ptr(p), pages + i))
1840 return -EFAULT;
1841 }
1842 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1843}
1844#endif /* CONFIG_COMPAT */
1845
1846#ifdef CONFIG_NUMA_BALANCING
1847/*
1848 * Returns true if this is a safe migration target node for misplaced NUMA
1849 * pages. Currently it only checks the watermarks which crude
1850 */
1851static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1852 unsigned long nr_migrate_pages)
1853{
1854 int z;
1855
1856 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1857 struct zone *zone = pgdat->node_zones + z;
1858
1859 if (!populated_zone(zone))
1860 continue;
1861
1862 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1863 if (!zone_watermark_ok(zone, 0,
1864 high_wmark_pages(zone) +
1865 nr_migrate_pages,
1866 0, 0))
1867 continue;
1868 return true;
1869 }
1870 return false;
1871}
1872
1873static struct page *alloc_misplaced_dst_page(struct page *page,
1874 unsigned long data)
1875{
1876 int nid = (int) data;
1877 struct page *newpage;
1878
1879 newpage = __alloc_pages_node(nid,
1880 (GFP_HIGHUSER_MOVABLE |
1881 __GFP_THISNODE | __GFP_NOMEMALLOC |
1882 __GFP_NORETRY | __GFP_NOWARN) &
1883 ~__GFP_RECLAIM, 0);
1884
1885 return newpage;
1886}
1887
1888static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1889{
1890 int page_lru;
1891
1892 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1893
1894 /* Avoid migrating to a node that is nearly full */
1895 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
1896 return 0;
1897
1898 if (isolate_lru_page(page))
1899 return 0;
1900
1901 /*
1902 * migrate_misplaced_transhuge_page() skips page migration's usual
1903 * check on page_count(), so we must do it here, now that the page
1904 * has been isolated: a GUP pin, or any other pin, prevents migration.
1905 * The expected page count is 3: 1 for page's mapcount and 1 for the
1906 * caller's pin and 1 for the reference taken by isolate_lru_page().
1907 */
1908 if (PageTransHuge(page) && page_count(page) != 3) {
1909 putback_lru_page(page);
1910 return 0;
1911 }
1912
1913 page_lru = page_is_file_cache(page);
1914 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1915 hpage_nr_pages(page));
1916
1917 /*
1918 * Isolating the page has taken another reference, so the
1919 * caller's reference can be safely dropped without the page
1920 * disappearing underneath us during migration.
1921 */
1922 put_page(page);
1923 return 1;
1924}
1925
1926bool pmd_trans_migrating(pmd_t pmd)
1927{
1928 struct page *page = pmd_page(pmd);
1929 return PageLocked(page);
1930}
1931
1932/*
1933 * Attempt to migrate a misplaced page to the specified destination
1934 * node. Caller is expected to have an elevated reference count on
1935 * the page that will be dropped by this function before returning.
1936 */
1937int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1938 int node)
1939{
1940 pg_data_t *pgdat = NODE_DATA(node);
1941 int isolated;
1942 int nr_remaining;
1943 LIST_HEAD(migratepages);
1944
1945 /*
1946 * Don't migrate file pages that are mapped in multiple processes
1947 * with execute permissions as they are probably shared libraries.
1948 */
1949 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1950 (vma->vm_flags & VM_EXEC))
1951 goto out;
1952
1953 /*
1954 * Also do not migrate dirty pages as not all filesystems can move
1955 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1956 */
1957 if (page_is_file_cache(page) && PageDirty(page))
1958 goto out;
1959
1960 isolated = numamigrate_isolate_page(pgdat, page);
1961 if (!isolated)
1962 goto out;
1963
1964 list_add(&page->lru, &migratepages);
1965 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1966 NULL, node, MIGRATE_ASYNC,
1967 MR_NUMA_MISPLACED);
1968 if (nr_remaining) {
1969 if (!list_empty(&migratepages)) {
1970 list_del(&page->lru);
1971 dec_node_page_state(page, NR_ISOLATED_ANON +
1972 page_is_file_cache(page));
1973 putback_lru_page(page);
1974 }
1975 isolated = 0;
1976 } else
1977 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1978 BUG_ON(!list_empty(&migratepages));
1979 return isolated;
1980
1981out:
1982 put_page(page);
1983 return 0;
1984}
1985#endif /* CONFIG_NUMA_BALANCING */
1986
1987#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1988/*
1989 * Migrates a THP to a given target node. page must be locked and is unlocked
1990 * before returning.
1991 */
1992int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1993 struct vm_area_struct *vma,
1994 pmd_t *pmd, pmd_t entry,
1995 unsigned long address,
1996 struct page *page, int node)
1997{
1998 spinlock_t *ptl;
1999 pg_data_t *pgdat = NODE_DATA(node);
2000 int isolated = 0;
2001 struct page *new_page = NULL;
2002 int page_lru = page_is_file_cache(page);
2003 unsigned long start = address & HPAGE_PMD_MASK;
2004
2005 new_page = alloc_pages_node(node,
2006 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2007 HPAGE_PMD_ORDER);
2008 if (!new_page)
2009 goto out_fail;
2010 prep_transhuge_page(new_page);
2011
2012 isolated = numamigrate_isolate_page(pgdat, page);
2013 if (!isolated) {
2014 put_page(new_page);
2015 goto out_fail;
2016 }
2017
2018 /* Prepare a page as a migration target */
2019 __SetPageLocked(new_page);
2020 if (PageSwapBacked(page))
2021 __SetPageSwapBacked(new_page);
2022
2023 /* anon mapping, we can simply copy page->mapping to the new page: */
2024 new_page->mapping = page->mapping;
2025 new_page->index = page->index;
2026 /* flush the cache before copying using the kernel virtual address */
2027 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2028 migrate_page_copy(new_page, page);
2029 WARN_ON(PageLRU(new_page));
2030
2031 /* Recheck the target PMD */
2032 ptl = pmd_lock(mm, pmd);
2033 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2034 spin_unlock(ptl);
2035
2036 /* Reverse changes made by migrate_page_copy() */
2037 if (TestClearPageActive(new_page))
2038 SetPageActive(page);
2039 if (TestClearPageUnevictable(new_page))
2040 SetPageUnevictable(page);
2041
2042 unlock_page(new_page);
2043 put_page(new_page); /* Free it */
2044
2045 /* Retake the callers reference and putback on LRU */
2046 get_page(page);
2047 putback_lru_page(page);
2048 mod_node_page_state(page_pgdat(page),
2049 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2050
2051 goto out_unlock;
2052 }
2053
2054 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2055 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2056
2057 /*
2058 * Overwrite the old entry under pagetable lock and establish
2059 * the new PTE. Any parallel GUP will either observe the old
2060 * page blocking on the page lock, block on the page table
2061 * lock or observe the new page. The SetPageUptodate on the
2062 * new page and page_add_new_anon_rmap guarantee the copy is
2063 * visible before the pagetable update.
2064 */
2065 page_add_anon_rmap(new_page, vma, start, true);
2066 /*
2067 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2068 * has already been flushed globally. So no TLB can be currently
2069 * caching this non present pmd mapping. There's no need to clear the
2070 * pmd before doing set_pmd_at(), nor to flush the TLB after
2071 * set_pmd_at(). Clearing the pmd here would introduce a race
2072 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2073 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2074 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2075 * pmd.
2076 */
2077 set_pmd_at(mm, start, pmd, entry);
2078 update_mmu_cache_pmd(vma, address, &entry);
2079
2080 page_ref_unfreeze(page, 2);
2081 mlock_migrate_page(new_page, page);
2082 page_remove_rmap(page, true);
2083 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2084
2085 spin_unlock(ptl);
2086
2087 /* Take an "isolate" reference and put new page on the LRU. */
2088 get_page(new_page);
2089 putback_lru_page(new_page);
2090
2091 unlock_page(new_page);
2092 unlock_page(page);
2093 put_page(page); /* Drop the rmap reference */
2094 put_page(page); /* Drop the LRU isolation reference */
2095
2096 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2097 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2098
2099 mod_node_page_state(page_pgdat(page),
2100 NR_ISOLATED_ANON + page_lru,
2101 -HPAGE_PMD_NR);
2102 return isolated;
2103
2104out_fail:
2105 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2106 ptl = pmd_lock(mm, pmd);
2107 if (pmd_same(*pmd, entry)) {
2108 entry = pmd_modify(entry, vma->vm_page_prot);
2109 set_pmd_at(mm, start, pmd, entry);
2110 update_mmu_cache_pmd(vma, address, &entry);
2111 }
2112 spin_unlock(ptl);
2113
2114out_unlock:
2115 unlock_page(page);
2116 put_page(page);
2117 return 0;
2118}
2119#endif /* CONFIG_NUMA_BALANCING */
2120
2121#endif /* CONFIG_NUMA */
2122
2123#ifdef CONFIG_DEVICE_PRIVATE
2124static int migrate_vma_collect_hole(unsigned long start,
2125 unsigned long end,
2126 struct mm_walk *walk)
2127{
2128 struct migrate_vma *migrate = walk->private;
2129 unsigned long addr;
2130
2131 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2132 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2133 migrate->dst[migrate->npages] = 0;
2134 migrate->npages++;
2135 migrate->cpages++;
2136 }
2137
2138 return 0;
2139}
2140
2141static int migrate_vma_collect_skip(unsigned long start,
2142 unsigned long end,
2143 struct mm_walk *walk)
2144{
2145 struct migrate_vma *migrate = walk->private;
2146 unsigned long addr;
2147
2148 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2149 migrate->dst[migrate->npages] = 0;
2150 migrate->src[migrate->npages++] = 0;
2151 }
2152
2153 return 0;
2154}
2155
2156static int migrate_vma_collect_pmd(pmd_t *pmdp,
2157 unsigned long start,
2158 unsigned long end,
2159 struct mm_walk *walk)
2160{
2161 struct migrate_vma *migrate = walk->private;
2162 struct vm_area_struct *vma = walk->vma;
2163 struct mm_struct *mm = vma->vm_mm;
2164 unsigned long addr = start, unmapped = 0;
2165 spinlock_t *ptl;
2166 pte_t *ptep;
2167
2168again:
2169 if (pmd_none(*pmdp))
2170 return migrate_vma_collect_hole(start, end, walk);
2171
2172 if (pmd_trans_huge(*pmdp)) {
2173 struct page *page;
2174
2175 ptl = pmd_lock(mm, pmdp);
2176 if (unlikely(!pmd_trans_huge(*pmdp))) {
2177 spin_unlock(ptl);
2178 goto again;
2179 }
2180
2181 page = pmd_page(*pmdp);
2182 if (is_huge_zero_page(page)) {
2183 spin_unlock(ptl);
2184 split_huge_pmd(vma, pmdp, addr);
2185 if (pmd_trans_unstable(pmdp))
2186 return migrate_vma_collect_skip(start, end,
2187 walk);
2188 } else {
2189 int ret;
2190
2191 get_page(page);
2192 spin_unlock(ptl);
2193 if (unlikely(!trylock_page(page)))
2194 return migrate_vma_collect_skip(start, end,
2195 walk);
2196 ret = split_huge_page(page);
2197 unlock_page(page);
2198 put_page(page);
2199 if (ret)
2200 return migrate_vma_collect_skip(start, end,
2201 walk);
2202 if (pmd_none(*pmdp))
2203 return migrate_vma_collect_hole(start, end,
2204 walk);
2205 }
2206 }
2207
2208 if (unlikely(pmd_bad(*pmdp)))
2209 return migrate_vma_collect_skip(start, end, walk);
2210
2211 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2212 arch_enter_lazy_mmu_mode();
2213
2214 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2215 unsigned long mpfn, pfn;
2216 struct page *page;
2217 swp_entry_t entry;
2218 pte_t pte;
2219
2220 pte = *ptep;
2221
2222 if (pte_none(pte)) {
2223 mpfn = MIGRATE_PFN_MIGRATE;
2224 migrate->cpages++;
2225 goto next;
2226 }
2227
2228 if (!pte_present(pte)) {
2229 mpfn = 0;
2230
2231 /*
2232 * Only care about unaddressable device page special
2233 * page table entry. Other special swap entries are not
2234 * migratable, and we ignore regular swapped page.
2235 */
2236 entry = pte_to_swp_entry(pte);
2237 if (!is_device_private_entry(entry))
2238 goto next;
2239
2240 page = device_private_entry_to_page(entry);
2241 mpfn = migrate_pfn(page_to_pfn(page)) |
2242 MIGRATE_PFN_MIGRATE;
2243 if (is_write_device_private_entry(entry))
2244 mpfn |= MIGRATE_PFN_WRITE;
2245 } else {
2246 pfn = pte_pfn(pte);
2247 if (is_zero_pfn(pfn)) {
2248 mpfn = MIGRATE_PFN_MIGRATE;
2249 migrate->cpages++;
2250 goto next;
2251 }
2252 page = vm_normal_page(migrate->vma, addr, pte);
2253 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2254 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2255 }
2256
2257 /* FIXME support THP */
2258 if (!page || !page->mapping || PageTransCompound(page)) {
2259 mpfn = 0;
2260 goto next;
2261 }
2262
2263 /*
2264 * By getting a reference on the page we pin it and that blocks
2265 * any kind of migration. Side effect is that it "freezes" the
2266 * pte.
2267 *
2268 * We drop this reference after isolating the page from the lru
2269 * for non device page (device page are not on the lru and thus
2270 * can't be dropped from it).
2271 */
2272 get_page(page);
2273 migrate->cpages++;
2274
2275 /*
2276 * Optimize for the common case where page is only mapped once
2277 * in one process. If we can lock the page, then we can safely
2278 * set up a special migration page table entry now.
2279 */
2280 if (trylock_page(page)) {
2281 pte_t swp_pte;
2282
2283 mpfn |= MIGRATE_PFN_LOCKED;
2284 ptep_get_and_clear(mm, addr, ptep);
2285
2286 /* Setup special migration page table entry */
2287 entry = make_migration_entry(page, mpfn &
2288 MIGRATE_PFN_WRITE);
2289 swp_pte = swp_entry_to_pte(entry);
2290 if (pte_soft_dirty(pte))
2291 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2292 set_pte_at(mm, addr, ptep, swp_pte);
2293
2294 /*
2295 * This is like regular unmap: we remove the rmap and
2296 * drop page refcount. Page won't be freed, as we took
2297 * a reference just above.
2298 */
2299 page_remove_rmap(page, false);
2300 put_page(page);
2301
2302 if (pte_present(pte))
2303 unmapped++;
2304 }
2305
2306next:
2307 migrate->dst[migrate->npages] = 0;
2308 migrate->src[migrate->npages++] = mpfn;
2309 }
2310 arch_leave_lazy_mmu_mode();
2311 pte_unmap_unlock(ptep - 1, ptl);
2312
2313 /* Only flush the TLB if we actually modified any entries */
2314 if (unmapped)
2315 flush_tlb_range(walk->vma, start, end);
2316
2317 return 0;
2318}
2319
2320static const struct mm_walk_ops migrate_vma_walk_ops = {
2321 .pmd_entry = migrate_vma_collect_pmd,
2322 .pte_hole = migrate_vma_collect_hole,
2323};
2324
2325/*
2326 * migrate_vma_collect() - collect pages over a range of virtual addresses
2327 * @migrate: migrate struct containing all migration information
2328 *
2329 * This will walk the CPU page table. For each virtual address backed by a
2330 * valid page, it updates the src array and takes a reference on the page, in
2331 * order to pin the page until we lock it and unmap it.
2332 */
2333static void migrate_vma_collect(struct migrate_vma *migrate)
2334{
2335 struct mmu_notifier_range range;
2336
2337 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL,
2338 migrate->vma->vm_mm, migrate->start, migrate->end);
2339 mmu_notifier_invalidate_range_start(&range);
2340
2341 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2342 &migrate_vma_walk_ops, migrate);
2343
2344 mmu_notifier_invalidate_range_end(&range);
2345 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2346}
2347
2348/*
2349 * migrate_vma_check_page() - check if page is pinned or not
2350 * @page: struct page to check
2351 *
2352 * Pinned pages cannot be migrated. This is the same test as in
2353 * migrate_page_move_mapping(), except that here we allow migration of a
2354 * ZONE_DEVICE page.
2355 */
2356static bool migrate_vma_check_page(struct page *page)
2357{
2358 /*
2359 * One extra ref because caller holds an extra reference, either from
2360 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2361 * a device page.
2362 */
2363 int extra = 1;
2364
2365 /*
2366 * FIXME support THP (transparent huge page), it is bit more complex to
2367 * check them than regular pages, because they can be mapped with a pmd
2368 * or with a pte (split pte mapping).
2369 */
2370 if (PageCompound(page))
2371 return false;
2372
2373 /* Page from ZONE_DEVICE have one extra reference */
2374 if (is_zone_device_page(page)) {
2375 /*
2376 * Private page can never be pin as they have no valid pte and
2377 * GUP will fail for those. Yet if there is a pending migration
2378 * a thread might try to wait on the pte migration entry and
2379 * will bump the page reference count. Sadly there is no way to
2380 * differentiate a regular pin from migration wait. Hence to
2381 * avoid 2 racing thread trying to migrate back to CPU to enter
2382 * infinite loop (one stoping migration because the other is
2383 * waiting on pte migration entry). We always return true here.
2384 *
2385 * FIXME proper solution is to rework migration_entry_wait() so
2386 * it does not need to take a reference on page.
2387 */
2388 return is_device_private_page(page);
2389 }
2390
2391 /* For file back page */
2392 if (page_mapping(page))
2393 extra += 1 + page_has_private(page);
2394
2395 if ((page_count(page) - extra) > page_mapcount(page))
2396 return false;
2397
2398 return true;
2399}
2400
2401/*
2402 * migrate_vma_prepare() - lock pages and isolate them from the lru
2403 * @migrate: migrate struct containing all migration information
2404 *
2405 * This locks pages that have been collected by migrate_vma_collect(). Once each
2406 * page is locked it is isolated from the lru (for non-device pages). Finally,
2407 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2408 * migrated by concurrent kernel threads.
2409 */
2410static void migrate_vma_prepare(struct migrate_vma *migrate)
2411{
2412 const unsigned long npages = migrate->npages;
2413 const unsigned long start = migrate->start;
2414 unsigned long addr, i, restore = 0;
2415 bool allow_drain = true;
2416
2417 lru_add_drain();
2418
2419 for (i = 0; (i < npages) && migrate->cpages; i++) {
2420 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2421 bool remap = true;
2422
2423 if (!page)
2424 continue;
2425
2426 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2427 /*
2428 * Because we are migrating several pages there can be
2429 * a deadlock between 2 concurrent migration where each
2430 * are waiting on each other page lock.
2431 *
2432 * Make migrate_vma() a best effort thing and backoff
2433 * for any page we can not lock right away.
2434 */
2435 if (!trylock_page(page)) {
2436 migrate->src[i] = 0;
2437 migrate->cpages--;
2438 put_page(page);
2439 continue;
2440 }
2441 remap = false;
2442 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2443 }
2444
2445 /* ZONE_DEVICE pages are not on LRU */
2446 if (!is_zone_device_page(page)) {
2447 if (!PageLRU(page) && allow_drain) {
2448 /* Drain CPU's pagevec */
2449 lru_add_drain_all();
2450 allow_drain = false;
2451 }
2452
2453 if (isolate_lru_page(page)) {
2454 if (remap) {
2455 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2456 migrate->cpages--;
2457 restore++;
2458 } else {
2459 migrate->src[i] = 0;
2460 unlock_page(page);
2461 migrate->cpages--;
2462 put_page(page);
2463 }
2464 continue;
2465 }
2466
2467 /* Drop the reference we took in collect */
2468 put_page(page);
2469 }
2470
2471 if (!migrate_vma_check_page(page)) {
2472 if (remap) {
2473 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2474 migrate->cpages--;
2475 restore++;
2476
2477 if (!is_zone_device_page(page)) {
2478 get_page(page);
2479 putback_lru_page(page);
2480 }
2481 } else {
2482 migrate->src[i] = 0;
2483 unlock_page(page);
2484 migrate->cpages--;
2485
2486 if (!is_zone_device_page(page))
2487 putback_lru_page(page);
2488 else
2489 put_page(page);
2490 }
2491 }
2492 }
2493
2494 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2495 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2496
2497 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2498 continue;
2499
2500 remove_migration_pte(page, migrate->vma, addr, page);
2501
2502 migrate->src[i] = 0;
2503 unlock_page(page);
2504 put_page(page);
2505 restore--;
2506 }
2507}
2508
2509/*
2510 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2511 * @migrate: migrate struct containing all migration information
2512 *
2513 * Replace page mapping (CPU page table pte) with a special migration pte entry
2514 * and check again if it has been pinned. Pinned pages are restored because we
2515 * cannot migrate them.
2516 *
2517 * This is the last step before we call the device driver callback to allocate
2518 * destination memory and copy contents of original page over to new page.
2519 */
2520static void migrate_vma_unmap(struct migrate_vma *migrate)
2521{
2522 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2523 const unsigned long npages = migrate->npages;
2524 const unsigned long start = migrate->start;
2525 unsigned long addr, i, restore = 0;
2526
2527 for (i = 0; i < npages; i++) {
2528 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2529
2530 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2531 continue;
2532
2533 if (page_mapped(page)) {
2534 try_to_unmap(page, flags);
2535 if (page_mapped(page))
2536 goto restore;
2537 }
2538
2539 if (migrate_vma_check_page(page))
2540 continue;
2541
2542restore:
2543 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2544 migrate->cpages--;
2545 restore++;
2546 }
2547
2548 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2549 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2550
2551 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2552 continue;
2553
2554 remove_migration_ptes(page, page, false);
2555
2556 migrate->src[i] = 0;
2557 unlock_page(page);
2558 restore--;
2559
2560 if (is_zone_device_page(page))
2561 put_page(page);
2562 else
2563 putback_lru_page(page);
2564 }
2565}
2566
2567/**
2568 * migrate_vma_setup() - prepare to migrate a range of memory
2569 * @args: contains the vma, start, and and pfns arrays for the migration
2570 *
2571 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2572 * without an error.
2573 *
2574 * Prepare to migrate a range of memory virtual address range by collecting all
2575 * the pages backing each virtual address in the range, saving them inside the
2576 * src array. Then lock those pages and unmap them. Once the pages are locked
2577 * and unmapped, check whether each page is pinned or not. Pages that aren't
2578 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2579 * corresponding src array entry. Then restores any pages that are pinned, by
2580 * remapping and unlocking those pages.
2581 *
2582 * The caller should then allocate destination memory and copy source memory to
2583 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2584 * flag set). Once these are allocated and copied, the caller must update each
2585 * corresponding entry in the dst array with the pfn value of the destination
2586 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2587 * (destination pages must have their struct pages locked, via lock_page()).
2588 *
2589 * Note that the caller does not have to migrate all the pages that are marked
2590 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2591 * device memory to system memory. If the caller cannot migrate a device page
2592 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2593 * consequences for the userspace process, so it must be avoided if at all
2594 * possible.
2595 *
2596 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2597 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2598 * allowing the caller to allocate device memory for those unback virtual
2599 * address. For this the caller simply has to allocate device memory and
2600 * properly set the destination entry like for regular migration. Note that
2601 * this can still fails and thus inside the device driver must check if the
2602 * migration was successful for those entries after calling migrate_vma_pages()
2603 * just like for regular migration.
2604 *
2605 * After that, the callers must call migrate_vma_pages() to go over each entry
2606 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2607 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2608 * then migrate_vma_pages() to migrate struct page information from the source
2609 * struct page to the destination struct page. If it fails to migrate the
2610 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2611 * src array.
2612 *
2613 * At this point all successfully migrated pages have an entry in the src
2614 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2615 * array entry with MIGRATE_PFN_VALID flag set.
2616 *
2617 * Once migrate_vma_pages() returns the caller may inspect which pages were
2618 * successfully migrated, and which were not. Successfully migrated pages will
2619 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2620 *
2621 * It is safe to update device page table after migrate_vma_pages() because
2622 * both destination and source page are still locked, and the mmap_sem is held
2623 * in read mode (hence no one can unmap the range being migrated).
2624 *
2625 * Once the caller is done cleaning up things and updating its page table (if it
2626 * chose to do so, this is not an obligation) it finally calls
2627 * migrate_vma_finalize() to update the CPU page table to point to new pages
2628 * for successfully migrated pages or otherwise restore the CPU page table to
2629 * point to the original source pages.
2630 */
2631int migrate_vma_setup(struct migrate_vma *args)
2632{
2633 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2634
2635 args->start &= PAGE_MASK;
2636 args->end &= PAGE_MASK;
2637 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2638 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2639 return -EINVAL;
2640 if (nr_pages <= 0)
2641 return -EINVAL;
2642 if (args->start < args->vma->vm_start ||
2643 args->start >= args->vma->vm_end)
2644 return -EINVAL;
2645 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2646 return -EINVAL;
2647 if (!args->src || !args->dst)
2648 return -EINVAL;
2649
2650 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2651 args->cpages = 0;
2652 args->npages = 0;
2653
2654 migrate_vma_collect(args);
2655
2656 if (args->cpages)
2657 migrate_vma_prepare(args);
2658 if (args->cpages)
2659 migrate_vma_unmap(args);
2660
2661 /*
2662 * At this point pages are locked and unmapped, and thus they have
2663 * stable content and can safely be copied to destination memory that
2664 * is allocated by the drivers.
2665 */
2666 return 0;
2667
2668}
2669EXPORT_SYMBOL(migrate_vma_setup);
2670
2671static void migrate_vma_insert_page(struct migrate_vma *migrate,
2672 unsigned long addr,
2673 struct page *page,
2674 unsigned long *src,
2675 unsigned long *dst)
2676{
2677 struct vm_area_struct *vma = migrate->vma;
2678 struct mm_struct *mm = vma->vm_mm;
2679 struct mem_cgroup *memcg;
2680 bool flush = false;
2681 spinlock_t *ptl;
2682 pte_t entry;
2683 pgd_t *pgdp;
2684 p4d_t *p4dp;
2685 pud_t *pudp;
2686 pmd_t *pmdp;
2687 pte_t *ptep;
2688
2689 /* Only allow populating anonymous memory */
2690 if (!vma_is_anonymous(vma))
2691 goto abort;
2692
2693 pgdp = pgd_offset(mm, addr);
2694 p4dp = p4d_alloc(mm, pgdp, addr);
2695 if (!p4dp)
2696 goto abort;
2697 pudp = pud_alloc(mm, p4dp, addr);
2698 if (!pudp)
2699 goto abort;
2700 pmdp = pmd_alloc(mm, pudp, addr);
2701 if (!pmdp)
2702 goto abort;
2703
2704 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2705 goto abort;
2706
2707 /*
2708 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2709 * pte_offset_map() on pmds where a huge pmd might be created
2710 * from a different thread.
2711 *
2712 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2713 * parallel threads are excluded by other means.
2714 *
2715 * Here we only have down_read(mmap_sem).
2716 */
2717 if (pte_alloc(mm, pmdp))
2718 goto abort;
2719
2720 /* See the comment in pte_alloc_one_map() */
2721 if (unlikely(pmd_trans_unstable(pmdp)))
2722 goto abort;
2723
2724 if (unlikely(anon_vma_prepare(vma)))
2725 goto abort;
2726 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2727 goto abort;
2728
2729 /*
2730 * The memory barrier inside __SetPageUptodate makes sure that
2731 * preceding stores to the page contents become visible before
2732 * the set_pte_at() write.
2733 */
2734 __SetPageUptodate(page);
2735
2736 if (is_zone_device_page(page)) {
2737 if (is_device_private_page(page)) {
2738 swp_entry_t swp_entry;
2739
2740 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2741 entry = swp_entry_to_pte(swp_entry);
2742 }
2743 } else {
2744 entry = mk_pte(page, vma->vm_page_prot);
2745 if (vma->vm_flags & VM_WRITE)
2746 entry = pte_mkwrite(pte_mkdirty(entry));
2747 }
2748
2749 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2750
2751 if (pte_present(*ptep)) {
2752 unsigned long pfn = pte_pfn(*ptep);
2753
2754 if (!is_zero_pfn(pfn)) {
2755 pte_unmap_unlock(ptep, ptl);
2756 mem_cgroup_cancel_charge(page, memcg, false);
2757 goto abort;
2758 }
2759 flush = true;
2760 } else if (!pte_none(*ptep)) {
2761 pte_unmap_unlock(ptep, ptl);
2762 mem_cgroup_cancel_charge(page, memcg, false);
2763 goto abort;
2764 }
2765
2766 /*
2767 * Check for usefaultfd but do not deliver the fault. Instead,
2768 * just back off.
2769 */
2770 if (userfaultfd_missing(vma)) {
2771 pte_unmap_unlock(ptep, ptl);
2772 mem_cgroup_cancel_charge(page, memcg, false);
2773 goto abort;
2774 }
2775
2776 inc_mm_counter(mm, MM_ANONPAGES);
2777 page_add_new_anon_rmap(page, vma, addr, false);
2778 mem_cgroup_commit_charge(page, memcg, false, false);
2779 if (!is_zone_device_page(page))
2780 lru_cache_add_active_or_unevictable(page, vma);
2781 get_page(page);
2782
2783 if (flush) {
2784 flush_cache_page(vma, addr, pte_pfn(*ptep));
2785 ptep_clear_flush_notify(vma, addr, ptep);
2786 set_pte_at_notify(mm, addr, ptep, entry);
2787 update_mmu_cache(vma, addr, ptep);
2788 } else {
2789 /* No need to invalidate - it was non-present before */
2790 set_pte_at(mm, addr, ptep, entry);
2791 update_mmu_cache(vma, addr, ptep);
2792 }
2793
2794 pte_unmap_unlock(ptep, ptl);
2795 *src = MIGRATE_PFN_MIGRATE;
2796 return;
2797
2798abort:
2799 *src &= ~MIGRATE_PFN_MIGRATE;
2800}
2801
2802/**
2803 * migrate_vma_pages() - migrate meta-data from src page to dst page
2804 * @migrate: migrate struct containing all migration information
2805 *
2806 * This migrates struct page meta-data from source struct page to destination
2807 * struct page. This effectively finishes the migration from source page to the
2808 * destination page.
2809 */
2810void migrate_vma_pages(struct migrate_vma *migrate)
2811{
2812 const unsigned long npages = migrate->npages;
2813 const unsigned long start = migrate->start;
2814 struct mmu_notifier_range range;
2815 unsigned long addr, i;
2816 bool notified = false;
2817
2818 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2819 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2820 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2821 struct address_space *mapping;
2822 int r;
2823
2824 if (!newpage) {
2825 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2826 continue;
2827 }
2828
2829 if (!page) {
2830 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2831 continue;
2832 }
2833 if (!notified) {
2834 notified = true;
2835
2836 mmu_notifier_range_init(&range,
2837 MMU_NOTIFY_CLEAR, 0,
2838 NULL,
2839 migrate->vma->vm_mm,
2840 addr, migrate->end);
2841 mmu_notifier_invalidate_range_start(&range);
2842 }
2843 migrate_vma_insert_page(migrate, addr, newpage,
2844 &migrate->src[i],
2845 &migrate->dst[i]);
2846 continue;
2847 }
2848
2849 mapping = page_mapping(page);
2850
2851 if (is_zone_device_page(newpage)) {
2852 if (is_device_private_page(newpage)) {
2853 /*
2854 * For now only support private anonymous when
2855 * migrating to un-addressable device memory.
2856 */
2857 if (mapping) {
2858 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2859 continue;
2860 }
2861 } else {
2862 /*
2863 * Other types of ZONE_DEVICE page are not
2864 * supported.
2865 */
2866 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2867 continue;
2868 }
2869 }
2870
2871 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2872 if (r != MIGRATEPAGE_SUCCESS)
2873 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2874 }
2875
2876 /*
2877 * No need to double call mmu_notifier->invalidate_range() callback as
2878 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2879 * did already call it.
2880 */
2881 if (notified)
2882 mmu_notifier_invalidate_range_only_end(&range);
2883}
2884EXPORT_SYMBOL(migrate_vma_pages);
2885
2886/**
2887 * migrate_vma_finalize() - restore CPU page table entry
2888 * @migrate: migrate struct containing all migration information
2889 *
2890 * This replaces the special migration pte entry with either a mapping to the
2891 * new page if migration was successful for that page, or to the original page
2892 * otherwise.
2893 *
2894 * This also unlocks the pages and puts them back on the lru, or drops the extra
2895 * refcount, for device pages.
2896 */
2897void migrate_vma_finalize(struct migrate_vma *migrate)
2898{
2899 const unsigned long npages = migrate->npages;
2900 unsigned long i;
2901
2902 for (i = 0; i < npages; i++) {
2903 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2904 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2905
2906 if (!page) {
2907 if (newpage) {
2908 unlock_page(newpage);
2909 put_page(newpage);
2910 }
2911 continue;
2912 }
2913
2914 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2915 if (newpage) {
2916 unlock_page(newpage);
2917 put_page(newpage);
2918 }
2919 newpage = page;
2920 }
2921
2922 remove_migration_ptes(page, newpage, false);
2923 unlock_page(page);
2924 migrate->cpages--;
2925
2926 if (is_zone_device_page(page))
2927 put_page(page);
2928 else
2929 putback_lru_page(page);
2930
2931 if (newpage != page) {
2932 unlock_page(newpage);
2933 if (is_zone_device_page(newpage))
2934 put_page(newpage);
2935 else
2936 putback_lru_page(newpage);
2937 }
2938 }
2939}
2940EXPORT_SYMBOL(migrate_vma_finalize);
2941#endif /* CONFIG_DEVICE_PRIVATE */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory Migration functionality - linux/mm/migrate.c
4 *
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 *
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
9 *
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
13 * Christoph Lameter
14 */
15
16#include <linux/migrate.h>
17#include <linux/export.h>
18#include <linux/swap.h>
19#include <linux/swapops.h>
20#include <linux/pagemap.h>
21#include <linux/buffer_head.h>
22#include <linux/mm_inline.h>
23#include <linux/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 if (pte_swp_soft_dirty(old_pte))
215 pte = pte_mksoft_dirty(pte);
216
217 entry = pte_to_swp_entry(old_pte);
218 if (!is_migration_entry_young(entry))
219 pte = pte_mkold(pte);
220 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
221 pte = pte_mkdirty(pte);
222 if (is_writable_migration_entry(entry))
223 pte = pte_mkwrite(pte, vma);
224 else if (pte_swp_uffd_wp(old_pte))
225 pte = pte_mkuffd_wp(pte);
226
227 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
228 rmap_flags |= RMAP_EXCLUSIVE;
229
230 if (unlikely(is_device_private_page(new))) {
231 if (pte_write(pte))
232 entry = make_writable_device_private_entry(
233 page_to_pfn(new));
234 else
235 entry = make_readable_device_private_entry(
236 page_to_pfn(new));
237 pte = swp_entry_to_pte(entry);
238 if (pte_swp_soft_dirty(old_pte))
239 pte = pte_swp_mksoft_dirty(pte);
240 if (pte_swp_uffd_wp(old_pte))
241 pte = pte_swp_mkuffd_wp(pte);
242 }
243
244#ifdef CONFIG_HUGETLB_PAGE
245 if (folio_test_hugetlb(folio)) {
246 struct hstate *h = hstate_vma(vma);
247 unsigned int shift = huge_page_shift(h);
248 unsigned long psize = huge_page_size(h);
249
250 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
251 if (folio_test_anon(folio))
252 hugetlb_add_anon_rmap(folio, vma, pvmw.address,
253 rmap_flags);
254 else
255 hugetlb_add_file_rmap(folio);
256 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
257 psize);
258 } else
259#endif
260 {
261 if (folio_test_anon(folio))
262 folio_add_anon_rmap_pte(folio, new, vma,
263 pvmw.address, rmap_flags);
264 else
265 folio_add_file_rmap_pte(folio, new, vma);
266 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
267 }
268 if (vma->vm_flags & VM_LOCKED)
269 mlock_drain_local();
270
271 trace_remove_migration_pte(pvmw.address, pte_val(pte),
272 compound_order(new));
273
274 /* No need to invalidate - it was non-present before */
275 update_mmu_cache(vma, pvmw.address, pvmw.pte);
276 }
277
278 return true;
279}
280
281/*
282 * Get rid of all migration entries and replace them by
283 * references to the indicated page.
284 */
285void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
286{
287 struct rmap_walk_control rwc = {
288 .rmap_one = remove_migration_pte,
289 .arg = src,
290 };
291
292 if (locked)
293 rmap_walk_locked(dst, &rwc);
294 else
295 rmap_walk(dst, &rwc);
296}
297
298/*
299 * Something used the pte of a page under migration. We need to
300 * get to the page and wait until migration is finished.
301 * When we return from this function the fault will be retried.
302 */
303void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
304 unsigned long address)
305{
306 spinlock_t *ptl;
307 pte_t *ptep;
308 pte_t pte;
309 swp_entry_t entry;
310
311 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
312 if (!ptep)
313 return;
314
315 pte = ptep_get(ptep);
316 pte_unmap(ptep);
317
318 if (!is_swap_pte(pte))
319 goto out;
320
321 entry = pte_to_swp_entry(pte);
322 if (!is_migration_entry(entry))
323 goto out;
324
325 migration_entry_wait_on_locked(entry, ptl);
326 return;
327out:
328 spin_unlock(ptl);
329}
330
331#ifdef CONFIG_HUGETLB_PAGE
332/*
333 * The vma read lock must be held upon entry. Holding that lock prevents either
334 * the pte or the ptl from being freed.
335 *
336 * This function will release the vma lock before returning.
337 */
338void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
339{
340 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
341 pte_t pte;
342
343 hugetlb_vma_assert_locked(vma);
344 spin_lock(ptl);
345 pte = huge_ptep_get(ptep);
346
347 if (unlikely(!is_hugetlb_entry_migration(pte))) {
348 spin_unlock(ptl);
349 hugetlb_vma_unlock_read(vma);
350 } else {
351 /*
352 * If migration entry existed, safe to release vma lock
353 * here because the pgtable page won't be freed without the
354 * pgtable lock released. See comment right above pgtable
355 * lock release in migration_entry_wait_on_locked().
356 */
357 hugetlb_vma_unlock_read(vma);
358 migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
359 }
360}
361#endif
362
363#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
364void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
365{
366 spinlock_t *ptl;
367
368 ptl = pmd_lock(mm, pmd);
369 if (!is_pmd_migration_entry(*pmd))
370 goto unlock;
371 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
372 return;
373unlock:
374 spin_unlock(ptl);
375}
376#endif
377
378static int folio_expected_refs(struct address_space *mapping,
379 struct folio *folio)
380{
381 int refs = 1;
382 if (!mapping)
383 return refs;
384
385 refs += folio_nr_pages(folio);
386 if (folio_test_private(folio))
387 refs++;
388
389 return refs;
390}
391
392/*
393 * Replace the page in the mapping.
394 *
395 * The number of remaining references must be:
396 * 1 for anonymous pages without a mapping
397 * 2 for pages with a mapping
398 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
399 */
400int folio_migrate_mapping(struct address_space *mapping,
401 struct folio *newfolio, struct folio *folio, int extra_count)
402{
403 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
404 struct zone *oldzone, *newzone;
405 int dirty;
406 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
407 long nr = folio_nr_pages(folio);
408 long entries, i;
409
410 if (!mapping) {
411 /* Anonymous page without mapping */
412 if (folio_ref_count(folio) != expected_count)
413 return -EAGAIN;
414
415 /* No turning back from here */
416 newfolio->index = folio->index;
417 newfolio->mapping = folio->mapping;
418 if (folio_test_swapbacked(folio))
419 __folio_set_swapbacked(newfolio);
420
421 return MIGRATEPAGE_SUCCESS;
422 }
423
424 oldzone = folio_zone(folio);
425 newzone = folio_zone(newfolio);
426
427 xas_lock_irq(&xas);
428 if (!folio_ref_freeze(folio, expected_count)) {
429 xas_unlock_irq(&xas);
430 return -EAGAIN;
431 }
432
433 /*
434 * Now we know that no one else is looking at the folio:
435 * no turning back from here.
436 */
437 newfolio->index = folio->index;
438 newfolio->mapping = folio->mapping;
439 folio_ref_add(newfolio, nr); /* add cache reference */
440 if (folio_test_swapbacked(folio)) {
441 __folio_set_swapbacked(newfolio);
442 if (folio_test_swapcache(folio)) {
443 folio_set_swapcache(newfolio);
444 newfolio->private = folio_get_private(folio);
445 }
446 entries = nr;
447 } else {
448 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
449 entries = 1;
450 }
451
452 /* Move dirty while page refs frozen and newpage not yet exposed */
453 dirty = folio_test_dirty(folio);
454 if (dirty) {
455 folio_clear_dirty(folio);
456 folio_set_dirty(newfolio);
457 }
458
459 /* Swap cache still stores N entries instead of a high-order entry */
460 for (i = 0; i < entries; i++) {
461 xas_store(&xas, newfolio);
462 xas_next(&xas);
463 }
464
465 /*
466 * Drop cache reference from old page by unfreezing
467 * to one less reference.
468 * We know this isn't the last reference.
469 */
470 folio_ref_unfreeze(folio, expected_count - nr);
471
472 xas_unlock(&xas);
473 /* Leave irq disabled to prevent preemption while updating stats */
474
475 /*
476 * If moved to a different zone then also account
477 * the page for that zone. Other VM counters will be
478 * taken care of when we establish references to the
479 * new page and drop references to the old page.
480 *
481 * Note that anonymous pages are accounted for
482 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
483 * are mapped to swap space.
484 */
485 if (newzone != oldzone) {
486 struct lruvec *old_lruvec, *new_lruvec;
487 struct mem_cgroup *memcg;
488
489 memcg = folio_memcg(folio);
490 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
491 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
492
493 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
494 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
495 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
496 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
497 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
498
499 if (folio_test_pmd_mappable(folio)) {
500 __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
501 __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
502 }
503 }
504#ifdef CONFIG_SWAP
505 if (folio_test_swapcache(folio)) {
506 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
507 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
508 }
509#endif
510 if (dirty && mapping_can_writeback(mapping)) {
511 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
512 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
513 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
514 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
515 }
516 }
517 local_irq_enable();
518
519 return MIGRATEPAGE_SUCCESS;
520}
521EXPORT_SYMBOL(folio_migrate_mapping);
522
523/*
524 * The expected number of remaining references is the same as that
525 * of folio_migrate_mapping().
526 */
527int migrate_huge_page_move_mapping(struct address_space *mapping,
528 struct folio *dst, struct folio *src)
529{
530 XA_STATE(xas, &mapping->i_pages, folio_index(src));
531 int expected_count;
532
533 xas_lock_irq(&xas);
534 expected_count = folio_expected_refs(mapping, src);
535 if (!folio_ref_freeze(src, expected_count)) {
536 xas_unlock_irq(&xas);
537 return -EAGAIN;
538 }
539
540 dst->index = src->index;
541 dst->mapping = src->mapping;
542
543 folio_ref_add(dst, folio_nr_pages(dst));
544
545 xas_store(&xas, dst);
546
547 folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
548
549 xas_unlock_irq(&xas);
550
551 return MIGRATEPAGE_SUCCESS;
552}
553
554/*
555 * Copy the flags and some other ancillary information
556 */
557void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
558{
559 int cpupid;
560
561 if (folio_test_error(folio))
562 folio_set_error(newfolio);
563 if (folio_test_referenced(folio))
564 folio_set_referenced(newfolio);
565 if (folio_test_uptodate(folio))
566 folio_mark_uptodate(newfolio);
567 if (folio_test_clear_active(folio)) {
568 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
569 folio_set_active(newfolio);
570 } else if (folio_test_clear_unevictable(folio))
571 folio_set_unevictable(newfolio);
572 if (folio_test_workingset(folio))
573 folio_set_workingset(newfolio);
574 if (folio_test_checked(folio))
575 folio_set_checked(newfolio);
576 /*
577 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
578 * migration entries. We can still have PG_anon_exclusive set on an
579 * effectively unmapped and unreferenced first sub-pages of an
580 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
581 */
582 if (folio_test_mappedtodisk(folio))
583 folio_set_mappedtodisk(newfolio);
584
585 /* Move dirty on pages not done by folio_migrate_mapping() */
586 if (folio_test_dirty(folio))
587 folio_set_dirty(newfolio);
588
589 if (folio_test_young(folio))
590 folio_set_young(newfolio);
591 if (folio_test_idle(folio))
592 folio_set_idle(newfolio);
593
594 /*
595 * Copy NUMA information to the new page, to prevent over-eager
596 * future migrations of this same page.
597 */
598 cpupid = folio_xchg_last_cpupid(folio, -1);
599 /*
600 * For memory tiering mode, when migrate between slow and fast
601 * memory node, reset cpupid, because that is used to record
602 * page access time in slow memory node.
603 */
604 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
605 bool f_toptier = node_is_toptier(folio_nid(folio));
606 bool t_toptier = node_is_toptier(folio_nid(newfolio));
607
608 if (f_toptier != t_toptier)
609 cpupid = -1;
610 }
611 folio_xchg_last_cpupid(newfolio, cpupid);
612
613 folio_migrate_ksm(newfolio, folio);
614 /*
615 * Please do not reorder this without considering how mm/ksm.c's
616 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
617 */
618 if (folio_test_swapcache(folio))
619 folio_clear_swapcache(folio);
620 folio_clear_private(folio);
621
622 /* page->private contains hugetlb specific flags */
623 if (!folio_test_hugetlb(folio))
624 folio->private = NULL;
625
626 /*
627 * If any waiters have accumulated on the new page then
628 * wake them up.
629 */
630 if (folio_test_writeback(newfolio))
631 folio_end_writeback(newfolio);
632
633 /*
634 * PG_readahead shares the same bit with PG_reclaim. The above
635 * end_page_writeback() may clear PG_readahead mistakenly, so set the
636 * bit after that.
637 */
638 if (folio_test_readahead(folio))
639 folio_set_readahead(newfolio);
640
641 folio_copy_owner(newfolio, folio);
642
643 mem_cgroup_migrate(folio, newfolio);
644}
645EXPORT_SYMBOL(folio_migrate_flags);
646
647void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
648{
649 folio_copy(newfolio, folio);
650 folio_migrate_flags(newfolio, folio);
651}
652EXPORT_SYMBOL(folio_migrate_copy);
653
654/************************************************************
655 * Migration functions
656 ***********************************************************/
657
658int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
659 struct folio *src, enum migrate_mode mode, int extra_count)
660{
661 int rc;
662
663 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
664
665 rc = folio_migrate_mapping(mapping, dst, src, extra_count);
666
667 if (rc != MIGRATEPAGE_SUCCESS)
668 return rc;
669
670 if (mode != MIGRATE_SYNC_NO_COPY)
671 folio_migrate_copy(dst, src);
672 else
673 folio_migrate_flags(dst, src);
674 return MIGRATEPAGE_SUCCESS;
675}
676
677/**
678 * migrate_folio() - Simple folio migration.
679 * @mapping: The address_space containing the folio.
680 * @dst: The folio to migrate the data to.
681 * @src: The folio containing the current data.
682 * @mode: How to migrate the page.
683 *
684 * Common logic to directly migrate a single LRU folio suitable for
685 * folios that do not use PagePrivate/PagePrivate2.
686 *
687 * Folios are locked upon entry and exit.
688 */
689int migrate_folio(struct address_space *mapping, struct folio *dst,
690 struct folio *src, enum migrate_mode mode)
691{
692 return migrate_folio_extra(mapping, dst, src, mode, 0);
693}
694EXPORT_SYMBOL(migrate_folio);
695
696#ifdef CONFIG_BUFFER_HEAD
697/* Returns true if all buffers are successfully locked */
698static bool buffer_migrate_lock_buffers(struct buffer_head *head,
699 enum migrate_mode mode)
700{
701 struct buffer_head *bh = head;
702 struct buffer_head *failed_bh;
703
704 do {
705 if (!trylock_buffer(bh)) {
706 if (mode == MIGRATE_ASYNC)
707 goto unlock;
708 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
709 goto unlock;
710 lock_buffer(bh);
711 }
712
713 bh = bh->b_this_page;
714 } while (bh != head);
715
716 return true;
717
718unlock:
719 /* We failed to lock the buffer and cannot stall. */
720 failed_bh = bh;
721 bh = head;
722 while (bh != failed_bh) {
723 unlock_buffer(bh);
724 bh = bh->b_this_page;
725 }
726
727 return false;
728}
729
730static int __buffer_migrate_folio(struct address_space *mapping,
731 struct folio *dst, struct folio *src, enum migrate_mode mode,
732 bool check_refs)
733{
734 struct buffer_head *bh, *head;
735 int rc;
736 int expected_count;
737
738 head = folio_buffers(src);
739 if (!head)
740 return migrate_folio(mapping, dst, src, mode);
741
742 /* Check whether page does not have extra refs before we do more work */
743 expected_count = folio_expected_refs(mapping, src);
744 if (folio_ref_count(src) != expected_count)
745 return -EAGAIN;
746
747 if (!buffer_migrate_lock_buffers(head, mode))
748 return -EAGAIN;
749
750 if (check_refs) {
751 bool busy;
752 bool invalidated = false;
753
754recheck_buffers:
755 busy = false;
756 spin_lock(&mapping->i_private_lock);
757 bh = head;
758 do {
759 if (atomic_read(&bh->b_count)) {
760 busy = true;
761 break;
762 }
763 bh = bh->b_this_page;
764 } while (bh != head);
765 if (busy) {
766 if (invalidated) {
767 rc = -EAGAIN;
768 goto unlock_buffers;
769 }
770 spin_unlock(&mapping->i_private_lock);
771 invalidate_bh_lrus();
772 invalidated = true;
773 goto recheck_buffers;
774 }
775 }
776
777 rc = folio_migrate_mapping(mapping, dst, src, 0);
778 if (rc != MIGRATEPAGE_SUCCESS)
779 goto unlock_buffers;
780
781 folio_attach_private(dst, folio_detach_private(src));
782
783 bh = head;
784 do {
785 folio_set_bh(bh, dst, bh_offset(bh));
786 bh = bh->b_this_page;
787 } while (bh != head);
788
789 if (mode != MIGRATE_SYNC_NO_COPY)
790 folio_migrate_copy(dst, src);
791 else
792 folio_migrate_flags(dst, src);
793
794 rc = MIGRATEPAGE_SUCCESS;
795unlock_buffers:
796 if (check_refs)
797 spin_unlock(&mapping->i_private_lock);
798 bh = head;
799 do {
800 unlock_buffer(bh);
801 bh = bh->b_this_page;
802 } while (bh != head);
803
804 return rc;
805}
806
807/**
808 * buffer_migrate_folio() - Migration function for folios with buffers.
809 * @mapping: The address space containing @src.
810 * @dst: The folio to migrate to.
811 * @src: The folio to migrate from.
812 * @mode: How to migrate the folio.
813 *
814 * This function can only be used if the underlying filesystem guarantees
815 * that no other references to @src exist. For example attached buffer
816 * heads are accessed only under the folio lock. If your filesystem cannot
817 * provide this guarantee, buffer_migrate_folio_norefs() may be more
818 * appropriate.
819 *
820 * Return: 0 on success or a negative errno on failure.
821 */
822int buffer_migrate_folio(struct address_space *mapping,
823 struct folio *dst, struct folio *src, enum migrate_mode mode)
824{
825 return __buffer_migrate_folio(mapping, dst, src, mode, false);
826}
827EXPORT_SYMBOL(buffer_migrate_folio);
828
829/**
830 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
831 * @mapping: The address space containing @src.
832 * @dst: The folio to migrate to.
833 * @src: The folio to migrate from.
834 * @mode: How to migrate the folio.
835 *
836 * Like buffer_migrate_folio() except that this variant is more careful
837 * and checks that there are also no buffer head references. This function
838 * is the right one for mappings where buffer heads are directly looked
839 * up and referenced (such as block device mappings).
840 *
841 * Return: 0 on success or a negative errno on failure.
842 */
843int buffer_migrate_folio_norefs(struct address_space *mapping,
844 struct folio *dst, struct folio *src, enum migrate_mode mode)
845{
846 return __buffer_migrate_folio(mapping, dst, src, mode, true);
847}
848EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
849#endif /* CONFIG_BUFFER_HEAD */
850
851int filemap_migrate_folio(struct address_space *mapping,
852 struct folio *dst, struct folio *src, enum migrate_mode mode)
853{
854 int ret;
855
856 ret = folio_migrate_mapping(mapping, dst, src, 0);
857 if (ret != MIGRATEPAGE_SUCCESS)
858 return ret;
859
860 if (folio_get_private(src))
861 folio_attach_private(dst, folio_detach_private(src));
862
863 if (mode != MIGRATE_SYNC_NO_COPY)
864 folio_migrate_copy(dst, src);
865 else
866 folio_migrate_flags(dst, src);
867 return MIGRATEPAGE_SUCCESS;
868}
869EXPORT_SYMBOL_GPL(filemap_migrate_folio);
870
871/*
872 * Writeback a folio to clean the dirty state
873 */
874static int writeout(struct address_space *mapping, struct folio *folio)
875{
876 struct writeback_control wbc = {
877 .sync_mode = WB_SYNC_NONE,
878 .nr_to_write = 1,
879 .range_start = 0,
880 .range_end = LLONG_MAX,
881 .for_reclaim = 1
882 };
883 int rc;
884
885 if (!mapping->a_ops->writepage)
886 /* No write method for the address space */
887 return -EINVAL;
888
889 if (!folio_clear_dirty_for_io(folio))
890 /* Someone else already triggered a write */
891 return -EAGAIN;
892
893 /*
894 * A dirty folio may imply that the underlying filesystem has
895 * the folio on some queue. So the folio must be clean for
896 * migration. Writeout may mean we lose the lock and the
897 * folio state is no longer what we checked for earlier.
898 * At this point we know that the migration attempt cannot
899 * be successful.
900 */
901 remove_migration_ptes(folio, folio, false);
902
903 rc = mapping->a_ops->writepage(&folio->page, &wbc);
904
905 if (rc != AOP_WRITEPAGE_ACTIVATE)
906 /* unlocked. Relock */
907 folio_lock(folio);
908
909 return (rc < 0) ? -EIO : -EAGAIN;
910}
911
912/*
913 * Default handling if a filesystem does not provide a migration function.
914 */
915static int fallback_migrate_folio(struct address_space *mapping,
916 struct folio *dst, struct folio *src, enum migrate_mode mode)
917{
918 if (folio_test_dirty(src)) {
919 /* Only writeback folios in full synchronous migration */
920 switch (mode) {
921 case MIGRATE_SYNC:
922 case MIGRATE_SYNC_NO_COPY:
923 break;
924 default:
925 return -EBUSY;
926 }
927 return writeout(mapping, src);
928 }
929
930 /*
931 * Buffers may be managed in a filesystem specific way.
932 * We must have no buffers or drop them.
933 */
934 if (!filemap_release_folio(src, GFP_KERNEL))
935 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
936
937 return migrate_folio(mapping, dst, src, mode);
938}
939
940/*
941 * Move a page to a newly allocated page
942 * The page is locked and all ptes have been successfully removed.
943 *
944 * The new page will have replaced the old page if this function
945 * is successful.
946 *
947 * Return value:
948 * < 0 - error code
949 * MIGRATEPAGE_SUCCESS - success
950 */
951static int move_to_new_folio(struct folio *dst, struct folio *src,
952 enum migrate_mode mode)
953{
954 int rc = -EAGAIN;
955 bool is_lru = !__folio_test_movable(src);
956
957 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
958 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
959
960 if (likely(is_lru)) {
961 struct address_space *mapping = folio_mapping(src);
962
963 if (!mapping)
964 rc = migrate_folio(mapping, dst, src, mode);
965 else if (mapping_unmovable(mapping))
966 rc = -EOPNOTSUPP;
967 else if (mapping->a_ops->migrate_folio)
968 /*
969 * Most folios have a mapping and most filesystems
970 * provide a migrate_folio callback. Anonymous folios
971 * are part of swap space which also has its own
972 * migrate_folio callback. This is the most common path
973 * for page migration.
974 */
975 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
976 mode);
977 else
978 rc = fallback_migrate_folio(mapping, dst, src, mode);
979 } else {
980 const struct movable_operations *mops;
981
982 /*
983 * In case of non-lru page, it could be released after
984 * isolation step. In that case, we shouldn't try migration.
985 */
986 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
987 if (!folio_test_movable(src)) {
988 rc = MIGRATEPAGE_SUCCESS;
989 folio_clear_isolated(src);
990 goto out;
991 }
992
993 mops = folio_movable_ops(src);
994 rc = mops->migrate_page(&dst->page, &src->page, mode);
995 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
996 !folio_test_isolated(src));
997 }
998
999 /*
1000 * When successful, old pagecache src->mapping must be cleared before
1001 * src is freed; but stats require that PageAnon be left as PageAnon.
1002 */
1003 if (rc == MIGRATEPAGE_SUCCESS) {
1004 if (__folio_test_movable(src)) {
1005 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1006
1007 /*
1008 * We clear PG_movable under page_lock so any compactor
1009 * cannot try to migrate this page.
1010 */
1011 folio_clear_isolated(src);
1012 }
1013
1014 /*
1015 * Anonymous and movable src->mapping will be cleared by
1016 * free_pages_prepare so don't reset it here for keeping
1017 * the type to work PageAnon, for example.
1018 */
1019 if (!folio_mapping_flags(src))
1020 src->mapping = NULL;
1021
1022 if (likely(!folio_is_zone_device(dst)))
1023 flush_dcache_folio(dst);
1024 }
1025out:
1026 return rc;
1027}
1028
1029/*
1030 * To record some information during migration, we use unused private
1031 * field of struct folio of the newly allocated destination folio.
1032 * This is safe because nobody is using it except us.
1033 */
1034enum {
1035 PAGE_WAS_MAPPED = BIT(0),
1036 PAGE_WAS_MLOCKED = BIT(1),
1037 PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1038};
1039
1040static void __migrate_folio_record(struct folio *dst,
1041 int old_page_state,
1042 struct anon_vma *anon_vma)
1043{
1044 dst->private = (void *)anon_vma + old_page_state;
1045}
1046
1047static void __migrate_folio_extract(struct folio *dst,
1048 int *old_page_state,
1049 struct anon_vma **anon_vmap)
1050{
1051 unsigned long private = (unsigned long)dst->private;
1052
1053 *anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1054 *old_page_state = private & PAGE_OLD_STATES;
1055 dst->private = NULL;
1056}
1057
1058/* Restore the source folio to the original state upon failure */
1059static void migrate_folio_undo_src(struct folio *src,
1060 int page_was_mapped,
1061 struct anon_vma *anon_vma,
1062 bool locked,
1063 struct list_head *ret)
1064{
1065 if (page_was_mapped)
1066 remove_migration_ptes(src, src, false);
1067 /* Drop an anon_vma reference if we took one */
1068 if (anon_vma)
1069 put_anon_vma(anon_vma);
1070 if (locked)
1071 folio_unlock(src);
1072 if (ret)
1073 list_move_tail(&src->lru, ret);
1074}
1075
1076/* Restore the destination folio to the original state upon failure */
1077static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1078 free_folio_t put_new_folio, unsigned long private)
1079{
1080 if (locked)
1081 folio_unlock(dst);
1082 if (put_new_folio)
1083 put_new_folio(dst, private);
1084 else
1085 folio_put(dst);
1086}
1087
1088/* Cleanup src folio upon migration success */
1089static void migrate_folio_done(struct folio *src,
1090 enum migrate_reason reason)
1091{
1092 /*
1093 * Compaction can migrate also non-LRU pages which are
1094 * not accounted to NR_ISOLATED_*. They can be recognized
1095 * as __folio_test_movable
1096 */
1097 if (likely(!__folio_test_movable(src)))
1098 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1099 folio_is_file_lru(src), -folio_nr_pages(src));
1100
1101 if (reason != MR_MEMORY_FAILURE)
1102 /* We release the page in page_handle_poison. */
1103 folio_put(src);
1104}
1105
1106/* Obtain the lock on page, remove all ptes. */
1107static int migrate_folio_unmap(new_folio_t get_new_folio,
1108 free_folio_t put_new_folio, unsigned long private,
1109 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1110 enum migrate_reason reason, struct list_head *ret)
1111{
1112 struct folio *dst;
1113 int rc = -EAGAIN;
1114 int old_page_state = 0;
1115 struct anon_vma *anon_vma = NULL;
1116 bool is_lru = !__folio_test_movable(src);
1117 bool locked = false;
1118 bool dst_locked = false;
1119
1120 if (folio_ref_count(src) == 1) {
1121 /* Folio was freed from under us. So we are done. */
1122 folio_clear_active(src);
1123 folio_clear_unevictable(src);
1124 /* free_pages_prepare() will clear PG_isolated. */
1125 list_del(&src->lru);
1126 migrate_folio_done(src, reason);
1127 return MIGRATEPAGE_SUCCESS;
1128 }
1129
1130 dst = get_new_folio(src, private);
1131 if (!dst)
1132 return -ENOMEM;
1133 *dstp = dst;
1134
1135 dst->private = NULL;
1136
1137 if (!folio_trylock(src)) {
1138 if (mode == MIGRATE_ASYNC)
1139 goto out;
1140
1141 /*
1142 * It's not safe for direct compaction to call lock_page.
1143 * For example, during page readahead pages are added locked
1144 * to the LRU. Later, when the IO completes the pages are
1145 * marked uptodate and unlocked. However, the queueing
1146 * could be merging multiple pages for one bio (e.g.
1147 * mpage_readahead). If an allocation happens for the
1148 * second or third page, the process can end up locking
1149 * the same page twice and deadlocking. Rather than
1150 * trying to be clever about what pages can be locked,
1151 * avoid the use of lock_page for direct compaction
1152 * altogether.
1153 */
1154 if (current->flags & PF_MEMALLOC)
1155 goto out;
1156
1157 /*
1158 * In "light" mode, we can wait for transient locks (eg
1159 * inserting a page into the page table), but it's not
1160 * worth waiting for I/O.
1161 */
1162 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1163 goto out;
1164
1165 folio_lock(src);
1166 }
1167 locked = true;
1168 if (folio_test_mlocked(src))
1169 old_page_state |= PAGE_WAS_MLOCKED;
1170
1171 if (folio_test_writeback(src)) {
1172 /*
1173 * Only in the case of a full synchronous migration is it
1174 * necessary to wait for PageWriteback. In the async case,
1175 * the retry loop is too short and in the sync-light case,
1176 * the overhead of stalling is too much
1177 */
1178 switch (mode) {
1179 case MIGRATE_SYNC:
1180 case MIGRATE_SYNC_NO_COPY:
1181 break;
1182 default:
1183 rc = -EBUSY;
1184 goto out;
1185 }
1186 folio_wait_writeback(src);
1187 }
1188
1189 /*
1190 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1191 * we cannot notice that anon_vma is freed while we migrate a page.
1192 * This get_anon_vma() delays freeing anon_vma pointer until the end
1193 * of migration. File cache pages are no problem because of page_lock()
1194 * File Caches may use write_page() or lock_page() in migration, then,
1195 * just care Anon page here.
1196 *
1197 * Only folio_get_anon_vma() understands the subtleties of
1198 * getting a hold on an anon_vma from outside one of its mms.
1199 * But if we cannot get anon_vma, then we won't need it anyway,
1200 * because that implies that the anon page is no longer mapped
1201 * (and cannot be remapped so long as we hold the page lock).
1202 */
1203 if (folio_test_anon(src) && !folio_test_ksm(src))
1204 anon_vma = folio_get_anon_vma(src);
1205
1206 /*
1207 * Block others from accessing the new page when we get around to
1208 * establishing additional references. We are usually the only one
1209 * holding a reference to dst at this point. We used to have a BUG
1210 * here if folio_trylock(dst) fails, but would like to allow for
1211 * cases where there might be a race with the previous use of dst.
1212 * This is much like races on refcount of oldpage: just don't BUG().
1213 */
1214 if (unlikely(!folio_trylock(dst)))
1215 goto out;
1216 dst_locked = true;
1217
1218 if (unlikely(!is_lru)) {
1219 __migrate_folio_record(dst, old_page_state, anon_vma);
1220 return MIGRATEPAGE_UNMAP;
1221 }
1222
1223 /*
1224 * Corner case handling:
1225 * 1. When a new swap-cache page is read into, it is added to the LRU
1226 * and treated as swapcache but it has no rmap yet.
1227 * Calling try_to_unmap() against a src->mapping==NULL page will
1228 * trigger a BUG. So handle it here.
1229 * 2. An orphaned page (see truncate_cleanup_page) might have
1230 * fs-private metadata. The page can be picked up due to memory
1231 * offlining. Everywhere else except page reclaim, the page is
1232 * invisible to the vm, so the page can not be migrated. So try to
1233 * free the metadata, so the page can be freed.
1234 */
1235 if (!src->mapping) {
1236 if (folio_test_private(src)) {
1237 try_to_free_buffers(src);
1238 goto out;
1239 }
1240 } else if (folio_mapped(src)) {
1241 /* Establish migration ptes */
1242 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1243 !folio_test_ksm(src) && !anon_vma, src);
1244 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1245 old_page_state |= PAGE_WAS_MAPPED;
1246 }
1247
1248 if (!folio_mapped(src)) {
1249 __migrate_folio_record(dst, old_page_state, anon_vma);
1250 return MIGRATEPAGE_UNMAP;
1251 }
1252
1253out:
1254 /*
1255 * A folio that has not been unmapped will be restored to
1256 * right list unless we want to retry.
1257 */
1258 if (rc == -EAGAIN)
1259 ret = NULL;
1260
1261 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1262 anon_vma, locked, ret);
1263 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1264
1265 return rc;
1266}
1267
1268/* Migrate the folio to the newly allocated folio in dst. */
1269static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1270 struct folio *src, struct folio *dst,
1271 enum migrate_mode mode, enum migrate_reason reason,
1272 struct list_head *ret)
1273{
1274 int rc;
1275 int old_page_state = 0;
1276 struct anon_vma *anon_vma = NULL;
1277 bool is_lru = !__folio_test_movable(src);
1278 struct list_head *prev;
1279
1280 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1281 prev = dst->lru.prev;
1282 list_del(&dst->lru);
1283
1284 rc = move_to_new_folio(dst, src, mode);
1285 if (rc)
1286 goto out;
1287
1288 if (unlikely(!is_lru))
1289 goto out_unlock_both;
1290
1291 /*
1292 * When successful, push dst to LRU immediately: so that if it
1293 * turns out to be an mlocked page, remove_migration_ptes() will
1294 * automatically build up the correct dst->mlock_count for it.
1295 *
1296 * We would like to do something similar for the old page, when
1297 * unsuccessful, and other cases when a page has been temporarily
1298 * isolated from the unevictable LRU: but this case is the easiest.
1299 */
1300 folio_add_lru(dst);
1301 if (old_page_state & PAGE_WAS_MLOCKED)
1302 lru_add_drain();
1303
1304 if (old_page_state & PAGE_WAS_MAPPED)
1305 remove_migration_ptes(src, dst, false);
1306
1307out_unlock_both:
1308 folio_unlock(dst);
1309 set_page_owner_migrate_reason(&dst->page, reason);
1310 /*
1311 * If migration is successful, decrease refcount of dst,
1312 * which will not free the page because new page owner increased
1313 * refcounter.
1314 */
1315 folio_put(dst);
1316
1317 /*
1318 * A folio that has been migrated has all references removed
1319 * and will be freed.
1320 */
1321 list_del(&src->lru);
1322 /* Drop an anon_vma reference if we took one */
1323 if (anon_vma)
1324 put_anon_vma(anon_vma);
1325 folio_unlock(src);
1326 migrate_folio_done(src, reason);
1327
1328 return rc;
1329out:
1330 /*
1331 * A folio that has not been migrated will be restored to
1332 * right list unless we want to retry.
1333 */
1334 if (rc == -EAGAIN) {
1335 list_add(&dst->lru, prev);
1336 __migrate_folio_record(dst, old_page_state, anon_vma);
1337 return rc;
1338 }
1339
1340 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1341 anon_vma, true, ret);
1342 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1343
1344 return rc;
1345}
1346
1347/*
1348 * Counterpart of unmap_and_move_page() for hugepage migration.
1349 *
1350 * This function doesn't wait the completion of hugepage I/O
1351 * because there is no race between I/O and migration for hugepage.
1352 * Note that currently hugepage I/O occurs only in direct I/O
1353 * where no lock is held and PG_writeback is irrelevant,
1354 * and writeback status of all subpages are counted in the reference
1355 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1356 * under direct I/O, the reference of the head page is 512 and a bit more.)
1357 * This means that when we try to migrate hugepage whose subpages are
1358 * doing direct I/O, some references remain after try_to_unmap() and
1359 * hugepage migration fails without data corruption.
1360 *
1361 * There is also no race when direct I/O is issued on the page under migration,
1362 * because then pte is replaced with migration swap entry and direct I/O code
1363 * will wait in the page fault for migration to complete.
1364 */
1365static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1366 free_folio_t put_new_folio, unsigned long private,
1367 struct folio *src, int force, enum migrate_mode mode,
1368 int reason, struct list_head *ret)
1369{
1370 struct folio *dst;
1371 int rc = -EAGAIN;
1372 int page_was_mapped = 0;
1373 struct anon_vma *anon_vma = NULL;
1374 struct address_space *mapping = NULL;
1375
1376 if (folio_ref_count(src) == 1) {
1377 /* page was freed from under us. So we are done. */
1378 folio_putback_active_hugetlb(src);
1379 return MIGRATEPAGE_SUCCESS;
1380 }
1381
1382 dst = get_new_folio(src, private);
1383 if (!dst)
1384 return -ENOMEM;
1385
1386 if (!folio_trylock(src)) {
1387 if (!force)
1388 goto out;
1389 switch (mode) {
1390 case MIGRATE_SYNC:
1391 case MIGRATE_SYNC_NO_COPY:
1392 break;
1393 default:
1394 goto out;
1395 }
1396 folio_lock(src);
1397 }
1398
1399 /*
1400 * Check for pages which are in the process of being freed. Without
1401 * folio_mapping() set, hugetlbfs specific move page routine will not
1402 * be called and we could leak usage counts for subpools.
1403 */
1404 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1405 rc = -EBUSY;
1406 goto out_unlock;
1407 }
1408
1409 if (folio_test_anon(src))
1410 anon_vma = folio_get_anon_vma(src);
1411
1412 if (unlikely(!folio_trylock(dst)))
1413 goto put_anon;
1414
1415 if (folio_mapped(src)) {
1416 enum ttu_flags ttu = 0;
1417
1418 if (!folio_test_anon(src)) {
1419 /*
1420 * In shared mappings, try_to_unmap could potentially
1421 * call huge_pmd_unshare. Because of this, take
1422 * semaphore in write mode here and set TTU_RMAP_LOCKED
1423 * to let lower levels know we have taken the lock.
1424 */
1425 mapping = hugetlb_page_mapping_lock_write(&src->page);
1426 if (unlikely(!mapping))
1427 goto unlock_put_anon;
1428
1429 ttu = TTU_RMAP_LOCKED;
1430 }
1431
1432 try_to_migrate(src, ttu);
1433 page_was_mapped = 1;
1434
1435 if (ttu & TTU_RMAP_LOCKED)
1436 i_mmap_unlock_write(mapping);
1437 }
1438
1439 if (!folio_mapped(src))
1440 rc = move_to_new_folio(dst, src, mode);
1441
1442 if (page_was_mapped)
1443 remove_migration_ptes(src,
1444 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1445
1446unlock_put_anon:
1447 folio_unlock(dst);
1448
1449put_anon:
1450 if (anon_vma)
1451 put_anon_vma(anon_vma);
1452
1453 if (rc == MIGRATEPAGE_SUCCESS) {
1454 move_hugetlb_state(src, dst, reason);
1455 put_new_folio = NULL;
1456 }
1457
1458out_unlock:
1459 folio_unlock(src);
1460out:
1461 if (rc == MIGRATEPAGE_SUCCESS)
1462 folio_putback_active_hugetlb(src);
1463 else if (rc != -EAGAIN)
1464 list_move_tail(&src->lru, ret);
1465
1466 /*
1467 * If migration was not successful and there's a freeing callback, use
1468 * it. Otherwise, put_page() will drop the reference grabbed during
1469 * isolation.
1470 */
1471 if (put_new_folio)
1472 put_new_folio(dst, private);
1473 else
1474 folio_putback_active_hugetlb(dst);
1475
1476 return rc;
1477}
1478
1479static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1480{
1481 int rc;
1482
1483 folio_lock(folio);
1484 rc = split_folio_to_list(folio, split_folios);
1485 folio_unlock(folio);
1486 if (!rc)
1487 list_move_tail(&folio->lru, split_folios);
1488
1489 return rc;
1490}
1491
1492#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1493#define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1494#else
1495#define NR_MAX_BATCHED_MIGRATION 512
1496#endif
1497#define NR_MAX_MIGRATE_PAGES_RETRY 10
1498#define NR_MAX_MIGRATE_ASYNC_RETRY 3
1499#define NR_MAX_MIGRATE_SYNC_RETRY \
1500 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1501
1502struct migrate_pages_stats {
1503 int nr_succeeded; /* Normal and large folios migrated successfully, in
1504 units of base pages */
1505 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1506 units of base pages. Untried folios aren't counted */
1507 int nr_thp_succeeded; /* THP migrated successfully */
1508 int nr_thp_failed; /* THP failed to be migrated */
1509 int nr_thp_split; /* THP split before migrating */
1510 int nr_split; /* Large folio (include THP) split before migrating */
1511};
1512
1513/*
1514 * Returns the number of hugetlb folios that were not migrated, or an error code
1515 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1516 * any more because the list has become empty or no retryable hugetlb folios
1517 * exist any more. It is caller's responsibility to call putback_movable_pages()
1518 * only if ret != 0.
1519 */
1520static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1521 free_folio_t put_new_folio, unsigned long private,
1522 enum migrate_mode mode, int reason,
1523 struct migrate_pages_stats *stats,
1524 struct list_head *ret_folios)
1525{
1526 int retry = 1;
1527 int nr_failed = 0;
1528 int nr_retry_pages = 0;
1529 int pass = 0;
1530 struct folio *folio, *folio2;
1531 int rc, nr_pages;
1532
1533 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1534 retry = 0;
1535 nr_retry_pages = 0;
1536
1537 list_for_each_entry_safe(folio, folio2, from, lru) {
1538 if (!folio_test_hugetlb(folio))
1539 continue;
1540
1541 nr_pages = folio_nr_pages(folio);
1542
1543 cond_resched();
1544
1545 /*
1546 * Migratability of hugepages depends on architectures and
1547 * their size. This check is necessary because some callers
1548 * of hugepage migration like soft offline and memory
1549 * hotremove don't walk through page tables or check whether
1550 * the hugepage is pmd-based or not before kicking migration.
1551 */
1552 if (!hugepage_migration_supported(folio_hstate(folio))) {
1553 nr_failed++;
1554 stats->nr_failed_pages += nr_pages;
1555 list_move_tail(&folio->lru, ret_folios);
1556 continue;
1557 }
1558
1559 rc = unmap_and_move_huge_page(get_new_folio,
1560 put_new_folio, private,
1561 folio, pass > 2, mode,
1562 reason, ret_folios);
1563 /*
1564 * The rules are:
1565 * Success: hugetlb folio will be put back
1566 * -EAGAIN: stay on the from list
1567 * -ENOMEM: stay on the from list
1568 * Other errno: put on ret_folios list
1569 */
1570 switch(rc) {
1571 case -ENOMEM:
1572 /*
1573 * When memory is low, don't bother to try to migrate
1574 * other folios, just exit.
1575 */
1576 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1577 return -ENOMEM;
1578 case -EAGAIN:
1579 retry++;
1580 nr_retry_pages += nr_pages;
1581 break;
1582 case MIGRATEPAGE_SUCCESS:
1583 stats->nr_succeeded += nr_pages;
1584 break;
1585 default:
1586 /*
1587 * Permanent failure (-EBUSY, etc.):
1588 * unlike -EAGAIN case, the failed folio is
1589 * removed from migration folio list and not
1590 * retried in the next outer loop.
1591 */
1592 nr_failed++;
1593 stats->nr_failed_pages += nr_pages;
1594 break;
1595 }
1596 }
1597 }
1598 /*
1599 * nr_failed is number of hugetlb folios failed to be migrated. After
1600 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1601 * folios as failed.
1602 */
1603 nr_failed += retry;
1604 stats->nr_failed_pages += nr_retry_pages;
1605
1606 return nr_failed;
1607}
1608
1609/*
1610 * migrate_pages_batch() first unmaps folios in the from list as many as
1611 * possible, then move the unmapped folios.
1612 *
1613 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1614 * lock or bit when we have locked more than one folio. Which may cause
1615 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1616 * length of the from list must be <= 1.
1617 */
1618static int migrate_pages_batch(struct list_head *from,
1619 new_folio_t get_new_folio, free_folio_t put_new_folio,
1620 unsigned long private, enum migrate_mode mode, int reason,
1621 struct list_head *ret_folios, struct list_head *split_folios,
1622 struct migrate_pages_stats *stats, int nr_pass)
1623{
1624 int retry = 1;
1625 int thp_retry = 1;
1626 int nr_failed = 0;
1627 int nr_retry_pages = 0;
1628 int pass = 0;
1629 bool is_thp = false;
1630 bool is_large = false;
1631 struct folio *folio, *folio2, *dst = NULL, *dst2;
1632 int rc, rc_saved = 0, nr_pages;
1633 LIST_HEAD(unmap_folios);
1634 LIST_HEAD(dst_folios);
1635 bool nosplit = (reason == MR_NUMA_MISPLACED);
1636
1637 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1638 !list_empty(from) && !list_is_singular(from));
1639
1640 for (pass = 0; pass < nr_pass && retry; pass++) {
1641 retry = 0;
1642 thp_retry = 0;
1643 nr_retry_pages = 0;
1644
1645 list_for_each_entry_safe(folio, folio2, from, lru) {
1646 is_large = folio_test_large(folio);
1647 is_thp = is_large && folio_test_pmd_mappable(folio);
1648 nr_pages = folio_nr_pages(folio);
1649
1650 cond_resched();
1651
1652 /*
1653 * Large folio migration might be unsupported or
1654 * the allocation might be failed so we should retry
1655 * on the same folio with the large folio split
1656 * to normal folios.
1657 *
1658 * Split folios are put in split_folios, and
1659 * we will migrate them after the rest of the
1660 * list is processed.
1661 */
1662 if (!thp_migration_supported() && is_thp) {
1663 nr_failed++;
1664 stats->nr_thp_failed++;
1665 if (!try_split_folio(folio, split_folios)) {
1666 stats->nr_thp_split++;
1667 stats->nr_split++;
1668 continue;
1669 }
1670 stats->nr_failed_pages += nr_pages;
1671 list_move_tail(&folio->lru, ret_folios);
1672 continue;
1673 }
1674
1675 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1676 private, folio, &dst, mode, reason,
1677 ret_folios);
1678 /*
1679 * The rules are:
1680 * Success: folio will be freed
1681 * Unmap: folio will be put on unmap_folios list,
1682 * dst folio put on dst_folios list
1683 * -EAGAIN: stay on the from list
1684 * -ENOMEM: stay on the from list
1685 * Other errno: put on ret_folios list
1686 */
1687 switch(rc) {
1688 case -ENOMEM:
1689 /*
1690 * When memory is low, don't bother to try to migrate
1691 * other folios, move unmapped folios, then exit.
1692 */
1693 nr_failed++;
1694 stats->nr_thp_failed += is_thp;
1695 /* Large folio NUMA faulting doesn't split to retry. */
1696 if (is_large && !nosplit) {
1697 int ret = try_split_folio(folio, split_folios);
1698
1699 if (!ret) {
1700 stats->nr_thp_split += is_thp;
1701 stats->nr_split++;
1702 break;
1703 } else if (reason == MR_LONGTERM_PIN &&
1704 ret == -EAGAIN) {
1705 /*
1706 * Try again to split large folio to
1707 * mitigate the failure of longterm pinning.
1708 */
1709 retry++;
1710 thp_retry += is_thp;
1711 nr_retry_pages += nr_pages;
1712 /* Undo duplicated failure counting. */
1713 nr_failed--;
1714 stats->nr_thp_failed -= is_thp;
1715 break;
1716 }
1717 }
1718
1719 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1720 /* nr_failed isn't updated for not used */
1721 stats->nr_thp_failed += thp_retry;
1722 rc_saved = rc;
1723 if (list_empty(&unmap_folios))
1724 goto out;
1725 else
1726 goto move;
1727 case -EAGAIN:
1728 retry++;
1729 thp_retry += is_thp;
1730 nr_retry_pages += nr_pages;
1731 break;
1732 case MIGRATEPAGE_SUCCESS:
1733 stats->nr_succeeded += nr_pages;
1734 stats->nr_thp_succeeded += is_thp;
1735 break;
1736 case MIGRATEPAGE_UNMAP:
1737 list_move_tail(&folio->lru, &unmap_folios);
1738 list_add_tail(&dst->lru, &dst_folios);
1739 break;
1740 default:
1741 /*
1742 * Permanent failure (-EBUSY, etc.):
1743 * unlike -EAGAIN case, the failed folio is
1744 * removed from migration folio list and not
1745 * retried in the next outer loop.
1746 */
1747 nr_failed++;
1748 stats->nr_thp_failed += is_thp;
1749 stats->nr_failed_pages += nr_pages;
1750 break;
1751 }
1752 }
1753 }
1754 nr_failed += retry;
1755 stats->nr_thp_failed += thp_retry;
1756 stats->nr_failed_pages += nr_retry_pages;
1757move:
1758 /* Flush TLBs for all unmapped folios */
1759 try_to_unmap_flush();
1760
1761 retry = 1;
1762 for (pass = 0; pass < nr_pass && retry; pass++) {
1763 retry = 0;
1764 thp_retry = 0;
1765 nr_retry_pages = 0;
1766
1767 dst = list_first_entry(&dst_folios, struct folio, lru);
1768 dst2 = list_next_entry(dst, lru);
1769 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1770 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1771 nr_pages = folio_nr_pages(folio);
1772
1773 cond_resched();
1774
1775 rc = migrate_folio_move(put_new_folio, private,
1776 folio, dst, mode,
1777 reason, ret_folios);
1778 /*
1779 * The rules are:
1780 * Success: folio will be freed
1781 * -EAGAIN: stay on the unmap_folios list
1782 * Other errno: put on ret_folios list
1783 */
1784 switch(rc) {
1785 case -EAGAIN:
1786 retry++;
1787 thp_retry += is_thp;
1788 nr_retry_pages += nr_pages;
1789 break;
1790 case MIGRATEPAGE_SUCCESS:
1791 stats->nr_succeeded += nr_pages;
1792 stats->nr_thp_succeeded += is_thp;
1793 break;
1794 default:
1795 nr_failed++;
1796 stats->nr_thp_failed += is_thp;
1797 stats->nr_failed_pages += nr_pages;
1798 break;
1799 }
1800 dst = dst2;
1801 dst2 = list_next_entry(dst, lru);
1802 }
1803 }
1804 nr_failed += retry;
1805 stats->nr_thp_failed += thp_retry;
1806 stats->nr_failed_pages += nr_retry_pages;
1807
1808 rc = rc_saved ? : nr_failed;
1809out:
1810 /* Cleanup remaining folios */
1811 dst = list_first_entry(&dst_folios, struct folio, lru);
1812 dst2 = list_next_entry(dst, lru);
1813 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1814 int old_page_state = 0;
1815 struct anon_vma *anon_vma = NULL;
1816
1817 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1818 migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1819 anon_vma, true, ret_folios);
1820 list_del(&dst->lru);
1821 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1822 dst = dst2;
1823 dst2 = list_next_entry(dst, lru);
1824 }
1825
1826 return rc;
1827}
1828
1829static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1830 free_folio_t put_new_folio, unsigned long private,
1831 enum migrate_mode mode, int reason,
1832 struct list_head *ret_folios, struct list_head *split_folios,
1833 struct migrate_pages_stats *stats)
1834{
1835 int rc, nr_failed = 0;
1836 LIST_HEAD(folios);
1837 struct migrate_pages_stats astats;
1838
1839 memset(&astats, 0, sizeof(astats));
1840 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1841 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1842 reason, &folios, split_folios, &astats,
1843 NR_MAX_MIGRATE_ASYNC_RETRY);
1844 stats->nr_succeeded += astats.nr_succeeded;
1845 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1846 stats->nr_thp_split += astats.nr_thp_split;
1847 stats->nr_split += astats.nr_split;
1848 if (rc < 0) {
1849 stats->nr_failed_pages += astats.nr_failed_pages;
1850 stats->nr_thp_failed += astats.nr_thp_failed;
1851 list_splice_tail(&folios, ret_folios);
1852 return rc;
1853 }
1854 stats->nr_thp_failed += astats.nr_thp_split;
1855 /*
1856 * Do not count rc, as pages will be retried below.
1857 * Count nr_split only, since it includes nr_thp_split.
1858 */
1859 nr_failed += astats.nr_split;
1860 /*
1861 * Fall back to migrate all failed folios one by one synchronously. All
1862 * failed folios except split THPs will be retried, so their failure
1863 * isn't counted
1864 */
1865 list_splice_tail_init(&folios, from);
1866 while (!list_empty(from)) {
1867 list_move(from->next, &folios);
1868 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1869 private, mode, reason, ret_folios,
1870 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1871 list_splice_tail_init(&folios, ret_folios);
1872 if (rc < 0)
1873 return rc;
1874 nr_failed += rc;
1875 }
1876
1877 return nr_failed;
1878}
1879
1880/*
1881 * migrate_pages - migrate the folios specified in a list, to the free folios
1882 * supplied as the target for the page migration
1883 *
1884 * @from: The list of folios to be migrated.
1885 * @get_new_folio: The function used to allocate free folios to be used
1886 * as the target of the folio migration.
1887 * @put_new_folio: The function used to free target folios if migration
1888 * fails, or NULL if no special handling is necessary.
1889 * @private: Private data to be passed on to get_new_folio()
1890 * @mode: The migration mode that specifies the constraints for
1891 * folio migration, if any.
1892 * @reason: The reason for folio migration.
1893 * @ret_succeeded: Set to the number of folios migrated successfully if
1894 * the caller passes a non-NULL pointer.
1895 *
1896 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1897 * are movable any more because the list has become empty or no retryable folios
1898 * exist any more. It is caller's responsibility to call putback_movable_pages()
1899 * only if ret != 0.
1900 *
1901 * Returns the number of {normal folio, large folio, hugetlb} that were not
1902 * migrated, or an error code. The number of large folio splits will be
1903 * considered as the number of non-migrated large folio, no matter how many
1904 * split folios of the large folio are migrated successfully.
1905 */
1906int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1907 free_folio_t put_new_folio, unsigned long private,
1908 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1909{
1910 int rc, rc_gather;
1911 int nr_pages;
1912 struct folio *folio, *folio2;
1913 LIST_HEAD(folios);
1914 LIST_HEAD(ret_folios);
1915 LIST_HEAD(split_folios);
1916 struct migrate_pages_stats stats;
1917
1918 trace_mm_migrate_pages_start(mode, reason);
1919
1920 memset(&stats, 0, sizeof(stats));
1921
1922 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1923 mode, reason, &stats, &ret_folios);
1924 if (rc_gather < 0)
1925 goto out;
1926
1927again:
1928 nr_pages = 0;
1929 list_for_each_entry_safe(folio, folio2, from, lru) {
1930 /* Retried hugetlb folios will be kept in list */
1931 if (folio_test_hugetlb(folio)) {
1932 list_move_tail(&folio->lru, &ret_folios);
1933 continue;
1934 }
1935
1936 nr_pages += folio_nr_pages(folio);
1937 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1938 break;
1939 }
1940 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1941 list_cut_before(&folios, from, &folio2->lru);
1942 else
1943 list_splice_init(from, &folios);
1944 if (mode == MIGRATE_ASYNC)
1945 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1946 private, mode, reason, &ret_folios,
1947 &split_folios, &stats,
1948 NR_MAX_MIGRATE_PAGES_RETRY);
1949 else
1950 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1951 private, mode, reason, &ret_folios,
1952 &split_folios, &stats);
1953 list_splice_tail_init(&folios, &ret_folios);
1954 if (rc < 0) {
1955 rc_gather = rc;
1956 list_splice_tail(&split_folios, &ret_folios);
1957 goto out;
1958 }
1959 if (!list_empty(&split_folios)) {
1960 /*
1961 * Failure isn't counted since all split folios of a large folio
1962 * is counted as 1 failure already. And, we only try to migrate
1963 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1964 */
1965 migrate_pages_batch(&split_folios, get_new_folio,
1966 put_new_folio, private, MIGRATE_ASYNC, reason,
1967 &ret_folios, NULL, &stats, 1);
1968 list_splice_tail_init(&split_folios, &ret_folios);
1969 }
1970 rc_gather += rc;
1971 if (!list_empty(from))
1972 goto again;
1973out:
1974 /*
1975 * Put the permanent failure folio back to migration list, they
1976 * will be put back to the right list by the caller.
1977 */
1978 list_splice(&ret_folios, from);
1979
1980 /*
1981 * Return 0 in case all split folios of fail-to-migrate large folios
1982 * are migrated successfully.
1983 */
1984 if (list_empty(from))
1985 rc_gather = 0;
1986
1987 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1988 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1989 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1990 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1991 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1992 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1993 stats.nr_thp_succeeded, stats.nr_thp_failed,
1994 stats.nr_thp_split, stats.nr_split, mode,
1995 reason);
1996
1997 if (ret_succeeded)
1998 *ret_succeeded = stats.nr_succeeded;
1999
2000 return rc_gather;
2001}
2002
2003struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2004{
2005 struct migration_target_control *mtc;
2006 gfp_t gfp_mask;
2007 unsigned int order = 0;
2008 int nid;
2009 int zidx;
2010
2011 mtc = (struct migration_target_control *)private;
2012 gfp_mask = mtc->gfp_mask;
2013 nid = mtc->nid;
2014 if (nid == NUMA_NO_NODE)
2015 nid = folio_nid(src);
2016
2017 if (folio_test_hugetlb(src)) {
2018 struct hstate *h = folio_hstate(src);
2019
2020 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2021 return alloc_hugetlb_folio_nodemask(h, nid,
2022 mtc->nmask, gfp_mask);
2023 }
2024
2025 if (folio_test_large(src)) {
2026 /*
2027 * clear __GFP_RECLAIM to make the migration callback
2028 * consistent with regular THP allocations.
2029 */
2030 gfp_mask &= ~__GFP_RECLAIM;
2031 gfp_mask |= GFP_TRANSHUGE;
2032 order = folio_order(src);
2033 }
2034 zidx = zone_idx(folio_zone(src));
2035 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2036 gfp_mask |= __GFP_HIGHMEM;
2037
2038 return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2039}
2040
2041#ifdef CONFIG_NUMA
2042
2043static int store_status(int __user *status, int start, int value, int nr)
2044{
2045 while (nr-- > 0) {
2046 if (put_user(value, status + start))
2047 return -EFAULT;
2048 start++;
2049 }
2050
2051 return 0;
2052}
2053
2054static int do_move_pages_to_node(struct list_head *pagelist, int node)
2055{
2056 int err;
2057 struct migration_target_control mtc = {
2058 .nid = node,
2059 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2060 };
2061
2062 err = migrate_pages(pagelist, alloc_migration_target, NULL,
2063 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2064 if (err)
2065 putback_movable_pages(pagelist);
2066 return err;
2067}
2068
2069/*
2070 * Resolves the given address to a struct page, isolates it from the LRU and
2071 * puts it to the given pagelist.
2072 * Returns:
2073 * errno - if the page cannot be found/isolated
2074 * 0 - when it doesn't have to be migrated because it is already on the
2075 * target node
2076 * 1 - when it has been queued
2077 */
2078static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2079 int node, struct list_head *pagelist, bool migrate_all)
2080{
2081 struct vm_area_struct *vma;
2082 unsigned long addr;
2083 struct page *page;
2084 struct folio *folio;
2085 int err;
2086
2087 mmap_read_lock(mm);
2088 addr = (unsigned long)untagged_addr_remote(mm, p);
2089
2090 err = -EFAULT;
2091 vma = vma_lookup(mm, addr);
2092 if (!vma || !vma_migratable(vma))
2093 goto out;
2094
2095 /* FOLL_DUMP to ignore special (like zero) pages */
2096 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2097
2098 err = PTR_ERR(page);
2099 if (IS_ERR(page))
2100 goto out;
2101
2102 err = -ENOENT;
2103 if (!page)
2104 goto out;
2105
2106 folio = page_folio(page);
2107 if (folio_is_zone_device(folio))
2108 goto out_putfolio;
2109
2110 err = 0;
2111 if (folio_nid(folio) == node)
2112 goto out_putfolio;
2113
2114 err = -EACCES;
2115 if (page_mapcount(page) > 1 && !migrate_all)
2116 goto out_putfolio;
2117
2118 err = -EBUSY;
2119 if (folio_test_hugetlb(folio)) {
2120 if (isolate_hugetlb(folio, pagelist))
2121 err = 1;
2122 } else {
2123 if (!folio_isolate_lru(folio))
2124 goto out_putfolio;
2125
2126 err = 1;
2127 list_add_tail(&folio->lru, pagelist);
2128 node_stat_mod_folio(folio,
2129 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2130 folio_nr_pages(folio));
2131 }
2132out_putfolio:
2133 /*
2134 * Either remove the duplicate refcount from folio_isolate_lru()
2135 * or drop the folio ref if it was not isolated.
2136 */
2137 folio_put(folio);
2138out:
2139 mmap_read_unlock(mm);
2140 return err;
2141}
2142
2143static int move_pages_and_store_status(int node,
2144 struct list_head *pagelist, int __user *status,
2145 int start, int i, unsigned long nr_pages)
2146{
2147 int err;
2148
2149 if (list_empty(pagelist))
2150 return 0;
2151
2152 err = do_move_pages_to_node(pagelist, node);
2153 if (err) {
2154 /*
2155 * Positive err means the number of failed
2156 * pages to migrate. Since we are going to
2157 * abort and return the number of non-migrated
2158 * pages, so need to include the rest of the
2159 * nr_pages that have not been attempted as
2160 * well.
2161 */
2162 if (err > 0)
2163 err += nr_pages - i;
2164 return err;
2165 }
2166 return store_status(status, start, node, i - start);
2167}
2168
2169/*
2170 * Migrate an array of page address onto an array of nodes and fill
2171 * the corresponding array of status.
2172 */
2173static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2174 unsigned long nr_pages,
2175 const void __user * __user *pages,
2176 const int __user *nodes,
2177 int __user *status, int flags)
2178{
2179 compat_uptr_t __user *compat_pages = (void __user *)pages;
2180 int current_node = NUMA_NO_NODE;
2181 LIST_HEAD(pagelist);
2182 int start, i;
2183 int err = 0, err1;
2184
2185 lru_cache_disable();
2186
2187 for (i = start = 0; i < nr_pages; i++) {
2188 const void __user *p;
2189 int node;
2190
2191 err = -EFAULT;
2192 if (in_compat_syscall()) {
2193 compat_uptr_t cp;
2194
2195 if (get_user(cp, compat_pages + i))
2196 goto out_flush;
2197
2198 p = compat_ptr(cp);
2199 } else {
2200 if (get_user(p, pages + i))
2201 goto out_flush;
2202 }
2203 if (get_user(node, nodes + i))
2204 goto out_flush;
2205
2206 err = -ENODEV;
2207 if (node < 0 || node >= MAX_NUMNODES)
2208 goto out_flush;
2209 if (!node_state(node, N_MEMORY))
2210 goto out_flush;
2211
2212 err = -EACCES;
2213 if (!node_isset(node, task_nodes))
2214 goto out_flush;
2215
2216 if (current_node == NUMA_NO_NODE) {
2217 current_node = node;
2218 start = i;
2219 } else if (node != current_node) {
2220 err = move_pages_and_store_status(current_node,
2221 &pagelist, status, start, i, nr_pages);
2222 if (err)
2223 goto out;
2224 start = i;
2225 current_node = node;
2226 }
2227
2228 /*
2229 * Errors in the page lookup or isolation are not fatal and we simply
2230 * report them via status
2231 */
2232 err = add_page_for_migration(mm, p, current_node, &pagelist,
2233 flags & MPOL_MF_MOVE_ALL);
2234
2235 if (err > 0) {
2236 /* The page is successfully queued for migration */
2237 continue;
2238 }
2239
2240 /*
2241 * The move_pages() man page does not have an -EEXIST choice, so
2242 * use -EFAULT instead.
2243 */
2244 if (err == -EEXIST)
2245 err = -EFAULT;
2246
2247 /*
2248 * If the page is already on the target node (!err), store the
2249 * node, otherwise, store the err.
2250 */
2251 err = store_status(status, i, err ? : current_node, 1);
2252 if (err)
2253 goto out_flush;
2254
2255 err = move_pages_and_store_status(current_node, &pagelist,
2256 status, start, i, nr_pages);
2257 if (err) {
2258 /* We have accounted for page i */
2259 if (err > 0)
2260 err--;
2261 goto out;
2262 }
2263 current_node = NUMA_NO_NODE;
2264 }
2265out_flush:
2266 /* Make sure we do not overwrite the existing error */
2267 err1 = move_pages_and_store_status(current_node, &pagelist,
2268 status, start, i, nr_pages);
2269 if (err >= 0)
2270 err = err1;
2271out:
2272 lru_cache_enable();
2273 return err;
2274}
2275
2276/*
2277 * Determine the nodes of an array of pages and store it in an array of status.
2278 */
2279static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2280 const void __user **pages, int *status)
2281{
2282 unsigned long i;
2283
2284 mmap_read_lock(mm);
2285
2286 for (i = 0; i < nr_pages; i++) {
2287 unsigned long addr = (unsigned long)(*pages);
2288 struct vm_area_struct *vma;
2289 struct page *page;
2290 int err = -EFAULT;
2291
2292 vma = vma_lookup(mm, addr);
2293 if (!vma)
2294 goto set_status;
2295
2296 /* FOLL_DUMP to ignore special (like zero) pages */
2297 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2298
2299 err = PTR_ERR(page);
2300 if (IS_ERR(page))
2301 goto set_status;
2302
2303 err = -ENOENT;
2304 if (!page)
2305 goto set_status;
2306
2307 if (!is_zone_device_page(page))
2308 err = page_to_nid(page);
2309
2310 put_page(page);
2311set_status:
2312 *status = err;
2313
2314 pages++;
2315 status++;
2316 }
2317
2318 mmap_read_unlock(mm);
2319}
2320
2321static int get_compat_pages_array(const void __user *chunk_pages[],
2322 const void __user * __user *pages,
2323 unsigned long chunk_nr)
2324{
2325 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2326 compat_uptr_t p;
2327 int i;
2328
2329 for (i = 0; i < chunk_nr; i++) {
2330 if (get_user(p, pages32 + i))
2331 return -EFAULT;
2332 chunk_pages[i] = compat_ptr(p);
2333 }
2334
2335 return 0;
2336}
2337
2338/*
2339 * Determine the nodes of a user array of pages and store it in
2340 * a user array of status.
2341 */
2342static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2343 const void __user * __user *pages,
2344 int __user *status)
2345{
2346#define DO_PAGES_STAT_CHUNK_NR 16UL
2347 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2348 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2349
2350 while (nr_pages) {
2351 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2352
2353 if (in_compat_syscall()) {
2354 if (get_compat_pages_array(chunk_pages, pages,
2355 chunk_nr))
2356 break;
2357 } else {
2358 if (copy_from_user(chunk_pages, pages,
2359 chunk_nr * sizeof(*chunk_pages)))
2360 break;
2361 }
2362
2363 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2364
2365 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2366 break;
2367
2368 pages += chunk_nr;
2369 status += chunk_nr;
2370 nr_pages -= chunk_nr;
2371 }
2372 return nr_pages ? -EFAULT : 0;
2373}
2374
2375static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2376{
2377 struct task_struct *task;
2378 struct mm_struct *mm;
2379
2380 /*
2381 * There is no need to check if current process has the right to modify
2382 * the specified process when they are same.
2383 */
2384 if (!pid) {
2385 mmget(current->mm);
2386 *mem_nodes = cpuset_mems_allowed(current);
2387 return current->mm;
2388 }
2389
2390 /* Find the mm_struct */
2391 rcu_read_lock();
2392 task = find_task_by_vpid(pid);
2393 if (!task) {
2394 rcu_read_unlock();
2395 return ERR_PTR(-ESRCH);
2396 }
2397 get_task_struct(task);
2398
2399 /*
2400 * Check if this process has the right to modify the specified
2401 * process. Use the regular "ptrace_may_access()" checks.
2402 */
2403 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2404 rcu_read_unlock();
2405 mm = ERR_PTR(-EPERM);
2406 goto out;
2407 }
2408 rcu_read_unlock();
2409
2410 mm = ERR_PTR(security_task_movememory(task));
2411 if (IS_ERR(mm))
2412 goto out;
2413 *mem_nodes = cpuset_mems_allowed(task);
2414 mm = get_task_mm(task);
2415out:
2416 put_task_struct(task);
2417 if (!mm)
2418 mm = ERR_PTR(-EINVAL);
2419 return mm;
2420}
2421
2422/*
2423 * Move a list of pages in the address space of the currently executing
2424 * process.
2425 */
2426static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2427 const void __user * __user *pages,
2428 const int __user *nodes,
2429 int __user *status, int flags)
2430{
2431 struct mm_struct *mm;
2432 int err;
2433 nodemask_t task_nodes;
2434
2435 /* Check flags */
2436 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2437 return -EINVAL;
2438
2439 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2440 return -EPERM;
2441
2442 mm = find_mm_struct(pid, &task_nodes);
2443 if (IS_ERR(mm))
2444 return PTR_ERR(mm);
2445
2446 if (nodes)
2447 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2448 nodes, status, flags);
2449 else
2450 err = do_pages_stat(mm, nr_pages, pages, status);
2451
2452 mmput(mm);
2453 return err;
2454}
2455
2456SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2457 const void __user * __user *, pages,
2458 const int __user *, nodes,
2459 int __user *, status, int, flags)
2460{
2461 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2462}
2463
2464#ifdef CONFIG_NUMA_BALANCING
2465/*
2466 * Returns true if this is a safe migration target node for misplaced NUMA
2467 * pages. Currently it only checks the watermarks which is crude.
2468 */
2469static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2470 unsigned long nr_migrate_pages)
2471{
2472 int z;
2473
2474 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2475 struct zone *zone = pgdat->node_zones + z;
2476
2477 if (!managed_zone(zone))
2478 continue;
2479
2480 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2481 if (!zone_watermark_ok(zone, 0,
2482 high_wmark_pages(zone) +
2483 nr_migrate_pages,
2484 ZONE_MOVABLE, 0))
2485 continue;
2486 return true;
2487 }
2488 return false;
2489}
2490
2491static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2492 unsigned long data)
2493{
2494 int nid = (int) data;
2495 int order = folio_order(src);
2496 gfp_t gfp = __GFP_THISNODE;
2497
2498 if (order > 0)
2499 gfp |= GFP_TRANSHUGE_LIGHT;
2500 else {
2501 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2502 __GFP_NOWARN;
2503 gfp &= ~__GFP_RECLAIM;
2504 }
2505 return __folio_alloc_node(gfp, order, nid);
2506}
2507
2508static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2509{
2510 int nr_pages = folio_nr_pages(folio);
2511
2512 /* Avoid migrating to a node that is nearly full */
2513 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2514 int z;
2515
2516 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2517 return 0;
2518 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2519 if (managed_zone(pgdat->node_zones + z))
2520 break;
2521 }
2522
2523 /*
2524 * If there are no managed zones, it should not proceed
2525 * further.
2526 */
2527 if (z < 0)
2528 return 0;
2529
2530 wakeup_kswapd(pgdat->node_zones + z, 0,
2531 folio_order(folio), ZONE_MOVABLE);
2532 return 0;
2533 }
2534
2535 if (!folio_isolate_lru(folio))
2536 return 0;
2537
2538 node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2539 nr_pages);
2540
2541 /*
2542 * Isolating the folio has taken another reference, so the
2543 * caller's reference can be safely dropped without the folio
2544 * disappearing underneath us during migration.
2545 */
2546 folio_put(folio);
2547 return 1;
2548}
2549
2550/*
2551 * Attempt to migrate a misplaced folio to the specified destination
2552 * node. Caller is expected to have an elevated reference count on
2553 * the folio that will be dropped by this function before returning.
2554 */
2555int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2556 int node)
2557{
2558 pg_data_t *pgdat = NODE_DATA(node);
2559 int isolated;
2560 int nr_remaining;
2561 unsigned int nr_succeeded;
2562 LIST_HEAD(migratepages);
2563 int nr_pages = folio_nr_pages(folio);
2564
2565 /*
2566 * Don't migrate file folios that are mapped in multiple processes
2567 * with execute permissions as they are probably shared libraries.
2568 * To check if the folio is shared, ideally we want to make sure
2569 * every page is mapped to the same process. Doing that is very
2570 * expensive, so check the estimated mapcount of the folio instead.
2571 */
2572 if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2573 (vma->vm_flags & VM_EXEC))
2574 goto out;
2575
2576 /*
2577 * Also do not migrate dirty folios as not all filesystems can move
2578 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2579 */
2580 if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2581 goto out;
2582
2583 isolated = numamigrate_isolate_folio(pgdat, folio);
2584 if (!isolated)
2585 goto out;
2586
2587 list_add(&folio->lru, &migratepages);
2588 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2589 NULL, node, MIGRATE_ASYNC,
2590 MR_NUMA_MISPLACED, &nr_succeeded);
2591 if (nr_remaining) {
2592 if (!list_empty(&migratepages)) {
2593 list_del(&folio->lru);
2594 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
2595 folio_is_file_lru(folio), -nr_pages);
2596 folio_putback_lru(folio);
2597 }
2598 isolated = 0;
2599 }
2600 if (nr_succeeded) {
2601 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2602 if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
2603 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2604 nr_succeeded);
2605 }
2606 BUG_ON(!list_empty(&migratepages));
2607 return isolated;
2608
2609out:
2610 folio_put(folio);
2611 return 0;
2612}
2613#endif /* CONFIG_NUMA_BALANCING */
2614#endif /* CONFIG_NUMA */