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