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/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 */
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter
13 */
14
15#include <linux/migrate.h>
16#include <linux/export.h>
17#include <linux/swap.h>
18#include <linux/swapops.h>
19#include <linux/pagemap.h>
20#include <linux/buffer_head.h>
21#include <linux/mm_inline.h>
22#include <linux/nsproxy.h>
23#include <linux/pagevec.h>
24#include <linux/ksm.h>
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
29#include <linux/writeback.h>
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
32#include <linux/security.h>
33#include <linux/memcontrol.h>
34#include <linux/syscalls.h>
35#include <linux/hugetlb.h>
36#include <linux/hugetlb_cgroup.h>
37#include <linux/gfp.h>
38#include <linux/balloon_compaction.h>
39#include <linux/mmu_notifier.h>
40
41#include <asm/tlbflush.h>
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
46#include "internal.h"
47
48/*
49 * migrate_prep() needs to be called before we start compiling a list of pages
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
52 */
53int migrate_prep(void)
54{
55 /*
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
60 */
61 lru_add_drain_all();
62
63 return 0;
64}
65
66/* Do the necessary work of migrate_prep but not if it involves other CPUs */
67int migrate_prep_local(void)
68{
69 lru_add_drain();
70
71 return 0;
72}
73
74/*
75 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
78 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
81 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
87 list_for_each_entry_safe(page, page2, l, lru) {
88 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
92 list_del(&page->lru);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
95 if (unlikely(isolated_balloon_page(page)))
96 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
99 }
100}
101
102/*
103 * Restore a potential migration pte to a working pte entry
104 */
105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
119 } else {
120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
122 goto out;
123 if (pmd_trans_huge(*pmd))
124 goto out;
125
126 ptep = pte_offset_map(pmd, addr);
127
128 /*
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
131 */
132
133 ptl = pte_lockptr(mm, pmd);
134 }
135
136 spin_lock(ptl);
137 pte = *ptep;
138 if (!is_swap_pte(pte))
139 goto unlock;
140
141 entry = pte_to_swp_entry(pte);
142
143 if (!is_migration_entry(entry) ||
144 migration_entry_to_page(entry) != old)
145 goto unlock;
146
147 get_page(new);
148 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
149 if (pte_swp_soft_dirty(*ptep))
150 pte = pte_mksoft_dirty(pte);
151 if (is_write_migration_entry(entry))
152 pte = pte_mkwrite(pte);
153#ifdef CONFIG_HUGETLB_PAGE
154 if (PageHuge(new)) {
155 pte = pte_mkhuge(pte);
156 pte = arch_make_huge_pte(pte, vma, new, 0);
157 }
158#endif
159 flush_dcache_page(new);
160 set_pte_at(mm, addr, ptep, pte);
161
162 if (PageHuge(new)) {
163 if (PageAnon(new))
164 hugepage_add_anon_rmap(new, vma, addr);
165 else
166 page_dup_rmap(new);
167 } else if (PageAnon(new))
168 page_add_anon_rmap(new, vma, addr);
169 else
170 page_add_file_rmap(new);
171
172 /* No need to invalidate - it was non-present before */
173 update_mmu_cache(vma, addr, ptep);
174unlock:
175 pte_unmap_unlock(ptep, ptl);
176out:
177 return SWAP_AGAIN;
178}
179
180/*
181 * Congratulations to trinity for discovering this bug.
182 * mm/fremap.c's remap_file_pages() accepts any range within a single vma to
183 * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then
184 * replace the specified range by file ptes throughout (maybe populated after).
185 * If page migration finds a page within that range, while it's still located
186 * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem:
187 * zap_pte() clears the temporary migration entry before mmap_sem is dropped.
188 * But if the migrating page is in a part of the vma outside the range to be
189 * remapped, then it will not be cleared, and remove_migration_ptes() needs to
190 * deal with it. Fortunately, this part of the vma is of course still linear,
191 * so we just need to use linear location on the nonlinear list.
192 */
193static int remove_linear_migration_ptes_from_nonlinear(struct page *page,
194 struct address_space *mapping, void *arg)
195{
196 struct vm_area_struct *vma;
197 /* hugetlbfs does not support remap_pages, so no huge pgoff worries */
198 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
199 unsigned long addr;
200
201 list_for_each_entry(vma,
202 &mapping->i_mmap_nonlinear, shared.nonlinear) {
203
204 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
205 if (addr >= vma->vm_start && addr < vma->vm_end)
206 remove_migration_pte(page, vma, addr, arg);
207 }
208 return SWAP_AGAIN;
209}
210
211/*
212 * Get rid of all migration entries and replace them by
213 * references to the indicated page.
214 */
215static void remove_migration_ptes(struct page *old, struct page *new)
216{
217 struct rmap_walk_control rwc = {
218 .rmap_one = remove_migration_pte,
219 .arg = old,
220 .file_nonlinear = remove_linear_migration_ptes_from_nonlinear,
221 };
222
223 rmap_walk(new, &rwc);
224}
225
226/*
227 * Something used the pte of a page under migration. We need to
228 * get to the page and wait until migration is finished.
229 * When we return from this function the fault will be retried.
230 */
231static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
232 spinlock_t *ptl)
233{
234 pte_t pte;
235 swp_entry_t entry;
236 struct page *page;
237
238 spin_lock(ptl);
239 pte = *ptep;
240 if (!is_swap_pte(pte))
241 goto out;
242
243 entry = pte_to_swp_entry(pte);
244 if (!is_migration_entry(entry))
245 goto out;
246
247 page = migration_entry_to_page(entry);
248
249 /*
250 * Once radix-tree replacement of page migration started, page_count
251 * *must* be zero. And, we don't want to call wait_on_page_locked()
252 * against a page without get_page().
253 * So, we use get_page_unless_zero(), here. Even failed, page fault
254 * will occur again.
255 */
256 if (!get_page_unless_zero(page))
257 goto out;
258 pte_unmap_unlock(ptep, ptl);
259 wait_on_page_locked(page);
260 put_page(page);
261 return;
262out:
263 pte_unmap_unlock(ptep, ptl);
264}
265
266void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
267 unsigned long address)
268{
269 spinlock_t *ptl = pte_lockptr(mm, pmd);
270 pte_t *ptep = pte_offset_map(pmd, address);
271 __migration_entry_wait(mm, ptep, ptl);
272}
273
274void migration_entry_wait_huge(struct vm_area_struct *vma,
275 struct mm_struct *mm, pte_t *pte)
276{
277 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
278 __migration_entry_wait(mm, pte, ptl);
279}
280
281#ifdef CONFIG_BLOCK
282/* Returns true if all buffers are successfully locked */
283static bool buffer_migrate_lock_buffers(struct buffer_head *head,
284 enum migrate_mode mode)
285{
286 struct buffer_head *bh = head;
287
288 /* Simple case, sync compaction */
289 if (mode != MIGRATE_ASYNC) {
290 do {
291 get_bh(bh);
292 lock_buffer(bh);
293 bh = bh->b_this_page;
294
295 } while (bh != head);
296
297 return true;
298 }
299
300 /* async case, we cannot block on lock_buffer so use trylock_buffer */
301 do {
302 get_bh(bh);
303 if (!trylock_buffer(bh)) {
304 /*
305 * We failed to lock the buffer and cannot stall in
306 * async migration. Release the taken locks
307 */
308 struct buffer_head *failed_bh = bh;
309 put_bh(failed_bh);
310 bh = head;
311 while (bh != failed_bh) {
312 unlock_buffer(bh);
313 put_bh(bh);
314 bh = bh->b_this_page;
315 }
316 return false;
317 }
318
319 bh = bh->b_this_page;
320 } while (bh != head);
321 return true;
322}
323#else
324static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
325 enum migrate_mode mode)
326{
327 return true;
328}
329#endif /* CONFIG_BLOCK */
330
331/*
332 * Replace the page in the mapping.
333 *
334 * The number of remaining references must be:
335 * 1 for anonymous pages without a mapping
336 * 2 for pages with a mapping
337 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
338 */
339int migrate_page_move_mapping(struct address_space *mapping,
340 struct page *newpage, struct page *page,
341 struct buffer_head *head, enum migrate_mode mode,
342 int extra_count)
343{
344 int expected_count = 1 + extra_count;
345 void **pslot;
346
347 if (!mapping) {
348 /* Anonymous page without mapping */
349 if (page_count(page) != expected_count)
350 return -EAGAIN;
351 return MIGRATEPAGE_SUCCESS;
352 }
353
354 spin_lock_irq(&mapping->tree_lock);
355
356 pslot = radix_tree_lookup_slot(&mapping->page_tree,
357 page_index(page));
358
359 expected_count += 1 + page_has_private(page);
360 if (page_count(page) != expected_count ||
361 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
362 spin_unlock_irq(&mapping->tree_lock);
363 return -EAGAIN;
364 }
365
366 if (!page_freeze_refs(page, expected_count)) {
367 spin_unlock_irq(&mapping->tree_lock);
368 return -EAGAIN;
369 }
370
371 /*
372 * In the async migration case of moving a page with buffers, lock the
373 * buffers using trylock before the mapping is moved. If the mapping
374 * was moved, we later failed to lock the buffers and could not move
375 * the mapping back due to an elevated page count, we would have to
376 * block waiting on other references to be dropped.
377 */
378 if (mode == MIGRATE_ASYNC && head &&
379 !buffer_migrate_lock_buffers(head, mode)) {
380 page_unfreeze_refs(page, expected_count);
381 spin_unlock_irq(&mapping->tree_lock);
382 return -EAGAIN;
383 }
384
385 /*
386 * Now we know that no one else is looking at the page.
387 */
388 get_page(newpage); /* add cache reference */
389 if (PageSwapCache(page)) {
390 SetPageSwapCache(newpage);
391 set_page_private(newpage, page_private(page));
392 }
393
394 radix_tree_replace_slot(pslot, newpage);
395
396 /*
397 * Drop cache reference from old page by unfreezing
398 * to one less reference.
399 * We know this isn't the last reference.
400 */
401 page_unfreeze_refs(page, expected_count - 1);
402
403 /*
404 * If moved to a different zone then also account
405 * the page for that zone. Other VM counters will be
406 * taken care of when we establish references to the
407 * new page and drop references to the old page.
408 *
409 * Note that anonymous pages are accounted for
410 * via NR_FILE_PAGES and NR_ANON_PAGES if they
411 * are mapped to swap space.
412 */
413 __dec_zone_page_state(page, NR_FILE_PAGES);
414 __inc_zone_page_state(newpage, NR_FILE_PAGES);
415 if (!PageSwapCache(page) && PageSwapBacked(page)) {
416 __dec_zone_page_state(page, NR_SHMEM);
417 __inc_zone_page_state(newpage, NR_SHMEM);
418 }
419 spin_unlock_irq(&mapping->tree_lock);
420
421 return MIGRATEPAGE_SUCCESS;
422}
423
424/*
425 * The expected number of remaining references is the same as that
426 * of migrate_page_move_mapping().
427 */
428int migrate_huge_page_move_mapping(struct address_space *mapping,
429 struct page *newpage, struct page *page)
430{
431 int expected_count;
432 void **pslot;
433
434 if (!mapping) {
435 if (page_count(page) != 1)
436 return -EAGAIN;
437 return MIGRATEPAGE_SUCCESS;
438 }
439
440 spin_lock_irq(&mapping->tree_lock);
441
442 pslot = radix_tree_lookup_slot(&mapping->page_tree,
443 page_index(page));
444
445 expected_count = 2 + page_has_private(page);
446 if (page_count(page) != expected_count ||
447 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
448 spin_unlock_irq(&mapping->tree_lock);
449 return -EAGAIN;
450 }
451
452 if (!page_freeze_refs(page, expected_count)) {
453 spin_unlock_irq(&mapping->tree_lock);
454 return -EAGAIN;
455 }
456
457 get_page(newpage);
458
459 radix_tree_replace_slot(pslot, newpage);
460
461 page_unfreeze_refs(page, expected_count - 1);
462
463 spin_unlock_irq(&mapping->tree_lock);
464 return MIGRATEPAGE_SUCCESS;
465}
466
467/*
468 * Gigantic pages are so large that we do not guarantee that page++ pointer
469 * arithmetic will work across the entire page. We need something more
470 * specialized.
471 */
472static void __copy_gigantic_page(struct page *dst, struct page *src,
473 int nr_pages)
474{
475 int i;
476 struct page *dst_base = dst;
477 struct page *src_base = src;
478
479 for (i = 0; i < nr_pages; ) {
480 cond_resched();
481 copy_highpage(dst, src);
482
483 i++;
484 dst = mem_map_next(dst, dst_base, i);
485 src = mem_map_next(src, src_base, i);
486 }
487}
488
489static void copy_huge_page(struct page *dst, struct page *src)
490{
491 int i;
492 int nr_pages;
493
494 if (PageHuge(src)) {
495 /* hugetlbfs page */
496 struct hstate *h = page_hstate(src);
497 nr_pages = pages_per_huge_page(h);
498
499 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
500 __copy_gigantic_page(dst, src, nr_pages);
501 return;
502 }
503 } else {
504 /* thp page */
505 BUG_ON(!PageTransHuge(src));
506 nr_pages = hpage_nr_pages(src);
507 }
508
509 for (i = 0; i < nr_pages; i++) {
510 cond_resched();
511 copy_highpage(dst + i, src + i);
512 }
513}
514
515/*
516 * Copy the page to its new location
517 */
518void migrate_page_copy(struct page *newpage, struct page *page)
519{
520 int cpupid;
521
522 if (PageHuge(page) || PageTransHuge(page))
523 copy_huge_page(newpage, page);
524 else
525 copy_highpage(newpage, page);
526
527 if (PageError(page))
528 SetPageError(newpage);
529 if (PageReferenced(page))
530 SetPageReferenced(newpage);
531 if (PageUptodate(page))
532 SetPageUptodate(newpage);
533 if (TestClearPageActive(page)) {
534 VM_BUG_ON_PAGE(PageUnevictable(page), page);
535 SetPageActive(newpage);
536 } else if (TestClearPageUnevictable(page))
537 SetPageUnevictable(newpage);
538 if (PageChecked(page))
539 SetPageChecked(newpage);
540 if (PageMappedToDisk(page))
541 SetPageMappedToDisk(newpage);
542
543 if (PageDirty(page)) {
544 clear_page_dirty_for_io(page);
545 /*
546 * Want to mark the page and the radix tree as dirty, and
547 * redo the accounting that clear_page_dirty_for_io undid,
548 * but we can't use set_page_dirty because that function
549 * is actually a signal that all of the page has become dirty.
550 * Whereas only part of our page may be dirty.
551 */
552 if (PageSwapBacked(page))
553 SetPageDirty(newpage);
554 else
555 __set_page_dirty_nobuffers(newpage);
556 }
557
558 /*
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
561 */
562 cpupid = page_cpupid_xchg_last(page, -1);
563 page_cpupid_xchg_last(newpage, cpupid);
564
565 mlock_migrate_page(newpage, page);
566 ksm_migrate_page(newpage, page);
567 /*
568 * Please do not reorder this without considering how mm/ksm.c's
569 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
570 */
571 ClearPageSwapCache(page);
572 ClearPagePrivate(page);
573 set_page_private(page, 0);
574
575 /*
576 * If any waiters have accumulated on the new page then
577 * wake them up.
578 */
579 if (PageWriteback(newpage))
580 end_page_writeback(newpage);
581}
582
583/************************************************************
584 * Migration functions
585 ***********************************************************/
586
587/*
588 * Common logic to directly migrate a single page suitable for
589 * pages that do not use PagePrivate/PagePrivate2.
590 *
591 * Pages are locked upon entry and exit.
592 */
593int migrate_page(struct address_space *mapping,
594 struct page *newpage, struct page *page,
595 enum migrate_mode mode)
596{
597 int rc;
598
599 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
600
601 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
602
603 if (rc != MIGRATEPAGE_SUCCESS)
604 return rc;
605
606 migrate_page_copy(newpage, page);
607 return MIGRATEPAGE_SUCCESS;
608}
609EXPORT_SYMBOL(migrate_page);
610
611#ifdef CONFIG_BLOCK
612/*
613 * Migration function for pages with buffers. This function can only be used
614 * if the underlying filesystem guarantees that no other references to "page"
615 * exist.
616 */
617int buffer_migrate_page(struct address_space *mapping,
618 struct page *newpage, struct page *page, enum migrate_mode mode)
619{
620 struct buffer_head *bh, *head;
621 int rc;
622
623 if (!page_has_buffers(page))
624 return migrate_page(mapping, newpage, page, mode);
625
626 head = page_buffers(page);
627
628 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
629
630 if (rc != MIGRATEPAGE_SUCCESS)
631 return rc;
632
633 /*
634 * In the async case, migrate_page_move_mapping locked the buffers
635 * with an IRQ-safe spinlock held. In the sync case, the buffers
636 * need to be locked now
637 */
638 if (mode != MIGRATE_ASYNC)
639 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
640
641 ClearPagePrivate(page);
642 set_page_private(newpage, page_private(page));
643 set_page_private(page, 0);
644 put_page(page);
645 get_page(newpage);
646
647 bh = head;
648 do {
649 set_bh_page(bh, newpage, bh_offset(bh));
650 bh = bh->b_this_page;
651
652 } while (bh != head);
653
654 SetPagePrivate(newpage);
655
656 migrate_page_copy(newpage, page);
657
658 bh = head;
659 do {
660 unlock_buffer(bh);
661 put_bh(bh);
662 bh = bh->b_this_page;
663
664 } while (bh != head);
665
666 return MIGRATEPAGE_SUCCESS;
667}
668EXPORT_SYMBOL(buffer_migrate_page);
669#endif
670
671/*
672 * Writeback a page to clean the dirty state
673 */
674static int writeout(struct address_space *mapping, struct page *page)
675{
676 struct writeback_control wbc = {
677 .sync_mode = WB_SYNC_NONE,
678 .nr_to_write = 1,
679 .range_start = 0,
680 .range_end = LLONG_MAX,
681 .for_reclaim = 1
682 };
683 int rc;
684
685 if (!mapping->a_ops->writepage)
686 /* No write method for the address space */
687 return -EINVAL;
688
689 if (!clear_page_dirty_for_io(page))
690 /* Someone else already triggered a write */
691 return -EAGAIN;
692
693 /*
694 * A dirty page may imply that the underlying filesystem has
695 * the page on some queue. So the page must be clean for
696 * migration. Writeout may mean we loose the lock and the
697 * page state is no longer what we checked for earlier.
698 * At this point we know that the migration attempt cannot
699 * be successful.
700 */
701 remove_migration_ptes(page, page);
702
703 rc = mapping->a_ops->writepage(page, &wbc);
704
705 if (rc != AOP_WRITEPAGE_ACTIVATE)
706 /* unlocked. Relock */
707 lock_page(page);
708
709 return (rc < 0) ? -EIO : -EAGAIN;
710}
711
712/*
713 * Default handling if a filesystem does not provide a migration function.
714 */
715static int fallback_migrate_page(struct address_space *mapping,
716 struct page *newpage, struct page *page, enum migrate_mode mode)
717{
718 if (PageDirty(page)) {
719 /* Only writeback pages in full synchronous migration */
720 if (mode != MIGRATE_SYNC)
721 return -EBUSY;
722 return writeout(mapping, page);
723 }
724
725 /*
726 * Buffers may be managed in a filesystem specific way.
727 * We must have no buffers or drop them.
728 */
729 if (page_has_private(page) &&
730 !try_to_release_page(page, GFP_KERNEL))
731 return -EAGAIN;
732
733 return migrate_page(mapping, newpage, page, mode);
734}
735
736/*
737 * Move a page to a newly allocated page
738 * The page is locked and all ptes have been successfully removed.
739 *
740 * The new page will have replaced the old page if this function
741 * is successful.
742 *
743 * Return value:
744 * < 0 - error code
745 * MIGRATEPAGE_SUCCESS - success
746 */
747static int move_to_new_page(struct page *newpage, struct page *page,
748 int remap_swapcache, enum migrate_mode mode)
749{
750 struct address_space *mapping;
751 int rc;
752
753 /*
754 * Block others from accessing the page when we get around to
755 * establishing additional references. We are the only one
756 * holding a reference to the new page at this point.
757 */
758 if (!trylock_page(newpage))
759 BUG();
760
761 /* Prepare mapping for the new page.*/
762 newpage->index = page->index;
763 newpage->mapping = page->mapping;
764 if (PageSwapBacked(page))
765 SetPageSwapBacked(newpage);
766
767 mapping = page_mapping(page);
768 if (!mapping)
769 rc = migrate_page(mapping, newpage, page, mode);
770 else if (mapping->a_ops->migratepage)
771 /*
772 * Most pages have a mapping and most filesystems provide a
773 * migratepage callback. Anonymous pages are part of swap
774 * space which also has its own migratepage callback. This
775 * is the most common path for page migration.
776 */
777 rc = mapping->a_ops->migratepage(mapping,
778 newpage, page, mode);
779 else
780 rc = fallback_migrate_page(mapping, newpage, page, mode);
781
782 if (rc != MIGRATEPAGE_SUCCESS) {
783 newpage->mapping = NULL;
784 } else {
785 if (remap_swapcache)
786 remove_migration_ptes(page, newpage);
787 page->mapping = NULL;
788 }
789
790 unlock_page(newpage);
791
792 return rc;
793}
794
795static int __unmap_and_move(struct page *page, struct page *newpage,
796 int force, enum migrate_mode mode)
797{
798 int rc = -EAGAIN;
799 int remap_swapcache = 1;
800 struct mem_cgroup *mem;
801 struct anon_vma *anon_vma = NULL;
802
803 if (!trylock_page(page)) {
804 if (!force || mode == MIGRATE_ASYNC)
805 goto out;
806
807 /*
808 * It's not safe for direct compaction to call lock_page.
809 * For example, during page readahead pages are added locked
810 * to the LRU. Later, when the IO completes the pages are
811 * marked uptodate and unlocked. However, the queueing
812 * could be merging multiple pages for one bio (e.g.
813 * mpage_readpages). If an allocation happens for the
814 * second or third page, the process can end up locking
815 * the same page twice and deadlocking. Rather than
816 * trying to be clever about what pages can be locked,
817 * avoid the use of lock_page for direct compaction
818 * altogether.
819 */
820 if (current->flags & PF_MEMALLOC)
821 goto out;
822
823 lock_page(page);
824 }
825
826 /* charge against new page */
827 mem_cgroup_prepare_migration(page, newpage, &mem);
828
829 if (PageWriteback(page)) {
830 /*
831 * Only in the case of a full synchronous migration is it
832 * necessary to wait for PageWriteback. In the async case,
833 * the retry loop is too short and in the sync-light case,
834 * the overhead of stalling is too much
835 */
836 if (mode != MIGRATE_SYNC) {
837 rc = -EBUSY;
838 goto uncharge;
839 }
840 if (!force)
841 goto uncharge;
842 wait_on_page_writeback(page);
843 }
844 /*
845 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
846 * we cannot notice that anon_vma is freed while we migrates a page.
847 * This get_anon_vma() delays freeing anon_vma pointer until the end
848 * of migration. File cache pages are no problem because of page_lock()
849 * File Caches may use write_page() or lock_page() in migration, then,
850 * just care Anon page here.
851 */
852 if (PageAnon(page) && !PageKsm(page)) {
853 /*
854 * Only page_lock_anon_vma_read() understands the subtleties of
855 * getting a hold on an anon_vma from outside one of its mms.
856 */
857 anon_vma = page_get_anon_vma(page);
858 if (anon_vma) {
859 /*
860 * Anon page
861 */
862 } else if (PageSwapCache(page)) {
863 /*
864 * We cannot be sure that the anon_vma of an unmapped
865 * swapcache page is safe to use because we don't
866 * know in advance if the VMA that this page belonged
867 * to still exists. If the VMA and others sharing the
868 * data have been freed, then the anon_vma could
869 * already be invalid.
870 *
871 * To avoid this possibility, swapcache pages get
872 * migrated but are not remapped when migration
873 * completes
874 */
875 remap_swapcache = 0;
876 } else {
877 goto uncharge;
878 }
879 }
880
881 if (unlikely(balloon_page_movable(page))) {
882 /*
883 * A ballooned page does not need any special attention from
884 * physical to virtual reverse mapping procedures.
885 * Skip any attempt to unmap PTEs or to remap swap cache,
886 * in order to avoid burning cycles at rmap level, and perform
887 * the page migration right away (proteced by page lock).
888 */
889 rc = balloon_page_migrate(newpage, page, mode);
890 goto uncharge;
891 }
892
893 /*
894 * Corner case handling:
895 * 1. When a new swap-cache page is read into, it is added to the LRU
896 * and treated as swapcache but it has no rmap yet.
897 * Calling try_to_unmap() against a page->mapping==NULL page will
898 * trigger a BUG. So handle it here.
899 * 2. An orphaned page (see truncate_complete_page) might have
900 * fs-private metadata. The page can be picked up due to memory
901 * offlining. Everywhere else except page reclaim, the page is
902 * invisible to the vm, so the page can not be migrated. So try to
903 * free the metadata, so the page can be freed.
904 */
905 if (!page->mapping) {
906 VM_BUG_ON_PAGE(PageAnon(page), page);
907 if (page_has_private(page)) {
908 try_to_free_buffers(page);
909 goto uncharge;
910 }
911 goto skip_unmap;
912 }
913
914 /* Establish migration ptes or remove ptes */
915 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
916
917skip_unmap:
918 if (!page_mapped(page))
919 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
920
921 if (rc && remap_swapcache)
922 remove_migration_ptes(page, page);
923
924 /* Drop an anon_vma reference if we took one */
925 if (anon_vma)
926 put_anon_vma(anon_vma);
927
928uncharge:
929 mem_cgroup_end_migration(mem, page, newpage,
930 (rc == MIGRATEPAGE_SUCCESS ||
931 rc == MIGRATEPAGE_BALLOON_SUCCESS));
932 unlock_page(page);
933out:
934 return rc;
935}
936
937/*
938 * Obtain the lock on page, remove all ptes and migrate the page
939 * to the newly allocated page in newpage.
940 */
941static int unmap_and_move(new_page_t get_new_page, unsigned long private,
942 struct page *page, int force, enum migrate_mode mode)
943{
944 int rc = 0;
945 int *result = NULL;
946 struct page *newpage = get_new_page(page, private, &result);
947
948 if (!newpage)
949 return -ENOMEM;
950
951 if (page_count(page) == 1) {
952 /* page was freed from under us. So we are done. */
953 goto out;
954 }
955
956 if (unlikely(PageTransHuge(page)))
957 if (unlikely(split_huge_page(page)))
958 goto out;
959
960 rc = __unmap_and_move(page, newpage, force, mode);
961
962 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
963 /*
964 * A ballooned page has been migrated already.
965 * Now, it's the time to wrap-up counters,
966 * handle the page back to Buddy and return.
967 */
968 dec_zone_page_state(page, NR_ISOLATED_ANON +
969 page_is_file_cache(page));
970 balloon_page_free(page);
971 return MIGRATEPAGE_SUCCESS;
972 }
973out:
974 if (rc != -EAGAIN) {
975 /*
976 * A page that has been migrated has all references
977 * removed and will be freed. A page that has not been
978 * migrated will have kepts its references and be
979 * restored.
980 */
981 list_del(&page->lru);
982 dec_zone_page_state(page, NR_ISOLATED_ANON +
983 page_is_file_cache(page));
984 putback_lru_page(page);
985 }
986 /*
987 * Move the new page to the LRU. If migration was not successful
988 * then this will free the page.
989 */
990 putback_lru_page(newpage);
991 if (result) {
992 if (rc)
993 *result = rc;
994 else
995 *result = page_to_nid(newpage);
996 }
997 return rc;
998}
999
1000/*
1001 * Counterpart of unmap_and_move_page() for hugepage migration.
1002 *
1003 * This function doesn't wait the completion of hugepage I/O
1004 * because there is no race between I/O and migration for hugepage.
1005 * Note that currently hugepage I/O occurs only in direct I/O
1006 * where no lock is held and PG_writeback is irrelevant,
1007 * and writeback status of all subpages are counted in the reference
1008 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1009 * under direct I/O, the reference of the head page is 512 and a bit more.)
1010 * This means that when we try to migrate hugepage whose subpages are
1011 * doing direct I/O, some references remain after try_to_unmap() and
1012 * hugepage migration fails without data corruption.
1013 *
1014 * There is also no race when direct I/O is issued on the page under migration,
1015 * because then pte is replaced with migration swap entry and direct I/O code
1016 * will wait in the page fault for migration to complete.
1017 */
1018static int unmap_and_move_huge_page(new_page_t get_new_page,
1019 unsigned long private, struct page *hpage,
1020 int force, enum migrate_mode mode)
1021{
1022 int rc = 0;
1023 int *result = NULL;
1024 struct page *new_hpage;
1025 struct anon_vma *anon_vma = NULL;
1026
1027 /*
1028 * Movability of hugepages depends on architectures and hugepage size.
1029 * This check is necessary because some callers of hugepage migration
1030 * like soft offline and memory hotremove don't walk through page
1031 * tables or check whether the hugepage is pmd-based or not before
1032 * kicking migration.
1033 */
1034 if (!hugepage_migration_support(page_hstate(hpage))) {
1035 putback_active_hugepage(hpage);
1036 return -ENOSYS;
1037 }
1038
1039 new_hpage = get_new_page(hpage, private, &result);
1040 if (!new_hpage)
1041 return -ENOMEM;
1042
1043 rc = -EAGAIN;
1044
1045 if (!trylock_page(hpage)) {
1046 if (!force || mode != MIGRATE_SYNC)
1047 goto out;
1048 lock_page(hpage);
1049 }
1050
1051 if (PageAnon(hpage))
1052 anon_vma = page_get_anon_vma(hpage);
1053
1054 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1055
1056 if (!page_mapped(hpage))
1057 rc = move_to_new_page(new_hpage, hpage, 1, mode);
1058
1059 if (rc)
1060 remove_migration_ptes(hpage, hpage);
1061
1062 if (anon_vma)
1063 put_anon_vma(anon_vma);
1064
1065 if (!rc)
1066 hugetlb_cgroup_migrate(hpage, new_hpage);
1067
1068 unlock_page(hpage);
1069out:
1070 if (rc != -EAGAIN)
1071 putback_active_hugepage(hpage);
1072 put_page(new_hpage);
1073 if (result) {
1074 if (rc)
1075 *result = rc;
1076 else
1077 *result = page_to_nid(new_hpage);
1078 }
1079 return rc;
1080}
1081
1082/*
1083 * migrate_pages - migrate the pages specified in a list, to the free pages
1084 * supplied as the target for the page migration
1085 *
1086 * @from: The list of pages to be migrated.
1087 * @get_new_page: The function used to allocate free pages to be used
1088 * as the target of the page migration.
1089 * @private: Private data to be passed on to get_new_page()
1090 * @mode: The migration mode that specifies the constraints for
1091 * page migration, if any.
1092 * @reason: The reason for page migration.
1093 *
1094 * The function returns after 10 attempts or if no pages are movable any more
1095 * because the list has become empty or no retryable pages exist any more.
1096 * The caller should call putback_lru_pages() to return pages to the LRU
1097 * or free list only if ret != 0.
1098 *
1099 * Returns the number of pages that were not migrated, or an error code.
1100 */
1101int migrate_pages(struct list_head *from, new_page_t get_new_page,
1102 unsigned long private, enum migrate_mode mode, int reason)
1103{
1104 int retry = 1;
1105 int nr_failed = 0;
1106 int nr_succeeded = 0;
1107 int pass = 0;
1108 struct page *page;
1109 struct page *page2;
1110 int swapwrite = current->flags & PF_SWAPWRITE;
1111 int rc;
1112
1113 if (!swapwrite)
1114 current->flags |= PF_SWAPWRITE;
1115
1116 for(pass = 0; pass < 10 && retry; pass++) {
1117 retry = 0;
1118
1119 list_for_each_entry_safe(page, page2, from, lru) {
1120 cond_resched();
1121
1122 if (PageHuge(page))
1123 rc = unmap_and_move_huge_page(get_new_page,
1124 private, page, pass > 2, mode);
1125 else
1126 rc = unmap_and_move(get_new_page, private,
1127 page, pass > 2, mode);
1128
1129 switch(rc) {
1130 case -ENOMEM:
1131 goto out;
1132 case -EAGAIN:
1133 retry++;
1134 break;
1135 case MIGRATEPAGE_SUCCESS:
1136 nr_succeeded++;
1137 break;
1138 default:
1139 /*
1140 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1141 * unlike -EAGAIN case, the failed page is
1142 * removed from migration page list and not
1143 * retried in the next outer loop.
1144 */
1145 nr_failed++;
1146 break;
1147 }
1148 }
1149 }
1150 rc = nr_failed + retry;
1151out:
1152 if (nr_succeeded)
1153 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1154 if (nr_failed)
1155 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1156 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1157
1158 if (!swapwrite)
1159 current->flags &= ~PF_SWAPWRITE;
1160
1161 return rc;
1162}
1163
1164#ifdef CONFIG_NUMA
1165/*
1166 * Move a list of individual pages
1167 */
1168struct page_to_node {
1169 unsigned long addr;
1170 struct page *page;
1171 int node;
1172 int status;
1173};
1174
1175static struct page *new_page_node(struct page *p, unsigned long private,
1176 int **result)
1177{
1178 struct page_to_node *pm = (struct page_to_node *)private;
1179
1180 while (pm->node != MAX_NUMNODES && pm->page != p)
1181 pm++;
1182
1183 if (pm->node == MAX_NUMNODES)
1184 return NULL;
1185
1186 *result = &pm->status;
1187
1188 if (PageHuge(p))
1189 return alloc_huge_page_node(page_hstate(compound_head(p)),
1190 pm->node);
1191 else
1192 return alloc_pages_exact_node(pm->node,
1193 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1194}
1195
1196/*
1197 * Move a set of pages as indicated in the pm array. The addr
1198 * field must be set to the virtual address of the page to be moved
1199 * and the node number must contain a valid target node.
1200 * The pm array ends with node = MAX_NUMNODES.
1201 */
1202static int do_move_page_to_node_array(struct mm_struct *mm,
1203 struct page_to_node *pm,
1204 int migrate_all)
1205{
1206 int err;
1207 struct page_to_node *pp;
1208 LIST_HEAD(pagelist);
1209
1210 down_read(&mm->mmap_sem);
1211
1212 /*
1213 * Build a list of pages to migrate
1214 */
1215 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1216 struct vm_area_struct *vma;
1217 struct page *page;
1218
1219 err = -EFAULT;
1220 vma = find_vma(mm, pp->addr);
1221 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1222 goto set_status;
1223
1224 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1225
1226 err = PTR_ERR(page);
1227 if (IS_ERR(page))
1228 goto set_status;
1229
1230 err = -ENOENT;
1231 if (!page)
1232 goto set_status;
1233
1234 /* Use PageReserved to check for zero page */
1235 if (PageReserved(page))
1236 goto put_and_set;
1237
1238 pp->page = page;
1239 err = page_to_nid(page);
1240
1241 if (err == pp->node)
1242 /*
1243 * Node already in the right place
1244 */
1245 goto put_and_set;
1246
1247 err = -EACCES;
1248 if (page_mapcount(page) > 1 &&
1249 !migrate_all)
1250 goto put_and_set;
1251
1252 if (PageHuge(page)) {
1253 isolate_huge_page(page, &pagelist);
1254 goto put_and_set;
1255 }
1256
1257 err = isolate_lru_page(page);
1258 if (!err) {
1259 list_add_tail(&page->lru, &pagelist);
1260 inc_zone_page_state(page, NR_ISOLATED_ANON +
1261 page_is_file_cache(page));
1262 }
1263put_and_set:
1264 /*
1265 * Either remove the duplicate refcount from
1266 * isolate_lru_page() or drop the page ref if it was
1267 * not isolated.
1268 */
1269 put_page(page);
1270set_status:
1271 pp->status = err;
1272 }
1273
1274 err = 0;
1275 if (!list_empty(&pagelist)) {
1276 err = migrate_pages(&pagelist, new_page_node,
1277 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1278 if (err)
1279 putback_movable_pages(&pagelist);
1280 }
1281
1282 up_read(&mm->mmap_sem);
1283 return err;
1284}
1285
1286/*
1287 * Migrate an array of page address onto an array of nodes and fill
1288 * the corresponding array of status.
1289 */
1290static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1291 unsigned long nr_pages,
1292 const void __user * __user *pages,
1293 const int __user *nodes,
1294 int __user *status, int flags)
1295{
1296 struct page_to_node *pm;
1297 unsigned long chunk_nr_pages;
1298 unsigned long chunk_start;
1299 int err;
1300
1301 err = -ENOMEM;
1302 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1303 if (!pm)
1304 goto out;
1305
1306 migrate_prep();
1307
1308 /*
1309 * Store a chunk of page_to_node array in a page,
1310 * but keep the last one as a marker
1311 */
1312 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1313
1314 for (chunk_start = 0;
1315 chunk_start < nr_pages;
1316 chunk_start += chunk_nr_pages) {
1317 int j;
1318
1319 if (chunk_start + chunk_nr_pages > nr_pages)
1320 chunk_nr_pages = nr_pages - chunk_start;
1321
1322 /* fill the chunk pm with addrs and nodes from user-space */
1323 for (j = 0; j < chunk_nr_pages; j++) {
1324 const void __user *p;
1325 int node;
1326
1327 err = -EFAULT;
1328 if (get_user(p, pages + j + chunk_start))
1329 goto out_pm;
1330 pm[j].addr = (unsigned long) p;
1331
1332 if (get_user(node, nodes + j + chunk_start))
1333 goto out_pm;
1334
1335 err = -ENODEV;
1336 if (node < 0 || node >= MAX_NUMNODES)
1337 goto out_pm;
1338
1339 if (!node_state(node, N_MEMORY))
1340 goto out_pm;
1341
1342 err = -EACCES;
1343 if (!node_isset(node, task_nodes))
1344 goto out_pm;
1345
1346 pm[j].node = node;
1347 }
1348
1349 /* End marker for this chunk */
1350 pm[chunk_nr_pages].node = MAX_NUMNODES;
1351
1352 /* Migrate this chunk */
1353 err = do_move_page_to_node_array(mm, pm,
1354 flags & MPOL_MF_MOVE_ALL);
1355 if (err < 0)
1356 goto out_pm;
1357
1358 /* Return status information */
1359 for (j = 0; j < chunk_nr_pages; j++)
1360 if (put_user(pm[j].status, status + j + chunk_start)) {
1361 err = -EFAULT;
1362 goto out_pm;
1363 }
1364 }
1365 err = 0;
1366
1367out_pm:
1368 free_page((unsigned long)pm);
1369out:
1370 return err;
1371}
1372
1373/*
1374 * Determine the nodes of an array of pages and store it in an array of status.
1375 */
1376static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1377 const void __user **pages, int *status)
1378{
1379 unsigned long i;
1380
1381 down_read(&mm->mmap_sem);
1382
1383 for (i = 0; i < nr_pages; i++) {
1384 unsigned long addr = (unsigned long)(*pages);
1385 struct vm_area_struct *vma;
1386 struct page *page;
1387 int err = -EFAULT;
1388
1389 vma = find_vma(mm, addr);
1390 if (!vma || addr < vma->vm_start)
1391 goto set_status;
1392
1393 page = follow_page(vma, addr, 0);
1394
1395 err = PTR_ERR(page);
1396 if (IS_ERR(page))
1397 goto set_status;
1398
1399 err = -ENOENT;
1400 /* Use PageReserved to check for zero page */
1401 if (!page || PageReserved(page))
1402 goto set_status;
1403
1404 err = page_to_nid(page);
1405set_status:
1406 *status = err;
1407
1408 pages++;
1409 status++;
1410 }
1411
1412 up_read(&mm->mmap_sem);
1413}
1414
1415/*
1416 * Determine the nodes of a user array of pages and store it in
1417 * a user array of status.
1418 */
1419static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1420 const void __user * __user *pages,
1421 int __user *status)
1422{
1423#define DO_PAGES_STAT_CHUNK_NR 16
1424 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1425 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1426
1427 while (nr_pages) {
1428 unsigned long chunk_nr;
1429
1430 chunk_nr = nr_pages;
1431 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1432 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1433
1434 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1435 break;
1436
1437 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1438
1439 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1440 break;
1441
1442 pages += chunk_nr;
1443 status += chunk_nr;
1444 nr_pages -= chunk_nr;
1445 }
1446 return nr_pages ? -EFAULT : 0;
1447}
1448
1449/*
1450 * Move a list of pages in the address space of the currently executing
1451 * process.
1452 */
1453SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1454 const void __user * __user *, pages,
1455 const int __user *, nodes,
1456 int __user *, status, int, flags)
1457{
1458 const struct cred *cred = current_cred(), *tcred;
1459 struct task_struct *task;
1460 struct mm_struct *mm;
1461 int err;
1462 nodemask_t task_nodes;
1463
1464 /* Check flags */
1465 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1466 return -EINVAL;
1467
1468 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1469 return -EPERM;
1470
1471 /* Find the mm_struct */
1472 rcu_read_lock();
1473 task = pid ? find_task_by_vpid(pid) : current;
1474 if (!task) {
1475 rcu_read_unlock();
1476 return -ESRCH;
1477 }
1478 get_task_struct(task);
1479
1480 /*
1481 * Check if this process has the right to modify the specified
1482 * process. The right exists if the process has administrative
1483 * capabilities, superuser privileges or the same
1484 * userid as the target process.
1485 */
1486 tcred = __task_cred(task);
1487 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1488 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1489 !capable(CAP_SYS_NICE)) {
1490 rcu_read_unlock();
1491 err = -EPERM;
1492 goto out;
1493 }
1494 rcu_read_unlock();
1495
1496 err = security_task_movememory(task);
1497 if (err)
1498 goto out;
1499
1500 task_nodes = cpuset_mems_allowed(task);
1501 mm = get_task_mm(task);
1502 put_task_struct(task);
1503
1504 if (!mm)
1505 return -EINVAL;
1506
1507 if (nodes)
1508 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1509 nodes, status, flags);
1510 else
1511 err = do_pages_stat(mm, nr_pages, pages, status);
1512
1513 mmput(mm);
1514 return err;
1515
1516out:
1517 put_task_struct(task);
1518 return err;
1519}
1520
1521/*
1522 * Call migration functions in the vma_ops that may prepare
1523 * memory in a vm for migration. migration functions may perform
1524 * the migration for vmas that do not have an underlying page struct.
1525 */
1526int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1527 const nodemask_t *from, unsigned long flags)
1528{
1529 struct vm_area_struct *vma;
1530 int err = 0;
1531
1532 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1533 if (vma->vm_ops && vma->vm_ops->migrate) {
1534 err = vma->vm_ops->migrate(vma, to, from, flags);
1535 if (err)
1536 break;
1537 }
1538 }
1539 return err;
1540}
1541
1542#ifdef CONFIG_NUMA_BALANCING
1543/*
1544 * Returns true if this is a safe migration target node for misplaced NUMA
1545 * pages. Currently it only checks the watermarks which crude
1546 */
1547static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1548 unsigned long nr_migrate_pages)
1549{
1550 int z;
1551 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1552 struct zone *zone = pgdat->node_zones + z;
1553
1554 if (!populated_zone(zone))
1555 continue;
1556
1557 if (!zone_reclaimable(zone))
1558 continue;
1559
1560 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1561 if (!zone_watermark_ok(zone, 0,
1562 high_wmark_pages(zone) +
1563 nr_migrate_pages,
1564 0, 0))
1565 continue;
1566 return true;
1567 }
1568 return false;
1569}
1570
1571static struct page *alloc_misplaced_dst_page(struct page *page,
1572 unsigned long data,
1573 int **result)
1574{
1575 int nid = (int) data;
1576 struct page *newpage;
1577
1578 newpage = alloc_pages_exact_node(nid,
1579 (GFP_HIGHUSER_MOVABLE |
1580 __GFP_THISNODE | __GFP_NOMEMALLOC |
1581 __GFP_NORETRY | __GFP_NOWARN) &
1582 ~GFP_IOFS, 0);
1583
1584 return newpage;
1585}
1586
1587/*
1588 * page migration rate limiting control.
1589 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1590 * window of time. Default here says do not migrate more than 1280M per second.
1591 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1592 * as it is faults that reset the window, pte updates will happen unconditionally
1593 * if there has not been a fault since @pteupdate_interval_millisecs after the
1594 * throttle window closed.
1595 */
1596static unsigned int migrate_interval_millisecs __read_mostly = 100;
1597static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1598static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1599
1600/* Returns true if NUMA migration is currently rate limited */
1601bool migrate_ratelimited(int node)
1602{
1603 pg_data_t *pgdat = NODE_DATA(node);
1604
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1606 msecs_to_jiffies(pteupdate_interval_millisecs)))
1607 return false;
1608
1609 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1610 return false;
1611
1612 return true;
1613}
1614
1615/* Returns true if the node is migrate rate-limited after the update */
1616static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1617 unsigned long nr_pages)
1618{
1619 /*
1620 * Rate-limit the amount of data that is being migrated to a node.
1621 * Optimal placement is no good if the memory bus is saturated and
1622 * all the time is being spent migrating!
1623 */
1624 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1625 spin_lock(&pgdat->numabalancing_migrate_lock);
1626 pgdat->numabalancing_migrate_nr_pages = 0;
1627 pgdat->numabalancing_migrate_next_window = jiffies +
1628 msecs_to_jiffies(migrate_interval_millisecs);
1629 spin_unlock(&pgdat->numabalancing_migrate_lock);
1630 }
1631 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1632 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1633 nr_pages);
1634 return true;
1635 }
1636
1637 /*
1638 * This is an unlocked non-atomic update so errors are possible.
1639 * The consequences are failing to migrate when we potentiall should
1640 * have which is not severe enough to warrant locking. If it is ever
1641 * a problem, it can be converted to a per-cpu counter.
1642 */
1643 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1644 return false;
1645}
1646
1647static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1648{
1649 int page_lru;
1650
1651 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1652
1653 /* Avoid migrating to a node that is nearly full */
1654 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1655 return 0;
1656
1657 if (isolate_lru_page(page))
1658 return 0;
1659
1660 /*
1661 * migrate_misplaced_transhuge_page() skips page migration's usual
1662 * check on page_count(), so we must do it here, now that the page
1663 * has been isolated: a GUP pin, or any other pin, prevents migration.
1664 * The expected page count is 3: 1 for page's mapcount and 1 for the
1665 * caller's pin and 1 for the reference taken by isolate_lru_page().
1666 */
1667 if (PageTransHuge(page) && page_count(page) != 3) {
1668 putback_lru_page(page);
1669 return 0;
1670 }
1671
1672 page_lru = page_is_file_cache(page);
1673 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1674 hpage_nr_pages(page));
1675
1676 /*
1677 * Isolating the page has taken another reference, so the
1678 * caller's reference can be safely dropped without the page
1679 * disappearing underneath us during migration.
1680 */
1681 put_page(page);
1682 return 1;
1683}
1684
1685bool pmd_trans_migrating(pmd_t pmd)
1686{
1687 struct page *page = pmd_page(pmd);
1688 return PageLocked(page);
1689}
1690
1691void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1692{
1693 struct page *page = pmd_page(*pmd);
1694 wait_on_page_locked(page);
1695}
1696
1697/*
1698 * Attempt to migrate a misplaced page to the specified destination
1699 * node. Caller is expected to have an elevated reference count on
1700 * the page that will be dropped by this function before returning.
1701 */
1702int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1703 int node)
1704{
1705 pg_data_t *pgdat = NODE_DATA(node);
1706 int isolated;
1707 int nr_remaining;
1708 LIST_HEAD(migratepages);
1709
1710 /*
1711 * Don't migrate file pages that are mapped in multiple processes
1712 * with execute permissions as they are probably shared libraries.
1713 */
1714 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1715 (vma->vm_flags & VM_EXEC))
1716 goto out;
1717
1718 /*
1719 * Rate-limit the amount of data that is being migrated to a node.
1720 * Optimal placement is no good if the memory bus is saturated and
1721 * all the time is being spent migrating!
1722 */
1723 if (numamigrate_update_ratelimit(pgdat, 1))
1724 goto out;
1725
1726 isolated = numamigrate_isolate_page(pgdat, page);
1727 if (!isolated)
1728 goto out;
1729
1730 list_add(&page->lru, &migratepages);
1731 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1732 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1733 if (nr_remaining) {
1734 if (!list_empty(&migratepages)) {
1735 list_del(&page->lru);
1736 dec_zone_page_state(page, NR_ISOLATED_ANON +
1737 page_is_file_cache(page));
1738 putback_lru_page(page);
1739 }
1740 isolated = 0;
1741 } else
1742 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1743 BUG_ON(!list_empty(&migratepages));
1744 return isolated;
1745
1746out:
1747 put_page(page);
1748 return 0;
1749}
1750#endif /* CONFIG_NUMA_BALANCING */
1751
1752#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1753/*
1754 * Migrates a THP to a given target node. page must be locked and is unlocked
1755 * before returning.
1756 */
1757int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1758 struct vm_area_struct *vma,
1759 pmd_t *pmd, pmd_t entry,
1760 unsigned long address,
1761 struct page *page, int node)
1762{
1763 spinlock_t *ptl;
1764 pg_data_t *pgdat = NODE_DATA(node);
1765 int isolated = 0;
1766 struct page *new_page = NULL;
1767 struct mem_cgroup *memcg = NULL;
1768 int page_lru = page_is_file_cache(page);
1769 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1770 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1771 pmd_t orig_entry;
1772
1773 /*
1774 * Rate-limit the amount of data that is being migrated to a node.
1775 * Optimal placement is no good if the memory bus is saturated and
1776 * all the time is being spent migrating!
1777 */
1778 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1779 goto out_dropref;
1780
1781 new_page = alloc_pages_node(node,
1782 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1783 HPAGE_PMD_ORDER);
1784 if (!new_page)
1785 goto out_fail;
1786
1787 isolated = numamigrate_isolate_page(pgdat, page);
1788 if (!isolated) {
1789 put_page(new_page);
1790 goto out_fail;
1791 }
1792
1793 if (mm_tlb_flush_pending(mm))
1794 flush_tlb_range(vma, mmun_start, mmun_end);
1795
1796 /* Prepare a page as a migration target */
1797 __set_page_locked(new_page);
1798 SetPageSwapBacked(new_page);
1799
1800 /* anon mapping, we can simply copy page->mapping to the new page: */
1801 new_page->mapping = page->mapping;
1802 new_page->index = page->index;
1803 migrate_page_copy(new_page, page);
1804 WARN_ON(PageLRU(new_page));
1805
1806 /* Recheck the target PMD */
1807 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1808 ptl = pmd_lock(mm, pmd);
1809 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1810fail_putback:
1811 spin_unlock(ptl);
1812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1813
1814 /* Reverse changes made by migrate_page_copy() */
1815 if (TestClearPageActive(new_page))
1816 SetPageActive(page);
1817 if (TestClearPageUnevictable(new_page))
1818 SetPageUnevictable(page);
1819 mlock_migrate_page(page, new_page);
1820
1821 unlock_page(new_page);
1822 put_page(new_page); /* Free it */
1823
1824 /* Retake the callers reference and putback on LRU */
1825 get_page(page);
1826 putback_lru_page(page);
1827 mod_zone_page_state(page_zone(page),
1828 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1829
1830 goto out_unlock;
1831 }
1832
1833 /*
1834 * Traditional migration needs to prepare the memcg charge
1835 * transaction early to prevent the old page from being
1836 * uncharged when installing migration entries. Here we can
1837 * save the potential rollback and start the charge transfer
1838 * only when migration is already known to end successfully.
1839 */
1840 mem_cgroup_prepare_migration(page, new_page, &memcg);
1841
1842 orig_entry = *pmd;
1843 entry = mk_pmd(new_page, vma->vm_page_prot);
1844 entry = pmd_mkhuge(entry);
1845 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1846
1847 /*
1848 * Clear the old entry under pagetable lock and establish the new PTE.
1849 * Any parallel GUP will either observe the old page blocking on the
1850 * page lock, block on the page table lock or observe the new page.
1851 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1852 * guarantee the copy is visible before the pagetable update.
1853 */
1854 flush_cache_range(vma, mmun_start, mmun_end);
1855 page_add_new_anon_rmap(new_page, vma, mmun_start);
1856 pmdp_clear_flush(vma, mmun_start, pmd);
1857 set_pmd_at(mm, mmun_start, pmd, entry);
1858 flush_tlb_range(vma, mmun_start, mmun_end);
1859 update_mmu_cache_pmd(vma, address, &entry);
1860
1861 if (page_count(page) != 2) {
1862 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1863 flush_tlb_range(vma, mmun_start, mmun_end);
1864 update_mmu_cache_pmd(vma, address, &entry);
1865 page_remove_rmap(new_page);
1866 goto fail_putback;
1867 }
1868
1869 page_remove_rmap(page);
1870
1871 /*
1872 * Finish the charge transaction under the page table lock to
1873 * prevent split_huge_page() from dividing up the charge
1874 * before it's fully transferred to the new page.
1875 */
1876 mem_cgroup_end_migration(memcg, page, new_page, true);
1877 spin_unlock(ptl);
1878 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1879
1880 unlock_page(new_page);
1881 unlock_page(page);
1882 put_page(page); /* Drop the rmap reference */
1883 put_page(page); /* Drop the LRU isolation reference */
1884
1885 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1886 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1887
1888 mod_zone_page_state(page_zone(page),
1889 NR_ISOLATED_ANON + page_lru,
1890 -HPAGE_PMD_NR);
1891 return isolated;
1892
1893out_fail:
1894 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1895out_dropref:
1896 ptl = pmd_lock(mm, pmd);
1897 if (pmd_same(*pmd, entry)) {
1898 entry = pmd_mknonnuma(entry);
1899 set_pmd_at(mm, mmun_start, pmd, entry);
1900 update_mmu_cache_pmd(vma, address, &entry);
1901 }
1902 spin_unlock(ptl);
1903
1904out_unlock:
1905 unlock_page(page);
1906 put_page(page);
1907 return 0;
1908}
1909#endif /* CONFIG_NUMA_BALANCING */
1910
1911#endif /* CONFIG_NUMA */