xref: /openbmc/linux/mm/migrate.c (revision 0cb4228f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Memory Migration functionality - linux/mm/migrate.c
4  *
5  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6  *
7  * Page migration was first developed in the context of the memory hotplug
8  * project. The main authors of the migration code are:
9  *
10  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11  * Hirokazu Takahashi <taka@valinux.co.jp>
12  * Dave Hansen <haveblue@us.ibm.com>
13  * Christoph Lameter
14  */
15 
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
53 #include <linux/memory-tiers.h>
54 
55 #include <asm/tlbflush.h>
56 
57 #include <trace/events/migrate.h>
58 
59 #include "internal.h"
60 
61 int isolate_movable_page(struct page *page, isolate_mode_t mode)
62 {
63 	const struct movable_operations *mops;
64 
65 	/*
66 	 * Avoid burning cycles with pages that are yet under __free_pages(),
67 	 * or just got freed under us.
68 	 *
69 	 * In case we 'win' a race for a movable page being freed under us and
70 	 * raise its refcount preventing __free_pages() from doing its job
71 	 * the put_page() at the end of this block will take care of
72 	 * release this page, thus avoiding a nasty leakage.
73 	 */
74 	if (unlikely(!get_page_unless_zero(page)))
75 		goto out;
76 
77 	if (unlikely(PageSlab(page)))
78 		goto out_putpage;
79 	/* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
80 	smp_rmb();
81 	/*
82 	 * Check movable flag before taking the page lock because
83 	 * we use non-atomic bitops on newly allocated page flags so
84 	 * unconditionally grabbing the lock ruins page's owner side.
85 	 */
86 	if (unlikely(!__PageMovable(page)))
87 		goto out_putpage;
88 	/* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
89 	smp_rmb();
90 	if (unlikely(PageSlab(page)))
91 		goto out_putpage;
92 
93 	/*
94 	 * As movable pages are not isolated from LRU lists, concurrent
95 	 * compaction threads can race against page migration functions
96 	 * as well as race against the releasing a page.
97 	 *
98 	 * In order to avoid having an already isolated movable page
99 	 * being (wrongly) re-isolated while it is under migration,
100 	 * or to avoid attempting to isolate pages being released,
101 	 * lets be sure we have the page lock
102 	 * before proceeding with the movable page isolation steps.
103 	 */
104 	if (unlikely(!trylock_page(page)))
105 		goto out_putpage;
106 
107 	if (!PageMovable(page) || PageIsolated(page))
108 		goto out_no_isolated;
109 
110 	mops = page_movable_ops(page);
111 	VM_BUG_ON_PAGE(!mops, page);
112 
113 	if (!mops->isolate_page(page, mode))
114 		goto out_no_isolated;
115 
116 	/* Driver shouldn't use PG_isolated bit of page->flags */
117 	WARN_ON_ONCE(PageIsolated(page));
118 	SetPageIsolated(page);
119 	unlock_page(page);
120 
121 	return 0;
122 
123 out_no_isolated:
124 	unlock_page(page);
125 out_putpage:
126 	put_page(page);
127 out:
128 	return -EBUSY;
129 }
130 
131 static void putback_movable_page(struct page *page)
132 {
133 	const struct movable_operations *mops = page_movable_ops(page);
134 
135 	mops->putback_page(page);
136 	ClearPageIsolated(page);
137 }
138 
139 /*
140  * Put previously isolated pages back onto the appropriate lists
141  * from where they were once taken off for compaction/migration.
142  *
143  * This function shall be used whenever the isolated pageset has been
144  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
145  * and isolate_hugetlb().
146  */
147 void putback_movable_pages(struct list_head *l)
148 {
149 	struct page *page;
150 	struct page *page2;
151 
152 	list_for_each_entry_safe(page, page2, l, lru) {
153 		if (unlikely(PageHuge(page))) {
154 			putback_active_hugepage(page);
155 			continue;
156 		}
157 		list_del(&page->lru);
158 		/*
159 		 * We isolated non-lru movable page so here we can use
160 		 * __PageMovable because LRU page's mapping cannot have
161 		 * PAGE_MAPPING_MOVABLE.
162 		 */
163 		if (unlikely(__PageMovable(page))) {
164 			VM_BUG_ON_PAGE(!PageIsolated(page), page);
165 			lock_page(page);
166 			if (PageMovable(page))
167 				putback_movable_page(page);
168 			else
169 				ClearPageIsolated(page);
170 			unlock_page(page);
171 			put_page(page);
172 		} else {
173 			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
174 					page_is_file_lru(page), -thp_nr_pages(page));
175 			putback_lru_page(page);
176 		}
177 	}
178 }
179 
180 /*
181  * Restore a potential migration pte to a working pte entry
182  */
183 static bool remove_migration_pte(struct folio *folio,
184 		struct vm_area_struct *vma, unsigned long addr, void *old)
185 {
186 	DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
187 
188 	while (page_vma_mapped_walk(&pvmw)) {
189 		rmap_t rmap_flags = RMAP_NONE;
190 		pte_t pte;
191 		swp_entry_t entry;
192 		struct page *new;
193 		unsigned long idx = 0;
194 
195 		/* pgoff is invalid for ksm pages, but they are never large */
196 		if (folio_test_large(folio) && !folio_test_hugetlb(folio))
197 			idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
198 		new = folio_page(folio, idx);
199 
200 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
201 		/* PMD-mapped THP migration entry */
202 		if (!pvmw.pte) {
203 			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
204 					!folio_test_pmd_mappable(folio), folio);
205 			remove_migration_pmd(&pvmw, new);
206 			continue;
207 		}
208 #endif
209 
210 		folio_get(folio);
211 		pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
212 		if (pte_swp_soft_dirty(*pvmw.pte))
213 			pte = pte_mksoft_dirty(pte);
214 
215 		/*
216 		 * Recheck VMA as permissions can change since migration started
217 		 */
218 		entry = pte_to_swp_entry(*pvmw.pte);
219 		if (!is_migration_entry_young(entry))
220 			pte = pte_mkold(pte);
221 		if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
222 			pte = pte_mkdirty(pte);
223 		if (is_writable_migration_entry(entry))
224 			pte = maybe_mkwrite(pte, vma);
225 		else if (pte_swp_uffd_wp(*pvmw.pte))
226 			pte = pte_mkuffd_wp(pte);
227 
228 		if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
229 			rmap_flags |= RMAP_EXCLUSIVE;
230 
231 		if (unlikely(is_device_private_page(new))) {
232 			if (pte_write(pte))
233 				entry = make_writable_device_private_entry(
234 							page_to_pfn(new));
235 			else
236 				entry = make_readable_device_private_entry(
237 							page_to_pfn(new));
238 			pte = swp_entry_to_pte(entry);
239 			if (pte_swp_soft_dirty(*pvmw.pte))
240 				pte = pte_swp_mksoft_dirty(pte);
241 			if (pte_swp_uffd_wp(*pvmw.pte))
242 				pte = pte_swp_mkuffd_wp(pte);
243 		}
244 
245 #ifdef CONFIG_HUGETLB_PAGE
246 		if (folio_test_hugetlb(folio)) {
247 			unsigned int shift = huge_page_shift(hstate_vma(vma));
248 
249 			pte = pte_mkhuge(pte);
250 			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
251 			if (folio_test_anon(folio))
252 				hugepage_add_anon_rmap(new, vma, pvmw.address,
253 						       rmap_flags);
254 			else
255 				page_dup_file_rmap(new, true);
256 			set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
257 		} else
258 #endif
259 		{
260 			if (folio_test_anon(folio))
261 				page_add_anon_rmap(new, vma, pvmw.address,
262 						   rmap_flags);
263 			else
264 				page_add_file_rmap(new, vma, false);
265 			set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
266 		}
267 		if (vma->vm_flags & VM_LOCKED)
268 			mlock_page_drain_local();
269 
270 		trace_remove_migration_pte(pvmw.address, pte_val(pte),
271 					   compound_order(new));
272 
273 		/* No need to invalidate - it was non-present before */
274 		update_mmu_cache(vma, pvmw.address, pvmw.pte);
275 	}
276 
277 	return true;
278 }
279 
280 /*
281  * Get rid of all migration entries and replace them by
282  * references to the indicated page.
283  */
284 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
285 {
286 	struct rmap_walk_control rwc = {
287 		.rmap_one = remove_migration_pte,
288 		.arg = src,
289 	};
290 
291 	if (locked)
292 		rmap_walk_locked(dst, &rwc);
293 	else
294 		rmap_walk(dst, &rwc);
295 }
296 
297 /*
298  * Something used the pte of a page under migration. We need to
299  * get to the page and wait until migration is finished.
300  * When we return from this function the fault will be retried.
301  */
302 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
303 				spinlock_t *ptl)
304 {
305 	pte_t pte;
306 	swp_entry_t entry;
307 
308 	spin_lock(ptl);
309 	pte = *ptep;
310 	if (!is_swap_pte(pte))
311 		goto out;
312 
313 	entry = pte_to_swp_entry(pte);
314 	if (!is_migration_entry(entry))
315 		goto out;
316 
317 	migration_entry_wait_on_locked(entry, ptep, ptl);
318 	return;
319 out:
320 	pte_unmap_unlock(ptep, ptl);
321 }
322 
323 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
324 				unsigned long address)
325 {
326 	spinlock_t *ptl = pte_lockptr(mm, pmd);
327 	pte_t *ptep = pte_offset_map(pmd, address);
328 	__migration_entry_wait(mm, ptep, ptl);
329 }
330 
331 #ifdef CONFIG_HUGETLB_PAGE
332 void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl)
333 {
334 	pte_t pte;
335 
336 	spin_lock(ptl);
337 	pte = huge_ptep_get(ptep);
338 
339 	if (unlikely(!is_hugetlb_entry_migration(pte)))
340 		spin_unlock(ptl);
341 	else
342 		migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
343 }
344 
345 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte)
346 {
347 	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte);
348 
349 	__migration_entry_wait_huge(pte, ptl);
350 }
351 #endif
352 
353 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
354 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
355 {
356 	spinlock_t *ptl;
357 
358 	ptl = pmd_lock(mm, pmd);
359 	if (!is_pmd_migration_entry(*pmd))
360 		goto unlock;
361 	migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
362 	return;
363 unlock:
364 	spin_unlock(ptl);
365 }
366 #endif
367 
368 static int folio_expected_refs(struct address_space *mapping,
369 		struct folio *folio)
370 {
371 	int refs = 1;
372 	if (!mapping)
373 		return refs;
374 
375 	refs += folio_nr_pages(folio);
376 	if (folio_test_private(folio))
377 		refs++;
378 
379 	return refs;
380 }
381 
382 /*
383  * Replace the page in the mapping.
384  *
385  * The number of remaining references must be:
386  * 1 for anonymous pages without a mapping
387  * 2 for pages with a mapping
388  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
389  */
390 int folio_migrate_mapping(struct address_space *mapping,
391 		struct folio *newfolio, struct folio *folio, int extra_count)
392 {
393 	XA_STATE(xas, &mapping->i_pages, folio_index(folio));
394 	struct zone *oldzone, *newzone;
395 	int dirty;
396 	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
397 	long nr = folio_nr_pages(folio);
398 
399 	if (!mapping) {
400 		/* Anonymous page without mapping */
401 		if (folio_ref_count(folio) != expected_count)
402 			return -EAGAIN;
403 
404 		/* No turning back from here */
405 		newfolio->index = folio->index;
406 		newfolio->mapping = folio->mapping;
407 		if (folio_test_swapbacked(folio))
408 			__folio_set_swapbacked(newfolio);
409 
410 		return MIGRATEPAGE_SUCCESS;
411 	}
412 
413 	oldzone = folio_zone(folio);
414 	newzone = folio_zone(newfolio);
415 
416 	xas_lock_irq(&xas);
417 	if (!folio_ref_freeze(folio, expected_count)) {
418 		xas_unlock_irq(&xas);
419 		return -EAGAIN;
420 	}
421 
422 	/*
423 	 * Now we know that no one else is looking at the folio:
424 	 * no turning back from here.
425 	 */
426 	newfolio->index = folio->index;
427 	newfolio->mapping = folio->mapping;
428 	folio_ref_add(newfolio, nr); /* add cache reference */
429 	if (folio_test_swapbacked(folio)) {
430 		__folio_set_swapbacked(newfolio);
431 		if (folio_test_swapcache(folio)) {
432 			folio_set_swapcache(newfolio);
433 			newfolio->private = folio_get_private(folio);
434 		}
435 	} else {
436 		VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
437 	}
438 
439 	/* Move dirty while page refs frozen and newpage not yet exposed */
440 	dirty = folio_test_dirty(folio);
441 	if (dirty) {
442 		folio_clear_dirty(folio);
443 		folio_set_dirty(newfolio);
444 	}
445 
446 	xas_store(&xas, newfolio);
447 
448 	/*
449 	 * Drop cache reference from old page by unfreezing
450 	 * to one less reference.
451 	 * We know this isn't the last reference.
452 	 */
453 	folio_ref_unfreeze(folio, expected_count - nr);
454 
455 	xas_unlock(&xas);
456 	/* Leave irq disabled to prevent preemption while updating stats */
457 
458 	/*
459 	 * If moved to a different zone then also account
460 	 * the page for that zone. Other VM counters will be
461 	 * taken care of when we establish references to the
462 	 * new page and drop references to the old page.
463 	 *
464 	 * Note that anonymous pages are accounted for
465 	 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
466 	 * are mapped to swap space.
467 	 */
468 	if (newzone != oldzone) {
469 		struct lruvec *old_lruvec, *new_lruvec;
470 		struct mem_cgroup *memcg;
471 
472 		memcg = folio_memcg(folio);
473 		old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
474 		new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
475 
476 		__mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
477 		__mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
478 		if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
479 			__mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
480 			__mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
481 		}
482 #ifdef CONFIG_SWAP
483 		if (folio_test_swapcache(folio)) {
484 			__mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
485 			__mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
486 		}
487 #endif
488 		if (dirty && mapping_can_writeback(mapping)) {
489 			__mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
490 			__mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
491 			__mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
492 			__mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
493 		}
494 	}
495 	local_irq_enable();
496 
497 	return MIGRATEPAGE_SUCCESS;
498 }
499 EXPORT_SYMBOL(folio_migrate_mapping);
500 
501 /*
502  * The expected number of remaining references is the same as that
503  * of folio_migrate_mapping().
504  */
505 int migrate_huge_page_move_mapping(struct address_space *mapping,
506 				   struct folio *dst, struct folio *src)
507 {
508 	XA_STATE(xas, &mapping->i_pages, folio_index(src));
509 	int expected_count;
510 
511 	xas_lock_irq(&xas);
512 	expected_count = 2 + folio_has_private(src);
513 	if (!folio_ref_freeze(src, expected_count)) {
514 		xas_unlock_irq(&xas);
515 		return -EAGAIN;
516 	}
517 
518 	dst->index = src->index;
519 	dst->mapping = src->mapping;
520 
521 	folio_get(dst);
522 
523 	xas_store(&xas, dst);
524 
525 	folio_ref_unfreeze(src, expected_count - 1);
526 
527 	xas_unlock_irq(&xas);
528 
529 	return MIGRATEPAGE_SUCCESS;
530 }
531 
532 /*
533  * Copy the flags and some other ancillary information
534  */
535 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
536 {
537 	int cpupid;
538 
539 	if (folio_test_error(folio))
540 		folio_set_error(newfolio);
541 	if (folio_test_referenced(folio))
542 		folio_set_referenced(newfolio);
543 	if (folio_test_uptodate(folio))
544 		folio_mark_uptodate(newfolio);
545 	if (folio_test_clear_active(folio)) {
546 		VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
547 		folio_set_active(newfolio);
548 	} else if (folio_test_clear_unevictable(folio))
549 		folio_set_unevictable(newfolio);
550 	if (folio_test_workingset(folio))
551 		folio_set_workingset(newfolio);
552 	if (folio_test_checked(folio))
553 		folio_set_checked(newfolio);
554 	/*
555 	 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
556 	 * migration entries. We can still have PG_anon_exclusive set on an
557 	 * effectively unmapped and unreferenced first sub-pages of an
558 	 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
559 	 */
560 	if (folio_test_mappedtodisk(folio))
561 		folio_set_mappedtodisk(newfolio);
562 
563 	/* Move dirty on pages not done by folio_migrate_mapping() */
564 	if (folio_test_dirty(folio))
565 		folio_set_dirty(newfolio);
566 
567 	if (folio_test_young(folio))
568 		folio_set_young(newfolio);
569 	if (folio_test_idle(folio))
570 		folio_set_idle(newfolio);
571 
572 	/*
573 	 * Copy NUMA information to the new page, to prevent over-eager
574 	 * future migrations of this same page.
575 	 */
576 	cpupid = page_cpupid_xchg_last(&folio->page, -1);
577 	/*
578 	 * For memory tiering mode, when migrate between slow and fast
579 	 * memory node, reset cpupid, because that is used to record
580 	 * page access time in slow memory node.
581 	 */
582 	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
583 		bool f_toptier = node_is_toptier(page_to_nid(&folio->page));
584 		bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page));
585 
586 		if (f_toptier != t_toptier)
587 			cpupid = -1;
588 	}
589 	page_cpupid_xchg_last(&newfolio->page, cpupid);
590 
591 	folio_migrate_ksm(newfolio, folio);
592 	/*
593 	 * Please do not reorder this without considering how mm/ksm.c's
594 	 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
595 	 */
596 	if (folio_test_swapcache(folio))
597 		folio_clear_swapcache(folio);
598 	folio_clear_private(folio);
599 
600 	/* page->private contains hugetlb specific flags */
601 	if (!folio_test_hugetlb(folio))
602 		folio->private = NULL;
603 
604 	/*
605 	 * If any waiters have accumulated on the new page then
606 	 * wake them up.
607 	 */
608 	if (folio_test_writeback(newfolio))
609 		folio_end_writeback(newfolio);
610 
611 	/*
612 	 * PG_readahead shares the same bit with PG_reclaim.  The above
613 	 * end_page_writeback() may clear PG_readahead mistakenly, so set the
614 	 * bit after that.
615 	 */
616 	if (folio_test_readahead(folio))
617 		folio_set_readahead(newfolio);
618 
619 	folio_copy_owner(newfolio, folio);
620 
621 	if (!folio_test_hugetlb(folio))
622 		mem_cgroup_migrate(folio, newfolio);
623 }
624 EXPORT_SYMBOL(folio_migrate_flags);
625 
626 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
627 {
628 	folio_copy(newfolio, folio);
629 	folio_migrate_flags(newfolio, folio);
630 }
631 EXPORT_SYMBOL(folio_migrate_copy);
632 
633 /************************************************************
634  *                    Migration functions
635  ***********************************************************/
636 
637 int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
638 		struct folio *src, enum migrate_mode mode, int extra_count)
639 {
640 	int rc;
641 
642 	BUG_ON(folio_test_writeback(src));	/* Writeback must be complete */
643 
644 	rc = folio_migrate_mapping(mapping, dst, src, extra_count);
645 
646 	if (rc != MIGRATEPAGE_SUCCESS)
647 		return rc;
648 
649 	if (mode != MIGRATE_SYNC_NO_COPY)
650 		folio_migrate_copy(dst, src);
651 	else
652 		folio_migrate_flags(dst, src);
653 	return MIGRATEPAGE_SUCCESS;
654 }
655 
656 /**
657  * migrate_folio() - Simple folio migration.
658  * @mapping: The address_space containing the folio.
659  * @dst: The folio to migrate the data to.
660  * @src: The folio containing the current data.
661  * @mode: How to migrate the page.
662  *
663  * Common logic to directly migrate a single LRU folio suitable for
664  * folios that do not use PagePrivate/PagePrivate2.
665  *
666  * Folios are locked upon entry and exit.
667  */
668 int migrate_folio(struct address_space *mapping, struct folio *dst,
669 		struct folio *src, enum migrate_mode mode)
670 {
671 	return migrate_folio_extra(mapping, dst, src, mode, 0);
672 }
673 EXPORT_SYMBOL(migrate_folio);
674 
675 #ifdef CONFIG_BLOCK
676 /* Returns true if all buffers are successfully locked */
677 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
678 							enum migrate_mode mode)
679 {
680 	struct buffer_head *bh = head;
681 
682 	/* Simple case, sync compaction */
683 	if (mode != MIGRATE_ASYNC) {
684 		do {
685 			lock_buffer(bh);
686 			bh = bh->b_this_page;
687 
688 		} while (bh != head);
689 
690 		return true;
691 	}
692 
693 	/* async case, we cannot block on lock_buffer so use trylock_buffer */
694 	do {
695 		if (!trylock_buffer(bh)) {
696 			/*
697 			 * We failed to lock the buffer and cannot stall in
698 			 * async migration. Release the taken locks
699 			 */
700 			struct buffer_head *failed_bh = bh;
701 			bh = head;
702 			while (bh != failed_bh) {
703 				unlock_buffer(bh);
704 				bh = bh->b_this_page;
705 			}
706 			return false;
707 		}
708 
709 		bh = bh->b_this_page;
710 	} while (bh != head);
711 	return true;
712 }
713 
714 static int __buffer_migrate_folio(struct address_space *mapping,
715 		struct folio *dst, struct folio *src, enum migrate_mode mode,
716 		bool check_refs)
717 {
718 	struct buffer_head *bh, *head;
719 	int rc;
720 	int expected_count;
721 
722 	head = folio_buffers(src);
723 	if (!head)
724 		return migrate_folio(mapping, dst, src, mode);
725 
726 	/* Check whether page does not have extra refs before we do more work */
727 	expected_count = folio_expected_refs(mapping, src);
728 	if (folio_ref_count(src) != expected_count)
729 		return -EAGAIN;
730 
731 	if (!buffer_migrate_lock_buffers(head, mode))
732 		return -EAGAIN;
733 
734 	if (check_refs) {
735 		bool busy;
736 		bool invalidated = false;
737 
738 recheck_buffers:
739 		busy = false;
740 		spin_lock(&mapping->private_lock);
741 		bh = head;
742 		do {
743 			if (atomic_read(&bh->b_count)) {
744 				busy = true;
745 				break;
746 			}
747 			bh = bh->b_this_page;
748 		} while (bh != head);
749 		if (busy) {
750 			if (invalidated) {
751 				rc = -EAGAIN;
752 				goto unlock_buffers;
753 			}
754 			spin_unlock(&mapping->private_lock);
755 			invalidate_bh_lrus();
756 			invalidated = true;
757 			goto recheck_buffers;
758 		}
759 	}
760 
761 	rc = folio_migrate_mapping(mapping, dst, src, 0);
762 	if (rc != MIGRATEPAGE_SUCCESS)
763 		goto unlock_buffers;
764 
765 	folio_attach_private(dst, folio_detach_private(src));
766 
767 	bh = head;
768 	do {
769 		set_bh_page(bh, &dst->page, bh_offset(bh));
770 		bh = bh->b_this_page;
771 	} while (bh != head);
772 
773 	if (mode != MIGRATE_SYNC_NO_COPY)
774 		folio_migrate_copy(dst, src);
775 	else
776 		folio_migrate_flags(dst, src);
777 
778 	rc = MIGRATEPAGE_SUCCESS;
779 unlock_buffers:
780 	if (check_refs)
781 		spin_unlock(&mapping->private_lock);
782 	bh = head;
783 	do {
784 		unlock_buffer(bh);
785 		bh = bh->b_this_page;
786 	} while (bh != head);
787 
788 	return rc;
789 }
790 
791 /**
792  * buffer_migrate_folio() - Migration function for folios with buffers.
793  * @mapping: The address space containing @src.
794  * @dst: The folio to migrate to.
795  * @src: The folio to migrate from.
796  * @mode: How to migrate the folio.
797  *
798  * This function can only be used if the underlying filesystem guarantees
799  * that no other references to @src exist. For example attached buffer
800  * heads are accessed only under the folio lock.  If your filesystem cannot
801  * provide this guarantee, buffer_migrate_folio_norefs() may be more
802  * appropriate.
803  *
804  * Return: 0 on success or a negative errno on failure.
805  */
806 int buffer_migrate_folio(struct address_space *mapping,
807 		struct folio *dst, struct folio *src, enum migrate_mode mode)
808 {
809 	return __buffer_migrate_folio(mapping, dst, src, mode, false);
810 }
811 EXPORT_SYMBOL(buffer_migrate_folio);
812 
813 /**
814  * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
815  * @mapping: The address space containing @src.
816  * @dst: The folio to migrate to.
817  * @src: The folio to migrate from.
818  * @mode: How to migrate the folio.
819  *
820  * Like buffer_migrate_folio() except that this variant is more careful
821  * and checks that there are also no buffer head references. This function
822  * is the right one for mappings where buffer heads are directly looked
823  * up and referenced (such as block device mappings).
824  *
825  * Return: 0 on success or a negative errno on failure.
826  */
827 int buffer_migrate_folio_norefs(struct address_space *mapping,
828 		struct folio *dst, struct folio *src, enum migrate_mode mode)
829 {
830 	return __buffer_migrate_folio(mapping, dst, src, mode, true);
831 }
832 EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
833 #endif
834 
835 int filemap_migrate_folio(struct address_space *mapping,
836 		struct folio *dst, struct folio *src, enum migrate_mode mode)
837 {
838 	int ret;
839 
840 	ret = folio_migrate_mapping(mapping, dst, src, 0);
841 	if (ret != MIGRATEPAGE_SUCCESS)
842 		return ret;
843 
844 	if (folio_get_private(src))
845 		folio_attach_private(dst, folio_detach_private(src));
846 
847 	if (mode != MIGRATE_SYNC_NO_COPY)
848 		folio_migrate_copy(dst, src);
849 	else
850 		folio_migrate_flags(dst, src);
851 	return MIGRATEPAGE_SUCCESS;
852 }
853 EXPORT_SYMBOL_GPL(filemap_migrate_folio);
854 
855 /*
856  * Writeback a folio to clean the dirty state
857  */
858 static int writeout(struct address_space *mapping, struct folio *folio)
859 {
860 	struct writeback_control wbc = {
861 		.sync_mode = WB_SYNC_NONE,
862 		.nr_to_write = 1,
863 		.range_start = 0,
864 		.range_end = LLONG_MAX,
865 		.for_reclaim = 1
866 	};
867 	int rc;
868 
869 	if (!mapping->a_ops->writepage)
870 		/* No write method for the address space */
871 		return -EINVAL;
872 
873 	if (!folio_clear_dirty_for_io(folio))
874 		/* Someone else already triggered a write */
875 		return -EAGAIN;
876 
877 	/*
878 	 * A dirty folio may imply that the underlying filesystem has
879 	 * the folio on some queue. So the folio must be clean for
880 	 * migration. Writeout may mean we lose the lock and the
881 	 * folio state is no longer what we checked for earlier.
882 	 * At this point we know that the migration attempt cannot
883 	 * be successful.
884 	 */
885 	remove_migration_ptes(folio, folio, false);
886 
887 	rc = mapping->a_ops->writepage(&folio->page, &wbc);
888 
889 	if (rc != AOP_WRITEPAGE_ACTIVATE)
890 		/* unlocked. Relock */
891 		folio_lock(folio);
892 
893 	return (rc < 0) ? -EIO : -EAGAIN;
894 }
895 
896 /*
897  * Default handling if a filesystem does not provide a migration function.
898  */
899 static int fallback_migrate_folio(struct address_space *mapping,
900 		struct folio *dst, struct folio *src, enum migrate_mode mode)
901 {
902 	if (folio_test_dirty(src)) {
903 		/* Only writeback folios in full synchronous migration */
904 		switch (mode) {
905 		case MIGRATE_SYNC:
906 		case MIGRATE_SYNC_NO_COPY:
907 			break;
908 		default:
909 			return -EBUSY;
910 		}
911 		return writeout(mapping, src);
912 	}
913 
914 	/*
915 	 * Buffers may be managed in a filesystem specific way.
916 	 * We must have no buffers or drop them.
917 	 */
918 	if (folio_test_private(src) &&
919 	    !filemap_release_folio(src, GFP_KERNEL))
920 		return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
921 
922 	return migrate_folio(mapping, dst, src, mode);
923 }
924 
925 /*
926  * Move a page to a newly allocated page
927  * The page is locked and all ptes have been successfully removed.
928  *
929  * The new page will have replaced the old page if this function
930  * is successful.
931  *
932  * Return value:
933  *   < 0 - error code
934  *  MIGRATEPAGE_SUCCESS - success
935  */
936 static int move_to_new_folio(struct folio *dst, struct folio *src,
937 				enum migrate_mode mode)
938 {
939 	int rc = -EAGAIN;
940 	bool is_lru = !__PageMovable(&src->page);
941 
942 	VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
943 	VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
944 
945 	if (likely(is_lru)) {
946 		struct address_space *mapping = folio_mapping(src);
947 
948 		if (!mapping)
949 			rc = migrate_folio(mapping, dst, src, mode);
950 		else if (mapping->a_ops->migrate_folio)
951 			/*
952 			 * Most folios have a mapping and most filesystems
953 			 * provide a migrate_folio callback. Anonymous folios
954 			 * are part of swap space which also has its own
955 			 * migrate_folio callback. This is the most common path
956 			 * for page migration.
957 			 */
958 			rc = mapping->a_ops->migrate_folio(mapping, dst, src,
959 								mode);
960 		else
961 			rc = fallback_migrate_folio(mapping, dst, src, mode);
962 	} else {
963 		const struct movable_operations *mops;
964 
965 		/*
966 		 * In case of non-lru page, it could be released after
967 		 * isolation step. In that case, we shouldn't try migration.
968 		 */
969 		VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
970 		if (!folio_test_movable(src)) {
971 			rc = MIGRATEPAGE_SUCCESS;
972 			folio_clear_isolated(src);
973 			goto out;
974 		}
975 
976 		mops = page_movable_ops(&src->page);
977 		rc = mops->migrate_page(&dst->page, &src->page, mode);
978 		WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
979 				!folio_test_isolated(src));
980 	}
981 
982 	/*
983 	 * When successful, old pagecache src->mapping must be cleared before
984 	 * src is freed; but stats require that PageAnon be left as PageAnon.
985 	 */
986 	if (rc == MIGRATEPAGE_SUCCESS) {
987 		if (__PageMovable(&src->page)) {
988 			VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
989 
990 			/*
991 			 * We clear PG_movable under page_lock so any compactor
992 			 * cannot try to migrate this page.
993 			 */
994 			folio_clear_isolated(src);
995 		}
996 
997 		/*
998 		 * Anonymous and movable src->mapping will be cleared by
999 		 * free_pages_prepare so don't reset it here for keeping
1000 		 * the type to work PageAnon, for example.
1001 		 */
1002 		if (!folio_mapping_flags(src))
1003 			src->mapping = NULL;
1004 
1005 		if (likely(!folio_is_zone_device(dst)))
1006 			flush_dcache_folio(dst);
1007 	}
1008 out:
1009 	return rc;
1010 }
1011 
1012 static int __unmap_and_move(struct folio *src, struct folio *dst,
1013 				int force, enum migrate_mode mode)
1014 {
1015 	int rc = -EAGAIN;
1016 	bool page_was_mapped = false;
1017 	struct anon_vma *anon_vma = NULL;
1018 	bool is_lru = !__PageMovable(&src->page);
1019 
1020 	if (!folio_trylock(src)) {
1021 		if (!force || mode == MIGRATE_ASYNC)
1022 			goto out;
1023 
1024 		/*
1025 		 * It's not safe for direct compaction to call lock_page.
1026 		 * For example, during page readahead pages are added locked
1027 		 * to the LRU. Later, when the IO completes the pages are
1028 		 * marked uptodate and unlocked. However, the queueing
1029 		 * could be merging multiple pages for one bio (e.g.
1030 		 * mpage_readahead). If an allocation happens for the
1031 		 * second or third page, the process can end up locking
1032 		 * the same page twice and deadlocking. Rather than
1033 		 * trying to be clever about what pages can be locked,
1034 		 * avoid the use of lock_page for direct compaction
1035 		 * altogether.
1036 		 */
1037 		if (current->flags & PF_MEMALLOC)
1038 			goto out;
1039 
1040 		folio_lock(src);
1041 	}
1042 
1043 	if (folio_test_writeback(src)) {
1044 		/*
1045 		 * Only in the case of a full synchronous migration is it
1046 		 * necessary to wait for PageWriteback. In the async case,
1047 		 * the retry loop is too short and in the sync-light case,
1048 		 * the overhead of stalling is too much
1049 		 */
1050 		switch (mode) {
1051 		case MIGRATE_SYNC:
1052 		case MIGRATE_SYNC_NO_COPY:
1053 			break;
1054 		default:
1055 			rc = -EBUSY;
1056 			goto out_unlock;
1057 		}
1058 		if (!force)
1059 			goto out_unlock;
1060 		folio_wait_writeback(src);
1061 	}
1062 
1063 	/*
1064 	 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1065 	 * we cannot notice that anon_vma is freed while we migrate a page.
1066 	 * This get_anon_vma() delays freeing anon_vma pointer until the end
1067 	 * of migration. File cache pages are no problem because of page_lock()
1068 	 * File Caches may use write_page() or lock_page() in migration, then,
1069 	 * just care Anon page here.
1070 	 *
1071 	 * Only folio_get_anon_vma() understands the subtleties of
1072 	 * getting a hold on an anon_vma from outside one of its mms.
1073 	 * But if we cannot get anon_vma, then we won't need it anyway,
1074 	 * because that implies that the anon page is no longer mapped
1075 	 * (and cannot be remapped so long as we hold the page lock).
1076 	 */
1077 	if (folio_test_anon(src) && !folio_test_ksm(src))
1078 		anon_vma = folio_get_anon_vma(src);
1079 
1080 	/*
1081 	 * Block others from accessing the new page when we get around to
1082 	 * establishing additional references. We are usually the only one
1083 	 * holding a reference to dst at this point. We used to have a BUG
1084 	 * here if folio_trylock(dst) fails, but would like to allow for
1085 	 * cases where there might be a race with the previous use of dst.
1086 	 * This is much like races on refcount of oldpage: just don't BUG().
1087 	 */
1088 	if (unlikely(!folio_trylock(dst)))
1089 		goto out_unlock;
1090 
1091 	if (unlikely(!is_lru)) {
1092 		rc = move_to_new_folio(dst, src, mode);
1093 		goto out_unlock_both;
1094 	}
1095 
1096 	/*
1097 	 * Corner case handling:
1098 	 * 1. When a new swap-cache page is read into, it is added to the LRU
1099 	 * and treated as swapcache but it has no rmap yet.
1100 	 * Calling try_to_unmap() against a src->mapping==NULL page will
1101 	 * trigger a BUG.  So handle it here.
1102 	 * 2. An orphaned page (see truncate_cleanup_page) might have
1103 	 * fs-private metadata. The page can be picked up due to memory
1104 	 * offlining.  Everywhere else except page reclaim, the page is
1105 	 * invisible to the vm, so the page can not be migrated.  So try to
1106 	 * free the metadata, so the page can be freed.
1107 	 */
1108 	if (!src->mapping) {
1109 		if (folio_test_private(src)) {
1110 			try_to_free_buffers(src);
1111 			goto out_unlock_both;
1112 		}
1113 	} else if (folio_mapped(src)) {
1114 		/* Establish migration ptes */
1115 		VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1116 			       !folio_test_ksm(src) && !anon_vma, src);
1117 		try_to_migrate(src, 0);
1118 		page_was_mapped = true;
1119 	}
1120 
1121 	if (!folio_mapped(src))
1122 		rc = move_to_new_folio(dst, src, mode);
1123 
1124 	/*
1125 	 * When successful, push dst to LRU immediately: so that if it
1126 	 * turns out to be an mlocked page, remove_migration_ptes() will
1127 	 * automatically build up the correct dst->mlock_count for it.
1128 	 *
1129 	 * We would like to do something similar for the old page, when
1130 	 * unsuccessful, and other cases when a page has been temporarily
1131 	 * isolated from the unevictable LRU: but this case is the easiest.
1132 	 */
1133 	if (rc == MIGRATEPAGE_SUCCESS) {
1134 		folio_add_lru(dst);
1135 		if (page_was_mapped)
1136 			lru_add_drain();
1137 	}
1138 
1139 	if (page_was_mapped)
1140 		remove_migration_ptes(src,
1141 			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1142 
1143 out_unlock_both:
1144 	folio_unlock(dst);
1145 out_unlock:
1146 	/* Drop an anon_vma reference if we took one */
1147 	if (anon_vma)
1148 		put_anon_vma(anon_vma);
1149 	folio_unlock(src);
1150 out:
1151 	/*
1152 	 * If migration is successful, decrease refcount of dst,
1153 	 * which will not free the page because new page owner increased
1154 	 * refcounter.
1155 	 */
1156 	if (rc == MIGRATEPAGE_SUCCESS)
1157 		folio_put(dst);
1158 
1159 	return rc;
1160 }
1161 
1162 /*
1163  * Obtain the lock on folio, remove all ptes and migrate the folio
1164  * to the newly allocated folio in dst.
1165  */
1166 static int unmap_and_move(new_page_t get_new_page,
1167 				   free_page_t put_new_page,
1168 				   unsigned long private, struct folio *src,
1169 				   int force, enum migrate_mode mode,
1170 				   enum migrate_reason reason,
1171 				   struct list_head *ret)
1172 {
1173 	struct folio *dst;
1174 	int rc = MIGRATEPAGE_SUCCESS;
1175 	struct page *newpage = NULL;
1176 
1177 	if (!thp_migration_supported() && folio_test_transhuge(src))
1178 		return -ENOSYS;
1179 
1180 	if (folio_ref_count(src) == 1) {
1181 		/* Folio was freed from under us. So we are done. */
1182 		folio_clear_active(src);
1183 		folio_clear_unevictable(src);
1184 		/* free_pages_prepare() will clear PG_isolated. */
1185 		goto out;
1186 	}
1187 
1188 	newpage = get_new_page(&src->page, private);
1189 	if (!newpage)
1190 		return -ENOMEM;
1191 	dst = page_folio(newpage);
1192 
1193 	dst->private = NULL;
1194 	rc = __unmap_and_move(src, dst, force, mode);
1195 	if (rc == MIGRATEPAGE_SUCCESS)
1196 		set_page_owner_migrate_reason(&dst->page, reason);
1197 
1198 out:
1199 	if (rc != -EAGAIN) {
1200 		/*
1201 		 * A folio that has been migrated has all references
1202 		 * removed and will be freed. A folio that has not been
1203 		 * migrated will have kept its references and be restored.
1204 		 */
1205 		list_del(&src->lru);
1206 	}
1207 
1208 	/*
1209 	 * If migration is successful, releases reference grabbed during
1210 	 * isolation. Otherwise, restore the folio to right list unless
1211 	 * we want to retry.
1212 	 */
1213 	if (rc == MIGRATEPAGE_SUCCESS) {
1214 		/*
1215 		 * Compaction can migrate also non-LRU folios which are
1216 		 * not accounted to NR_ISOLATED_*. They can be recognized
1217 		 * as __folio_test_movable
1218 		 */
1219 		if (likely(!__folio_test_movable(src)))
1220 			mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1221 					folio_is_file_lru(src), -folio_nr_pages(src));
1222 
1223 		if (reason != MR_MEMORY_FAILURE)
1224 			/*
1225 			 * We release the folio in page_handle_poison.
1226 			 */
1227 			folio_put(src);
1228 	} else {
1229 		if (rc != -EAGAIN)
1230 			list_add_tail(&src->lru, ret);
1231 
1232 		if (put_new_page)
1233 			put_new_page(&dst->page, private);
1234 		else
1235 			folio_put(dst);
1236 	}
1237 
1238 	return rc;
1239 }
1240 
1241 /*
1242  * Counterpart of unmap_and_move_page() for hugepage migration.
1243  *
1244  * This function doesn't wait the completion of hugepage I/O
1245  * because there is no race between I/O and migration for hugepage.
1246  * Note that currently hugepage I/O occurs only in direct I/O
1247  * where no lock is held and PG_writeback is irrelevant,
1248  * and writeback status of all subpages are counted in the reference
1249  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1250  * under direct I/O, the reference of the head page is 512 and a bit more.)
1251  * This means that when we try to migrate hugepage whose subpages are
1252  * doing direct I/O, some references remain after try_to_unmap() and
1253  * hugepage migration fails without data corruption.
1254  *
1255  * There is also no race when direct I/O is issued on the page under migration,
1256  * because then pte is replaced with migration swap entry and direct I/O code
1257  * will wait in the page fault for migration to complete.
1258  */
1259 static int unmap_and_move_huge_page(new_page_t get_new_page,
1260 				free_page_t put_new_page, unsigned long private,
1261 				struct page *hpage, int force,
1262 				enum migrate_mode mode, int reason,
1263 				struct list_head *ret)
1264 {
1265 	struct folio *dst, *src = page_folio(hpage);
1266 	int rc = -EAGAIN;
1267 	int page_was_mapped = 0;
1268 	struct page *new_hpage;
1269 	struct anon_vma *anon_vma = NULL;
1270 	struct address_space *mapping = NULL;
1271 
1272 	/*
1273 	 * Migratability of hugepages depends on architectures and their size.
1274 	 * This check is necessary because some callers of hugepage migration
1275 	 * like soft offline and memory hotremove don't walk through page
1276 	 * tables or check whether the hugepage is pmd-based or not before
1277 	 * kicking migration.
1278 	 */
1279 	if (!hugepage_migration_supported(page_hstate(hpage)))
1280 		return -ENOSYS;
1281 
1282 	if (folio_ref_count(src) == 1) {
1283 		/* page was freed from under us. So we are done. */
1284 		putback_active_hugepage(hpage);
1285 		return MIGRATEPAGE_SUCCESS;
1286 	}
1287 
1288 	new_hpage = get_new_page(hpage, private);
1289 	if (!new_hpage)
1290 		return -ENOMEM;
1291 	dst = page_folio(new_hpage);
1292 
1293 	if (!folio_trylock(src)) {
1294 		if (!force)
1295 			goto out;
1296 		switch (mode) {
1297 		case MIGRATE_SYNC:
1298 		case MIGRATE_SYNC_NO_COPY:
1299 			break;
1300 		default:
1301 			goto out;
1302 		}
1303 		folio_lock(src);
1304 	}
1305 
1306 	/*
1307 	 * Check for pages which are in the process of being freed.  Without
1308 	 * folio_mapping() set, hugetlbfs specific move page routine will not
1309 	 * be called and we could leak usage counts for subpools.
1310 	 */
1311 	if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1312 		rc = -EBUSY;
1313 		goto out_unlock;
1314 	}
1315 
1316 	if (folio_test_anon(src))
1317 		anon_vma = folio_get_anon_vma(src);
1318 
1319 	if (unlikely(!folio_trylock(dst)))
1320 		goto put_anon;
1321 
1322 	if (folio_mapped(src)) {
1323 		enum ttu_flags ttu = 0;
1324 
1325 		if (!folio_test_anon(src)) {
1326 			/*
1327 			 * In shared mappings, try_to_unmap could potentially
1328 			 * call huge_pmd_unshare.  Because of this, take
1329 			 * semaphore in write mode here and set TTU_RMAP_LOCKED
1330 			 * to let lower levels know we have taken the lock.
1331 			 */
1332 			mapping = hugetlb_page_mapping_lock_write(hpage);
1333 			if (unlikely(!mapping))
1334 				goto unlock_put_anon;
1335 
1336 			ttu = TTU_RMAP_LOCKED;
1337 		}
1338 
1339 		try_to_migrate(src, ttu);
1340 		page_was_mapped = 1;
1341 
1342 		if (ttu & TTU_RMAP_LOCKED)
1343 			i_mmap_unlock_write(mapping);
1344 	}
1345 
1346 	if (!folio_mapped(src))
1347 		rc = move_to_new_folio(dst, src, mode);
1348 
1349 	if (page_was_mapped)
1350 		remove_migration_ptes(src,
1351 			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1352 
1353 unlock_put_anon:
1354 	folio_unlock(dst);
1355 
1356 put_anon:
1357 	if (anon_vma)
1358 		put_anon_vma(anon_vma);
1359 
1360 	if (rc == MIGRATEPAGE_SUCCESS) {
1361 		move_hugetlb_state(src, dst, reason);
1362 		put_new_page = NULL;
1363 	}
1364 
1365 out_unlock:
1366 	folio_unlock(src);
1367 out:
1368 	if (rc == MIGRATEPAGE_SUCCESS)
1369 		putback_active_hugepage(hpage);
1370 	else if (rc != -EAGAIN)
1371 		list_move_tail(&src->lru, ret);
1372 
1373 	/*
1374 	 * If migration was not successful and there's a freeing callback, use
1375 	 * it.  Otherwise, put_page() will drop the reference grabbed during
1376 	 * isolation.
1377 	 */
1378 	if (put_new_page)
1379 		put_new_page(new_hpage, private);
1380 	else
1381 		putback_active_hugepage(new_hpage);
1382 
1383 	return rc;
1384 }
1385 
1386 static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1387 {
1388 	int rc;
1389 
1390 	folio_lock(folio);
1391 	rc = split_folio_to_list(folio, split_folios);
1392 	folio_unlock(folio);
1393 	if (!rc)
1394 		list_move_tail(&folio->lru, split_folios);
1395 
1396 	return rc;
1397 }
1398 
1399 /*
1400  * migrate_pages - migrate the folios specified in a list, to the free folios
1401  *		   supplied as the target for the page migration
1402  *
1403  * @from:		The list of folios to be migrated.
1404  * @get_new_page:	The function used to allocate free folios to be used
1405  *			as the target of the folio migration.
1406  * @put_new_page:	The function used to free target folios if migration
1407  *			fails, or NULL if no special handling is necessary.
1408  * @private:		Private data to be passed on to get_new_page()
1409  * @mode:		The migration mode that specifies the constraints for
1410  *			folio migration, if any.
1411  * @reason:		The reason for folio migration.
1412  * @ret_succeeded:	Set to the number of folios migrated successfully if
1413  *			the caller passes a non-NULL pointer.
1414  *
1415  * The function returns after 10 attempts or if no folios are movable any more
1416  * because the list has become empty or no retryable folios exist any more.
1417  * It is caller's responsibility to call putback_movable_pages() to return folios
1418  * to the LRU or free list only if ret != 0.
1419  *
1420  * Returns the number of {normal folio, large folio, hugetlb} that were not
1421  * migrated, or an error code. The number of large folio splits will be
1422  * considered as the number of non-migrated large folio, no matter how many
1423  * split folios of the large folio are migrated successfully.
1424  */
1425 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1426 		free_page_t put_new_page, unsigned long private,
1427 		enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1428 {
1429 	int retry = 1;
1430 	int large_retry = 1;
1431 	int thp_retry = 1;
1432 	int nr_failed = 0;
1433 	int nr_failed_pages = 0;
1434 	int nr_retry_pages = 0;
1435 	int nr_succeeded = 0;
1436 	int nr_thp_succeeded = 0;
1437 	int nr_large_failed = 0;
1438 	int nr_thp_failed = 0;
1439 	int nr_thp_split = 0;
1440 	int pass = 0;
1441 	bool is_large = false;
1442 	bool is_thp = false;
1443 	struct folio *folio, *folio2;
1444 	int rc, nr_pages;
1445 	LIST_HEAD(ret_folios);
1446 	LIST_HEAD(split_folios);
1447 	bool nosplit = (reason == MR_NUMA_MISPLACED);
1448 	bool no_split_folio_counting = false;
1449 
1450 	trace_mm_migrate_pages_start(mode, reason);
1451 
1452 split_folio_migration:
1453 	for (pass = 0; pass < 10 && (retry || large_retry); pass++) {
1454 		retry = 0;
1455 		large_retry = 0;
1456 		thp_retry = 0;
1457 		nr_retry_pages = 0;
1458 
1459 		list_for_each_entry_safe(folio, folio2, from, lru) {
1460 			/*
1461 			 * Large folio statistics is based on the source large
1462 			 * folio. Capture required information that might get
1463 			 * lost during migration.
1464 			 */
1465 			is_large = folio_test_large(folio) && !folio_test_hugetlb(folio);
1466 			is_thp = is_large && folio_test_pmd_mappable(folio);
1467 			nr_pages = folio_nr_pages(folio);
1468 			cond_resched();
1469 
1470 			if (folio_test_hugetlb(folio))
1471 				rc = unmap_and_move_huge_page(get_new_page,
1472 						put_new_page, private,
1473 						&folio->page, pass > 2, mode,
1474 						reason,
1475 						&ret_folios);
1476 			else
1477 				rc = unmap_and_move(get_new_page, put_new_page,
1478 						private, folio, pass > 2, mode,
1479 						reason, &ret_folios);
1480 			/*
1481 			 * The rules are:
1482 			 *	Success: non hugetlb folio will be freed, hugetlb
1483 			 *		 folio will be put back
1484 			 *	-EAGAIN: stay on the from list
1485 			 *	-ENOMEM: stay on the from list
1486 			 *	-ENOSYS: stay on the from list
1487 			 *	Other errno: put on ret_folios list then splice to
1488 			 *		     from list
1489 			 */
1490 			switch(rc) {
1491 			/*
1492 			 * Large folio migration might be unsupported or
1493 			 * the allocation could've failed so we should retry
1494 			 * on the same folio with the large folio split
1495 			 * to normal folios.
1496 			 *
1497 			 * Split folios are put in split_folios, and
1498 			 * we will migrate them after the rest of the
1499 			 * list is processed.
1500 			 */
1501 			case -ENOSYS:
1502 				/* Large folio migration is unsupported */
1503 				if (is_large) {
1504 					nr_large_failed++;
1505 					nr_thp_failed += is_thp;
1506 					if (!try_split_folio(folio, &split_folios)) {
1507 						nr_thp_split += is_thp;
1508 						break;
1509 					}
1510 				/* Hugetlb migration is unsupported */
1511 				} else if (!no_split_folio_counting) {
1512 					nr_failed++;
1513 				}
1514 
1515 				nr_failed_pages += nr_pages;
1516 				list_move_tail(&folio->lru, &ret_folios);
1517 				break;
1518 			case -ENOMEM:
1519 				/*
1520 				 * When memory is low, don't bother to try to migrate
1521 				 * other folios, just exit.
1522 				 */
1523 				if (is_large) {
1524 					nr_large_failed++;
1525 					nr_thp_failed += is_thp;
1526 					/* Large folio NUMA faulting doesn't split to retry. */
1527 					if (!nosplit) {
1528 						int ret = try_split_folio(folio, &split_folios);
1529 
1530 						if (!ret) {
1531 							nr_thp_split += is_thp;
1532 							break;
1533 						} else if (reason == MR_LONGTERM_PIN &&
1534 							   ret == -EAGAIN) {
1535 							/*
1536 							 * Try again to split large folio to
1537 							 * mitigate the failure of longterm pinning.
1538 							 */
1539 							large_retry++;
1540 							thp_retry += is_thp;
1541 							nr_retry_pages += nr_pages;
1542 							break;
1543 						}
1544 					}
1545 				} else if (!no_split_folio_counting) {
1546 					nr_failed++;
1547 				}
1548 
1549 				nr_failed_pages += nr_pages + nr_retry_pages;
1550 				/*
1551 				 * There might be some split folios of fail-to-migrate large
1552 				 * folios left in split_folios list. Move them back to migration
1553 				 * list so that they could be put back to the right list by
1554 				 * the caller otherwise the folio refcnt will be leaked.
1555 				 */
1556 				list_splice_init(&split_folios, from);
1557 				/* nr_failed isn't updated for not used */
1558 				nr_large_failed += large_retry;
1559 				nr_thp_failed += thp_retry;
1560 				goto out;
1561 			case -EAGAIN:
1562 				if (is_large) {
1563 					large_retry++;
1564 					thp_retry += is_thp;
1565 				} else if (!no_split_folio_counting) {
1566 					retry++;
1567 				}
1568 				nr_retry_pages += nr_pages;
1569 				break;
1570 			case MIGRATEPAGE_SUCCESS:
1571 				nr_succeeded += nr_pages;
1572 				nr_thp_succeeded += is_thp;
1573 				break;
1574 			default:
1575 				/*
1576 				 * Permanent failure (-EBUSY, etc.):
1577 				 * unlike -EAGAIN case, the failed folio is
1578 				 * removed from migration folio list and not
1579 				 * retried in the next outer loop.
1580 				 */
1581 				if (is_large) {
1582 					nr_large_failed++;
1583 					nr_thp_failed += is_thp;
1584 				} else if (!no_split_folio_counting) {
1585 					nr_failed++;
1586 				}
1587 
1588 				nr_failed_pages += nr_pages;
1589 				break;
1590 			}
1591 		}
1592 	}
1593 	nr_failed += retry;
1594 	nr_large_failed += large_retry;
1595 	nr_thp_failed += thp_retry;
1596 	nr_failed_pages += nr_retry_pages;
1597 	/*
1598 	 * Try to migrate split folios of fail-to-migrate large folios, no
1599 	 * nr_failed counting in this round, since all split folios of a
1600 	 * large folio is counted as 1 failure in the first round.
1601 	 */
1602 	if (!list_empty(&split_folios)) {
1603 		/*
1604 		 * Move non-migrated folios (after 10 retries) to ret_folios
1605 		 * to avoid migrating them again.
1606 		 */
1607 		list_splice_init(from, &ret_folios);
1608 		list_splice_init(&split_folios, from);
1609 		no_split_folio_counting = true;
1610 		retry = 1;
1611 		goto split_folio_migration;
1612 	}
1613 
1614 	rc = nr_failed + nr_large_failed;
1615 out:
1616 	/*
1617 	 * Put the permanent failure folio back to migration list, they
1618 	 * will be put back to the right list by the caller.
1619 	 */
1620 	list_splice(&ret_folios, from);
1621 
1622 	/*
1623 	 * Return 0 in case all split folios of fail-to-migrate large folios
1624 	 * are migrated successfully.
1625 	 */
1626 	if (list_empty(from))
1627 		rc = 0;
1628 
1629 	count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1630 	count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1631 	count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1632 	count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1633 	count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1634 	trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1635 			       nr_thp_failed, nr_thp_split, mode, reason);
1636 
1637 	if (ret_succeeded)
1638 		*ret_succeeded = nr_succeeded;
1639 
1640 	return rc;
1641 }
1642 
1643 struct page *alloc_migration_target(struct page *page, unsigned long private)
1644 {
1645 	struct folio *folio = page_folio(page);
1646 	struct migration_target_control *mtc;
1647 	gfp_t gfp_mask;
1648 	unsigned int order = 0;
1649 	struct folio *new_folio = NULL;
1650 	int nid;
1651 	int zidx;
1652 
1653 	mtc = (struct migration_target_control *)private;
1654 	gfp_mask = mtc->gfp_mask;
1655 	nid = mtc->nid;
1656 	if (nid == NUMA_NO_NODE)
1657 		nid = folio_nid(folio);
1658 
1659 	if (folio_test_hugetlb(folio)) {
1660 		struct hstate *h = folio_hstate(folio);
1661 
1662 		gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1663 		return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1664 	}
1665 
1666 	if (folio_test_large(folio)) {
1667 		/*
1668 		 * clear __GFP_RECLAIM to make the migration callback
1669 		 * consistent with regular THP allocations.
1670 		 */
1671 		gfp_mask &= ~__GFP_RECLAIM;
1672 		gfp_mask |= GFP_TRANSHUGE;
1673 		order = folio_order(folio);
1674 	}
1675 	zidx = zone_idx(folio_zone(folio));
1676 	if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1677 		gfp_mask |= __GFP_HIGHMEM;
1678 
1679 	new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1680 
1681 	return &new_folio->page;
1682 }
1683 
1684 #ifdef CONFIG_NUMA
1685 
1686 static int store_status(int __user *status, int start, int value, int nr)
1687 {
1688 	while (nr-- > 0) {
1689 		if (put_user(value, status + start))
1690 			return -EFAULT;
1691 		start++;
1692 	}
1693 
1694 	return 0;
1695 }
1696 
1697 static int do_move_pages_to_node(struct mm_struct *mm,
1698 		struct list_head *pagelist, int node)
1699 {
1700 	int err;
1701 	struct migration_target_control mtc = {
1702 		.nid = node,
1703 		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1704 	};
1705 
1706 	err = migrate_pages(pagelist, alloc_migration_target, NULL,
1707 		(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1708 	if (err)
1709 		putback_movable_pages(pagelist);
1710 	return err;
1711 }
1712 
1713 /*
1714  * Resolves the given address to a struct page, isolates it from the LRU and
1715  * puts it to the given pagelist.
1716  * Returns:
1717  *     errno - if the page cannot be found/isolated
1718  *     0 - when it doesn't have to be migrated because it is already on the
1719  *         target node
1720  *     1 - when it has been queued
1721  */
1722 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1723 		int node, struct list_head *pagelist, bool migrate_all)
1724 {
1725 	struct vm_area_struct *vma;
1726 	struct page *page;
1727 	int err;
1728 
1729 	mmap_read_lock(mm);
1730 	err = -EFAULT;
1731 	vma = vma_lookup(mm, addr);
1732 	if (!vma || !vma_migratable(vma))
1733 		goto out;
1734 
1735 	/* FOLL_DUMP to ignore special (like zero) pages */
1736 	page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1737 
1738 	err = PTR_ERR(page);
1739 	if (IS_ERR(page))
1740 		goto out;
1741 
1742 	err = -ENOENT;
1743 	if (!page)
1744 		goto out;
1745 
1746 	if (is_zone_device_page(page))
1747 		goto out_putpage;
1748 
1749 	err = 0;
1750 	if (page_to_nid(page) == node)
1751 		goto out_putpage;
1752 
1753 	err = -EACCES;
1754 	if (page_mapcount(page) > 1 && !migrate_all)
1755 		goto out_putpage;
1756 
1757 	if (PageHuge(page)) {
1758 		if (PageHead(page)) {
1759 			err = isolate_hugetlb(page, pagelist);
1760 			if (!err)
1761 				err = 1;
1762 		}
1763 	} else {
1764 		struct page *head;
1765 
1766 		head = compound_head(page);
1767 		err = isolate_lru_page(head);
1768 		if (err)
1769 			goto out_putpage;
1770 
1771 		err = 1;
1772 		list_add_tail(&head->lru, pagelist);
1773 		mod_node_page_state(page_pgdat(head),
1774 			NR_ISOLATED_ANON + page_is_file_lru(head),
1775 			thp_nr_pages(head));
1776 	}
1777 out_putpage:
1778 	/*
1779 	 * Either remove the duplicate refcount from
1780 	 * isolate_lru_page() or drop the page ref if it was
1781 	 * not isolated.
1782 	 */
1783 	put_page(page);
1784 out:
1785 	mmap_read_unlock(mm);
1786 	return err;
1787 }
1788 
1789 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1790 		struct list_head *pagelist, int __user *status,
1791 		int start, int i, unsigned long nr_pages)
1792 {
1793 	int err;
1794 
1795 	if (list_empty(pagelist))
1796 		return 0;
1797 
1798 	err = do_move_pages_to_node(mm, pagelist, node);
1799 	if (err) {
1800 		/*
1801 		 * Positive err means the number of failed
1802 		 * pages to migrate.  Since we are going to
1803 		 * abort and return the number of non-migrated
1804 		 * pages, so need to include the rest of the
1805 		 * nr_pages that have not been attempted as
1806 		 * well.
1807 		 */
1808 		if (err > 0)
1809 			err += nr_pages - i;
1810 		return err;
1811 	}
1812 	return store_status(status, start, node, i - start);
1813 }
1814 
1815 /*
1816  * Migrate an array of page address onto an array of nodes and fill
1817  * the corresponding array of status.
1818  */
1819 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1820 			 unsigned long nr_pages,
1821 			 const void __user * __user *pages,
1822 			 const int __user *nodes,
1823 			 int __user *status, int flags)
1824 {
1825 	int current_node = NUMA_NO_NODE;
1826 	LIST_HEAD(pagelist);
1827 	int start, i;
1828 	int err = 0, err1;
1829 
1830 	lru_cache_disable();
1831 
1832 	for (i = start = 0; i < nr_pages; i++) {
1833 		const void __user *p;
1834 		unsigned long addr;
1835 		int node;
1836 
1837 		err = -EFAULT;
1838 		if (get_user(p, pages + i))
1839 			goto out_flush;
1840 		if (get_user(node, nodes + i))
1841 			goto out_flush;
1842 		addr = (unsigned long)untagged_addr(p);
1843 
1844 		err = -ENODEV;
1845 		if (node < 0 || node >= MAX_NUMNODES)
1846 			goto out_flush;
1847 		if (!node_state(node, N_MEMORY))
1848 			goto out_flush;
1849 
1850 		err = -EACCES;
1851 		if (!node_isset(node, task_nodes))
1852 			goto out_flush;
1853 
1854 		if (current_node == NUMA_NO_NODE) {
1855 			current_node = node;
1856 			start = i;
1857 		} else if (node != current_node) {
1858 			err = move_pages_and_store_status(mm, current_node,
1859 					&pagelist, status, start, i, nr_pages);
1860 			if (err)
1861 				goto out;
1862 			start = i;
1863 			current_node = node;
1864 		}
1865 
1866 		/*
1867 		 * Errors in the page lookup or isolation are not fatal and we simply
1868 		 * report them via status
1869 		 */
1870 		err = add_page_for_migration(mm, addr, current_node,
1871 				&pagelist, flags & MPOL_MF_MOVE_ALL);
1872 
1873 		if (err > 0) {
1874 			/* The page is successfully queued for migration */
1875 			continue;
1876 		}
1877 
1878 		/*
1879 		 * The move_pages() man page does not have an -EEXIST choice, so
1880 		 * use -EFAULT instead.
1881 		 */
1882 		if (err == -EEXIST)
1883 			err = -EFAULT;
1884 
1885 		/*
1886 		 * If the page is already on the target node (!err), store the
1887 		 * node, otherwise, store the err.
1888 		 */
1889 		err = store_status(status, i, err ? : current_node, 1);
1890 		if (err)
1891 			goto out_flush;
1892 
1893 		err = move_pages_and_store_status(mm, current_node, &pagelist,
1894 				status, start, i, nr_pages);
1895 		if (err) {
1896 			/* We have accounted for page i */
1897 			if (err > 0)
1898 				err--;
1899 			goto out;
1900 		}
1901 		current_node = NUMA_NO_NODE;
1902 	}
1903 out_flush:
1904 	/* Make sure we do not overwrite the existing error */
1905 	err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1906 				status, start, i, nr_pages);
1907 	if (err >= 0)
1908 		err = err1;
1909 out:
1910 	lru_cache_enable();
1911 	return err;
1912 }
1913 
1914 /*
1915  * Determine the nodes of an array of pages and store it in an array of status.
1916  */
1917 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1918 				const void __user **pages, int *status)
1919 {
1920 	unsigned long i;
1921 
1922 	mmap_read_lock(mm);
1923 
1924 	for (i = 0; i < nr_pages; i++) {
1925 		unsigned long addr = (unsigned long)(*pages);
1926 		struct vm_area_struct *vma;
1927 		struct page *page;
1928 		int err = -EFAULT;
1929 
1930 		vma = vma_lookup(mm, addr);
1931 		if (!vma)
1932 			goto set_status;
1933 
1934 		/* FOLL_DUMP to ignore special (like zero) pages */
1935 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1936 
1937 		err = PTR_ERR(page);
1938 		if (IS_ERR(page))
1939 			goto set_status;
1940 
1941 		err = -ENOENT;
1942 		if (!page)
1943 			goto set_status;
1944 
1945 		if (!is_zone_device_page(page))
1946 			err = page_to_nid(page);
1947 
1948 		put_page(page);
1949 set_status:
1950 		*status = err;
1951 
1952 		pages++;
1953 		status++;
1954 	}
1955 
1956 	mmap_read_unlock(mm);
1957 }
1958 
1959 static int get_compat_pages_array(const void __user *chunk_pages[],
1960 				  const void __user * __user *pages,
1961 				  unsigned long chunk_nr)
1962 {
1963 	compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1964 	compat_uptr_t p;
1965 	int i;
1966 
1967 	for (i = 0; i < chunk_nr; i++) {
1968 		if (get_user(p, pages32 + i))
1969 			return -EFAULT;
1970 		chunk_pages[i] = compat_ptr(p);
1971 	}
1972 
1973 	return 0;
1974 }
1975 
1976 /*
1977  * Determine the nodes of a user array of pages and store it in
1978  * a user array of status.
1979  */
1980 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1981 			 const void __user * __user *pages,
1982 			 int __user *status)
1983 {
1984 #define DO_PAGES_STAT_CHUNK_NR 16UL
1985 	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1986 	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1987 
1988 	while (nr_pages) {
1989 		unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1990 
1991 		if (in_compat_syscall()) {
1992 			if (get_compat_pages_array(chunk_pages, pages,
1993 						   chunk_nr))
1994 				break;
1995 		} else {
1996 			if (copy_from_user(chunk_pages, pages,
1997 				      chunk_nr * sizeof(*chunk_pages)))
1998 				break;
1999 		}
2000 
2001 		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2002 
2003 		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2004 			break;
2005 
2006 		pages += chunk_nr;
2007 		status += chunk_nr;
2008 		nr_pages -= chunk_nr;
2009 	}
2010 	return nr_pages ? -EFAULT : 0;
2011 }
2012 
2013 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2014 {
2015 	struct task_struct *task;
2016 	struct mm_struct *mm;
2017 
2018 	/*
2019 	 * There is no need to check if current process has the right to modify
2020 	 * the specified process when they are same.
2021 	 */
2022 	if (!pid) {
2023 		mmget(current->mm);
2024 		*mem_nodes = cpuset_mems_allowed(current);
2025 		return current->mm;
2026 	}
2027 
2028 	/* Find the mm_struct */
2029 	rcu_read_lock();
2030 	task = find_task_by_vpid(pid);
2031 	if (!task) {
2032 		rcu_read_unlock();
2033 		return ERR_PTR(-ESRCH);
2034 	}
2035 	get_task_struct(task);
2036 
2037 	/*
2038 	 * Check if this process has the right to modify the specified
2039 	 * process. Use the regular "ptrace_may_access()" checks.
2040 	 */
2041 	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2042 		rcu_read_unlock();
2043 		mm = ERR_PTR(-EPERM);
2044 		goto out;
2045 	}
2046 	rcu_read_unlock();
2047 
2048 	mm = ERR_PTR(security_task_movememory(task));
2049 	if (IS_ERR(mm))
2050 		goto out;
2051 	*mem_nodes = cpuset_mems_allowed(task);
2052 	mm = get_task_mm(task);
2053 out:
2054 	put_task_struct(task);
2055 	if (!mm)
2056 		mm = ERR_PTR(-EINVAL);
2057 	return mm;
2058 }
2059 
2060 /*
2061  * Move a list of pages in the address space of the currently executing
2062  * process.
2063  */
2064 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2065 			     const void __user * __user *pages,
2066 			     const int __user *nodes,
2067 			     int __user *status, int flags)
2068 {
2069 	struct mm_struct *mm;
2070 	int err;
2071 	nodemask_t task_nodes;
2072 
2073 	/* Check flags */
2074 	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2075 		return -EINVAL;
2076 
2077 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2078 		return -EPERM;
2079 
2080 	mm = find_mm_struct(pid, &task_nodes);
2081 	if (IS_ERR(mm))
2082 		return PTR_ERR(mm);
2083 
2084 	if (nodes)
2085 		err = do_pages_move(mm, task_nodes, nr_pages, pages,
2086 				    nodes, status, flags);
2087 	else
2088 		err = do_pages_stat(mm, nr_pages, pages, status);
2089 
2090 	mmput(mm);
2091 	return err;
2092 }
2093 
2094 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2095 		const void __user * __user *, pages,
2096 		const int __user *, nodes,
2097 		int __user *, status, int, flags)
2098 {
2099 	return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2100 }
2101 
2102 #ifdef CONFIG_NUMA_BALANCING
2103 /*
2104  * Returns true if this is a safe migration target node for misplaced NUMA
2105  * pages. Currently it only checks the watermarks which is crude.
2106  */
2107 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2108 				   unsigned long nr_migrate_pages)
2109 {
2110 	int z;
2111 
2112 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2113 		struct zone *zone = pgdat->node_zones + z;
2114 
2115 		if (!managed_zone(zone))
2116 			continue;
2117 
2118 		/* Avoid waking kswapd by allocating pages_to_migrate pages. */
2119 		if (!zone_watermark_ok(zone, 0,
2120 				       high_wmark_pages(zone) +
2121 				       nr_migrate_pages,
2122 				       ZONE_MOVABLE, 0))
2123 			continue;
2124 		return true;
2125 	}
2126 	return false;
2127 }
2128 
2129 static struct page *alloc_misplaced_dst_page(struct page *page,
2130 					   unsigned long data)
2131 {
2132 	int nid = (int) data;
2133 	int order = compound_order(page);
2134 	gfp_t gfp = __GFP_THISNODE;
2135 	struct folio *new;
2136 
2137 	if (order > 0)
2138 		gfp |= GFP_TRANSHUGE_LIGHT;
2139 	else {
2140 		gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2141 			__GFP_NOWARN;
2142 		gfp &= ~__GFP_RECLAIM;
2143 	}
2144 	new = __folio_alloc_node(gfp, order, nid);
2145 
2146 	return &new->page;
2147 }
2148 
2149 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2150 {
2151 	int nr_pages = thp_nr_pages(page);
2152 	int order = compound_order(page);
2153 
2154 	VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2155 
2156 	/* Do not migrate THP mapped by multiple processes */
2157 	if (PageTransHuge(page) && total_mapcount(page) > 1)
2158 		return 0;
2159 
2160 	/* Avoid migrating to a node that is nearly full */
2161 	if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2162 		int z;
2163 
2164 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2165 			return 0;
2166 		for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2167 			if (managed_zone(pgdat->node_zones + z))
2168 				break;
2169 		}
2170 		wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2171 		return 0;
2172 	}
2173 
2174 	if (isolate_lru_page(page))
2175 		return 0;
2176 
2177 	mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2178 			    nr_pages);
2179 
2180 	/*
2181 	 * Isolating the page has taken another reference, so the
2182 	 * caller's reference can be safely dropped without the page
2183 	 * disappearing underneath us during migration.
2184 	 */
2185 	put_page(page);
2186 	return 1;
2187 }
2188 
2189 /*
2190  * Attempt to migrate a misplaced page to the specified destination
2191  * node. Caller is expected to have an elevated reference count on
2192  * the page that will be dropped by this function before returning.
2193  */
2194 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2195 			   int node)
2196 {
2197 	pg_data_t *pgdat = NODE_DATA(node);
2198 	int isolated;
2199 	int nr_remaining;
2200 	unsigned int nr_succeeded;
2201 	LIST_HEAD(migratepages);
2202 	int nr_pages = thp_nr_pages(page);
2203 
2204 	/*
2205 	 * Don't migrate file pages that are mapped in multiple processes
2206 	 * with execute permissions as they are probably shared libraries.
2207 	 */
2208 	if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2209 	    (vma->vm_flags & VM_EXEC))
2210 		goto out;
2211 
2212 	/*
2213 	 * Also do not migrate dirty pages as not all filesystems can move
2214 	 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2215 	 */
2216 	if (page_is_file_lru(page) && PageDirty(page))
2217 		goto out;
2218 
2219 	isolated = numamigrate_isolate_page(pgdat, page);
2220 	if (!isolated)
2221 		goto out;
2222 
2223 	list_add(&page->lru, &migratepages);
2224 	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2225 				     NULL, node, MIGRATE_ASYNC,
2226 				     MR_NUMA_MISPLACED, &nr_succeeded);
2227 	if (nr_remaining) {
2228 		if (!list_empty(&migratepages)) {
2229 			list_del(&page->lru);
2230 			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2231 					page_is_file_lru(page), -nr_pages);
2232 			putback_lru_page(page);
2233 		}
2234 		isolated = 0;
2235 	}
2236 	if (nr_succeeded) {
2237 		count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2238 		if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2239 			mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2240 					    nr_succeeded);
2241 	}
2242 	BUG_ON(!list_empty(&migratepages));
2243 	return isolated;
2244 
2245 out:
2246 	put_page(page);
2247 	return 0;
2248 }
2249 #endif /* CONFIG_NUMA_BALANCING */
2250 #endif /* CONFIG_NUMA */
2251