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/migrate.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/backing-dev.h>
34#include <linux/compaction.h>
35#include <linux/syscalls.h>
36#include <linux/hugetlb.h>
37#include <linux/hugetlb_cgroup.h>
38#include <linux/gfp.h>
39#include <linux/balloon_compaction.h>
40#include <linux/mmu_notifier.h>
41#include <linux/page_idle.h>
42#include <linux/page_owner.h>
43
44#include <asm/tlbflush.h>
45
46#define CREATE_TRACE_POINTS
47#include <trace/events/migrate.h>
48
49#include "internal.h"
50
51/*
52 * migrate_prep() needs to be called before we start compiling a list of pages
53 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
54 * undesirable, use migrate_prep_local()
55 */
56int migrate_prep(void)
57{
58 /*
59 * Clear the LRU lists so pages can be isolated.
60 * Note that pages may be moved off the LRU after we have
61 * drained them. Those pages will fail to migrate like other
62 * pages that may be busy.
63 */
64 lru_add_drain_all();
65
66 return 0;
67}
68
69/* Do the necessary work of migrate_prep but not if it involves other CPUs */
70int migrate_prep_local(void)
71{
72 lru_add_drain();
73
74 return 0;
75}
76
77bool isolate_movable_page(struct page *page, isolate_mode_t mode)
78{
79 struct address_space *mapping;
80
81 /*
82 * Avoid burning cycles with pages that are yet under __free_pages(),
83 * or just got freed under us.
84 *
85 * In case we 'win' a race for a movable page being freed under us and
86 * raise its refcount preventing __free_pages() from doing its job
87 * the put_page() at the end of this block will take care of
88 * release this page, thus avoiding a nasty leakage.
89 */
90 if (unlikely(!get_page_unless_zero(page)))
91 goto out;
92
93 /*
94 * Check PageMovable before holding a PG_lock because page's owner
95 * assumes anybody doesn't touch PG_lock of newly allocated page
96 * so unconditionally grapping the lock ruins page's owner side.
97 */
98 if (unlikely(!__PageMovable(page)))
99 goto out_putpage;
100 /*
101 * As movable pages are not isolated from LRU lists, concurrent
102 * compaction threads can race against page migration functions
103 * as well as race against the releasing a page.
104 *
105 * In order to avoid having an already isolated movable page
106 * being (wrongly) re-isolated while it is under migration,
107 * or to avoid attempting to isolate pages being released,
108 * lets be sure we have the page lock
109 * before proceeding with the movable page isolation steps.
110 */
111 if (unlikely(!trylock_page(page)))
112 goto out_putpage;
113
114 if (!PageMovable(page) || PageIsolated(page))
115 goto out_no_isolated;
116
117 mapping = page_mapping(page);
118 VM_BUG_ON_PAGE(!mapping, page);
119
120 if (!mapping->a_ops->isolate_page(page, mode))
121 goto out_no_isolated;
122
123 /* Driver shouldn't use PG_isolated bit of page->flags */
124 WARN_ON_ONCE(PageIsolated(page));
125 __SetPageIsolated(page);
126 unlock_page(page);
127
128 return true;
129
130out_no_isolated:
131 unlock_page(page);
132out_putpage:
133 put_page(page);
134out:
135 return false;
136}
137
138/* It should be called on page which is PG_movable */
139void putback_movable_page(struct page *page)
140{
141 struct address_space *mapping;
142
143 VM_BUG_ON_PAGE(!PageLocked(page), page);
144 VM_BUG_ON_PAGE(!PageMovable(page), page);
145 VM_BUG_ON_PAGE(!PageIsolated(page), page);
146
147 mapping = page_mapping(page);
148 mapping->a_ops->putback_page(page);
149 __ClearPageIsolated(page);
150}
151
152/*
153 * Put previously isolated pages back onto the appropriate lists
154 * from where they were once taken off for compaction/migration.
155 *
156 * This function shall be used whenever the isolated pageset has been
157 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
158 * and isolate_huge_page().
159 */
160void putback_movable_pages(struct list_head *l)
161{
162 struct page *page;
163 struct page *page2;
164
165 list_for_each_entry_safe(page, page2, l, lru) {
166 if (unlikely(PageHuge(page))) {
167 putback_active_hugepage(page);
168 continue;
169 }
170 list_del(&page->lru);
171 /*
172 * We isolated non-lru movable page so here we can use
173 * __PageMovable because LRU page's mapping cannot have
174 * PAGE_MAPPING_MOVABLE.
175 */
176 if (unlikely(__PageMovable(page))) {
177 VM_BUG_ON_PAGE(!PageIsolated(page), page);
178 lock_page(page);
179 if (PageMovable(page))
180 putback_movable_page(page);
181 else
182 __ClearPageIsolated(page);
183 unlock_page(page);
184 put_page(page);
185 } else {
186 putback_lru_page(page);
187 dec_node_page_state(page, NR_ISOLATED_ANON +
188 page_is_file_cache(page));
189 }
190 }
191}
192
193/*
194 * Restore a potential migration pte to a working pte entry
195 */
196static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
197 unsigned long addr, void *old)
198{
199 struct mm_struct *mm = vma->vm_mm;
200 swp_entry_t entry;
201 pmd_t *pmd;
202 pte_t *ptep, pte;
203 spinlock_t *ptl;
204
205 if (unlikely(PageHuge(new))) {
206 ptep = huge_pte_offset(mm, addr);
207 if (!ptep)
208 goto out;
209 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
210 } else {
211 pmd = mm_find_pmd(mm, addr);
212 if (!pmd)
213 goto out;
214
215 ptep = pte_offset_map(pmd, addr);
216
217 /*
218 * Peek to check is_swap_pte() before taking ptlock? No, we
219 * can race mremap's move_ptes(), which skips anon_vma lock.
220 */
221
222 ptl = pte_lockptr(mm, pmd);
223 }
224
225 spin_lock(ptl);
226 pte = *ptep;
227 if (!is_swap_pte(pte))
228 goto unlock;
229
230 entry = pte_to_swp_entry(pte);
231
232 if (!is_migration_entry(entry) ||
233 migration_entry_to_page(entry) != old)
234 goto unlock;
235
236 get_page(new);
237 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
238 if (pte_swp_soft_dirty(*ptep))
239 pte = pte_mksoft_dirty(pte);
240
241 /* Recheck VMA as permissions can change since migration started */
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
244
245#ifdef CONFIG_HUGETLB_PAGE
246 if (PageHuge(new)) {
247 pte = pte_mkhuge(pte);
248 pte = arch_make_huge_pte(pte, vma, new, 0);
249 }
250#endif
251 flush_dcache_page(new);
252 set_pte_at(mm, addr, ptep, pte);
253
254 if (PageHuge(new)) {
255 if (PageAnon(new))
256 hugepage_add_anon_rmap(new, vma, addr);
257 else
258 page_dup_rmap(new, true);
259 } else if (PageAnon(new))
260 page_add_anon_rmap(new, vma, addr, false);
261 else
262 page_add_file_rmap(new, false);
263
264 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
265 mlock_vma_page(new);
266
267 /* No need to invalidate - it was non-present before */
268 update_mmu_cache(vma, addr, ptep);
269unlock:
270 pte_unmap_unlock(ptep, ptl);
271out:
272 return SWAP_AGAIN;
273}
274
275/*
276 * Get rid of all migration entries and replace them by
277 * references to the indicated page.
278 */
279void remove_migration_ptes(struct page *old, struct page *new, bool locked)
280{
281 struct rmap_walk_control rwc = {
282 .rmap_one = remove_migration_pte,
283 .arg = old,
284 };
285
286 if (locked)
287 rmap_walk_locked(new, &rwc);
288 else
289 rmap_walk(new, &rwc);
290}
291
292/*
293 * Something used the pte of a page under migration. We need to
294 * get to the page and wait until migration is finished.
295 * When we return from this function the fault will be retried.
296 */
297void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
298 spinlock_t *ptl)
299{
300 pte_t pte;
301 swp_entry_t entry;
302 struct page *page;
303
304 spin_lock(ptl);
305 pte = *ptep;
306 if (!is_swap_pte(pte))
307 goto out;
308
309 entry = pte_to_swp_entry(pte);
310 if (!is_migration_entry(entry))
311 goto out;
312
313 page = migration_entry_to_page(entry);
314
315 /*
316 * Once radix-tree replacement of page migration started, page_count
317 * *must* be zero. And, we don't want to call wait_on_page_locked()
318 * against a page without get_page().
319 * So, we use get_page_unless_zero(), here. Even failed, page fault
320 * will occur again.
321 */
322 if (!get_page_unless_zero(page))
323 goto out;
324 pte_unmap_unlock(ptep, ptl);
325 wait_on_page_locked(page);
326 put_page(page);
327 return;
328out:
329 pte_unmap_unlock(ptep, ptl);
330}
331
332void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
333 unsigned long address)
334{
335 spinlock_t *ptl = pte_lockptr(mm, pmd);
336 pte_t *ptep = pte_offset_map(pmd, address);
337 __migration_entry_wait(mm, ptep, ptl);
338}
339
340void migration_entry_wait_huge(struct vm_area_struct *vma,
341 struct mm_struct *mm, pte_t *pte)
342{
343 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
344 __migration_entry_wait(mm, pte, ptl);
345}
346
347#ifdef CONFIG_BLOCK
348/* Returns true if all buffers are successfully locked */
349static bool buffer_migrate_lock_buffers(struct buffer_head *head,
350 enum migrate_mode mode)
351{
352 struct buffer_head *bh = head;
353
354 /* Simple case, sync compaction */
355 if (mode != MIGRATE_ASYNC) {
356 do {
357 get_bh(bh);
358 lock_buffer(bh);
359 bh = bh->b_this_page;
360
361 } while (bh != head);
362
363 return true;
364 }
365
366 /* async case, we cannot block on lock_buffer so use trylock_buffer */
367 do {
368 get_bh(bh);
369 if (!trylock_buffer(bh)) {
370 /*
371 * We failed to lock the buffer and cannot stall in
372 * async migration. Release the taken locks
373 */
374 struct buffer_head *failed_bh = bh;
375 put_bh(failed_bh);
376 bh = head;
377 while (bh != failed_bh) {
378 unlock_buffer(bh);
379 put_bh(bh);
380 bh = bh->b_this_page;
381 }
382 return false;
383 }
384
385 bh = bh->b_this_page;
386 } while (bh != head);
387 return true;
388}
389#else
390static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
391 enum migrate_mode mode)
392{
393 return true;
394}
395#endif /* CONFIG_BLOCK */
396
397/*
398 * Replace the page in the mapping.
399 *
400 * The number of remaining references must be:
401 * 1 for anonymous pages without a mapping
402 * 2 for pages with a mapping
403 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
404 */
405int migrate_page_move_mapping(struct address_space *mapping,
406 struct page *newpage, struct page *page,
407 struct buffer_head *head, enum migrate_mode mode,
408 int extra_count)
409{
410 struct zone *oldzone, *newzone;
411 int dirty;
412 int expected_count = 1 + extra_count;
413 void **pslot;
414
415 if (!mapping) {
416 /* Anonymous page without mapping */
417 if (page_count(page) != expected_count)
418 return -EAGAIN;
419
420 /* No turning back from here */
421 newpage->index = page->index;
422 newpage->mapping = page->mapping;
423 if (PageSwapBacked(page))
424 __SetPageSwapBacked(newpage);
425
426 return MIGRATEPAGE_SUCCESS;
427 }
428
429 oldzone = page_zone(page);
430 newzone = page_zone(newpage);
431
432 spin_lock_irq(&mapping->tree_lock);
433
434 pslot = radix_tree_lookup_slot(&mapping->page_tree,
435 page_index(page));
436
437 expected_count += 1 + page_has_private(page);
438 if (page_count(page) != expected_count ||
439 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
440 spin_unlock_irq(&mapping->tree_lock);
441 return -EAGAIN;
442 }
443
444 if (!page_ref_freeze(page, expected_count)) {
445 spin_unlock_irq(&mapping->tree_lock);
446 return -EAGAIN;
447 }
448
449 /*
450 * In the async migration case of moving a page with buffers, lock the
451 * buffers using trylock before the mapping is moved. If the mapping
452 * was moved, we later failed to lock the buffers and could not move
453 * the mapping back due to an elevated page count, we would have to
454 * block waiting on other references to be dropped.
455 */
456 if (mode == MIGRATE_ASYNC && head &&
457 !buffer_migrate_lock_buffers(head, mode)) {
458 page_ref_unfreeze(page, expected_count);
459 spin_unlock_irq(&mapping->tree_lock);
460 return -EAGAIN;
461 }
462
463 /*
464 * Now we know that no one else is looking at the page:
465 * no turning back from here.
466 */
467 newpage->index = page->index;
468 newpage->mapping = page->mapping;
469 get_page(newpage); /* add cache reference */
470 if (PageSwapBacked(page)) {
471 __SetPageSwapBacked(newpage);
472 if (PageSwapCache(page)) {
473 SetPageSwapCache(newpage);
474 set_page_private(newpage, page_private(page));
475 }
476 } else {
477 VM_BUG_ON_PAGE(PageSwapCache(page), page);
478 }
479
480 /* Move dirty while page refs frozen and newpage not yet exposed */
481 dirty = PageDirty(page);
482 if (dirty) {
483 ClearPageDirty(page);
484 SetPageDirty(newpage);
485 }
486
487 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
488
489 /*
490 * Drop cache reference from old page by unfreezing
491 * to one less reference.
492 * We know this isn't the last reference.
493 */
494 page_ref_unfreeze(page, expected_count - 1);
495
496 spin_unlock(&mapping->tree_lock);
497 /* Leave irq disabled to prevent preemption while updating stats */
498
499 /*
500 * If moved to a different zone then also account
501 * the page for that zone. Other VM counters will be
502 * taken care of when we establish references to the
503 * new page and drop references to the old page.
504 *
505 * Note that anonymous pages are accounted for
506 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
507 * are mapped to swap space.
508 */
509 if (newzone != oldzone) {
510 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
511 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
512 if (PageSwapBacked(page) && !PageSwapCache(page)) {
513 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
514 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
515 }
516 if (dirty && mapping_cap_account_dirty(mapping)) {
517 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
518 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
519 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
520 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
521 }
522 }
523 local_irq_enable();
524
525 return MIGRATEPAGE_SUCCESS;
526}
527EXPORT_SYMBOL(migrate_page_move_mapping);
528
529/*
530 * The expected number of remaining references is the same as that
531 * of migrate_page_move_mapping().
532 */
533int migrate_huge_page_move_mapping(struct address_space *mapping,
534 struct page *newpage, struct page *page)
535{
536 int expected_count;
537 void **pslot;
538
539 spin_lock_irq(&mapping->tree_lock);
540
541 pslot = radix_tree_lookup_slot(&mapping->page_tree,
542 page_index(page));
543
544 expected_count = 2 + page_has_private(page);
545 if (page_count(page) != expected_count ||
546 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
547 spin_unlock_irq(&mapping->tree_lock);
548 return -EAGAIN;
549 }
550
551 if (!page_ref_freeze(page, expected_count)) {
552 spin_unlock_irq(&mapping->tree_lock);
553 return -EAGAIN;
554 }
555
556 newpage->index = page->index;
557 newpage->mapping = page->mapping;
558
559 get_page(newpage);
560
561 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
562
563 page_ref_unfreeze(page, expected_count - 1);
564
565 spin_unlock_irq(&mapping->tree_lock);
566
567 return MIGRATEPAGE_SUCCESS;
568}
569
570/*
571 * Gigantic pages are so large that we do not guarantee that page++ pointer
572 * arithmetic will work across the entire page. We need something more
573 * specialized.
574 */
575static void __copy_gigantic_page(struct page *dst, struct page *src,
576 int nr_pages)
577{
578 int i;
579 struct page *dst_base = dst;
580 struct page *src_base = src;
581
582 for (i = 0; i < nr_pages; ) {
583 cond_resched();
584 copy_highpage(dst, src);
585
586 i++;
587 dst = mem_map_next(dst, dst_base, i);
588 src = mem_map_next(src, src_base, i);
589 }
590}
591
592static void copy_huge_page(struct page *dst, struct page *src)
593{
594 int i;
595 int nr_pages;
596
597 if (PageHuge(src)) {
598 /* hugetlbfs page */
599 struct hstate *h = page_hstate(src);
600 nr_pages = pages_per_huge_page(h);
601
602 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
603 __copy_gigantic_page(dst, src, nr_pages);
604 return;
605 }
606 } else {
607 /* thp page */
608 BUG_ON(!PageTransHuge(src));
609 nr_pages = hpage_nr_pages(src);
610 }
611
612 for (i = 0; i < nr_pages; i++) {
613 cond_resched();
614 copy_highpage(dst + i, src + i);
615 }
616}
617
618/*
619 * Copy the page to its new location
620 */
621void migrate_page_copy(struct page *newpage, struct page *page)
622{
623 int cpupid;
624
625 if (PageHuge(page) || PageTransHuge(page))
626 copy_huge_page(newpage, page);
627 else
628 copy_highpage(newpage, page);
629
630 if (PageError(page))
631 SetPageError(newpage);
632 if (PageReferenced(page))
633 SetPageReferenced(newpage);
634 if (PageUptodate(page))
635 SetPageUptodate(newpage);
636 if (TestClearPageActive(page)) {
637 VM_BUG_ON_PAGE(PageUnevictable(page), page);
638 SetPageActive(newpage);
639 } else if (TestClearPageUnevictable(page))
640 SetPageUnevictable(newpage);
641 if (PageChecked(page))
642 SetPageChecked(newpage);
643 if (PageMappedToDisk(page))
644 SetPageMappedToDisk(newpage);
645
646 /* Move dirty on pages not done by migrate_page_move_mapping() */
647 if (PageDirty(page))
648 SetPageDirty(newpage);
649
650 if (page_is_young(page))
651 set_page_young(newpage);
652 if (page_is_idle(page))
653 set_page_idle(newpage);
654
655 /*
656 * Copy NUMA information to the new page, to prevent over-eager
657 * future migrations of this same page.
658 */
659 cpupid = page_cpupid_xchg_last(page, -1);
660 page_cpupid_xchg_last(newpage, cpupid);
661
662 ksm_migrate_page(newpage, page);
663 /*
664 * Please do not reorder this without considering how mm/ksm.c's
665 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
666 */
667 if (PageSwapCache(page))
668 ClearPageSwapCache(page);
669 ClearPagePrivate(page);
670 set_page_private(page, 0);
671
672 /*
673 * If any waiters have accumulated on the new page then
674 * wake them up.
675 */
676 if (PageWriteback(newpage))
677 end_page_writeback(newpage);
678
679 copy_page_owner(page, newpage);
680
681 mem_cgroup_migrate(page, newpage);
682}
683EXPORT_SYMBOL(migrate_page_copy);
684
685/************************************************************
686 * Migration functions
687 ***********************************************************/
688
689/*
690 * Common logic to directly migrate a single LRU page suitable for
691 * pages that do not use PagePrivate/PagePrivate2.
692 *
693 * Pages are locked upon entry and exit.
694 */
695int migrate_page(struct address_space *mapping,
696 struct page *newpage, struct page *page,
697 enum migrate_mode mode)
698{
699 int rc;
700
701 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
702
703 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
704
705 if (rc != MIGRATEPAGE_SUCCESS)
706 return rc;
707
708 migrate_page_copy(newpage, page);
709 return MIGRATEPAGE_SUCCESS;
710}
711EXPORT_SYMBOL(migrate_page);
712
713#ifdef CONFIG_BLOCK
714/*
715 * Migration function for pages with buffers. This function can only be used
716 * if the underlying filesystem guarantees that no other references to "page"
717 * exist.
718 */
719int buffer_migrate_page(struct address_space *mapping,
720 struct page *newpage, struct page *page, enum migrate_mode mode)
721{
722 struct buffer_head *bh, *head;
723 int rc;
724
725 if (!page_has_buffers(page))
726 return migrate_page(mapping, newpage, page, mode);
727
728 head = page_buffers(page);
729
730 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
731
732 if (rc != MIGRATEPAGE_SUCCESS)
733 return rc;
734
735 /*
736 * In the async case, migrate_page_move_mapping locked the buffers
737 * with an IRQ-safe spinlock held. In the sync case, the buffers
738 * need to be locked now
739 */
740 if (mode != MIGRATE_ASYNC)
741 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
742
743 ClearPagePrivate(page);
744 set_page_private(newpage, page_private(page));
745 set_page_private(page, 0);
746 put_page(page);
747 get_page(newpage);
748
749 bh = head;
750 do {
751 set_bh_page(bh, newpage, bh_offset(bh));
752 bh = bh->b_this_page;
753
754 } while (bh != head);
755
756 SetPagePrivate(newpage);
757
758 migrate_page_copy(newpage, page);
759
760 bh = head;
761 do {
762 unlock_buffer(bh);
763 put_bh(bh);
764 bh = bh->b_this_page;
765
766 } while (bh != head);
767
768 return MIGRATEPAGE_SUCCESS;
769}
770EXPORT_SYMBOL(buffer_migrate_page);
771#endif
772
773/*
774 * Writeback a page to clean the dirty state
775 */
776static int writeout(struct address_space *mapping, struct page *page)
777{
778 struct writeback_control wbc = {
779 .sync_mode = WB_SYNC_NONE,
780 .nr_to_write = 1,
781 .range_start = 0,
782 .range_end = LLONG_MAX,
783 .for_reclaim = 1
784 };
785 int rc;
786
787 if (!mapping->a_ops->writepage)
788 /* No write method for the address space */
789 return -EINVAL;
790
791 if (!clear_page_dirty_for_io(page))
792 /* Someone else already triggered a write */
793 return -EAGAIN;
794
795 /*
796 * A dirty page may imply that the underlying filesystem has
797 * the page on some queue. So the page must be clean for
798 * migration. Writeout may mean we loose the lock and the
799 * page state is no longer what we checked for earlier.
800 * At this point we know that the migration attempt cannot
801 * be successful.
802 */
803 remove_migration_ptes(page, page, false);
804
805 rc = mapping->a_ops->writepage(page, &wbc);
806
807 if (rc != AOP_WRITEPAGE_ACTIVATE)
808 /* unlocked. Relock */
809 lock_page(page);
810
811 return (rc < 0) ? -EIO : -EAGAIN;
812}
813
814/*
815 * Default handling if a filesystem does not provide a migration function.
816 */
817static int fallback_migrate_page(struct address_space *mapping,
818 struct page *newpage, struct page *page, enum migrate_mode mode)
819{
820 if (PageDirty(page)) {
821 /* Only writeback pages in full synchronous migration */
822 if (mode != MIGRATE_SYNC)
823 return -EBUSY;
824 return writeout(mapping, page);
825 }
826
827 /*
828 * Buffers may be managed in a filesystem specific way.
829 * We must have no buffers or drop them.
830 */
831 if (page_has_private(page) &&
832 !try_to_release_page(page, GFP_KERNEL))
833 return -EAGAIN;
834
835 return migrate_page(mapping, newpage, page, mode);
836}
837
838/*
839 * Move a page to a newly allocated page
840 * The page is locked and all ptes have been successfully removed.
841 *
842 * The new page will have replaced the old page if this function
843 * is successful.
844 *
845 * Return value:
846 * < 0 - error code
847 * MIGRATEPAGE_SUCCESS - success
848 */
849static int move_to_new_page(struct page *newpage, struct page *page,
850 enum migrate_mode mode)
851{
852 struct address_space *mapping;
853 int rc = -EAGAIN;
854 bool is_lru = !__PageMovable(page);
855
856 VM_BUG_ON_PAGE(!PageLocked(page), page);
857 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
858
859 mapping = page_mapping(page);
860
861 if (likely(is_lru)) {
862 if (!mapping)
863 rc = migrate_page(mapping, newpage, page, mode);
864 else if (mapping->a_ops->migratepage)
865 /*
866 * Most pages have a mapping and most filesystems
867 * provide a migratepage callback. Anonymous pages
868 * are part of swap space which also has its own
869 * migratepage callback. This is the most common path
870 * for page migration.
871 */
872 rc = mapping->a_ops->migratepage(mapping, newpage,
873 page, mode);
874 else
875 rc = fallback_migrate_page(mapping, newpage,
876 page, mode);
877 } else {
878 /*
879 * In case of non-lru page, it could be released after
880 * isolation step. In that case, we shouldn't try migration.
881 */
882 VM_BUG_ON_PAGE(!PageIsolated(page), page);
883 if (!PageMovable(page)) {
884 rc = MIGRATEPAGE_SUCCESS;
885 __ClearPageIsolated(page);
886 goto out;
887 }
888
889 rc = mapping->a_ops->migratepage(mapping, newpage,
890 page, mode);
891 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
892 !PageIsolated(page));
893 }
894
895 /*
896 * When successful, old pagecache page->mapping must be cleared before
897 * page is freed; but stats require that PageAnon be left as PageAnon.
898 */
899 if (rc == MIGRATEPAGE_SUCCESS) {
900 if (__PageMovable(page)) {
901 VM_BUG_ON_PAGE(!PageIsolated(page), page);
902
903 /*
904 * We clear PG_movable under page_lock so any compactor
905 * cannot try to migrate this page.
906 */
907 __ClearPageIsolated(page);
908 }
909
910 /*
911 * Anonymous and movable page->mapping will be cleard by
912 * free_pages_prepare so don't reset it here for keeping
913 * the type to work PageAnon, for example.
914 */
915 if (!PageMappingFlags(page))
916 page->mapping = NULL;
917 }
918out:
919 return rc;
920}
921
922static int __unmap_and_move(struct page *page, struct page *newpage,
923 int force, enum migrate_mode mode)
924{
925 int rc = -EAGAIN;
926 int page_was_mapped = 0;
927 struct anon_vma *anon_vma = NULL;
928 bool is_lru = !__PageMovable(page);
929
930 if (!trylock_page(page)) {
931 if (!force || mode == MIGRATE_ASYNC)
932 goto out;
933
934 /*
935 * It's not safe for direct compaction to call lock_page.
936 * For example, during page readahead pages are added locked
937 * to the LRU. Later, when the IO completes the pages are
938 * marked uptodate and unlocked. However, the queueing
939 * could be merging multiple pages for one bio (e.g.
940 * mpage_readpages). If an allocation happens for the
941 * second or third page, the process can end up locking
942 * the same page twice and deadlocking. Rather than
943 * trying to be clever about what pages can be locked,
944 * avoid the use of lock_page for direct compaction
945 * altogether.
946 */
947 if (current->flags & PF_MEMALLOC)
948 goto out;
949
950 lock_page(page);
951 }
952
953 if (PageWriteback(page)) {
954 /*
955 * Only in the case of a full synchronous migration is it
956 * necessary to wait for PageWriteback. In the async case,
957 * the retry loop is too short and in the sync-light case,
958 * the overhead of stalling is too much
959 */
960 if (mode != MIGRATE_SYNC) {
961 rc = -EBUSY;
962 goto out_unlock;
963 }
964 if (!force)
965 goto out_unlock;
966 wait_on_page_writeback(page);
967 }
968
969 /*
970 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
971 * we cannot notice that anon_vma is freed while we migrates a page.
972 * This get_anon_vma() delays freeing anon_vma pointer until the end
973 * of migration. File cache pages are no problem because of page_lock()
974 * File Caches may use write_page() or lock_page() in migration, then,
975 * just care Anon page here.
976 *
977 * Only page_get_anon_vma() understands the subtleties of
978 * getting a hold on an anon_vma from outside one of its mms.
979 * But if we cannot get anon_vma, then we won't need it anyway,
980 * because that implies that the anon page is no longer mapped
981 * (and cannot be remapped so long as we hold the page lock).
982 */
983 if (PageAnon(page) && !PageKsm(page))
984 anon_vma = page_get_anon_vma(page);
985
986 /*
987 * Block others from accessing the new page when we get around to
988 * establishing additional references. We are usually the only one
989 * holding a reference to newpage at this point. We used to have a BUG
990 * here if trylock_page(newpage) fails, but would like to allow for
991 * cases where there might be a race with the previous use of newpage.
992 * This is much like races on refcount of oldpage: just don't BUG().
993 */
994 if (unlikely(!trylock_page(newpage)))
995 goto out_unlock;
996
997 if (unlikely(!is_lru)) {
998 rc = move_to_new_page(newpage, page, mode);
999 goto out_unlock_both;
1000 }
1001
1002 /*
1003 * Corner case handling:
1004 * 1. When a new swap-cache page is read into, it is added to the LRU
1005 * and treated as swapcache but it has no rmap yet.
1006 * Calling try_to_unmap() against a page->mapping==NULL page will
1007 * trigger a BUG. So handle it here.
1008 * 2. An orphaned page (see truncate_complete_page) might have
1009 * fs-private metadata. The page can be picked up due to memory
1010 * offlining. Everywhere else except page reclaim, the page is
1011 * invisible to the vm, so the page can not be migrated. So try to
1012 * free the metadata, so the page can be freed.
1013 */
1014 if (!page->mapping) {
1015 VM_BUG_ON_PAGE(PageAnon(page), page);
1016 if (page_has_private(page)) {
1017 try_to_free_buffers(page);
1018 goto out_unlock_both;
1019 }
1020 } else if (page_mapped(page)) {
1021 /* Establish migration ptes */
1022 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1023 page);
1024 try_to_unmap(page,
1025 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1026 page_was_mapped = 1;
1027 }
1028
1029 if (!page_mapped(page))
1030 rc = move_to_new_page(newpage, page, mode);
1031
1032 if (page_was_mapped)
1033 remove_migration_ptes(page,
1034 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1035
1036out_unlock_both:
1037 unlock_page(newpage);
1038out_unlock:
1039 /* Drop an anon_vma reference if we took one */
1040 if (anon_vma)
1041 put_anon_vma(anon_vma);
1042 unlock_page(page);
1043out:
1044 /*
1045 * If migration is successful, decrease refcount of the newpage
1046 * which will not free the page because new page owner increased
1047 * refcounter. As well, if it is LRU page, add the page to LRU
1048 * list in here.
1049 */
1050 if (rc == MIGRATEPAGE_SUCCESS) {
1051 if (unlikely(__PageMovable(newpage)))
1052 put_page(newpage);
1053 else
1054 putback_lru_page(newpage);
1055 }
1056
1057 return rc;
1058}
1059
1060/*
1061 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1062 * around it.
1063 */
1064#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1065#define ICE_noinline noinline
1066#else
1067#define ICE_noinline
1068#endif
1069
1070/*
1071 * Obtain the lock on page, remove all ptes and migrate the page
1072 * to the newly allocated page in newpage.
1073 */
1074static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1075 free_page_t put_new_page,
1076 unsigned long private, struct page *page,
1077 int force, enum migrate_mode mode,
1078 enum migrate_reason reason)
1079{
1080 int rc = MIGRATEPAGE_SUCCESS;
1081 int *result = NULL;
1082 struct page *newpage;
1083
1084 newpage = get_new_page(page, private, &result);
1085 if (!newpage)
1086 return -ENOMEM;
1087
1088 if (page_count(page) == 1) {
1089 /* page was freed from under us. So we are done. */
1090 ClearPageActive(page);
1091 ClearPageUnevictable(page);
1092 if (unlikely(__PageMovable(page))) {
1093 lock_page(page);
1094 if (!PageMovable(page))
1095 __ClearPageIsolated(page);
1096 unlock_page(page);
1097 }
1098 if (put_new_page)
1099 put_new_page(newpage, private);
1100 else
1101 put_page(newpage);
1102 goto out;
1103 }
1104
1105 if (unlikely(PageTransHuge(page))) {
1106 lock_page(page);
1107 rc = split_huge_page(page);
1108 unlock_page(page);
1109 if (rc)
1110 goto out;
1111 }
1112
1113 rc = __unmap_and_move(page, newpage, force, mode);
1114 if (rc == MIGRATEPAGE_SUCCESS)
1115 set_page_owner_migrate_reason(newpage, reason);
1116
1117out:
1118 if (rc != -EAGAIN) {
1119 /*
1120 * A page that has been migrated has all references
1121 * removed and will be freed. A page that has not been
1122 * migrated will have kepts its references and be
1123 * restored.
1124 */
1125 list_del(&page->lru);
1126
1127 /*
1128 * Compaction can migrate also non-LRU pages which are
1129 * not accounted to NR_ISOLATED_*. They can be recognized
1130 * as __PageMovable
1131 */
1132 if (likely(!__PageMovable(page)))
1133 dec_node_page_state(page, NR_ISOLATED_ANON +
1134 page_is_file_cache(page));
1135 }
1136
1137 /*
1138 * If migration is successful, releases reference grabbed during
1139 * isolation. Otherwise, restore the page to right list unless
1140 * we want to retry.
1141 */
1142 if (rc == MIGRATEPAGE_SUCCESS) {
1143 put_page(page);
1144 if (reason == MR_MEMORY_FAILURE) {
1145 /*
1146 * Set PG_HWPoison on just freed page
1147 * intentionally. Although it's rather weird,
1148 * it's how HWPoison flag works at the moment.
1149 */
1150 if (!test_set_page_hwpoison(page))
1151 num_poisoned_pages_inc();
1152 }
1153 } else {
1154 if (rc != -EAGAIN) {
1155 if (likely(!__PageMovable(page))) {
1156 putback_lru_page(page);
1157 goto put_new;
1158 }
1159
1160 lock_page(page);
1161 if (PageMovable(page))
1162 putback_movable_page(page);
1163 else
1164 __ClearPageIsolated(page);
1165 unlock_page(page);
1166 put_page(page);
1167 }
1168put_new:
1169 if (put_new_page)
1170 put_new_page(newpage, private);
1171 else
1172 put_page(newpage);
1173 }
1174
1175 if (result) {
1176 if (rc)
1177 *result = rc;
1178 else
1179 *result = page_to_nid(newpage);
1180 }
1181 return rc;
1182}
1183
1184/*
1185 * Counterpart of unmap_and_move_page() for hugepage migration.
1186 *
1187 * This function doesn't wait the completion of hugepage I/O
1188 * because there is no race between I/O and migration for hugepage.
1189 * Note that currently hugepage I/O occurs only in direct I/O
1190 * where no lock is held and PG_writeback is irrelevant,
1191 * and writeback status of all subpages are counted in the reference
1192 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1193 * under direct I/O, the reference of the head page is 512 and a bit more.)
1194 * This means that when we try to migrate hugepage whose subpages are
1195 * doing direct I/O, some references remain after try_to_unmap() and
1196 * hugepage migration fails without data corruption.
1197 *
1198 * There is also no race when direct I/O is issued on the page under migration,
1199 * because then pte is replaced with migration swap entry and direct I/O code
1200 * will wait in the page fault for migration to complete.
1201 */
1202static int unmap_and_move_huge_page(new_page_t get_new_page,
1203 free_page_t put_new_page, unsigned long private,
1204 struct page *hpage, int force,
1205 enum migrate_mode mode, int reason)
1206{
1207 int rc = -EAGAIN;
1208 int *result = NULL;
1209 int page_was_mapped = 0;
1210 struct page *new_hpage;
1211 struct anon_vma *anon_vma = NULL;
1212
1213 /*
1214 * Movability of hugepages depends on architectures and hugepage size.
1215 * This check is necessary because some callers of hugepage migration
1216 * like soft offline and memory hotremove don't walk through page
1217 * tables or check whether the hugepage is pmd-based or not before
1218 * kicking migration.
1219 */
1220 if (!hugepage_migration_supported(page_hstate(hpage))) {
1221 putback_active_hugepage(hpage);
1222 return -ENOSYS;
1223 }
1224
1225 new_hpage = get_new_page(hpage, private, &result);
1226 if (!new_hpage)
1227 return -ENOMEM;
1228
1229 if (!trylock_page(hpage)) {
1230 if (!force || mode != MIGRATE_SYNC)
1231 goto out;
1232 lock_page(hpage);
1233 }
1234
1235 if (PageAnon(hpage))
1236 anon_vma = page_get_anon_vma(hpage);
1237
1238 if (unlikely(!trylock_page(new_hpage)))
1239 goto put_anon;
1240
1241 if (page_mapped(hpage)) {
1242 try_to_unmap(hpage,
1243 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1244 page_was_mapped = 1;
1245 }
1246
1247 if (!page_mapped(hpage))
1248 rc = move_to_new_page(new_hpage, hpage, mode);
1249
1250 if (page_was_mapped)
1251 remove_migration_ptes(hpage,
1252 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1253
1254 unlock_page(new_hpage);
1255
1256put_anon:
1257 if (anon_vma)
1258 put_anon_vma(anon_vma);
1259
1260 if (rc == MIGRATEPAGE_SUCCESS) {
1261 hugetlb_cgroup_migrate(hpage, new_hpage);
1262 put_new_page = NULL;
1263 set_page_owner_migrate_reason(new_hpage, reason);
1264 }
1265
1266 unlock_page(hpage);
1267out:
1268 if (rc != -EAGAIN)
1269 putback_active_hugepage(hpage);
1270
1271 /*
1272 * If migration was not successful and there's a freeing callback, use
1273 * it. Otherwise, put_page() will drop the reference grabbed during
1274 * isolation.
1275 */
1276 if (put_new_page)
1277 put_new_page(new_hpage, private);
1278 else
1279 putback_active_hugepage(new_hpage);
1280
1281 if (result) {
1282 if (rc)
1283 *result = rc;
1284 else
1285 *result = page_to_nid(new_hpage);
1286 }
1287 return rc;
1288}
1289
1290/*
1291 * migrate_pages - migrate the pages specified in a list, to the free pages
1292 * supplied as the target for the page migration
1293 *
1294 * @from: The list of pages to be migrated.
1295 * @get_new_page: The function used to allocate free pages to be used
1296 * as the target of the page migration.
1297 * @put_new_page: The function used to free target pages if migration
1298 * fails, or NULL if no special handling is necessary.
1299 * @private: Private data to be passed on to get_new_page()
1300 * @mode: The migration mode that specifies the constraints for
1301 * page migration, if any.
1302 * @reason: The reason for page migration.
1303 *
1304 * The function returns after 10 attempts or if no pages are movable any more
1305 * because the list has become empty or no retryable pages exist any more.
1306 * The caller should call putback_movable_pages() to return pages to the LRU
1307 * or free list only if ret != 0.
1308 *
1309 * Returns the number of pages that were not migrated, or an error code.
1310 */
1311int migrate_pages(struct list_head *from, new_page_t get_new_page,
1312 free_page_t put_new_page, unsigned long private,
1313 enum migrate_mode mode, int reason)
1314{
1315 int retry = 1;
1316 int nr_failed = 0;
1317 int nr_succeeded = 0;
1318 int pass = 0;
1319 struct page *page;
1320 struct page *page2;
1321 int swapwrite = current->flags & PF_SWAPWRITE;
1322 int rc;
1323
1324 if (!swapwrite)
1325 current->flags |= PF_SWAPWRITE;
1326
1327 for(pass = 0; pass < 10 && retry; pass++) {
1328 retry = 0;
1329
1330 list_for_each_entry_safe(page, page2, from, lru) {
1331 cond_resched();
1332
1333 if (PageHuge(page))
1334 rc = unmap_and_move_huge_page(get_new_page,
1335 put_new_page, private, page,
1336 pass > 2, mode, reason);
1337 else
1338 rc = unmap_and_move(get_new_page, put_new_page,
1339 private, page, pass > 2, mode,
1340 reason);
1341
1342 switch(rc) {
1343 case -ENOMEM:
1344 nr_failed++;
1345 goto out;
1346 case -EAGAIN:
1347 retry++;
1348 break;
1349 case MIGRATEPAGE_SUCCESS:
1350 nr_succeeded++;
1351 break;
1352 default:
1353 /*
1354 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1355 * unlike -EAGAIN case, the failed page is
1356 * removed from migration page list and not
1357 * retried in the next outer loop.
1358 */
1359 nr_failed++;
1360 break;
1361 }
1362 }
1363 }
1364 nr_failed += retry;
1365 rc = nr_failed;
1366out:
1367 if (nr_succeeded)
1368 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1369 if (nr_failed)
1370 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1371 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1372
1373 if (!swapwrite)
1374 current->flags &= ~PF_SWAPWRITE;
1375
1376 return rc;
1377}
1378
1379#ifdef CONFIG_NUMA
1380/*
1381 * Move a list of individual pages
1382 */
1383struct page_to_node {
1384 unsigned long addr;
1385 struct page *page;
1386 int node;
1387 int status;
1388};
1389
1390static struct page *new_page_node(struct page *p, unsigned long private,
1391 int **result)
1392{
1393 struct page_to_node *pm = (struct page_to_node *)private;
1394
1395 while (pm->node != MAX_NUMNODES && pm->page != p)
1396 pm++;
1397
1398 if (pm->node == MAX_NUMNODES)
1399 return NULL;
1400
1401 *result = &pm->status;
1402
1403 if (PageHuge(p))
1404 return alloc_huge_page_node(page_hstate(compound_head(p)),
1405 pm->node);
1406 else
1407 return __alloc_pages_node(pm->node,
1408 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1409}
1410
1411/*
1412 * Move a set of pages as indicated in the pm array. The addr
1413 * field must be set to the virtual address of the page to be moved
1414 * and the node number must contain a valid target node.
1415 * The pm array ends with node = MAX_NUMNODES.
1416 */
1417static int do_move_page_to_node_array(struct mm_struct *mm,
1418 struct page_to_node *pm,
1419 int migrate_all)
1420{
1421 int err;
1422 struct page_to_node *pp;
1423 LIST_HEAD(pagelist);
1424
1425 down_read(&mm->mmap_sem);
1426
1427 /*
1428 * Build a list of pages to migrate
1429 */
1430 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1431 struct vm_area_struct *vma;
1432 struct page *page;
1433
1434 err = -EFAULT;
1435 vma = find_vma(mm, pp->addr);
1436 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1437 goto set_status;
1438
1439 /* FOLL_DUMP to ignore special (like zero) pages */
1440 page = follow_page(vma, pp->addr,
1441 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1442
1443 err = PTR_ERR(page);
1444 if (IS_ERR(page))
1445 goto set_status;
1446
1447 err = -ENOENT;
1448 if (!page)
1449 goto set_status;
1450
1451 pp->page = page;
1452 err = page_to_nid(page);
1453
1454 if (err == pp->node)
1455 /*
1456 * Node already in the right place
1457 */
1458 goto put_and_set;
1459
1460 err = -EACCES;
1461 if (page_mapcount(page) > 1 &&
1462 !migrate_all)
1463 goto put_and_set;
1464
1465 if (PageHuge(page)) {
1466 if (PageHead(page))
1467 isolate_huge_page(page, &pagelist);
1468 goto put_and_set;
1469 }
1470
1471 err = isolate_lru_page(page);
1472 if (!err) {
1473 list_add_tail(&page->lru, &pagelist);
1474 inc_node_page_state(page, NR_ISOLATED_ANON +
1475 page_is_file_cache(page));
1476 }
1477put_and_set:
1478 /*
1479 * Either remove the duplicate refcount from
1480 * isolate_lru_page() or drop the page ref if it was
1481 * not isolated.
1482 */
1483 put_page(page);
1484set_status:
1485 pp->status = err;
1486 }
1487
1488 err = 0;
1489 if (!list_empty(&pagelist)) {
1490 err = migrate_pages(&pagelist, new_page_node, NULL,
1491 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1492 if (err)
1493 putback_movable_pages(&pagelist);
1494 }
1495
1496 up_read(&mm->mmap_sem);
1497 return err;
1498}
1499
1500/*
1501 * Migrate an array of page address onto an array of nodes and fill
1502 * the corresponding array of status.
1503 */
1504static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1505 unsigned long nr_pages,
1506 const void __user * __user *pages,
1507 const int __user *nodes,
1508 int __user *status, int flags)
1509{
1510 struct page_to_node *pm;
1511 unsigned long chunk_nr_pages;
1512 unsigned long chunk_start;
1513 int err;
1514
1515 err = -ENOMEM;
1516 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1517 if (!pm)
1518 goto out;
1519
1520 migrate_prep();
1521
1522 /*
1523 * Store a chunk of page_to_node array in a page,
1524 * but keep the last one as a marker
1525 */
1526 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1527
1528 for (chunk_start = 0;
1529 chunk_start < nr_pages;
1530 chunk_start += chunk_nr_pages) {
1531 int j;
1532
1533 if (chunk_start + chunk_nr_pages > nr_pages)
1534 chunk_nr_pages = nr_pages - chunk_start;
1535
1536 /* fill the chunk pm with addrs and nodes from user-space */
1537 for (j = 0; j < chunk_nr_pages; j++) {
1538 const void __user *p;
1539 int node;
1540
1541 err = -EFAULT;
1542 if (get_user(p, pages + j + chunk_start))
1543 goto out_pm;
1544 pm[j].addr = (unsigned long) p;
1545
1546 if (get_user(node, nodes + j + chunk_start))
1547 goto out_pm;
1548
1549 err = -ENODEV;
1550 if (node < 0 || node >= MAX_NUMNODES)
1551 goto out_pm;
1552
1553 if (!node_state(node, N_MEMORY))
1554 goto out_pm;
1555
1556 err = -EACCES;
1557 if (!node_isset(node, task_nodes))
1558 goto out_pm;
1559
1560 pm[j].node = node;
1561 }
1562
1563 /* End marker for this chunk */
1564 pm[chunk_nr_pages].node = MAX_NUMNODES;
1565
1566 /* Migrate this chunk */
1567 err = do_move_page_to_node_array(mm, pm,
1568 flags & MPOL_MF_MOVE_ALL);
1569 if (err < 0)
1570 goto out_pm;
1571
1572 /* Return status information */
1573 for (j = 0; j < chunk_nr_pages; j++)
1574 if (put_user(pm[j].status, status + j + chunk_start)) {
1575 err = -EFAULT;
1576 goto out_pm;
1577 }
1578 }
1579 err = 0;
1580
1581out_pm:
1582 free_page((unsigned long)pm);
1583out:
1584 return err;
1585}
1586
1587/*
1588 * Determine the nodes of an array of pages and store it in an array of status.
1589 */
1590static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1591 const void __user **pages, int *status)
1592{
1593 unsigned long i;
1594
1595 down_read(&mm->mmap_sem);
1596
1597 for (i = 0; i < nr_pages; i++) {
1598 unsigned long addr = (unsigned long)(*pages);
1599 struct vm_area_struct *vma;
1600 struct page *page;
1601 int err = -EFAULT;
1602
1603 vma = find_vma(mm, addr);
1604 if (!vma || addr < vma->vm_start)
1605 goto set_status;
1606
1607 /* FOLL_DUMP to ignore special (like zero) pages */
1608 page = follow_page(vma, addr, FOLL_DUMP);
1609
1610 err = PTR_ERR(page);
1611 if (IS_ERR(page))
1612 goto set_status;
1613
1614 err = page ? page_to_nid(page) : -ENOENT;
1615set_status:
1616 *status = err;
1617
1618 pages++;
1619 status++;
1620 }
1621
1622 up_read(&mm->mmap_sem);
1623}
1624
1625/*
1626 * Determine the nodes of a user array of pages and store it in
1627 * a user array of status.
1628 */
1629static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1630 const void __user * __user *pages,
1631 int __user *status)
1632{
1633#define DO_PAGES_STAT_CHUNK_NR 16
1634 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1635 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1636
1637 while (nr_pages) {
1638 unsigned long chunk_nr;
1639
1640 chunk_nr = nr_pages;
1641 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1642 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1643
1644 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1645 break;
1646
1647 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1648
1649 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1650 break;
1651
1652 pages += chunk_nr;
1653 status += chunk_nr;
1654 nr_pages -= chunk_nr;
1655 }
1656 return nr_pages ? -EFAULT : 0;
1657}
1658
1659/*
1660 * Move a list of pages in the address space of the currently executing
1661 * process.
1662 */
1663SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1664 const void __user * __user *, pages,
1665 const int __user *, nodes,
1666 int __user *, status, int, flags)
1667{
1668 const struct cred *cred = current_cred(), *tcred;
1669 struct task_struct *task;
1670 struct mm_struct *mm;
1671 int err;
1672 nodemask_t task_nodes;
1673
1674 /* Check flags */
1675 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1676 return -EINVAL;
1677
1678 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1679 return -EPERM;
1680
1681 /* Find the mm_struct */
1682 rcu_read_lock();
1683 task = pid ? find_task_by_vpid(pid) : current;
1684 if (!task) {
1685 rcu_read_unlock();
1686 return -ESRCH;
1687 }
1688 get_task_struct(task);
1689
1690 /*
1691 * Check if this process has the right to modify the specified
1692 * process. The right exists if the process has administrative
1693 * capabilities, superuser privileges or the same
1694 * userid as the target process.
1695 */
1696 tcred = __task_cred(task);
1697 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1698 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1699 !capable(CAP_SYS_NICE)) {
1700 rcu_read_unlock();
1701 err = -EPERM;
1702 goto out;
1703 }
1704 rcu_read_unlock();
1705
1706 err = security_task_movememory(task);
1707 if (err)
1708 goto out;
1709
1710 task_nodes = cpuset_mems_allowed(task);
1711 mm = get_task_mm(task);
1712 put_task_struct(task);
1713
1714 if (!mm)
1715 return -EINVAL;
1716
1717 if (nodes)
1718 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1719 nodes, status, flags);
1720 else
1721 err = do_pages_stat(mm, nr_pages, pages, status);
1722
1723 mmput(mm);
1724 return err;
1725
1726out:
1727 put_task_struct(task);
1728 return err;
1729}
1730
1731#ifdef CONFIG_NUMA_BALANCING
1732/*
1733 * Returns true if this is a safe migration target node for misplaced NUMA
1734 * pages. Currently it only checks the watermarks which crude
1735 */
1736static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1737 unsigned long nr_migrate_pages)
1738{
1739 int z;
1740
1741 if (!pgdat_reclaimable(pgdat))
1742 return false;
1743
1744 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1745 struct zone *zone = pgdat->node_zones + z;
1746
1747 if (!populated_zone(zone))
1748 continue;
1749
1750 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1751 if (!zone_watermark_ok(zone, 0,
1752 high_wmark_pages(zone) +
1753 nr_migrate_pages,
1754 0, 0))
1755 continue;
1756 return true;
1757 }
1758 return false;
1759}
1760
1761static struct page *alloc_misplaced_dst_page(struct page *page,
1762 unsigned long data,
1763 int **result)
1764{
1765 int nid = (int) data;
1766 struct page *newpage;
1767
1768 newpage = __alloc_pages_node(nid,
1769 (GFP_HIGHUSER_MOVABLE |
1770 __GFP_THISNODE | __GFP_NOMEMALLOC |
1771 __GFP_NORETRY | __GFP_NOWARN) &
1772 ~__GFP_RECLAIM, 0);
1773
1774 return newpage;
1775}
1776
1777/*
1778 * page migration rate limiting control.
1779 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1780 * window of time. Default here says do not migrate more than 1280M per second.
1781 */
1782static unsigned int migrate_interval_millisecs __read_mostly = 100;
1783static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1784
1785/* Returns true if the node is migrate rate-limited after the update */
1786static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1787 unsigned long nr_pages)
1788{
1789 /*
1790 * Rate-limit the amount of data that is being migrated to a node.
1791 * Optimal placement is no good if the memory bus is saturated and
1792 * all the time is being spent migrating!
1793 */
1794 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1795 spin_lock(&pgdat->numabalancing_migrate_lock);
1796 pgdat->numabalancing_migrate_nr_pages = 0;
1797 pgdat->numabalancing_migrate_next_window = jiffies +
1798 msecs_to_jiffies(migrate_interval_millisecs);
1799 spin_unlock(&pgdat->numabalancing_migrate_lock);
1800 }
1801 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1802 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1803 nr_pages);
1804 return true;
1805 }
1806
1807 /*
1808 * This is an unlocked non-atomic update so errors are possible.
1809 * The consequences are failing to migrate when we potentiall should
1810 * have which is not severe enough to warrant locking. If it is ever
1811 * a problem, it can be converted to a per-cpu counter.
1812 */
1813 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1814 return false;
1815}
1816
1817static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1818{
1819 int page_lru;
1820
1821 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1822
1823 /* Avoid migrating to a node that is nearly full */
1824 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1825 return 0;
1826
1827 if (isolate_lru_page(page))
1828 return 0;
1829
1830 /*
1831 * migrate_misplaced_transhuge_page() skips page migration's usual
1832 * check on page_count(), so we must do it here, now that the page
1833 * has been isolated: a GUP pin, or any other pin, prevents migration.
1834 * The expected page count is 3: 1 for page's mapcount and 1 for the
1835 * caller's pin and 1 for the reference taken by isolate_lru_page().
1836 */
1837 if (PageTransHuge(page) && page_count(page) != 3) {
1838 putback_lru_page(page);
1839 return 0;
1840 }
1841
1842 page_lru = page_is_file_cache(page);
1843 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1844 hpage_nr_pages(page));
1845
1846 /*
1847 * Isolating the page has taken another reference, so the
1848 * caller's reference can be safely dropped without the page
1849 * disappearing underneath us during migration.
1850 */
1851 put_page(page);
1852 return 1;
1853}
1854
1855bool pmd_trans_migrating(pmd_t pmd)
1856{
1857 struct page *page = pmd_page(pmd);
1858 return PageLocked(page);
1859}
1860
1861/*
1862 * Attempt to migrate a misplaced page to the specified destination
1863 * node. Caller is expected to have an elevated reference count on
1864 * the page that will be dropped by this function before returning.
1865 */
1866int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1867 int node)
1868{
1869 pg_data_t *pgdat = NODE_DATA(node);
1870 int isolated;
1871 int nr_remaining;
1872 LIST_HEAD(migratepages);
1873
1874 /*
1875 * Don't migrate file pages that are mapped in multiple processes
1876 * with execute permissions as they are probably shared libraries.
1877 */
1878 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1879 (vma->vm_flags & VM_EXEC))
1880 goto out;
1881
1882 /*
1883 * Rate-limit the amount of data that is being migrated to a node.
1884 * Optimal placement is no good if the memory bus is saturated and
1885 * all the time is being spent migrating!
1886 */
1887 if (numamigrate_update_ratelimit(pgdat, 1))
1888 goto out;
1889
1890 isolated = numamigrate_isolate_page(pgdat, page);
1891 if (!isolated)
1892 goto out;
1893
1894 list_add(&page->lru, &migratepages);
1895 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1896 NULL, node, MIGRATE_ASYNC,
1897 MR_NUMA_MISPLACED);
1898 if (nr_remaining) {
1899 if (!list_empty(&migratepages)) {
1900 list_del(&page->lru);
1901 dec_node_page_state(page, NR_ISOLATED_ANON +
1902 page_is_file_cache(page));
1903 putback_lru_page(page);
1904 }
1905 isolated = 0;
1906 } else
1907 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1908 BUG_ON(!list_empty(&migratepages));
1909 return isolated;
1910
1911out:
1912 put_page(page);
1913 return 0;
1914}
1915#endif /* CONFIG_NUMA_BALANCING */
1916
1917#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1918/*
1919 * Migrates a THP to a given target node. page must be locked and is unlocked
1920 * before returning.
1921 */
1922int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1923 struct vm_area_struct *vma,
1924 pmd_t *pmd, pmd_t entry,
1925 unsigned long address,
1926 struct page *page, int node)
1927{
1928 spinlock_t *ptl;
1929 pg_data_t *pgdat = NODE_DATA(node);
1930 int isolated = 0;
1931 struct page *new_page = NULL;
1932 int page_lru = page_is_file_cache(page);
1933 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1934 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1935 pmd_t orig_entry;
1936
1937 /*
1938 * Rate-limit the amount of data that is being migrated to a node.
1939 * Optimal placement is no good if the memory bus is saturated and
1940 * all the time is being spent migrating!
1941 */
1942 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1943 goto out_dropref;
1944
1945 new_page = alloc_pages_node(node,
1946 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1947 HPAGE_PMD_ORDER);
1948 if (!new_page)
1949 goto out_fail;
1950 prep_transhuge_page(new_page);
1951
1952 isolated = numamigrate_isolate_page(pgdat, page);
1953 if (!isolated) {
1954 put_page(new_page);
1955 goto out_fail;
1956 }
1957 /*
1958 * We are not sure a pending tlb flush here is for a huge page
1959 * mapping or not. Hence use the tlb range variant
1960 */
1961 if (mm_tlb_flush_pending(mm))
1962 flush_tlb_range(vma, mmun_start, mmun_end);
1963
1964 /* Prepare a page as a migration target */
1965 __SetPageLocked(new_page);
1966 __SetPageSwapBacked(new_page);
1967
1968 /* anon mapping, we can simply copy page->mapping to the new page: */
1969 new_page->mapping = page->mapping;
1970 new_page->index = page->index;
1971 migrate_page_copy(new_page, page);
1972 WARN_ON(PageLRU(new_page));
1973
1974 /* Recheck the target PMD */
1975 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1976 ptl = pmd_lock(mm, pmd);
1977 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1978fail_putback:
1979 spin_unlock(ptl);
1980 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1981
1982 /* Reverse changes made by migrate_page_copy() */
1983 if (TestClearPageActive(new_page))
1984 SetPageActive(page);
1985 if (TestClearPageUnevictable(new_page))
1986 SetPageUnevictable(page);
1987
1988 unlock_page(new_page);
1989 put_page(new_page); /* Free it */
1990
1991 /* Retake the callers reference and putback on LRU */
1992 get_page(page);
1993 putback_lru_page(page);
1994 mod_node_page_state(page_pgdat(page),
1995 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1996
1997 goto out_unlock;
1998 }
1999
2000 orig_entry = *pmd;
2001 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2002 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2003
2004 /*
2005 * Clear the old entry under pagetable lock and establish the new PTE.
2006 * Any parallel GUP will either observe the old page blocking on the
2007 * page lock, block on the page table lock or observe the new page.
2008 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2009 * guarantee the copy is visible before the pagetable update.
2010 */
2011 flush_cache_range(vma, mmun_start, mmun_end);
2012 page_add_anon_rmap(new_page, vma, mmun_start, true);
2013 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2014 set_pmd_at(mm, mmun_start, pmd, entry);
2015 update_mmu_cache_pmd(vma, address, &entry);
2016
2017 if (page_count(page) != 2) {
2018 set_pmd_at(mm, mmun_start, pmd, orig_entry);
2019 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
2020 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2021 update_mmu_cache_pmd(vma, address, &entry);
2022 page_remove_rmap(new_page, true);
2023 goto fail_putback;
2024 }
2025
2026 mlock_migrate_page(new_page, page);
2027 page_remove_rmap(page, true);
2028 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2029
2030 spin_unlock(ptl);
2031 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2032
2033 /* Take an "isolate" reference and put new page on the LRU. */
2034 get_page(new_page);
2035 putback_lru_page(new_page);
2036
2037 unlock_page(new_page);
2038 unlock_page(page);
2039 put_page(page); /* Drop the rmap reference */
2040 put_page(page); /* Drop the LRU isolation reference */
2041
2042 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2043 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2044
2045 mod_node_page_state(page_pgdat(page),
2046 NR_ISOLATED_ANON + page_lru,
2047 -HPAGE_PMD_NR);
2048 return isolated;
2049
2050out_fail:
2051 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2052out_dropref:
2053 ptl = pmd_lock(mm, pmd);
2054 if (pmd_same(*pmd, entry)) {
2055 entry = pmd_modify(entry, vma->vm_page_prot);
2056 set_pmd_at(mm, mmun_start, pmd, entry);
2057 update_mmu_cache_pmd(vma, address, &entry);
2058 }
2059 spin_unlock(ptl);
2060
2061out_unlock:
2062 unlock_page(page);
2063 put_page(page);
2064 return 0;
2065}
2066#endif /* CONFIG_NUMA_BALANCING */
2067
2068#endif /* CONFIG_NUMA */