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(entry: &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, item: NR_ISOLATED_ANON +
174 folio_is_file_lru(folio), nr: -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: &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, address: 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: &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(ptep: pvmw.pte);
214 if (pte_swp_soft_dirty(pte: old_pte))
215 pte = pte_mksoft_dirty(pte);
216
217 entry = pte_to_swp_entry(pte: 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(pte: 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(pte: old_pte))
239 pte = pte_swp_mksoft_dirty(pte);
240 if (pte_swp_uffd_wp(pte: 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(entry: pte, shift, flags: vma->vm_flags);
251 if (folio_test_anon(folio))
252 hugepage_add_anon_rmap(folio, vma, address: pvmw.address,
253 flags: rmap_flags);
254 else
255 page_dup_file_rmap(page: new, compound: true);
256 set_huge_pte_at(mm: vma->vm_mm, addr: pvmw.address, ptep: pvmw.pte, pte,
257 sz: psize);
258 } else
259#endif
260 {
261 if (folio_test_anon(folio))
262 page_add_anon_rmap(new, vma, address: pvmw.address,
263 flags: rmap_flags);
264 else
265 page_add_file_rmap(new, vma, compound: false);
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(addr: pvmw.address, pte: pte_val(pte),
272 order: compound_order(page: new));
273
274 /* No need to invalidate - it was non-present before */
275 update_mmu_cache(vma, addr: pvmw.address, ptep: 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(folio: dst, rwc: &rwc);
294 else
295 rmap_walk(folio: dst, rwc: &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, addr: address, ptlp: &ptl);
312 if (!ptep)
313 return;
314
315 pte = ptep_get(ptep);
316 pte_unmap(pte: 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(lock: 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(h: hstate_vma(vma), mm: vma->vm_mm, pte: ptep);
341 pte_t pte;
342
343 hugetlb_vma_assert_locked(vma);
344 spin_lock(lock: ptl);
345 pte = huge_ptep_get(ptep);
346
347 if (unlikely(!is_hugetlb_entry_migration(pte))) {
348 spin_unlock(lock: 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(entry: 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: *pmd))
370 goto unlock;
371 migration_entry_wait_on_locked(entry: pmd_to_swp_entry(pmd: *pmd), ptl);
372 return;
373unlock:
374 spin_unlock(lock: 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
409 if (!mapping) {
410 /* Anonymous page without mapping */
411 if (folio_ref_count(folio) != expected_count)
412 return -EAGAIN;
413
414 /* No turning back from here */
415 newfolio->index = folio->index;
416 newfolio->mapping = folio->mapping;
417 if (folio_test_swapbacked(folio))
418 __folio_set_swapbacked(folio: newfolio);
419
420 return MIGRATEPAGE_SUCCESS;
421 }
422
423 oldzone = folio_zone(folio);
424 newzone = folio_zone(folio: newfolio);
425
426 xas_lock_irq(&xas);
427 if (!folio_ref_freeze(folio, count: expected_count)) {
428 xas_unlock_irq(&xas);
429 return -EAGAIN;
430 }
431
432 /*
433 * Now we know that no one else is looking at the folio:
434 * no turning back from here.
435 */
436 newfolio->index = folio->index;
437 newfolio->mapping = folio->mapping;
438 folio_ref_add(folio: newfolio, nr); /* add cache reference */
439 if (folio_test_swapbacked(folio)) {
440 __folio_set_swapbacked(folio: newfolio);
441 if (folio_test_swapcache(folio)) {
442 folio_set_swapcache(folio: newfolio);
443 newfolio->private = folio_get_private(folio);
444 }
445 } else {
446 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
447 }
448
449 /* Move dirty while page refs frozen and newpage not yet exposed */
450 dirty = folio_test_dirty(folio);
451 if (dirty) {
452 folio_clear_dirty(folio);
453 folio_set_dirty(folio: newfolio);
454 }
455
456 xas_store(&xas, entry: newfolio);
457
458 /*
459 * Drop cache reference from old page by unfreezing
460 * to one less reference.
461 * We know this isn't the last reference.
462 */
463 folio_ref_unfreeze(folio, count: expected_count - nr);
464
465 xas_unlock(&xas);
466 /* Leave irq disabled to prevent preemption while updating stats */
467
468 /*
469 * If moved to a different zone then also account
470 * the page for that zone. Other VM counters will be
471 * taken care of when we establish references to the
472 * new page and drop references to the old page.
473 *
474 * Note that anonymous pages are accounted for
475 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
476 * are mapped to swap space.
477 */
478 if (newzone != oldzone) {
479 struct lruvec *old_lruvec, *new_lruvec;
480 struct mem_cgroup *memcg;
481
482 memcg = folio_memcg(folio);
483 old_lruvec = mem_cgroup_lruvec(memcg, pgdat: oldzone->zone_pgdat);
484 new_lruvec = mem_cgroup_lruvec(memcg, pgdat: newzone->zone_pgdat);
485
486 __mod_lruvec_state(lruvec: old_lruvec, idx: NR_FILE_PAGES, val: -nr);
487 __mod_lruvec_state(lruvec: new_lruvec, idx: NR_FILE_PAGES, val: nr);
488 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
489 __mod_lruvec_state(lruvec: old_lruvec, idx: NR_SHMEM, val: -nr);
490 __mod_lruvec_state(lruvec: new_lruvec, idx: NR_SHMEM, val: nr);
491
492 if (folio_test_pmd_mappable(folio)) {
493 __mod_lruvec_state(lruvec: old_lruvec, idx: NR_SHMEM_THPS, val: -nr);
494 __mod_lruvec_state(lruvec: new_lruvec, idx: NR_SHMEM_THPS, val: nr);
495 }
496 }
497#ifdef CONFIG_SWAP
498 if (folio_test_swapcache(folio)) {
499 __mod_lruvec_state(lruvec: old_lruvec, idx: NR_SWAPCACHE, val: -nr);
500 __mod_lruvec_state(lruvec: new_lruvec, idx: NR_SWAPCACHE, val: nr);
501 }
502#endif
503 if (dirty && mapping_can_writeback(mapping)) {
504 __mod_lruvec_state(lruvec: old_lruvec, idx: NR_FILE_DIRTY, val: -nr);
505 __mod_zone_page_state(oldzone, item: NR_ZONE_WRITE_PENDING, -nr);
506 __mod_lruvec_state(lruvec: new_lruvec, idx: NR_FILE_DIRTY, val: nr);
507 __mod_zone_page_state(newzone, item: NR_ZONE_WRITE_PENDING, nr);
508 }
509 }
510 local_irq_enable();
511
512 return MIGRATEPAGE_SUCCESS;
513}
514EXPORT_SYMBOL(folio_migrate_mapping);
515
516/*
517 * The expected number of remaining references is the same as that
518 * of folio_migrate_mapping().
519 */
520int migrate_huge_page_move_mapping(struct address_space *mapping,
521 struct folio *dst, struct folio *src)
522{
523 XA_STATE(xas, &mapping->i_pages, folio_index(src));
524 int expected_count;
525
526 xas_lock_irq(&xas);
527 expected_count = folio_expected_refs(mapping, folio: src);
528 if (!folio_ref_freeze(folio: src, count: expected_count)) {
529 xas_unlock_irq(&xas);
530 return -EAGAIN;
531 }
532
533 dst->index = src->index;
534 dst->mapping = src->mapping;
535
536 folio_ref_add(folio: dst, nr: folio_nr_pages(folio: dst));
537
538 xas_store(&xas, entry: dst);
539
540 folio_ref_unfreeze(folio: src, count: expected_count - folio_nr_pages(folio: src));
541
542 xas_unlock_irq(&xas);
543
544 return MIGRATEPAGE_SUCCESS;
545}
546
547/*
548 * Copy the flags and some other ancillary information
549 */
550void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
551{
552 int cpupid;
553
554 if (folio_test_error(folio))
555 folio_set_error(folio: newfolio);
556 if (folio_test_referenced(folio))
557 folio_set_referenced(folio: newfolio);
558 if (folio_test_uptodate(folio))
559 folio_mark_uptodate(folio: newfolio);
560 if (folio_test_clear_active(folio)) {
561 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
562 folio_set_active(folio: newfolio);
563 } else if (folio_test_clear_unevictable(folio))
564 folio_set_unevictable(folio: newfolio);
565 if (folio_test_workingset(folio))
566 folio_set_workingset(folio: newfolio);
567 if (folio_test_checked(folio))
568 folio_set_checked(folio: newfolio);
569 /*
570 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
571 * migration entries. We can still have PG_anon_exclusive set on an
572 * effectively unmapped and unreferenced first sub-pages of an
573 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
574 */
575 if (folio_test_mappedtodisk(folio))
576 folio_set_mappedtodisk(folio: newfolio);
577
578 /* Move dirty on pages not done by folio_migrate_mapping() */
579 if (folio_test_dirty(folio))
580 folio_set_dirty(folio: newfolio);
581
582 if (folio_test_young(folio))
583 folio_set_young(folio: newfolio);
584 if (folio_test_idle(folio))
585 folio_set_idle(folio: newfolio);
586
587 /*
588 * Copy NUMA information to the new page, to prevent over-eager
589 * future migrations of this same page.
590 */
591 cpupid = folio_xchg_last_cpupid(folio, cpupid: -1);
592 /*
593 * For memory tiering mode, when migrate between slow and fast
594 * memory node, reset cpupid, because that is used to record
595 * page access time in slow memory node.
596 */
597 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
598 bool f_toptier = node_is_toptier(node: folio_nid(folio));
599 bool t_toptier = node_is_toptier(node: folio_nid(folio: newfolio));
600
601 if (f_toptier != t_toptier)
602 cpupid = -1;
603 }
604 folio_xchg_last_cpupid(folio: newfolio, cpupid);
605
606 folio_migrate_ksm(newfolio, folio);
607 /*
608 * Please do not reorder this without considering how mm/ksm.c's
609 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
610 */
611 if (folio_test_swapcache(folio))
612 folio_clear_swapcache(folio);
613 folio_clear_private(folio);
614
615 /* page->private contains hugetlb specific flags */
616 if (!folio_test_hugetlb(folio))
617 folio->private = NULL;
618
619 /*
620 * If any waiters have accumulated on the new page then
621 * wake them up.
622 */
623 if (folio_test_writeback(folio: newfolio))
624 folio_end_writeback(folio: newfolio);
625
626 /*
627 * PG_readahead shares the same bit with PG_reclaim. The above
628 * end_page_writeback() may clear PG_readahead mistakenly, so set the
629 * bit after that.
630 */
631 if (folio_test_readahead(folio))
632 folio_set_readahead(folio: newfolio);
633
634 folio_copy_owner(newfolio, old: folio);
635
636 mem_cgroup_migrate(old: folio, new: newfolio);
637}
638EXPORT_SYMBOL(folio_migrate_flags);
639
640void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
641{
642 folio_copy(dst: newfolio, src: folio);
643 folio_migrate_flags(newfolio, folio);
644}
645EXPORT_SYMBOL(folio_migrate_copy);
646
647/************************************************************
648 * Migration functions
649 ***********************************************************/
650
651int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
652 struct folio *src, enum migrate_mode mode, int extra_count)
653{
654 int rc;
655
656 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
657
658 rc = folio_migrate_mapping(mapping, dst, src, extra_count);
659
660 if (rc != MIGRATEPAGE_SUCCESS)
661 return rc;
662
663 if (mode != MIGRATE_SYNC_NO_COPY)
664 folio_migrate_copy(dst, src);
665 else
666 folio_migrate_flags(dst, src);
667 return MIGRATEPAGE_SUCCESS;
668}
669
670/**
671 * migrate_folio() - Simple folio migration.
672 * @mapping: The address_space containing the folio.
673 * @dst: The folio to migrate the data to.
674 * @src: The folio containing the current data.
675 * @mode: How to migrate the page.
676 *
677 * Common logic to directly migrate a single LRU folio suitable for
678 * folios that do not use PagePrivate/PagePrivate2.
679 *
680 * Folios are locked upon entry and exit.
681 */
682int migrate_folio(struct address_space *mapping, struct folio *dst,
683 struct folio *src, enum migrate_mode mode)
684{
685 return migrate_folio_extra(mapping, dst, src, mode, extra_count: 0);
686}
687EXPORT_SYMBOL(migrate_folio);
688
689#ifdef CONFIG_BUFFER_HEAD
690/* Returns true if all buffers are successfully locked */
691static bool buffer_migrate_lock_buffers(struct buffer_head *head,
692 enum migrate_mode mode)
693{
694 struct buffer_head *bh = head;
695 struct buffer_head *failed_bh;
696
697 do {
698 if (!trylock_buffer(bh)) {
699 if (mode == MIGRATE_ASYNC)
700 goto unlock;
701 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
702 goto unlock;
703 lock_buffer(bh);
704 }
705
706 bh = bh->b_this_page;
707 } while (bh != head);
708
709 return true;
710
711unlock:
712 /* We failed to lock the buffer and cannot stall. */
713 failed_bh = bh;
714 bh = head;
715 while (bh != failed_bh) {
716 unlock_buffer(bh);
717 bh = bh->b_this_page;
718 }
719
720 return false;
721}
722
723static int __buffer_migrate_folio(struct address_space *mapping,
724 struct folio *dst, struct folio *src, enum migrate_mode mode,
725 bool check_refs)
726{
727 struct buffer_head *bh, *head;
728 int rc;
729 int expected_count;
730
731 head = folio_buffers(src);
732 if (!head)
733 return migrate_folio(mapping, dst, src, mode);
734
735 /* Check whether page does not have extra refs before we do more work */
736 expected_count = folio_expected_refs(mapping, folio: src);
737 if (folio_ref_count(folio: src) != expected_count)
738 return -EAGAIN;
739
740 if (!buffer_migrate_lock_buffers(head, mode))
741 return -EAGAIN;
742
743 if (check_refs) {
744 bool busy;
745 bool invalidated = false;
746
747recheck_buffers:
748 busy = false;
749 spin_lock(lock: &mapping->private_lock);
750 bh = head;
751 do {
752 if (atomic_read(v: &bh->b_count)) {
753 busy = true;
754 break;
755 }
756 bh = bh->b_this_page;
757 } while (bh != head);
758 if (busy) {
759 if (invalidated) {
760 rc = -EAGAIN;
761 goto unlock_buffers;
762 }
763 spin_unlock(lock: &mapping->private_lock);
764 invalidate_bh_lrus();
765 invalidated = true;
766 goto recheck_buffers;
767 }
768 }
769
770 rc = folio_migrate_mapping(mapping, dst, src, 0);
771 if (rc != MIGRATEPAGE_SUCCESS)
772 goto unlock_buffers;
773
774 folio_attach_private(folio: dst, data: folio_detach_private(folio: src));
775
776 bh = head;
777 do {
778 folio_set_bh(bh, folio: dst, offset: bh_offset(bh));
779 bh = bh->b_this_page;
780 } while (bh != head);
781
782 if (mode != MIGRATE_SYNC_NO_COPY)
783 folio_migrate_copy(dst, src);
784 else
785 folio_migrate_flags(dst, src);
786
787 rc = MIGRATEPAGE_SUCCESS;
788unlock_buffers:
789 if (check_refs)
790 spin_unlock(lock: &mapping->private_lock);
791 bh = head;
792 do {
793 unlock_buffer(bh);
794 bh = bh->b_this_page;
795 } while (bh != head);
796
797 return rc;
798}
799
800/**
801 * buffer_migrate_folio() - Migration function for folios with buffers.
802 * @mapping: The address space containing @src.
803 * @dst: The folio to migrate to.
804 * @src: The folio to migrate from.
805 * @mode: How to migrate the folio.
806 *
807 * This function can only be used if the underlying filesystem guarantees
808 * that no other references to @src exist. For example attached buffer
809 * heads are accessed only under the folio lock. If your filesystem cannot
810 * provide this guarantee, buffer_migrate_folio_norefs() may be more
811 * appropriate.
812 *
813 * Return: 0 on success or a negative errno on failure.
814 */
815int buffer_migrate_folio(struct address_space *mapping,
816 struct folio *dst, struct folio *src, enum migrate_mode mode)
817{
818 return __buffer_migrate_folio(mapping, dst, src, mode, check_refs: false);
819}
820EXPORT_SYMBOL(buffer_migrate_folio);
821
822/**
823 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
824 * @mapping: The address space containing @src.
825 * @dst: The folio to migrate to.
826 * @src: The folio to migrate from.
827 * @mode: How to migrate the folio.
828 *
829 * Like buffer_migrate_folio() except that this variant is more careful
830 * and checks that there are also no buffer head references. This function
831 * is the right one for mappings where buffer heads are directly looked
832 * up and referenced (such as block device mappings).
833 *
834 * Return: 0 on success or a negative errno on failure.
835 */
836int buffer_migrate_folio_norefs(struct address_space *mapping,
837 struct folio *dst, struct folio *src, enum migrate_mode mode)
838{
839 return __buffer_migrate_folio(mapping, dst, src, mode, check_refs: true);
840}
841EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
842#endif /* CONFIG_BUFFER_HEAD */
843
844int filemap_migrate_folio(struct address_space *mapping,
845 struct folio *dst, struct folio *src, enum migrate_mode mode)
846{
847 int ret;
848
849 ret = folio_migrate_mapping(mapping, dst, src, 0);
850 if (ret != MIGRATEPAGE_SUCCESS)
851 return ret;
852
853 if (folio_get_private(folio: src))
854 folio_attach_private(folio: dst, data: folio_detach_private(folio: src));
855
856 if (mode != MIGRATE_SYNC_NO_COPY)
857 folio_migrate_copy(dst, src);
858 else
859 folio_migrate_flags(dst, src);
860 return MIGRATEPAGE_SUCCESS;
861}
862EXPORT_SYMBOL_GPL(filemap_migrate_folio);
863
864/*
865 * Writeback a folio to clean the dirty state
866 */
867static int writeout(struct address_space *mapping, struct folio *folio)
868{
869 struct writeback_control wbc = {
870 .sync_mode = WB_SYNC_NONE,
871 .nr_to_write = 1,
872 .range_start = 0,
873 .range_end = LLONG_MAX,
874 .for_reclaim = 1
875 };
876 int rc;
877
878 if (!mapping->a_ops->writepage)
879 /* No write method for the address space */
880 return -EINVAL;
881
882 if (!folio_clear_dirty_for_io(folio))
883 /* Someone else already triggered a write */
884 return -EAGAIN;
885
886 /*
887 * A dirty folio may imply that the underlying filesystem has
888 * the folio on some queue. So the folio must be clean for
889 * migration. Writeout may mean we lose the lock and the
890 * folio state is no longer what we checked for earlier.
891 * At this point we know that the migration attempt cannot
892 * be successful.
893 */
894 remove_migration_ptes(src: folio, dst: folio, locked: false);
895
896 rc = mapping->a_ops->writepage(&folio->page, &wbc);
897
898 if (rc != AOP_WRITEPAGE_ACTIVATE)
899 /* unlocked. Relock */
900 folio_lock(folio);
901
902 return (rc < 0) ? -EIO : -EAGAIN;
903}
904
905/*
906 * Default handling if a filesystem does not provide a migration function.
907 */
908static int fallback_migrate_folio(struct address_space *mapping,
909 struct folio *dst, struct folio *src, enum migrate_mode mode)
910{
911 if (folio_test_dirty(folio: src)) {
912 /* Only writeback folios in full synchronous migration */
913 switch (mode) {
914 case MIGRATE_SYNC:
915 case MIGRATE_SYNC_NO_COPY:
916 break;
917 default:
918 return -EBUSY;
919 }
920 return writeout(mapping, folio: src);
921 }
922
923 /*
924 * Buffers may be managed in a filesystem specific way.
925 * We must have no buffers or drop them.
926 */
927 if (!filemap_release_folio(folio: src, GFP_KERNEL))
928 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
929
930 return migrate_folio(mapping, dst, src, mode);
931}
932
933/*
934 * Move a page to a newly allocated page
935 * The page is locked and all ptes have been successfully removed.
936 *
937 * The new page will have replaced the old page if this function
938 * is successful.
939 *
940 * Return value:
941 * < 0 - error code
942 * MIGRATEPAGE_SUCCESS - success
943 */
944static int move_to_new_folio(struct folio *dst, struct folio *src,
945 enum migrate_mode mode)
946{
947 int rc = -EAGAIN;
948 bool is_lru = !__folio_test_movable(folio: src);
949
950 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
951 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
952
953 if (likely(is_lru)) {
954 struct address_space *mapping = folio_mapping(src);
955
956 if (!mapping)
957 rc = migrate_folio(mapping, dst, src, mode);
958 else if (mapping->a_ops->migrate_folio)
959 /*
960 * Most folios have a mapping and most filesystems
961 * provide a migrate_folio callback. Anonymous folios
962 * are part of swap space which also has its own
963 * migrate_folio callback. This is the most common path
964 * for page migration.
965 */
966 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
967 mode);
968 else
969 rc = fallback_migrate_folio(mapping, dst, src, mode);
970 } else {
971 const struct movable_operations *mops;
972
973 /*
974 * In case of non-lru page, it could be released after
975 * isolation step. In that case, we shouldn't try migration.
976 */
977 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
978 if (!folio_test_movable(folio: src)) {
979 rc = MIGRATEPAGE_SUCCESS;
980 folio_clear_isolated(folio: src);
981 goto out;
982 }
983
984 mops = folio_movable_ops(folio: src);
985 rc = mops->migrate_page(&dst->page, &src->page, mode);
986 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
987 !folio_test_isolated(src));
988 }
989
990 /*
991 * When successful, old pagecache src->mapping must be cleared before
992 * src is freed; but stats require that PageAnon be left as PageAnon.
993 */
994 if (rc == MIGRATEPAGE_SUCCESS) {
995 if (__folio_test_movable(folio: src)) {
996 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
997
998 /*
999 * We clear PG_movable under page_lock so any compactor
1000 * cannot try to migrate this page.
1001 */
1002 folio_clear_isolated(folio: src);
1003 }
1004
1005 /*
1006 * Anonymous and movable src->mapping will be cleared by
1007 * free_pages_prepare so don't reset it here for keeping
1008 * the type to work PageAnon, for example.
1009 */
1010 if (!folio_mapping_flags(folio: src))
1011 src->mapping = NULL;
1012
1013 if (likely(!folio_is_zone_device(dst)))
1014 flush_dcache_folio(folio: dst);
1015 }
1016out:
1017 return rc;
1018}
1019
1020/*
1021 * To record some information during migration, we use some unused
1022 * fields (mapping and private) of struct folio of the newly allocated
1023 * destination folio. This is safe because nobody is using them
1024 * except us.
1025 */
1026union migration_ptr {
1027 struct anon_vma *anon_vma;
1028 struct address_space *mapping;
1029};
1030
1031enum {
1032 PAGE_WAS_MAPPED = BIT(0),
1033 PAGE_WAS_MLOCKED = BIT(1),
1034};
1035
1036static void __migrate_folio_record(struct folio *dst,
1037 unsigned long old_page_state,
1038 struct anon_vma *anon_vma)
1039{
1040 union migration_ptr ptr = { .anon_vma = anon_vma };
1041 dst->mapping = ptr.mapping;
1042 dst->private = (void *)old_page_state;
1043}
1044
1045static void __migrate_folio_extract(struct folio *dst,
1046 int *old_page_state,
1047 struct anon_vma **anon_vmap)
1048{
1049 union migration_ptr ptr = { .mapping = dst->mapping };
1050 *anon_vmap = ptr.anon_vma;
1051 *old_page_state = (unsigned long)dst->private;
1052 dst->mapping = NULL;
1053 dst->private = NULL;
1054}
1055
1056/* Restore the source folio to the original state upon failure */
1057static void migrate_folio_undo_src(struct folio *src,
1058 int page_was_mapped,
1059 struct anon_vma *anon_vma,
1060 bool locked,
1061 struct list_head *ret)
1062{
1063 if (page_was_mapped)
1064 remove_migration_ptes(src, dst: src, locked: false);
1065 /* Drop an anon_vma reference if we took one */
1066 if (anon_vma)
1067 put_anon_vma(anon_vma);
1068 if (locked)
1069 folio_unlock(folio: src);
1070 if (ret)
1071 list_move_tail(list: &src->lru, head: ret);
1072}
1073
1074/* Restore the destination folio to the original state upon failure */
1075static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1076 free_folio_t put_new_folio, unsigned long private)
1077{
1078 if (locked)
1079 folio_unlock(folio: dst);
1080 if (put_new_folio)
1081 put_new_folio(dst, private);
1082 else
1083 folio_put(folio: dst);
1084}
1085
1086/* Cleanup src folio upon migration success */
1087static void migrate_folio_done(struct folio *src,
1088 enum migrate_reason reason)
1089{
1090 /*
1091 * Compaction can migrate also non-LRU pages which are
1092 * not accounted to NR_ISOLATED_*. They can be recognized
1093 * as __folio_test_movable
1094 */
1095 if (likely(!__folio_test_movable(src)))
1096 mod_node_page_state(folio_pgdat(folio: src), NR_ISOLATED_ANON +
1097 folio_is_file_lru(folio: src), -folio_nr_pages(folio: src));
1098
1099 if (reason != MR_MEMORY_FAILURE)
1100 /* We release the page in page_handle_poison. */
1101 folio_put(folio: src);
1102}
1103
1104/* Obtain the lock on page, remove all ptes. */
1105static int migrate_folio_unmap(new_folio_t get_new_folio,
1106 free_folio_t put_new_folio, unsigned long private,
1107 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1108 enum migrate_reason reason, struct list_head *ret)
1109{
1110 struct folio *dst;
1111 int rc = -EAGAIN;
1112 int old_page_state = 0;
1113 struct anon_vma *anon_vma = NULL;
1114 bool is_lru = !__folio_test_movable(folio: src);
1115 bool locked = false;
1116 bool dst_locked = false;
1117
1118 if (folio_ref_count(folio: src) == 1) {
1119 /* Folio was freed from under us. So we are done. */
1120 folio_clear_active(folio: src);
1121 folio_clear_unevictable(folio: src);
1122 /* free_pages_prepare() will clear PG_isolated. */
1123 list_del(entry: &src->lru);
1124 migrate_folio_done(src, reason);
1125 return MIGRATEPAGE_SUCCESS;
1126 }
1127
1128 dst = get_new_folio(src, private);
1129 if (!dst)
1130 return -ENOMEM;
1131 *dstp = dst;
1132
1133 dst->private = NULL;
1134
1135 if (!folio_trylock(folio: src)) {
1136 if (mode == MIGRATE_ASYNC)
1137 goto out;
1138
1139 /*
1140 * It's not safe for direct compaction to call lock_page.
1141 * For example, during page readahead pages are added locked
1142 * to the LRU. Later, when the IO completes the pages are
1143 * marked uptodate and unlocked. However, the queueing
1144 * could be merging multiple pages for one bio (e.g.
1145 * mpage_readahead). If an allocation happens for the
1146 * second or third page, the process can end up locking
1147 * the same page twice and deadlocking. Rather than
1148 * trying to be clever about what pages can be locked,
1149 * avoid the use of lock_page for direct compaction
1150 * altogether.
1151 */
1152 if (current->flags & PF_MEMALLOC)
1153 goto out;
1154
1155 /*
1156 * In "light" mode, we can wait for transient locks (eg
1157 * inserting a page into the page table), but it's not
1158 * worth waiting for I/O.
1159 */
1160 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(folio: src))
1161 goto out;
1162
1163 folio_lock(folio: src);
1164 }
1165 locked = true;
1166 if (folio_test_mlocked(folio: src))
1167 old_page_state |= PAGE_WAS_MLOCKED;
1168
1169 if (folio_test_writeback(folio: src)) {
1170 /*
1171 * Only in the case of a full synchronous migration is it
1172 * necessary to wait for PageWriteback. In the async case,
1173 * the retry loop is too short and in the sync-light case,
1174 * the overhead of stalling is too much
1175 */
1176 switch (mode) {
1177 case MIGRATE_SYNC:
1178 case MIGRATE_SYNC_NO_COPY:
1179 break;
1180 default:
1181 rc = -EBUSY;
1182 goto out;
1183 }
1184 folio_wait_writeback(folio: src);
1185 }
1186
1187 /*
1188 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1189 * we cannot notice that anon_vma is freed while we migrate a page.
1190 * This get_anon_vma() delays freeing anon_vma pointer until the end
1191 * of migration. File cache pages are no problem because of page_lock()
1192 * File Caches may use write_page() or lock_page() in migration, then,
1193 * just care Anon page here.
1194 *
1195 * Only folio_get_anon_vma() understands the subtleties of
1196 * getting a hold on an anon_vma from outside one of its mms.
1197 * But if we cannot get anon_vma, then we won't need it anyway,
1198 * because that implies that the anon page is no longer mapped
1199 * (and cannot be remapped so long as we hold the page lock).
1200 */
1201 if (folio_test_anon(folio: src) && !folio_test_ksm(folio: src))
1202 anon_vma = folio_get_anon_vma(folio: src);
1203
1204 /*
1205 * Block others from accessing the new page when we get around to
1206 * establishing additional references. We are usually the only one
1207 * holding a reference to dst at this point. We used to have a BUG
1208 * here if folio_trylock(dst) fails, but would like to allow for
1209 * cases where there might be a race with the previous use of dst.
1210 * This is much like races on refcount of oldpage: just don't BUG().
1211 */
1212 if (unlikely(!folio_trylock(dst)))
1213 goto out;
1214 dst_locked = true;
1215
1216 if (unlikely(!is_lru)) {
1217 __migrate_folio_record(dst, old_page_state, anon_vma);
1218 return MIGRATEPAGE_UNMAP;
1219 }
1220
1221 /*
1222 * Corner case handling:
1223 * 1. When a new swap-cache page is read into, it is added to the LRU
1224 * and treated as swapcache but it has no rmap yet.
1225 * Calling try_to_unmap() against a src->mapping==NULL page will
1226 * trigger a BUG. So handle it here.
1227 * 2. An orphaned page (see truncate_cleanup_page) might have
1228 * fs-private metadata. The page can be picked up due to memory
1229 * offlining. Everywhere else except page reclaim, the page is
1230 * invisible to the vm, so the page can not be migrated. So try to
1231 * free the metadata, so the page can be freed.
1232 */
1233 if (!src->mapping) {
1234 if (folio_test_private(folio: src)) {
1235 try_to_free_buffers(folio: src);
1236 goto out;
1237 }
1238 } else if (folio_mapped(folio: src)) {
1239 /* Establish migration ptes */
1240 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1241 !folio_test_ksm(src) && !anon_vma, src);
1242 try_to_migrate(folio: src, flags: mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1243 old_page_state |= PAGE_WAS_MAPPED;
1244 }
1245
1246 if (!folio_mapped(folio: src)) {
1247 __migrate_folio_record(dst, old_page_state, anon_vma);
1248 return MIGRATEPAGE_UNMAP;
1249 }
1250
1251out:
1252 /*
1253 * A folio that has not been unmapped will be restored to
1254 * right list unless we want to retry.
1255 */
1256 if (rc == -EAGAIN)
1257 ret = NULL;
1258
1259 migrate_folio_undo_src(src, page_was_mapped: old_page_state & PAGE_WAS_MAPPED,
1260 anon_vma, locked, ret);
1261 migrate_folio_undo_dst(dst, locked: dst_locked, put_new_folio, private);
1262
1263 return rc;
1264}
1265
1266/* Migrate the folio to the newly allocated folio in dst. */
1267static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1268 struct folio *src, struct folio *dst,
1269 enum migrate_mode mode, enum migrate_reason reason,
1270 struct list_head *ret)
1271{
1272 int rc;
1273 int old_page_state = 0;
1274 struct anon_vma *anon_vma = NULL;
1275 bool is_lru = !__folio_test_movable(folio: src);
1276 struct list_head *prev;
1277
1278 __migrate_folio_extract(dst, old_page_state: &old_page_state, anon_vmap: &anon_vma);
1279 prev = dst->lru.prev;
1280 list_del(entry: &dst->lru);
1281
1282 rc = move_to_new_folio(dst, src, mode);
1283 if (rc)
1284 goto out;
1285
1286 if (unlikely(!is_lru))
1287 goto out_unlock_both;
1288
1289 /*
1290 * When successful, push dst to LRU immediately: so that if it
1291 * turns out to be an mlocked page, remove_migration_ptes() will
1292 * automatically build up the correct dst->mlock_count for it.
1293 *
1294 * We would like to do something similar for the old page, when
1295 * unsuccessful, and other cases when a page has been temporarily
1296 * isolated from the unevictable LRU: but this case is the easiest.
1297 */
1298 folio_add_lru(dst);
1299 if (old_page_state & PAGE_WAS_MLOCKED)
1300 lru_add_drain();
1301
1302 if (old_page_state & PAGE_WAS_MAPPED)
1303 remove_migration_ptes(src, dst, locked: false);
1304
1305out_unlock_both:
1306 folio_unlock(folio: dst);
1307 set_page_owner_migrate_reason(page: &dst->page, reason);
1308 /*
1309 * If migration is successful, decrease refcount of dst,
1310 * which will not free the page because new page owner increased
1311 * refcounter.
1312 */
1313 folio_put(folio: dst);
1314
1315 /*
1316 * A folio that has been migrated has all references removed
1317 * and will be freed.
1318 */
1319 list_del(entry: &src->lru);
1320 /* Drop an anon_vma reference if we took one */
1321 if (anon_vma)
1322 put_anon_vma(anon_vma);
1323 folio_unlock(folio: src);
1324 migrate_folio_done(src, reason);
1325
1326 return rc;
1327out:
1328 /*
1329 * A folio that has not been migrated will be restored to
1330 * right list unless we want to retry.
1331 */
1332 if (rc == -EAGAIN) {
1333 list_add(new: &dst->lru, head: prev);
1334 __migrate_folio_record(dst, old_page_state, anon_vma);
1335 return rc;
1336 }
1337
1338 migrate_folio_undo_src(src, page_was_mapped: old_page_state & PAGE_WAS_MAPPED,
1339 anon_vma, locked: true, ret);
1340 migrate_folio_undo_dst(dst, locked: true, put_new_folio, private);
1341
1342 return rc;
1343}
1344
1345/*
1346 * Counterpart of unmap_and_move_page() for hugepage migration.
1347 *
1348 * This function doesn't wait the completion of hugepage I/O
1349 * because there is no race between I/O and migration for hugepage.
1350 * Note that currently hugepage I/O occurs only in direct I/O
1351 * where no lock is held and PG_writeback is irrelevant,
1352 * and writeback status of all subpages are counted in the reference
1353 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1354 * under direct I/O, the reference of the head page is 512 and a bit more.)
1355 * This means that when we try to migrate hugepage whose subpages are
1356 * doing direct I/O, some references remain after try_to_unmap() and
1357 * hugepage migration fails without data corruption.
1358 *
1359 * There is also no race when direct I/O is issued on the page under migration,
1360 * because then pte is replaced with migration swap entry and direct I/O code
1361 * will wait in the page fault for migration to complete.
1362 */
1363static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1364 free_folio_t put_new_folio, unsigned long private,
1365 struct folio *src, int force, enum migrate_mode mode,
1366 int reason, struct list_head *ret)
1367{
1368 struct folio *dst;
1369 int rc = -EAGAIN;
1370 int page_was_mapped = 0;
1371 struct anon_vma *anon_vma = NULL;
1372 struct address_space *mapping = NULL;
1373
1374 if (folio_ref_count(folio: src) == 1) {
1375 /* page was freed from under us. So we are done. */
1376 folio_putback_active_hugetlb(folio: src);
1377 return MIGRATEPAGE_SUCCESS;
1378 }
1379
1380 dst = get_new_folio(src, private);
1381 if (!dst)
1382 return -ENOMEM;
1383
1384 if (!folio_trylock(folio: src)) {
1385 if (!force)
1386 goto out;
1387 switch (mode) {
1388 case MIGRATE_SYNC:
1389 case MIGRATE_SYNC_NO_COPY:
1390 break;
1391 default:
1392 goto out;
1393 }
1394 folio_lock(folio: src);
1395 }
1396
1397 /*
1398 * Check for pages which are in the process of being freed. Without
1399 * folio_mapping() set, hugetlbfs specific move page routine will not
1400 * be called and we could leak usage counts for subpools.
1401 */
1402 if (hugetlb_folio_subpool(folio: src) && !folio_mapping(src)) {
1403 rc = -EBUSY;
1404 goto out_unlock;
1405 }
1406
1407 if (folio_test_anon(folio: src))
1408 anon_vma = folio_get_anon_vma(folio: src);
1409
1410 if (unlikely(!folio_trylock(dst)))
1411 goto put_anon;
1412
1413 if (folio_mapped(folio: src)) {
1414 enum ttu_flags ttu = 0;
1415
1416 if (!folio_test_anon(folio: src)) {
1417 /*
1418 * In shared mappings, try_to_unmap could potentially
1419 * call huge_pmd_unshare. Because of this, take
1420 * semaphore in write mode here and set TTU_RMAP_LOCKED
1421 * to let lower levels know we have taken the lock.
1422 */
1423 mapping = hugetlb_page_mapping_lock_write(hpage: &src->page);
1424 if (unlikely(!mapping))
1425 goto unlock_put_anon;
1426
1427 ttu = TTU_RMAP_LOCKED;
1428 }
1429
1430 try_to_migrate(folio: src, flags: ttu);
1431 page_was_mapped = 1;
1432
1433 if (ttu & TTU_RMAP_LOCKED)
1434 i_mmap_unlock_write(mapping);
1435 }
1436
1437 if (!folio_mapped(folio: src))
1438 rc = move_to_new_folio(dst, src, mode);
1439
1440 if (page_was_mapped)
1441 remove_migration_ptes(src,
1442 dst: rc == MIGRATEPAGE_SUCCESS ? dst : src, locked: false);
1443
1444unlock_put_anon:
1445 folio_unlock(folio: dst);
1446
1447put_anon:
1448 if (anon_vma)
1449 put_anon_vma(anon_vma);
1450
1451 if (rc == MIGRATEPAGE_SUCCESS) {
1452 move_hugetlb_state(old_folio: src, new_folio: dst, reason);
1453 put_new_folio = NULL;
1454 }
1455
1456out_unlock:
1457 folio_unlock(folio: src);
1458out:
1459 if (rc == MIGRATEPAGE_SUCCESS)
1460 folio_putback_active_hugetlb(folio: src);
1461 else if (rc != -EAGAIN)
1462 list_move_tail(list: &src->lru, head: ret);
1463
1464 /*
1465 * If migration was not successful and there's a freeing callback, use
1466 * it. Otherwise, put_page() will drop the reference grabbed during
1467 * isolation.
1468 */
1469 if (put_new_folio)
1470 put_new_folio(dst, private);
1471 else
1472 folio_putback_active_hugetlb(folio: dst);
1473
1474 return rc;
1475}
1476
1477static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1478{
1479 int rc;
1480
1481 folio_lock(folio);
1482 rc = split_folio_to_list(folio, list: split_folios);
1483 folio_unlock(folio);
1484 if (!rc)
1485 list_move_tail(list: &folio->lru, head: split_folios);
1486
1487 return rc;
1488}
1489
1490#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1491#define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1492#else
1493#define NR_MAX_BATCHED_MIGRATION 512
1494#endif
1495#define NR_MAX_MIGRATE_PAGES_RETRY 10
1496#define NR_MAX_MIGRATE_ASYNC_RETRY 3
1497#define NR_MAX_MIGRATE_SYNC_RETRY \
1498 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1499
1500struct migrate_pages_stats {
1501 int nr_succeeded; /* Normal and large folios migrated successfully, in
1502 units of base pages */
1503 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1504 units of base pages. Untried folios aren't counted */
1505 int nr_thp_succeeded; /* THP migrated successfully */
1506 int nr_thp_failed; /* THP failed to be migrated */
1507 int nr_thp_split; /* THP split before migrating */
1508 int nr_split; /* Large folio (include THP) split before migrating */
1509};
1510
1511/*
1512 * Returns the number of hugetlb folios that were not migrated, or an error code
1513 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1514 * any more because the list has become empty or no retryable hugetlb folios
1515 * exist any more. It is caller's responsibility to call putback_movable_pages()
1516 * only if ret != 0.
1517 */
1518static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1519 free_folio_t put_new_folio, unsigned long private,
1520 enum migrate_mode mode, int reason,
1521 struct migrate_pages_stats *stats,
1522 struct list_head *ret_folios)
1523{
1524 int retry = 1;
1525 int nr_failed = 0;
1526 int nr_retry_pages = 0;
1527 int pass = 0;
1528 struct folio *folio, *folio2;
1529 int rc, nr_pages;
1530
1531 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1532 retry = 0;
1533 nr_retry_pages = 0;
1534
1535 list_for_each_entry_safe(folio, folio2, from, lru) {
1536 if (!folio_test_hugetlb(folio))
1537 continue;
1538
1539 nr_pages = folio_nr_pages(folio);
1540
1541 cond_resched();
1542
1543 /*
1544 * Migratability of hugepages depends on architectures and
1545 * their size. This check is necessary because some callers
1546 * of hugepage migration like soft offline and memory
1547 * hotremove don't walk through page tables or check whether
1548 * the hugepage is pmd-based or not before kicking migration.
1549 */
1550 if (!hugepage_migration_supported(h: folio_hstate(folio))) {
1551 nr_failed++;
1552 stats->nr_failed_pages += nr_pages;
1553 list_move_tail(list: &folio->lru, head: ret_folios);
1554 continue;
1555 }
1556
1557 rc = unmap_and_move_huge_page(get_new_folio,
1558 put_new_folio, private,
1559 src: folio, force: pass > 2, mode,
1560 reason, ret: ret_folios);
1561 /*
1562 * The rules are:
1563 * Success: hugetlb folio will be put back
1564 * -EAGAIN: stay on the from list
1565 * -ENOMEM: stay on the from list
1566 * Other errno: put on ret_folios list
1567 */
1568 switch(rc) {
1569 case -ENOMEM:
1570 /*
1571 * When memory is low, don't bother to try to migrate
1572 * other folios, just exit.
1573 */
1574 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1575 return -ENOMEM;
1576 case -EAGAIN:
1577 retry++;
1578 nr_retry_pages += nr_pages;
1579 break;
1580 case MIGRATEPAGE_SUCCESS:
1581 stats->nr_succeeded += nr_pages;
1582 break;
1583 default:
1584 /*
1585 * Permanent failure (-EBUSY, etc.):
1586 * unlike -EAGAIN case, the failed folio is
1587 * removed from migration folio list and not
1588 * retried in the next outer loop.
1589 */
1590 nr_failed++;
1591 stats->nr_failed_pages += nr_pages;
1592 break;
1593 }
1594 }
1595 }
1596 /*
1597 * nr_failed is number of hugetlb folios failed to be migrated. After
1598 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1599 * folios as failed.
1600 */
1601 nr_failed += retry;
1602 stats->nr_failed_pages += nr_retry_pages;
1603
1604 return nr_failed;
1605}
1606
1607/*
1608 * migrate_pages_batch() first unmaps folios in the from list as many as
1609 * possible, then move the unmapped folios.
1610 *
1611 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1612 * lock or bit when we have locked more than one folio. Which may cause
1613 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1614 * length of the from list must be <= 1.
1615 */
1616static int migrate_pages_batch(struct list_head *from,
1617 new_folio_t get_new_folio, free_folio_t put_new_folio,
1618 unsigned long private, enum migrate_mode mode, int reason,
1619 struct list_head *ret_folios, struct list_head *split_folios,
1620 struct migrate_pages_stats *stats, int nr_pass)
1621{
1622 int retry = 1;
1623 int thp_retry = 1;
1624 int nr_failed = 0;
1625 int nr_retry_pages = 0;
1626 int pass = 0;
1627 bool is_thp = false;
1628 bool is_large = false;
1629 struct folio *folio, *folio2, *dst = NULL, *dst2;
1630 int rc, rc_saved = 0, nr_pages;
1631 LIST_HEAD(unmap_folios);
1632 LIST_HEAD(dst_folios);
1633 bool nosplit = (reason == MR_NUMA_MISPLACED);
1634
1635 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1636 !list_empty(from) && !list_is_singular(from));
1637
1638 for (pass = 0; pass < nr_pass && retry; pass++) {
1639 retry = 0;
1640 thp_retry = 0;
1641 nr_retry_pages = 0;
1642
1643 list_for_each_entry_safe(folio, folio2, from, lru) {
1644 is_large = folio_test_large(folio);
1645 is_thp = is_large && folio_test_pmd_mappable(folio);
1646 nr_pages = folio_nr_pages(folio);
1647
1648 cond_resched();
1649
1650 /*
1651 * Large folio migration might be unsupported or
1652 * the allocation might be failed so we should retry
1653 * on the same folio with the large folio split
1654 * to normal folios.
1655 *
1656 * Split folios are put in split_folios, and
1657 * we will migrate them after the rest of the
1658 * list is processed.
1659 */
1660 if (!thp_migration_supported() && is_thp) {
1661 nr_failed++;
1662 stats->nr_thp_failed++;
1663 if (!try_split_folio(folio, split_folios)) {
1664 stats->nr_thp_split++;
1665 stats->nr_split++;
1666 continue;
1667 }
1668 stats->nr_failed_pages += nr_pages;
1669 list_move_tail(list: &folio->lru, head: ret_folios);
1670 continue;
1671 }
1672
1673 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1674 private, src: folio, dstp: &dst, mode, reason,
1675 ret: ret_folios);
1676 /*
1677 * The rules are:
1678 * Success: folio will be freed
1679 * Unmap: folio will be put on unmap_folios list,
1680 * dst folio put on dst_folios list
1681 * -EAGAIN: stay on the from list
1682 * -ENOMEM: stay on the from list
1683 * Other errno: put on ret_folios list
1684 */
1685 switch(rc) {
1686 case -ENOMEM:
1687 /*
1688 * When memory is low, don't bother to try to migrate
1689 * other folios, move unmapped folios, then exit.
1690 */
1691 nr_failed++;
1692 stats->nr_thp_failed += is_thp;
1693 /* Large folio NUMA faulting doesn't split to retry. */
1694 if (is_large && !nosplit) {
1695 int ret = try_split_folio(folio, split_folios);
1696
1697 if (!ret) {
1698 stats->nr_thp_split += is_thp;
1699 stats->nr_split++;
1700 break;
1701 } else if (reason == MR_LONGTERM_PIN &&
1702 ret == -EAGAIN) {
1703 /*
1704 * Try again to split large folio to
1705 * mitigate the failure of longterm pinning.
1706 */
1707 retry++;
1708 thp_retry += is_thp;
1709 nr_retry_pages += nr_pages;
1710 /* Undo duplicated failure counting. */
1711 nr_failed--;
1712 stats->nr_thp_failed -= is_thp;
1713 break;
1714 }
1715 }
1716
1717 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1718 /* nr_failed isn't updated for not used */
1719 stats->nr_thp_failed += thp_retry;
1720 rc_saved = rc;
1721 if (list_empty(head: &unmap_folios))
1722 goto out;
1723 else
1724 goto move;
1725 case -EAGAIN:
1726 retry++;
1727 thp_retry += is_thp;
1728 nr_retry_pages += nr_pages;
1729 break;
1730 case MIGRATEPAGE_SUCCESS:
1731 stats->nr_succeeded += nr_pages;
1732 stats->nr_thp_succeeded += is_thp;
1733 break;
1734 case MIGRATEPAGE_UNMAP:
1735 list_move_tail(list: &folio->lru, head: &unmap_folios);
1736 list_add_tail(new: &dst->lru, head: &dst_folios);
1737 break;
1738 default:
1739 /*
1740 * Permanent failure (-EBUSY, etc.):
1741 * unlike -EAGAIN case, the failed folio is
1742 * removed from migration folio list and not
1743 * retried in the next outer loop.
1744 */
1745 nr_failed++;
1746 stats->nr_thp_failed += is_thp;
1747 stats->nr_failed_pages += nr_pages;
1748 break;
1749 }
1750 }
1751 }
1752 nr_failed += retry;
1753 stats->nr_thp_failed += thp_retry;
1754 stats->nr_failed_pages += nr_retry_pages;
1755move:
1756 /* Flush TLBs for all unmapped folios */
1757 try_to_unmap_flush();
1758
1759 retry = 1;
1760 for (pass = 0; pass < nr_pass && retry; pass++) {
1761 retry = 0;
1762 thp_retry = 0;
1763 nr_retry_pages = 0;
1764
1765 dst = list_first_entry(&dst_folios, struct folio, lru);
1766 dst2 = list_next_entry(dst, lru);
1767 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1768 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1769 nr_pages = folio_nr_pages(folio);
1770
1771 cond_resched();
1772
1773 rc = migrate_folio_move(put_new_folio, private,
1774 src: folio, dst, mode,
1775 reason, ret: ret_folios);
1776 /*
1777 * The rules are:
1778 * Success: folio will be freed
1779 * -EAGAIN: stay on the unmap_folios list
1780 * Other errno: put on ret_folios list
1781 */
1782 switch(rc) {
1783 case -EAGAIN:
1784 retry++;
1785 thp_retry += is_thp;
1786 nr_retry_pages += nr_pages;
1787 break;
1788 case MIGRATEPAGE_SUCCESS:
1789 stats->nr_succeeded += nr_pages;
1790 stats->nr_thp_succeeded += is_thp;
1791 break;
1792 default:
1793 nr_failed++;
1794 stats->nr_thp_failed += is_thp;
1795 stats->nr_failed_pages += nr_pages;
1796 break;
1797 }
1798 dst = dst2;
1799 dst2 = list_next_entry(dst, lru);
1800 }
1801 }
1802 nr_failed += retry;
1803 stats->nr_thp_failed += thp_retry;
1804 stats->nr_failed_pages += nr_retry_pages;
1805
1806 rc = rc_saved ? : nr_failed;
1807out:
1808 /* Cleanup remaining folios */
1809 dst = list_first_entry(&dst_folios, struct folio, lru);
1810 dst2 = list_next_entry(dst, lru);
1811 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1812 int old_page_state = 0;
1813 struct anon_vma *anon_vma = NULL;
1814
1815 __migrate_folio_extract(dst, old_page_state: &old_page_state, anon_vmap: &anon_vma);
1816 migrate_folio_undo_src(src: folio, page_was_mapped: old_page_state & PAGE_WAS_MAPPED,
1817 anon_vma, locked: true, ret: ret_folios);
1818 list_del(entry: &dst->lru);
1819 migrate_folio_undo_dst(dst, locked: true, put_new_folio, private);
1820 dst = dst2;
1821 dst2 = list_next_entry(dst, lru);
1822 }
1823
1824 return rc;
1825}
1826
1827static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1828 free_folio_t put_new_folio, unsigned long private,
1829 enum migrate_mode mode, int reason,
1830 struct list_head *ret_folios, struct list_head *split_folios,
1831 struct migrate_pages_stats *stats)
1832{
1833 int rc, nr_failed = 0;
1834 LIST_HEAD(folios);
1835 struct migrate_pages_stats astats;
1836
1837 memset(&astats, 0, sizeof(astats));
1838 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1839 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, mode: MIGRATE_ASYNC,
1840 reason, ret_folios: &folios, split_folios, stats: &astats,
1841 NR_MAX_MIGRATE_ASYNC_RETRY);
1842 stats->nr_succeeded += astats.nr_succeeded;
1843 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1844 stats->nr_thp_split += astats.nr_thp_split;
1845 stats->nr_split += astats.nr_split;
1846 if (rc < 0) {
1847 stats->nr_failed_pages += astats.nr_failed_pages;
1848 stats->nr_thp_failed += astats.nr_thp_failed;
1849 list_splice_tail(list: &folios, head: ret_folios);
1850 return rc;
1851 }
1852 stats->nr_thp_failed += astats.nr_thp_split;
1853 /*
1854 * Do not count rc, as pages will be retried below.
1855 * Count nr_split only, since it includes nr_thp_split.
1856 */
1857 nr_failed += astats.nr_split;
1858 /*
1859 * Fall back to migrate all failed folios one by one synchronously. All
1860 * failed folios except split THPs will be retried, so their failure
1861 * isn't counted
1862 */
1863 list_splice_tail_init(list: &folios, head: from);
1864 while (!list_empty(head: from)) {
1865 list_move(list: from->next, head: &folios);
1866 rc = migrate_pages_batch(from: &folios, get_new_folio, put_new_folio,
1867 private, mode, reason, ret_folios,
1868 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1869 list_splice_tail_init(list: &folios, head: ret_folios);
1870 if (rc < 0)
1871 return rc;
1872 nr_failed += rc;
1873 }
1874
1875 return nr_failed;
1876}
1877
1878/*
1879 * migrate_pages - migrate the folios specified in a list, to the free folios
1880 * supplied as the target for the page migration
1881 *
1882 * @from: The list of folios to be migrated.
1883 * @get_new_folio: The function used to allocate free folios to be used
1884 * as the target of the folio migration.
1885 * @put_new_folio: The function used to free target folios if migration
1886 * fails, or NULL if no special handling is necessary.
1887 * @private: Private data to be passed on to get_new_folio()
1888 * @mode: The migration mode that specifies the constraints for
1889 * folio migration, if any.
1890 * @reason: The reason for folio migration.
1891 * @ret_succeeded: Set to the number of folios migrated successfully if
1892 * the caller passes a non-NULL pointer.
1893 *
1894 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1895 * are movable any more because the list has become empty or no retryable folios
1896 * exist any more. It is caller's responsibility to call putback_movable_pages()
1897 * only if ret != 0.
1898 *
1899 * Returns the number of {normal folio, large folio, hugetlb} that were not
1900 * migrated, or an error code. The number of large folio splits will be
1901 * considered as the number of non-migrated large folio, no matter how many
1902 * split folios of the large folio are migrated successfully.
1903 */
1904int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1905 free_folio_t put_new_folio, unsigned long private,
1906 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1907{
1908 int rc, rc_gather;
1909 int nr_pages;
1910 struct folio *folio, *folio2;
1911 LIST_HEAD(folios);
1912 LIST_HEAD(ret_folios);
1913 LIST_HEAD(split_folios);
1914 struct migrate_pages_stats stats;
1915
1916 trace_mm_migrate_pages_start(mode, reason);
1917
1918 memset(&stats, 0, sizeof(stats));
1919
1920 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1921 mode, reason, stats: &stats, ret_folios: &ret_folios);
1922 if (rc_gather < 0)
1923 goto out;
1924
1925again:
1926 nr_pages = 0;
1927 list_for_each_entry_safe(folio, folio2, from, lru) {
1928 /* Retried hugetlb folios will be kept in list */
1929 if (folio_test_hugetlb(folio)) {
1930 list_move_tail(list: &folio->lru, head: &ret_folios);
1931 continue;
1932 }
1933
1934 nr_pages += folio_nr_pages(folio);
1935 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1936 break;
1937 }
1938 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1939 list_cut_before(list: &folios, head: from, entry: &folio2->lru);
1940 else
1941 list_splice_init(list: from, head: &folios);
1942 if (mode == MIGRATE_ASYNC)
1943 rc = migrate_pages_batch(from: &folios, get_new_folio, put_new_folio,
1944 private, mode, reason, ret_folios: &ret_folios,
1945 split_folios: &split_folios, stats: &stats,
1946 NR_MAX_MIGRATE_PAGES_RETRY);
1947 else
1948 rc = migrate_pages_sync(from: &folios, get_new_folio, put_new_folio,
1949 private, mode, reason, ret_folios: &ret_folios,
1950 split_folios: &split_folios, stats: &stats);
1951 list_splice_tail_init(list: &folios, head: &ret_folios);
1952 if (rc < 0) {
1953 rc_gather = rc;
1954 list_splice_tail(list: &split_folios, head: &ret_folios);
1955 goto out;
1956 }
1957 if (!list_empty(head: &split_folios)) {
1958 /*
1959 * Failure isn't counted since all split folios of a large folio
1960 * is counted as 1 failure already. And, we only try to migrate
1961 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1962 */
1963 migrate_pages_batch(from: &split_folios, get_new_folio,
1964 put_new_folio, private, mode: MIGRATE_ASYNC, reason,
1965 ret_folios: &ret_folios, NULL, stats: &stats, nr_pass: 1);
1966 list_splice_tail_init(list: &split_folios, head: &ret_folios);
1967 }
1968 rc_gather += rc;
1969 if (!list_empty(head: from))
1970 goto again;
1971out:
1972 /*
1973 * Put the permanent failure folio back to migration list, they
1974 * will be put back to the right list by the caller.
1975 */
1976 list_splice(list: &ret_folios, head: from);
1977
1978 /*
1979 * Return 0 in case all split folios of fail-to-migrate large folios
1980 * are migrated successfully.
1981 */
1982 if (list_empty(head: from))
1983 rc_gather = 0;
1984
1985 count_vm_events(item: PGMIGRATE_SUCCESS, delta: stats.nr_succeeded);
1986 count_vm_events(item: PGMIGRATE_FAIL, delta: stats.nr_failed_pages);
1987 count_vm_events(item: THP_MIGRATION_SUCCESS, delta: stats.nr_thp_succeeded);
1988 count_vm_events(item: THP_MIGRATION_FAIL, delta: stats.nr_thp_failed);
1989 count_vm_events(item: THP_MIGRATION_SPLIT, delta: stats.nr_thp_split);
1990 trace_mm_migrate_pages(succeeded: stats.nr_succeeded, failed: stats.nr_failed_pages,
1991 thp_succeeded: stats.nr_thp_succeeded, thp_failed: stats.nr_thp_failed,
1992 thp_split: stats.nr_thp_split, large_folio_split: stats.nr_split, mode,
1993 reason);
1994
1995 if (ret_succeeded)
1996 *ret_succeeded = stats.nr_succeeded;
1997
1998 return rc_gather;
1999}
2000
2001struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2002{
2003 struct migration_target_control *mtc;
2004 gfp_t gfp_mask;
2005 unsigned int order = 0;
2006 int nid;
2007 int zidx;
2008
2009 mtc = (struct migration_target_control *)private;
2010 gfp_mask = mtc->gfp_mask;
2011 nid = mtc->nid;
2012 if (nid == NUMA_NO_NODE)
2013 nid = folio_nid(folio: src);
2014
2015 if (folio_test_hugetlb(folio: src)) {
2016 struct hstate *h = folio_hstate(folio: src);
2017
2018 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2019 return alloc_hugetlb_folio_nodemask(h, preferred_nid: nid,
2020 nmask: mtc->nmask, gfp_mask);
2021 }
2022
2023 if (folio_test_large(folio: src)) {
2024 /*
2025 * clear __GFP_RECLAIM to make the migration callback
2026 * consistent with regular THP allocations.
2027 */
2028 gfp_mask &= ~__GFP_RECLAIM;
2029 gfp_mask |= GFP_TRANSHUGE;
2030 order = folio_order(folio: src);
2031 }
2032 zidx = zone_idx(folio_zone(src));
2033 if (is_highmem_idx(idx: zidx) || zidx == ZONE_MOVABLE)
2034 gfp_mask |= __GFP_HIGHMEM;
2035
2036 return __folio_alloc(gfp: gfp_mask, order, preferred_nid: nid, nodemask: mtc->nmask);
2037}
2038
2039#ifdef CONFIG_NUMA
2040
2041static int store_status(int __user *status, int start, int value, int nr)
2042{
2043 while (nr-- > 0) {
2044 if (put_user(value, status + start))
2045 return -EFAULT;
2046 start++;
2047 }
2048
2049 return 0;
2050}
2051
2052static int do_move_pages_to_node(struct list_head *pagelist, int node)
2053{
2054 int err;
2055 struct migration_target_control mtc = {
2056 .nid = node,
2057 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2058 };
2059
2060 err = migrate_pages(from: pagelist, get_new_folio: alloc_migration_target, NULL,
2061 private: (unsigned long)&mtc, mode: MIGRATE_SYNC, reason: MR_SYSCALL, NULL);
2062 if (err)
2063 putback_movable_pages(l: pagelist);
2064 return err;
2065}
2066
2067/*
2068 * Resolves the given address to a struct page, isolates it from the LRU and
2069 * puts it to the given pagelist.
2070 * Returns:
2071 * errno - if the page cannot be found/isolated
2072 * 0 - when it doesn't have to be migrated because it is already on the
2073 * target node
2074 * 1 - when it has been queued
2075 */
2076static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2077 int node, struct list_head *pagelist, bool migrate_all)
2078{
2079 struct vm_area_struct *vma;
2080 unsigned long addr;
2081 struct page *page;
2082 struct folio *folio;
2083 int err;
2084
2085 mmap_read_lock(mm);
2086 addr = (unsigned long)untagged_addr_remote(mm, p);
2087
2088 err = -EFAULT;
2089 vma = vma_lookup(mm, addr);
2090 if (!vma || !vma_migratable(vma))
2091 goto out;
2092
2093 /* FOLL_DUMP to ignore special (like zero) pages */
2094 page = follow_page(vma, address: addr, foll_flags: FOLL_GET | FOLL_DUMP);
2095
2096 err = PTR_ERR(ptr: page);
2097 if (IS_ERR(ptr: page))
2098 goto out;
2099
2100 err = -ENOENT;
2101 if (!page)
2102 goto out;
2103
2104 folio = page_folio(page);
2105 if (folio_is_zone_device(folio))
2106 goto out_putfolio;
2107
2108 err = 0;
2109 if (folio_nid(folio) == node)
2110 goto out_putfolio;
2111
2112 err = -EACCES;
2113 if (page_mapcount(page) > 1 && !migrate_all)
2114 goto out_putfolio;
2115
2116 err = -EBUSY;
2117 if (folio_test_hugetlb(folio)) {
2118 if (isolate_hugetlb(folio, list: pagelist))
2119 err = 1;
2120 } else {
2121 if (!folio_isolate_lru(folio))
2122 goto out_putfolio;
2123
2124 err = 1;
2125 list_add_tail(new: &folio->lru, head: pagelist);
2126 node_stat_mod_folio(folio,
2127 item: NR_ISOLATED_ANON + folio_is_file_lru(folio),
2128 nr: folio_nr_pages(folio));
2129 }
2130out_putfolio:
2131 /*
2132 * Either remove the duplicate refcount from folio_isolate_lru()
2133 * or drop the folio ref if it was not isolated.
2134 */
2135 folio_put(folio);
2136out:
2137 mmap_read_unlock(mm);
2138 return err;
2139}
2140
2141static int move_pages_and_store_status(int node,
2142 struct list_head *pagelist, int __user *status,
2143 int start, int i, unsigned long nr_pages)
2144{
2145 int err;
2146
2147 if (list_empty(head: pagelist))
2148 return 0;
2149
2150 err = do_move_pages_to_node(pagelist, node);
2151 if (err) {
2152 /*
2153 * Positive err means the number of failed
2154 * pages to migrate. Since we are going to
2155 * abort and return the number of non-migrated
2156 * pages, so need to include the rest of the
2157 * nr_pages that have not been attempted as
2158 * well.
2159 */
2160 if (err > 0)
2161 err += nr_pages - i;
2162 return err;
2163 }
2164 return store_status(status, start, value: node, nr: i - start);
2165}
2166
2167/*
2168 * Migrate an array of page address onto an array of nodes and fill
2169 * the corresponding array of status.
2170 */
2171static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2172 unsigned long nr_pages,
2173 const void __user * __user *pages,
2174 const int __user *nodes,
2175 int __user *status, int flags)
2176{
2177 compat_uptr_t __user *compat_pages = (void __user *)pages;
2178 int current_node = NUMA_NO_NODE;
2179 LIST_HEAD(pagelist);
2180 int start, i;
2181 int err = 0, err1;
2182
2183 lru_cache_disable();
2184
2185 for (i = start = 0; i < nr_pages; i++) {
2186 const void __user *p;
2187 int node;
2188
2189 err = -EFAULT;
2190 if (in_compat_syscall()) {
2191 compat_uptr_t cp;
2192
2193 if (get_user(cp, compat_pages + i))
2194 goto out_flush;
2195
2196 p = compat_ptr(uptr: cp);
2197 } else {
2198 if (get_user(p, pages + i))
2199 goto out_flush;
2200 }
2201 if (get_user(node, nodes + i))
2202 goto out_flush;
2203
2204 err = -ENODEV;
2205 if (node < 0 || node >= MAX_NUMNODES)
2206 goto out_flush;
2207 if (!node_state(node, state: N_MEMORY))
2208 goto out_flush;
2209
2210 err = -EACCES;
2211 if (!node_isset(node, task_nodes))
2212 goto out_flush;
2213
2214 if (current_node == NUMA_NO_NODE) {
2215 current_node = node;
2216 start = i;
2217 } else if (node != current_node) {
2218 err = move_pages_and_store_status(node: current_node,
2219 pagelist: &pagelist, status, start, i, nr_pages);
2220 if (err)
2221 goto out;
2222 start = i;
2223 current_node = node;
2224 }
2225
2226 /*
2227 * Errors in the page lookup or isolation are not fatal and we simply
2228 * report them via status
2229 */
2230 err = add_page_for_migration(mm, p, node: current_node, pagelist: &pagelist,
2231 migrate_all: flags & MPOL_MF_MOVE_ALL);
2232
2233 if (err > 0) {
2234 /* The page is successfully queued for migration */
2235 continue;
2236 }
2237
2238 /*
2239 * The move_pages() man page does not have an -EEXIST choice, so
2240 * use -EFAULT instead.
2241 */
2242 if (err == -EEXIST)
2243 err = -EFAULT;
2244
2245 /*
2246 * If the page is already on the target node (!err), store the
2247 * node, otherwise, store the err.
2248 */
2249 err = store_status(status, start: i, value: err ? : current_node, nr: 1);
2250 if (err)
2251 goto out_flush;
2252
2253 err = move_pages_and_store_status(node: current_node, pagelist: &pagelist,
2254 status, start, i, nr_pages);
2255 if (err) {
2256 /* We have accounted for page i */
2257 if (err > 0)
2258 err--;
2259 goto out;
2260 }
2261 current_node = NUMA_NO_NODE;
2262 }
2263out_flush:
2264 /* Make sure we do not overwrite the existing error */
2265 err1 = move_pages_and_store_status(node: current_node, pagelist: &pagelist,
2266 status, start, i, nr_pages);
2267 if (err >= 0)
2268 err = err1;
2269out:
2270 lru_cache_enable();
2271 return err;
2272}
2273
2274/*
2275 * Determine the nodes of an array of pages and store it in an array of status.
2276 */
2277static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2278 const void __user **pages, int *status)
2279{
2280 unsigned long i;
2281
2282 mmap_read_lock(mm);
2283
2284 for (i = 0; i < nr_pages; i++) {
2285 unsigned long addr = (unsigned long)(*pages);
2286 struct vm_area_struct *vma;
2287 struct page *page;
2288 int err = -EFAULT;
2289
2290 vma = vma_lookup(mm, addr);
2291 if (!vma)
2292 goto set_status;
2293
2294 /* FOLL_DUMP to ignore special (like zero) pages */
2295 page = follow_page(vma, address: addr, foll_flags: FOLL_GET | FOLL_DUMP);
2296
2297 err = PTR_ERR(ptr: page);
2298 if (IS_ERR(ptr: page))
2299 goto set_status;
2300
2301 err = -ENOENT;
2302 if (!page)
2303 goto set_status;
2304
2305 if (!is_zone_device_page(page))
2306 err = page_to_nid(page);
2307
2308 put_page(page);
2309set_status:
2310 *status = err;
2311
2312 pages++;
2313 status++;
2314 }
2315
2316 mmap_read_unlock(mm);
2317}
2318
2319static int get_compat_pages_array(const void __user *chunk_pages[],
2320 const void __user * __user *pages,
2321 unsigned long chunk_nr)
2322{
2323 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2324 compat_uptr_t p;
2325 int i;
2326
2327 for (i = 0; i < chunk_nr; i++) {
2328 if (get_user(p, pages32 + i))
2329 return -EFAULT;
2330 chunk_pages[i] = compat_ptr(uptr: p);
2331 }
2332
2333 return 0;
2334}
2335
2336/*
2337 * Determine the nodes of a user array of pages and store it in
2338 * a user array of status.
2339 */
2340static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2341 const void __user * __user *pages,
2342 int __user *status)
2343{
2344#define DO_PAGES_STAT_CHUNK_NR 16UL
2345 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2346 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2347
2348 while (nr_pages) {
2349 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2350
2351 if (in_compat_syscall()) {
2352 if (get_compat_pages_array(chunk_pages, pages,
2353 chunk_nr))
2354 break;
2355 } else {
2356 if (copy_from_user(to: chunk_pages, from: pages,
2357 n: chunk_nr * sizeof(*chunk_pages)))
2358 break;
2359 }
2360
2361 do_pages_stat_array(mm, nr_pages: chunk_nr, pages: chunk_pages, status: chunk_status);
2362
2363 if (copy_to_user(to: status, from: chunk_status, n: chunk_nr * sizeof(*status)))
2364 break;
2365
2366 pages += chunk_nr;
2367 status += chunk_nr;
2368 nr_pages -= chunk_nr;
2369 }
2370 return nr_pages ? -EFAULT : 0;
2371}
2372
2373static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2374{
2375 struct task_struct *task;
2376 struct mm_struct *mm;
2377
2378 /*
2379 * There is no need to check if current process has the right to modify
2380 * the specified process when they are same.
2381 */
2382 if (!pid) {
2383 mmget(current->mm);
2384 *mem_nodes = cpuset_mems_allowed(current);
2385 return current->mm;
2386 }
2387
2388 /* Find the mm_struct */
2389 rcu_read_lock();
2390 task = find_task_by_vpid(nr: pid);
2391 if (!task) {
2392 rcu_read_unlock();
2393 return ERR_PTR(error: -ESRCH);
2394 }
2395 get_task_struct(t: task);
2396
2397 /*
2398 * Check if this process has the right to modify the specified
2399 * process. Use the regular "ptrace_may_access()" checks.
2400 */
2401 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2402 rcu_read_unlock();
2403 mm = ERR_PTR(error: -EPERM);
2404 goto out;
2405 }
2406 rcu_read_unlock();
2407
2408 mm = ERR_PTR(error: security_task_movememory(p: task));
2409 if (IS_ERR(ptr: mm))
2410 goto out;
2411 *mem_nodes = cpuset_mems_allowed(p: task);
2412 mm = get_task_mm(task);
2413out:
2414 put_task_struct(t: task);
2415 if (!mm)
2416 mm = ERR_PTR(error: -EINVAL);
2417 return mm;
2418}
2419
2420/*
2421 * Move a list of pages in the address space of the currently executing
2422 * process.
2423 */
2424static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2425 const void __user * __user *pages,
2426 const int __user *nodes,
2427 int __user *status, int flags)
2428{
2429 struct mm_struct *mm;
2430 int err;
2431 nodemask_t task_nodes;
2432
2433 /* Check flags */
2434 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2435 return -EINVAL;
2436
2437 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2438 return -EPERM;
2439
2440 mm = find_mm_struct(pid, mem_nodes: &task_nodes);
2441 if (IS_ERR(ptr: mm))
2442 return PTR_ERR(ptr: mm);
2443
2444 if (nodes)
2445 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2446 nodes, status, flags);
2447 else
2448 err = do_pages_stat(mm, nr_pages, pages, status);
2449
2450 mmput(mm);
2451 return err;
2452}
2453
2454SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2455 const void __user * __user *, pages,
2456 const int __user *, nodes,
2457 int __user *, status, int, flags)
2458{
2459 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2460}
2461
2462#ifdef CONFIG_NUMA_BALANCING
2463/*
2464 * Returns true if this is a safe migration target node for misplaced NUMA
2465 * pages. Currently it only checks the watermarks which is crude.
2466 */
2467static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2468 unsigned long nr_migrate_pages)
2469{
2470 int z;
2471
2472 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2473 struct zone *zone = pgdat->node_zones + z;
2474
2475 if (!managed_zone(zone))
2476 continue;
2477
2478 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2479 if (!zone_watermark_ok(z: zone, order: 0,
2480 high_wmark_pages(zone) +
2481 nr_migrate_pages,
2482 highest_zoneidx: ZONE_MOVABLE, alloc_flags: 0))
2483 continue;
2484 return true;
2485 }
2486 return false;
2487}
2488
2489static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2490 unsigned long data)
2491{
2492 int nid = (int) data;
2493 int order = folio_order(folio: src);
2494 gfp_t gfp = __GFP_THISNODE;
2495
2496 if (order > 0)
2497 gfp |= GFP_TRANSHUGE_LIGHT;
2498 else {
2499 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2500 __GFP_NOWARN;
2501 gfp &= ~__GFP_RECLAIM;
2502 }
2503 return __folio_alloc_node(gfp, order, nid);
2504}
2505
2506static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2507{
2508 int nr_pages = folio_nr_pages(folio);
2509
2510 /* Avoid migrating to a node that is nearly full */
2511 if (!migrate_balanced_pgdat(pgdat, nr_migrate_pages: nr_pages)) {
2512 int z;
2513
2514 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2515 return 0;
2516 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2517 if (managed_zone(zone: pgdat->node_zones + z))
2518 break;
2519 }
2520 wakeup_kswapd(zone: pgdat->node_zones + z, gfp_mask: 0,
2521 order: folio_order(folio), highest_zoneidx: ZONE_MOVABLE);
2522 return 0;
2523 }
2524
2525 if (!folio_isolate_lru(folio))
2526 return 0;
2527
2528 node_stat_mod_folio(folio, item: NR_ISOLATED_ANON + folio_is_file_lru(folio),
2529 nr: nr_pages);
2530
2531 /*
2532 * Isolating the folio has taken another reference, so the
2533 * caller's reference can be safely dropped without the folio
2534 * disappearing underneath us during migration.
2535 */
2536 folio_put(folio);
2537 return 1;
2538}
2539
2540/*
2541 * Attempt to migrate a misplaced folio to the specified destination
2542 * node. Caller is expected to have an elevated reference count on
2543 * the folio that will be dropped by this function before returning.
2544 */
2545int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2546 int node)
2547{
2548 pg_data_t *pgdat = NODE_DATA(node);
2549 int isolated;
2550 int nr_remaining;
2551 unsigned int nr_succeeded;
2552 LIST_HEAD(migratepages);
2553 int nr_pages = folio_nr_pages(folio);
2554
2555 /*
2556 * Don't migrate file folios that are mapped in multiple processes
2557 * with execute permissions as they are probably shared libraries.
2558 * To check if the folio is shared, ideally we want to make sure
2559 * every page is mapped to the same process. Doing that is very
2560 * expensive, so check the estimated mapcount of the folio instead.
2561 */
2562 if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2563 (vma->vm_flags & VM_EXEC))
2564 goto out;
2565
2566 /*
2567 * Also do not migrate dirty folios as not all filesystems can move
2568 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2569 */
2570 if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2571 goto out;
2572
2573 isolated = numamigrate_isolate_folio(pgdat, folio);
2574 if (!isolated)
2575 goto out;
2576
2577 list_add(new: &folio->lru, head: &migratepages);
2578 nr_remaining = migrate_pages(from: &migratepages, get_new_folio: alloc_misplaced_dst_folio,
2579 NULL, private: node, mode: MIGRATE_ASYNC,
2580 reason: MR_NUMA_MISPLACED, ret_succeeded: &nr_succeeded);
2581 if (nr_remaining) {
2582 if (!list_empty(head: &migratepages)) {
2583 list_del(entry: &folio->lru);
2584 node_stat_mod_folio(folio, item: NR_ISOLATED_ANON +
2585 folio_is_file_lru(folio), nr: -nr_pages);
2586 folio_putback_lru(folio);
2587 }
2588 isolated = 0;
2589 }
2590 if (nr_succeeded) {
2591 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2592 if (!node_is_toptier(node: folio_nid(folio)) && node_is_toptier(node))
2593 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2594 nr_succeeded);
2595 }
2596 BUG_ON(!list_empty(&migratepages));
2597 return isolated;
2598
2599out:
2600 folio_put(folio);
2601 return 0;
2602}
2603#endif /* CONFIG_NUMA_BALANCING */
2604#endif /* CONFIG_NUMA */
2605

source code of linux/mm/migrate.c