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