Line data Source code
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 : #include <linux/memory-tiers.h>
54 :
55 : #include <asm/tlbflush.h>
56 :
57 : #include <trace/events/migrate.h>
58 :
59 : #include "internal.h"
60 :
61 0 : bool isolate_movable_page(struct page *page, isolate_mode_t mode)
62 : {
63 0 : struct folio *folio = folio_get_nontail_page(page);
64 : const struct movable_operations *mops;
65 :
66 : /*
67 : * Avoid burning cycles with pages that are yet under __free_pages(),
68 : * or just got freed under us.
69 : *
70 : * In case we 'win' a race for a movable page being freed under us and
71 : * raise its refcount preventing __free_pages() from doing its job
72 : * the put_page() at the end of this block will take care of
73 : * release this page, thus avoiding a nasty leakage.
74 : */
75 0 : if (!folio)
76 : goto out;
77 :
78 0 : if (unlikely(folio_test_slab(folio)))
79 : goto out_putfolio;
80 : /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
81 0 : smp_rmb();
82 : /*
83 : * Check movable flag before taking the page lock because
84 : * we use non-atomic bitops on newly allocated page flags so
85 : * unconditionally grabbing the lock ruins page's owner side.
86 : */
87 0 : if (unlikely(!__folio_test_movable(folio)))
88 : goto out_putfolio;
89 : /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
90 0 : smp_rmb();
91 0 : if (unlikely(folio_test_slab(folio)))
92 : goto out_putfolio;
93 :
94 : /*
95 : * As movable pages are not isolated from LRU lists, concurrent
96 : * compaction threads can race against page migration functions
97 : * as well as race against the releasing a page.
98 : *
99 : * In order to avoid having an already isolated movable page
100 : * being (wrongly) re-isolated while it is under migration,
101 : * or to avoid attempting to isolate pages being released,
102 : * lets be sure we have the page lock
103 : * before proceeding with the movable page isolation steps.
104 : */
105 0 : if (unlikely(!folio_trylock(folio)))
106 : goto out_putfolio;
107 :
108 0 : if (!folio_test_movable(folio) || folio_test_isolated(folio))
109 : goto out_no_isolated;
110 :
111 0 : mops = folio_movable_ops(folio);
112 : VM_BUG_ON_FOLIO(!mops, folio);
113 :
114 0 : if (!mops->isolate_page(&folio->page, mode))
115 : goto out_no_isolated;
116 :
117 : /* Driver shouldn't use PG_isolated bit of page->flags */
118 0 : WARN_ON_ONCE(folio_test_isolated(folio));
119 0 : folio_set_isolated(folio);
120 0 : folio_unlock(folio);
121 :
122 0 : return true;
123 :
124 : out_no_isolated:
125 0 : folio_unlock(folio);
126 : out_putfolio:
127 : folio_put(folio);
128 : out:
129 : return false;
130 : }
131 :
132 : static void putback_movable_folio(struct folio *folio)
133 : {
134 0 : const struct movable_operations *mops = folio_movable_ops(folio);
135 :
136 0 : mops->putback_page(&folio->page);
137 0 : folio_clear_isolated(folio);
138 : }
139 :
140 : /*
141 : * Put previously isolated pages back onto the appropriate lists
142 : * from where they were once taken off for compaction/migration.
143 : *
144 : * This function shall be used whenever the isolated pageset has been
145 : * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
146 : * and isolate_hugetlb().
147 : */
148 0 : void putback_movable_pages(struct list_head *l)
149 : {
150 : struct folio *folio;
151 : struct folio *folio2;
152 :
153 0 : list_for_each_entry_safe(folio, folio2, l, lru) {
154 0 : if (unlikely(folio_test_hugetlb(folio))) {
155 : folio_putback_active_hugetlb(folio);
156 : continue;
157 : }
158 0 : list_del(&folio->lru);
159 : /*
160 : * We isolated non-lru movable folio so here we can use
161 : * __PageMovable because LRU folio's mapping cannot have
162 : * PAGE_MAPPING_MOVABLE.
163 : */
164 0 : if (unlikely(__folio_test_movable(folio))) {
165 : VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
166 0 : folio_lock(folio);
167 0 : if (folio_test_movable(folio))
168 : putback_movable_folio(folio);
169 : else
170 : folio_clear_isolated(folio);
171 0 : folio_unlock(folio);
172 : folio_put(folio);
173 : } else {
174 0 : node_stat_mod_folio(folio, NR_ISOLATED_ANON +
175 0 : folio_is_file_lru(folio), -folio_nr_pages(folio));
176 0 : folio_putback_lru(folio);
177 : }
178 : }
179 0 : }
180 :
181 : /*
182 : * Restore a potential migration pte to a working pte entry
183 : */
184 0 : static bool remove_migration_pte(struct folio *folio,
185 : struct vm_area_struct *vma, unsigned long addr, void *old)
186 : {
187 0 : DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
188 :
189 0 : while (page_vma_mapped_walk(&pvmw)) {
190 0 : rmap_t rmap_flags = RMAP_NONE;
191 : pte_t pte;
192 : swp_entry_t entry;
193 : struct page *new;
194 0 : unsigned long idx = 0;
195 :
196 : /* pgoff is invalid for ksm pages, but they are never large */
197 0 : if (folio_test_large(folio) && !folio_test_hugetlb(folio))
198 0 : idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
199 0 : 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 0 : folio_get(folio);
212 0 : pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
213 0 : if (pte_swp_soft_dirty(*pvmw.pte))
214 : pte = pte_mksoft_dirty(pte);
215 :
216 : /*
217 : * Recheck VMA as permissions can change since migration started
218 : */
219 0 : entry = pte_to_swp_entry(*pvmw.pte);
220 0 : if (!is_migration_entry_young(entry))
221 : pte = pte_mkold(pte);
222 0 : if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
223 : pte = pte_mkdirty(pte);
224 0 : if (is_writable_migration_entry(entry))
225 0 : pte = maybe_mkwrite(pte, vma);
226 : else if (pte_swp_uffd_wp(*pvmw.pte))
227 : pte = pte_mkuffd_wp(pte);
228 : else
229 : pte = pte_wrprotect(pte);
230 :
231 0 : if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
232 0 : rmap_flags |= RMAP_EXCLUSIVE;
233 :
234 0 : if (unlikely(is_device_private_page(new))) {
235 : if (pte_write(pte))
236 : entry = make_writable_device_private_entry(
237 : page_to_pfn(new));
238 : else
239 : entry = make_readable_device_private_entry(
240 : page_to_pfn(new));
241 : pte = swp_entry_to_pte(entry);
242 : if (pte_swp_soft_dirty(*pvmw.pte))
243 : pte = pte_swp_mksoft_dirty(pte);
244 : if (pte_swp_uffd_wp(*pvmw.pte))
245 : pte = pte_swp_mkuffd_wp(pte);
246 : }
247 :
248 : #ifdef CONFIG_HUGETLB_PAGE
249 : if (folio_test_hugetlb(folio)) {
250 : unsigned int shift = huge_page_shift(hstate_vma(vma));
251 :
252 : pte = pte_mkhuge(pte);
253 : pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
254 : if (folio_test_anon(folio))
255 : hugepage_add_anon_rmap(new, vma, pvmw.address,
256 : rmap_flags);
257 : else
258 : page_dup_file_rmap(new, true);
259 : set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
260 : } else
261 : #endif
262 : {
263 0 : if (folio_test_anon(folio))
264 0 : page_add_anon_rmap(new, vma, pvmw.address,
265 : rmap_flags);
266 : else
267 0 : page_add_file_rmap(new, vma, false);
268 0 : set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269 : }
270 0 : if (vma->vm_flags & VM_LOCKED)
271 0 : mlock_drain_local();
272 :
273 : trace_remove_migration_pte(pvmw.address, pte_val(pte),
274 : compound_order(new));
275 :
276 : /* No need to invalidate - it was non-present before */
277 : update_mmu_cache(vma, pvmw.address, pvmw.pte);
278 : }
279 :
280 0 : return true;
281 : }
282 :
283 : /*
284 : * Get rid of all migration entries and replace them by
285 : * references to the indicated page.
286 : */
287 0 : void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
288 : {
289 0 : struct rmap_walk_control rwc = {
290 : .rmap_one = remove_migration_pte,
291 : .arg = src,
292 : };
293 :
294 0 : if (locked)
295 0 : rmap_walk_locked(dst, &rwc);
296 : else
297 0 : rmap_walk(dst, &rwc);
298 0 : }
299 :
300 : /*
301 : * Something used the pte of a page under migration. We need to
302 : * get to the page and wait until migration is finished.
303 : * When we return from this function the fault will be retried.
304 : */
305 0 : void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
306 : spinlock_t *ptl)
307 : {
308 : pte_t pte;
309 : swp_entry_t entry;
310 :
311 0 : spin_lock(ptl);
312 0 : pte = *ptep;
313 0 : if (!is_swap_pte(pte))
314 : goto out;
315 :
316 0 : entry = pte_to_swp_entry(pte);
317 0 : if (!is_migration_entry(entry))
318 : goto out;
319 :
320 0 : migration_entry_wait_on_locked(entry, ptep, ptl);
321 0 : return;
322 : out:
323 0 : pte_unmap_unlock(ptep, ptl);
324 : }
325 :
326 0 : void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
327 : unsigned long address)
328 : {
329 0 : spinlock_t *ptl = pte_lockptr(mm, pmd);
330 0 : pte_t *ptep = pte_offset_map(pmd, address);
331 0 : __migration_entry_wait(mm, ptep, ptl);
332 0 : }
333 :
334 : #ifdef CONFIG_HUGETLB_PAGE
335 : /*
336 : * The vma read lock must be held upon entry. Holding that lock prevents either
337 : * the pte or the ptl from being freed.
338 : *
339 : * This function will release the vma lock before returning.
340 : */
341 : void __migration_entry_wait_huge(struct vm_area_struct *vma,
342 : pte_t *ptep, spinlock_t *ptl)
343 : {
344 : pte_t pte;
345 :
346 : hugetlb_vma_assert_locked(vma);
347 : spin_lock(ptl);
348 : pte = huge_ptep_get(ptep);
349 :
350 : if (unlikely(!is_hugetlb_entry_migration(pte))) {
351 : spin_unlock(ptl);
352 : hugetlb_vma_unlock_read(vma);
353 : } else {
354 : /*
355 : * If migration entry existed, safe to release vma lock
356 : * here because the pgtable page won't be freed without the
357 : * pgtable lock released. See comment right above pgtable
358 : * lock release in migration_entry_wait_on_locked().
359 : */
360 : hugetlb_vma_unlock_read(vma);
361 : migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
362 : }
363 : }
364 :
365 : void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte)
366 : {
367 : spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte);
368 :
369 : __migration_entry_wait_huge(vma, pte, ptl);
370 : }
371 : #endif
372 :
373 : #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
374 : void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
375 : {
376 : spinlock_t *ptl;
377 :
378 : ptl = pmd_lock(mm, pmd);
379 : if (!is_pmd_migration_entry(*pmd))
380 : goto unlock;
381 : migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
382 : return;
383 : unlock:
384 : spin_unlock(ptl);
385 : }
386 : #endif
387 :
388 : static int folio_expected_refs(struct address_space *mapping,
389 : struct folio *folio)
390 : {
391 0 : int refs = 1;
392 0 : if (!mapping)
393 : return refs;
394 :
395 0 : refs += folio_nr_pages(folio);
396 0 : if (folio_test_private(folio))
397 0 : refs++;
398 :
399 : return refs;
400 : }
401 :
402 : /*
403 : * Replace the page in the mapping.
404 : *
405 : * The number of remaining references must be:
406 : * 1 for anonymous pages without a mapping
407 : * 2 for pages with a mapping
408 : * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
409 : */
410 0 : int folio_migrate_mapping(struct address_space *mapping,
411 : struct folio *newfolio, struct folio *folio, int extra_count)
412 : {
413 0 : XA_STATE(xas, &mapping->i_pages, folio_index(folio));
414 : struct zone *oldzone, *newzone;
415 : int dirty;
416 0 : int expected_count = folio_expected_refs(mapping, folio) + extra_count;
417 0 : long nr = folio_nr_pages(folio);
418 :
419 0 : if (!mapping) {
420 : /* Anonymous page without mapping */
421 0 : if (folio_ref_count(folio) != expected_count)
422 : return -EAGAIN;
423 :
424 : /* No turning back from here */
425 0 : newfolio->index = folio->index;
426 0 : newfolio->mapping = folio->mapping;
427 0 : if (folio_test_swapbacked(folio))
428 : __folio_set_swapbacked(newfolio);
429 :
430 : return MIGRATEPAGE_SUCCESS;
431 : }
432 :
433 0 : oldzone = folio_zone(folio);
434 0 : newzone = folio_zone(newfolio);
435 :
436 0 : xas_lock_irq(&xas);
437 0 : if (!folio_ref_freeze(folio, expected_count)) {
438 0 : xas_unlock_irq(&xas);
439 0 : return -EAGAIN;
440 : }
441 :
442 : /*
443 : * Now we know that no one else is looking at the folio:
444 : * no turning back from here.
445 : */
446 0 : newfolio->index = folio->index;
447 0 : newfolio->mapping = folio->mapping;
448 0 : folio_ref_add(newfolio, nr); /* add cache reference */
449 0 : if (folio_test_swapbacked(folio)) {
450 0 : __folio_set_swapbacked(newfolio);
451 0 : if (folio_test_swapcache(folio)) {
452 0 : folio_set_swapcache(newfolio);
453 0 : newfolio->private = folio_get_private(folio);
454 : }
455 : } else {
456 : VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
457 : }
458 :
459 : /* Move dirty while page refs frozen and newpage not yet exposed */
460 0 : dirty = folio_test_dirty(folio);
461 0 : if (dirty) {
462 0 : folio_clear_dirty(folio);
463 : folio_set_dirty(newfolio);
464 : }
465 :
466 0 : xas_store(&xas, newfolio);
467 :
468 : /*
469 : * Drop cache reference from old page by unfreezing
470 : * to one less reference.
471 : * We know this isn't the last reference.
472 : */
473 0 : folio_ref_unfreeze(folio, expected_count - nr);
474 :
475 0 : xas_unlock(&xas);
476 : /* Leave irq disabled to prevent preemption while updating stats */
477 :
478 : /*
479 : * If moved to a different zone then also account
480 : * the page for that zone. Other VM counters will be
481 : * taken care of when we establish references to the
482 : * new page and drop references to the old page.
483 : *
484 : * Note that anonymous pages are accounted for
485 : * via NR_FILE_PAGES and NR_ANON_MAPPED if they
486 : * are mapped to swap space.
487 : */
488 0 : if (newzone != oldzone) {
489 : struct lruvec *old_lruvec, *new_lruvec;
490 : struct mem_cgroup *memcg;
491 :
492 0 : memcg = folio_memcg(folio);
493 0 : old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
494 0 : new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
495 :
496 0 : __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
497 0 : __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
498 0 : if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
499 0 : __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
500 0 : __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
501 : }
502 : #ifdef CONFIG_SWAP
503 0 : if (folio_test_swapcache(folio)) {
504 0 : __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
505 0 : __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
506 : }
507 : #endif
508 0 : if (dirty && mapping_can_writeback(mapping)) {
509 0 : __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
510 0 : __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
511 0 : __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
512 : __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
513 : }
514 : }
515 : local_irq_enable();
516 :
517 0 : return MIGRATEPAGE_SUCCESS;
518 : }
519 : EXPORT_SYMBOL(folio_migrate_mapping);
520 :
521 : /*
522 : * The expected number of remaining references is the same as that
523 : * of folio_migrate_mapping().
524 : */
525 0 : int migrate_huge_page_move_mapping(struct address_space *mapping,
526 : struct folio *dst, struct folio *src)
527 : {
528 0 : XA_STATE(xas, &mapping->i_pages, folio_index(src));
529 : int expected_count;
530 :
531 0 : xas_lock_irq(&xas);
532 0 : expected_count = 2 + folio_has_private(src);
533 0 : if (!folio_ref_freeze(src, expected_count)) {
534 0 : xas_unlock_irq(&xas);
535 0 : return -EAGAIN;
536 : }
537 :
538 0 : dst->index = src->index;
539 0 : dst->mapping = src->mapping;
540 :
541 0 : folio_get(dst);
542 :
543 0 : xas_store(&xas, dst);
544 :
545 0 : folio_ref_unfreeze(src, expected_count - 1);
546 :
547 0 : xas_unlock_irq(&xas);
548 :
549 0 : return MIGRATEPAGE_SUCCESS;
550 : }
551 :
552 : /*
553 : * Copy the flags and some other ancillary information
554 : */
555 0 : void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
556 : {
557 : int cpupid;
558 :
559 0 : if (folio_test_error(folio))
560 : folio_set_error(newfolio);
561 0 : if (folio_test_referenced(folio))
562 : folio_set_referenced(newfolio);
563 0 : if (folio_test_uptodate(folio))
564 : folio_mark_uptodate(newfolio);
565 0 : if (folio_test_clear_active(folio)) {
566 : VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
567 : folio_set_active(newfolio);
568 0 : } else if (folio_test_clear_unevictable(folio))
569 : folio_set_unevictable(newfolio);
570 0 : if (folio_test_workingset(folio))
571 : folio_set_workingset(newfolio);
572 0 : if (folio_test_checked(folio))
573 : folio_set_checked(newfolio);
574 : /*
575 : * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
576 : * migration entries. We can still have PG_anon_exclusive set on an
577 : * effectively unmapped and unreferenced first sub-pages of an
578 : * anonymous THP: we can simply copy it here via PG_mappedtodisk.
579 : */
580 0 : if (folio_test_mappedtodisk(folio))
581 : folio_set_mappedtodisk(newfolio);
582 :
583 : /* Move dirty on pages not done by folio_migrate_mapping() */
584 0 : if (folio_test_dirty(folio))
585 : folio_set_dirty(newfolio);
586 :
587 0 : if (folio_test_young(folio))
588 : folio_set_young(newfolio);
589 0 : if (folio_test_idle(folio))
590 : folio_set_idle(newfolio);
591 :
592 : /*
593 : * Copy NUMA information to the new page, to prevent over-eager
594 : * future migrations of this same page.
595 : */
596 0 : cpupid = page_cpupid_xchg_last(&folio->page, -1);
597 : /*
598 : * For memory tiering mode, when migrate between slow and fast
599 : * memory node, reset cpupid, because that is used to record
600 : * page access time in slow memory node.
601 : */
602 : if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
603 : bool f_toptier = node_is_toptier(page_to_nid(&folio->page));
604 : bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page));
605 :
606 : if (f_toptier != t_toptier)
607 : cpupid = -1;
608 : }
609 0 : page_cpupid_xchg_last(&newfolio->page, cpupid);
610 :
611 0 : folio_migrate_ksm(newfolio, folio);
612 : /*
613 : * Please do not reorder this without considering how mm/ksm.c's
614 : * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
615 : */
616 0 : if (folio_test_swapcache(folio))
617 : folio_clear_swapcache(folio);
618 0 : folio_clear_private(folio);
619 :
620 : /* page->private contains hugetlb specific flags */
621 0 : if (!folio_test_hugetlb(folio))
622 0 : folio->private = NULL;
623 :
624 : /*
625 : * If any waiters have accumulated on the new page then
626 : * wake them up.
627 : */
628 0 : if (folio_test_writeback(newfolio))
629 0 : folio_end_writeback(newfolio);
630 :
631 : /*
632 : * PG_readahead shares the same bit with PG_reclaim. The above
633 : * end_page_writeback() may clear PG_readahead mistakenly, so set the
634 : * bit after that.
635 : */
636 0 : if (folio_test_readahead(folio))
637 : folio_set_readahead(newfolio);
638 :
639 0 : folio_copy_owner(newfolio, folio);
640 :
641 0 : if (!folio_test_hugetlb(folio))
642 : mem_cgroup_migrate(folio, newfolio);
643 0 : }
644 : EXPORT_SYMBOL(folio_migrate_flags);
645 :
646 0 : void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
647 : {
648 0 : folio_copy(newfolio, folio);
649 0 : folio_migrate_flags(newfolio, folio);
650 0 : }
651 : EXPORT_SYMBOL(folio_migrate_copy);
652 :
653 : /************************************************************
654 : * Migration functions
655 : ***********************************************************/
656 :
657 0 : int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
658 : struct folio *src, enum migrate_mode mode, int extra_count)
659 : {
660 : int rc;
661 :
662 0 : BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
663 :
664 0 : rc = folio_migrate_mapping(mapping, dst, src, extra_count);
665 :
666 0 : if (rc != MIGRATEPAGE_SUCCESS)
667 : return rc;
668 :
669 0 : if (mode != MIGRATE_SYNC_NO_COPY)
670 : folio_migrate_copy(dst, src);
671 : else
672 0 : folio_migrate_flags(dst, src);
673 : return MIGRATEPAGE_SUCCESS;
674 : }
675 :
676 : /**
677 : * migrate_folio() - Simple folio migration.
678 : * @mapping: The address_space containing the folio.
679 : * @dst: The folio to migrate the data to.
680 : * @src: The folio containing the current data.
681 : * @mode: How to migrate the page.
682 : *
683 : * Common logic to directly migrate a single LRU folio suitable for
684 : * folios that do not use PagePrivate/PagePrivate2.
685 : *
686 : * Folios are locked upon entry and exit.
687 : */
688 0 : int migrate_folio(struct address_space *mapping, struct folio *dst,
689 : struct folio *src, enum migrate_mode mode)
690 : {
691 0 : return migrate_folio_extra(mapping, dst, src, mode, 0);
692 : }
693 : EXPORT_SYMBOL(migrate_folio);
694 :
695 : #ifdef CONFIG_BLOCK
696 : /* Returns true if all buffers are successfully locked */
697 0 : static bool buffer_migrate_lock_buffers(struct buffer_head *head,
698 : enum migrate_mode mode)
699 : {
700 0 : struct buffer_head *bh = head;
701 :
702 : /* Simple case, sync compaction */
703 0 : if (mode != MIGRATE_ASYNC) {
704 : do {
705 0 : lock_buffer(bh);
706 0 : bh = bh->b_this_page;
707 :
708 0 : } while (bh != head);
709 :
710 : return true;
711 : }
712 :
713 : /* async case, we cannot block on lock_buffer so use trylock_buffer */
714 : do {
715 0 : if (!trylock_buffer(bh)) {
716 : /*
717 : * We failed to lock the buffer and cannot stall in
718 : * async migration. Release the taken locks
719 : */
720 : struct buffer_head *failed_bh = bh;
721 : bh = head;
722 0 : while (bh != failed_bh) {
723 0 : unlock_buffer(bh);
724 0 : bh = bh->b_this_page;
725 : }
726 : return false;
727 : }
728 :
729 0 : bh = bh->b_this_page;
730 0 : } while (bh != head);
731 : return true;
732 : }
733 :
734 0 : static int __buffer_migrate_folio(struct address_space *mapping,
735 : struct folio *dst, struct folio *src, enum migrate_mode mode,
736 : bool check_refs)
737 : {
738 : struct buffer_head *bh, *head;
739 : int rc;
740 : int expected_count;
741 :
742 0 : head = folio_buffers(src);
743 0 : if (!head)
744 0 : return migrate_folio(mapping, dst, src, mode);
745 :
746 : /* Check whether page does not have extra refs before we do more work */
747 0 : expected_count = folio_expected_refs(mapping, src);
748 0 : if (folio_ref_count(src) != expected_count)
749 : return -EAGAIN;
750 :
751 0 : if (!buffer_migrate_lock_buffers(head, mode))
752 : return -EAGAIN;
753 :
754 0 : if (check_refs) {
755 : bool busy;
756 : bool invalidated = false;
757 :
758 : recheck_buffers:
759 0 : busy = false;
760 0 : spin_lock(&mapping->private_lock);
761 0 : bh = head;
762 : do {
763 0 : if (atomic_read(&bh->b_count)) {
764 : busy = true;
765 : break;
766 : }
767 0 : bh = bh->b_this_page;
768 0 : } while (bh != head);
769 0 : if (busy) {
770 0 : if (invalidated) {
771 : rc = -EAGAIN;
772 : goto unlock_buffers;
773 : }
774 0 : spin_unlock(&mapping->private_lock);
775 0 : invalidate_bh_lrus();
776 0 : invalidated = true;
777 0 : goto recheck_buffers;
778 : }
779 : }
780 :
781 0 : rc = folio_migrate_mapping(mapping, dst, src, 0);
782 0 : if (rc != MIGRATEPAGE_SUCCESS)
783 : goto unlock_buffers;
784 :
785 0 : folio_attach_private(dst, folio_detach_private(src));
786 :
787 0 : bh = head;
788 : do {
789 0 : set_bh_page(bh, &dst->page, bh_offset(bh));
790 0 : bh = bh->b_this_page;
791 0 : } while (bh != head);
792 :
793 0 : if (mode != MIGRATE_SYNC_NO_COPY)
794 : folio_migrate_copy(dst, src);
795 : else
796 0 : folio_migrate_flags(dst, src);
797 :
798 : rc = MIGRATEPAGE_SUCCESS;
799 : unlock_buffers:
800 0 : if (check_refs)
801 0 : spin_unlock(&mapping->private_lock);
802 : bh = head;
803 : do {
804 0 : unlock_buffer(bh);
805 0 : bh = bh->b_this_page;
806 0 : } while (bh != head);
807 :
808 : return rc;
809 : }
810 :
811 : /**
812 : * buffer_migrate_folio() - Migration function for folios with buffers.
813 : * @mapping: The address space containing @src.
814 : * @dst: The folio to migrate to.
815 : * @src: The folio to migrate from.
816 : * @mode: How to migrate the folio.
817 : *
818 : * This function can only be used if the underlying filesystem guarantees
819 : * that no other references to @src exist. For example attached buffer
820 : * heads are accessed only under the folio lock. If your filesystem cannot
821 : * provide this guarantee, buffer_migrate_folio_norefs() may be more
822 : * appropriate.
823 : *
824 : * Return: 0 on success or a negative errno on failure.
825 : */
826 0 : int buffer_migrate_folio(struct address_space *mapping,
827 : struct folio *dst, struct folio *src, enum migrate_mode mode)
828 : {
829 0 : return __buffer_migrate_folio(mapping, dst, src, mode, false);
830 : }
831 : EXPORT_SYMBOL(buffer_migrate_folio);
832 :
833 : /**
834 : * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
835 : * @mapping: The address space containing @src.
836 : * @dst: The folio to migrate to.
837 : * @src: The folio to migrate from.
838 : * @mode: How to migrate the folio.
839 : *
840 : * Like buffer_migrate_folio() except that this variant is more careful
841 : * and checks that there are also no buffer head references. This function
842 : * is the right one for mappings where buffer heads are directly looked
843 : * up and referenced (such as block device mappings).
844 : *
845 : * Return: 0 on success or a negative errno on failure.
846 : */
847 0 : int buffer_migrate_folio_norefs(struct address_space *mapping,
848 : struct folio *dst, struct folio *src, enum migrate_mode mode)
849 : {
850 0 : return __buffer_migrate_folio(mapping, dst, src, mode, true);
851 : }
852 : EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
853 : #endif
854 :
855 0 : int filemap_migrate_folio(struct address_space *mapping,
856 : struct folio *dst, struct folio *src, enum migrate_mode mode)
857 : {
858 : int ret;
859 :
860 0 : ret = folio_migrate_mapping(mapping, dst, src, 0);
861 0 : if (ret != MIGRATEPAGE_SUCCESS)
862 : return ret;
863 :
864 0 : if (folio_get_private(src))
865 0 : folio_attach_private(dst, folio_detach_private(src));
866 :
867 0 : if (mode != MIGRATE_SYNC_NO_COPY)
868 : folio_migrate_copy(dst, src);
869 : else
870 0 : folio_migrate_flags(dst, src);
871 : return MIGRATEPAGE_SUCCESS;
872 : }
873 : EXPORT_SYMBOL_GPL(filemap_migrate_folio);
874 :
875 : /*
876 : * Writeback a folio to clean the dirty state
877 : */
878 0 : static int writeout(struct address_space *mapping, struct folio *folio)
879 : {
880 0 : struct writeback_control wbc = {
881 : .sync_mode = WB_SYNC_NONE,
882 : .nr_to_write = 1,
883 : .range_start = 0,
884 : .range_end = LLONG_MAX,
885 : .for_reclaim = 1
886 : };
887 : int rc;
888 :
889 0 : if (!mapping->a_ops->writepage)
890 : /* No write method for the address space */
891 : return -EINVAL;
892 :
893 0 : if (!folio_clear_dirty_for_io(folio))
894 : /* Someone else already triggered a write */
895 : return -EAGAIN;
896 :
897 : /*
898 : * A dirty folio may imply that the underlying filesystem has
899 : * the folio on some queue. So the folio must be clean for
900 : * migration. Writeout may mean we lose the lock and the
901 : * folio state is no longer what we checked for earlier.
902 : * At this point we know that the migration attempt cannot
903 : * be successful.
904 : */
905 0 : remove_migration_ptes(folio, folio, false);
906 :
907 0 : rc = mapping->a_ops->writepage(&folio->page, &wbc);
908 :
909 0 : if (rc != AOP_WRITEPAGE_ACTIVATE)
910 : /* unlocked. Relock */
911 : folio_lock(folio);
912 :
913 0 : return (rc < 0) ? -EIO : -EAGAIN;
914 : }
915 :
916 : /*
917 : * Default handling if a filesystem does not provide a migration function.
918 : */
919 0 : static int fallback_migrate_folio(struct address_space *mapping,
920 : struct folio *dst, struct folio *src, enum migrate_mode mode)
921 : {
922 0 : if (folio_test_dirty(src)) {
923 : /* Only writeback folios in full synchronous migration */
924 0 : switch (mode) {
925 : case MIGRATE_SYNC:
926 : case MIGRATE_SYNC_NO_COPY:
927 : break;
928 : default:
929 : return -EBUSY;
930 : }
931 0 : return writeout(mapping, src);
932 : }
933 :
934 : /*
935 : * Buffers may be managed in a filesystem specific way.
936 : * We must have no buffers or drop them.
937 : */
938 0 : if (folio_test_private(src) &&
939 0 : !filemap_release_folio(src, GFP_KERNEL))
940 0 : return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
941 :
942 0 : return migrate_folio(mapping, dst, src, mode);
943 : }
944 :
945 : /*
946 : * Move a page to a newly allocated page
947 : * The page is locked and all ptes have been successfully removed.
948 : *
949 : * The new page will have replaced the old page if this function
950 : * is successful.
951 : *
952 : * Return value:
953 : * < 0 - error code
954 : * MIGRATEPAGE_SUCCESS - success
955 : */
956 0 : static int move_to_new_folio(struct folio *dst, struct folio *src,
957 : enum migrate_mode mode)
958 : {
959 0 : int rc = -EAGAIN;
960 0 : bool is_lru = !__PageMovable(&src->page);
961 :
962 : VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
963 : VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
964 :
965 0 : if (likely(is_lru)) {
966 0 : struct address_space *mapping = folio_mapping(src);
967 :
968 0 : if (!mapping)
969 0 : rc = migrate_folio(mapping, dst, src, mode);
970 0 : else if (mapping->a_ops->migrate_folio)
971 : /*
972 : * Most folios have a mapping and most filesystems
973 : * provide a migrate_folio callback. Anonymous folios
974 : * are part of swap space which also has its own
975 : * migrate_folio callback. This is the most common path
976 : * for page migration.
977 : */
978 0 : rc = mapping->a_ops->migrate_folio(mapping, dst, src,
979 : mode);
980 : else
981 0 : rc = fallback_migrate_folio(mapping, dst, src, mode);
982 : } else {
983 : const struct movable_operations *mops;
984 :
985 : /*
986 : * In case of non-lru page, it could be released after
987 : * isolation step. In that case, we shouldn't try migration.
988 : */
989 : VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
990 0 : if (!folio_test_movable(src)) {
991 0 : rc = MIGRATEPAGE_SUCCESS;
992 : folio_clear_isolated(src);
993 : goto out;
994 : }
995 :
996 0 : mops = folio_movable_ops(src);
997 0 : rc = mops->migrate_page(&dst->page, &src->page, mode);
998 0 : WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
999 : !folio_test_isolated(src));
1000 : }
1001 :
1002 : /*
1003 : * When successful, old pagecache src->mapping must be cleared before
1004 : * src is freed; but stats require that PageAnon be left as PageAnon.
1005 : */
1006 0 : if (rc == MIGRATEPAGE_SUCCESS) {
1007 0 : if (__PageMovable(&src->page)) {
1008 : VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1009 :
1010 : /*
1011 : * We clear PG_movable under page_lock so any compactor
1012 : * cannot try to migrate this page.
1013 : */
1014 : folio_clear_isolated(src);
1015 : }
1016 :
1017 : /*
1018 : * Anonymous and movable src->mapping will be cleared by
1019 : * free_pages_prepare so don't reset it here for keeping
1020 : * the type to work PageAnon, for example.
1021 : */
1022 0 : if (!folio_mapping_flags(src))
1023 0 : src->mapping = NULL;
1024 :
1025 : if (likely(!folio_is_zone_device(dst)))
1026 : flush_dcache_folio(dst);
1027 : }
1028 : out:
1029 0 : return rc;
1030 : }
1031 :
1032 : /*
1033 : * To record some information during migration, we use some unused
1034 : * fields (mapping and private) of struct folio of the newly allocated
1035 : * destination folio. This is safe because nobody is using them
1036 : * except us.
1037 : */
1038 : union migration_ptr {
1039 : struct anon_vma *anon_vma;
1040 : struct address_space *mapping;
1041 : };
1042 : static void __migrate_folio_record(struct folio *dst,
1043 : unsigned long page_was_mapped,
1044 : struct anon_vma *anon_vma)
1045 : {
1046 0 : union migration_ptr ptr = { .anon_vma = anon_vma };
1047 0 : dst->mapping = ptr.mapping;
1048 0 : dst->private = (void *)page_was_mapped;
1049 : }
1050 :
1051 : static void __migrate_folio_extract(struct folio *dst,
1052 : int *page_was_mappedp,
1053 : struct anon_vma **anon_vmap)
1054 : {
1055 0 : union migration_ptr ptr = { .mapping = dst->mapping };
1056 0 : *anon_vmap = ptr.anon_vma;
1057 0 : *page_was_mappedp = (unsigned long)dst->private;
1058 0 : dst->mapping = NULL;
1059 0 : dst->private = NULL;
1060 : }
1061 :
1062 : /* Restore the source folio to the original state upon failure */
1063 0 : static void migrate_folio_undo_src(struct folio *src,
1064 : int page_was_mapped,
1065 : struct anon_vma *anon_vma,
1066 : bool locked,
1067 : struct list_head *ret)
1068 : {
1069 0 : if (page_was_mapped)
1070 0 : remove_migration_ptes(src, src, false);
1071 : /* Drop an anon_vma reference if we took one */
1072 0 : if (anon_vma)
1073 : put_anon_vma(anon_vma);
1074 0 : if (locked)
1075 0 : folio_unlock(src);
1076 0 : if (ret)
1077 0 : list_move_tail(&src->lru, ret);
1078 0 : }
1079 :
1080 : /* Restore the destination folio to the original state upon failure */
1081 0 : static void migrate_folio_undo_dst(struct folio *dst,
1082 : bool locked,
1083 : free_page_t put_new_page,
1084 : unsigned long private)
1085 : {
1086 0 : if (locked)
1087 0 : folio_unlock(dst);
1088 0 : if (put_new_page)
1089 0 : put_new_page(&dst->page, private);
1090 : else
1091 : folio_put(dst);
1092 0 : }
1093 :
1094 : /* Cleanup src folio upon migration success */
1095 0 : static void migrate_folio_done(struct folio *src,
1096 : enum migrate_reason reason)
1097 : {
1098 : /*
1099 : * Compaction can migrate also non-LRU pages which are
1100 : * not accounted to NR_ISOLATED_*. They can be recognized
1101 : * as __PageMovable
1102 : */
1103 0 : if (likely(!__folio_test_movable(src)))
1104 0 : mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1105 0 : folio_is_file_lru(src), -folio_nr_pages(src));
1106 :
1107 0 : if (reason != MR_MEMORY_FAILURE)
1108 : /* We release the page in page_handle_poison. */
1109 : folio_put(src);
1110 0 : }
1111 :
1112 : /* Obtain the lock on page, remove all ptes. */
1113 0 : static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page,
1114 : unsigned long private, struct folio *src,
1115 : struct folio **dstp, int force, bool avoid_force_lock,
1116 : enum migrate_mode mode, enum migrate_reason reason,
1117 : struct list_head *ret)
1118 : {
1119 : struct folio *dst;
1120 0 : int rc = -EAGAIN;
1121 0 : struct page *newpage = NULL;
1122 0 : int page_was_mapped = 0;
1123 0 : struct anon_vma *anon_vma = NULL;
1124 0 : bool is_lru = !__PageMovable(&src->page);
1125 0 : bool locked = false;
1126 0 : bool dst_locked = false;
1127 :
1128 0 : if (folio_ref_count(src) == 1) {
1129 : /* Folio was freed from under us. So we are done. */
1130 0 : folio_clear_active(src);
1131 0 : folio_clear_unevictable(src);
1132 : /* free_pages_prepare() will clear PG_isolated. */
1133 0 : list_del(&src->lru);
1134 0 : migrate_folio_done(src, reason);
1135 0 : return MIGRATEPAGE_SUCCESS;
1136 : }
1137 :
1138 0 : newpage = get_new_page(&src->page, private);
1139 0 : if (!newpage)
1140 : return -ENOMEM;
1141 0 : dst = page_folio(newpage);
1142 0 : *dstp = dst;
1143 :
1144 0 : dst->private = NULL;
1145 :
1146 0 : if (!folio_trylock(src)) {
1147 0 : if (!force || mode == MIGRATE_ASYNC)
1148 : goto out;
1149 :
1150 : /*
1151 : * It's not safe for direct compaction to call lock_page.
1152 : * For example, during page readahead pages are added locked
1153 : * to the LRU. Later, when the IO completes the pages are
1154 : * marked uptodate and unlocked. However, the queueing
1155 : * could be merging multiple pages for one bio (e.g.
1156 : * mpage_readahead). If an allocation happens for the
1157 : * second or third page, the process can end up locking
1158 : * the same page twice and deadlocking. Rather than
1159 : * trying to be clever about what pages can be locked,
1160 : * avoid the use of lock_page for direct compaction
1161 : * altogether.
1162 : */
1163 0 : if (current->flags & PF_MEMALLOC)
1164 : goto out;
1165 :
1166 : /*
1167 : * We have locked some folios and are going to wait to lock
1168 : * this folio. To avoid a potential deadlock, let's bail
1169 : * out and not do that. The locked folios will be moved and
1170 : * unlocked, then we can wait to lock this folio.
1171 : */
1172 0 : if (avoid_force_lock) {
1173 : rc = -EDEADLOCK;
1174 : goto out;
1175 : }
1176 :
1177 : folio_lock(src);
1178 : }
1179 0 : locked = true;
1180 :
1181 0 : if (folio_test_writeback(src)) {
1182 : /*
1183 : * Only in the case of a full synchronous migration is it
1184 : * necessary to wait for PageWriteback. In the async case,
1185 : * the retry loop is too short and in the sync-light case,
1186 : * the overhead of stalling is too much
1187 : */
1188 0 : switch (mode) {
1189 : case MIGRATE_SYNC:
1190 : case MIGRATE_SYNC_NO_COPY:
1191 : break;
1192 : default:
1193 : rc = -EBUSY;
1194 : goto out;
1195 : }
1196 0 : if (!force)
1197 : goto out;
1198 0 : folio_wait_writeback(src);
1199 : }
1200 :
1201 : /*
1202 : * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1203 : * we cannot notice that anon_vma is freed while we migrate a page.
1204 : * This get_anon_vma() delays freeing anon_vma pointer until the end
1205 : * of migration. File cache pages are no problem because of page_lock()
1206 : * File Caches may use write_page() or lock_page() in migration, then,
1207 : * just care Anon page here.
1208 : *
1209 : * Only folio_get_anon_vma() understands the subtleties of
1210 : * getting a hold on an anon_vma from outside one of its mms.
1211 : * But if we cannot get anon_vma, then we won't need it anyway,
1212 : * because that implies that the anon page is no longer mapped
1213 : * (and cannot be remapped so long as we hold the page lock).
1214 : */
1215 0 : if (folio_test_anon(src) && !folio_test_ksm(src))
1216 0 : anon_vma = folio_get_anon_vma(src);
1217 :
1218 : /*
1219 : * Block others from accessing the new page when we get around to
1220 : * establishing additional references. We are usually the only one
1221 : * holding a reference to dst at this point. We used to have a BUG
1222 : * here if folio_trylock(dst) fails, but would like to allow for
1223 : * cases where there might be a race with the previous use of dst.
1224 : * This is much like races on refcount of oldpage: just don't BUG().
1225 : */
1226 0 : if (unlikely(!folio_trylock(dst)))
1227 : goto out;
1228 0 : dst_locked = true;
1229 :
1230 0 : if (unlikely(!is_lru)) {
1231 0 : __migrate_folio_record(dst, page_was_mapped, anon_vma);
1232 0 : return MIGRATEPAGE_UNMAP;
1233 : }
1234 :
1235 : /*
1236 : * Corner case handling:
1237 : * 1. When a new swap-cache page is read into, it is added to the LRU
1238 : * and treated as swapcache but it has no rmap yet.
1239 : * Calling try_to_unmap() against a src->mapping==NULL page will
1240 : * trigger a BUG. So handle it here.
1241 : * 2. An orphaned page (see truncate_cleanup_page) might have
1242 : * fs-private metadata. The page can be picked up due to memory
1243 : * offlining. Everywhere else except page reclaim, the page is
1244 : * invisible to the vm, so the page can not be migrated. So try to
1245 : * free the metadata, so the page can be freed.
1246 : */
1247 0 : if (!src->mapping) {
1248 0 : if (folio_test_private(src)) {
1249 0 : try_to_free_buffers(src);
1250 0 : goto out;
1251 : }
1252 0 : } else if (folio_mapped(src)) {
1253 : /* Establish migration ptes */
1254 : VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1255 : !folio_test_ksm(src) && !anon_vma, src);
1256 0 : try_to_migrate(src, TTU_BATCH_FLUSH);
1257 0 : page_was_mapped = 1;
1258 : }
1259 :
1260 0 : if (!folio_mapped(src)) {
1261 0 : __migrate_folio_record(dst, page_was_mapped, anon_vma);
1262 0 : return MIGRATEPAGE_UNMAP;
1263 : }
1264 :
1265 : out:
1266 : /*
1267 : * A folio that has not been unmapped will be restored to
1268 : * right list unless we want to retry.
1269 : */
1270 0 : if (rc == -EAGAIN || rc == -EDEADLOCK)
1271 0 : ret = NULL;
1272 :
1273 0 : migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret);
1274 0 : migrate_folio_undo_dst(dst, dst_locked, put_new_page, private);
1275 :
1276 0 : return rc;
1277 : }
1278 :
1279 : /* Migrate the folio to the newly allocated folio in dst. */
1280 0 : static int migrate_folio_move(free_page_t put_new_page, unsigned long private,
1281 : struct folio *src, struct folio *dst,
1282 : enum migrate_mode mode, enum migrate_reason reason,
1283 : struct list_head *ret)
1284 : {
1285 : int rc;
1286 0 : int page_was_mapped = 0;
1287 0 : struct anon_vma *anon_vma = NULL;
1288 0 : bool is_lru = !__PageMovable(&src->page);
1289 : struct list_head *prev;
1290 :
1291 0 : __migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
1292 0 : prev = dst->lru.prev;
1293 0 : list_del(&dst->lru);
1294 :
1295 0 : rc = move_to_new_folio(dst, src, mode);
1296 0 : if (rc)
1297 : goto out;
1298 :
1299 0 : if (unlikely(!is_lru))
1300 : goto out_unlock_both;
1301 :
1302 : /*
1303 : * When successful, push dst to LRU immediately: so that if it
1304 : * turns out to be an mlocked page, remove_migration_ptes() will
1305 : * automatically build up the correct dst->mlock_count for it.
1306 : *
1307 : * We would like to do something similar for the old page, when
1308 : * unsuccessful, and other cases when a page has been temporarily
1309 : * isolated from the unevictable LRU: but this case is the easiest.
1310 : */
1311 0 : folio_add_lru(dst);
1312 0 : if (page_was_mapped)
1313 0 : lru_add_drain();
1314 :
1315 0 : if (page_was_mapped)
1316 0 : remove_migration_ptes(src, dst, false);
1317 :
1318 : out_unlock_both:
1319 0 : folio_unlock(dst);
1320 0 : set_page_owner_migrate_reason(&dst->page, reason);
1321 : /*
1322 : * If migration is successful, decrease refcount of dst,
1323 : * which will not free the page because new page owner increased
1324 : * refcounter.
1325 : */
1326 0 : folio_put(dst);
1327 :
1328 : /*
1329 : * A folio that has been migrated has all references removed
1330 : * and will be freed.
1331 : */
1332 0 : list_del(&src->lru);
1333 : /* Drop an anon_vma reference if we took one */
1334 0 : if (anon_vma)
1335 0 : put_anon_vma(anon_vma);
1336 0 : folio_unlock(src);
1337 0 : migrate_folio_done(src, reason);
1338 :
1339 0 : return rc;
1340 : out:
1341 : /*
1342 : * A folio that has not been migrated will be restored to
1343 : * right list unless we want to retry.
1344 : */
1345 0 : if (rc == -EAGAIN) {
1346 0 : list_add(&dst->lru, prev);
1347 0 : __migrate_folio_record(dst, page_was_mapped, anon_vma);
1348 0 : return rc;
1349 : }
1350 :
1351 0 : migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret);
1352 0 : migrate_folio_undo_dst(dst, true, put_new_page, private);
1353 :
1354 0 : return rc;
1355 : }
1356 :
1357 : /*
1358 : * Counterpart of unmap_and_move_page() for hugepage migration.
1359 : *
1360 : * This function doesn't wait the completion of hugepage I/O
1361 : * because there is no race between I/O and migration for hugepage.
1362 : * Note that currently hugepage I/O occurs only in direct I/O
1363 : * where no lock is held and PG_writeback is irrelevant,
1364 : * and writeback status of all subpages are counted in the reference
1365 : * count of the head page (i.e. if all subpages of a 2MB hugepage are
1366 : * under direct I/O, the reference of the head page is 512 and a bit more.)
1367 : * This means that when we try to migrate hugepage whose subpages are
1368 : * doing direct I/O, some references remain after try_to_unmap() and
1369 : * hugepage migration fails without data corruption.
1370 : *
1371 : * There is also no race when direct I/O is issued on the page under migration,
1372 : * because then pte is replaced with migration swap entry and direct I/O code
1373 : * will wait in the page fault for migration to complete.
1374 : */
1375 : static int unmap_and_move_huge_page(new_page_t get_new_page,
1376 : free_page_t put_new_page, unsigned long private,
1377 : struct page *hpage, int force,
1378 : enum migrate_mode mode, int reason,
1379 : struct list_head *ret)
1380 : {
1381 : struct folio *dst, *src = page_folio(hpage);
1382 : int rc = -EAGAIN;
1383 : int page_was_mapped = 0;
1384 : struct page *new_hpage;
1385 : struct anon_vma *anon_vma = NULL;
1386 : struct address_space *mapping = NULL;
1387 :
1388 : if (folio_ref_count(src) == 1) {
1389 : /* page was freed from under us. So we are done. */
1390 : folio_putback_active_hugetlb(src);
1391 : return MIGRATEPAGE_SUCCESS;
1392 : }
1393 :
1394 : new_hpage = get_new_page(hpage, private);
1395 : if (!new_hpage)
1396 : return -ENOMEM;
1397 : dst = page_folio(new_hpage);
1398 :
1399 : if (!folio_trylock(src)) {
1400 : if (!force)
1401 : goto out;
1402 : switch (mode) {
1403 : case MIGRATE_SYNC:
1404 : case MIGRATE_SYNC_NO_COPY:
1405 : break;
1406 : default:
1407 : goto out;
1408 : }
1409 : folio_lock(src);
1410 : }
1411 :
1412 : /*
1413 : * Check for pages which are in the process of being freed. Without
1414 : * folio_mapping() set, hugetlbfs specific move page routine will not
1415 : * be called and we could leak usage counts for subpools.
1416 : */
1417 : if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1418 : rc = -EBUSY;
1419 : goto out_unlock;
1420 : }
1421 :
1422 : if (folio_test_anon(src))
1423 : anon_vma = folio_get_anon_vma(src);
1424 :
1425 : if (unlikely(!folio_trylock(dst)))
1426 : goto put_anon;
1427 :
1428 : if (folio_mapped(src)) {
1429 : enum ttu_flags ttu = 0;
1430 :
1431 : if (!folio_test_anon(src)) {
1432 : /*
1433 : * In shared mappings, try_to_unmap could potentially
1434 : * call huge_pmd_unshare. Because of this, take
1435 : * semaphore in write mode here and set TTU_RMAP_LOCKED
1436 : * to let lower levels know we have taken the lock.
1437 : */
1438 : mapping = hugetlb_page_mapping_lock_write(hpage);
1439 : if (unlikely(!mapping))
1440 : goto unlock_put_anon;
1441 :
1442 : ttu = TTU_RMAP_LOCKED;
1443 : }
1444 :
1445 : try_to_migrate(src, ttu);
1446 : page_was_mapped = 1;
1447 :
1448 : if (ttu & TTU_RMAP_LOCKED)
1449 : i_mmap_unlock_write(mapping);
1450 : }
1451 :
1452 : if (!folio_mapped(src))
1453 : rc = move_to_new_folio(dst, src, mode);
1454 :
1455 : if (page_was_mapped)
1456 : remove_migration_ptes(src,
1457 : rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1458 :
1459 : unlock_put_anon:
1460 : folio_unlock(dst);
1461 :
1462 : put_anon:
1463 : if (anon_vma)
1464 : put_anon_vma(anon_vma);
1465 :
1466 : if (rc == MIGRATEPAGE_SUCCESS) {
1467 : move_hugetlb_state(src, dst, reason);
1468 : put_new_page = NULL;
1469 : }
1470 :
1471 : out_unlock:
1472 : folio_unlock(src);
1473 : out:
1474 : if (rc == MIGRATEPAGE_SUCCESS)
1475 : folio_putback_active_hugetlb(src);
1476 : else if (rc != -EAGAIN)
1477 : list_move_tail(&src->lru, ret);
1478 :
1479 : /*
1480 : * If migration was not successful and there's a freeing callback, use
1481 : * it. Otherwise, put_page() will drop the reference grabbed during
1482 : * isolation.
1483 : */
1484 : if (put_new_page)
1485 : put_new_page(new_hpage, private);
1486 : else
1487 : folio_putback_active_hugetlb(dst);
1488 :
1489 : return rc;
1490 : }
1491 :
1492 0 : static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1493 : {
1494 : int rc;
1495 :
1496 0 : folio_lock(folio);
1497 0 : rc = split_folio_to_list(folio, split_folios);
1498 0 : folio_unlock(folio);
1499 : if (!rc)
1500 0 : list_move_tail(&folio->lru, split_folios);
1501 :
1502 0 : return rc;
1503 : }
1504 :
1505 : #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1506 : #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1507 : #else
1508 : #define NR_MAX_BATCHED_MIGRATION 512
1509 : #endif
1510 : #define NR_MAX_MIGRATE_PAGES_RETRY 10
1511 :
1512 : struct migrate_pages_stats {
1513 : int nr_succeeded; /* Normal and large folios migrated successfully, in
1514 : units of base pages */
1515 : int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1516 : units of base pages. Untried folios aren't counted */
1517 : int nr_thp_succeeded; /* THP migrated successfully */
1518 : int nr_thp_failed; /* THP failed to be migrated */
1519 : int nr_thp_split; /* THP split before migrating */
1520 : };
1521 :
1522 : /*
1523 : * Returns the number of hugetlb folios that were not migrated, or an error code
1524 : * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1525 : * any more because the list has become empty or no retryable hugetlb folios
1526 : * exist any more. It is caller's responsibility to call putback_movable_pages()
1527 : * only if ret != 0.
1528 : */
1529 : static int migrate_hugetlbs(struct list_head *from, new_page_t get_new_page,
1530 : free_page_t put_new_page, unsigned long private,
1531 : enum migrate_mode mode, int reason,
1532 : struct migrate_pages_stats *stats,
1533 : struct list_head *ret_folios)
1534 : {
1535 : int retry = 1;
1536 : int nr_failed = 0;
1537 : int nr_retry_pages = 0;
1538 : int pass = 0;
1539 : struct folio *folio, *folio2;
1540 : int rc, nr_pages;
1541 :
1542 0 : for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1543 0 : retry = 0;
1544 0 : nr_retry_pages = 0;
1545 :
1546 0 : list_for_each_entry_safe(folio, folio2, from, lru) {
1547 0 : if (!folio_test_hugetlb(folio))
1548 0 : continue;
1549 :
1550 : nr_pages = folio_nr_pages(folio);
1551 :
1552 : cond_resched();
1553 :
1554 : /*
1555 : * Migratability of hugepages depends on architectures and
1556 : * their size. This check is necessary because some callers
1557 : * of hugepage migration like soft offline and memory
1558 : * hotremove don't walk through page tables or check whether
1559 : * the hugepage is pmd-based or not before kicking migration.
1560 : */
1561 : if (!hugepage_migration_supported(folio_hstate(folio))) {
1562 : nr_failed++;
1563 : stats->nr_failed_pages += nr_pages;
1564 : list_move_tail(&folio->lru, ret_folios);
1565 : continue;
1566 : }
1567 :
1568 : rc = unmap_and_move_huge_page(get_new_page,
1569 : put_new_page, private,
1570 : &folio->page, pass > 2, mode,
1571 : reason, ret_folios);
1572 : /*
1573 : * The rules are:
1574 : * Success: hugetlb folio will be put back
1575 : * -EAGAIN: stay on the from list
1576 : * -ENOMEM: stay on the from list
1577 : * Other errno: put on ret_folios list
1578 : */
1579 : switch(rc) {
1580 : case -ENOMEM:
1581 : /*
1582 : * When memory is low, don't bother to try to migrate
1583 : * other folios, just exit.
1584 : */
1585 : stats->nr_failed_pages += nr_pages + nr_retry_pages;
1586 : return -ENOMEM;
1587 : case -EAGAIN:
1588 : retry++;
1589 : nr_retry_pages += nr_pages;
1590 : break;
1591 : case MIGRATEPAGE_SUCCESS:
1592 : stats->nr_succeeded += nr_pages;
1593 : break;
1594 : default:
1595 : /*
1596 : * Permanent failure (-EBUSY, etc.):
1597 : * unlike -EAGAIN case, the failed folio is
1598 : * removed from migration folio list and not
1599 : * retried in the next outer loop.
1600 : */
1601 : nr_failed++;
1602 : stats->nr_failed_pages += nr_pages;
1603 : break;
1604 : }
1605 : }
1606 : }
1607 : /*
1608 : * nr_failed is number of hugetlb folios failed to be migrated. After
1609 : * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1610 : * folios as failed.
1611 : */
1612 : nr_failed += retry;
1613 : stats->nr_failed_pages += nr_retry_pages;
1614 :
1615 : return nr_failed;
1616 : }
1617 :
1618 : /*
1619 : * migrate_pages_batch() first unmaps folios in the from list as many as
1620 : * possible, then move the unmapped folios.
1621 : */
1622 0 : static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
1623 : free_page_t put_new_page, unsigned long private,
1624 : enum migrate_mode mode, int reason, struct list_head *ret_folios,
1625 : struct migrate_pages_stats *stats)
1626 : {
1627 : int retry;
1628 0 : int large_retry = 1;
1629 0 : int thp_retry = 1;
1630 0 : int nr_failed = 0;
1631 0 : int nr_retry_pages = 0;
1632 0 : int nr_large_failed = 0;
1633 0 : int pass = 0;
1634 0 : bool is_large = false;
1635 0 : bool is_thp = false;
1636 0 : struct folio *folio, *folio2, *dst = NULL, *dst2;
1637 : int rc, rc_saved, nr_pages;
1638 0 : LIST_HEAD(split_folios);
1639 0 : LIST_HEAD(unmap_folios);
1640 0 : LIST_HEAD(dst_folios);
1641 0 : bool nosplit = (reason == MR_NUMA_MISPLACED);
1642 0 : bool no_split_folio_counting = false;
1643 : bool avoid_force_lock;
1644 :
1645 : retry:
1646 0 : rc_saved = 0;
1647 0 : avoid_force_lock = false;
1648 0 : retry = 1;
1649 0 : for (pass = 0;
1650 0 : pass < NR_MAX_MIGRATE_PAGES_RETRY && (retry || large_retry);
1651 0 : pass++) {
1652 0 : retry = 0;
1653 0 : large_retry = 0;
1654 0 : thp_retry = 0;
1655 0 : nr_retry_pages = 0;
1656 :
1657 0 : list_for_each_entry_safe(folio, folio2, from, lru) {
1658 : /*
1659 : * Large folio statistics is based on the source large
1660 : * folio. Capture required information that might get
1661 : * lost during migration.
1662 : */
1663 0 : is_large = folio_test_large(folio);
1664 0 : is_thp = is_large && folio_test_pmd_mappable(folio);
1665 0 : nr_pages = folio_nr_pages(folio);
1666 :
1667 0 : cond_resched();
1668 :
1669 : /*
1670 : * Large folio migration might be unsupported or
1671 : * the allocation might be failed so we should retry
1672 : * on the same folio with the large folio split
1673 : * to normal folios.
1674 : *
1675 : * Split folios are put in split_folios, and
1676 : * we will migrate them after the rest of the
1677 : * list is processed.
1678 : */
1679 : if (!thp_migration_supported() && is_thp) {
1680 : nr_large_failed++;
1681 : stats->nr_thp_failed++;
1682 : if (!try_split_folio(folio, &split_folios)) {
1683 : stats->nr_thp_split++;
1684 : continue;
1685 : }
1686 : stats->nr_failed_pages += nr_pages;
1687 : list_move_tail(&folio->lru, ret_folios);
1688 : continue;
1689 : }
1690 :
1691 0 : rc = migrate_folio_unmap(get_new_page, put_new_page, private,
1692 : folio, &dst, pass > 2, avoid_force_lock,
1693 : mode, reason, ret_folios);
1694 : /*
1695 : * The rules are:
1696 : * Success: folio will be freed
1697 : * Unmap: folio will be put on unmap_folios list,
1698 : * dst folio put on dst_folios list
1699 : * -EAGAIN: stay on the from list
1700 : * -EDEADLOCK: stay on the from list
1701 : * -ENOMEM: stay on the from list
1702 : * Other errno: put on ret_folios list
1703 : */
1704 0 : switch(rc) {
1705 : case -ENOMEM:
1706 : /*
1707 : * When memory is low, don't bother to try to migrate
1708 : * other folios, move unmapped folios, then exit.
1709 : */
1710 0 : if (is_large) {
1711 0 : nr_large_failed++;
1712 : stats->nr_thp_failed += is_thp;
1713 : /* Large folio NUMA faulting doesn't split to retry. */
1714 0 : if (!nosplit) {
1715 0 : int ret = try_split_folio(folio, &split_folios);
1716 :
1717 0 : if (!ret) {
1718 : stats->nr_thp_split += is_thp;
1719 : break;
1720 0 : } else if (reason == MR_LONGTERM_PIN &&
1721 0 : ret == -EAGAIN) {
1722 : /*
1723 : * Try again to split large folio to
1724 : * mitigate the failure of longterm pinning.
1725 : */
1726 0 : large_retry++;
1727 0 : thp_retry += is_thp;
1728 0 : nr_retry_pages += nr_pages;
1729 0 : break;
1730 : }
1731 : }
1732 0 : } else if (!no_split_folio_counting) {
1733 0 : nr_failed++;
1734 : }
1735 :
1736 0 : stats->nr_failed_pages += nr_pages + nr_retry_pages;
1737 : /*
1738 : * There might be some split folios of fail-to-migrate large
1739 : * folios left in split_folios list. Move them to ret_folios
1740 : * list so that they could be put back to the right list by
1741 : * the caller otherwise the folio refcnt will be leaked.
1742 : */
1743 0 : list_splice_init(&split_folios, ret_folios);
1744 : /* nr_failed isn't updated for not used */
1745 0 : nr_large_failed += large_retry;
1746 : stats->nr_thp_failed += thp_retry;
1747 0 : rc_saved = rc;
1748 0 : if (list_empty(&unmap_folios))
1749 : goto out;
1750 : else
1751 : goto move;
1752 : case -EDEADLOCK:
1753 : /*
1754 : * The folio cannot be locked for potential deadlock.
1755 : * Go move (and unlock) all locked folios. Then we can
1756 : * try again.
1757 : */
1758 : rc_saved = rc;
1759 : goto move;
1760 : case -EAGAIN:
1761 0 : if (is_large) {
1762 0 : large_retry++;
1763 0 : thp_retry += is_thp;
1764 0 : } else if (!no_split_folio_counting) {
1765 0 : retry++;
1766 : }
1767 0 : nr_retry_pages += nr_pages;
1768 0 : break;
1769 : case MIGRATEPAGE_SUCCESS:
1770 0 : stats->nr_succeeded += nr_pages;
1771 : stats->nr_thp_succeeded += is_thp;
1772 0 : break;
1773 : case MIGRATEPAGE_UNMAP:
1774 : /*
1775 : * We have locked some folios, don't force lock
1776 : * to avoid deadlock.
1777 : */
1778 0 : avoid_force_lock = true;
1779 0 : list_move_tail(&folio->lru, &unmap_folios);
1780 0 : list_add_tail(&dst->lru, &dst_folios);
1781 : break;
1782 : default:
1783 : /*
1784 : * Permanent failure (-EBUSY, etc.):
1785 : * unlike -EAGAIN case, the failed folio is
1786 : * removed from migration folio list and not
1787 : * retried in the next outer loop.
1788 : */
1789 0 : if (is_large) {
1790 0 : nr_large_failed++;
1791 : stats->nr_thp_failed += is_thp;
1792 0 : } else if (!no_split_folio_counting) {
1793 0 : nr_failed++;
1794 : }
1795 :
1796 0 : stats->nr_failed_pages += nr_pages;
1797 0 : break;
1798 : }
1799 : }
1800 : }
1801 0 : nr_failed += retry;
1802 0 : nr_large_failed += large_retry;
1803 0 : stats->nr_thp_failed += thp_retry;
1804 0 : stats->nr_failed_pages += nr_retry_pages;
1805 : move:
1806 : /* Flush TLBs for all unmapped folios */
1807 : try_to_unmap_flush();
1808 :
1809 0 : retry = 1;
1810 0 : for (pass = 0;
1811 0 : pass < NR_MAX_MIGRATE_PAGES_RETRY && (retry || large_retry);
1812 0 : pass++) {
1813 0 : retry = 0;
1814 0 : large_retry = 0;
1815 0 : thp_retry = 0;
1816 0 : nr_retry_pages = 0;
1817 :
1818 0 : dst = list_first_entry(&dst_folios, struct folio, lru);
1819 0 : dst2 = list_next_entry(dst, lru);
1820 0 : list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1821 0 : is_large = folio_test_large(folio);
1822 0 : is_thp = is_large && folio_test_pmd_mappable(folio);
1823 0 : nr_pages = folio_nr_pages(folio);
1824 :
1825 0 : cond_resched();
1826 :
1827 0 : rc = migrate_folio_move(put_new_page, private,
1828 : folio, dst, mode,
1829 : reason, ret_folios);
1830 : /*
1831 : * The rules are:
1832 : * Success: folio will be freed
1833 : * -EAGAIN: stay on the unmap_folios list
1834 : * Other errno: put on ret_folios list
1835 : */
1836 0 : switch(rc) {
1837 : case -EAGAIN:
1838 0 : if (is_large) {
1839 0 : large_retry++;
1840 0 : thp_retry += is_thp;
1841 0 : } else if (!no_split_folio_counting) {
1842 0 : retry++;
1843 : }
1844 0 : nr_retry_pages += nr_pages;
1845 0 : break;
1846 : case MIGRATEPAGE_SUCCESS:
1847 0 : stats->nr_succeeded += nr_pages;
1848 : stats->nr_thp_succeeded += is_thp;
1849 0 : break;
1850 : default:
1851 0 : if (is_large) {
1852 0 : nr_large_failed++;
1853 : stats->nr_thp_failed += is_thp;
1854 0 : } else if (!no_split_folio_counting) {
1855 0 : nr_failed++;
1856 : }
1857 :
1858 0 : stats->nr_failed_pages += nr_pages;
1859 0 : break;
1860 : }
1861 0 : dst = dst2;
1862 0 : dst2 = list_next_entry(dst, lru);
1863 : }
1864 : }
1865 0 : nr_failed += retry;
1866 0 : nr_large_failed += large_retry;
1867 0 : stats->nr_thp_failed += thp_retry;
1868 0 : stats->nr_failed_pages += nr_retry_pages;
1869 :
1870 0 : if (rc_saved)
1871 : rc = rc_saved;
1872 : else
1873 0 : rc = nr_failed + nr_large_failed;
1874 : out:
1875 : /* Cleanup remaining folios */
1876 0 : dst = list_first_entry(&dst_folios, struct folio, lru);
1877 0 : dst2 = list_next_entry(dst, lru);
1878 0 : list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1879 0 : int page_was_mapped = 0;
1880 0 : struct anon_vma *anon_vma = NULL;
1881 :
1882 0 : __migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
1883 0 : migrate_folio_undo_src(folio, page_was_mapped, anon_vma,
1884 : true, ret_folios);
1885 0 : list_del(&dst->lru);
1886 0 : migrate_folio_undo_dst(dst, true, put_new_page, private);
1887 0 : dst = dst2;
1888 0 : dst2 = list_next_entry(dst, lru);
1889 : }
1890 :
1891 : /*
1892 : * Try to migrate split folios of fail-to-migrate large folios, no
1893 : * nr_failed counting in this round, since all split folios of a
1894 : * large folio is counted as 1 failure in the first round.
1895 : */
1896 0 : if (rc >= 0 && !list_empty(&split_folios)) {
1897 : /*
1898 : * Move non-migrated folios (after NR_MAX_MIGRATE_PAGES_RETRY
1899 : * retries) to ret_folios to avoid migrating them again.
1900 : */
1901 0 : list_splice_init(from, ret_folios);
1902 : list_splice_init(&split_folios, from);
1903 : no_split_folio_counting = true;
1904 : goto retry;
1905 : }
1906 :
1907 : /*
1908 : * We have unlocked all locked folios, so we can force lock now, let's
1909 : * try again.
1910 : */
1911 0 : if (rc == -EDEADLOCK)
1912 : goto retry;
1913 :
1914 0 : return rc;
1915 : }
1916 :
1917 : /*
1918 : * migrate_pages - migrate the folios specified in a list, to the free folios
1919 : * supplied as the target for the page migration
1920 : *
1921 : * @from: The list of folios to be migrated.
1922 : * @get_new_page: The function used to allocate free folios to be used
1923 : * as the target of the folio migration.
1924 : * @put_new_page: The function used to free target folios if migration
1925 : * fails, or NULL if no special handling is necessary.
1926 : * @private: Private data to be passed on to get_new_page()
1927 : * @mode: The migration mode that specifies the constraints for
1928 : * folio migration, if any.
1929 : * @reason: The reason for folio migration.
1930 : * @ret_succeeded: Set to the number of folios migrated successfully if
1931 : * the caller passes a non-NULL pointer.
1932 : *
1933 : * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1934 : * are movable any more because the list has become empty or no retryable folios
1935 : * exist any more. It is caller's responsibility to call putback_movable_pages()
1936 : * only if ret != 0.
1937 : *
1938 : * Returns the number of {normal folio, large folio, hugetlb} that were not
1939 : * migrated, or an error code. The number of large folio splits will be
1940 : * considered as the number of non-migrated large folio, no matter how many
1941 : * split folios of the large folio are migrated successfully.
1942 : */
1943 0 : int migrate_pages(struct list_head *from, new_page_t get_new_page,
1944 : free_page_t put_new_page, unsigned long private,
1945 : enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1946 : {
1947 : int rc, rc_gather;
1948 : int nr_pages;
1949 : struct folio *folio, *folio2;
1950 0 : LIST_HEAD(folios);
1951 0 : LIST_HEAD(ret_folios);
1952 : struct migrate_pages_stats stats;
1953 :
1954 0 : trace_mm_migrate_pages_start(mode, reason);
1955 :
1956 0 : memset(&stats, 0, sizeof(stats));
1957 :
1958 0 : rc_gather = migrate_hugetlbs(from, get_new_page, put_new_page, private,
1959 : mode, reason, &stats, &ret_folios);
1960 : if (rc_gather < 0)
1961 : goto out;
1962 : again:
1963 0 : nr_pages = 0;
1964 0 : list_for_each_entry_safe(folio, folio2, from, lru) {
1965 : /* Retried hugetlb folios will be kept in list */
1966 0 : if (folio_test_hugetlb(folio)) {
1967 : list_move_tail(&folio->lru, &ret_folios);
1968 : continue;
1969 : }
1970 :
1971 0 : nr_pages += folio_nr_pages(folio);
1972 0 : if (nr_pages > NR_MAX_BATCHED_MIGRATION)
1973 : break;
1974 : }
1975 0 : if (nr_pages > NR_MAX_BATCHED_MIGRATION)
1976 0 : list_cut_before(&folios, from, &folio->lru);
1977 : else
1978 : list_splice_init(from, &folios);
1979 0 : rc = migrate_pages_batch(&folios, get_new_page, put_new_page, private,
1980 : mode, reason, &ret_folios, &stats);
1981 0 : list_splice_tail_init(&folios, &ret_folios);
1982 0 : if (rc < 0) {
1983 : rc_gather = rc;
1984 : goto out;
1985 : }
1986 0 : rc_gather += rc;
1987 0 : if (!list_empty(from))
1988 : goto again;
1989 : out:
1990 : /*
1991 : * Put the permanent failure folio back to migration list, they
1992 : * will be put back to the right list by the caller.
1993 : */
1994 0 : list_splice(&ret_folios, from);
1995 :
1996 : /*
1997 : * Return 0 in case all split folios of fail-to-migrate large folios
1998 : * are migrated successfully.
1999 : */
2000 0 : if (list_empty(from))
2001 0 : rc_gather = 0;
2002 :
2003 0 : count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
2004 0 : count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
2005 0 : count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
2006 0 : count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
2007 0 : count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
2008 0 : trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
2009 0 : stats.nr_thp_succeeded, stats.nr_thp_failed,
2010 0 : stats.nr_thp_split, mode, reason);
2011 :
2012 0 : if (ret_succeeded)
2013 0 : *ret_succeeded = stats.nr_succeeded;
2014 :
2015 0 : return rc_gather;
2016 : }
2017 :
2018 0 : struct page *alloc_migration_target(struct page *page, unsigned long private)
2019 : {
2020 0 : struct folio *folio = page_folio(page);
2021 : struct migration_target_control *mtc;
2022 : gfp_t gfp_mask;
2023 0 : unsigned int order = 0;
2024 0 : struct folio *hugetlb_folio = NULL;
2025 0 : struct folio *new_folio = NULL;
2026 : int nid;
2027 : int zidx;
2028 :
2029 0 : mtc = (struct migration_target_control *)private;
2030 0 : gfp_mask = mtc->gfp_mask;
2031 0 : nid = mtc->nid;
2032 0 : if (nid == NUMA_NO_NODE)
2033 0 : nid = folio_nid(folio);
2034 :
2035 0 : if (folio_test_hugetlb(folio)) {
2036 : struct hstate *h = folio_hstate(folio);
2037 :
2038 : gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2039 : hugetlb_folio = alloc_hugetlb_folio_nodemask(h, nid,
2040 : mtc->nmask, gfp_mask);
2041 : return &hugetlb_folio->page;
2042 : }
2043 :
2044 0 : if (folio_test_large(folio)) {
2045 : /*
2046 : * clear __GFP_RECLAIM to make the migration callback
2047 : * consistent with regular THP allocations.
2048 : */
2049 0 : gfp_mask &= ~__GFP_RECLAIM;
2050 0 : gfp_mask |= GFP_TRANSHUGE;
2051 : order = folio_order(folio);
2052 : }
2053 0 : zidx = zone_idx(folio_zone(folio));
2054 0 : if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2055 0 : gfp_mask |= __GFP_HIGHMEM;
2056 :
2057 0 : new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2058 :
2059 0 : return &new_folio->page;
2060 : }
2061 :
2062 : #ifdef CONFIG_NUMA
2063 :
2064 : static int store_status(int __user *status, int start, int value, int nr)
2065 : {
2066 : while (nr-- > 0) {
2067 : if (put_user(value, status + start))
2068 : return -EFAULT;
2069 : start++;
2070 : }
2071 :
2072 : return 0;
2073 : }
2074 :
2075 : static int do_move_pages_to_node(struct mm_struct *mm,
2076 : struct list_head *pagelist, int node)
2077 : {
2078 : int err;
2079 : struct migration_target_control mtc = {
2080 : .nid = node,
2081 : .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2082 : };
2083 :
2084 : err = migrate_pages(pagelist, alloc_migration_target, NULL,
2085 : (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2086 : if (err)
2087 : putback_movable_pages(pagelist);
2088 : return err;
2089 : }
2090 :
2091 : /*
2092 : * Resolves the given address to a struct page, isolates it from the LRU and
2093 : * puts it to the given pagelist.
2094 : * Returns:
2095 : * errno - if the page cannot be found/isolated
2096 : * 0 - when it doesn't have to be migrated because it is already on the
2097 : * target node
2098 : * 1 - when it has been queued
2099 : */
2100 : static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
2101 : int node, struct list_head *pagelist, bool migrate_all)
2102 : {
2103 : struct vm_area_struct *vma;
2104 : struct page *page;
2105 : int err;
2106 : bool isolated;
2107 :
2108 : mmap_read_lock(mm);
2109 : err = -EFAULT;
2110 : vma = vma_lookup(mm, addr);
2111 : if (!vma || !vma_migratable(vma))
2112 : goto out;
2113 :
2114 : /* FOLL_DUMP to ignore special (like zero) pages */
2115 : page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2116 :
2117 : err = PTR_ERR(page);
2118 : if (IS_ERR(page))
2119 : goto out;
2120 :
2121 : err = -ENOENT;
2122 : if (!page)
2123 : goto out;
2124 :
2125 : if (is_zone_device_page(page))
2126 : goto out_putpage;
2127 :
2128 : err = 0;
2129 : if (page_to_nid(page) == node)
2130 : goto out_putpage;
2131 :
2132 : err = -EACCES;
2133 : if (page_mapcount(page) > 1 && !migrate_all)
2134 : goto out_putpage;
2135 :
2136 : if (PageHuge(page)) {
2137 : if (PageHead(page)) {
2138 : isolated = isolate_hugetlb(page_folio(page), pagelist);
2139 : err = isolated ? 1 : -EBUSY;
2140 : }
2141 : } else {
2142 : struct page *head;
2143 :
2144 : head = compound_head(page);
2145 : isolated = isolate_lru_page(head);
2146 : if (!isolated) {
2147 : err = -EBUSY;
2148 : goto out_putpage;
2149 : }
2150 :
2151 : err = 1;
2152 : list_add_tail(&head->lru, pagelist);
2153 : mod_node_page_state(page_pgdat(head),
2154 : NR_ISOLATED_ANON + page_is_file_lru(head),
2155 : thp_nr_pages(head));
2156 : }
2157 : out_putpage:
2158 : /*
2159 : * Either remove the duplicate refcount from
2160 : * isolate_lru_page() or drop the page ref if it was
2161 : * not isolated.
2162 : */
2163 : put_page(page);
2164 : out:
2165 : mmap_read_unlock(mm);
2166 : return err;
2167 : }
2168 :
2169 : static int move_pages_and_store_status(struct mm_struct *mm, int node,
2170 : struct list_head *pagelist, int __user *status,
2171 : int start, int i, unsigned long nr_pages)
2172 : {
2173 : int err;
2174 :
2175 : if (list_empty(pagelist))
2176 : return 0;
2177 :
2178 : err = do_move_pages_to_node(mm, pagelist, node);
2179 : if (err) {
2180 : /*
2181 : * Positive err means the number of failed
2182 : * pages to migrate. Since we are going to
2183 : * abort and return the number of non-migrated
2184 : * pages, so need to include the rest of the
2185 : * nr_pages that have not been attempted as
2186 : * well.
2187 : */
2188 : if (err > 0)
2189 : err += nr_pages - i;
2190 : return err;
2191 : }
2192 : return store_status(status, start, node, i - start);
2193 : }
2194 :
2195 : /*
2196 : * Migrate an array of page address onto an array of nodes and fill
2197 : * the corresponding array of status.
2198 : */
2199 : static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2200 : unsigned long nr_pages,
2201 : const void __user * __user *pages,
2202 : const int __user *nodes,
2203 : int __user *status, int flags)
2204 : {
2205 : int current_node = NUMA_NO_NODE;
2206 : LIST_HEAD(pagelist);
2207 : int start, i;
2208 : int err = 0, err1;
2209 :
2210 : lru_cache_disable();
2211 :
2212 : for (i = start = 0; i < nr_pages; i++) {
2213 : const void __user *p;
2214 : unsigned long addr;
2215 : int node;
2216 :
2217 : err = -EFAULT;
2218 : if (get_user(p, pages + i))
2219 : goto out_flush;
2220 : if (get_user(node, nodes + i))
2221 : goto out_flush;
2222 : addr = (unsigned long)untagged_addr(p);
2223 :
2224 : err = -ENODEV;
2225 : if (node < 0 || node >= MAX_NUMNODES)
2226 : goto out_flush;
2227 : if (!node_state(node, N_MEMORY))
2228 : goto out_flush;
2229 :
2230 : err = -EACCES;
2231 : if (!node_isset(node, task_nodes))
2232 : goto out_flush;
2233 :
2234 : if (current_node == NUMA_NO_NODE) {
2235 : current_node = node;
2236 : start = i;
2237 : } else if (node != current_node) {
2238 : err = move_pages_and_store_status(mm, current_node,
2239 : &pagelist, status, start, i, nr_pages);
2240 : if (err)
2241 : goto out;
2242 : start = i;
2243 : current_node = node;
2244 : }
2245 :
2246 : /*
2247 : * Errors in the page lookup or isolation are not fatal and we simply
2248 : * report them via status
2249 : */
2250 : err = add_page_for_migration(mm, addr, current_node,
2251 : &pagelist, flags & MPOL_MF_MOVE_ALL);
2252 :
2253 : if (err > 0) {
2254 : /* The page is successfully queued for migration */
2255 : continue;
2256 : }
2257 :
2258 : /*
2259 : * The move_pages() man page does not have an -EEXIST choice, so
2260 : * use -EFAULT instead.
2261 : */
2262 : if (err == -EEXIST)
2263 : err = -EFAULT;
2264 :
2265 : /*
2266 : * If the page is already on the target node (!err), store the
2267 : * node, otherwise, store the err.
2268 : */
2269 : err = store_status(status, i, err ? : current_node, 1);
2270 : if (err)
2271 : goto out_flush;
2272 :
2273 : err = move_pages_and_store_status(mm, current_node, &pagelist,
2274 : status, start, i, nr_pages);
2275 : if (err) {
2276 : /* We have accounted for page i */
2277 : if (err > 0)
2278 : err--;
2279 : goto out;
2280 : }
2281 : current_node = NUMA_NO_NODE;
2282 : }
2283 : out_flush:
2284 : /* Make sure we do not overwrite the existing error */
2285 : err1 = move_pages_and_store_status(mm, current_node, &pagelist,
2286 : status, start, i, nr_pages);
2287 : if (err >= 0)
2288 : err = err1;
2289 : out:
2290 : lru_cache_enable();
2291 : return err;
2292 : }
2293 :
2294 : /*
2295 : * Determine the nodes of an array of pages and store it in an array of status.
2296 : */
2297 : static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2298 : const void __user **pages, int *status)
2299 : {
2300 : unsigned long i;
2301 :
2302 : mmap_read_lock(mm);
2303 :
2304 : for (i = 0; i < nr_pages; i++) {
2305 : unsigned long addr = (unsigned long)(*pages);
2306 : struct vm_area_struct *vma;
2307 : struct page *page;
2308 : int err = -EFAULT;
2309 :
2310 : vma = vma_lookup(mm, addr);
2311 : if (!vma)
2312 : goto set_status;
2313 :
2314 : /* FOLL_DUMP to ignore special (like zero) pages */
2315 : page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2316 :
2317 : err = PTR_ERR(page);
2318 : if (IS_ERR(page))
2319 : goto set_status;
2320 :
2321 : err = -ENOENT;
2322 : if (!page)
2323 : goto set_status;
2324 :
2325 : if (!is_zone_device_page(page))
2326 : err = page_to_nid(page);
2327 :
2328 : put_page(page);
2329 : set_status:
2330 : *status = err;
2331 :
2332 : pages++;
2333 : status++;
2334 : }
2335 :
2336 : mmap_read_unlock(mm);
2337 : }
2338 :
2339 : static int get_compat_pages_array(const void __user *chunk_pages[],
2340 : const void __user * __user *pages,
2341 : unsigned long chunk_nr)
2342 : {
2343 : compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2344 : compat_uptr_t p;
2345 : int i;
2346 :
2347 : for (i = 0; i < chunk_nr; i++) {
2348 : if (get_user(p, pages32 + i))
2349 : return -EFAULT;
2350 : chunk_pages[i] = compat_ptr(p);
2351 : }
2352 :
2353 : return 0;
2354 : }
2355 :
2356 : /*
2357 : * Determine the nodes of a user array of pages and store it in
2358 : * a user array of status.
2359 : */
2360 : static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2361 : const void __user * __user *pages,
2362 : int __user *status)
2363 : {
2364 : #define DO_PAGES_STAT_CHUNK_NR 16UL
2365 : const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2366 : int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2367 :
2368 : while (nr_pages) {
2369 : unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2370 :
2371 : if (in_compat_syscall()) {
2372 : if (get_compat_pages_array(chunk_pages, pages,
2373 : chunk_nr))
2374 : break;
2375 : } else {
2376 : if (copy_from_user(chunk_pages, pages,
2377 : chunk_nr * sizeof(*chunk_pages)))
2378 : break;
2379 : }
2380 :
2381 : do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2382 :
2383 : if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2384 : break;
2385 :
2386 : pages += chunk_nr;
2387 : status += chunk_nr;
2388 : nr_pages -= chunk_nr;
2389 : }
2390 : return nr_pages ? -EFAULT : 0;
2391 : }
2392 :
2393 : static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2394 : {
2395 : struct task_struct *task;
2396 : struct mm_struct *mm;
2397 :
2398 : /*
2399 : * There is no need to check if current process has the right to modify
2400 : * the specified process when they are same.
2401 : */
2402 : if (!pid) {
2403 : mmget(current->mm);
2404 : *mem_nodes = cpuset_mems_allowed(current);
2405 : return current->mm;
2406 : }
2407 :
2408 : /* Find the mm_struct */
2409 : rcu_read_lock();
2410 : task = find_task_by_vpid(pid);
2411 : if (!task) {
2412 : rcu_read_unlock();
2413 : return ERR_PTR(-ESRCH);
2414 : }
2415 : get_task_struct(task);
2416 :
2417 : /*
2418 : * Check if this process has the right to modify the specified
2419 : * process. Use the regular "ptrace_may_access()" checks.
2420 : */
2421 : if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2422 : rcu_read_unlock();
2423 : mm = ERR_PTR(-EPERM);
2424 : goto out;
2425 : }
2426 : rcu_read_unlock();
2427 :
2428 : mm = ERR_PTR(security_task_movememory(task));
2429 : if (IS_ERR(mm))
2430 : goto out;
2431 : *mem_nodes = cpuset_mems_allowed(task);
2432 : mm = get_task_mm(task);
2433 : out:
2434 : put_task_struct(task);
2435 : if (!mm)
2436 : mm = ERR_PTR(-EINVAL);
2437 : return mm;
2438 : }
2439 :
2440 : /*
2441 : * Move a list of pages in the address space of the currently executing
2442 : * process.
2443 : */
2444 : static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2445 : const void __user * __user *pages,
2446 : const int __user *nodes,
2447 : int __user *status, int flags)
2448 : {
2449 : struct mm_struct *mm;
2450 : int err;
2451 : nodemask_t task_nodes;
2452 :
2453 : /* Check flags */
2454 : if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2455 : return -EINVAL;
2456 :
2457 : if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2458 : return -EPERM;
2459 :
2460 : mm = find_mm_struct(pid, &task_nodes);
2461 : if (IS_ERR(mm))
2462 : return PTR_ERR(mm);
2463 :
2464 : if (nodes)
2465 : err = do_pages_move(mm, task_nodes, nr_pages, pages,
2466 : nodes, status, flags);
2467 : else
2468 : err = do_pages_stat(mm, nr_pages, pages, status);
2469 :
2470 : mmput(mm);
2471 : return err;
2472 : }
2473 :
2474 : SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2475 : const void __user * __user *, pages,
2476 : const int __user *, nodes,
2477 : int __user *, status, int, flags)
2478 : {
2479 : return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2480 : }
2481 :
2482 : #ifdef CONFIG_NUMA_BALANCING
2483 : /*
2484 : * Returns true if this is a safe migration target node for misplaced NUMA
2485 : * pages. Currently it only checks the watermarks which is crude.
2486 : */
2487 : static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2488 : unsigned long nr_migrate_pages)
2489 : {
2490 : int z;
2491 :
2492 : for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2493 : struct zone *zone = pgdat->node_zones + z;
2494 :
2495 : if (!managed_zone(zone))
2496 : continue;
2497 :
2498 : /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2499 : if (!zone_watermark_ok(zone, 0,
2500 : high_wmark_pages(zone) +
2501 : nr_migrate_pages,
2502 : ZONE_MOVABLE, 0))
2503 : continue;
2504 : return true;
2505 : }
2506 : return false;
2507 : }
2508 :
2509 : static struct page *alloc_misplaced_dst_page(struct page *page,
2510 : unsigned long data)
2511 : {
2512 : int nid = (int) data;
2513 : int order = compound_order(page);
2514 : gfp_t gfp = __GFP_THISNODE;
2515 : struct folio *new;
2516 :
2517 : if (order > 0)
2518 : gfp |= GFP_TRANSHUGE_LIGHT;
2519 : else {
2520 : gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2521 : __GFP_NOWARN;
2522 : gfp &= ~__GFP_RECLAIM;
2523 : }
2524 : new = __folio_alloc_node(gfp, order, nid);
2525 :
2526 : return &new->page;
2527 : }
2528 :
2529 : static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2530 : {
2531 : int nr_pages = thp_nr_pages(page);
2532 : int order = compound_order(page);
2533 :
2534 : VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2535 :
2536 : /* Do not migrate THP mapped by multiple processes */
2537 : if (PageTransHuge(page) && total_mapcount(page) > 1)
2538 : return 0;
2539 :
2540 : /* Avoid migrating to a node that is nearly full */
2541 : if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2542 : int z;
2543 :
2544 : if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2545 : return 0;
2546 : for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2547 : if (managed_zone(pgdat->node_zones + z))
2548 : break;
2549 : }
2550 : wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2551 : return 0;
2552 : }
2553 :
2554 : if (!isolate_lru_page(page))
2555 : return 0;
2556 :
2557 : mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2558 : nr_pages);
2559 :
2560 : /*
2561 : * Isolating the page has taken another reference, so the
2562 : * caller's reference can be safely dropped without the page
2563 : * disappearing underneath us during migration.
2564 : */
2565 : put_page(page);
2566 : return 1;
2567 : }
2568 :
2569 : /*
2570 : * Attempt to migrate a misplaced page to the specified destination
2571 : * node. Caller is expected to have an elevated reference count on
2572 : * the page that will be dropped by this function before returning.
2573 : */
2574 : int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2575 : int node)
2576 : {
2577 : pg_data_t *pgdat = NODE_DATA(node);
2578 : int isolated;
2579 : int nr_remaining;
2580 : unsigned int nr_succeeded;
2581 : LIST_HEAD(migratepages);
2582 : int nr_pages = thp_nr_pages(page);
2583 :
2584 : /*
2585 : * Don't migrate file pages that are mapped in multiple processes
2586 : * with execute permissions as they are probably shared libraries.
2587 : */
2588 : if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2589 : (vma->vm_flags & VM_EXEC))
2590 : goto out;
2591 :
2592 : /*
2593 : * Also do not migrate dirty pages as not all filesystems can move
2594 : * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2595 : */
2596 : if (page_is_file_lru(page) && PageDirty(page))
2597 : goto out;
2598 :
2599 : isolated = numamigrate_isolate_page(pgdat, page);
2600 : if (!isolated)
2601 : goto out;
2602 :
2603 : list_add(&page->lru, &migratepages);
2604 : nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2605 : NULL, node, MIGRATE_ASYNC,
2606 : MR_NUMA_MISPLACED, &nr_succeeded);
2607 : if (nr_remaining) {
2608 : if (!list_empty(&migratepages)) {
2609 : list_del(&page->lru);
2610 : mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2611 : page_is_file_lru(page), -nr_pages);
2612 : putback_lru_page(page);
2613 : }
2614 : isolated = 0;
2615 : }
2616 : if (nr_succeeded) {
2617 : count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2618 : if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2619 : mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2620 : nr_succeeded);
2621 : }
2622 : BUG_ON(!list_empty(&migratepages));
2623 : return isolated;
2624 :
2625 : out:
2626 : put_page(page);
2627 : return 0;
2628 : }
2629 : #endif /* CONFIG_NUMA_BALANCING */
2630 : #endif /* CONFIG_NUMA */
|
