xref: /linux/mm/migrate.c (revision 24bce201d79807b668bf9d9e0aca801c5c0d5f78)
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 int isolate_movable_page(struct page *page, isolate_mode_t mode)
61 {
62 	struct address_space *mapping;
63 
64 	/*
65 	 * Avoid burning cycles with pages that are yet under __free_pages(),
66 	 * or just got freed under us.
67 	 *
68 	 * In case we 'win' a race for a movable page being freed under us and
69 	 * raise its refcount preventing __free_pages() from doing its job
70 	 * the put_page() at the end of this block will take care of
71 	 * release this page, thus avoiding a nasty leakage.
72 	 */
73 	if (unlikely(!get_page_unless_zero(page)))
74 		goto out;
75 
76 	/*
77 	 * Check PageMovable before holding a PG_lock because page's owner
78 	 * assumes anybody doesn't touch PG_lock of newly allocated page
79 	 * so unconditionally grabbing the lock ruins page's owner side.
80 	 */
81 	if (unlikely(!__PageMovable(page)))
82 		goto out_putpage;
83 	/*
84 	 * As movable pages are not isolated from LRU lists, concurrent
85 	 * compaction threads can race against page migration functions
86 	 * as well as race against the releasing a page.
87 	 *
88 	 * In order to avoid having an already isolated movable page
89 	 * being (wrongly) re-isolated while it is under migration,
90 	 * or to avoid attempting to isolate pages being released,
91 	 * lets be sure we have the page lock
92 	 * before proceeding with the movable page isolation steps.
93 	 */
94 	if (unlikely(!trylock_page(page)))
95 		goto out_putpage;
96 
97 	if (!PageMovable(page) || PageIsolated(page))
98 		goto out_no_isolated;
99 
100 	mapping = page_mapping(page);
101 	VM_BUG_ON_PAGE(!mapping, page);
102 
103 	if (!mapping->a_ops->isolate_page(page, mode))
104 		goto out_no_isolated;
105 
106 	/* Driver shouldn't use PG_isolated bit of page->flags */
107 	WARN_ON_ONCE(PageIsolated(page));
108 	SetPageIsolated(page);
109 	unlock_page(page);
110 
111 	return 0;
112 
113 out_no_isolated:
114 	unlock_page(page);
115 out_putpage:
116 	put_page(page);
117 out:
118 	return -EBUSY;
119 }
120 
121 static void putback_movable_page(struct page *page)
122 {
123 	struct address_space *mapping;
124 
125 	mapping = page_mapping(page);
126 	mapping->a_ops->putback_page(page);
127 	ClearPageIsolated(page);
128 }
129 
130 /*
131  * Put previously isolated pages back onto the appropriate lists
132  * from where they were once taken off for compaction/migration.
133  *
134  * This function shall be used whenever the isolated pageset has been
135  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136  * and isolate_huge_page().
137  */
138 void putback_movable_pages(struct list_head *l)
139 {
140 	struct page *page;
141 	struct page *page2;
142 
143 	list_for_each_entry_safe(page, page2, l, lru) {
144 		if (unlikely(PageHuge(page))) {
145 			putback_active_hugepage(page);
146 			continue;
147 		}
148 		list_del(&page->lru);
149 		/*
150 		 * We isolated non-lru movable page so here we can use
151 		 * __PageMovable because LRU page's mapping cannot have
152 		 * PAGE_MAPPING_MOVABLE.
153 		 */
154 		if (unlikely(__PageMovable(page))) {
155 			VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 			lock_page(page);
157 			if (PageMovable(page))
158 				putback_movable_page(page);
159 			else
160 				ClearPageIsolated(page);
161 			unlock_page(page);
162 			put_page(page);
163 		} else {
164 			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
165 					page_is_file_lru(page), -thp_nr_pages(page));
166 			putback_lru_page(page);
167 		}
168 	}
169 }
170 
171 /*
172  * Restore a potential migration pte to a working pte entry
173  */
174 static bool remove_migration_pte(struct folio *folio,
175 		struct vm_area_struct *vma, unsigned long addr, void *old)
176 {
177 	DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
178 
179 	while (page_vma_mapped_walk(&pvmw)) {
180 		rmap_t rmap_flags = RMAP_NONE;
181 		pte_t pte;
182 		swp_entry_t entry;
183 		struct page *new;
184 		unsigned long idx = 0;
185 
186 		/* pgoff is invalid for ksm pages, but they are never large */
187 		if (folio_test_large(folio) && !folio_test_hugetlb(folio))
188 			idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
189 		new = folio_page(folio, idx);
190 
191 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
192 		/* PMD-mapped THP migration entry */
193 		if (!pvmw.pte) {
194 			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
195 					!folio_test_pmd_mappable(folio), folio);
196 			remove_migration_pmd(&pvmw, new);
197 			continue;
198 		}
199 #endif
200 
201 		folio_get(folio);
202 		pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
203 		if (pte_swp_soft_dirty(*pvmw.pte))
204 			pte = pte_mksoft_dirty(pte);
205 
206 		/*
207 		 * Recheck VMA as permissions can change since migration started
208 		 */
209 		entry = pte_to_swp_entry(*pvmw.pte);
210 		if (is_writable_migration_entry(entry))
211 			pte = maybe_mkwrite(pte, vma);
212 		else if (pte_swp_uffd_wp(*pvmw.pte))
213 			pte = pte_mkuffd_wp(pte);
214 
215 		if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
216 			rmap_flags |= RMAP_EXCLUSIVE;
217 
218 		if (unlikely(is_device_private_page(new))) {
219 			if (pte_write(pte))
220 				entry = make_writable_device_private_entry(
221 							page_to_pfn(new));
222 			else
223 				entry = make_readable_device_private_entry(
224 							page_to_pfn(new));
225 			pte = swp_entry_to_pte(entry);
226 			if (pte_swp_soft_dirty(*pvmw.pte))
227 				pte = pte_swp_mksoft_dirty(pte);
228 			if (pte_swp_uffd_wp(*pvmw.pte))
229 				pte = pte_swp_mkuffd_wp(pte);
230 		}
231 
232 #ifdef CONFIG_HUGETLB_PAGE
233 		if (folio_test_hugetlb(folio)) {
234 			unsigned int shift = huge_page_shift(hstate_vma(vma));
235 
236 			pte = pte_mkhuge(pte);
237 			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
238 			if (folio_test_anon(folio))
239 				hugepage_add_anon_rmap(new, vma, pvmw.address,
240 						       rmap_flags);
241 			else
242 				page_dup_file_rmap(new, true);
243 			set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
244 		} else
245 #endif
246 		{
247 			if (folio_test_anon(folio))
248 				page_add_anon_rmap(new, vma, pvmw.address,
249 						   rmap_flags);
250 			else
251 				page_add_file_rmap(new, vma, false);
252 			set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
253 		}
254 		if (vma->vm_flags & VM_LOCKED)
255 			mlock_page_drain_local();
256 
257 		trace_remove_migration_pte(pvmw.address, pte_val(pte),
258 					   compound_order(new));
259 
260 		/* No need to invalidate - it was non-present before */
261 		update_mmu_cache(vma, pvmw.address, pvmw.pte);
262 	}
263 
264 	return true;
265 }
266 
267 /*
268  * Get rid of all migration entries and replace them by
269  * references to the indicated page.
270  */
271 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
272 {
273 	struct rmap_walk_control rwc = {
274 		.rmap_one = remove_migration_pte,
275 		.arg = src,
276 	};
277 
278 	if (locked)
279 		rmap_walk_locked(dst, &rwc);
280 	else
281 		rmap_walk(dst, &rwc);
282 }
283 
284 /*
285  * Something used the pte of a page under migration. We need to
286  * get to the page and wait until migration is finished.
287  * When we return from this function the fault will be retried.
288  */
289 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
290 				spinlock_t *ptl)
291 {
292 	pte_t pte;
293 	swp_entry_t entry;
294 
295 	spin_lock(ptl);
296 	pte = *ptep;
297 	if (!is_swap_pte(pte))
298 		goto out;
299 
300 	entry = pte_to_swp_entry(pte);
301 	if (!is_migration_entry(entry))
302 		goto out;
303 
304 	migration_entry_wait_on_locked(entry, ptep, ptl);
305 	return;
306 out:
307 	pte_unmap_unlock(ptep, ptl);
308 }
309 
310 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
311 				unsigned long address)
312 {
313 	spinlock_t *ptl = pte_lockptr(mm, pmd);
314 	pte_t *ptep = pte_offset_map(pmd, address);
315 	__migration_entry_wait(mm, ptep, ptl);
316 }
317 
318 void migration_entry_wait_huge(struct vm_area_struct *vma,
319 		struct mm_struct *mm, pte_t *pte)
320 {
321 	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
322 	__migration_entry_wait(mm, pte, ptl);
323 }
324 
325 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
326 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
327 {
328 	spinlock_t *ptl;
329 
330 	ptl = pmd_lock(mm, pmd);
331 	if (!is_pmd_migration_entry(*pmd))
332 		goto unlock;
333 	migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
334 	return;
335 unlock:
336 	spin_unlock(ptl);
337 }
338 #endif
339 
340 static int expected_page_refs(struct address_space *mapping, struct page *page)
341 {
342 	int expected_count = 1;
343 
344 	if (mapping)
345 		expected_count += compound_nr(page) + page_has_private(page);
346 	return expected_count;
347 }
348 
349 /*
350  * Replace the page in the mapping.
351  *
352  * The number of remaining references must be:
353  * 1 for anonymous pages without a mapping
354  * 2 for pages with a mapping
355  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
356  */
357 int folio_migrate_mapping(struct address_space *mapping,
358 		struct folio *newfolio, struct folio *folio, int extra_count)
359 {
360 	XA_STATE(xas, &mapping->i_pages, folio_index(folio));
361 	struct zone *oldzone, *newzone;
362 	int dirty;
363 	int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
364 	long nr = folio_nr_pages(folio);
365 
366 	if (!mapping) {
367 		/* Anonymous page without mapping */
368 		if (folio_ref_count(folio) != expected_count)
369 			return -EAGAIN;
370 
371 		/* No turning back from here */
372 		newfolio->index = folio->index;
373 		newfolio->mapping = folio->mapping;
374 		if (folio_test_swapbacked(folio))
375 			__folio_set_swapbacked(newfolio);
376 
377 		return MIGRATEPAGE_SUCCESS;
378 	}
379 
380 	oldzone = folio_zone(folio);
381 	newzone = folio_zone(newfolio);
382 
383 	xas_lock_irq(&xas);
384 	if (!folio_ref_freeze(folio, expected_count)) {
385 		xas_unlock_irq(&xas);
386 		return -EAGAIN;
387 	}
388 
389 	/*
390 	 * Now we know that no one else is looking at the folio:
391 	 * no turning back from here.
392 	 */
393 	newfolio->index = folio->index;
394 	newfolio->mapping = folio->mapping;
395 	folio_ref_add(newfolio, nr); /* add cache reference */
396 	if (folio_test_swapbacked(folio)) {
397 		__folio_set_swapbacked(newfolio);
398 		if (folio_test_swapcache(folio)) {
399 			folio_set_swapcache(newfolio);
400 			newfolio->private = folio_get_private(folio);
401 		}
402 	} else {
403 		VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
404 	}
405 
406 	/* Move dirty while page refs frozen and newpage not yet exposed */
407 	dirty = folio_test_dirty(folio);
408 	if (dirty) {
409 		folio_clear_dirty(folio);
410 		folio_set_dirty(newfolio);
411 	}
412 
413 	xas_store(&xas, newfolio);
414 
415 	/*
416 	 * Drop cache reference from old page by unfreezing
417 	 * to one less reference.
418 	 * We know this isn't the last reference.
419 	 */
420 	folio_ref_unfreeze(folio, expected_count - nr);
421 
422 	xas_unlock(&xas);
423 	/* Leave irq disabled to prevent preemption while updating stats */
424 
425 	/*
426 	 * If moved to a different zone then also account
427 	 * the page for that zone. Other VM counters will be
428 	 * taken care of when we establish references to the
429 	 * new page and drop references to the old page.
430 	 *
431 	 * Note that anonymous pages are accounted for
432 	 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
433 	 * are mapped to swap space.
434 	 */
435 	if (newzone != oldzone) {
436 		struct lruvec *old_lruvec, *new_lruvec;
437 		struct mem_cgroup *memcg;
438 
439 		memcg = folio_memcg(folio);
440 		old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
441 		new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
442 
443 		__mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
444 		__mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
445 		if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
446 			__mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
447 			__mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
448 		}
449 #ifdef CONFIG_SWAP
450 		if (folio_test_swapcache(folio)) {
451 			__mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
452 			__mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
453 		}
454 #endif
455 		if (dirty && mapping_can_writeback(mapping)) {
456 			__mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
457 			__mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
458 			__mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
459 			__mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
460 		}
461 	}
462 	local_irq_enable();
463 
464 	return MIGRATEPAGE_SUCCESS;
465 }
466 EXPORT_SYMBOL(folio_migrate_mapping);
467 
468 /*
469  * The expected number of remaining references is the same as that
470  * of folio_migrate_mapping().
471  */
472 int migrate_huge_page_move_mapping(struct address_space *mapping,
473 				   struct page *newpage, struct page *page)
474 {
475 	XA_STATE(xas, &mapping->i_pages, page_index(page));
476 	int expected_count;
477 
478 	xas_lock_irq(&xas);
479 	expected_count = 2 + page_has_private(page);
480 	if (!page_ref_freeze(page, expected_count)) {
481 		xas_unlock_irq(&xas);
482 		return -EAGAIN;
483 	}
484 
485 	newpage->index = page->index;
486 	newpage->mapping = page->mapping;
487 
488 	get_page(newpage);
489 
490 	xas_store(&xas, newpage);
491 
492 	page_ref_unfreeze(page, expected_count - 1);
493 
494 	xas_unlock_irq(&xas);
495 
496 	return MIGRATEPAGE_SUCCESS;
497 }
498 
499 /*
500  * Copy the flags and some other ancillary information
501  */
502 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
503 {
504 	int cpupid;
505 
506 	if (folio_test_error(folio))
507 		folio_set_error(newfolio);
508 	if (folio_test_referenced(folio))
509 		folio_set_referenced(newfolio);
510 	if (folio_test_uptodate(folio))
511 		folio_mark_uptodate(newfolio);
512 	if (folio_test_clear_active(folio)) {
513 		VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
514 		folio_set_active(newfolio);
515 	} else if (folio_test_clear_unevictable(folio))
516 		folio_set_unevictable(newfolio);
517 	if (folio_test_workingset(folio))
518 		folio_set_workingset(newfolio);
519 	if (folio_test_checked(folio))
520 		folio_set_checked(newfolio);
521 	/*
522 	 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
523 	 * migration entries. We can still have PG_anon_exclusive set on an
524 	 * effectively unmapped and unreferenced first sub-pages of an
525 	 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
526 	 */
527 	if (folio_test_mappedtodisk(folio))
528 		folio_set_mappedtodisk(newfolio);
529 
530 	/* Move dirty on pages not done by folio_migrate_mapping() */
531 	if (folio_test_dirty(folio))
532 		folio_set_dirty(newfolio);
533 
534 	if (folio_test_young(folio))
535 		folio_set_young(newfolio);
536 	if (folio_test_idle(folio))
537 		folio_set_idle(newfolio);
538 
539 	/*
540 	 * Copy NUMA information to the new page, to prevent over-eager
541 	 * future migrations of this same page.
542 	 */
543 	cpupid = page_cpupid_xchg_last(&folio->page, -1);
544 	page_cpupid_xchg_last(&newfolio->page, cpupid);
545 
546 	folio_migrate_ksm(newfolio, folio);
547 	/*
548 	 * Please do not reorder this without considering how mm/ksm.c's
549 	 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
550 	 */
551 	if (folio_test_swapcache(folio))
552 		folio_clear_swapcache(folio);
553 	folio_clear_private(folio);
554 
555 	/* page->private contains hugetlb specific flags */
556 	if (!folio_test_hugetlb(folio))
557 		folio->private = NULL;
558 
559 	/*
560 	 * If any waiters have accumulated on the new page then
561 	 * wake them up.
562 	 */
563 	if (folio_test_writeback(newfolio))
564 		folio_end_writeback(newfolio);
565 
566 	/*
567 	 * PG_readahead shares the same bit with PG_reclaim.  The above
568 	 * end_page_writeback() may clear PG_readahead mistakenly, so set the
569 	 * bit after that.
570 	 */
571 	if (folio_test_readahead(folio))
572 		folio_set_readahead(newfolio);
573 
574 	folio_copy_owner(newfolio, folio);
575 
576 	if (!folio_test_hugetlb(folio))
577 		mem_cgroup_migrate(folio, newfolio);
578 }
579 EXPORT_SYMBOL(folio_migrate_flags);
580 
581 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
582 {
583 	folio_copy(newfolio, folio);
584 	folio_migrate_flags(newfolio, folio);
585 }
586 EXPORT_SYMBOL(folio_migrate_copy);
587 
588 /************************************************************
589  *                    Migration functions
590  ***********************************************************/
591 
592 /*
593  * Common logic to directly migrate a single LRU page suitable for
594  * pages that do not use PagePrivate/PagePrivate2.
595  *
596  * Pages are locked upon entry and exit.
597  */
598 int migrate_page(struct address_space *mapping,
599 		struct page *newpage, struct page *page,
600 		enum migrate_mode mode)
601 {
602 	struct folio *newfolio = page_folio(newpage);
603 	struct folio *folio = page_folio(page);
604 	int rc;
605 
606 	BUG_ON(folio_test_writeback(folio));	/* Writeback must be complete */
607 
608 	rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
609 
610 	if (rc != MIGRATEPAGE_SUCCESS)
611 		return rc;
612 
613 	if (mode != MIGRATE_SYNC_NO_COPY)
614 		folio_migrate_copy(newfolio, folio);
615 	else
616 		folio_migrate_flags(newfolio, folio);
617 	return MIGRATEPAGE_SUCCESS;
618 }
619 EXPORT_SYMBOL(migrate_page);
620 
621 #ifdef CONFIG_BLOCK
622 /* Returns true if all buffers are successfully locked */
623 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
624 							enum migrate_mode mode)
625 {
626 	struct buffer_head *bh = head;
627 
628 	/* Simple case, sync compaction */
629 	if (mode != MIGRATE_ASYNC) {
630 		do {
631 			lock_buffer(bh);
632 			bh = bh->b_this_page;
633 
634 		} while (bh != head);
635 
636 		return true;
637 	}
638 
639 	/* async case, we cannot block on lock_buffer so use trylock_buffer */
640 	do {
641 		if (!trylock_buffer(bh)) {
642 			/*
643 			 * We failed to lock the buffer and cannot stall in
644 			 * async migration. Release the taken locks
645 			 */
646 			struct buffer_head *failed_bh = bh;
647 			bh = head;
648 			while (bh != failed_bh) {
649 				unlock_buffer(bh);
650 				bh = bh->b_this_page;
651 			}
652 			return false;
653 		}
654 
655 		bh = bh->b_this_page;
656 	} while (bh != head);
657 	return true;
658 }
659 
660 static int __buffer_migrate_page(struct address_space *mapping,
661 		struct page *newpage, struct page *page, enum migrate_mode mode,
662 		bool check_refs)
663 {
664 	struct buffer_head *bh, *head;
665 	int rc;
666 	int expected_count;
667 
668 	if (!page_has_buffers(page))
669 		return migrate_page(mapping, newpage, page, mode);
670 
671 	/* Check whether page does not have extra refs before we do more work */
672 	expected_count = expected_page_refs(mapping, page);
673 	if (page_count(page) != expected_count)
674 		return -EAGAIN;
675 
676 	head = page_buffers(page);
677 	if (!buffer_migrate_lock_buffers(head, mode))
678 		return -EAGAIN;
679 
680 	if (check_refs) {
681 		bool busy;
682 		bool invalidated = false;
683 
684 recheck_buffers:
685 		busy = false;
686 		spin_lock(&mapping->private_lock);
687 		bh = head;
688 		do {
689 			if (atomic_read(&bh->b_count)) {
690 				busy = true;
691 				break;
692 			}
693 			bh = bh->b_this_page;
694 		} while (bh != head);
695 		if (busy) {
696 			if (invalidated) {
697 				rc = -EAGAIN;
698 				goto unlock_buffers;
699 			}
700 			spin_unlock(&mapping->private_lock);
701 			invalidate_bh_lrus();
702 			invalidated = true;
703 			goto recheck_buffers;
704 		}
705 	}
706 
707 	rc = migrate_page_move_mapping(mapping, newpage, page, 0);
708 	if (rc != MIGRATEPAGE_SUCCESS)
709 		goto unlock_buffers;
710 
711 	attach_page_private(newpage, detach_page_private(page));
712 
713 	bh = head;
714 	do {
715 		set_bh_page(bh, newpage, bh_offset(bh));
716 		bh = bh->b_this_page;
717 
718 	} while (bh != head);
719 
720 	if (mode != MIGRATE_SYNC_NO_COPY)
721 		migrate_page_copy(newpage, page);
722 	else
723 		migrate_page_states(newpage, page);
724 
725 	rc = MIGRATEPAGE_SUCCESS;
726 unlock_buffers:
727 	if (check_refs)
728 		spin_unlock(&mapping->private_lock);
729 	bh = head;
730 	do {
731 		unlock_buffer(bh);
732 		bh = bh->b_this_page;
733 
734 	} while (bh != head);
735 
736 	return rc;
737 }
738 
739 /*
740  * Migration function for pages with buffers. This function can only be used
741  * if the underlying filesystem guarantees that no other references to "page"
742  * exist. For example attached buffer heads are accessed only under page lock.
743  */
744 int buffer_migrate_page(struct address_space *mapping,
745 		struct page *newpage, struct page *page, enum migrate_mode mode)
746 {
747 	return __buffer_migrate_page(mapping, newpage, page, mode, false);
748 }
749 EXPORT_SYMBOL(buffer_migrate_page);
750 
751 /*
752  * Same as above except that this variant is more careful and checks that there
753  * are also no buffer head references. This function is the right one for
754  * mappings where buffer heads are directly looked up and referenced (such as
755  * block device mappings).
756  */
757 int buffer_migrate_page_norefs(struct address_space *mapping,
758 		struct page *newpage, struct page *page, enum migrate_mode mode)
759 {
760 	return __buffer_migrate_page(mapping, newpage, page, mode, true);
761 }
762 #endif
763 
764 /*
765  * Writeback a page to clean the dirty state
766  */
767 static int writeout(struct address_space *mapping, struct page *page)
768 {
769 	struct folio *folio = page_folio(page);
770 	struct writeback_control wbc = {
771 		.sync_mode = WB_SYNC_NONE,
772 		.nr_to_write = 1,
773 		.range_start = 0,
774 		.range_end = LLONG_MAX,
775 		.for_reclaim = 1
776 	};
777 	int rc;
778 
779 	if (!mapping->a_ops->writepage)
780 		/* No write method for the address space */
781 		return -EINVAL;
782 
783 	if (!clear_page_dirty_for_io(page))
784 		/* Someone else already triggered a write */
785 		return -EAGAIN;
786 
787 	/*
788 	 * A dirty page may imply that the underlying filesystem has
789 	 * the page on some queue. So the page must be clean for
790 	 * migration. Writeout may mean we loose the lock and the
791 	 * page state is no longer what we checked for earlier.
792 	 * At this point we know that the migration attempt cannot
793 	 * be successful.
794 	 */
795 	remove_migration_ptes(folio, folio, false);
796 
797 	rc = mapping->a_ops->writepage(page, &wbc);
798 
799 	if (rc != AOP_WRITEPAGE_ACTIVATE)
800 		/* unlocked. Relock */
801 		lock_page(page);
802 
803 	return (rc < 0) ? -EIO : -EAGAIN;
804 }
805 
806 /*
807  * Default handling if a filesystem does not provide a migration function.
808  */
809 static int fallback_migrate_page(struct address_space *mapping,
810 	struct page *newpage, struct page *page, enum migrate_mode mode)
811 {
812 	if (PageDirty(page)) {
813 		/* Only writeback pages in full synchronous migration */
814 		switch (mode) {
815 		case MIGRATE_SYNC:
816 		case MIGRATE_SYNC_NO_COPY:
817 			break;
818 		default:
819 			return -EBUSY;
820 		}
821 		return writeout(mapping, page);
822 	}
823 
824 	/*
825 	 * Buffers may be managed in a filesystem specific way.
826 	 * We must have no buffers or drop them.
827 	 */
828 	if (page_has_private(page) &&
829 	    !try_to_release_page(page, GFP_KERNEL))
830 		return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
831 
832 	return migrate_page(mapping, newpage, page, mode);
833 }
834 
835 /*
836  * Move a page to a newly allocated page
837  * The page is locked and all ptes have been successfully removed.
838  *
839  * The new page will have replaced the old page if this function
840  * is successful.
841  *
842  * Return value:
843  *   < 0 - error code
844  *  MIGRATEPAGE_SUCCESS - success
845  */
846 static int move_to_new_folio(struct folio *dst, struct folio *src,
847 				enum migrate_mode mode)
848 {
849 	struct address_space *mapping;
850 	int rc = -EAGAIN;
851 	bool is_lru = !__PageMovable(&src->page);
852 
853 	VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
854 	VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
855 
856 	mapping = folio_mapping(src);
857 
858 	if (likely(is_lru)) {
859 		if (!mapping)
860 			rc = migrate_page(mapping, &dst->page, &src->page, mode);
861 		else if (mapping->a_ops->migratepage)
862 			/*
863 			 * Most pages have a mapping and most filesystems
864 			 * provide a migratepage callback. Anonymous pages
865 			 * are part of swap space which also has its own
866 			 * migratepage callback. This is the most common path
867 			 * for page migration.
868 			 */
869 			rc = mapping->a_ops->migratepage(mapping, &dst->page,
870 							&src->page, mode);
871 		else
872 			rc = fallback_migrate_page(mapping, &dst->page,
873 							&src->page, mode);
874 	} else {
875 		/*
876 		 * In case of non-lru page, it could be released after
877 		 * isolation step. In that case, we shouldn't try migration.
878 		 */
879 		VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
880 		if (!folio_test_movable(src)) {
881 			rc = MIGRATEPAGE_SUCCESS;
882 			folio_clear_isolated(src);
883 			goto out;
884 		}
885 
886 		rc = mapping->a_ops->migratepage(mapping, &dst->page,
887 						&src->page, mode);
888 		WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
889 				!folio_test_isolated(src));
890 	}
891 
892 	/*
893 	 * When successful, old pagecache src->mapping must be cleared before
894 	 * src is freed; but stats require that PageAnon be left as PageAnon.
895 	 */
896 	if (rc == MIGRATEPAGE_SUCCESS) {
897 		if (__PageMovable(&src->page)) {
898 			VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
899 
900 			/*
901 			 * We clear PG_movable under page_lock so any compactor
902 			 * cannot try to migrate this page.
903 			 */
904 			folio_clear_isolated(src);
905 		}
906 
907 		/*
908 		 * Anonymous and movable src->mapping will be cleared by
909 		 * free_pages_prepare so don't reset it here for keeping
910 		 * the type to work PageAnon, for example.
911 		 */
912 		if (!folio_mapping_flags(src))
913 			src->mapping = NULL;
914 
915 		if (likely(!folio_is_zone_device(dst)))
916 			flush_dcache_folio(dst);
917 	}
918 out:
919 	return rc;
920 }
921 
922 static int __unmap_and_move(struct page *page, struct page *newpage,
923 				int force, enum migrate_mode mode)
924 {
925 	struct folio *folio = page_folio(page);
926 	struct folio *dst = page_folio(newpage);
927 	int rc = -EAGAIN;
928 	bool page_was_mapped = false;
929 	struct anon_vma *anon_vma = NULL;
930 	bool is_lru = !__PageMovable(page);
931 
932 	if (!trylock_page(page)) {
933 		if (!force || mode == MIGRATE_ASYNC)
934 			goto out;
935 
936 		/*
937 		 * It's not safe for direct compaction to call lock_page.
938 		 * For example, during page readahead pages are added locked
939 		 * to the LRU. Later, when the IO completes the pages are
940 		 * marked uptodate and unlocked. However, the queueing
941 		 * could be merging multiple pages for one bio (e.g.
942 		 * mpage_readahead). If an allocation happens for the
943 		 * second or third page, the process can end up locking
944 		 * the same page twice and deadlocking. Rather than
945 		 * trying to be clever about what pages can be locked,
946 		 * avoid the use of lock_page for direct compaction
947 		 * altogether.
948 		 */
949 		if (current->flags & PF_MEMALLOC)
950 			goto out;
951 
952 		lock_page(page);
953 	}
954 
955 	if (PageWriteback(page)) {
956 		/*
957 		 * Only in the case of a full synchronous migration is it
958 		 * necessary to wait for PageWriteback. In the async case,
959 		 * the retry loop is too short and in the sync-light case,
960 		 * the overhead of stalling is too much
961 		 */
962 		switch (mode) {
963 		case MIGRATE_SYNC:
964 		case MIGRATE_SYNC_NO_COPY:
965 			break;
966 		default:
967 			rc = -EBUSY;
968 			goto out_unlock;
969 		}
970 		if (!force)
971 			goto out_unlock;
972 		wait_on_page_writeback(page);
973 	}
974 
975 	/*
976 	 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
977 	 * we cannot notice that anon_vma is freed while we migrates a page.
978 	 * This get_anon_vma() delays freeing anon_vma pointer until the end
979 	 * of migration. File cache pages are no problem because of page_lock()
980 	 * File Caches may use write_page() or lock_page() in migration, then,
981 	 * just care Anon page here.
982 	 *
983 	 * Only page_get_anon_vma() understands the subtleties of
984 	 * getting a hold on an anon_vma from outside one of its mms.
985 	 * But if we cannot get anon_vma, then we won't need it anyway,
986 	 * because that implies that the anon page is no longer mapped
987 	 * (and cannot be remapped so long as we hold the page lock).
988 	 */
989 	if (PageAnon(page) && !PageKsm(page))
990 		anon_vma = page_get_anon_vma(page);
991 
992 	/*
993 	 * Block others from accessing the new page when we get around to
994 	 * establishing additional references. We are usually the only one
995 	 * holding a reference to newpage at this point. We used to have a BUG
996 	 * here if trylock_page(newpage) fails, but would like to allow for
997 	 * cases where there might be a race with the previous use of newpage.
998 	 * This is much like races on refcount of oldpage: just don't BUG().
999 	 */
1000 	if (unlikely(!trylock_page(newpage)))
1001 		goto out_unlock;
1002 
1003 	if (unlikely(!is_lru)) {
1004 		rc = move_to_new_folio(dst, folio, mode);
1005 		goto out_unlock_both;
1006 	}
1007 
1008 	/*
1009 	 * Corner case handling:
1010 	 * 1. When a new swap-cache page is read into, it is added to the LRU
1011 	 * and treated as swapcache but it has no rmap yet.
1012 	 * Calling try_to_unmap() against a page->mapping==NULL page will
1013 	 * trigger a BUG.  So handle it here.
1014 	 * 2. An orphaned page (see truncate_cleanup_page) might have
1015 	 * fs-private metadata. The page can be picked up due to memory
1016 	 * offlining.  Everywhere else except page reclaim, the page is
1017 	 * invisible to the vm, so the page can not be migrated.  So try to
1018 	 * free the metadata, so the page can be freed.
1019 	 */
1020 	if (!page->mapping) {
1021 		VM_BUG_ON_PAGE(PageAnon(page), page);
1022 		if (page_has_private(page)) {
1023 			try_to_free_buffers(folio);
1024 			goto out_unlock_both;
1025 		}
1026 	} else if (page_mapped(page)) {
1027 		/* Establish migration ptes */
1028 		VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1029 				page);
1030 		try_to_migrate(folio, 0);
1031 		page_was_mapped = true;
1032 	}
1033 
1034 	if (!page_mapped(page))
1035 		rc = move_to_new_folio(dst, folio, mode);
1036 
1037 	/*
1038 	 * When successful, push newpage to LRU immediately: so that if it
1039 	 * turns out to be an mlocked page, remove_migration_ptes() will
1040 	 * automatically build up the correct newpage->mlock_count for it.
1041 	 *
1042 	 * We would like to do something similar for the old page, when
1043 	 * unsuccessful, and other cases when a page has been temporarily
1044 	 * isolated from the unevictable LRU: but this case is the easiest.
1045 	 */
1046 	if (rc == MIGRATEPAGE_SUCCESS) {
1047 		lru_cache_add(newpage);
1048 		if (page_was_mapped)
1049 			lru_add_drain();
1050 	}
1051 
1052 	if (page_was_mapped)
1053 		remove_migration_ptes(folio,
1054 			rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1055 
1056 out_unlock_both:
1057 	unlock_page(newpage);
1058 out_unlock:
1059 	/* Drop an anon_vma reference if we took one */
1060 	if (anon_vma)
1061 		put_anon_vma(anon_vma);
1062 	unlock_page(page);
1063 out:
1064 	/*
1065 	 * If migration is successful, decrease refcount of the newpage,
1066 	 * which will not free the page because new page owner increased
1067 	 * refcounter.
1068 	 */
1069 	if (rc == MIGRATEPAGE_SUCCESS)
1070 		put_page(newpage);
1071 
1072 	return rc;
1073 }
1074 
1075 /*
1076  * Obtain the lock on page, remove all ptes and migrate the page
1077  * to the newly allocated page in newpage.
1078  */
1079 static int unmap_and_move(new_page_t get_new_page,
1080 				   free_page_t put_new_page,
1081 				   unsigned long private, struct page *page,
1082 				   int force, enum migrate_mode mode,
1083 				   enum migrate_reason reason,
1084 				   struct list_head *ret)
1085 {
1086 	int rc = MIGRATEPAGE_SUCCESS;
1087 	struct page *newpage = NULL;
1088 
1089 	if (!thp_migration_supported() && PageTransHuge(page))
1090 		return -ENOSYS;
1091 
1092 	if (page_count(page) == 1) {
1093 		/* page was freed from under us. So we are done. */
1094 		ClearPageActive(page);
1095 		ClearPageUnevictable(page);
1096 		if (unlikely(__PageMovable(page))) {
1097 			lock_page(page);
1098 			if (!PageMovable(page))
1099 				ClearPageIsolated(page);
1100 			unlock_page(page);
1101 		}
1102 		goto out;
1103 	}
1104 
1105 	newpage = get_new_page(page, private);
1106 	if (!newpage)
1107 		return -ENOMEM;
1108 
1109 	newpage->private = 0;
1110 	rc = __unmap_and_move(page, newpage, force, mode);
1111 	if (rc == MIGRATEPAGE_SUCCESS)
1112 		set_page_owner_migrate_reason(newpage, reason);
1113 
1114 out:
1115 	if (rc != -EAGAIN) {
1116 		/*
1117 		 * A page that has been migrated has all references
1118 		 * removed and will be freed. A page that has not been
1119 		 * migrated will have kept its references and be restored.
1120 		 */
1121 		list_del(&page->lru);
1122 	}
1123 
1124 	/*
1125 	 * If migration is successful, releases reference grabbed during
1126 	 * isolation. Otherwise, restore the page to right list unless
1127 	 * we want to retry.
1128 	 */
1129 	if (rc == MIGRATEPAGE_SUCCESS) {
1130 		/*
1131 		 * Compaction can migrate also non-LRU pages which are
1132 		 * not accounted to NR_ISOLATED_*. They can be recognized
1133 		 * as __PageMovable
1134 		 */
1135 		if (likely(!__PageMovable(page)))
1136 			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1137 					page_is_file_lru(page), -thp_nr_pages(page));
1138 
1139 		if (reason != MR_MEMORY_FAILURE)
1140 			/*
1141 			 * We release the page in page_handle_poison.
1142 			 */
1143 			put_page(page);
1144 	} else {
1145 		if (rc != -EAGAIN)
1146 			list_add_tail(&page->lru, ret);
1147 
1148 		if (put_new_page)
1149 			put_new_page(newpage, private);
1150 		else
1151 			put_page(newpage);
1152 	}
1153 
1154 	return rc;
1155 }
1156 
1157 /*
1158  * Counterpart of unmap_and_move_page() for hugepage migration.
1159  *
1160  * This function doesn't wait the completion of hugepage I/O
1161  * because there is no race between I/O and migration for hugepage.
1162  * Note that currently hugepage I/O occurs only in direct I/O
1163  * where no lock is held and PG_writeback is irrelevant,
1164  * and writeback status of all subpages are counted in the reference
1165  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1166  * under direct I/O, the reference of the head page is 512 and a bit more.)
1167  * This means that when we try to migrate hugepage whose subpages are
1168  * doing direct I/O, some references remain after try_to_unmap() and
1169  * hugepage migration fails without data corruption.
1170  *
1171  * There is also no race when direct I/O is issued on the page under migration,
1172  * because then pte is replaced with migration swap entry and direct I/O code
1173  * will wait in the page fault for migration to complete.
1174  */
1175 static int unmap_and_move_huge_page(new_page_t get_new_page,
1176 				free_page_t put_new_page, unsigned long private,
1177 				struct page *hpage, int force,
1178 				enum migrate_mode mode, int reason,
1179 				struct list_head *ret)
1180 {
1181 	struct folio *dst, *src = page_folio(hpage);
1182 	int rc = -EAGAIN;
1183 	int page_was_mapped = 0;
1184 	struct page *new_hpage;
1185 	struct anon_vma *anon_vma = NULL;
1186 	struct address_space *mapping = NULL;
1187 
1188 	/*
1189 	 * Migratability of hugepages depends on architectures and their size.
1190 	 * This check is necessary because some callers of hugepage migration
1191 	 * like soft offline and memory hotremove don't walk through page
1192 	 * tables or check whether the hugepage is pmd-based or not before
1193 	 * kicking migration.
1194 	 */
1195 	if (!hugepage_migration_supported(page_hstate(hpage))) {
1196 		list_move_tail(&hpage->lru, ret);
1197 		return -ENOSYS;
1198 	}
1199 
1200 	if (page_count(hpage) == 1) {
1201 		/* page was freed from under us. So we are done. */
1202 		putback_active_hugepage(hpage);
1203 		return MIGRATEPAGE_SUCCESS;
1204 	}
1205 
1206 	new_hpage = get_new_page(hpage, private);
1207 	if (!new_hpage)
1208 		return -ENOMEM;
1209 	dst = page_folio(new_hpage);
1210 
1211 	if (!trylock_page(hpage)) {
1212 		if (!force)
1213 			goto out;
1214 		switch (mode) {
1215 		case MIGRATE_SYNC:
1216 		case MIGRATE_SYNC_NO_COPY:
1217 			break;
1218 		default:
1219 			goto out;
1220 		}
1221 		lock_page(hpage);
1222 	}
1223 
1224 	/*
1225 	 * Check for pages which are in the process of being freed.  Without
1226 	 * page_mapping() set, hugetlbfs specific move page routine will not
1227 	 * be called and we could leak usage counts for subpools.
1228 	 */
1229 	if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1230 		rc = -EBUSY;
1231 		goto out_unlock;
1232 	}
1233 
1234 	if (PageAnon(hpage))
1235 		anon_vma = page_get_anon_vma(hpage);
1236 
1237 	if (unlikely(!trylock_page(new_hpage)))
1238 		goto put_anon;
1239 
1240 	if (page_mapped(hpage)) {
1241 		enum ttu_flags ttu = 0;
1242 
1243 		if (!PageAnon(hpage)) {
1244 			/*
1245 			 * In shared mappings, try_to_unmap could potentially
1246 			 * call huge_pmd_unshare.  Because of this, take
1247 			 * semaphore in write mode here and set TTU_RMAP_LOCKED
1248 			 * to let lower levels know we have taken the lock.
1249 			 */
1250 			mapping = hugetlb_page_mapping_lock_write(hpage);
1251 			if (unlikely(!mapping))
1252 				goto unlock_put_anon;
1253 
1254 			ttu = TTU_RMAP_LOCKED;
1255 		}
1256 
1257 		try_to_migrate(src, ttu);
1258 		page_was_mapped = 1;
1259 
1260 		if (ttu & TTU_RMAP_LOCKED)
1261 			i_mmap_unlock_write(mapping);
1262 	}
1263 
1264 	if (!page_mapped(hpage))
1265 		rc = move_to_new_folio(dst, src, mode);
1266 
1267 	if (page_was_mapped)
1268 		remove_migration_ptes(src,
1269 			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1270 
1271 unlock_put_anon:
1272 	unlock_page(new_hpage);
1273 
1274 put_anon:
1275 	if (anon_vma)
1276 		put_anon_vma(anon_vma);
1277 
1278 	if (rc == MIGRATEPAGE_SUCCESS) {
1279 		move_hugetlb_state(hpage, new_hpage, reason);
1280 		put_new_page = NULL;
1281 	}
1282 
1283 out_unlock:
1284 	unlock_page(hpage);
1285 out:
1286 	if (rc == MIGRATEPAGE_SUCCESS)
1287 		putback_active_hugepage(hpage);
1288 	else if (rc != -EAGAIN)
1289 		list_move_tail(&hpage->lru, ret);
1290 
1291 	/*
1292 	 * If migration was not successful and there's a freeing callback, use
1293 	 * it.  Otherwise, put_page() will drop the reference grabbed during
1294 	 * isolation.
1295 	 */
1296 	if (put_new_page)
1297 		put_new_page(new_hpage, private);
1298 	else
1299 		putback_active_hugepage(new_hpage);
1300 
1301 	return rc;
1302 }
1303 
1304 static inline int try_split_thp(struct page *page, struct page **page2,
1305 				struct list_head *from)
1306 {
1307 	int rc = 0;
1308 
1309 	lock_page(page);
1310 	rc = split_huge_page_to_list(page, from);
1311 	unlock_page(page);
1312 	if (!rc)
1313 		list_safe_reset_next(page, *page2, lru);
1314 
1315 	return rc;
1316 }
1317 
1318 /*
1319  * migrate_pages - migrate the pages specified in a list, to the free pages
1320  *		   supplied as the target for the page migration
1321  *
1322  * @from:		The list of pages to be migrated.
1323  * @get_new_page:	The function used to allocate free pages to be used
1324  *			as the target of the page migration.
1325  * @put_new_page:	The function used to free target pages if migration
1326  *			fails, or NULL if no special handling is necessary.
1327  * @private:		Private data to be passed on to get_new_page()
1328  * @mode:		The migration mode that specifies the constraints for
1329  *			page migration, if any.
1330  * @reason:		The reason for page migration.
1331  * @ret_succeeded:	Set to the number of normal pages migrated successfully if
1332  *			the caller passes a non-NULL pointer.
1333  *
1334  * The function returns after 10 attempts or if no pages are movable any more
1335  * because the list has become empty or no retryable pages exist any more.
1336  * It is caller's responsibility to call putback_movable_pages() to return pages
1337  * to the LRU or free list only if ret != 0.
1338  *
1339  * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1340  * an error code. The number of THP splits will be considered as the number of
1341  * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1342  */
1343 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1344 		free_page_t put_new_page, unsigned long private,
1345 		enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1346 {
1347 	int retry = 1;
1348 	int thp_retry = 1;
1349 	int nr_failed = 0;
1350 	int nr_failed_pages = 0;
1351 	int nr_succeeded = 0;
1352 	int nr_thp_succeeded = 0;
1353 	int nr_thp_failed = 0;
1354 	int nr_thp_split = 0;
1355 	int pass = 0;
1356 	bool is_thp = false;
1357 	struct page *page;
1358 	struct page *page2;
1359 	int rc, nr_subpages;
1360 	LIST_HEAD(ret_pages);
1361 	LIST_HEAD(thp_split_pages);
1362 	bool nosplit = (reason == MR_NUMA_MISPLACED);
1363 	bool no_subpage_counting = false;
1364 
1365 	trace_mm_migrate_pages_start(mode, reason);
1366 
1367 thp_subpage_migration:
1368 	for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1369 		retry = 0;
1370 		thp_retry = 0;
1371 
1372 		list_for_each_entry_safe(page, page2, from, lru) {
1373 retry:
1374 			/*
1375 			 * THP statistics is based on the source huge page.
1376 			 * Capture required information that might get lost
1377 			 * during migration.
1378 			 */
1379 			is_thp = PageTransHuge(page) && !PageHuge(page);
1380 			nr_subpages = compound_nr(page);
1381 			cond_resched();
1382 
1383 			if (PageHuge(page))
1384 				rc = unmap_and_move_huge_page(get_new_page,
1385 						put_new_page, private, page,
1386 						pass > 2, mode, reason,
1387 						&ret_pages);
1388 			else
1389 				rc = unmap_and_move(get_new_page, put_new_page,
1390 						private, page, pass > 2, mode,
1391 						reason, &ret_pages);
1392 			/*
1393 			 * The rules are:
1394 			 *	Success: non hugetlb page will be freed, hugetlb
1395 			 *		 page will be put back
1396 			 *	-EAGAIN: stay on the from list
1397 			 *	-ENOMEM: stay on the from list
1398 			 *	Other errno: put on ret_pages list then splice to
1399 			 *		     from list
1400 			 */
1401 			switch(rc) {
1402 			/*
1403 			 * THP migration might be unsupported or the
1404 			 * allocation could've failed so we should
1405 			 * retry on the same page with the THP split
1406 			 * to base pages.
1407 			 *
1408 			 * Head page is retried immediately and tail
1409 			 * pages are added to the tail of the list so
1410 			 * we encounter them after the rest of the list
1411 			 * is processed.
1412 			 */
1413 			case -ENOSYS:
1414 				/* THP migration is unsupported */
1415 				if (is_thp) {
1416 					nr_thp_failed++;
1417 					if (!try_split_thp(page, &page2, &thp_split_pages)) {
1418 						nr_thp_split++;
1419 						goto retry;
1420 					}
1421 				/* Hugetlb migration is unsupported */
1422 				} else if (!no_subpage_counting) {
1423 					nr_failed++;
1424 				}
1425 
1426 				nr_failed_pages += nr_subpages;
1427 				break;
1428 			case -ENOMEM:
1429 				/*
1430 				 * When memory is low, don't bother to try to migrate
1431 				 * other pages, just exit.
1432 				 * THP NUMA faulting doesn't split THP to retry.
1433 				 */
1434 				if (is_thp && !nosplit) {
1435 					nr_thp_failed++;
1436 					if (!try_split_thp(page, &page2, &thp_split_pages)) {
1437 						nr_thp_split++;
1438 						goto retry;
1439 					}
1440 				} else if (!no_subpage_counting) {
1441 					nr_failed++;
1442 				}
1443 
1444 				nr_failed_pages += nr_subpages;
1445 				/*
1446 				 * There might be some subpages of fail-to-migrate THPs
1447 				 * left in thp_split_pages list. Move them back to migration
1448 				 * list so that they could be put back to the right list by
1449 				 * the caller otherwise the page refcnt will be leaked.
1450 				 */
1451 				list_splice_init(&thp_split_pages, from);
1452 				nr_thp_failed += thp_retry;
1453 				goto out;
1454 			case -EAGAIN:
1455 				if (is_thp)
1456 					thp_retry++;
1457 				else
1458 					retry++;
1459 				break;
1460 			case MIGRATEPAGE_SUCCESS:
1461 				nr_succeeded += nr_subpages;
1462 				if (is_thp)
1463 					nr_thp_succeeded++;
1464 				break;
1465 			default:
1466 				/*
1467 				 * Permanent failure (-EBUSY, etc.):
1468 				 * unlike -EAGAIN case, the failed page is
1469 				 * removed from migration page list and not
1470 				 * retried in the next outer loop.
1471 				 */
1472 				if (is_thp)
1473 					nr_thp_failed++;
1474 				else if (!no_subpage_counting)
1475 					nr_failed++;
1476 
1477 				nr_failed_pages += nr_subpages;
1478 				break;
1479 			}
1480 		}
1481 	}
1482 	nr_failed += retry;
1483 	nr_thp_failed += thp_retry;
1484 	/*
1485 	 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1486 	 * counting in this round, since all subpages of a THP is counted
1487 	 * as 1 failure in the first round.
1488 	 */
1489 	if (!list_empty(&thp_split_pages)) {
1490 		/*
1491 		 * Move non-migrated pages (after 10 retries) to ret_pages
1492 		 * to avoid migrating them again.
1493 		 */
1494 		list_splice_init(from, &ret_pages);
1495 		list_splice_init(&thp_split_pages, from);
1496 		no_subpage_counting = true;
1497 		retry = 1;
1498 		goto thp_subpage_migration;
1499 	}
1500 
1501 	rc = nr_failed + nr_thp_failed;
1502 out:
1503 	/*
1504 	 * Put the permanent failure page back to migration list, they
1505 	 * will be put back to the right list by the caller.
1506 	 */
1507 	list_splice(&ret_pages, from);
1508 
1509 	count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1510 	count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1511 	count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1512 	count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1513 	count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1514 	trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1515 			       nr_thp_failed, nr_thp_split, mode, reason);
1516 
1517 	if (ret_succeeded)
1518 		*ret_succeeded = nr_succeeded;
1519 
1520 	return rc;
1521 }
1522 
1523 struct page *alloc_migration_target(struct page *page, unsigned long private)
1524 {
1525 	struct folio *folio = page_folio(page);
1526 	struct migration_target_control *mtc;
1527 	gfp_t gfp_mask;
1528 	unsigned int order = 0;
1529 	struct folio *new_folio = NULL;
1530 	int nid;
1531 	int zidx;
1532 
1533 	mtc = (struct migration_target_control *)private;
1534 	gfp_mask = mtc->gfp_mask;
1535 	nid = mtc->nid;
1536 	if (nid == NUMA_NO_NODE)
1537 		nid = folio_nid(folio);
1538 
1539 	if (folio_test_hugetlb(folio)) {
1540 		struct hstate *h = page_hstate(&folio->page);
1541 
1542 		gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1543 		return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1544 	}
1545 
1546 	if (folio_test_large(folio)) {
1547 		/*
1548 		 * clear __GFP_RECLAIM to make the migration callback
1549 		 * consistent with regular THP allocations.
1550 		 */
1551 		gfp_mask &= ~__GFP_RECLAIM;
1552 		gfp_mask |= GFP_TRANSHUGE;
1553 		order = folio_order(folio);
1554 	}
1555 	zidx = zone_idx(folio_zone(folio));
1556 	if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1557 		gfp_mask |= __GFP_HIGHMEM;
1558 
1559 	new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1560 
1561 	return &new_folio->page;
1562 }
1563 
1564 #ifdef CONFIG_NUMA
1565 
1566 static int store_status(int __user *status, int start, int value, int nr)
1567 {
1568 	while (nr-- > 0) {
1569 		if (put_user(value, status + start))
1570 			return -EFAULT;
1571 		start++;
1572 	}
1573 
1574 	return 0;
1575 }
1576 
1577 static int do_move_pages_to_node(struct mm_struct *mm,
1578 		struct list_head *pagelist, int node)
1579 {
1580 	int err;
1581 	struct migration_target_control mtc = {
1582 		.nid = node,
1583 		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1584 	};
1585 
1586 	err = migrate_pages(pagelist, alloc_migration_target, NULL,
1587 		(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1588 	if (err)
1589 		putback_movable_pages(pagelist);
1590 	return err;
1591 }
1592 
1593 /*
1594  * Resolves the given address to a struct page, isolates it from the LRU and
1595  * puts it to the given pagelist.
1596  * Returns:
1597  *     errno - if the page cannot be found/isolated
1598  *     0 - when it doesn't have to be migrated because it is already on the
1599  *         target node
1600  *     1 - when it has been queued
1601  */
1602 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1603 		int node, struct list_head *pagelist, bool migrate_all)
1604 {
1605 	struct vm_area_struct *vma;
1606 	struct page *page;
1607 	int err;
1608 
1609 	mmap_read_lock(mm);
1610 	err = -EFAULT;
1611 	vma = vma_lookup(mm, addr);
1612 	if (!vma || !vma_migratable(vma))
1613 		goto out;
1614 
1615 	/* FOLL_DUMP to ignore special (like zero) pages */
1616 	page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1617 
1618 	err = PTR_ERR(page);
1619 	if (IS_ERR(page))
1620 		goto out;
1621 
1622 	err = -ENOENT;
1623 	if (!page)
1624 		goto out;
1625 
1626 	err = 0;
1627 	if (page_to_nid(page) == node)
1628 		goto out_putpage;
1629 
1630 	err = -EACCES;
1631 	if (page_mapcount(page) > 1 && !migrate_all)
1632 		goto out_putpage;
1633 
1634 	if (PageHuge(page)) {
1635 		if (PageHead(page)) {
1636 			isolate_huge_page(page, pagelist);
1637 			err = 1;
1638 		}
1639 	} else {
1640 		struct page *head;
1641 
1642 		head = compound_head(page);
1643 		err = isolate_lru_page(head);
1644 		if (err)
1645 			goto out_putpage;
1646 
1647 		err = 1;
1648 		list_add_tail(&head->lru, pagelist);
1649 		mod_node_page_state(page_pgdat(head),
1650 			NR_ISOLATED_ANON + page_is_file_lru(head),
1651 			thp_nr_pages(head));
1652 	}
1653 out_putpage:
1654 	/*
1655 	 * Either remove the duplicate refcount from
1656 	 * isolate_lru_page() or drop the page ref if it was
1657 	 * not isolated.
1658 	 */
1659 	put_page(page);
1660 out:
1661 	mmap_read_unlock(mm);
1662 	return err;
1663 }
1664 
1665 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1666 		struct list_head *pagelist, int __user *status,
1667 		int start, int i, unsigned long nr_pages)
1668 {
1669 	int err;
1670 
1671 	if (list_empty(pagelist))
1672 		return 0;
1673 
1674 	err = do_move_pages_to_node(mm, pagelist, node);
1675 	if (err) {
1676 		/*
1677 		 * Positive err means the number of failed
1678 		 * pages to migrate.  Since we are going to
1679 		 * abort and return the number of non-migrated
1680 		 * pages, so need to include the rest of the
1681 		 * nr_pages that have not been attempted as
1682 		 * well.
1683 		 */
1684 		if (err > 0)
1685 			err += nr_pages - i - 1;
1686 		return err;
1687 	}
1688 	return store_status(status, start, node, i - start);
1689 }
1690 
1691 /*
1692  * Migrate an array of page address onto an array of nodes and fill
1693  * the corresponding array of status.
1694  */
1695 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1696 			 unsigned long nr_pages,
1697 			 const void __user * __user *pages,
1698 			 const int __user *nodes,
1699 			 int __user *status, int flags)
1700 {
1701 	int current_node = NUMA_NO_NODE;
1702 	LIST_HEAD(pagelist);
1703 	int start, i;
1704 	int err = 0, err1;
1705 
1706 	lru_cache_disable();
1707 
1708 	for (i = start = 0; i < nr_pages; i++) {
1709 		const void __user *p;
1710 		unsigned long addr;
1711 		int node;
1712 
1713 		err = -EFAULT;
1714 		if (get_user(p, pages + i))
1715 			goto out_flush;
1716 		if (get_user(node, nodes + i))
1717 			goto out_flush;
1718 		addr = (unsigned long)untagged_addr(p);
1719 
1720 		err = -ENODEV;
1721 		if (node < 0 || node >= MAX_NUMNODES)
1722 			goto out_flush;
1723 		if (!node_state(node, N_MEMORY))
1724 			goto out_flush;
1725 
1726 		err = -EACCES;
1727 		if (!node_isset(node, task_nodes))
1728 			goto out_flush;
1729 
1730 		if (current_node == NUMA_NO_NODE) {
1731 			current_node = node;
1732 			start = i;
1733 		} else if (node != current_node) {
1734 			err = move_pages_and_store_status(mm, current_node,
1735 					&pagelist, status, start, i, nr_pages);
1736 			if (err)
1737 				goto out;
1738 			start = i;
1739 			current_node = node;
1740 		}
1741 
1742 		/*
1743 		 * Errors in the page lookup or isolation are not fatal and we simply
1744 		 * report them via status
1745 		 */
1746 		err = add_page_for_migration(mm, addr, current_node,
1747 				&pagelist, flags & MPOL_MF_MOVE_ALL);
1748 
1749 		if (err > 0) {
1750 			/* The page is successfully queued for migration */
1751 			continue;
1752 		}
1753 
1754 		/*
1755 		 * The move_pages() man page does not have an -EEXIST choice, so
1756 		 * use -EFAULT instead.
1757 		 */
1758 		if (err == -EEXIST)
1759 			err = -EFAULT;
1760 
1761 		/*
1762 		 * If the page is already on the target node (!err), store the
1763 		 * node, otherwise, store the err.
1764 		 */
1765 		err = store_status(status, i, err ? : current_node, 1);
1766 		if (err)
1767 			goto out_flush;
1768 
1769 		err = move_pages_and_store_status(mm, current_node, &pagelist,
1770 				status, start, i, nr_pages);
1771 		if (err)
1772 			goto out;
1773 		current_node = NUMA_NO_NODE;
1774 	}
1775 out_flush:
1776 	/* Make sure we do not overwrite the existing error */
1777 	err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1778 				status, start, i, nr_pages);
1779 	if (err >= 0)
1780 		err = err1;
1781 out:
1782 	lru_cache_enable();
1783 	return err;
1784 }
1785 
1786 /*
1787  * Determine the nodes of an array of pages and store it in an array of status.
1788  */
1789 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1790 				const void __user **pages, int *status)
1791 {
1792 	unsigned long i;
1793 
1794 	mmap_read_lock(mm);
1795 
1796 	for (i = 0; i < nr_pages; i++) {
1797 		unsigned long addr = (unsigned long)(*pages);
1798 		struct vm_area_struct *vma;
1799 		struct page *page;
1800 		int err = -EFAULT;
1801 
1802 		vma = vma_lookup(mm, addr);
1803 		if (!vma)
1804 			goto set_status;
1805 
1806 		/* FOLL_DUMP to ignore special (like zero) pages */
1807 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1808 
1809 		err = PTR_ERR(page);
1810 		if (IS_ERR(page))
1811 			goto set_status;
1812 
1813 		if (page) {
1814 			err = page_to_nid(page);
1815 			put_page(page);
1816 		} else {
1817 			err = -ENOENT;
1818 		}
1819 set_status:
1820 		*status = err;
1821 
1822 		pages++;
1823 		status++;
1824 	}
1825 
1826 	mmap_read_unlock(mm);
1827 }
1828 
1829 static int get_compat_pages_array(const void __user *chunk_pages[],
1830 				  const void __user * __user *pages,
1831 				  unsigned long chunk_nr)
1832 {
1833 	compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1834 	compat_uptr_t p;
1835 	int i;
1836 
1837 	for (i = 0; i < chunk_nr; i++) {
1838 		if (get_user(p, pages32 + i))
1839 			return -EFAULT;
1840 		chunk_pages[i] = compat_ptr(p);
1841 	}
1842 
1843 	return 0;
1844 }
1845 
1846 /*
1847  * Determine the nodes of a user array of pages and store it in
1848  * a user array of status.
1849  */
1850 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1851 			 const void __user * __user *pages,
1852 			 int __user *status)
1853 {
1854 #define DO_PAGES_STAT_CHUNK_NR 16UL
1855 	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1856 	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1857 
1858 	while (nr_pages) {
1859 		unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1860 
1861 		if (in_compat_syscall()) {
1862 			if (get_compat_pages_array(chunk_pages, pages,
1863 						   chunk_nr))
1864 				break;
1865 		} else {
1866 			if (copy_from_user(chunk_pages, pages,
1867 				      chunk_nr * sizeof(*chunk_pages)))
1868 				break;
1869 		}
1870 
1871 		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1872 
1873 		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1874 			break;
1875 
1876 		pages += chunk_nr;
1877 		status += chunk_nr;
1878 		nr_pages -= chunk_nr;
1879 	}
1880 	return nr_pages ? -EFAULT : 0;
1881 }
1882 
1883 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1884 {
1885 	struct task_struct *task;
1886 	struct mm_struct *mm;
1887 
1888 	/*
1889 	 * There is no need to check if current process has the right to modify
1890 	 * the specified process when they are same.
1891 	 */
1892 	if (!pid) {
1893 		mmget(current->mm);
1894 		*mem_nodes = cpuset_mems_allowed(current);
1895 		return current->mm;
1896 	}
1897 
1898 	/* Find the mm_struct */
1899 	rcu_read_lock();
1900 	task = find_task_by_vpid(pid);
1901 	if (!task) {
1902 		rcu_read_unlock();
1903 		return ERR_PTR(-ESRCH);
1904 	}
1905 	get_task_struct(task);
1906 
1907 	/*
1908 	 * Check if this process has the right to modify the specified
1909 	 * process. Use the regular "ptrace_may_access()" checks.
1910 	 */
1911 	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1912 		rcu_read_unlock();
1913 		mm = ERR_PTR(-EPERM);
1914 		goto out;
1915 	}
1916 	rcu_read_unlock();
1917 
1918 	mm = ERR_PTR(security_task_movememory(task));
1919 	if (IS_ERR(mm))
1920 		goto out;
1921 	*mem_nodes = cpuset_mems_allowed(task);
1922 	mm = get_task_mm(task);
1923 out:
1924 	put_task_struct(task);
1925 	if (!mm)
1926 		mm = ERR_PTR(-EINVAL);
1927 	return mm;
1928 }
1929 
1930 /*
1931  * Move a list of pages in the address space of the currently executing
1932  * process.
1933  */
1934 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1935 			     const void __user * __user *pages,
1936 			     const int __user *nodes,
1937 			     int __user *status, int flags)
1938 {
1939 	struct mm_struct *mm;
1940 	int err;
1941 	nodemask_t task_nodes;
1942 
1943 	/* Check flags */
1944 	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1945 		return -EINVAL;
1946 
1947 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1948 		return -EPERM;
1949 
1950 	mm = find_mm_struct(pid, &task_nodes);
1951 	if (IS_ERR(mm))
1952 		return PTR_ERR(mm);
1953 
1954 	if (nodes)
1955 		err = do_pages_move(mm, task_nodes, nr_pages, pages,
1956 				    nodes, status, flags);
1957 	else
1958 		err = do_pages_stat(mm, nr_pages, pages, status);
1959 
1960 	mmput(mm);
1961 	return err;
1962 }
1963 
1964 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1965 		const void __user * __user *, pages,
1966 		const int __user *, nodes,
1967 		int __user *, status, int, flags)
1968 {
1969 	return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1970 }
1971 
1972 #ifdef CONFIG_NUMA_BALANCING
1973 /*
1974  * Returns true if this is a safe migration target node for misplaced NUMA
1975  * pages. Currently it only checks the watermarks which is crude.
1976  */
1977 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1978 				   unsigned long nr_migrate_pages)
1979 {
1980 	int z;
1981 
1982 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1983 		struct zone *zone = pgdat->node_zones + z;
1984 
1985 		if (!managed_zone(zone))
1986 			continue;
1987 
1988 		/* Avoid waking kswapd by allocating pages_to_migrate pages. */
1989 		if (!zone_watermark_ok(zone, 0,
1990 				       high_wmark_pages(zone) +
1991 				       nr_migrate_pages,
1992 				       ZONE_MOVABLE, 0))
1993 			continue;
1994 		return true;
1995 	}
1996 	return false;
1997 }
1998 
1999 static struct page *alloc_misplaced_dst_page(struct page *page,
2000 					   unsigned long data)
2001 {
2002 	int nid = (int) data;
2003 	int order = compound_order(page);
2004 	gfp_t gfp = __GFP_THISNODE;
2005 	struct folio *new;
2006 
2007 	if (order > 0)
2008 		gfp |= GFP_TRANSHUGE_LIGHT;
2009 	else {
2010 		gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2011 			__GFP_NOWARN;
2012 		gfp &= ~__GFP_RECLAIM;
2013 	}
2014 	new = __folio_alloc_node(gfp, order, nid);
2015 
2016 	return &new->page;
2017 }
2018 
2019 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2020 {
2021 	int nr_pages = thp_nr_pages(page);
2022 	int order = compound_order(page);
2023 
2024 	VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2025 
2026 	/* Do not migrate THP mapped by multiple processes */
2027 	if (PageTransHuge(page) && total_mapcount(page) > 1)
2028 		return 0;
2029 
2030 	/* Avoid migrating to a node that is nearly full */
2031 	if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2032 		int z;
2033 
2034 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2035 			return 0;
2036 		for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2037 			if (managed_zone(pgdat->node_zones + z))
2038 				break;
2039 		}
2040 		wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2041 		return 0;
2042 	}
2043 
2044 	if (isolate_lru_page(page))
2045 		return 0;
2046 
2047 	mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2048 			    nr_pages);
2049 
2050 	/*
2051 	 * Isolating the page has taken another reference, so the
2052 	 * caller's reference can be safely dropped without the page
2053 	 * disappearing underneath us during migration.
2054 	 */
2055 	put_page(page);
2056 	return 1;
2057 }
2058 
2059 /*
2060  * Attempt to migrate a misplaced page to the specified destination
2061  * node. Caller is expected to have an elevated reference count on
2062  * the page that will be dropped by this function before returning.
2063  */
2064 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2065 			   int node)
2066 {
2067 	pg_data_t *pgdat = NODE_DATA(node);
2068 	int isolated;
2069 	int nr_remaining;
2070 	unsigned int nr_succeeded;
2071 	LIST_HEAD(migratepages);
2072 	int nr_pages = thp_nr_pages(page);
2073 
2074 	/*
2075 	 * Don't migrate file pages that are mapped in multiple processes
2076 	 * with execute permissions as they are probably shared libraries.
2077 	 */
2078 	if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2079 	    (vma->vm_flags & VM_EXEC))
2080 		goto out;
2081 
2082 	/*
2083 	 * Also do not migrate dirty pages as not all filesystems can move
2084 	 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2085 	 */
2086 	if (page_is_file_lru(page) && PageDirty(page))
2087 		goto out;
2088 
2089 	isolated = numamigrate_isolate_page(pgdat, page);
2090 	if (!isolated)
2091 		goto out;
2092 
2093 	list_add(&page->lru, &migratepages);
2094 	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2095 				     NULL, node, MIGRATE_ASYNC,
2096 				     MR_NUMA_MISPLACED, &nr_succeeded);
2097 	if (nr_remaining) {
2098 		if (!list_empty(&migratepages)) {
2099 			list_del(&page->lru);
2100 			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2101 					page_is_file_lru(page), -nr_pages);
2102 			putback_lru_page(page);
2103 		}
2104 		isolated = 0;
2105 	}
2106 	if (nr_succeeded) {
2107 		count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2108 		if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2109 			mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2110 					    nr_succeeded);
2111 	}
2112 	BUG_ON(!list_empty(&migratepages));
2113 	return isolated;
2114 
2115 out:
2116 	put_page(page);
2117 	return 0;
2118 }
2119 #endif /* CONFIG_NUMA_BALANCING */
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 int next_demotion_node(int node)
2200 {
2201 	struct demotion_nodes *nd;
2202 	unsigned short target_nr, index;
2203 	int target;
2204 
2205 	if (!node_demotion)
2206 		return NUMA_NO_NODE;
2207 
2208 	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 	target_nr = READ_ONCE(nd->nr);
2221 
2222 	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 		index = get_random_int() % target_nr;
2243 		break;
2244 	}
2245 
2246 	target = READ_ONCE(nd->nodes[index]);
2247 
2248 out:
2249 	rcu_read_unlock();
2250 	return target;
2251 }
2252 
2253 /* Disable reclaim-based migration. */
2254 static void __disable_all_migrate_targets(void)
2255 {
2256 	int node, i;
2257 
2258 	if (!node_demotion)
2259 		return;
2260 
2261 	for_each_online_node(node) {
2262 		node_demotion[node].nr = 0;
2263 		for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2264 			node_demotion[node].nodes[i] = NUMA_NO_NODE;
2265 	}
2266 }
2267 
2268 static void disable_all_migrate_targets(void)
2269 {
2270 	__disable_all_migrate_targets();
2271 
2272 	/*
2273 	 * Ensure that the "disable" is visible across the system.
2274 	 * Readers will see either a combination of before+disable
2275 	 * state or disable+after.  They will never see before and
2276 	 * after state together.
2277 	 *
2278 	 * The before+after state together might have cycles and
2279 	 * could cause readers to do things like loop until this
2280 	 * function finishes.  This ensures they can only see a
2281 	 * single "bad" read and would, for instance, only loop
2282 	 * once.
2283 	 */
2284 	synchronize_rcu();
2285 }
2286 
2287 /*
2288  * Find an automatic demotion target for 'node'.
2289  * Failing here is OK.  It might just indicate
2290  * being at the end of a chain.
2291  */
2292 static int establish_migrate_target(int node, nodemask_t *used,
2293 				    int best_distance)
2294 {
2295 	int migration_target, index, val;
2296 	struct demotion_nodes *nd;
2297 
2298 	if (!node_demotion)
2299 		return NUMA_NO_NODE;
2300 
2301 	nd = &node_demotion[node];
2302 
2303 	migration_target = find_next_best_node(node, used);
2304 	if (migration_target == NUMA_NO_NODE)
2305 		return NUMA_NO_NODE;
2306 
2307 	/*
2308 	 * If the node has been set a migration target node before,
2309 	 * which means it's the best distance between them. Still
2310 	 * check if this node can be demoted to other target nodes
2311 	 * if they have a same best distance.
2312 	 */
2313 	if (best_distance != -1) {
2314 		val = node_distance(node, migration_target);
2315 		if (val > best_distance)
2316 			goto out_clear;
2317 	}
2318 
2319 	index = nd->nr;
2320 	if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2321 		      "Exceeds maximum demotion target nodes\n"))
2322 		goto out_clear;
2323 
2324 	nd->nodes[index] = migration_target;
2325 	nd->nr++;
2326 
2327 	return migration_target;
2328 out_clear:
2329 	node_clear(migration_target, *used);
2330 	return NUMA_NO_NODE;
2331 }
2332 
2333 /*
2334  * When memory fills up on a node, memory contents can be
2335  * automatically migrated to another node instead of
2336  * discarded at reclaim.
2337  *
2338  * Establish a "migration path" which will start at nodes
2339  * with CPUs and will follow the priorities used to build the
2340  * page allocator zonelists.
2341  *
2342  * The difference here is that cycles must be avoided.  If
2343  * node0 migrates to node1, then neither node1, nor anything
2344  * node1 migrates to can migrate to node0. Also one node can
2345  * be migrated to multiple nodes if the target nodes all have
2346  * a same best-distance against the source node.
2347  *
2348  * This function can run simultaneously with readers of
2349  * node_demotion[].  However, it can not run simultaneously
2350  * with itself.  Exclusion is provided by memory hotplug events
2351  * being single-threaded.
2352  */
2353 static void __set_migration_target_nodes(void)
2354 {
2355 	nodemask_t next_pass;
2356 	nodemask_t this_pass;
2357 	nodemask_t used_targets = NODE_MASK_NONE;
2358 	int node, best_distance;
2359 
2360 	/*
2361 	 * Avoid any oddities like cycles that could occur
2362 	 * from changes in the topology.  This will leave
2363 	 * a momentary gap when migration is disabled.
2364 	 */
2365 	disable_all_migrate_targets();
2366 
2367 	/*
2368 	 * Allocations go close to CPUs, first.  Assume that
2369 	 * the migration path starts at the nodes with CPUs.
2370 	 */
2371 	next_pass = node_states[N_CPU];
2372 again:
2373 	this_pass = next_pass;
2374 	next_pass = NODE_MASK_NONE;
2375 	/*
2376 	 * To avoid cycles in the migration "graph", ensure
2377 	 * that migration sources are not future targets by
2378 	 * setting them in 'used_targets'.  Do this only
2379 	 * once per pass so that multiple source nodes can
2380 	 * share a target node.
2381 	 *
2382 	 * 'used_targets' will become unavailable in future
2383 	 * passes.  This limits some opportunities for
2384 	 * multiple source nodes to share a destination.
2385 	 */
2386 	nodes_or(used_targets, used_targets, this_pass);
2387 
2388 	for_each_node_mask(node, this_pass) {
2389 		best_distance = -1;
2390 
2391 		/*
2392 		 * Try to set up the migration path for the node, and the target
2393 		 * migration nodes can be multiple, so doing a loop to find all
2394 		 * the target nodes if they all have a best node distance.
2395 		 */
2396 		do {
2397 			int target_node =
2398 				establish_migrate_target(node, &used_targets,
2399 							 best_distance);
2400 
2401 			if (target_node == NUMA_NO_NODE)
2402 				break;
2403 
2404 			if (best_distance == -1)
2405 				best_distance = node_distance(node, target_node);
2406 
2407 			/*
2408 			 * Visit targets from this pass in the next pass.
2409 			 * Eventually, every node will have been part of
2410 			 * a pass, and will become set in 'used_targets'.
2411 			 */
2412 			node_set(target_node, next_pass);
2413 		} while (1);
2414 	}
2415 	/*
2416 	 * 'next_pass' contains nodes which became migration
2417 	 * targets in this pass.  Make additional passes until
2418 	 * no more migrations targets are available.
2419 	 */
2420 	if (!nodes_empty(next_pass))
2421 		goto again;
2422 }
2423 
2424 /*
2425  * For callers that do not hold get_online_mems() already.
2426  */
2427 void set_migration_target_nodes(void)
2428 {
2429 	get_online_mems();
2430 	__set_migration_target_nodes();
2431 	put_online_mems();
2432 }
2433 
2434 /*
2435  * This leaves migrate-on-reclaim transiently disabled between
2436  * the MEM_GOING_OFFLINE and MEM_OFFLINE events.  This runs
2437  * whether reclaim-based migration is enabled or not, which
2438  * ensures that the user can turn reclaim-based migration at
2439  * any time without needing to recalculate migration targets.
2440  *
2441  * These callbacks already hold get_online_mems().  That is why
2442  * __set_migration_target_nodes() can be used as opposed to
2443  * set_migration_target_nodes().
2444  */
2445 #ifdef CONFIG_MEMORY_HOTPLUG
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 #endif
2492 
2493 void __init migrate_on_reclaim_init(void)
2494 {
2495 	node_demotion = kcalloc(nr_node_ids,
2496 				sizeof(struct demotion_nodes),
2497 				GFP_KERNEL);
2498 	WARN_ON(!node_demotion);
2499 #ifdef CONFIG_MEMORY_HOTPLUG
2500 	hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2501 #endif
2502 	/*
2503 	 * At this point, all numa nodes with memory/CPus have their state
2504 	 * properly set, so we can build the demotion order now.
2505 	 * Let us hold the cpu_hotplug lock just, as we could possibily have
2506 	 * CPU hotplug events during boot.
2507 	 */
2508 	cpus_read_lock();
2509 	set_migration_target_nodes();
2510 	cpus_read_unlock();
2511 }
2512 
2513 bool numa_demotion_enabled = false;
2514 
2515 #ifdef CONFIG_SYSFS
2516 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2517 					  struct kobj_attribute *attr, char *buf)
2518 {
2519 	return sysfs_emit(buf, "%s\n",
2520 			  numa_demotion_enabled ? "true" : "false");
2521 }
2522 
2523 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2524 					   struct kobj_attribute *attr,
2525 					   const char *buf, size_t count)
2526 {
2527 	ssize_t ret;
2528 
2529 	ret = kstrtobool(buf, &numa_demotion_enabled);
2530 	if (ret)
2531 		return ret;
2532 
2533 	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 static int __init numa_init_sysfs(void)
2550 {
2551 	int err;
2552 	struct kobject *numa_kobj;
2553 
2554 	numa_kobj = kobject_create_and_add("numa", mm_kobj);
2555 	if (!numa_kobj) {
2556 		pr_err("failed to create numa kobject\n");
2557 		return -ENOMEM;
2558 	}
2559 	err = sysfs_create_group(numa_kobj, &numa_attr_group);
2560 	if (err) {
2561 		pr_err("failed to register numa group\n");
2562 		goto delete_obj;
2563 	}
2564 	return 0;
2565 
2566 delete_obj:
2567 	kobject_put(numa_kobj);
2568 	return err;
2569 }
2570 subsys_initcall(numa_init_sysfs);
2571 #endif /* CONFIG_SYSFS */
2572 #endif /* CONFIG_NUMA */
2573