1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Memory Migration functionality - linux/mm/migrate.c
4 *
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 *
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
9 *
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
13 * Christoph Lameter
14 */
15
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
53
54 #include <asm/tlbflush.h>
55
56 #include <trace/events/migrate.h>
57
58 #include "internal.h"
59
isolate_movable_page(struct page * page,isolate_mode_t mode)60 int isolate_movable_page(struct page *page, isolate_mode_t mode)
61 {
62 struct address_space *mapping;
63
64 /*
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
67 *
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
72 */
73 if (unlikely(!get_page_unless_zero(page)))
74 goto out;
75
76 /*
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
80 */
81 if (unlikely(!__PageMovable(page)))
82 goto out_putpage;
83 /*
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
87 *
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
93 */
94 if (unlikely(!trylock_page(page)))
95 goto out_putpage;
96
97 if (!PageMovable(page) || PageIsolated(page))
98 goto out_no_isolated;
99
100 mapping = page_mapping(page);
101 VM_BUG_ON_PAGE(!mapping, page);
102
103 if (!mapping->a_ops->isolate_page(page, mode))
104 goto out_no_isolated;
105
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 SetPageIsolated(page);
109 unlock_page(page);
110
111 return 0;
112
113 out_no_isolated:
114 unlock_page(page);
115 out_putpage:
116 put_page(page);
117 out:
118 return -EBUSY;
119 }
120
putback_movable_page(struct page * page)121 static void putback_movable_page(struct page *page)
122 {
123 struct address_space *mapping;
124
125 mapping = page_mapping(page);
126 mapping->a_ops->putback_page(page);
127 ClearPageIsolated(page);
128 }
129
130 /*
131 * Put previously isolated pages back onto the appropriate lists
132 * from where they were once taken off for compaction/migration.
133 *
134 * This function shall be used whenever the isolated pageset has been
135 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136 * and isolate_hugetlb().
137 */
putback_movable_pages(struct list_head * l)138 void putback_movable_pages(struct list_head *l)
139 {
140 struct page *page;
141 struct page *page2;
142
143 list_for_each_entry_safe(page, page2, l, lru) {
144 if (unlikely(PageHuge(page))) {
145 putback_active_hugepage(page);
146 continue;
147 }
148 list_del(&page->lru);
149 /*
150 * We isolated non-lru movable page so here we can use
151 * __PageMovable because LRU page's mapping cannot have
152 * PAGE_MAPPING_MOVABLE.
153 */
154 if (unlikely(__PageMovable(page))) {
155 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 lock_page(page);
157 if (PageMovable(page))
158 putback_movable_page(page);
159 else
160 ClearPageIsolated(page);
161 unlock_page(page);
162 put_page(page);
163 } else {
164 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
165 page_is_file_lru(page), -thp_nr_pages(page));
166 putback_lru_page(page);
167 }
168 }
169 }
170
171 /*
172 * Restore a potential migration pte to a working pte entry
173 */
remove_migration_pte(struct folio * folio,struct vm_area_struct * vma,unsigned long addr,void * old)174 static bool remove_migration_pte(struct folio *folio,
175 struct vm_area_struct *vma, unsigned long addr, void *old)
176 {
177 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
178
179 while (page_vma_mapped_walk(&pvmw)) {
180 rmap_t rmap_flags = RMAP_NONE;
181 pte_t pte;
182 swp_entry_t entry;
183 struct page *new;
184 unsigned long idx = 0;
185
186 /* pgoff is invalid for ksm pages, but they are never large */
187 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
188 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
189 new = folio_page(folio, idx);
190
191 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
192 /* PMD-mapped THP migration entry */
193 if (!pvmw.pte) {
194 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
195 !folio_test_pmd_mappable(folio), folio);
196 remove_migration_pmd(&pvmw, new);
197 continue;
198 }
199 #endif
200
201 folio_get(folio);
202 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
203 if (pte_swp_soft_dirty(*pvmw.pte))
204 pte = pte_mksoft_dirty(pte);
205
206 /*
207 * Recheck VMA as permissions can change since migration started
208 */
209 entry = pte_to_swp_entry(*pvmw.pte);
210 if (is_writable_migration_entry(entry))
211 pte = maybe_mkwrite(pte, vma);
212 else if (pte_swp_uffd_wp(*pvmw.pte))
213 pte = pte_mkuffd_wp(pte);
214
215 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
216 rmap_flags |= RMAP_EXCLUSIVE;
217
218 if (unlikely(is_device_private_page(new))) {
219 if (pte_write(pte))
220 entry = make_writable_device_private_entry(
221 page_to_pfn(new));
222 else
223 entry = make_readable_device_private_entry(
224 page_to_pfn(new));
225 pte = swp_entry_to_pte(entry);
226 if (pte_swp_soft_dirty(*pvmw.pte))
227 pte = pte_swp_mksoft_dirty(pte);
228 if (pte_swp_uffd_wp(*pvmw.pte))
229 pte = pte_swp_mkuffd_wp(pte);
230 }
231
232 #ifdef CONFIG_HUGETLB_PAGE
233 if (folio_test_hugetlb(folio)) {
234 unsigned int shift = huge_page_shift(hstate_vma(vma));
235
236 pte = pte_mkhuge(pte);
237 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
238 if (folio_test_anon(folio))
239 hugepage_add_anon_rmap(new, vma, pvmw.address,
240 rmap_flags);
241 else
242 page_dup_file_rmap(new, true);
243 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
244 } else
245 #endif
246 {
247 if (folio_test_anon(folio))
248 page_add_anon_rmap(new, vma, pvmw.address,
249 rmap_flags);
250 else
251 page_add_file_rmap(new, vma, false);
252 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
253 }
254 if (vma->vm_flags & VM_LOCKED)
255 mlock_page_drain_local();
256
257 trace_remove_migration_pte(pvmw.address, pte_val(pte),
258 compound_order(new));
259
260 /* No need to invalidate - it was non-present before */
261 update_mmu_cache(vma, pvmw.address, pvmw.pte);
262 }
263
264 return true;
265 }
266
267 /*
268 * Get rid of all migration entries and replace them by
269 * references to the indicated page.
270 */
remove_migration_ptes(struct folio * src,struct folio * dst,bool locked)271 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
272 {
273 struct rmap_walk_control rwc = {
274 .rmap_one = remove_migration_pte,
275 .arg = src,
276 };
277
278 if (locked)
279 rmap_walk_locked(dst, &rwc);
280 else
281 rmap_walk(dst, &rwc);
282 }
283
284 /*
285 * Something used the pte of a page under migration. We need to
286 * get to the page and wait until migration is finished.
287 * When we return from this function the fault will be retried.
288 */
__migration_entry_wait(struct mm_struct * mm,pte_t * ptep,spinlock_t * ptl)289 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
290 spinlock_t *ptl)
291 {
292 pte_t pte;
293 swp_entry_t entry;
294
295 spin_lock(ptl);
296 pte = *ptep;
297 if (!is_swap_pte(pte))
298 goto out;
299
300 entry = pte_to_swp_entry(pte);
301 if (!is_migration_entry(entry))
302 goto out;
303
304 migration_entry_wait_on_locked(entry, ptep, ptl);
305 return;
306 out:
307 pte_unmap_unlock(ptep, ptl);
308 }
309
migration_entry_wait(struct mm_struct * mm,pmd_t * pmd,unsigned long address)310 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
311 unsigned long address)
312 {
313 spinlock_t *ptl = pte_lockptr(mm, pmd);
314 pte_t *ptep = pte_offset_map(pmd, address);
315 __migration_entry_wait(mm, ptep, ptl);
316 }
317
318 #ifdef CONFIG_HUGETLB_PAGE
__migration_entry_wait_huge(pte_t * ptep,spinlock_t * ptl)319 void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl)
320 {
321 pte_t pte;
322
323 spin_lock(ptl);
324 pte = huge_ptep_get(ptep);
325
326 if (unlikely(!is_hugetlb_entry_migration(pte)))
327 spin_unlock(ptl);
328 else
329 migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
330 }
331
migration_entry_wait_huge(struct vm_area_struct * vma,pte_t * pte)332 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte)
333 {
334 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte);
335
336 __migration_entry_wait_huge(pte, ptl);
337 }
338 #endif
339
340 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
pmd_migration_entry_wait(struct mm_struct * mm,pmd_t * pmd)341 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
342 {
343 spinlock_t *ptl;
344
345 ptl = pmd_lock(mm, pmd);
346 if (!is_pmd_migration_entry(*pmd))
347 goto unlock;
348 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
349 return;
350 unlock:
351 spin_unlock(ptl);
352 }
353 #endif
354
expected_page_refs(struct address_space * mapping,struct page * page)355 static int expected_page_refs(struct address_space *mapping, struct page *page)
356 {
357 int expected_count = 1;
358
359 if (mapping)
360 expected_count += compound_nr(page) + page_has_private(page);
361 return expected_count;
362 }
363
364 /*
365 * Replace the page in the mapping.
366 *
367 * The number of remaining references must be:
368 * 1 for anonymous pages without a mapping
369 * 2 for pages with a mapping
370 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
371 */
folio_migrate_mapping(struct address_space * mapping,struct folio * newfolio,struct folio * folio,int extra_count)372 int folio_migrate_mapping(struct address_space *mapping,
373 struct folio *newfolio, struct folio *folio, int extra_count)
374 {
375 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
376 struct zone *oldzone, *newzone;
377 int dirty;
378 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
379 long nr = folio_nr_pages(folio);
380
381 if (!mapping) {
382 /* Anonymous page without mapping */
383 if (folio_ref_count(folio) != expected_count)
384 return -EAGAIN;
385
386 /* No turning back from here */
387 newfolio->index = folio->index;
388 newfolio->mapping = folio->mapping;
389 if (folio_test_swapbacked(folio))
390 __folio_set_swapbacked(newfolio);
391
392 return MIGRATEPAGE_SUCCESS;
393 }
394
395 oldzone = folio_zone(folio);
396 newzone = folio_zone(newfolio);
397
398 xas_lock_irq(&xas);
399 if (!folio_ref_freeze(folio, expected_count)) {
400 xas_unlock_irq(&xas);
401 return -EAGAIN;
402 }
403
404 /*
405 * Now we know that no one else is looking at the folio:
406 * no turning back from here.
407 */
408 newfolio->index = folio->index;
409 newfolio->mapping = folio->mapping;
410 folio_ref_add(newfolio, nr); /* add cache reference */
411 if (folio_test_swapbacked(folio)) {
412 __folio_set_swapbacked(newfolio);
413 if (folio_test_swapcache(folio)) {
414 folio_set_swapcache(newfolio);
415 newfolio->private = folio_get_private(folio);
416 }
417 } else {
418 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
419 }
420
421 /* Move dirty while page refs frozen and newpage not yet exposed */
422 dirty = folio_test_dirty(folio);
423 if (dirty) {
424 folio_clear_dirty(folio);
425 folio_set_dirty(newfolio);
426 }
427
428 xas_store(&xas, newfolio);
429
430 /*
431 * Drop cache reference from old page by unfreezing
432 * to one less reference.
433 * We know this isn't the last reference.
434 */
435 folio_ref_unfreeze(folio, expected_count - nr);
436
437 xas_unlock(&xas);
438 /* Leave irq disabled to prevent preemption while updating stats */
439
440 /*
441 * If moved to a different zone then also account
442 * the page for that zone. Other VM counters will be
443 * taken care of when we establish references to the
444 * new page and drop references to the old page.
445 *
446 * Note that anonymous pages are accounted for
447 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
448 * are mapped to swap space.
449 */
450 if (newzone != oldzone) {
451 struct lruvec *old_lruvec, *new_lruvec;
452 struct mem_cgroup *memcg;
453
454 memcg = folio_memcg(folio);
455 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
456 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
457
458 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
459 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
460 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
461 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
462 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
463 }
464 #ifdef CONFIG_SWAP
465 if (folio_test_swapcache(folio)) {
466 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
467 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
468 }
469 #endif
470 if (dirty && mapping_can_writeback(mapping)) {
471 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
472 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
473 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
474 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
475 }
476 }
477 local_irq_enable();
478
479 return MIGRATEPAGE_SUCCESS;
480 }
481 EXPORT_SYMBOL(folio_migrate_mapping);
482
483 /*
484 * The expected number of remaining references is the same as that
485 * of folio_migrate_mapping().
486 */
migrate_huge_page_move_mapping(struct address_space * mapping,struct page * newpage,struct page * page)487 int migrate_huge_page_move_mapping(struct address_space *mapping,
488 struct page *newpage, struct page *page)
489 {
490 XA_STATE(xas, &mapping->i_pages, page_index(page));
491 int expected_count;
492
493 xas_lock_irq(&xas);
494 expected_count = 2 + page_has_private(page);
495 if (!page_ref_freeze(page, expected_count)) {
496 xas_unlock_irq(&xas);
497 return -EAGAIN;
498 }
499
500 newpage->index = page->index;
501 newpage->mapping = page->mapping;
502
503 get_page(newpage);
504
505 xas_store(&xas, newpage);
506
507 page_ref_unfreeze(page, expected_count - 1);
508
509 xas_unlock_irq(&xas);
510
511 return MIGRATEPAGE_SUCCESS;
512 }
513
514 /*
515 * Copy the flags and some other ancillary information
516 */
folio_migrate_flags(struct folio * newfolio,struct folio * folio)517 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
518 {
519 int cpupid;
520
521 if (folio_test_error(folio))
522 folio_set_error(newfolio);
523 if (folio_test_referenced(folio))
524 folio_set_referenced(newfolio);
525 if (folio_test_uptodate(folio))
526 folio_mark_uptodate(newfolio);
527 if (folio_test_clear_active(folio)) {
528 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
529 folio_set_active(newfolio);
530 } else if (folio_test_clear_unevictable(folio))
531 folio_set_unevictable(newfolio);
532 if (folio_test_workingset(folio))
533 folio_set_workingset(newfolio);
534 if (folio_test_checked(folio))
535 folio_set_checked(newfolio);
536 /*
537 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
538 * migration entries. We can still have PG_anon_exclusive set on an
539 * effectively unmapped and unreferenced first sub-pages of an
540 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
541 */
542 if (folio_test_mappedtodisk(folio))
543 folio_set_mappedtodisk(newfolio);
544
545 /* Move dirty on pages not done by folio_migrate_mapping() */
546 if (folio_test_dirty(folio))
547 folio_set_dirty(newfolio);
548
549 if (folio_test_young(folio))
550 folio_set_young(newfolio);
551 if (folio_test_idle(folio))
552 folio_set_idle(newfolio);
553
554 /*
555 * Copy NUMA information to the new page, to prevent over-eager
556 * future migrations of this same page.
557 */
558 cpupid = page_cpupid_xchg_last(&folio->page, -1);
559 page_cpupid_xchg_last(&newfolio->page, cpupid);
560
561 folio_migrate_ksm(newfolio, folio);
562 /*
563 * Please do not reorder this without considering how mm/ksm.c's
564 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
565 */
566 if (folio_test_swapcache(folio))
567 folio_clear_swapcache(folio);
568 folio_clear_private(folio);
569
570 /* page->private contains hugetlb specific flags */
571 if (!folio_test_hugetlb(folio))
572 folio->private = NULL;
573
574 /*
575 * If any waiters have accumulated on the new page then
576 * wake them up.
577 */
578 if (folio_test_writeback(newfolio))
579 folio_end_writeback(newfolio);
580
581 /*
582 * PG_readahead shares the same bit with PG_reclaim. The above
583 * end_page_writeback() may clear PG_readahead mistakenly, so set the
584 * bit after that.
585 */
586 if (folio_test_readahead(folio))
587 folio_set_readahead(newfolio);
588
589 folio_copy_owner(newfolio, folio);
590
591 if (!folio_test_hugetlb(folio))
592 mem_cgroup_migrate(folio, newfolio);
593 }
594 EXPORT_SYMBOL(folio_migrate_flags);
595
folio_migrate_copy(struct folio * newfolio,struct folio * folio)596 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
597 {
598 folio_copy(newfolio, folio);
599 folio_migrate_flags(newfolio, folio);
600 }
601 EXPORT_SYMBOL(folio_migrate_copy);
602
603 /************************************************************
604 * Migration functions
605 ***********************************************************/
606
607 /*
608 * Common logic to directly migrate a single LRU page suitable for
609 * pages that do not use PagePrivate/PagePrivate2.
610 *
611 * Pages are locked upon entry and exit.
612 */
migrate_page(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode)613 int migrate_page(struct address_space *mapping,
614 struct page *newpage, struct page *page,
615 enum migrate_mode mode)
616 {
617 struct folio *newfolio = page_folio(newpage);
618 struct folio *folio = page_folio(page);
619 int rc;
620
621 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
622
623 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
624
625 if (rc != MIGRATEPAGE_SUCCESS)
626 return rc;
627
628 if (mode != MIGRATE_SYNC_NO_COPY)
629 folio_migrate_copy(newfolio, folio);
630 else
631 folio_migrate_flags(newfolio, folio);
632 return MIGRATEPAGE_SUCCESS;
633 }
634 EXPORT_SYMBOL(migrate_page);
635
636 #ifdef CONFIG_BLOCK
637 /* Returns true if all buffers are successfully locked */
buffer_migrate_lock_buffers(struct buffer_head * head,enum migrate_mode mode)638 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
639 enum migrate_mode mode)
640 {
641 struct buffer_head *bh = head;
642
643 /* Simple case, sync compaction */
644 if (mode != MIGRATE_ASYNC) {
645 do {
646 lock_buffer(bh);
647 bh = bh->b_this_page;
648
649 } while (bh != head);
650
651 return true;
652 }
653
654 /* async case, we cannot block on lock_buffer so use trylock_buffer */
655 do {
656 if (!trylock_buffer(bh)) {
657 /*
658 * We failed to lock the buffer and cannot stall in
659 * async migration. Release the taken locks
660 */
661 struct buffer_head *failed_bh = bh;
662 bh = head;
663 while (bh != failed_bh) {
664 unlock_buffer(bh);
665 bh = bh->b_this_page;
666 }
667 return false;
668 }
669
670 bh = bh->b_this_page;
671 } while (bh != head);
672 return true;
673 }
674
__buffer_migrate_page(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode,bool check_refs)675 static int __buffer_migrate_page(struct address_space *mapping,
676 struct page *newpage, struct page *page, enum migrate_mode mode,
677 bool check_refs)
678 {
679 struct buffer_head *bh, *head;
680 int rc;
681 int expected_count;
682
683 if (!page_has_buffers(page))
684 return migrate_page(mapping, newpage, page, mode);
685
686 /* Check whether page does not have extra refs before we do more work */
687 expected_count = expected_page_refs(mapping, page);
688 if (page_count(page) != expected_count)
689 return -EAGAIN;
690
691 head = page_buffers(page);
692 if (!buffer_migrate_lock_buffers(head, mode))
693 return -EAGAIN;
694
695 if (check_refs) {
696 bool busy;
697 bool invalidated = false;
698
699 recheck_buffers:
700 busy = false;
701 spin_lock(&mapping->private_lock);
702 bh = head;
703 do {
704 if (atomic_read(&bh->b_count)) {
705 busy = true;
706 break;
707 }
708 bh = bh->b_this_page;
709 } while (bh != head);
710 if (busy) {
711 if (invalidated) {
712 rc = -EAGAIN;
713 goto unlock_buffers;
714 }
715 spin_unlock(&mapping->private_lock);
716 invalidate_bh_lrus();
717 invalidated = true;
718 goto recheck_buffers;
719 }
720 }
721
722 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
723 if (rc != MIGRATEPAGE_SUCCESS)
724 goto unlock_buffers;
725
726 attach_page_private(newpage, detach_page_private(page));
727
728 bh = head;
729 do {
730 set_bh_page(bh, newpage, bh_offset(bh));
731 bh = bh->b_this_page;
732
733 } while (bh != head);
734
735 if (mode != MIGRATE_SYNC_NO_COPY)
736 migrate_page_copy(newpage, page);
737 else
738 migrate_page_states(newpage, page);
739
740 rc = MIGRATEPAGE_SUCCESS;
741 unlock_buffers:
742 if (check_refs)
743 spin_unlock(&mapping->private_lock);
744 bh = head;
745 do {
746 unlock_buffer(bh);
747 bh = bh->b_this_page;
748
749 } while (bh != head);
750
751 return rc;
752 }
753
754 /*
755 * Migration function for pages with buffers. This function can only be used
756 * if the underlying filesystem guarantees that no other references to "page"
757 * exist. For example attached buffer heads are accessed only under page lock.
758 */
buffer_migrate_page(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode)759 int buffer_migrate_page(struct address_space *mapping,
760 struct page *newpage, struct page *page, enum migrate_mode mode)
761 {
762 return __buffer_migrate_page(mapping, newpage, page, mode, false);
763 }
764 EXPORT_SYMBOL(buffer_migrate_page);
765
766 /*
767 * Same as above except that this variant is more careful and checks that there
768 * are also no buffer head references. This function is the right one for
769 * mappings where buffer heads are directly looked up and referenced (such as
770 * block device mappings).
771 */
buffer_migrate_page_norefs(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode)772 int buffer_migrate_page_norefs(struct address_space *mapping,
773 struct page *newpage, struct page *page, enum migrate_mode mode)
774 {
775 return __buffer_migrate_page(mapping, newpage, page, mode, true);
776 }
777 #endif
778
779 /*
780 * Writeback a page to clean the dirty state
781 */
writeout(struct address_space * mapping,struct page * page)782 static int writeout(struct address_space *mapping, struct page *page)
783 {
784 struct folio *folio = page_folio(page);
785 struct writeback_control wbc = {
786 .sync_mode = WB_SYNC_NONE,
787 .nr_to_write = 1,
788 .range_start = 0,
789 .range_end = LLONG_MAX,
790 .for_reclaim = 1
791 };
792 int rc;
793
794 if (!mapping->a_ops->writepage)
795 /* No write method for the address space */
796 return -EINVAL;
797
798 if (!clear_page_dirty_for_io(page))
799 /* Someone else already triggered a write */
800 return -EAGAIN;
801
802 /*
803 * A dirty page may imply that the underlying filesystem has
804 * the page on some queue. So the page must be clean for
805 * migration. Writeout may mean we loose the lock and the
806 * page state is no longer what we checked for earlier.
807 * At this point we know that the migration attempt cannot
808 * be successful.
809 */
810 remove_migration_ptes(folio, folio, false);
811
812 rc = mapping->a_ops->writepage(page, &wbc);
813
814 if (rc != AOP_WRITEPAGE_ACTIVATE)
815 /* unlocked. Relock */
816 lock_page(page);
817
818 return (rc < 0) ? -EIO : -EAGAIN;
819 }
820
821 /*
822 * Default handling if a filesystem does not provide a migration function.
823 */
fallback_migrate_page(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode)824 static int fallback_migrate_page(struct address_space *mapping,
825 struct page *newpage, struct page *page, enum migrate_mode mode)
826 {
827 if (PageDirty(page)) {
828 /* Only writeback pages in full synchronous migration */
829 switch (mode) {
830 case MIGRATE_SYNC:
831 case MIGRATE_SYNC_NO_COPY:
832 break;
833 default:
834 return -EBUSY;
835 }
836 return writeout(mapping, page);
837 }
838
839 /*
840 * Buffers may be managed in a filesystem specific way.
841 * We must have no buffers or drop them.
842 */
843 if (page_has_private(page) &&
844 !try_to_release_page(page, GFP_KERNEL))
845 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
846
847 return migrate_page(mapping, newpage, page, mode);
848 }
849
850 /*
851 * Move a page to a newly allocated page
852 * The page is locked and all ptes have been successfully removed.
853 *
854 * The new page will have replaced the old page if this function
855 * is successful.
856 *
857 * Return value:
858 * < 0 - error code
859 * MIGRATEPAGE_SUCCESS - success
860 */
move_to_new_folio(struct folio * dst,struct folio * src,enum migrate_mode mode)861 static int move_to_new_folio(struct folio *dst, struct folio *src,
862 enum migrate_mode mode)
863 {
864 struct address_space *mapping;
865 int rc = -EAGAIN;
866 bool is_lru = !__PageMovable(&src->page);
867
868 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
869 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
870
871 mapping = folio_mapping(src);
872
873 if (likely(is_lru)) {
874 if (!mapping)
875 rc = migrate_page(mapping, &dst->page, &src->page, mode);
876 else if (mapping->a_ops->migratepage)
877 /*
878 * Most pages have a mapping and most filesystems
879 * provide a migratepage callback. Anonymous pages
880 * are part of swap space which also has its own
881 * migratepage callback. This is the most common path
882 * for page migration.
883 */
884 rc = mapping->a_ops->migratepage(mapping, &dst->page,
885 &src->page, mode);
886 else
887 rc = fallback_migrate_page(mapping, &dst->page,
888 &src->page, mode);
889 } else {
890 /*
891 * In case of non-lru page, it could be released after
892 * isolation step. In that case, we shouldn't try migration.
893 */
894 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
895 if (!folio_test_movable(src)) {
896 rc = MIGRATEPAGE_SUCCESS;
897 folio_clear_isolated(src);
898 goto out;
899 }
900
901 rc = mapping->a_ops->migratepage(mapping, &dst->page,
902 &src->page, mode);
903 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
904 !folio_test_isolated(src));
905 }
906
907 /*
908 * When successful, old pagecache src->mapping must be cleared before
909 * src is freed; but stats require that PageAnon be left as PageAnon.
910 */
911 if (rc == MIGRATEPAGE_SUCCESS) {
912 if (__PageMovable(&src->page)) {
913 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
914
915 /*
916 * We clear PG_movable under page_lock so any compactor
917 * cannot try to migrate this page.
918 */
919 folio_clear_isolated(src);
920 }
921
922 /*
923 * Anonymous and movable src->mapping will be cleared by
924 * free_pages_prepare so don't reset it here for keeping
925 * the type to work PageAnon, for example.
926 */
927 if (!folio_mapping_flags(src))
928 src->mapping = NULL;
929
930 if (likely(!folio_is_zone_device(dst)))
931 flush_dcache_folio(dst);
932 }
933 out:
934 return rc;
935 }
936
__unmap_and_move(struct page * page,struct page * newpage,int force,enum migrate_mode mode)937 static int __unmap_and_move(struct page *page, struct page *newpage,
938 int force, enum migrate_mode mode)
939 {
940 struct folio *folio = page_folio(page);
941 struct folio *dst = page_folio(newpage);
942 int rc = -EAGAIN;
943 bool page_was_mapped = false;
944 struct anon_vma *anon_vma = NULL;
945 bool is_lru = !__PageMovable(page);
946
947 if (!trylock_page(page)) {
948 if (!force || mode == MIGRATE_ASYNC)
949 goto out;
950
951 /*
952 * It's not safe for direct compaction to call lock_page.
953 * For example, during page readahead pages are added locked
954 * to the LRU. Later, when the IO completes the pages are
955 * marked uptodate and unlocked. However, the queueing
956 * could be merging multiple pages for one bio (e.g.
957 * mpage_readahead). If an allocation happens for the
958 * second or third page, the process can end up locking
959 * the same page twice and deadlocking. Rather than
960 * trying to be clever about what pages can be locked,
961 * avoid the use of lock_page for direct compaction
962 * altogether.
963 */
964 if (current->flags & PF_MEMALLOC)
965 goto out;
966
967 lock_page(page);
968 }
969
970 if (PageWriteback(page)) {
971 /*
972 * Only in the case of a full synchronous migration is it
973 * necessary to wait for PageWriteback. In the async case,
974 * the retry loop is too short and in the sync-light case,
975 * the overhead of stalling is too much
976 */
977 switch (mode) {
978 case MIGRATE_SYNC:
979 case MIGRATE_SYNC_NO_COPY:
980 break;
981 default:
982 rc = -EBUSY;
983 goto out_unlock;
984 }
985 if (!force)
986 goto out_unlock;
987 wait_on_page_writeback(page);
988 }
989
990 /*
991 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
992 * we cannot notice that anon_vma is freed while we migrates a page.
993 * This get_anon_vma() delays freeing anon_vma pointer until the end
994 * of migration. File cache pages are no problem because of page_lock()
995 * File Caches may use write_page() or lock_page() in migration, then,
996 * just care Anon page here.
997 *
998 * Only page_get_anon_vma() understands the subtleties of
999 * getting a hold on an anon_vma from outside one of its mms.
1000 * But if we cannot get anon_vma, then we won't need it anyway,
1001 * because that implies that the anon page is no longer mapped
1002 * (and cannot be remapped so long as we hold the page lock).
1003 */
1004 if (PageAnon(page) && !PageKsm(page))
1005 anon_vma = page_get_anon_vma(page);
1006
1007 /*
1008 * Block others from accessing the new page when we get around to
1009 * establishing additional references. We are usually the only one
1010 * holding a reference to newpage at this point. We used to have a BUG
1011 * here if trylock_page(newpage) fails, but would like to allow for
1012 * cases where there might be a race with the previous use of newpage.
1013 * This is much like races on refcount of oldpage: just don't BUG().
1014 */
1015 if (unlikely(!trylock_page(newpage)))
1016 goto out_unlock;
1017
1018 if (unlikely(!is_lru)) {
1019 rc = move_to_new_folio(dst, folio, mode);
1020 goto out_unlock_both;
1021 }
1022
1023 /*
1024 * Corner case handling:
1025 * 1. When a new swap-cache page is read into, it is added to the LRU
1026 * and treated as swapcache but it has no rmap yet.
1027 * Calling try_to_unmap() against a page->mapping==NULL page will
1028 * trigger a BUG. So handle it here.
1029 * 2. An orphaned page (see truncate_cleanup_page) might have
1030 * fs-private metadata. The page can be picked up due to memory
1031 * offlining. Everywhere else except page reclaim, the page is
1032 * invisible to the vm, so the page can not be migrated. So try to
1033 * free the metadata, so the page can be freed.
1034 */
1035 if (!page->mapping) {
1036 VM_BUG_ON_PAGE(PageAnon(page), page);
1037 if (page_has_private(page)) {
1038 try_to_free_buffers(folio);
1039 goto out_unlock_both;
1040 }
1041 } else if (page_mapped(page)) {
1042 /* Establish migration ptes */
1043 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1044 page);
1045 try_to_migrate(folio, 0);
1046 page_was_mapped = true;
1047 }
1048
1049 if (!page_mapped(page))
1050 rc = move_to_new_folio(dst, folio, mode);
1051
1052 /*
1053 * When successful, push newpage to LRU immediately: so that if it
1054 * turns out to be an mlocked page, remove_migration_ptes() will
1055 * automatically build up the correct newpage->mlock_count for it.
1056 *
1057 * We would like to do something similar for the old page, when
1058 * unsuccessful, and other cases when a page has been temporarily
1059 * isolated from the unevictable LRU: but this case is the easiest.
1060 */
1061 if (rc == MIGRATEPAGE_SUCCESS) {
1062 lru_cache_add(newpage);
1063 if (page_was_mapped)
1064 lru_add_drain();
1065 }
1066
1067 if (page_was_mapped)
1068 remove_migration_ptes(folio,
1069 rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1070
1071 out_unlock_both:
1072 unlock_page(newpage);
1073 out_unlock:
1074 /* Drop an anon_vma reference if we took one */
1075 if (anon_vma)
1076 put_anon_vma(anon_vma);
1077 unlock_page(page);
1078 out:
1079 /*
1080 * If migration is successful, decrease refcount of the newpage,
1081 * which will not free the page because new page owner increased
1082 * refcounter.
1083 */
1084 if (rc == MIGRATEPAGE_SUCCESS)
1085 put_page(newpage);
1086
1087 return rc;
1088 }
1089
1090 /*
1091 * Obtain the lock on page, remove all ptes and migrate the page
1092 * to the newly allocated page in newpage.
1093 */
unmap_and_move(new_page_t get_new_page,free_page_t put_new_page,unsigned long private,struct page * page,int force,enum migrate_mode mode,enum migrate_reason reason,struct list_head * ret)1094 static int unmap_and_move(new_page_t get_new_page,
1095 free_page_t put_new_page,
1096 unsigned long private, struct page *page,
1097 int force, enum migrate_mode mode,
1098 enum migrate_reason reason,
1099 struct list_head *ret)
1100 {
1101 int rc = MIGRATEPAGE_SUCCESS;
1102 struct page *newpage = NULL;
1103
1104 if (!thp_migration_supported() && PageTransHuge(page))
1105 return -ENOSYS;
1106
1107 if (page_count(page) == 1) {
1108 /* page was freed from under us. So we are done. */
1109 ClearPageActive(page);
1110 ClearPageUnevictable(page);
1111 if (unlikely(__PageMovable(page))) {
1112 lock_page(page);
1113 if (!PageMovable(page))
1114 ClearPageIsolated(page);
1115 unlock_page(page);
1116 }
1117 goto out;
1118 }
1119
1120 newpage = get_new_page(page, private);
1121 if (!newpage)
1122 return -ENOMEM;
1123
1124 newpage->private = 0;
1125 rc = __unmap_and_move(page, newpage, force, mode);
1126 if (rc == MIGRATEPAGE_SUCCESS)
1127 set_page_owner_migrate_reason(newpage, reason);
1128
1129 out:
1130 if (rc != -EAGAIN) {
1131 /*
1132 * A page that has been migrated has all references
1133 * removed and will be freed. A page that has not been
1134 * migrated will have kept its references and be restored.
1135 */
1136 list_del(&page->lru);
1137 }
1138
1139 /*
1140 * If migration is successful, releases reference grabbed during
1141 * isolation. Otherwise, restore the page to right list unless
1142 * we want to retry.
1143 */
1144 if (rc == MIGRATEPAGE_SUCCESS) {
1145 /*
1146 * Compaction can migrate also non-LRU pages which are
1147 * not accounted to NR_ISOLATED_*. They can be recognized
1148 * as __PageMovable
1149 */
1150 if (likely(!__PageMovable(page)))
1151 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1152 page_is_file_lru(page), -thp_nr_pages(page));
1153
1154 if (reason != MR_MEMORY_FAILURE)
1155 /*
1156 * We release the page in page_handle_poison.
1157 */
1158 put_page(page);
1159 } else {
1160 if (rc != -EAGAIN)
1161 list_add_tail(&page->lru, ret);
1162
1163 if (put_new_page)
1164 put_new_page(newpage, private);
1165 else
1166 put_page(newpage);
1167 }
1168
1169 return rc;
1170 }
1171
1172 /*
1173 * Counterpart of unmap_and_move_page() for hugepage migration.
1174 *
1175 * This function doesn't wait the completion of hugepage I/O
1176 * because there is no race between I/O and migration for hugepage.
1177 * Note that currently hugepage I/O occurs only in direct I/O
1178 * where no lock is held and PG_writeback is irrelevant,
1179 * and writeback status of all subpages are counted in the reference
1180 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1181 * under direct I/O, the reference of the head page is 512 and a bit more.)
1182 * This means that when we try to migrate hugepage whose subpages are
1183 * doing direct I/O, some references remain after try_to_unmap() and
1184 * hugepage migration fails without data corruption.
1185 *
1186 * There is also no race when direct I/O is issued on the page under migration,
1187 * because then pte is replaced with migration swap entry and direct I/O code
1188 * will wait in the page fault for migration to complete.
1189 */
unmap_and_move_huge_page(new_page_t get_new_page,free_page_t put_new_page,unsigned long private,struct page * hpage,int force,enum migrate_mode mode,int reason,struct list_head * ret)1190 static int unmap_and_move_huge_page(new_page_t get_new_page,
1191 free_page_t put_new_page, unsigned long private,
1192 struct page *hpage, int force,
1193 enum migrate_mode mode, int reason,
1194 struct list_head *ret)
1195 {
1196 struct folio *dst, *src = page_folio(hpage);
1197 int rc = -EAGAIN;
1198 int page_was_mapped = 0;
1199 struct page *new_hpage;
1200 struct anon_vma *anon_vma = NULL;
1201 struct address_space *mapping = NULL;
1202
1203 /*
1204 * Migratability of hugepages depends on architectures and their size.
1205 * This check is necessary because some callers of hugepage migration
1206 * like soft offline and memory hotremove don't walk through page
1207 * tables or check whether the hugepage is pmd-based or not before
1208 * kicking migration.
1209 */
1210 if (!hugepage_migration_supported(page_hstate(hpage))) {
1211 list_move_tail(&hpage->lru, ret);
1212 return -ENOSYS;
1213 }
1214
1215 if (page_count(hpage) == 1) {
1216 /* page was freed from under us. So we are done. */
1217 putback_active_hugepage(hpage);
1218 return MIGRATEPAGE_SUCCESS;
1219 }
1220
1221 new_hpage = get_new_page(hpage, private);
1222 if (!new_hpage)
1223 return -ENOMEM;
1224 dst = page_folio(new_hpage);
1225
1226 if (!trylock_page(hpage)) {
1227 if (!force)
1228 goto out;
1229 switch (mode) {
1230 case MIGRATE_SYNC:
1231 case MIGRATE_SYNC_NO_COPY:
1232 break;
1233 default:
1234 goto out;
1235 }
1236 lock_page(hpage);
1237 }
1238
1239 /*
1240 * Check for pages which are in the process of being freed. Without
1241 * page_mapping() set, hugetlbfs specific move page routine will not
1242 * be called and we could leak usage counts for subpools.
1243 */
1244 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1245 rc = -EBUSY;
1246 goto out_unlock;
1247 }
1248
1249 if (PageAnon(hpage))
1250 anon_vma = page_get_anon_vma(hpage);
1251
1252 if (unlikely(!trylock_page(new_hpage)))
1253 goto put_anon;
1254
1255 if (page_mapped(hpage)) {
1256 enum ttu_flags ttu = 0;
1257
1258 if (!PageAnon(hpage)) {
1259 /*
1260 * In shared mappings, try_to_unmap could potentially
1261 * call huge_pmd_unshare. Because of this, take
1262 * semaphore in write mode here and set TTU_RMAP_LOCKED
1263 * to let lower levels know we have taken the lock.
1264 */
1265 mapping = hugetlb_page_mapping_lock_write(hpage);
1266 if (unlikely(!mapping))
1267 goto unlock_put_anon;
1268
1269 ttu = TTU_RMAP_LOCKED;
1270 }
1271
1272 try_to_migrate(src, ttu);
1273 page_was_mapped = 1;
1274
1275 if (ttu & TTU_RMAP_LOCKED)
1276 i_mmap_unlock_write(mapping);
1277 }
1278
1279 if (!page_mapped(hpage))
1280 rc = move_to_new_folio(dst, src, mode);
1281
1282 if (page_was_mapped)
1283 remove_migration_ptes(src,
1284 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1285
1286 unlock_put_anon:
1287 unlock_page(new_hpage);
1288
1289 put_anon:
1290 if (anon_vma)
1291 put_anon_vma(anon_vma);
1292
1293 if (rc == MIGRATEPAGE_SUCCESS) {
1294 move_hugetlb_state(hpage, new_hpage, reason);
1295 put_new_page = NULL;
1296 }
1297
1298 out_unlock:
1299 unlock_page(hpage);
1300 out:
1301 if (rc == MIGRATEPAGE_SUCCESS)
1302 putback_active_hugepage(hpage);
1303 else if (rc != -EAGAIN)
1304 list_move_tail(&hpage->lru, ret);
1305
1306 /*
1307 * If migration was not successful and there's a freeing callback, use
1308 * it. Otherwise, put_page() will drop the reference grabbed during
1309 * isolation.
1310 */
1311 if (put_new_page)
1312 put_new_page(new_hpage, private);
1313 else
1314 putback_active_hugepage(new_hpage);
1315
1316 return rc;
1317 }
1318
try_split_thp(struct page * page,struct page ** page2,struct list_head * from)1319 static inline int try_split_thp(struct page *page, struct page **page2,
1320 struct list_head *from)
1321 {
1322 int rc = 0;
1323
1324 lock_page(page);
1325 rc = split_huge_page_to_list(page, from);
1326 unlock_page(page);
1327 if (!rc)
1328 list_safe_reset_next(page, *page2, lru);
1329
1330 return rc;
1331 }
1332
1333 /*
1334 * migrate_pages - migrate the pages specified in a list, to the free pages
1335 * supplied as the target for the page migration
1336 *
1337 * @from: The list of pages to be migrated.
1338 * @get_new_page: The function used to allocate free pages to be used
1339 * as the target of the page migration.
1340 * @put_new_page: The function used to free target pages if migration
1341 * fails, or NULL if no special handling is necessary.
1342 * @private: Private data to be passed on to get_new_page()
1343 * @mode: The migration mode that specifies the constraints for
1344 * page migration, if any.
1345 * @reason: The reason for page migration.
1346 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1347 * the caller passes a non-NULL pointer.
1348 *
1349 * The function returns after 10 attempts or if no pages are movable any more
1350 * because the list has become empty or no retryable pages exist any more.
1351 * It is caller's responsibility to call putback_movable_pages() to return pages
1352 * to the LRU or free list only if ret != 0.
1353 *
1354 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1355 * an error code. The number of THP splits will be considered as the number of
1356 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1357 */
migrate_pages(struct list_head * from,new_page_t get_new_page,free_page_t put_new_page,unsigned long private,enum migrate_mode mode,int reason,unsigned int * ret_succeeded)1358 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1359 free_page_t put_new_page, unsigned long private,
1360 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1361 {
1362 int retry = 1;
1363 int thp_retry = 1;
1364 int nr_failed = 0;
1365 int nr_failed_pages = 0;
1366 int nr_succeeded = 0;
1367 int nr_thp_succeeded = 0;
1368 int nr_thp_failed = 0;
1369 int nr_thp_split = 0;
1370 int pass = 0;
1371 bool is_thp = false;
1372 struct page *page;
1373 struct page *page2;
1374 int rc, nr_subpages;
1375 LIST_HEAD(ret_pages);
1376 LIST_HEAD(thp_split_pages);
1377 bool nosplit = (reason == MR_NUMA_MISPLACED);
1378 bool no_subpage_counting = false;
1379
1380 trace_mm_migrate_pages_start(mode, reason);
1381
1382 thp_subpage_migration:
1383 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1384 retry = 0;
1385 thp_retry = 0;
1386
1387 list_for_each_entry_safe(page, page2, from, lru) {
1388 retry:
1389 /*
1390 * THP statistics is based on the source huge page.
1391 * Capture required information that might get lost
1392 * during migration.
1393 */
1394 is_thp = PageTransHuge(page) && !PageHuge(page);
1395 nr_subpages = compound_nr(page);
1396 cond_resched();
1397
1398 if (PageHuge(page))
1399 rc = unmap_and_move_huge_page(get_new_page,
1400 put_new_page, private, page,
1401 pass > 2, mode, reason,
1402 &ret_pages);
1403 else
1404 rc = unmap_and_move(get_new_page, put_new_page,
1405 private, page, pass > 2, mode,
1406 reason, &ret_pages);
1407 /*
1408 * The rules are:
1409 * Success: non hugetlb page will be freed, hugetlb
1410 * page will be put back
1411 * -EAGAIN: stay on the from list
1412 * -ENOMEM: stay on the from list
1413 * Other errno: put on ret_pages list then splice to
1414 * from list
1415 */
1416 switch(rc) {
1417 /*
1418 * THP migration might be unsupported or the
1419 * allocation could've failed so we should
1420 * retry on the same page with the THP split
1421 * to base pages.
1422 *
1423 * Head page is retried immediately and tail
1424 * pages are added to the tail of the list so
1425 * we encounter them after the rest of the list
1426 * is processed.
1427 */
1428 case -ENOSYS:
1429 /* THP migration is unsupported */
1430 if (is_thp) {
1431 nr_thp_failed++;
1432 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1433 nr_thp_split++;
1434 goto retry;
1435 }
1436 /* Hugetlb migration is unsupported */
1437 } else if (!no_subpage_counting) {
1438 nr_failed++;
1439 }
1440
1441 nr_failed_pages += nr_subpages;
1442 break;
1443 case -ENOMEM:
1444 /*
1445 * When memory is low, don't bother to try to migrate
1446 * other pages, just exit.
1447 * THP NUMA faulting doesn't split THP to retry.
1448 */
1449 if (is_thp && !nosplit) {
1450 nr_thp_failed++;
1451 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1452 nr_thp_split++;
1453 goto retry;
1454 }
1455 } else if (!no_subpage_counting) {
1456 nr_failed++;
1457 }
1458
1459 nr_failed_pages += nr_subpages;
1460 /*
1461 * There might be some subpages of fail-to-migrate THPs
1462 * left in thp_split_pages list. Move them back to migration
1463 * list so that they could be put back to the right list by
1464 * the caller otherwise the page refcnt will be leaked.
1465 */
1466 list_splice_init(&thp_split_pages, from);
1467 nr_thp_failed += thp_retry;
1468 goto out;
1469 case -EAGAIN:
1470 if (is_thp)
1471 thp_retry++;
1472 else
1473 retry++;
1474 break;
1475 case MIGRATEPAGE_SUCCESS:
1476 nr_succeeded += nr_subpages;
1477 if (is_thp)
1478 nr_thp_succeeded++;
1479 break;
1480 default:
1481 /*
1482 * Permanent failure (-EBUSY, etc.):
1483 * unlike -EAGAIN case, the failed page is
1484 * removed from migration page list and not
1485 * retried in the next outer loop.
1486 */
1487 if (is_thp)
1488 nr_thp_failed++;
1489 else if (!no_subpage_counting)
1490 nr_failed++;
1491
1492 nr_failed_pages += nr_subpages;
1493 break;
1494 }
1495 }
1496 }
1497 nr_failed += retry;
1498 nr_thp_failed += thp_retry;
1499 /*
1500 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1501 * counting in this round, since all subpages of a THP is counted
1502 * as 1 failure in the first round.
1503 */
1504 if (!list_empty(&thp_split_pages)) {
1505 /*
1506 * Move non-migrated pages (after 10 retries) to ret_pages
1507 * to avoid migrating them again.
1508 */
1509 list_splice_init(from, &ret_pages);
1510 list_splice_init(&thp_split_pages, from);
1511 no_subpage_counting = true;
1512 retry = 1;
1513 goto thp_subpage_migration;
1514 }
1515
1516 rc = nr_failed + nr_thp_failed;
1517 out:
1518 /*
1519 * Put the permanent failure page back to migration list, they
1520 * will be put back to the right list by the caller.
1521 */
1522 list_splice(&ret_pages, from);
1523
1524 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1525 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1526 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1527 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1528 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1529 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1530 nr_thp_failed, nr_thp_split, mode, reason);
1531
1532 if (ret_succeeded)
1533 *ret_succeeded = nr_succeeded;
1534
1535 return rc;
1536 }
1537
alloc_migration_target(struct page * page,unsigned long private)1538 struct page *alloc_migration_target(struct page *page, unsigned long private)
1539 {
1540 struct folio *folio = page_folio(page);
1541 struct migration_target_control *mtc;
1542 gfp_t gfp_mask;
1543 unsigned int order = 0;
1544 struct folio *new_folio = NULL;
1545 int nid;
1546 int zidx;
1547
1548 mtc = (struct migration_target_control *)private;
1549 gfp_mask = mtc->gfp_mask;
1550 nid = mtc->nid;
1551 if (nid == NUMA_NO_NODE)
1552 nid = folio_nid(folio);
1553
1554 if (folio_test_hugetlb(folio)) {
1555 struct hstate *h = page_hstate(&folio->page);
1556
1557 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1558 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1559 }
1560
1561 if (folio_test_large(folio)) {
1562 /*
1563 * clear __GFP_RECLAIM to make the migration callback
1564 * consistent with regular THP allocations.
1565 */
1566 gfp_mask &= ~__GFP_RECLAIM;
1567 gfp_mask |= GFP_TRANSHUGE;
1568 order = folio_order(folio);
1569 }
1570 zidx = zone_idx(folio_zone(folio));
1571 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1572 gfp_mask |= __GFP_HIGHMEM;
1573
1574 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1575
1576 return &new_folio->page;
1577 }
1578
1579 #ifdef CONFIG_NUMA
1580
store_status(int __user * status,int start,int value,int nr)1581 static int store_status(int __user *status, int start, int value, int nr)
1582 {
1583 while (nr-- > 0) {
1584 if (put_user(value, status + start))
1585 return -EFAULT;
1586 start++;
1587 }
1588
1589 return 0;
1590 }
1591
do_move_pages_to_node(struct mm_struct * mm,struct list_head * pagelist,int node)1592 static int do_move_pages_to_node(struct mm_struct *mm,
1593 struct list_head *pagelist, int node)
1594 {
1595 int err;
1596 struct migration_target_control mtc = {
1597 .nid = node,
1598 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1599 };
1600
1601 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1602 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1603 if (err)
1604 putback_movable_pages(pagelist);
1605 return err;
1606 }
1607
1608 /*
1609 * Resolves the given address to a struct page, isolates it from the LRU and
1610 * puts it to the given pagelist.
1611 * Returns:
1612 * errno - if the page cannot be found/isolated
1613 * 0 - when it doesn't have to be migrated because it is already on the
1614 * target node
1615 * 1 - when it has been queued
1616 */
add_page_for_migration(struct mm_struct * mm,unsigned long addr,int node,struct list_head * pagelist,bool migrate_all)1617 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1618 int node, struct list_head *pagelist, bool migrate_all)
1619 {
1620 struct vm_area_struct *vma;
1621 struct page *page;
1622 int err;
1623
1624 mmap_read_lock(mm);
1625 err = -EFAULT;
1626 vma = vma_lookup(mm, addr);
1627 if (!vma || !vma_migratable(vma))
1628 goto out;
1629
1630 /* FOLL_DUMP to ignore special (like zero) pages */
1631 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1632
1633 err = PTR_ERR(page);
1634 if (IS_ERR(page))
1635 goto out;
1636
1637 err = -ENOENT;
1638 if (!page)
1639 goto out;
1640
1641 err = 0;
1642 if (page_to_nid(page) == node)
1643 goto out_putpage;
1644
1645 err = -EACCES;
1646 if (page_mapcount(page) > 1 && !migrate_all)
1647 goto out_putpage;
1648
1649 if (PageHuge(page)) {
1650 if (PageHead(page)) {
1651 err = isolate_hugetlb(page, pagelist);
1652 if (!err)
1653 err = 1;
1654 }
1655 } else {
1656 struct page *head;
1657
1658 head = compound_head(page);
1659 err = isolate_lru_page(head);
1660 if (err)
1661 goto out_putpage;
1662
1663 err = 1;
1664 list_add_tail(&head->lru, pagelist);
1665 mod_node_page_state(page_pgdat(head),
1666 NR_ISOLATED_ANON + page_is_file_lru(head),
1667 thp_nr_pages(head));
1668 }
1669 out_putpage:
1670 /*
1671 * Either remove the duplicate refcount from
1672 * isolate_lru_page() or drop the page ref if it was
1673 * not isolated.
1674 */
1675 put_page(page);
1676 out:
1677 mmap_read_unlock(mm);
1678 return err;
1679 }
1680
move_pages_and_store_status(struct mm_struct * mm,int node,struct list_head * pagelist,int __user * status,int start,int i,unsigned long nr_pages)1681 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1682 struct list_head *pagelist, int __user *status,
1683 int start, int i, unsigned long nr_pages)
1684 {
1685 int err;
1686
1687 if (list_empty(pagelist))
1688 return 0;
1689
1690 err = do_move_pages_to_node(mm, pagelist, node);
1691 if (err) {
1692 /*
1693 * Positive err means the number of failed
1694 * pages to migrate. Since we are going to
1695 * abort and return the number of non-migrated
1696 * pages, so need to include the rest of the
1697 * nr_pages that have not been attempted as
1698 * well.
1699 */
1700 if (err > 0)
1701 err += nr_pages - i - 1;
1702 return err;
1703 }
1704 return store_status(status, start, node, i - start);
1705 }
1706
1707 /*
1708 * Migrate an array of page address onto an array of nodes and fill
1709 * the corresponding array of status.
1710 */
do_pages_move(struct mm_struct * mm,nodemask_t task_nodes,unsigned long nr_pages,const void __user * __user * pages,const int __user * nodes,int __user * status,int flags)1711 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1712 unsigned long nr_pages,
1713 const void __user * __user *pages,
1714 const int __user *nodes,
1715 int __user *status, int flags)
1716 {
1717 int current_node = NUMA_NO_NODE;
1718 LIST_HEAD(pagelist);
1719 int start, i;
1720 int err = 0, err1;
1721
1722 lru_cache_disable();
1723
1724 for (i = start = 0; i < nr_pages; i++) {
1725 const void __user *p;
1726 unsigned long addr;
1727 int node;
1728
1729 err = -EFAULT;
1730 if (get_user(p, pages + i))
1731 goto out_flush;
1732 if (get_user(node, nodes + i))
1733 goto out_flush;
1734 addr = (unsigned long)untagged_addr(p);
1735
1736 err = -ENODEV;
1737 if (node < 0 || node >= MAX_NUMNODES)
1738 goto out_flush;
1739 if (!node_state(node, N_MEMORY))
1740 goto out_flush;
1741
1742 err = -EACCES;
1743 if (!node_isset(node, task_nodes))
1744 goto out_flush;
1745
1746 if (current_node == NUMA_NO_NODE) {
1747 current_node = node;
1748 start = i;
1749 } else if (node != current_node) {
1750 err = move_pages_and_store_status(mm, current_node,
1751 &pagelist, status, start, i, nr_pages);
1752 if (err)
1753 goto out;
1754 start = i;
1755 current_node = node;
1756 }
1757
1758 /*
1759 * Errors in the page lookup or isolation are not fatal and we simply
1760 * report them via status
1761 */
1762 err = add_page_for_migration(mm, addr, current_node,
1763 &pagelist, flags & MPOL_MF_MOVE_ALL);
1764
1765 if (err > 0) {
1766 /* The page is successfully queued for migration */
1767 continue;
1768 }
1769
1770 /*
1771 * The move_pages() man page does not have an -EEXIST choice, so
1772 * use -EFAULT instead.
1773 */
1774 if (err == -EEXIST)
1775 err = -EFAULT;
1776
1777 /*
1778 * If the page is already on the target node (!err), store the
1779 * node, otherwise, store the err.
1780 */
1781 err = store_status(status, i, err ? : current_node, 1);
1782 if (err)
1783 goto out_flush;
1784
1785 err = move_pages_and_store_status(mm, current_node, &pagelist,
1786 status, start, i, nr_pages);
1787 if (err)
1788 goto out;
1789 current_node = NUMA_NO_NODE;
1790 }
1791 out_flush:
1792 /* Make sure we do not overwrite the existing error */
1793 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1794 status, start, i, nr_pages);
1795 if (err >= 0)
1796 err = err1;
1797 out:
1798 lru_cache_enable();
1799 return err;
1800 }
1801
1802 /*
1803 * Determine the nodes of an array of pages and store it in an array of status.
1804 */
do_pages_stat_array(struct mm_struct * mm,unsigned long nr_pages,const void __user ** pages,int * status)1805 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1806 const void __user **pages, int *status)
1807 {
1808 unsigned long i;
1809
1810 mmap_read_lock(mm);
1811
1812 for (i = 0; i < nr_pages; i++) {
1813 unsigned long addr = (unsigned long)(*pages);
1814 struct vm_area_struct *vma;
1815 struct page *page;
1816 int err = -EFAULT;
1817
1818 vma = vma_lookup(mm, addr);
1819 if (!vma)
1820 goto set_status;
1821
1822 /* FOLL_DUMP to ignore special (like zero) pages */
1823 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1824
1825 err = PTR_ERR(page);
1826 if (IS_ERR(page))
1827 goto set_status;
1828
1829 if (page) {
1830 err = page_to_nid(page);
1831 put_page(page);
1832 } else {
1833 err = -ENOENT;
1834 }
1835 set_status:
1836 *status = err;
1837
1838 pages++;
1839 status++;
1840 }
1841
1842 mmap_read_unlock(mm);
1843 }
1844
get_compat_pages_array(const void __user * chunk_pages[],const void __user * __user * pages,unsigned long chunk_nr)1845 static int get_compat_pages_array(const void __user *chunk_pages[],
1846 const void __user * __user *pages,
1847 unsigned long chunk_nr)
1848 {
1849 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1850 compat_uptr_t p;
1851 int i;
1852
1853 for (i = 0; i < chunk_nr; i++) {
1854 if (get_user(p, pages32 + i))
1855 return -EFAULT;
1856 chunk_pages[i] = compat_ptr(p);
1857 }
1858
1859 return 0;
1860 }
1861
1862 /*
1863 * Determine the nodes of a user array of pages and store it in
1864 * a user array of status.
1865 */
do_pages_stat(struct mm_struct * mm,unsigned long nr_pages,const void __user * __user * pages,int __user * status)1866 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1867 const void __user * __user *pages,
1868 int __user *status)
1869 {
1870 #define DO_PAGES_STAT_CHUNK_NR 16UL
1871 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1872 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1873
1874 while (nr_pages) {
1875 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1876
1877 if (in_compat_syscall()) {
1878 if (get_compat_pages_array(chunk_pages, pages,
1879 chunk_nr))
1880 break;
1881 } else {
1882 if (copy_from_user(chunk_pages, pages,
1883 chunk_nr * sizeof(*chunk_pages)))
1884 break;
1885 }
1886
1887 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1888
1889 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1890 break;
1891
1892 pages += chunk_nr;
1893 status += chunk_nr;
1894 nr_pages -= chunk_nr;
1895 }
1896 return nr_pages ? -EFAULT : 0;
1897 }
1898
find_mm_struct(pid_t pid,nodemask_t * mem_nodes)1899 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1900 {
1901 struct task_struct *task;
1902 struct mm_struct *mm;
1903
1904 /*
1905 * There is no need to check if current process has the right to modify
1906 * the specified process when they are same.
1907 */
1908 if (!pid) {
1909 mmget(current->mm);
1910 *mem_nodes = cpuset_mems_allowed(current);
1911 return current->mm;
1912 }
1913
1914 /* Find the mm_struct */
1915 rcu_read_lock();
1916 task = find_task_by_vpid(pid);
1917 if (!task) {
1918 rcu_read_unlock();
1919 return ERR_PTR(-ESRCH);
1920 }
1921 get_task_struct(task);
1922
1923 /*
1924 * Check if this process has the right to modify the specified
1925 * process. Use the regular "ptrace_may_access()" checks.
1926 */
1927 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1928 rcu_read_unlock();
1929 mm = ERR_PTR(-EPERM);
1930 goto out;
1931 }
1932 rcu_read_unlock();
1933
1934 mm = ERR_PTR(security_task_movememory(task));
1935 if (IS_ERR(mm))
1936 goto out;
1937 *mem_nodes = cpuset_mems_allowed(task);
1938 mm = get_task_mm(task);
1939 out:
1940 put_task_struct(task);
1941 if (!mm)
1942 mm = ERR_PTR(-EINVAL);
1943 return mm;
1944 }
1945
1946 /*
1947 * Move a list of pages in the address space of the currently executing
1948 * process.
1949 */
kernel_move_pages(pid_t pid,unsigned long nr_pages,const void __user * __user * pages,const int __user * nodes,int __user * status,int flags)1950 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1951 const void __user * __user *pages,
1952 const int __user *nodes,
1953 int __user *status, int flags)
1954 {
1955 struct mm_struct *mm;
1956 int err;
1957 nodemask_t task_nodes;
1958
1959 /* Check flags */
1960 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1961 return -EINVAL;
1962
1963 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1964 return -EPERM;
1965
1966 mm = find_mm_struct(pid, &task_nodes);
1967 if (IS_ERR(mm))
1968 return PTR_ERR(mm);
1969
1970 if (nodes)
1971 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1972 nodes, status, flags);
1973 else
1974 err = do_pages_stat(mm, nr_pages, pages, status);
1975
1976 mmput(mm);
1977 return err;
1978 }
1979
SYSCALL_DEFINE6(move_pages,pid_t,pid,unsigned long,nr_pages,const void __user * __user *,pages,const int __user *,nodes,int __user *,status,int,flags)1980 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1981 const void __user * __user *, pages,
1982 const int __user *, nodes,
1983 int __user *, status, int, flags)
1984 {
1985 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1986 }
1987
1988 #ifdef CONFIG_NUMA_BALANCING
1989 /*
1990 * Returns true if this is a safe migration target node for misplaced NUMA
1991 * pages. Currently it only checks the watermarks which is crude.
1992 */
migrate_balanced_pgdat(struct pglist_data * pgdat,unsigned long nr_migrate_pages)1993 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1994 unsigned long nr_migrate_pages)
1995 {
1996 int z;
1997
1998 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1999 struct zone *zone = pgdat->node_zones + z;
2000
2001 if (!managed_zone(zone))
2002 continue;
2003
2004 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2005 if (!zone_watermark_ok(zone, 0,
2006 high_wmark_pages(zone) +
2007 nr_migrate_pages,
2008 ZONE_MOVABLE, 0))
2009 continue;
2010 return true;
2011 }
2012 return false;
2013 }
2014
alloc_misplaced_dst_page(struct page * page,unsigned long data)2015 static struct page *alloc_misplaced_dst_page(struct page *page,
2016 unsigned long data)
2017 {
2018 int nid = (int) data;
2019 int order = compound_order(page);
2020 gfp_t gfp = __GFP_THISNODE;
2021 struct folio *new;
2022
2023 if (order > 0)
2024 gfp |= GFP_TRANSHUGE_LIGHT;
2025 else {
2026 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2027 __GFP_NOWARN;
2028 gfp &= ~__GFP_RECLAIM;
2029 }
2030 new = __folio_alloc_node(gfp, order, nid);
2031
2032 return &new->page;
2033 }
2034
numamigrate_isolate_page(pg_data_t * pgdat,struct page * page)2035 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2036 {
2037 int nr_pages = thp_nr_pages(page);
2038 int order = compound_order(page);
2039
2040 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2041
2042 /* Do not migrate THP mapped by multiple processes */
2043 if (PageTransHuge(page) && total_mapcount(page) > 1)
2044 return 0;
2045
2046 /* Avoid migrating to a node that is nearly full */
2047 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2048 int z;
2049
2050 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2051 return 0;
2052 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2053 if (managed_zone(pgdat->node_zones + z))
2054 break;
2055 }
2056 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2057 return 0;
2058 }
2059
2060 if (isolate_lru_page(page))
2061 return 0;
2062
2063 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2064 nr_pages);
2065
2066 /*
2067 * Isolating the page has taken another reference, so the
2068 * caller's reference can be safely dropped without the page
2069 * disappearing underneath us during migration.
2070 */
2071 put_page(page);
2072 return 1;
2073 }
2074
2075 /*
2076 * Attempt to migrate a misplaced page to the specified destination
2077 * node. Caller is expected to have an elevated reference count on
2078 * the page that will be dropped by this function before returning.
2079 */
migrate_misplaced_page(struct page * page,struct vm_area_struct * vma,int node)2080 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2081 int node)
2082 {
2083 pg_data_t *pgdat = NODE_DATA(node);
2084 int isolated;
2085 int nr_remaining;
2086 unsigned int nr_succeeded;
2087 LIST_HEAD(migratepages);
2088 int nr_pages = thp_nr_pages(page);
2089
2090 /*
2091 * Don't migrate file pages that are mapped in multiple processes
2092 * with execute permissions as they are probably shared libraries.
2093 */
2094 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2095 (vma->vm_flags & VM_EXEC))
2096 goto out;
2097
2098 /*
2099 * Also do not migrate dirty pages as not all filesystems can move
2100 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2101 */
2102 if (page_is_file_lru(page) && PageDirty(page))
2103 goto out;
2104
2105 isolated = numamigrate_isolate_page(pgdat, page);
2106 if (!isolated)
2107 goto out;
2108
2109 list_add(&page->lru, &migratepages);
2110 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2111 NULL, node, MIGRATE_ASYNC,
2112 MR_NUMA_MISPLACED, &nr_succeeded);
2113 if (nr_remaining) {
2114 if (!list_empty(&migratepages)) {
2115 list_del(&page->lru);
2116 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2117 page_is_file_lru(page), -nr_pages);
2118 putback_lru_page(page);
2119 }
2120 isolated = 0;
2121 }
2122 if (nr_succeeded) {
2123 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2124 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2125 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2126 nr_succeeded);
2127 }
2128 BUG_ON(!list_empty(&migratepages));
2129 return isolated;
2130
2131 out:
2132 put_page(page);
2133 return 0;
2134 }
2135 #endif /* CONFIG_NUMA_BALANCING */
2136
2137 /*
2138 * node_demotion[] example:
2139 *
2140 * Consider a system with two sockets. Each socket has
2141 * three classes of memory attached: fast, medium and slow.
2142 * Each memory class is placed in its own NUMA node. The
2143 * CPUs are placed in the node with the "fast" memory. The
2144 * 6 NUMA nodes (0-5) might be split among the sockets like
2145 * this:
2146 *
2147 * Socket A: 0, 1, 2
2148 * Socket B: 3, 4, 5
2149 *
2150 * When Node 0 fills up, its memory should be migrated to
2151 * Node 1. When Node 1 fills up, it should be migrated to
2152 * Node 2. The migration path start on the nodes with the
2153 * processors (since allocations default to this node) and
2154 * fast memory, progress through medium and end with the
2155 * slow memory:
2156 *
2157 * 0 -> 1 -> 2 -> stop
2158 * 3 -> 4 -> 5 -> stop
2159 *
2160 * This is represented in the node_demotion[] like this:
2161 *
2162 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2163 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2164 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2165 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2166 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2167 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2168 *
2169 * Moreover some systems may have multiple slow memory nodes.
2170 * Suppose a system has one socket with 3 memory nodes, node 0
2171 * is fast memory type, and node 1/2 both are slow memory
2172 * type, and the distance between fast memory node and slow
2173 * memory node is same. So the migration path should be:
2174 *
2175 * 0 -> 1/2 -> stop
2176 *
2177 * This is represented in the node_demotion[] like this:
2178 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2179 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2180 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2181 */
2182
2183 /*
2184 * Writes to this array occur without locking. Cycles are
2185 * not allowed: Node X demotes to Y which demotes to X...
2186 *
2187 * If multiple reads are performed, a single rcu_read_lock()
2188 * must be held over all reads to ensure that no cycles are
2189 * observed.
2190 */
2191 #define DEFAULT_DEMOTION_TARGET_NODES 15
2192
2193 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2194 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2195 #else
2196 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2197 #endif
2198
2199 struct demotion_nodes {
2200 unsigned short nr;
2201 short nodes[DEMOTION_TARGET_NODES];
2202 };
2203
2204 static struct demotion_nodes *node_demotion __read_mostly;
2205
2206 /**
2207 * next_demotion_node() - Get the next node in the demotion path
2208 * @node: The starting node to lookup the next node
2209 *
2210 * Return: node id for next memory node in the demotion path hierarchy
2211 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2212 * @node online or guarantee that it *continues* to be the next demotion
2213 * target.
2214 */
next_demotion_node(int node)2215 int next_demotion_node(int node)
2216 {
2217 struct demotion_nodes *nd;
2218 unsigned short target_nr, index;
2219 int target;
2220
2221 if (!node_demotion)
2222 return NUMA_NO_NODE;
2223
2224 nd = &node_demotion[node];
2225
2226 /*
2227 * node_demotion[] is updated without excluding this
2228 * function from running. RCU doesn't provide any
2229 * compiler barriers, so the READ_ONCE() is required
2230 * to avoid compiler reordering or read merging.
2231 *
2232 * Make sure to use RCU over entire code blocks if
2233 * node_demotion[] reads need to be consistent.
2234 */
2235 rcu_read_lock();
2236 target_nr = READ_ONCE(nd->nr);
2237
2238 switch (target_nr) {
2239 case 0:
2240 target = NUMA_NO_NODE;
2241 goto out;
2242 case 1:
2243 index = 0;
2244 break;
2245 default:
2246 /*
2247 * If there are multiple target nodes, just select one
2248 * target node randomly.
2249 *
2250 * In addition, we can also use round-robin to select
2251 * target node, but we should introduce another variable
2252 * for node_demotion[] to record last selected target node,
2253 * that may cause cache ping-pong due to the changing of
2254 * last target node. Or introducing per-cpu data to avoid
2255 * caching issue, which seems more complicated. So selecting
2256 * target node randomly seems better until now.
2257 */
2258 index = get_random_int() % target_nr;
2259 break;
2260 }
2261
2262 target = READ_ONCE(nd->nodes[index]);
2263
2264 out:
2265 rcu_read_unlock();
2266 return target;
2267 }
2268
2269 /* Disable reclaim-based migration. */
__disable_all_migrate_targets(void)2270 static void __disable_all_migrate_targets(void)
2271 {
2272 int node, i;
2273
2274 if (!node_demotion)
2275 return;
2276
2277 for_each_online_node(node) {
2278 node_demotion[node].nr = 0;
2279 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2280 node_demotion[node].nodes[i] = NUMA_NO_NODE;
2281 }
2282 }
2283
disable_all_migrate_targets(void)2284 static void disable_all_migrate_targets(void)
2285 {
2286 __disable_all_migrate_targets();
2287
2288 /*
2289 * Ensure that the "disable" is visible across the system.
2290 * Readers will see either a combination of before+disable
2291 * state or disable+after. They will never see before and
2292 * after state together.
2293 *
2294 * The before+after state together might have cycles and
2295 * could cause readers to do things like loop until this
2296 * function finishes. This ensures they can only see a
2297 * single "bad" read and would, for instance, only loop
2298 * once.
2299 */
2300 synchronize_rcu();
2301 }
2302
2303 /*
2304 * Find an automatic demotion target for 'node'.
2305 * Failing here is OK. It might just indicate
2306 * being at the end of a chain.
2307 */
establish_migrate_target(int node,nodemask_t * used,int best_distance)2308 static int establish_migrate_target(int node, nodemask_t *used,
2309 int best_distance)
2310 {
2311 int migration_target, index, val;
2312 struct demotion_nodes *nd;
2313
2314 if (!node_demotion)
2315 return NUMA_NO_NODE;
2316
2317 nd = &node_demotion[node];
2318
2319 migration_target = find_next_best_node(node, used);
2320 if (migration_target == NUMA_NO_NODE)
2321 return NUMA_NO_NODE;
2322
2323 /*
2324 * If the node has been set a migration target node before,
2325 * which means it's the best distance between them. Still
2326 * check if this node can be demoted to other target nodes
2327 * if they have a same best distance.
2328 */
2329 if (best_distance != -1) {
2330 val = node_distance(node, migration_target);
2331 if (val > best_distance)
2332 goto out_clear;
2333 }
2334
2335 index = nd->nr;
2336 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2337 "Exceeds maximum demotion target nodes\n"))
2338 goto out_clear;
2339
2340 nd->nodes[index] = migration_target;
2341 nd->nr++;
2342
2343 return migration_target;
2344 out_clear:
2345 node_clear(migration_target, *used);
2346 return NUMA_NO_NODE;
2347 }
2348
2349 /*
2350 * When memory fills up on a node, memory contents can be
2351 * automatically migrated to another node instead of
2352 * discarded at reclaim.
2353 *
2354 * Establish a "migration path" which will start at nodes
2355 * with CPUs and will follow the priorities used to build the
2356 * page allocator zonelists.
2357 *
2358 * The difference here is that cycles must be avoided. If
2359 * node0 migrates to node1, then neither node1, nor anything
2360 * node1 migrates to can migrate to node0. Also one node can
2361 * be migrated to multiple nodes if the target nodes all have
2362 * a same best-distance against the source node.
2363 *
2364 * This function can run simultaneously with readers of
2365 * node_demotion[]. However, it can not run simultaneously
2366 * with itself. Exclusion is provided by memory hotplug events
2367 * being single-threaded.
2368 */
__set_migration_target_nodes(void)2369 static void __set_migration_target_nodes(void)
2370 {
2371 nodemask_t next_pass;
2372 nodemask_t this_pass;
2373 nodemask_t used_targets = NODE_MASK_NONE;
2374 int node, best_distance;
2375
2376 /*
2377 * Avoid any oddities like cycles that could occur
2378 * from changes in the topology. This will leave
2379 * a momentary gap when migration is disabled.
2380 */
2381 disable_all_migrate_targets();
2382
2383 /*
2384 * Allocations go close to CPUs, first. Assume that
2385 * the migration path starts at the nodes with CPUs.
2386 */
2387 next_pass = node_states[N_CPU];
2388 again:
2389 this_pass = next_pass;
2390 next_pass = NODE_MASK_NONE;
2391 /*
2392 * To avoid cycles in the migration "graph", ensure
2393 * that migration sources are not future targets by
2394 * setting them in 'used_targets'. Do this only
2395 * once per pass so that multiple source nodes can
2396 * share a target node.
2397 *
2398 * 'used_targets' will become unavailable in future
2399 * passes. This limits some opportunities for
2400 * multiple source nodes to share a destination.
2401 */
2402 nodes_or(used_targets, used_targets, this_pass);
2403
2404 for_each_node_mask(node, this_pass) {
2405 best_distance = -1;
2406
2407 /*
2408 * Try to set up the migration path for the node, and the target
2409 * migration nodes can be multiple, so doing a loop to find all
2410 * the target nodes if they all have a best node distance.
2411 */
2412 do {
2413 int target_node =
2414 establish_migrate_target(node, &used_targets,
2415 best_distance);
2416
2417 if (target_node == NUMA_NO_NODE)
2418 break;
2419
2420 if (best_distance == -1)
2421 best_distance = node_distance(node, target_node);
2422
2423 /*
2424 * Visit targets from this pass in the next pass.
2425 * Eventually, every node will have been part of
2426 * a pass, and will become set in 'used_targets'.
2427 */
2428 node_set(target_node, next_pass);
2429 } while (1);
2430 }
2431 /*
2432 * 'next_pass' contains nodes which became migration
2433 * targets in this pass. Make additional passes until
2434 * no more migrations targets are available.
2435 */
2436 if (!nodes_empty(next_pass))
2437 goto again;
2438 }
2439
2440 /*
2441 * For callers that do not hold get_online_mems() already.
2442 */
set_migration_target_nodes(void)2443 void set_migration_target_nodes(void)
2444 {
2445 get_online_mems();
2446 __set_migration_target_nodes();
2447 put_online_mems();
2448 }
2449
2450 /*
2451 * This leaves migrate-on-reclaim transiently disabled between
2452 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2453 * whether reclaim-based migration is enabled or not, which
2454 * ensures that the user can turn reclaim-based migration at
2455 * any time without needing to recalculate migration targets.
2456 *
2457 * These callbacks already hold get_online_mems(). That is why
2458 * __set_migration_target_nodes() can be used as opposed to
2459 * set_migration_target_nodes().
2460 */
2461 #ifdef CONFIG_MEMORY_HOTPLUG
migrate_on_reclaim_callback(struct notifier_block * self,unsigned long action,void * _arg)2462 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
2463 unsigned long action, void *_arg)
2464 {
2465 struct memory_notify *arg = _arg;
2466
2467 /*
2468 * Only update the node migration order when a node is
2469 * changing status, like online->offline. This avoids
2470 * the overhead of synchronize_rcu() in most cases.
2471 */
2472 if (arg->status_change_nid < 0)
2473 return notifier_from_errno(0);
2474
2475 switch (action) {
2476 case MEM_GOING_OFFLINE:
2477 /*
2478 * Make sure there are not transient states where
2479 * an offline node is a migration target. This
2480 * will leave migration disabled until the offline
2481 * completes and the MEM_OFFLINE case below runs.
2482 */
2483 disable_all_migrate_targets();
2484 break;
2485 case MEM_OFFLINE:
2486 case MEM_ONLINE:
2487 /*
2488 * Recalculate the target nodes once the node
2489 * reaches its final state (online or offline).
2490 */
2491 __set_migration_target_nodes();
2492 break;
2493 case MEM_CANCEL_OFFLINE:
2494 /*
2495 * MEM_GOING_OFFLINE disabled all the migration
2496 * targets. Reenable them.
2497 */
2498 __set_migration_target_nodes();
2499 break;
2500 case MEM_GOING_ONLINE:
2501 case MEM_CANCEL_ONLINE:
2502 break;
2503 }
2504
2505 return notifier_from_errno(0);
2506 }
2507 #endif
2508
migrate_on_reclaim_init(void)2509 void __init migrate_on_reclaim_init(void)
2510 {
2511 node_demotion = kcalloc(nr_node_ids,
2512 sizeof(struct demotion_nodes),
2513 GFP_KERNEL);
2514 WARN_ON(!node_demotion);
2515 #ifdef CONFIG_MEMORY_HOTPLUG
2516 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2517 #endif
2518 /*
2519 * At this point, all numa nodes with memory/CPus have their state
2520 * properly set, so we can build the demotion order now.
2521 * Let us hold the cpu_hotplug lock just, as we could possibily have
2522 * CPU hotplug events during boot.
2523 */
2524 cpus_read_lock();
2525 set_migration_target_nodes();
2526 cpus_read_unlock();
2527 }
2528
2529 bool numa_demotion_enabled = false;
2530
2531 #ifdef CONFIG_SYSFS
numa_demotion_enabled_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)2532 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2533 struct kobj_attribute *attr, char *buf)
2534 {
2535 return sysfs_emit(buf, "%s\n",
2536 numa_demotion_enabled ? "true" : "false");
2537 }
2538
numa_demotion_enabled_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)2539 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2540 struct kobj_attribute *attr,
2541 const char *buf, size_t count)
2542 {
2543 ssize_t ret;
2544
2545 ret = kstrtobool(buf, &numa_demotion_enabled);
2546 if (ret)
2547 return ret;
2548
2549 return count;
2550 }
2551
2552 static struct kobj_attribute numa_demotion_enabled_attr =
2553 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
2554 numa_demotion_enabled_store);
2555
2556 static struct attribute *numa_attrs[] = {
2557 &numa_demotion_enabled_attr.attr,
2558 NULL,
2559 };
2560
2561 static const struct attribute_group numa_attr_group = {
2562 .attrs = numa_attrs,
2563 };
2564
numa_init_sysfs(void)2565 static int __init numa_init_sysfs(void)
2566 {
2567 int err;
2568 struct kobject *numa_kobj;
2569
2570 numa_kobj = kobject_create_and_add("numa", mm_kobj);
2571 if (!numa_kobj) {
2572 pr_err("failed to create numa kobject\n");
2573 return -ENOMEM;
2574 }
2575 err = sysfs_create_group(numa_kobj, &numa_attr_group);
2576 if (err) {
2577 pr_err("failed to register numa group\n");
2578 goto delete_obj;
2579 }
2580 return 0;
2581
2582 delete_obj:
2583 kobject_put(numa_kobj);
2584 return err;
2585 }
2586 subsys_initcall(numa_init_sysfs);
2587 #endif /* CONFIG_SYSFS */
2588 #endif /* CONFIG_NUMA */
2589