Loading...
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/pagewalk.h>
42#include <linux/pfn_t.h>
43#include <linux/memremap.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/balloon_compaction.h>
46#include <linux/mmu_notifier.h>
47#include <linux/page_idle.h>
48#include <linux/page_owner.h>
49#include <linux/sched/mm.h>
50#include <linux/ptrace.h>
51
52#include <asm/tlbflush.h>
53
54#define CREATE_TRACE_POINTS
55#include <trace/events/migrate.h>
56
57#include "internal.h"
58
59/*
60 * migrate_prep() needs to be called before we start compiling a list of pages
61 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
62 * undesirable, use migrate_prep_local()
63 */
64int migrate_prep(void)
65{
66 /*
67 * Clear the LRU lists so pages can be isolated.
68 * Note that pages may be moved off the LRU after we have
69 * drained them. Those pages will fail to migrate like other
70 * pages that may be busy.
71 */
72 lru_add_drain_all();
73
74 return 0;
75}
76
77/* Do the necessary work of migrate_prep but not if it involves other CPUs */
78int migrate_prep_local(void)
79{
80 lru_add_drain();
81
82 return 0;
83}
84
85int isolate_movable_page(struct page *page, isolate_mode_t mode)
86{
87 struct address_space *mapping;
88
89 /*
90 * Avoid burning cycles with pages that are yet under __free_pages(),
91 * or just got freed under us.
92 *
93 * In case we 'win' a race for a movable page being freed under us and
94 * raise its refcount preventing __free_pages() from doing its job
95 * the put_page() at the end of this block will take care of
96 * release this page, thus avoiding a nasty leakage.
97 */
98 if (unlikely(!get_page_unless_zero(page)))
99 goto out;
100
101 /*
102 * Check PageMovable before holding a PG_lock because page's owner
103 * assumes anybody doesn't touch PG_lock of newly allocated page
104 * so unconditionally grabbing the lock ruins page's owner side.
105 */
106 if (unlikely(!__PageMovable(page)))
107 goto out_putpage;
108 /*
109 * As movable pages are not isolated from LRU lists, concurrent
110 * compaction threads can race against page migration functions
111 * as well as race against the releasing a page.
112 *
113 * In order to avoid having an already isolated movable page
114 * being (wrongly) re-isolated while it is under migration,
115 * or to avoid attempting to isolate pages being released,
116 * lets be sure we have the page lock
117 * before proceeding with the movable page isolation steps.
118 */
119 if (unlikely(!trylock_page(page)))
120 goto out_putpage;
121
122 if (!PageMovable(page) || PageIsolated(page))
123 goto out_no_isolated;
124
125 mapping = page_mapping(page);
126 VM_BUG_ON_PAGE(!mapping, page);
127
128 if (!mapping->a_ops->isolate_page(page, mode))
129 goto out_no_isolated;
130
131 /* Driver shouldn't use PG_isolated bit of page->flags */
132 WARN_ON_ONCE(PageIsolated(page));
133 __SetPageIsolated(page);
134 unlock_page(page);
135
136 return 0;
137
138out_no_isolated:
139 unlock_page(page);
140out_putpage:
141 put_page(page);
142out:
143 return -EBUSY;
144}
145
146/* It should be called on page which is PG_movable */
147void putback_movable_page(struct page *page)
148{
149 struct address_space *mapping;
150
151 VM_BUG_ON_PAGE(!PageLocked(page), page);
152 VM_BUG_ON_PAGE(!PageMovable(page), page);
153 VM_BUG_ON_PAGE(!PageIsolated(page), page);
154
155 mapping = page_mapping(page);
156 mapping->a_ops->putback_page(page);
157 __ClearPageIsolated(page);
158}
159
160/*
161 * Put previously isolated pages back onto the appropriate lists
162 * from where they were once taken off for compaction/migration.
163 *
164 * This function shall be used whenever the isolated pageset has been
165 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
166 * and isolate_huge_page().
167 */
168void putback_movable_pages(struct list_head *l)
169{
170 struct page *page;
171 struct page *page2;
172
173 list_for_each_entry_safe(page, page2, l, lru) {
174 if (unlikely(PageHuge(page))) {
175 putback_active_hugepage(page);
176 continue;
177 }
178 list_del(&page->lru);
179 /*
180 * We isolated non-lru movable page so here we can use
181 * __PageMovable because LRU page's mapping cannot have
182 * PAGE_MAPPING_MOVABLE.
183 */
184 if (unlikely(__PageMovable(page))) {
185 VM_BUG_ON_PAGE(!PageIsolated(page), page);
186 lock_page(page);
187 if (PageMovable(page))
188 putback_movable_page(page);
189 else
190 __ClearPageIsolated(page);
191 unlock_page(page);
192 put_page(page);
193 } else {
194 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
195 page_is_file_cache(page), -hpage_nr_pages(page));
196 putback_lru_page(page);
197 }
198 }
199}
200
201/*
202 * Restore a potential migration pte to a working pte entry
203 */
204static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
205 unsigned long addr, void *old)
206{
207 struct page_vma_mapped_walk pvmw = {
208 .page = old,
209 .vma = vma,
210 .address = addr,
211 .flags = PVMW_SYNC | PVMW_MIGRATION,
212 };
213 struct page *new;
214 pte_t pte;
215 swp_entry_t entry;
216
217 VM_BUG_ON_PAGE(PageTail(page), page);
218 while (page_vma_mapped_walk(&pvmw)) {
219 if (PageKsm(page))
220 new = page;
221 else
222 new = page - pvmw.page->index +
223 linear_page_index(vma, pvmw.address);
224
225#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
226 /* PMD-mapped THP migration entry */
227 if (!pvmw.pte) {
228 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
229 remove_migration_pmd(&pvmw, new);
230 continue;
231 }
232#endif
233
234 get_page(new);
235 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
236 if (pte_swp_soft_dirty(*pvmw.pte))
237 pte = pte_mksoft_dirty(pte);
238
239 /*
240 * Recheck VMA as permissions can change since migration started
241 */
242 entry = pte_to_swp_entry(*pvmw.pte);
243 if (is_write_migration_entry(entry))
244 pte = maybe_mkwrite(pte, vma);
245
246 if (unlikely(is_zone_device_page(new))) {
247 if (is_device_private_page(new)) {
248 entry = make_device_private_entry(new, pte_write(pte));
249 pte = swp_entry_to_pte(entry);
250 }
251 }
252
253#ifdef CONFIG_HUGETLB_PAGE
254 if (PageHuge(new)) {
255 pte = pte_mkhuge(pte);
256 pte = arch_make_huge_pte(pte, vma, new, 0);
257 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
258 if (PageAnon(new))
259 hugepage_add_anon_rmap(new, vma, pvmw.address);
260 else
261 page_dup_rmap(new, true);
262 } else
263#endif
264 {
265 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
266
267 if (PageAnon(new))
268 page_add_anon_rmap(new, vma, pvmw.address, false);
269 else
270 page_add_file_rmap(new, false);
271 }
272 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
273 mlock_vma_page(new);
274
275 if (PageTransHuge(page) && PageMlocked(page))
276 clear_page_mlock(page);
277
278 /* No need to invalidate - it was non-present before */
279 update_mmu_cache(vma, pvmw.address, pvmw.pte);
280 }
281
282 return true;
283}
284
285/*
286 * Get rid of all migration entries and replace them by
287 * references to the indicated page.
288 */
289void remove_migration_ptes(struct page *old, struct page *new, bool locked)
290{
291 struct rmap_walk_control rwc = {
292 .rmap_one = remove_migration_pte,
293 .arg = old,
294 };
295
296 if (locked)
297 rmap_walk_locked(new, &rwc);
298 else
299 rmap_walk(new, &rwc);
300}
301
302/*
303 * Something used the pte of a page under migration. We need to
304 * get to the page and wait until migration is finished.
305 * When we return from this function the fault will be retried.
306 */
307void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
308 spinlock_t *ptl)
309{
310 pte_t pte;
311 swp_entry_t entry;
312 struct page *page;
313
314 spin_lock(ptl);
315 pte = *ptep;
316 if (!is_swap_pte(pte))
317 goto out;
318
319 entry = pte_to_swp_entry(pte);
320 if (!is_migration_entry(entry))
321 goto out;
322
323 page = migration_entry_to_page(entry);
324
325 /*
326 * Once page cache replacement of page migration started, page_count
327 * is zero; but we must not call put_and_wait_on_page_locked() without
328 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
329 */
330 if (!get_page_unless_zero(page))
331 goto out;
332 pte_unmap_unlock(ptep, ptl);
333 put_and_wait_on_page_locked(page);
334 return;
335out:
336 pte_unmap_unlock(ptep, ptl);
337}
338
339void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
340 unsigned long address)
341{
342 spinlock_t *ptl = pte_lockptr(mm, pmd);
343 pte_t *ptep = pte_offset_map(pmd, address);
344 __migration_entry_wait(mm, ptep, ptl);
345}
346
347void migration_entry_wait_huge(struct vm_area_struct *vma,
348 struct mm_struct *mm, pte_t *pte)
349{
350 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
351 __migration_entry_wait(mm, pte, ptl);
352}
353
354#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
355void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
356{
357 spinlock_t *ptl;
358 struct page *page;
359
360 ptl = pmd_lock(mm, pmd);
361 if (!is_pmd_migration_entry(*pmd))
362 goto unlock;
363 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
364 if (!get_page_unless_zero(page))
365 goto unlock;
366 spin_unlock(ptl);
367 put_and_wait_on_page_locked(page);
368 return;
369unlock:
370 spin_unlock(ptl);
371}
372#endif
373
374static int expected_page_refs(struct address_space *mapping, struct page *page)
375{
376 int expected_count = 1;
377
378 /*
379 * Device public or private pages have an extra refcount as they are
380 * ZONE_DEVICE pages.
381 */
382 expected_count += is_device_private_page(page);
383 if (mapping)
384 expected_count += hpage_nr_pages(page) + page_has_private(page);
385
386 return expected_count;
387}
388
389/*
390 * Replace the page in the mapping.
391 *
392 * The number of remaining references must be:
393 * 1 for anonymous pages without a mapping
394 * 2 for pages with a mapping
395 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
396 */
397int migrate_page_move_mapping(struct address_space *mapping,
398 struct page *newpage, struct page *page, int extra_count)
399{
400 XA_STATE(xas, &mapping->i_pages, page_index(page));
401 struct zone *oldzone, *newzone;
402 int dirty;
403 int expected_count = expected_page_refs(mapping, page) + extra_count;
404
405 if (!mapping) {
406 /* Anonymous page without mapping */
407 if (page_count(page) != expected_count)
408 return -EAGAIN;
409
410 /* No turning back from here */
411 newpage->index = page->index;
412 newpage->mapping = page->mapping;
413 if (PageSwapBacked(page))
414 __SetPageSwapBacked(newpage);
415
416 return MIGRATEPAGE_SUCCESS;
417 }
418
419 oldzone = page_zone(page);
420 newzone = page_zone(newpage);
421
422 xas_lock_irq(&xas);
423 if (page_count(page) != expected_count || xas_load(&xas) != page) {
424 xas_unlock_irq(&xas);
425 return -EAGAIN;
426 }
427
428 if (!page_ref_freeze(page, expected_count)) {
429 xas_unlock_irq(&xas);
430 return -EAGAIN;
431 }
432
433 /*
434 * Now we know that no one else is looking at the page:
435 * no turning back from here.
436 */
437 newpage->index = page->index;
438 newpage->mapping = page->mapping;
439 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
440 if (PageSwapBacked(page)) {
441 __SetPageSwapBacked(newpage);
442 if (PageSwapCache(page)) {
443 SetPageSwapCache(newpage);
444 set_page_private(newpage, page_private(page));
445 }
446 } else {
447 VM_BUG_ON_PAGE(PageSwapCache(page), page);
448 }
449
450 /* Move dirty while page refs frozen and newpage not yet exposed */
451 dirty = PageDirty(page);
452 if (dirty) {
453 ClearPageDirty(page);
454 SetPageDirty(newpage);
455 }
456
457 xas_store(&xas, newpage);
458 if (PageTransHuge(page)) {
459 int i;
460
461 for (i = 1; i < HPAGE_PMD_NR; i++) {
462 xas_next(&xas);
463 xas_store(&xas, newpage);
464 }
465 }
466
467 /*
468 * Drop cache reference from old page by unfreezing
469 * to one less reference.
470 * We know this isn't the last reference.
471 */
472 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
473
474 xas_unlock(&xas);
475 /* Leave irq disabled to prevent preemption while updating stats */
476
477 /*
478 * If moved to a different zone then also account
479 * the page for that zone. Other VM counters will be
480 * taken care of when we establish references to the
481 * new page and drop references to the old page.
482 *
483 * Note that anonymous pages are accounted for
484 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
485 * are mapped to swap space.
486 */
487 if (newzone != oldzone) {
488 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
489 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
490 if (PageSwapBacked(page) && !PageSwapCache(page)) {
491 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
492 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
493 }
494 if (dirty && mapping_cap_account_dirty(mapping)) {
495 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
496 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
497 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
498 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
499 }
500 }
501 local_irq_enable();
502
503 return MIGRATEPAGE_SUCCESS;
504}
505EXPORT_SYMBOL(migrate_page_move_mapping);
506
507/*
508 * The expected number of remaining references is the same as that
509 * of migrate_page_move_mapping().
510 */
511int migrate_huge_page_move_mapping(struct address_space *mapping,
512 struct page *newpage, struct page *page)
513{
514 XA_STATE(xas, &mapping->i_pages, page_index(page));
515 int expected_count;
516
517 xas_lock_irq(&xas);
518 expected_count = 2 + page_has_private(page);
519 if (page_count(page) != expected_count || xas_load(&xas) != page) {
520 xas_unlock_irq(&xas);
521 return -EAGAIN;
522 }
523
524 if (!page_ref_freeze(page, expected_count)) {
525 xas_unlock_irq(&xas);
526 return -EAGAIN;
527 }
528
529 newpage->index = page->index;
530 newpage->mapping = page->mapping;
531
532 get_page(newpage);
533
534 xas_store(&xas, newpage);
535
536 page_ref_unfreeze(page, expected_count - 1);
537
538 xas_unlock_irq(&xas);
539
540 return MIGRATEPAGE_SUCCESS;
541}
542
543/*
544 * Gigantic pages are so large that we do not guarantee that page++ pointer
545 * arithmetic will work across the entire page. We need something more
546 * specialized.
547 */
548static void __copy_gigantic_page(struct page *dst, struct page *src,
549 int nr_pages)
550{
551 int i;
552 struct page *dst_base = dst;
553 struct page *src_base = src;
554
555 for (i = 0; i < nr_pages; ) {
556 cond_resched();
557 copy_highpage(dst, src);
558
559 i++;
560 dst = mem_map_next(dst, dst_base, i);
561 src = mem_map_next(src, src_base, i);
562 }
563}
564
565static void copy_huge_page(struct page *dst, struct page *src)
566{
567 int i;
568 int nr_pages;
569
570 if (PageHuge(src)) {
571 /* hugetlbfs page */
572 struct hstate *h = page_hstate(src);
573 nr_pages = pages_per_huge_page(h);
574
575 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
576 __copy_gigantic_page(dst, src, nr_pages);
577 return;
578 }
579 } else {
580 /* thp page */
581 BUG_ON(!PageTransHuge(src));
582 nr_pages = hpage_nr_pages(src);
583 }
584
585 for (i = 0; i < nr_pages; i++) {
586 cond_resched();
587 copy_highpage(dst + i, src + i);
588 }
589}
590
591/*
592 * Copy the page to its new location
593 */
594void migrate_page_states(struct page *newpage, struct page *page)
595{
596 int cpupid;
597
598 if (PageError(page))
599 SetPageError(newpage);
600 if (PageReferenced(page))
601 SetPageReferenced(newpage);
602 if (PageUptodate(page))
603 SetPageUptodate(newpage);
604 if (TestClearPageActive(page)) {
605 VM_BUG_ON_PAGE(PageUnevictable(page), page);
606 SetPageActive(newpage);
607 } else if (TestClearPageUnevictable(page))
608 SetPageUnevictable(newpage);
609 if (PageWorkingset(page))
610 SetPageWorkingset(newpage);
611 if (PageChecked(page))
612 SetPageChecked(newpage);
613 if (PageMappedToDisk(page))
614 SetPageMappedToDisk(newpage);
615
616 /* Move dirty on pages not done by migrate_page_move_mapping() */
617 if (PageDirty(page))
618 SetPageDirty(newpage);
619
620 if (page_is_young(page))
621 set_page_young(newpage);
622 if (page_is_idle(page))
623 set_page_idle(newpage);
624
625 /*
626 * Copy NUMA information to the new page, to prevent over-eager
627 * future migrations of this same page.
628 */
629 cpupid = page_cpupid_xchg_last(page, -1);
630 page_cpupid_xchg_last(newpage, cpupid);
631
632 ksm_migrate_page(newpage, page);
633 /*
634 * Please do not reorder this without considering how mm/ksm.c's
635 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
636 */
637 if (PageSwapCache(page))
638 ClearPageSwapCache(page);
639 ClearPagePrivate(page);
640 set_page_private(page, 0);
641
642 /*
643 * If any waiters have accumulated on the new page then
644 * wake them up.
645 */
646 if (PageWriteback(newpage))
647 end_page_writeback(newpage);
648
649 copy_page_owner(page, newpage);
650
651 mem_cgroup_migrate(page, newpage);
652}
653EXPORT_SYMBOL(migrate_page_states);
654
655void migrate_page_copy(struct page *newpage, struct page *page)
656{
657 if (PageHuge(page) || PageTransHuge(page))
658 copy_huge_page(newpage, page);
659 else
660 copy_highpage(newpage, page);
661
662 migrate_page_states(newpage, page);
663}
664EXPORT_SYMBOL(migrate_page_copy);
665
666/************************************************************
667 * Migration functions
668 ***********************************************************/
669
670/*
671 * Common logic to directly migrate a single LRU page suitable for
672 * pages that do not use PagePrivate/PagePrivate2.
673 *
674 * Pages are locked upon entry and exit.
675 */
676int migrate_page(struct address_space *mapping,
677 struct page *newpage, struct page *page,
678 enum migrate_mode mode)
679{
680 int rc;
681
682 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
683
684 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
685
686 if (rc != MIGRATEPAGE_SUCCESS)
687 return rc;
688
689 if (mode != MIGRATE_SYNC_NO_COPY)
690 migrate_page_copy(newpage, page);
691 else
692 migrate_page_states(newpage, page);
693 return MIGRATEPAGE_SUCCESS;
694}
695EXPORT_SYMBOL(migrate_page);
696
697#ifdef CONFIG_BLOCK
698/* Returns true if all buffers are successfully locked */
699static bool buffer_migrate_lock_buffers(struct buffer_head *head,
700 enum migrate_mode mode)
701{
702 struct buffer_head *bh = head;
703
704 /* Simple case, sync compaction */
705 if (mode != MIGRATE_ASYNC) {
706 do {
707 lock_buffer(bh);
708 bh = bh->b_this_page;
709
710 } while (bh != head);
711
712 return true;
713 }
714
715 /* async case, we cannot block on lock_buffer so use trylock_buffer */
716 do {
717 if (!trylock_buffer(bh)) {
718 /*
719 * We failed to lock the buffer and cannot stall in
720 * async migration. Release the taken locks
721 */
722 struct buffer_head *failed_bh = bh;
723 bh = head;
724 while (bh != failed_bh) {
725 unlock_buffer(bh);
726 bh = bh->b_this_page;
727 }
728 return false;
729 }
730
731 bh = bh->b_this_page;
732 } while (bh != head);
733 return true;
734}
735
736static int __buffer_migrate_page(struct address_space *mapping,
737 struct page *newpage, struct page *page, enum migrate_mode mode,
738 bool check_refs)
739{
740 struct buffer_head *bh, *head;
741 int rc;
742 int expected_count;
743
744 if (!page_has_buffers(page))
745 return migrate_page(mapping, newpage, page, mode);
746
747 /* Check whether page does not have extra refs before we do more work */
748 expected_count = expected_page_refs(mapping, page);
749 if (page_count(page) != expected_count)
750 return -EAGAIN;
751
752 head = page_buffers(page);
753 if (!buffer_migrate_lock_buffers(head, mode))
754 return -EAGAIN;
755
756 if (check_refs) {
757 bool busy;
758 bool invalidated = false;
759
760recheck_buffers:
761 busy = false;
762 spin_lock(&mapping->private_lock);
763 bh = head;
764 do {
765 if (atomic_read(&bh->b_count)) {
766 busy = true;
767 break;
768 }
769 bh = bh->b_this_page;
770 } while (bh != head);
771 if (busy) {
772 if (invalidated) {
773 rc = -EAGAIN;
774 goto unlock_buffers;
775 }
776 spin_unlock(&mapping->private_lock);
777 invalidate_bh_lrus();
778 invalidated = true;
779 goto recheck_buffers;
780 }
781 }
782
783 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
784 if (rc != MIGRATEPAGE_SUCCESS)
785 goto unlock_buffers;
786
787 ClearPagePrivate(page);
788 set_page_private(newpage, page_private(page));
789 set_page_private(page, 0);
790 put_page(page);
791 get_page(newpage);
792
793 bh = head;
794 do {
795 set_bh_page(bh, newpage, bh_offset(bh));
796 bh = bh->b_this_page;
797
798 } while (bh != head);
799
800 SetPagePrivate(newpage);
801
802 if (mode != MIGRATE_SYNC_NO_COPY)
803 migrate_page_copy(newpage, page);
804 else
805 migrate_page_states(newpage, page);
806
807 rc = MIGRATEPAGE_SUCCESS;
808unlock_buffers:
809 if (check_refs)
810 spin_unlock(&mapping->private_lock);
811 bh = head;
812 do {
813 unlock_buffer(bh);
814 bh = bh->b_this_page;
815
816 } while (bh != head);
817
818 return rc;
819}
820
821/*
822 * Migration function for pages with buffers. This function can only be used
823 * if the underlying filesystem guarantees that no other references to "page"
824 * exist. For example attached buffer heads are accessed only under page lock.
825 */
826int buffer_migrate_page(struct address_space *mapping,
827 struct page *newpage, struct page *page, enum migrate_mode mode)
828{
829 return __buffer_migrate_page(mapping, newpage, page, mode, false);
830}
831EXPORT_SYMBOL(buffer_migrate_page);
832
833/*
834 * Same as above except that this variant is more careful and checks that there
835 * are also no buffer head references. This function is the right one for
836 * mappings where buffer heads are directly looked up and referenced (such as
837 * block device mappings).
838 */
839int buffer_migrate_page_norefs(struct address_space *mapping,
840 struct page *newpage, struct page *page, enum migrate_mode mode)
841{
842 return __buffer_migrate_page(mapping, newpage, page, mode, true);
843}
844#endif
845
846/*
847 * Writeback a page to clean the dirty state
848 */
849static int writeout(struct address_space *mapping, struct page *page)
850{
851 struct writeback_control wbc = {
852 .sync_mode = WB_SYNC_NONE,
853 .nr_to_write = 1,
854 .range_start = 0,
855 .range_end = LLONG_MAX,
856 .for_reclaim = 1
857 };
858 int rc;
859
860 if (!mapping->a_ops->writepage)
861 /* No write method for the address space */
862 return -EINVAL;
863
864 if (!clear_page_dirty_for_io(page))
865 /* Someone else already triggered a write */
866 return -EAGAIN;
867
868 /*
869 * A dirty page may imply that the underlying filesystem has
870 * the page on some queue. So the page must be clean for
871 * migration. Writeout may mean we loose the lock and the
872 * page state is no longer what we checked for earlier.
873 * At this point we know that the migration attempt cannot
874 * be successful.
875 */
876 remove_migration_ptes(page, page, false);
877
878 rc = mapping->a_ops->writepage(page, &wbc);
879
880 if (rc != AOP_WRITEPAGE_ACTIVATE)
881 /* unlocked. Relock */
882 lock_page(page);
883
884 return (rc < 0) ? -EIO : -EAGAIN;
885}
886
887/*
888 * Default handling if a filesystem does not provide a migration function.
889 */
890static int fallback_migrate_page(struct address_space *mapping,
891 struct page *newpage, struct page *page, enum migrate_mode mode)
892{
893 if (PageDirty(page)) {
894 /* Only writeback pages in full synchronous migration */
895 switch (mode) {
896 case MIGRATE_SYNC:
897 case MIGRATE_SYNC_NO_COPY:
898 break;
899 default:
900 return -EBUSY;
901 }
902 return writeout(mapping, page);
903 }
904
905 /*
906 * Buffers may be managed in a filesystem specific way.
907 * We must have no buffers or drop them.
908 */
909 if (page_has_private(page) &&
910 !try_to_release_page(page, GFP_KERNEL))
911 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
912
913 return migrate_page(mapping, newpage, page, mode);
914}
915
916/*
917 * Move a page to a newly allocated page
918 * The page is locked and all ptes have been successfully removed.
919 *
920 * The new page will have replaced the old page if this function
921 * is successful.
922 *
923 * Return value:
924 * < 0 - error code
925 * MIGRATEPAGE_SUCCESS - success
926 */
927static int move_to_new_page(struct page *newpage, struct page *page,
928 enum migrate_mode mode)
929{
930 struct address_space *mapping;
931 int rc = -EAGAIN;
932 bool is_lru = !__PageMovable(page);
933
934 VM_BUG_ON_PAGE(!PageLocked(page), page);
935 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
936
937 mapping = page_mapping(page);
938
939 if (likely(is_lru)) {
940 if (!mapping)
941 rc = migrate_page(mapping, newpage, page, mode);
942 else if (mapping->a_ops->migratepage)
943 /*
944 * Most pages have a mapping and most filesystems
945 * provide a migratepage callback. Anonymous pages
946 * are part of swap space which also has its own
947 * migratepage callback. This is the most common path
948 * for page migration.
949 */
950 rc = mapping->a_ops->migratepage(mapping, newpage,
951 page, mode);
952 else
953 rc = fallback_migrate_page(mapping, newpage,
954 page, mode);
955 } else {
956 /*
957 * In case of non-lru page, it could be released after
958 * isolation step. In that case, we shouldn't try migration.
959 */
960 VM_BUG_ON_PAGE(!PageIsolated(page), page);
961 if (!PageMovable(page)) {
962 rc = MIGRATEPAGE_SUCCESS;
963 __ClearPageIsolated(page);
964 goto out;
965 }
966
967 rc = mapping->a_ops->migratepage(mapping, newpage,
968 page, mode);
969 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
970 !PageIsolated(page));
971 }
972
973 /*
974 * When successful, old pagecache page->mapping must be cleared before
975 * page is freed; but stats require that PageAnon be left as PageAnon.
976 */
977 if (rc == MIGRATEPAGE_SUCCESS) {
978 if (__PageMovable(page)) {
979 VM_BUG_ON_PAGE(!PageIsolated(page), page);
980
981 /*
982 * We clear PG_movable under page_lock so any compactor
983 * cannot try to migrate this page.
984 */
985 __ClearPageIsolated(page);
986 }
987
988 /*
989 * Anonymous and movable page->mapping will be cleard by
990 * free_pages_prepare so don't reset it here for keeping
991 * the type to work PageAnon, for example.
992 */
993 if (!PageMappingFlags(page))
994 page->mapping = NULL;
995
996 if (likely(!is_zone_device_page(newpage)))
997 flush_dcache_page(newpage);
998
999 }
1000out:
1001 return rc;
1002}
1003
1004static int __unmap_and_move(struct page *page, struct page *newpage,
1005 int force, enum migrate_mode mode)
1006{
1007 int rc = -EAGAIN;
1008 int page_was_mapped = 0;
1009 struct anon_vma *anon_vma = NULL;
1010 bool is_lru = !__PageMovable(page);
1011
1012 if (!trylock_page(page)) {
1013 if (!force || mode == MIGRATE_ASYNC)
1014 goto out;
1015
1016 /*
1017 * It's not safe for direct compaction to call lock_page.
1018 * For example, during page readahead pages are added locked
1019 * to the LRU. Later, when the IO completes the pages are
1020 * marked uptodate and unlocked. However, the queueing
1021 * could be merging multiple pages for one bio (e.g.
1022 * mpage_readpages). If an allocation happens for the
1023 * second or third page, the process can end up locking
1024 * the same page twice and deadlocking. Rather than
1025 * trying to be clever about what pages can be locked,
1026 * avoid the use of lock_page for direct compaction
1027 * altogether.
1028 */
1029 if (current->flags & PF_MEMALLOC)
1030 goto out;
1031
1032 lock_page(page);
1033 }
1034
1035 if (PageWriteback(page)) {
1036 /*
1037 * Only in the case of a full synchronous migration is it
1038 * necessary to wait for PageWriteback. In the async case,
1039 * the retry loop is too short and in the sync-light case,
1040 * the overhead of stalling is too much
1041 */
1042 switch (mode) {
1043 case MIGRATE_SYNC:
1044 case MIGRATE_SYNC_NO_COPY:
1045 break;
1046 default:
1047 rc = -EBUSY;
1048 goto out_unlock;
1049 }
1050 if (!force)
1051 goto out_unlock;
1052 wait_on_page_writeback(page);
1053 }
1054
1055 /*
1056 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1057 * we cannot notice that anon_vma is freed while we migrates a page.
1058 * This get_anon_vma() delays freeing anon_vma pointer until the end
1059 * of migration. File cache pages are no problem because of page_lock()
1060 * File Caches may use write_page() or lock_page() in migration, then,
1061 * just care Anon page here.
1062 *
1063 * Only page_get_anon_vma() understands the subtleties of
1064 * getting a hold on an anon_vma from outside one of its mms.
1065 * But if we cannot get anon_vma, then we won't need it anyway,
1066 * because that implies that the anon page is no longer mapped
1067 * (and cannot be remapped so long as we hold the page lock).
1068 */
1069 if (PageAnon(page) && !PageKsm(page))
1070 anon_vma = page_get_anon_vma(page);
1071
1072 /*
1073 * Block others from accessing the new page when we get around to
1074 * establishing additional references. We are usually the only one
1075 * holding a reference to newpage at this point. We used to have a BUG
1076 * here if trylock_page(newpage) fails, but would like to allow for
1077 * cases where there might be a race with the previous use of newpage.
1078 * This is much like races on refcount of oldpage: just don't BUG().
1079 */
1080 if (unlikely(!trylock_page(newpage)))
1081 goto out_unlock;
1082
1083 if (unlikely(!is_lru)) {
1084 rc = move_to_new_page(newpage, page, mode);
1085 goto out_unlock_both;
1086 }
1087
1088 /*
1089 * Corner case handling:
1090 * 1. When a new swap-cache page is read into, it is added to the LRU
1091 * and treated as swapcache but it has no rmap yet.
1092 * Calling try_to_unmap() against a page->mapping==NULL page will
1093 * trigger a BUG. So handle it here.
1094 * 2. An orphaned page (see truncate_complete_page) might have
1095 * fs-private metadata. The page can be picked up due to memory
1096 * offlining. Everywhere else except page reclaim, the page is
1097 * invisible to the vm, so the page can not be migrated. So try to
1098 * free the metadata, so the page can be freed.
1099 */
1100 if (!page->mapping) {
1101 VM_BUG_ON_PAGE(PageAnon(page), page);
1102 if (page_has_private(page)) {
1103 try_to_free_buffers(page);
1104 goto out_unlock_both;
1105 }
1106 } else if (page_mapped(page)) {
1107 /* Establish migration ptes */
1108 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1109 page);
1110 try_to_unmap(page,
1111 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1112 page_was_mapped = 1;
1113 }
1114
1115 if (!page_mapped(page))
1116 rc = move_to_new_page(newpage, page, mode);
1117
1118 if (page_was_mapped)
1119 remove_migration_ptes(page,
1120 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1121
1122out_unlock_both:
1123 unlock_page(newpage);
1124out_unlock:
1125 /* Drop an anon_vma reference if we took one */
1126 if (anon_vma)
1127 put_anon_vma(anon_vma);
1128 unlock_page(page);
1129out:
1130 /*
1131 * If migration is successful, decrease refcount of the newpage
1132 * which will not free the page because new page owner increased
1133 * refcounter. As well, if it is LRU page, add the page to LRU
1134 * list in here. Use the old state of the isolated source page to
1135 * determine if we migrated a LRU page. newpage was already unlocked
1136 * and possibly modified by its owner - don't rely on the page
1137 * state.
1138 */
1139 if (rc == MIGRATEPAGE_SUCCESS) {
1140 if (unlikely(!is_lru))
1141 put_page(newpage);
1142 else
1143 putback_lru_page(newpage);
1144 }
1145
1146 return rc;
1147}
1148
1149/*
1150 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1151 * around it.
1152 */
1153#if defined(CONFIG_ARM) && \
1154 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1155#define ICE_noinline noinline
1156#else
1157#define ICE_noinline
1158#endif
1159
1160/*
1161 * Obtain the lock on page, remove all ptes and migrate the page
1162 * to the newly allocated page in newpage.
1163 */
1164static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1165 free_page_t put_new_page,
1166 unsigned long private, struct page *page,
1167 int force, enum migrate_mode mode,
1168 enum migrate_reason reason)
1169{
1170 int rc = MIGRATEPAGE_SUCCESS;
1171 struct page *newpage;
1172
1173 if (!thp_migration_supported() && PageTransHuge(page))
1174 return -ENOMEM;
1175
1176 newpage = get_new_page(page, private);
1177 if (!newpage)
1178 return -ENOMEM;
1179
1180 if (page_count(page) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page);
1183 ClearPageUnevictable(page);
1184 if (unlikely(__PageMovable(page))) {
1185 lock_page(page);
1186 if (!PageMovable(page))
1187 __ClearPageIsolated(page);
1188 unlock_page(page);
1189 }
1190 if (put_new_page)
1191 put_new_page(newpage, private);
1192 else
1193 put_page(newpage);
1194 goto out;
1195 }
1196
1197 rc = __unmap_and_move(page, newpage, force, mode);
1198 if (rc == MIGRATEPAGE_SUCCESS)
1199 set_page_owner_migrate_reason(newpage, reason);
1200
1201out:
1202 if (rc != -EAGAIN) {
1203 /*
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kepts its references and be
1207 * restored.
1208 */
1209 list_del(&page->lru);
1210
1211 /*
1212 * Compaction can migrate also non-LRU pages which are
1213 * not accounted to NR_ISOLATED_*. They can be recognized
1214 * as __PageMovable
1215 */
1216 if (likely(!__PageMovable(page)))
1217 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1218 page_is_file_cache(page), -hpage_nr_pages(page));
1219 }
1220
1221 /*
1222 * If migration is successful, releases reference grabbed during
1223 * isolation. Otherwise, restore the page to right list unless
1224 * we want to retry.
1225 */
1226 if (rc == MIGRATEPAGE_SUCCESS) {
1227 put_page(page);
1228 if (reason == MR_MEMORY_FAILURE) {
1229 /*
1230 * Set PG_HWPoison on just freed page
1231 * intentionally. Although it's rather weird,
1232 * it's how HWPoison flag works at the moment.
1233 */
1234 if (set_hwpoison_free_buddy_page(page))
1235 num_poisoned_pages_inc();
1236 }
1237 } else {
1238 if (rc != -EAGAIN) {
1239 if (likely(!__PageMovable(page))) {
1240 putback_lru_page(page);
1241 goto put_new;
1242 }
1243
1244 lock_page(page);
1245 if (PageMovable(page))
1246 putback_movable_page(page);
1247 else
1248 __ClearPageIsolated(page);
1249 unlock_page(page);
1250 put_page(page);
1251 }
1252put_new:
1253 if (put_new_page)
1254 put_new_page(newpage, private);
1255 else
1256 put_page(newpage);
1257 }
1258
1259 return rc;
1260}
1261
1262/*
1263 * Counterpart of unmap_and_move_page() for hugepage migration.
1264 *
1265 * This function doesn't wait the completion of hugepage I/O
1266 * because there is no race between I/O and migration for hugepage.
1267 * Note that currently hugepage I/O occurs only in direct I/O
1268 * where no lock is held and PG_writeback is irrelevant,
1269 * and writeback status of all subpages are counted in the reference
1270 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1271 * under direct I/O, the reference of the head page is 512 and a bit more.)
1272 * This means that when we try to migrate hugepage whose subpages are
1273 * doing direct I/O, some references remain after try_to_unmap() and
1274 * hugepage migration fails without data corruption.
1275 *
1276 * There is also no race when direct I/O is issued on the page under migration,
1277 * because then pte is replaced with migration swap entry and direct I/O code
1278 * will wait in the page fault for migration to complete.
1279 */
1280static int unmap_and_move_huge_page(new_page_t get_new_page,
1281 free_page_t put_new_page, unsigned long private,
1282 struct page *hpage, int force,
1283 enum migrate_mode mode, int reason)
1284{
1285 int rc = -EAGAIN;
1286 int page_was_mapped = 0;
1287 struct page *new_hpage;
1288 struct anon_vma *anon_vma = NULL;
1289
1290 /*
1291 * Migratability of hugepages depends on architectures and their size.
1292 * This check is necessary because some callers of hugepage migration
1293 * like soft offline and memory hotremove don't walk through page
1294 * tables or check whether the hugepage is pmd-based or not before
1295 * kicking migration.
1296 */
1297 if (!hugepage_migration_supported(page_hstate(hpage))) {
1298 putback_active_hugepage(hpage);
1299 return -ENOSYS;
1300 }
1301
1302 new_hpage = get_new_page(hpage, private);
1303 if (!new_hpage)
1304 return -ENOMEM;
1305
1306 if (!trylock_page(hpage)) {
1307 if (!force)
1308 goto out;
1309 switch (mode) {
1310 case MIGRATE_SYNC:
1311 case MIGRATE_SYNC_NO_COPY:
1312 break;
1313 default:
1314 goto out;
1315 }
1316 lock_page(hpage);
1317 }
1318
1319 /*
1320 * Check for pages which are in the process of being freed. Without
1321 * page_mapping() set, hugetlbfs specific move page routine will not
1322 * be called and we could leak usage counts for subpools.
1323 */
1324 if (page_private(hpage) && !page_mapping(hpage)) {
1325 rc = -EBUSY;
1326 goto out_unlock;
1327 }
1328
1329 if (PageAnon(hpage))
1330 anon_vma = page_get_anon_vma(hpage);
1331
1332 if (unlikely(!trylock_page(new_hpage)))
1333 goto put_anon;
1334
1335 if (page_mapped(hpage)) {
1336 try_to_unmap(hpage,
1337 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1338 page_was_mapped = 1;
1339 }
1340
1341 if (!page_mapped(hpage))
1342 rc = move_to_new_page(new_hpage, hpage, mode);
1343
1344 if (page_was_mapped)
1345 remove_migration_ptes(hpage,
1346 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1347
1348 unlock_page(new_hpage);
1349
1350put_anon:
1351 if (anon_vma)
1352 put_anon_vma(anon_vma);
1353
1354 if (rc == MIGRATEPAGE_SUCCESS) {
1355 move_hugetlb_state(hpage, new_hpage, reason);
1356 put_new_page = NULL;
1357 }
1358
1359out_unlock:
1360 unlock_page(hpage);
1361out:
1362 if (rc != -EAGAIN)
1363 putback_active_hugepage(hpage);
1364
1365 /*
1366 * If migration was not successful and there's a freeing callback, use
1367 * it. Otherwise, put_page() will drop the reference grabbed during
1368 * isolation.
1369 */
1370 if (put_new_page)
1371 put_new_page(new_hpage, private);
1372 else
1373 putback_active_hugepage(new_hpage);
1374
1375 return rc;
1376}
1377
1378/*
1379 * migrate_pages - migrate the pages specified in a list, to the free pages
1380 * supplied as the target for the page migration
1381 *
1382 * @from: The list of pages to be migrated.
1383 * @get_new_page: The function used to allocate free pages to be used
1384 * as the target of the page migration.
1385 * @put_new_page: The function used to free target pages if migration
1386 * fails, or NULL if no special handling is necessary.
1387 * @private: Private data to be passed on to get_new_page()
1388 * @mode: The migration mode that specifies the constraints for
1389 * page migration, if any.
1390 * @reason: The reason for page migration.
1391 *
1392 * The function returns after 10 attempts or if no pages are movable any more
1393 * because the list has become empty or no retryable pages exist any more.
1394 * The caller should call putback_movable_pages() to return pages to the LRU
1395 * or free list only if ret != 0.
1396 *
1397 * Returns the number of pages that were not migrated, or an error code.
1398 */
1399int migrate_pages(struct list_head *from, new_page_t get_new_page,
1400 free_page_t put_new_page, unsigned long private,
1401 enum migrate_mode mode, int reason)
1402{
1403 int retry = 1;
1404 int nr_failed = 0;
1405 int nr_succeeded = 0;
1406 int pass = 0;
1407 struct page *page;
1408 struct page *page2;
1409 int swapwrite = current->flags & PF_SWAPWRITE;
1410 int rc;
1411
1412 if (!swapwrite)
1413 current->flags |= PF_SWAPWRITE;
1414
1415 for(pass = 0; pass < 10 && retry; pass++) {
1416 retry = 0;
1417
1418 list_for_each_entry_safe(page, page2, from, lru) {
1419retry:
1420 cond_resched();
1421
1422 if (PageHuge(page))
1423 rc = unmap_and_move_huge_page(get_new_page,
1424 put_new_page, private, page,
1425 pass > 2, mode, reason);
1426 else
1427 rc = unmap_and_move(get_new_page, put_new_page,
1428 private, page, pass > 2, mode,
1429 reason);
1430
1431 switch(rc) {
1432 case -ENOMEM:
1433 /*
1434 * THP migration might be unsupported or the
1435 * allocation could've failed so we should
1436 * retry on the same page with the THP split
1437 * to base pages.
1438 *
1439 * Head page is retried immediately and tail
1440 * pages are added to the tail of the list so
1441 * we encounter them after the rest of the list
1442 * is processed.
1443 */
1444 if (PageTransHuge(page) && !PageHuge(page)) {
1445 lock_page(page);
1446 rc = split_huge_page_to_list(page, from);
1447 unlock_page(page);
1448 if (!rc) {
1449 list_safe_reset_next(page, page2, lru);
1450 goto retry;
1451 }
1452 }
1453 nr_failed++;
1454 goto out;
1455 case -EAGAIN:
1456 retry++;
1457 break;
1458 case MIGRATEPAGE_SUCCESS:
1459 nr_succeeded++;
1460 break;
1461 default:
1462 /*
1463 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1464 * unlike -EAGAIN case, the failed page is
1465 * removed from migration page list and not
1466 * retried in the next outer loop.
1467 */
1468 nr_failed++;
1469 break;
1470 }
1471 }
1472 }
1473 nr_failed += retry;
1474 rc = nr_failed;
1475out:
1476 if (nr_succeeded)
1477 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1478 if (nr_failed)
1479 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1480 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1481
1482 if (!swapwrite)
1483 current->flags &= ~PF_SWAPWRITE;
1484
1485 return rc;
1486}
1487
1488#ifdef CONFIG_NUMA
1489
1490static int store_status(int __user *status, int start, int value, int nr)
1491{
1492 while (nr-- > 0) {
1493 if (put_user(value, status + start))
1494 return -EFAULT;
1495 start++;
1496 }
1497
1498 return 0;
1499}
1500
1501static int do_move_pages_to_node(struct mm_struct *mm,
1502 struct list_head *pagelist, int node)
1503{
1504 int err;
1505
1506 if (list_empty(pagelist))
1507 return 0;
1508
1509 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1510 MIGRATE_SYNC, MR_SYSCALL);
1511 if (err)
1512 putback_movable_pages(pagelist);
1513 return err;
1514}
1515
1516/*
1517 * Resolves the given address to a struct page, isolates it from the LRU and
1518 * puts it to the given pagelist.
1519 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1520 * queued or the page doesn't need to be migrated because it is already on
1521 * the target node
1522 */
1523static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1524 int node, struct list_head *pagelist, bool migrate_all)
1525{
1526 struct vm_area_struct *vma;
1527 struct page *page;
1528 unsigned int follflags;
1529 int err;
1530
1531 down_read(&mm->mmap_sem);
1532 err = -EFAULT;
1533 vma = find_vma(mm, addr);
1534 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1535 goto out;
1536
1537 /* FOLL_DUMP to ignore special (like zero) pages */
1538 follflags = FOLL_GET | FOLL_DUMP;
1539 page = follow_page(vma, addr, follflags);
1540
1541 err = PTR_ERR(page);
1542 if (IS_ERR(page))
1543 goto out;
1544
1545 err = -ENOENT;
1546 if (!page)
1547 goto out;
1548
1549 err = 0;
1550 if (page_to_nid(page) == node)
1551 goto out_putpage;
1552
1553 err = -EACCES;
1554 if (page_mapcount(page) > 1 && !migrate_all)
1555 goto out_putpage;
1556
1557 if (PageHuge(page)) {
1558 if (PageHead(page)) {
1559 isolate_huge_page(page, pagelist);
1560 err = 0;
1561 }
1562 } else {
1563 struct page *head;
1564
1565 head = compound_head(page);
1566 err = isolate_lru_page(head);
1567 if (err)
1568 goto out_putpage;
1569
1570 err = 0;
1571 list_add_tail(&head->lru, pagelist);
1572 mod_node_page_state(page_pgdat(head),
1573 NR_ISOLATED_ANON + page_is_file_cache(head),
1574 hpage_nr_pages(head));
1575 }
1576out_putpage:
1577 /*
1578 * Either remove the duplicate refcount from
1579 * isolate_lru_page() or drop the page ref if it was
1580 * not isolated.
1581 */
1582 put_page(page);
1583out:
1584 up_read(&mm->mmap_sem);
1585 return err;
1586}
1587
1588/*
1589 * Migrate an array of page address onto an array of nodes and fill
1590 * the corresponding array of status.
1591 */
1592static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1593 unsigned long nr_pages,
1594 const void __user * __user *pages,
1595 const int __user *nodes,
1596 int __user *status, int flags)
1597{
1598 int current_node = NUMA_NO_NODE;
1599 LIST_HEAD(pagelist);
1600 int start, i;
1601 int err = 0, err1;
1602
1603 migrate_prep();
1604
1605 for (i = start = 0; i < nr_pages; i++) {
1606 const void __user *p;
1607 unsigned long addr;
1608 int node;
1609
1610 err = -EFAULT;
1611 if (get_user(p, pages + i))
1612 goto out_flush;
1613 if (get_user(node, nodes + i))
1614 goto out_flush;
1615 addr = (unsigned long)untagged_addr(p);
1616
1617 err = -ENODEV;
1618 if (node < 0 || node >= MAX_NUMNODES)
1619 goto out_flush;
1620 if (!node_state(node, N_MEMORY))
1621 goto out_flush;
1622
1623 err = -EACCES;
1624 if (!node_isset(node, task_nodes))
1625 goto out_flush;
1626
1627 if (current_node == NUMA_NO_NODE) {
1628 current_node = node;
1629 start = i;
1630 } else if (node != current_node) {
1631 err = do_move_pages_to_node(mm, &pagelist, current_node);
1632 if (err)
1633 goto out;
1634 err = store_status(status, start, current_node, i - start);
1635 if (err)
1636 goto out;
1637 start = i;
1638 current_node = node;
1639 }
1640
1641 /*
1642 * Errors in the page lookup or isolation are not fatal and we simply
1643 * report them via status
1644 */
1645 err = add_page_for_migration(mm, addr, current_node,
1646 &pagelist, flags & MPOL_MF_MOVE_ALL);
1647 if (!err)
1648 continue;
1649
1650 err = store_status(status, i, err, 1);
1651 if (err)
1652 goto out_flush;
1653
1654 err = do_move_pages_to_node(mm, &pagelist, current_node);
1655 if (err)
1656 goto out;
1657 if (i > start) {
1658 err = store_status(status, start, current_node, i - start);
1659 if (err)
1660 goto out;
1661 }
1662 current_node = NUMA_NO_NODE;
1663 }
1664out_flush:
1665 if (list_empty(&pagelist))
1666 return err;
1667
1668 /* Make sure we do not overwrite the existing error */
1669 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1670 if (!err1)
1671 err1 = store_status(status, start, current_node, i - start);
1672 if (!err)
1673 err = err1;
1674out:
1675 return err;
1676}
1677
1678/*
1679 * Determine the nodes of an array of pages and store it in an array of status.
1680 */
1681static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1682 const void __user **pages, int *status)
1683{
1684 unsigned long i;
1685
1686 down_read(&mm->mmap_sem);
1687
1688 for (i = 0; i < nr_pages; i++) {
1689 unsigned long addr = (unsigned long)(*pages);
1690 struct vm_area_struct *vma;
1691 struct page *page;
1692 int err = -EFAULT;
1693
1694 vma = find_vma(mm, addr);
1695 if (!vma || addr < vma->vm_start)
1696 goto set_status;
1697
1698 /* FOLL_DUMP to ignore special (like zero) pages */
1699 page = follow_page(vma, addr, FOLL_DUMP);
1700
1701 err = PTR_ERR(page);
1702 if (IS_ERR(page))
1703 goto set_status;
1704
1705 err = page ? page_to_nid(page) : -ENOENT;
1706set_status:
1707 *status = err;
1708
1709 pages++;
1710 status++;
1711 }
1712
1713 up_read(&mm->mmap_sem);
1714}
1715
1716/*
1717 * Determine the nodes of a user array of pages and store it in
1718 * a user array of status.
1719 */
1720static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1721 const void __user * __user *pages,
1722 int __user *status)
1723{
1724#define DO_PAGES_STAT_CHUNK_NR 16
1725 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1726 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1727
1728 while (nr_pages) {
1729 unsigned long chunk_nr;
1730
1731 chunk_nr = nr_pages;
1732 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1733 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1734
1735 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1736 break;
1737
1738 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1739
1740 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1741 break;
1742
1743 pages += chunk_nr;
1744 status += chunk_nr;
1745 nr_pages -= chunk_nr;
1746 }
1747 return nr_pages ? -EFAULT : 0;
1748}
1749
1750/*
1751 * Move a list of pages in the address space of the currently executing
1752 * process.
1753 */
1754static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1755 const void __user * __user *pages,
1756 const int __user *nodes,
1757 int __user *status, int flags)
1758{
1759 struct task_struct *task;
1760 struct mm_struct *mm;
1761 int err;
1762 nodemask_t task_nodes;
1763
1764 /* Check flags */
1765 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1766 return -EINVAL;
1767
1768 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1769 return -EPERM;
1770
1771 /* Find the mm_struct */
1772 rcu_read_lock();
1773 task = pid ? find_task_by_vpid(pid) : current;
1774 if (!task) {
1775 rcu_read_unlock();
1776 return -ESRCH;
1777 }
1778 get_task_struct(task);
1779
1780 /*
1781 * Check if this process has the right to modify the specified
1782 * process. Use the regular "ptrace_may_access()" checks.
1783 */
1784 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1785 rcu_read_unlock();
1786 err = -EPERM;
1787 goto out;
1788 }
1789 rcu_read_unlock();
1790
1791 err = security_task_movememory(task);
1792 if (err)
1793 goto out;
1794
1795 task_nodes = cpuset_mems_allowed(task);
1796 mm = get_task_mm(task);
1797 put_task_struct(task);
1798
1799 if (!mm)
1800 return -EINVAL;
1801
1802 if (nodes)
1803 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1804 nodes, status, flags);
1805 else
1806 err = do_pages_stat(mm, nr_pages, pages, status);
1807
1808 mmput(mm);
1809 return err;
1810
1811out:
1812 put_task_struct(task);
1813 return err;
1814}
1815
1816SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1817 const void __user * __user *, pages,
1818 const int __user *, nodes,
1819 int __user *, status, int, flags)
1820{
1821 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1822}
1823
1824#ifdef CONFIG_COMPAT
1825COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1826 compat_uptr_t __user *, pages32,
1827 const int __user *, nodes,
1828 int __user *, status,
1829 int, flags)
1830{
1831 const void __user * __user *pages;
1832 int i;
1833
1834 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1835 for (i = 0; i < nr_pages; i++) {
1836 compat_uptr_t p;
1837
1838 if (get_user(p, pages32 + i) ||
1839 put_user(compat_ptr(p), pages + i))
1840 return -EFAULT;
1841 }
1842 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1843}
1844#endif /* CONFIG_COMPAT */
1845
1846#ifdef CONFIG_NUMA_BALANCING
1847/*
1848 * Returns true if this is a safe migration target node for misplaced NUMA
1849 * pages. Currently it only checks the watermarks which crude
1850 */
1851static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1852 unsigned long nr_migrate_pages)
1853{
1854 int z;
1855
1856 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1857 struct zone *zone = pgdat->node_zones + z;
1858
1859 if (!populated_zone(zone))
1860 continue;
1861
1862 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1863 if (!zone_watermark_ok(zone, 0,
1864 high_wmark_pages(zone) +
1865 nr_migrate_pages,
1866 0, 0))
1867 continue;
1868 return true;
1869 }
1870 return false;
1871}
1872
1873static struct page *alloc_misplaced_dst_page(struct page *page,
1874 unsigned long data)
1875{
1876 int nid = (int) data;
1877 struct page *newpage;
1878
1879 newpage = __alloc_pages_node(nid,
1880 (GFP_HIGHUSER_MOVABLE |
1881 __GFP_THISNODE | __GFP_NOMEMALLOC |
1882 __GFP_NORETRY | __GFP_NOWARN) &
1883 ~__GFP_RECLAIM, 0);
1884
1885 return newpage;
1886}
1887
1888static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1889{
1890 int page_lru;
1891
1892 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1893
1894 /* Avoid migrating to a node that is nearly full */
1895 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
1896 return 0;
1897
1898 if (isolate_lru_page(page))
1899 return 0;
1900
1901 /*
1902 * migrate_misplaced_transhuge_page() skips page migration's usual
1903 * check on page_count(), so we must do it here, now that the page
1904 * has been isolated: a GUP pin, or any other pin, prevents migration.
1905 * The expected page count is 3: 1 for page's mapcount and 1 for the
1906 * caller's pin and 1 for the reference taken by isolate_lru_page().
1907 */
1908 if (PageTransHuge(page) && page_count(page) != 3) {
1909 putback_lru_page(page);
1910 return 0;
1911 }
1912
1913 page_lru = page_is_file_cache(page);
1914 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1915 hpage_nr_pages(page));
1916
1917 /*
1918 * Isolating the page has taken another reference, so the
1919 * caller's reference can be safely dropped without the page
1920 * disappearing underneath us during migration.
1921 */
1922 put_page(page);
1923 return 1;
1924}
1925
1926bool pmd_trans_migrating(pmd_t pmd)
1927{
1928 struct page *page = pmd_page(pmd);
1929 return PageLocked(page);
1930}
1931
1932/*
1933 * Attempt to migrate a misplaced page to the specified destination
1934 * node. Caller is expected to have an elevated reference count on
1935 * the page that will be dropped by this function before returning.
1936 */
1937int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1938 int node)
1939{
1940 pg_data_t *pgdat = NODE_DATA(node);
1941 int isolated;
1942 int nr_remaining;
1943 LIST_HEAD(migratepages);
1944
1945 /*
1946 * Don't migrate file pages that are mapped in multiple processes
1947 * with execute permissions as they are probably shared libraries.
1948 */
1949 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1950 (vma->vm_flags & VM_EXEC))
1951 goto out;
1952
1953 /*
1954 * Also do not migrate dirty pages as not all filesystems can move
1955 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1956 */
1957 if (page_is_file_cache(page) && PageDirty(page))
1958 goto out;
1959
1960 isolated = numamigrate_isolate_page(pgdat, page);
1961 if (!isolated)
1962 goto out;
1963
1964 list_add(&page->lru, &migratepages);
1965 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1966 NULL, node, MIGRATE_ASYNC,
1967 MR_NUMA_MISPLACED);
1968 if (nr_remaining) {
1969 if (!list_empty(&migratepages)) {
1970 list_del(&page->lru);
1971 dec_node_page_state(page, NR_ISOLATED_ANON +
1972 page_is_file_cache(page));
1973 putback_lru_page(page);
1974 }
1975 isolated = 0;
1976 } else
1977 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1978 BUG_ON(!list_empty(&migratepages));
1979 return isolated;
1980
1981out:
1982 put_page(page);
1983 return 0;
1984}
1985#endif /* CONFIG_NUMA_BALANCING */
1986
1987#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1988/*
1989 * Migrates a THP to a given target node. page must be locked and is unlocked
1990 * before returning.
1991 */
1992int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1993 struct vm_area_struct *vma,
1994 pmd_t *pmd, pmd_t entry,
1995 unsigned long address,
1996 struct page *page, int node)
1997{
1998 spinlock_t *ptl;
1999 pg_data_t *pgdat = NODE_DATA(node);
2000 int isolated = 0;
2001 struct page *new_page = NULL;
2002 int page_lru = page_is_file_cache(page);
2003 unsigned long start = address & HPAGE_PMD_MASK;
2004
2005 new_page = alloc_pages_node(node,
2006 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2007 HPAGE_PMD_ORDER);
2008 if (!new_page)
2009 goto out_fail;
2010 prep_transhuge_page(new_page);
2011
2012 isolated = numamigrate_isolate_page(pgdat, page);
2013 if (!isolated) {
2014 put_page(new_page);
2015 goto out_fail;
2016 }
2017
2018 /* Prepare a page as a migration target */
2019 __SetPageLocked(new_page);
2020 if (PageSwapBacked(page))
2021 __SetPageSwapBacked(new_page);
2022
2023 /* anon mapping, we can simply copy page->mapping to the new page: */
2024 new_page->mapping = page->mapping;
2025 new_page->index = page->index;
2026 /* flush the cache before copying using the kernel virtual address */
2027 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2028 migrate_page_copy(new_page, page);
2029 WARN_ON(PageLRU(new_page));
2030
2031 /* Recheck the target PMD */
2032 ptl = pmd_lock(mm, pmd);
2033 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2034 spin_unlock(ptl);
2035
2036 /* Reverse changes made by migrate_page_copy() */
2037 if (TestClearPageActive(new_page))
2038 SetPageActive(page);
2039 if (TestClearPageUnevictable(new_page))
2040 SetPageUnevictable(page);
2041
2042 unlock_page(new_page);
2043 put_page(new_page); /* Free it */
2044
2045 /* Retake the callers reference and putback on LRU */
2046 get_page(page);
2047 putback_lru_page(page);
2048 mod_node_page_state(page_pgdat(page),
2049 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2050
2051 goto out_unlock;
2052 }
2053
2054 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2055 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2056
2057 /*
2058 * Overwrite the old entry under pagetable lock and establish
2059 * the new PTE. Any parallel GUP will either observe the old
2060 * page blocking on the page lock, block on the page table
2061 * lock or observe the new page. The SetPageUptodate on the
2062 * new page and page_add_new_anon_rmap guarantee the copy is
2063 * visible before the pagetable update.
2064 */
2065 page_add_anon_rmap(new_page, vma, start, true);
2066 /*
2067 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2068 * has already been flushed globally. So no TLB can be currently
2069 * caching this non present pmd mapping. There's no need to clear the
2070 * pmd before doing set_pmd_at(), nor to flush the TLB after
2071 * set_pmd_at(). Clearing the pmd here would introduce a race
2072 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2073 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2074 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2075 * pmd.
2076 */
2077 set_pmd_at(mm, start, pmd, entry);
2078 update_mmu_cache_pmd(vma, address, &entry);
2079
2080 page_ref_unfreeze(page, 2);
2081 mlock_migrate_page(new_page, page);
2082 page_remove_rmap(page, true);
2083 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2084
2085 spin_unlock(ptl);
2086
2087 /* Take an "isolate" reference and put new page on the LRU. */
2088 get_page(new_page);
2089 putback_lru_page(new_page);
2090
2091 unlock_page(new_page);
2092 unlock_page(page);
2093 put_page(page); /* Drop the rmap reference */
2094 put_page(page); /* Drop the LRU isolation reference */
2095
2096 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2097 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2098
2099 mod_node_page_state(page_pgdat(page),
2100 NR_ISOLATED_ANON + page_lru,
2101 -HPAGE_PMD_NR);
2102 return isolated;
2103
2104out_fail:
2105 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2106 ptl = pmd_lock(mm, pmd);
2107 if (pmd_same(*pmd, entry)) {
2108 entry = pmd_modify(entry, vma->vm_page_prot);
2109 set_pmd_at(mm, start, pmd, entry);
2110 update_mmu_cache_pmd(vma, address, &entry);
2111 }
2112 spin_unlock(ptl);
2113
2114out_unlock:
2115 unlock_page(page);
2116 put_page(page);
2117 return 0;
2118}
2119#endif /* CONFIG_NUMA_BALANCING */
2120
2121#endif /* CONFIG_NUMA */
2122
2123#ifdef CONFIG_DEVICE_PRIVATE
2124static int migrate_vma_collect_hole(unsigned long start,
2125 unsigned long end,
2126 struct mm_walk *walk)
2127{
2128 struct migrate_vma *migrate = walk->private;
2129 unsigned long addr;
2130
2131 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2132 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2133 migrate->dst[migrate->npages] = 0;
2134 migrate->npages++;
2135 migrate->cpages++;
2136 }
2137
2138 return 0;
2139}
2140
2141static int migrate_vma_collect_skip(unsigned long start,
2142 unsigned long end,
2143 struct mm_walk *walk)
2144{
2145 struct migrate_vma *migrate = walk->private;
2146 unsigned long addr;
2147
2148 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2149 migrate->dst[migrate->npages] = 0;
2150 migrate->src[migrate->npages++] = 0;
2151 }
2152
2153 return 0;
2154}
2155
2156static int migrate_vma_collect_pmd(pmd_t *pmdp,
2157 unsigned long start,
2158 unsigned long end,
2159 struct mm_walk *walk)
2160{
2161 struct migrate_vma *migrate = walk->private;
2162 struct vm_area_struct *vma = walk->vma;
2163 struct mm_struct *mm = vma->vm_mm;
2164 unsigned long addr = start, unmapped = 0;
2165 spinlock_t *ptl;
2166 pte_t *ptep;
2167
2168again:
2169 if (pmd_none(*pmdp))
2170 return migrate_vma_collect_hole(start, end, walk);
2171
2172 if (pmd_trans_huge(*pmdp)) {
2173 struct page *page;
2174
2175 ptl = pmd_lock(mm, pmdp);
2176 if (unlikely(!pmd_trans_huge(*pmdp))) {
2177 spin_unlock(ptl);
2178 goto again;
2179 }
2180
2181 page = pmd_page(*pmdp);
2182 if (is_huge_zero_page(page)) {
2183 spin_unlock(ptl);
2184 split_huge_pmd(vma, pmdp, addr);
2185 if (pmd_trans_unstable(pmdp))
2186 return migrate_vma_collect_skip(start, end,
2187 walk);
2188 } else {
2189 int ret;
2190
2191 get_page(page);
2192 spin_unlock(ptl);
2193 if (unlikely(!trylock_page(page)))
2194 return migrate_vma_collect_skip(start, end,
2195 walk);
2196 ret = split_huge_page(page);
2197 unlock_page(page);
2198 put_page(page);
2199 if (ret)
2200 return migrate_vma_collect_skip(start, end,
2201 walk);
2202 if (pmd_none(*pmdp))
2203 return migrate_vma_collect_hole(start, end,
2204 walk);
2205 }
2206 }
2207
2208 if (unlikely(pmd_bad(*pmdp)))
2209 return migrate_vma_collect_skip(start, end, walk);
2210
2211 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2212 arch_enter_lazy_mmu_mode();
2213
2214 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2215 unsigned long mpfn, pfn;
2216 struct page *page;
2217 swp_entry_t entry;
2218 pte_t pte;
2219
2220 pte = *ptep;
2221
2222 if (pte_none(pte)) {
2223 mpfn = MIGRATE_PFN_MIGRATE;
2224 migrate->cpages++;
2225 goto next;
2226 }
2227
2228 if (!pte_present(pte)) {
2229 mpfn = 0;
2230
2231 /*
2232 * Only care about unaddressable device page special
2233 * page table entry. Other special swap entries are not
2234 * migratable, and we ignore regular swapped page.
2235 */
2236 entry = pte_to_swp_entry(pte);
2237 if (!is_device_private_entry(entry))
2238 goto next;
2239
2240 page = device_private_entry_to_page(entry);
2241 mpfn = migrate_pfn(page_to_pfn(page)) |
2242 MIGRATE_PFN_MIGRATE;
2243 if (is_write_device_private_entry(entry))
2244 mpfn |= MIGRATE_PFN_WRITE;
2245 } else {
2246 pfn = pte_pfn(pte);
2247 if (is_zero_pfn(pfn)) {
2248 mpfn = MIGRATE_PFN_MIGRATE;
2249 migrate->cpages++;
2250 goto next;
2251 }
2252 page = vm_normal_page(migrate->vma, addr, pte);
2253 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2254 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2255 }
2256
2257 /* FIXME support THP */
2258 if (!page || !page->mapping || PageTransCompound(page)) {
2259 mpfn = 0;
2260 goto next;
2261 }
2262
2263 /*
2264 * By getting a reference on the page we pin it and that blocks
2265 * any kind of migration. Side effect is that it "freezes" the
2266 * pte.
2267 *
2268 * We drop this reference after isolating the page from the lru
2269 * for non device page (device page are not on the lru and thus
2270 * can't be dropped from it).
2271 */
2272 get_page(page);
2273 migrate->cpages++;
2274
2275 /*
2276 * Optimize for the common case where page is only mapped once
2277 * in one process. If we can lock the page, then we can safely
2278 * set up a special migration page table entry now.
2279 */
2280 if (trylock_page(page)) {
2281 pte_t swp_pte;
2282
2283 mpfn |= MIGRATE_PFN_LOCKED;
2284 ptep_get_and_clear(mm, addr, ptep);
2285
2286 /* Setup special migration page table entry */
2287 entry = make_migration_entry(page, mpfn &
2288 MIGRATE_PFN_WRITE);
2289 swp_pte = swp_entry_to_pte(entry);
2290 if (pte_soft_dirty(pte))
2291 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2292 set_pte_at(mm, addr, ptep, swp_pte);
2293
2294 /*
2295 * This is like regular unmap: we remove the rmap and
2296 * drop page refcount. Page won't be freed, as we took
2297 * a reference just above.
2298 */
2299 page_remove_rmap(page, false);
2300 put_page(page);
2301
2302 if (pte_present(pte))
2303 unmapped++;
2304 }
2305
2306next:
2307 migrate->dst[migrate->npages] = 0;
2308 migrate->src[migrate->npages++] = mpfn;
2309 }
2310 arch_leave_lazy_mmu_mode();
2311 pte_unmap_unlock(ptep - 1, ptl);
2312
2313 /* Only flush the TLB if we actually modified any entries */
2314 if (unmapped)
2315 flush_tlb_range(walk->vma, start, end);
2316
2317 return 0;
2318}
2319
2320static const struct mm_walk_ops migrate_vma_walk_ops = {
2321 .pmd_entry = migrate_vma_collect_pmd,
2322 .pte_hole = migrate_vma_collect_hole,
2323};
2324
2325/*
2326 * migrate_vma_collect() - collect pages over a range of virtual addresses
2327 * @migrate: migrate struct containing all migration information
2328 *
2329 * This will walk the CPU page table. For each virtual address backed by a
2330 * valid page, it updates the src array and takes a reference on the page, in
2331 * order to pin the page until we lock it and unmap it.
2332 */
2333static void migrate_vma_collect(struct migrate_vma *migrate)
2334{
2335 struct mmu_notifier_range range;
2336
2337 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL,
2338 migrate->vma->vm_mm, migrate->start, migrate->end);
2339 mmu_notifier_invalidate_range_start(&range);
2340
2341 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2342 &migrate_vma_walk_ops, migrate);
2343
2344 mmu_notifier_invalidate_range_end(&range);
2345 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2346}
2347
2348/*
2349 * migrate_vma_check_page() - check if page is pinned or not
2350 * @page: struct page to check
2351 *
2352 * Pinned pages cannot be migrated. This is the same test as in
2353 * migrate_page_move_mapping(), except that here we allow migration of a
2354 * ZONE_DEVICE page.
2355 */
2356static bool migrate_vma_check_page(struct page *page)
2357{
2358 /*
2359 * One extra ref because caller holds an extra reference, either from
2360 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2361 * a device page.
2362 */
2363 int extra = 1;
2364
2365 /*
2366 * FIXME support THP (transparent huge page), it is bit more complex to
2367 * check them than regular pages, because they can be mapped with a pmd
2368 * or with a pte (split pte mapping).
2369 */
2370 if (PageCompound(page))
2371 return false;
2372
2373 /* Page from ZONE_DEVICE have one extra reference */
2374 if (is_zone_device_page(page)) {
2375 /*
2376 * Private page can never be pin as they have no valid pte and
2377 * GUP will fail for those. Yet if there is a pending migration
2378 * a thread might try to wait on the pte migration entry and
2379 * will bump the page reference count. Sadly there is no way to
2380 * differentiate a regular pin from migration wait. Hence to
2381 * avoid 2 racing thread trying to migrate back to CPU to enter
2382 * infinite loop (one stoping migration because the other is
2383 * waiting on pte migration entry). We always return true here.
2384 *
2385 * FIXME proper solution is to rework migration_entry_wait() so
2386 * it does not need to take a reference on page.
2387 */
2388 return is_device_private_page(page);
2389 }
2390
2391 /* For file back page */
2392 if (page_mapping(page))
2393 extra += 1 + page_has_private(page);
2394
2395 if ((page_count(page) - extra) > page_mapcount(page))
2396 return false;
2397
2398 return true;
2399}
2400
2401/*
2402 * migrate_vma_prepare() - lock pages and isolate them from the lru
2403 * @migrate: migrate struct containing all migration information
2404 *
2405 * This locks pages that have been collected by migrate_vma_collect(). Once each
2406 * page is locked it is isolated from the lru (for non-device pages). Finally,
2407 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2408 * migrated by concurrent kernel threads.
2409 */
2410static void migrate_vma_prepare(struct migrate_vma *migrate)
2411{
2412 const unsigned long npages = migrate->npages;
2413 const unsigned long start = migrate->start;
2414 unsigned long addr, i, restore = 0;
2415 bool allow_drain = true;
2416
2417 lru_add_drain();
2418
2419 for (i = 0; (i < npages) && migrate->cpages; i++) {
2420 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2421 bool remap = true;
2422
2423 if (!page)
2424 continue;
2425
2426 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2427 /*
2428 * Because we are migrating several pages there can be
2429 * a deadlock between 2 concurrent migration where each
2430 * are waiting on each other page lock.
2431 *
2432 * Make migrate_vma() a best effort thing and backoff
2433 * for any page we can not lock right away.
2434 */
2435 if (!trylock_page(page)) {
2436 migrate->src[i] = 0;
2437 migrate->cpages--;
2438 put_page(page);
2439 continue;
2440 }
2441 remap = false;
2442 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2443 }
2444
2445 /* ZONE_DEVICE pages are not on LRU */
2446 if (!is_zone_device_page(page)) {
2447 if (!PageLRU(page) && allow_drain) {
2448 /* Drain CPU's pagevec */
2449 lru_add_drain_all();
2450 allow_drain = false;
2451 }
2452
2453 if (isolate_lru_page(page)) {
2454 if (remap) {
2455 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2456 migrate->cpages--;
2457 restore++;
2458 } else {
2459 migrate->src[i] = 0;
2460 unlock_page(page);
2461 migrate->cpages--;
2462 put_page(page);
2463 }
2464 continue;
2465 }
2466
2467 /* Drop the reference we took in collect */
2468 put_page(page);
2469 }
2470
2471 if (!migrate_vma_check_page(page)) {
2472 if (remap) {
2473 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2474 migrate->cpages--;
2475 restore++;
2476
2477 if (!is_zone_device_page(page)) {
2478 get_page(page);
2479 putback_lru_page(page);
2480 }
2481 } else {
2482 migrate->src[i] = 0;
2483 unlock_page(page);
2484 migrate->cpages--;
2485
2486 if (!is_zone_device_page(page))
2487 putback_lru_page(page);
2488 else
2489 put_page(page);
2490 }
2491 }
2492 }
2493
2494 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2495 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2496
2497 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2498 continue;
2499
2500 remove_migration_pte(page, migrate->vma, addr, page);
2501
2502 migrate->src[i] = 0;
2503 unlock_page(page);
2504 put_page(page);
2505 restore--;
2506 }
2507}
2508
2509/*
2510 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2511 * @migrate: migrate struct containing all migration information
2512 *
2513 * Replace page mapping (CPU page table pte) with a special migration pte entry
2514 * and check again if it has been pinned. Pinned pages are restored because we
2515 * cannot migrate them.
2516 *
2517 * This is the last step before we call the device driver callback to allocate
2518 * destination memory and copy contents of original page over to new page.
2519 */
2520static void migrate_vma_unmap(struct migrate_vma *migrate)
2521{
2522 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2523 const unsigned long npages = migrate->npages;
2524 const unsigned long start = migrate->start;
2525 unsigned long addr, i, restore = 0;
2526
2527 for (i = 0; i < npages; i++) {
2528 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2529
2530 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2531 continue;
2532
2533 if (page_mapped(page)) {
2534 try_to_unmap(page, flags);
2535 if (page_mapped(page))
2536 goto restore;
2537 }
2538
2539 if (migrate_vma_check_page(page))
2540 continue;
2541
2542restore:
2543 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2544 migrate->cpages--;
2545 restore++;
2546 }
2547
2548 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2549 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2550
2551 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2552 continue;
2553
2554 remove_migration_ptes(page, page, false);
2555
2556 migrate->src[i] = 0;
2557 unlock_page(page);
2558 restore--;
2559
2560 if (is_zone_device_page(page))
2561 put_page(page);
2562 else
2563 putback_lru_page(page);
2564 }
2565}
2566
2567/**
2568 * migrate_vma_setup() - prepare to migrate a range of memory
2569 * @args: contains the vma, start, and and pfns arrays for the migration
2570 *
2571 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2572 * without an error.
2573 *
2574 * Prepare to migrate a range of memory virtual address range by collecting all
2575 * the pages backing each virtual address in the range, saving them inside the
2576 * src array. Then lock those pages and unmap them. Once the pages are locked
2577 * and unmapped, check whether each page is pinned or not. Pages that aren't
2578 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2579 * corresponding src array entry. Then restores any pages that are pinned, by
2580 * remapping and unlocking those pages.
2581 *
2582 * The caller should then allocate destination memory and copy source memory to
2583 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2584 * flag set). Once these are allocated and copied, the caller must update each
2585 * corresponding entry in the dst array with the pfn value of the destination
2586 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2587 * (destination pages must have their struct pages locked, via lock_page()).
2588 *
2589 * Note that the caller does not have to migrate all the pages that are marked
2590 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2591 * device memory to system memory. If the caller cannot migrate a device page
2592 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2593 * consequences for the userspace process, so it must be avoided if at all
2594 * possible.
2595 *
2596 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2597 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2598 * allowing the caller to allocate device memory for those unback virtual
2599 * address. For this the caller simply has to allocate device memory and
2600 * properly set the destination entry like for regular migration. Note that
2601 * this can still fails and thus inside the device driver must check if the
2602 * migration was successful for those entries after calling migrate_vma_pages()
2603 * just like for regular migration.
2604 *
2605 * After that, the callers must call migrate_vma_pages() to go over each entry
2606 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2607 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2608 * then migrate_vma_pages() to migrate struct page information from the source
2609 * struct page to the destination struct page. If it fails to migrate the
2610 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2611 * src array.
2612 *
2613 * At this point all successfully migrated pages have an entry in the src
2614 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2615 * array entry with MIGRATE_PFN_VALID flag set.
2616 *
2617 * Once migrate_vma_pages() returns the caller may inspect which pages were
2618 * successfully migrated, and which were not. Successfully migrated pages will
2619 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2620 *
2621 * It is safe to update device page table after migrate_vma_pages() because
2622 * both destination and source page are still locked, and the mmap_sem is held
2623 * in read mode (hence no one can unmap the range being migrated).
2624 *
2625 * Once the caller is done cleaning up things and updating its page table (if it
2626 * chose to do so, this is not an obligation) it finally calls
2627 * migrate_vma_finalize() to update the CPU page table to point to new pages
2628 * for successfully migrated pages or otherwise restore the CPU page table to
2629 * point to the original source pages.
2630 */
2631int migrate_vma_setup(struct migrate_vma *args)
2632{
2633 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2634
2635 args->start &= PAGE_MASK;
2636 args->end &= PAGE_MASK;
2637 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2638 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2639 return -EINVAL;
2640 if (nr_pages <= 0)
2641 return -EINVAL;
2642 if (args->start < args->vma->vm_start ||
2643 args->start >= args->vma->vm_end)
2644 return -EINVAL;
2645 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2646 return -EINVAL;
2647 if (!args->src || !args->dst)
2648 return -EINVAL;
2649
2650 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2651 args->cpages = 0;
2652 args->npages = 0;
2653
2654 migrate_vma_collect(args);
2655
2656 if (args->cpages)
2657 migrate_vma_prepare(args);
2658 if (args->cpages)
2659 migrate_vma_unmap(args);
2660
2661 /*
2662 * At this point pages are locked and unmapped, and thus they have
2663 * stable content and can safely be copied to destination memory that
2664 * is allocated by the drivers.
2665 */
2666 return 0;
2667
2668}
2669EXPORT_SYMBOL(migrate_vma_setup);
2670
2671static void migrate_vma_insert_page(struct migrate_vma *migrate,
2672 unsigned long addr,
2673 struct page *page,
2674 unsigned long *src,
2675 unsigned long *dst)
2676{
2677 struct vm_area_struct *vma = migrate->vma;
2678 struct mm_struct *mm = vma->vm_mm;
2679 struct mem_cgroup *memcg;
2680 bool flush = false;
2681 spinlock_t *ptl;
2682 pte_t entry;
2683 pgd_t *pgdp;
2684 p4d_t *p4dp;
2685 pud_t *pudp;
2686 pmd_t *pmdp;
2687 pte_t *ptep;
2688
2689 /* Only allow populating anonymous memory */
2690 if (!vma_is_anonymous(vma))
2691 goto abort;
2692
2693 pgdp = pgd_offset(mm, addr);
2694 p4dp = p4d_alloc(mm, pgdp, addr);
2695 if (!p4dp)
2696 goto abort;
2697 pudp = pud_alloc(mm, p4dp, addr);
2698 if (!pudp)
2699 goto abort;
2700 pmdp = pmd_alloc(mm, pudp, addr);
2701 if (!pmdp)
2702 goto abort;
2703
2704 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2705 goto abort;
2706
2707 /*
2708 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2709 * pte_offset_map() on pmds where a huge pmd might be created
2710 * from a different thread.
2711 *
2712 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2713 * parallel threads are excluded by other means.
2714 *
2715 * Here we only have down_read(mmap_sem).
2716 */
2717 if (pte_alloc(mm, pmdp))
2718 goto abort;
2719
2720 /* See the comment in pte_alloc_one_map() */
2721 if (unlikely(pmd_trans_unstable(pmdp)))
2722 goto abort;
2723
2724 if (unlikely(anon_vma_prepare(vma)))
2725 goto abort;
2726 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2727 goto abort;
2728
2729 /*
2730 * The memory barrier inside __SetPageUptodate makes sure that
2731 * preceding stores to the page contents become visible before
2732 * the set_pte_at() write.
2733 */
2734 __SetPageUptodate(page);
2735
2736 if (is_zone_device_page(page)) {
2737 if (is_device_private_page(page)) {
2738 swp_entry_t swp_entry;
2739
2740 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2741 entry = swp_entry_to_pte(swp_entry);
2742 }
2743 } else {
2744 entry = mk_pte(page, vma->vm_page_prot);
2745 if (vma->vm_flags & VM_WRITE)
2746 entry = pte_mkwrite(pte_mkdirty(entry));
2747 }
2748
2749 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2750
2751 if (pte_present(*ptep)) {
2752 unsigned long pfn = pte_pfn(*ptep);
2753
2754 if (!is_zero_pfn(pfn)) {
2755 pte_unmap_unlock(ptep, ptl);
2756 mem_cgroup_cancel_charge(page, memcg, false);
2757 goto abort;
2758 }
2759 flush = true;
2760 } else if (!pte_none(*ptep)) {
2761 pte_unmap_unlock(ptep, ptl);
2762 mem_cgroup_cancel_charge(page, memcg, false);
2763 goto abort;
2764 }
2765
2766 /*
2767 * Check for usefaultfd but do not deliver the fault. Instead,
2768 * just back off.
2769 */
2770 if (userfaultfd_missing(vma)) {
2771 pte_unmap_unlock(ptep, ptl);
2772 mem_cgroup_cancel_charge(page, memcg, false);
2773 goto abort;
2774 }
2775
2776 inc_mm_counter(mm, MM_ANONPAGES);
2777 page_add_new_anon_rmap(page, vma, addr, false);
2778 mem_cgroup_commit_charge(page, memcg, false, false);
2779 if (!is_zone_device_page(page))
2780 lru_cache_add_active_or_unevictable(page, vma);
2781 get_page(page);
2782
2783 if (flush) {
2784 flush_cache_page(vma, addr, pte_pfn(*ptep));
2785 ptep_clear_flush_notify(vma, addr, ptep);
2786 set_pte_at_notify(mm, addr, ptep, entry);
2787 update_mmu_cache(vma, addr, ptep);
2788 } else {
2789 /* No need to invalidate - it was non-present before */
2790 set_pte_at(mm, addr, ptep, entry);
2791 update_mmu_cache(vma, addr, ptep);
2792 }
2793
2794 pte_unmap_unlock(ptep, ptl);
2795 *src = MIGRATE_PFN_MIGRATE;
2796 return;
2797
2798abort:
2799 *src &= ~MIGRATE_PFN_MIGRATE;
2800}
2801
2802/**
2803 * migrate_vma_pages() - migrate meta-data from src page to dst page
2804 * @migrate: migrate struct containing all migration information
2805 *
2806 * This migrates struct page meta-data from source struct page to destination
2807 * struct page. This effectively finishes the migration from source page to the
2808 * destination page.
2809 */
2810void migrate_vma_pages(struct migrate_vma *migrate)
2811{
2812 const unsigned long npages = migrate->npages;
2813 const unsigned long start = migrate->start;
2814 struct mmu_notifier_range range;
2815 unsigned long addr, i;
2816 bool notified = false;
2817
2818 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2819 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2820 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2821 struct address_space *mapping;
2822 int r;
2823
2824 if (!newpage) {
2825 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2826 continue;
2827 }
2828
2829 if (!page) {
2830 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2831 continue;
2832 }
2833 if (!notified) {
2834 notified = true;
2835
2836 mmu_notifier_range_init(&range,
2837 MMU_NOTIFY_CLEAR, 0,
2838 NULL,
2839 migrate->vma->vm_mm,
2840 addr, migrate->end);
2841 mmu_notifier_invalidate_range_start(&range);
2842 }
2843 migrate_vma_insert_page(migrate, addr, newpage,
2844 &migrate->src[i],
2845 &migrate->dst[i]);
2846 continue;
2847 }
2848
2849 mapping = page_mapping(page);
2850
2851 if (is_zone_device_page(newpage)) {
2852 if (is_device_private_page(newpage)) {
2853 /*
2854 * For now only support private anonymous when
2855 * migrating to un-addressable device memory.
2856 */
2857 if (mapping) {
2858 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2859 continue;
2860 }
2861 } else {
2862 /*
2863 * Other types of ZONE_DEVICE page are not
2864 * supported.
2865 */
2866 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2867 continue;
2868 }
2869 }
2870
2871 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2872 if (r != MIGRATEPAGE_SUCCESS)
2873 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2874 }
2875
2876 /*
2877 * No need to double call mmu_notifier->invalidate_range() callback as
2878 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2879 * did already call it.
2880 */
2881 if (notified)
2882 mmu_notifier_invalidate_range_only_end(&range);
2883}
2884EXPORT_SYMBOL(migrate_vma_pages);
2885
2886/**
2887 * migrate_vma_finalize() - restore CPU page table entry
2888 * @migrate: migrate struct containing all migration information
2889 *
2890 * This replaces the special migration pte entry with either a mapping to the
2891 * new page if migration was successful for that page, or to the original page
2892 * otherwise.
2893 *
2894 * This also unlocks the pages and puts them back on the lru, or drops the extra
2895 * refcount, for device pages.
2896 */
2897void migrate_vma_finalize(struct migrate_vma *migrate)
2898{
2899 const unsigned long npages = migrate->npages;
2900 unsigned long i;
2901
2902 for (i = 0; i < npages; i++) {
2903 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2904 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2905
2906 if (!page) {
2907 if (newpage) {
2908 unlock_page(newpage);
2909 put_page(newpage);
2910 }
2911 continue;
2912 }
2913
2914 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2915 if (newpage) {
2916 unlock_page(newpage);
2917 put_page(newpage);
2918 }
2919 newpage = page;
2920 }
2921
2922 remove_migration_ptes(page, newpage, false);
2923 unlock_page(page);
2924 migrate->cpages--;
2925
2926 if (is_zone_device_page(page))
2927 put_page(page);
2928 else
2929 putback_lru_page(page);
2930
2931 if (newpage != page) {
2932 unlock_page(newpage);
2933 if (is_zone_device_page(newpage))
2934 put_page(newpage);
2935 else
2936 putback_lru_page(newpage);
2937 }
2938 }
2939}
2940EXPORT_SYMBOL(migrate_vma_finalize);
2941#endif /* CONFIG_DEVICE_PRIVATE */
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter
13 */
14
15#include <linux/migrate.h>
16#include <linux/export.h>
17#include <linux/swap.h>
18#include <linux/swapops.h>
19#include <linux/pagemap.h>
20#include <linux/buffer_head.h>
21#include <linux/mm_inline.h>
22#include <linux/nsproxy.h>
23#include <linux/pagevec.h>
24#include <linux/ksm.h>
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
29#include <linux/writeback.h>
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
32#include <linux/security.h>
33#include <linux/memcontrol.h>
34#include <linux/syscalls.h>
35#include <linux/hugetlb.h>
36#include <linux/gfp.h>
37
38#include <asm/tlbflush.h>
39
40#include "internal.h"
41
42/*
43 * migrate_prep() needs to be called before we start compiling a list of pages
44 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
45 * undesirable, use migrate_prep_local()
46 */
47int migrate_prep(void)
48{
49 /*
50 * Clear the LRU lists so pages can be isolated.
51 * Note that pages may be moved off the LRU after we have
52 * drained them. Those pages will fail to migrate like other
53 * pages that may be busy.
54 */
55 lru_add_drain_all();
56
57 return 0;
58}
59
60/* Do the necessary work of migrate_prep but not if it involves other CPUs */
61int migrate_prep_local(void)
62{
63 lru_add_drain();
64
65 return 0;
66}
67
68/*
69 * Add isolated pages on the list back to the LRU under page lock
70 * to avoid leaking evictable pages back onto unevictable list.
71 */
72void putback_lru_pages(struct list_head *l)
73{
74 struct page *page;
75 struct page *page2;
76
77 list_for_each_entry_safe(page, page2, l, lru) {
78 list_del(&page->lru);
79 dec_zone_page_state(page, NR_ISOLATED_ANON +
80 page_is_file_cache(page));
81 putback_lru_page(page);
82 }
83}
84
85/*
86 * Restore a potential migration pte to a working pte entry
87 */
88static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
89 unsigned long addr, void *old)
90{
91 struct mm_struct *mm = vma->vm_mm;
92 swp_entry_t entry;
93 pgd_t *pgd;
94 pud_t *pud;
95 pmd_t *pmd;
96 pte_t *ptep, pte;
97 spinlock_t *ptl;
98
99 if (unlikely(PageHuge(new))) {
100 ptep = huge_pte_offset(mm, addr);
101 if (!ptep)
102 goto out;
103 ptl = &mm->page_table_lock;
104 } else {
105 pgd = pgd_offset(mm, addr);
106 if (!pgd_present(*pgd))
107 goto out;
108
109 pud = pud_offset(pgd, addr);
110 if (!pud_present(*pud))
111 goto out;
112
113 pmd = pmd_offset(pud, addr);
114 if (pmd_trans_huge(*pmd))
115 goto out;
116 if (!pmd_present(*pmd))
117 goto out;
118
119 ptep = pte_offset_map(pmd, addr);
120
121 /*
122 * Peek to check is_swap_pte() before taking ptlock? No, we
123 * can race mremap's move_ptes(), which skips anon_vma lock.
124 */
125
126 ptl = pte_lockptr(mm, pmd);
127 }
128
129 spin_lock(ptl);
130 pte = *ptep;
131 if (!is_swap_pte(pte))
132 goto unlock;
133
134 entry = pte_to_swp_entry(pte);
135
136 if (!is_migration_entry(entry) ||
137 migration_entry_to_page(entry) != old)
138 goto unlock;
139
140 get_page(new);
141 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
142 if (is_write_migration_entry(entry))
143 pte = pte_mkwrite(pte);
144#ifdef CONFIG_HUGETLB_PAGE
145 if (PageHuge(new))
146 pte = pte_mkhuge(pte);
147#endif
148 flush_cache_page(vma, addr, pte_pfn(pte));
149 set_pte_at(mm, addr, ptep, pte);
150
151 if (PageHuge(new)) {
152 if (PageAnon(new))
153 hugepage_add_anon_rmap(new, vma, addr);
154 else
155 page_dup_rmap(new);
156 } else if (PageAnon(new))
157 page_add_anon_rmap(new, vma, addr);
158 else
159 page_add_file_rmap(new);
160
161 /* No need to invalidate - it was non-present before */
162 update_mmu_cache(vma, addr, ptep);
163unlock:
164 pte_unmap_unlock(ptep, ptl);
165out:
166 return SWAP_AGAIN;
167}
168
169/*
170 * Get rid of all migration entries and replace them by
171 * references to the indicated page.
172 */
173static void remove_migration_ptes(struct page *old, struct page *new)
174{
175 rmap_walk(new, remove_migration_pte, old);
176}
177
178/*
179 * Something used the pte of a page under migration. We need to
180 * get to the page and wait until migration is finished.
181 * When we return from this function the fault will be retried.
182 */
183void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
184 unsigned long address)
185{
186 pte_t *ptep, pte;
187 spinlock_t *ptl;
188 swp_entry_t entry;
189 struct page *page;
190
191 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
192 pte = *ptep;
193 if (!is_swap_pte(pte))
194 goto out;
195
196 entry = pte_to_swp_entry(pte);
197 if (!is_migration_entry(entry))
198 goto out;
199
200 page = migration_entry_to_page(entry);
201
202 /*
203 * Once radix-tree replacement of page migration started, page_count
204 * *must* be zero. And, we don't want to call wait_on_page_locked()
205 * against a page without get_page().
206 * So, we use get_page_unless_zero(), here. Even failed, page fault
207 * will occur again.
208 */
209 if (!get_page_unless_zero(page))
210 goto out;
211 pte_unmap_unlock(ptep, ptl);
212 wait_on_page_locked(page);
213 put_page(page);
214 return;
215out:
216 pte_unmap_unlock(ptep, ptl);
217}
218
219#ifdef CONFIG_BLOCK
220/* Returns true if all buffers are successfully locked */
221static bool buffer_migrate_lock_buffers(struct buffer_head *head,
222 enum migrate_mode mode)
223{
224 struct buffer_head *bh = head;
225
226 /* Simple case, sync compaction */
227 if (mode != MIGRATE_ASYNC) {
228 do {
229 get_bh(bh);
230 lock_buffer(bh);
231 bh = bh->b_this_page;
232
233 } while (bh != head);
234
235 return true;
236 }
237
238 /* async case, we cannot block on lock_buffer so use trylock_buffer */
239 do {
240 get_bh(bh);
241 if (!trylock_buffer(bh)) {
242 /*
243 * We failed to lock the buffer and cannot stall in
244 * async migration. Release the taken locks
245 */
246 struct buffer_head *failed_bh = bh;
247 put_bh(failed_bh);
248 bh = head;
249 while (bh != failed_bh) {
250 unlock_buffer(bh);
251 put_bh(bh);
252 bh = bh->b_this_page;
253 }
254 return false;
255 }
256
257 bh = bh->b_this_page;
258 } while (bh != head);
259 return true;
260}
261#else
262static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
263 enum migrate_mode mode)
264{
265 return true;
266}
267#endif /* CONFIG_BLOCK */
268
269/*
270 * Replace the page in the mapping.
271 *
272 * The number of remaining references must be:
273 * 1 for anonymous pages without a mapping
274 * 2 for pages with a mapping
275 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
276 */
277static int migrate_page_move_mapping(struct address_space *mapping,
278 struct page *newpage, struct page *page,
279 struct buffer_head *head, enum migrate_mode mode)
280{
281 int expected_count;
282 void **pslot;
283
284 if (!mapping) {
285 /* Anonymous page without mapping */
286 if (page_count(page) != 1)
287 return -EAGAIN;
288 return 0;
289 }
290
291 spin_lock_irq(&mapping->tree_lock);
292
293 pslot = radix_tree_lookup_slot(&mapping->page_tree,
294 page_index(page));
295
296 expected_count = 2 + page_has_private(page);
297 if (page_count(page) != expected_count ||
298 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
299 spin_unlock_irq(&mapping->tree_lock);
300 return -EAGAIN;
301 }
302
303 if (!page_freeze_refs(page, expected_count)) {
304 spin_unlock_irq(&mapping->tree_lock);
305 return -EAGAIN;
306 }
307
308 /*
309 * In the async migration case of moving a page with buffers, lock the
310 * buffers using trylock before the mapping is moved. If the mapping
311 * was moved, we later failed to lock the buffers and could not move
312 * the mapping back due to an elevated page count, we would have to
313 * block waiting on other references to be dropped.
314 */
315 if (mode == MIGRATE_ASYNC && head &&
316 !buffer_migrate_lock_buffers(head, mode)) {
317 page_unfreeze_refs(page, expected_count);
318 spin_unlock_irq(&mapping->tree_lock);
319 return -EAGAIN;
320 }
321
322 /*
323 * Now we know that no one else is looking at the page.
324 */
325 get_page(newpage); /* add cache reference */
326 if (PageSwapCache(page)) {
327 SetPageSwapCache(newpage);
328 set_page_private(newpage, page_private(page));
329 }
330
331 radix_tree_replace_slot(pslot, newpage);
332
333 /*
334 * Drop cache reference from old page by unfreezing
335 * to one less reference.
336 * We know this isn't the last reference.
337 */
338 page_unfreeze_refs(page, expected_count - 1);
339
340 /*
341 * If moved to a different zone then also account
342 * the page for that zone. Other VM counters will be
343 * taken care of when we establish references to the
344 * new page and drop references to the old page.
345 *
346 * Note that anonymous pages are accounted for
347 * via NR_FILE_PAGES and NR_ANON_PAGES if they
348 * are mapped to swap space.
349 */
350 __dec_zone_page_state(page, NR_FILE_PAGES);
351 __inc_zone_page_state(newpage, NR_FILE_PAGES);
352 if (!PageSwapCache(page) && PageSwapBacked(page)) {
353 __dec_zone_page_state(page, NR_SHMEM);
354 __inc_zone_page_state(newpage, NR_SHMEM);
355 }
356 spin_unlock_irq(&mapping->tree_lock);
357
358 return 0;
359}
360
361/*
362 * The expected number of remaining references is the same as that
363 * of migrate_page_move_mapping().
364 */
365int migrate_huge_page_move_mapping(struct address_space *mapping,
366 struct page *newpage, struct page *page)
367{
368 int expected_count;
369 void **pslot;
370
371 if (!mapping) {
372 if (page_count(page) != 1)
373 return -EAGAIN;
374 return 0;
375 }
376
377 spin_lock_irq(&mapping->tree_lock);
378
379 pslot = radix_tree_lookup_slot(&mapping->page_tree,
380 page_index(page));
381
382 expected_count = 2 + page_has_private(page);
383 if (page_count(page) != expected_count ||
384 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
385 spin_unlock_irq(&mapping->tree_lock);
386 return -EAGAIN;
387 }
388
389 if (!page_freeze_refs(page, expected_count)) {
390 spin_unlock_irq(&mapping->tree_lock);
391 return -EAGAIN;
392 }
393
394 get_page(newpage);
395
396 radix_tree_replace_slot(pslot, newpage);
397
398 page_unfreeze_refs(page, expected_count - 1);
399
400 spin_unlock_irq(&mapping->tree_lock);
401 return 0;
402}
403
404/*
405 * Copy the page to its new location
406 */
407void migrate_page_copy(struct page *newpage, struct page *page)
408{
409 if (PageHuge(page))
410 copy_huge_page(newpage, page);
411 else
412 copy_highpage(newpage, page);
413
414 if (PageError(page))
415 SetPageError(newpage);
416 if (PageReferenced(page))
417 SetPageReferenced(newpage);
418 if (PageUptodate(page))
419 SetPageUptodate(newpage);
420 if (TestClearPageActive(page)) {
421 VM_BUG_ON(PageUnevictable(page));
422 SetPageActive(newpage);
423 } else if (TestClearPageUnevictable(page))
424 SetPageUnevictable(newpage);
425 if (PageChecked(page))
426 SetPageChecked(newpage);
427 if (PageMappedToDisk(page))
428 SetPageMappedToDisk(newpage);
429
430 if (PageDirty(page)) {
431 clear_page_dirty_for_io(page);
432 /*
433 * Want to mark the page and the radix tree as dirty, and
434 * redo the accounting that clear_page_dirty_for_io undid,
435 * but we can't use set_page_dirty because that function
436 * is actually a signal that all of the page has become dirty.
437 * Whereas only part of our page may be dirty.
438 */
439 if (PageSwapBacked(page))
440 SetPageDirty(newpage);
441 else
442 __set_page_dirty_nobuffers(newpage);
443 }
444
445 mlock_migrate_page(newpage, page);
446 ksm_migrate_page(newpage, page);
447
448 ClearPageSwapCache(page);
449 ClearPagePrivate(page);
450 set_page_private(page, 0);
451
452 /*
453 * If any waiters have accumulated on the new page then
454 * wake them up.
455 */
456 if (PageWriteback(newpage))
457 end_page_writeback(newpage);
458}
459
460/************************************************************
461 * Migration functions
462 ***********************************************************/
463
464/* Always fail migration. Used for mappings that are not movable */
465int fail_migrate_page(struct address_space *mapping,
466 struct page *newpage, struct page *page)
467{
468 return -EIO;
469}
470EXPORT_SYMBOL(fail_migrate_page);
471
472/*
473 * Common logic to directly migrate a single page suitable for
474 * pages that do not use PagePrivate/PagePrivate2.
475 *
476 * Pages are locked upon entry and exit.
477 */
478int migrate_page(struct address_space *mapping,
479 struct page *newpage, struct page *page,
480 enum migrate_mode mode)
481{
482 int rc;
483
484 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
485
486 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
487
488 if (rc)
489 return rc;
490
491 migrate_page_copy(newpage, page);
492 return 0;
493}
494EXPORT_SYMBOL(migrate_page);
495
496#ifdef CONFIG_BLOCK
497/*
498 * Migration function for pages with buffers. This function can only be used
499 * if the underlying filesystem guarantees that no other references to "page"
500 * exist.
501 */
502int buffer_migrate_page(struct address_space *mapping,
503 struct page *newpage, struct page *page, enum migrate_mode mode)
504{
505 struct buffer_head *bh, *head;
506 int rc;
507
508 if (!page_has_buffers(page))
509 return migrate_page(mapping, newpage, page, mode);
510
511 head = page_buffers(page);
512
513 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
514
515 if (rc)
516 return rc;
517
518 /*
519 * In the async case, migrate_page_move_mapping locked the buffers
520 * with an IRQ-safe spinlock held. In the sync case, the buffers
521 * need to be locked now
522 */
523 if (mode != MIGRATE_ASYNC)
524 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
525
526 ClearPagePrivate(page);
527 set_page_private(newpage, page_private(page));
528 set_page_private(page, 0);
529 put_page(page);
530 get_page(newpage);
531
532 bh = head;
533 do {
534 set_bh_page(bh, newpage, bh_offset(bh));
535 bh = bh->b_this_page;
536
537 } while (bh != head);
538
539 SetPagePrivate(newpage);
540
541 migrate_page_copy(newpage, page);
542
543 bh = head;
544 do {
545 unlock_buffer(bh);
546 put_bh(bh);
547 bh = bh->b_this_page;
548
549 } while (bh != head);
550
551 return 0;
552}
553EXPORT_SYMBOL(buffer_migrate_page);
554#endif
555
556/*
557 * Writeback a page to clean the dirty state
558 */
559static int writeout(struct address_space *mapping, struct page *page)
560{
561 struct writeback_control wbc = {
562 .sync_mode = WB_SYNC_NONE,
563 .nr_to_write = 1,
564 .range_start = 0,
565 .range_end = LLONG_MAX,
566 .for_reclaim = 1
567 };
568 int rc;
569
570 if (!mapping->a_ops->writepage)
571 /* No write method for the address space */
572 return -EINVAL;
573
574 if (!clear_page_dirty_for_io(page))
575 /* Someone else already triggered a write */
576 return -EAGAIN;
577
578 /*
579 * A dirty page may imply that the underlying filesystem has
580 * the page on some queue. So the page must be clean for
581 * migration. Writeout may mean we loose the lock and the
582 * page state is no longer what we checked for earlier.
583 * At this point we know that the migration attempt cannot
584 * be successful.
585 */
586 remove_migration_ptes(page, page);
587
588 rc = mapping->a_ops->writepage(page, &wbc);
589
590 if (rc != AOP_WRITEPAGE_ACTIVATE)
591 /* unlocked. Relock */
592 lock_page(page);
593
594 return (rc < 0) ? -EIO : -EAGAIN;
595}
596
597/*
598 * Default handling if a filesystem does not provide a migration function.
599 */
600static int fallback_migrate_page(struct address_space *mapping,
601 struct page *newpage, struct page *page, enum migrate_mode mode)
602{
603 if (PageDirty(page)) {
604 /* Only writeback pages in full synchronous migration */
605 if (mode != MIGRATE_SYNC)
606 return -EBUSY;
607 return writeout(mapping, page);
608 }
609
610 /*
611 * Buffers may be managed in a filesystem specific way.
612 * We must have no buffers or drop them.
613 */
614 if (page_has_private(page) &&
615 !try_to_release_page(page, GFP_KERNEL))
616 return -EAGAIN;
617
618 return migrate_page(mapping, newpage, page, mode);
619}
620
621/*
622 * Move a page to a newly allocated page
623 * The page is locked and all ptes have been successfully removed.
624 *
625 * The new page will have replaced the old page if this function
626 * is successful.
627 *
628 * Return value:
629 * < 0 - error code
630 * == 0 - success
631 */
632static int move_to_new_page(struct page *newpage, struct page *page,
633 int remap_swapcache, enum migrate_mode mode)
634{
635 struct address_space *mapping;
636 int rc;
637
638 /*
639 * Block others from accessing the page when we get around to
640 * establishing additional references. We are the only one
641 * holding a reference to the new page at this point.
642 */
643 if (!trylock_page(newpage))
644 BUG();
645
646 /* Prepare mapping for the new page.*/
647 newpage->index = page->index;
648 newpage->mapping = page->mapping;
649 if (PageSwapBacked(page))
650 SetPageSwapBacked(newpage);
651
652 mapping = page_mapping(page);
653 if (!mapping)
654 rc = migrate_page(mapping, newpage, page, mode);
655 else if (mapping->a_ops->migratepage)
656 /*
657 * Most pages have a mapping and most filesystems provide a
658 * migratepage callback. Anonymous pages are part of swap
659 * space which also has its own migratepage callback. This
660 * is the most common path for page migration.
661 */
662 rc = mapping->a_ops->migratepage(mapping,
663 newpage, page, mode);
664 else
665 rc = fallback_migrate_page(mapping, newpage, page, mode);
666
667 if (rc) {
668 newpage->mapping = NULL;
669 } else {
670 if (remap_swapcache)
671 remove_migration_ptes(page, newpage);
672 page->mapping = NULL;
673 }
674
675 unlock_page(newpage);
676
677 return rc;
678}
679
680static int __unmap_and_move(struct page *page, struct page *newpage,
681 int force, bool offlining, enum migrate_mode mode)
682{
683 int rc = -EAGAIN;
684 int remap_swapcache = 1;
685 int charge = 0;
686 struct mem_cgroup *mem;
687 struct anon_vma *anon_vma = NULL;
688
689 if (!trylock_page(page)) {
690 if (!force || mode == MIGRATE_ASYNC)
691 goto out;
692
693 /*
694 * It's not safe for direct compaction to call lock_page.
695 * For example, during page readahead pages are added locked
696 * to the LRU. Later, when the IO completes the pages are
697 * marked uptodate and unlocked. However, the queueing
698 * could be merging multiple pages for one bio (e.g.
699 * mpage_readpages). If an allocation happens for the
700 * second or third page, the process can end up locking
701 * the same page twice and deadlocking. Rather than
702 * trying to be clever about what pages can be locked,
703 * avoid the use of lock_page for direct compaction
704 * altogether.
705 */
706 if (current->flags & PF_MEMALLOC)
707 goto out;
708
709 lock_page(page);
710 }
711
712 /*
713 * Only memory hotplug's offline_pages() caller has locked out KSM,
714 * and can safely migrate a KSM page. The other cases have skipped
715 * PageKsm along with PageReserved - but it is only now when we have
716 * the page lock that we can be certain it will not go KSM beneath us
717 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
718 * its pagecount raised, but only here do we take the page lock which
719 * serializes that).
720 */
721 if (PageKsm(page) && !offlining) {
722 rc = -EBUSY;
723 goto unlock;
724 }
725
726 /* charge against new page */
727 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
728 if (charge == -ENOMEM) {
729 rc = -ENOMEM;
730 goto unlock;
731 }
732 BUG_ON(charge);
733
734 if (PageWriteback(page)) {
735 /*
736 * Only in the case of a full syncronous migration is it
737 * necessary to wait for PageWriteback. In the async case,
738 * the retry loop is too short and in the sync-light case,
739 * the overhead of stalling is too much
740 */
741 if (mode != MIGRATE_SYNC) {
742 rc = -EBUSY;
743 goto uncharge;
744 }
745 if (!force)
746 goto uncharge;
747 wait_on_page_writeback(page);
748 }
749 /*
750 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
751 * we cannot notice that anon_vma is freed while we migrates a page.
752 * This get_anon_vma() delays freeing anon_vma pointer until the end
753 * of migration. File cache pages are no problem because of page_lock()
754 * File Caches may use write_page() or lock_page() in migration, then,
755 * just care Anon page here.
756 */
757 if (PageAnon(page)) {
758 /*
759 * Only page_lock_anon_vma() understands the subtleties of
760 * getting a hold on an anon_vma from outside one of its mms.
761 */
762 anon_vma = page_get_anon_vma(page);
763 if (anon_vma) {
764 /*
765 * Anon page
766 */
767 } else if (PageSwapCache(page)) {
768 /*
769 * We cannot be sure that the anon_vma of an unmapped
770 * swapcache page is safe to use because we don't
771 * know in advance if the VMA that this page belonged
772 * to still exists. If the VMA and others sharing the
773 * data have been freed, then the anon_vma could
774 * already be invalid.
775 *
776 * To avoid this possibility, swapcache pages get
777 * migrated but are not remapped when migration
778 * completes
779 */
780 remap_swapcache = 0;
781 } else {
782 goto uncharge;
783 }
784 }
785
786 /*
787 * Corner case handling:
788 * 1. When a new swap-cache page is read into, it is added to the LRU
789 * and treated as swapcache but it has no rmap yet.
790 * Calling try_to_unmap() against a page->mapping==NULL page will
791 * trigger a BUG. So handle it here.
792 * 2. An orphaned page (see truncate_complete_page) might have
793 * fs-private metadata. The page can be picked up due to memory
794 * offlining. Everywhere else except page reclaim, the page is
795 * invisible to the vm, so the page can not be migrated. So try to
796 * free the metadata, so the page can be freed.
797 */
798 if (!page->mapping) {
799 VM_BUG_ON(PageAnon(page));
800 if (page_has_private(page)) {
801 try_to_free_buffers(page);
802 goto uncharge;
803 }
804 goto skip_unmap;
805 }
806
807 /* Establish migration ptes or remove ptes */
808 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
809
810skip_unmap:
811 if (!page_mapped(page))
812 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
813
814 if (rc && remap_swapcache)
815 remove_migration_ptes(page, page);
816
817 /* Drop an anon_vma reference if we took one */
818 if (anon_vma)
819 put_anon_vma(anon_vma);
820
821uncharge:
822 if (!charge)
823 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
824unlock:
825 unlock_page(page);
826out:
827 return rc;
828}
829
830/*
831 * Obtain the lock on page, remove all ptes and migrate the page
832 * to the newly allocated page in newpage.
833 */
834static int unmap_and_move(new_page_t get_new_page, unsigned long private,
835 struct page *page, int force, bool offlining,
836 enum migrate_mode mode)
837{
838 int rc = 0;
839 int *result = NULL;
840 struct page *newpage = get_new_page(page, private, &result);
841
842 if (!newpage)
843 return -ENOMEM;
844
845 if (page_count(page) == 1) {
846 /* page was freed from under us. So we are done. */
847 goto out;
848 }
849
850 if (unlikely(PageTransHuge(page)))
851 if (unlikely(split_huge_page(page)))
852 goto out;
853
854 rc = __unmap_and_move(page, newpage, force, offlining, mode);
855out:
856 if (rc != -EAGAIN) {
857 /*
858 * A page that has been migrated has all references
859 * removed and will be freed. A page that has not been
860 * migrated will have kepts its references and be
861 * restored.
862 */
863 list_del(&page->lru);
864 dec_zone_page_state(page, NR_ISOLATED_ANON +
865 page_is_file_cache(page));
866 putback_lru_page(page);
867 }
868 /*
869 * Move the new page to the LRU. If migration was not successful
870 * then this will free the page.
871 */
872 putback_lru_page(newpage);
873 if (result) {
874 if (rc)
875 *result = rc;
876 else
877 *result = page_to_nid(newpage);
878 }
879 return rc;
880}
881
882/*
883 * Counterpart of unmap_and_move_page() for hugepage migration.
884 *
885 * This function doesn't wait the completion of hugepage I/O
886 * because there is no race between I/O and migration for hugepage.
887 * Note that currently hugepage I/O occurs only in direct I/O
888 * where no lock is held and PG_writeback is irrelevant,
889 * and writeback status of all subpages are counted in the reference
890 * count of the head page (i.e. if all subpages of a 2MB hugepage are
891 * under direct I/O, the reference of the head page is 512 and a bit more.)
892 * This means that when we try to migrate hugepage whose subpages are
893 * doing direct I/O, some references remain after try_to_unmap() and
894 * hugepage migration fails without data corruption.
895 *
896 * There is also no race when direct I/O is issued on the page under migration,
897 * because then pte is replaced with migration swap entry and direct I/O code
898 * will wait in the page fault for migration to complete.
899 */
900static int unmap_and_move_huge_page(new_page_t get_new_page,
901 unsigned long private, struct page *hpage,
902 int force, bool offlining,
903 enum migrate_mode mode)
904{
905 int rc = 0;
906 int *result = NULL;
907 struct page *new_hpage = get_new_page(hpage, private, &result);
908 struct anon_vma *anon_vma = NULL;
909
910 if (!new_hpage)
911 return -ENOMEM;
912
913 rc = -EAGAIN;
914
915 if (!trylock_page(hpage)) {
916 if (!force || mode != MIGRATE_SYNC)
917 goto out;
918 lock_page(hpage);
919 }
920
921 if (PageAnon(hpage))
922 anon_vma = page_get_anon_vma(hpage);
923
924 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
925
926 if (!page_mapped(hpage))
927 rc = move_to_new_page(new_hpage, hpage, 1, mode);
928
929 if (rc)
930 remove_migration_ptes(hpage, hpage);
931
932 if (anon_vma)
933 put_anon_vma(anon_vma);
934 unlock_page(hpage);
935
936out:
937 if (rc != -EAGAIN) {
938 list_del(&hpage->lru);
939 put_page(hpage);
940 }
941
942 put_page(new_hpage);
943
944 if (result) {
945 if (rc)
946 *result = rc;
947 else
948 *result = page_to_nid(new_hpage);
949 }
950 return rc;
951}
952
953/*
954 * migrate_pages
955 *
956 * The function takes one list of pages to migrate and a function
957 * that determines from the page to be migrated and the private data
958 * the target of the move and allocates the page.
959 *
960 * The function returns after 10 attempts or if no pages
961 * are movable anymore because to has become empty
962 * or no retryable pages exist anymore.
963 * Caller should call putback_lru_pages to return pages to the LRU
964 * or free list only if ret != 0.
965 *
966 * Return: Number of pages not migrated or error code.
967 */
968int migrate_pages(struct list_head *from,
969 new_page_t get_new_page, unsigned long private, bool offlining,
970 enum migrate_mode mode)
971{
972 int retry = 1;
973 int nr_failed = 0;
974 int pass = 0;
975 struct page *page;
976 struct page *page2;
977 int swapwrite = current->flags & PF_SWAPWRITE;
978 int rc;
979
980 if (!swapwrite)
981 current->flags |= PF_SWAPWRITE;
982
983 for(pass = 0; pass < 10 && retry; pass++) {
984 retry = 0;
985
986 list_for_each_entry_safe(page, page2, from, lru) {
987 cond_resched();
988
989 rc = unmap_and_move(get_new_page, private,
990 page, pass > 2, offlining,
991 mode);
992
993 switch(rc) {
994 case -ENOMEM:
995 goto out;
996 case -EAGAIN:
997 retry++;
998 break;
999 case 0:
1000 break;
1001 default:
1002 /* Permanent failure */
1003 nr_failed++;
1004 break;
1005 }
1006 }
1007 }
1008 rc = 0;
1009out:
1010 if (!swapwrite)
1011 current->flags &= ~PF_SWAPWRITE;
1012
1013 if (rc)
1014 return rc;
1015
1016 return nr_failed + retry;
1017}
1018
1019int migrate_huge_pages(struct list_head *from,
1020 new_page_t get_new_page, unsigned long private, bool offlining,
1021 enum migrate_mode mode)
1022{
1023 int retry = 1;
1024 int nr_failed = 0;
1025 int pass = 0;
1026 struct page *page;
1027 struct page *page2;
1028 int rc;
1029
1030 for (pass = 0; pass < 10 && retry; pass++) {
1031 retry = 0;
1032
1033 list_for_each_entry_safe(page, page2, from, lru) {
1034 cond_resched();
1035
1036 rc = unmap_and_move_huge_page(get_new_page,
1037 private, page, pass > 2, offlining,
1038 mode);
1039
1040 switch(rc) {
1041 case -ENOMEM:
1042 goto out;
1043 case -EAGAIN:
1044 retry++;
1045 break;
1046 case 0:
1047 break;
1048 default:
1049 /* Permanent failure */
1050 nr_failed++;
1051 break;
1052 }
1053 }
1054 }
1055 rc = 0;
1056out:
1057 if (rc)
1058 return rc;
1059
1060 return nr_failed + retry;
1061}
1062
1063#ifdef CONFIG_NUMA
1064/*
1065 * Move a list of individual pages
1066 */
1067struct page_to_node {
1068 unsigned long addr;
1069 struct page *page;
1070 int node;
1071 int status;
1072};
1073
1074static struct page *new_page_node(struct page *p, unsigned long private,
1075 int **result)
1076{
1077 struct page_to_node *pm = (struct page_to_node *)private;
1078
1079 while (pm->node != MAX_NUMNODES && pm->page != p)
1080 pm++;
1081
1082 if (pm->node == MAX_NUMNODES)
1083 return NULL;
1084
1085 *result = &pm->status;
1086
1087 return alloc_pages_exact_node(pm->node,
1088 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1089}
1090
1091/*
1092 * Move a set of pages as indicated in the pm array. The addr
1093 * field must be set to the virtual address of the page to be moved
1094 * and the node number must contain a valid target node.
1095 * The pm array ends with node = MAX_NUMNODES.
1096 */
1097static int do_move_page_to_node_array(struct mm_struct *mm,
1098 struct page_to_node *pm,
1099 int migrate_all)
1100{
1101 int err;
1102 struct page_to_node *pp;
1103 LIST_HEAD(pagelist);
1104
1105 down_read(&mm->mmap_sem);
1106
1107 /*
1108 * Build a list of pages to migrate
1109 */
1110 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1111 struct vm_area_struct *vma;
1112 struct page *page;
1113
1114 err = -EFAULT;
1115 vma = find_vma(mm, pp->addr);
1116 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1117 goto set_status;
1118
1119 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1120
1121 err = PTR_ERR(page);
1122 if (IS_ERR(page))
1123 goto set_status;
1124
1125 err = -ENOENT;
1126 if (!page)
1127 goto set_status;
1128
1129 /* Use PageReserved to check for zero page */
1130 if (PageReserved(page) || PageKsm(page))
1131 goto put_and_set;
1132
1133 pp->page = page;
1134 err = page_to_nid(page);
1135
1136 if (err == pp->node)
1137 /*
1138 * Node already in the right place
1139 */
1140 goto put_and_set;
1141
1142 err = -EACCES;
1143 if (page_mapcount(page) > 1 &&
1144 !migrate_all)
1145 goto put_and_set;
1146
1147 err = isolate_lru_page(page);
1148 if (!err) {
1149 list_add_tail(&page->lru, &pagelist);
1150 inc_zone_page_state(page, NR_ISOLATED_ANON +
1151 page_is_file_cache(page));
1152 }
1153put_and_set:
1154 /*
1155 * Either remove the duplicate refcount from
1156 * isolate_lru_page() or drop the page ref if it was
1157 * not isolated.
1158 */
1159 put_page(page);
1160set_status:
1161 pp->status = err;
1162 }
1163
1164 err = 0;
1165 if (!list_empty(&pagelist)) {
1166 err = migrate_pages(&pagelist, new_page_node,
1167 (unsigned long)pm, 0, MIGRATE_SYNC);
1168 if (err)
1169 putback_lru_pages(&pagelist);
1170 }
1171
1172 up_read(&mm->mmap_sem);
1173 return err;
1174}
1175
1176/*
1177 * Migrate an array of page address onto an array of nodes and fill
1178 * the corresponding array of status.
1179 */
1180static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1181 unsigned long nr_pages,
1182 const void __user * __user *pages,
1183 const int __user *nodes,
1184 int __user *status, int flags)
1185{
1186 struct page_to_node *pm;
1187 unsigned long chunk_nr_pages;
1188 unsigned long chunk_start;
1189 int err;
1190
1191 err = -ENOMEM;
1192 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1193 if (!pm)
1194 goto out;
1195
1196 migrate_prep();
1197
1198 /*
1199 * Store a chunk of page_to_node array in a page,
1200 * but keep the last one as a marker
1201 */
1202 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1203
1204 for (chunk_start = 0;
1205 chunk_start < nr_pages;
1206 chunk_start += chunk_nr_pages) {
1207 int j;
1208
1209 if (chunk_start + chunk_nr_pages > nr_pages)
1210 chunk_nr_pages = nr_pages - chunk_start;
1211
1212 /* fill the chunk pm with addrs and nodes from user-space */
1213 for (j = 0; j < chunk_nr_pages; j++) {
1214 const void __user *p;
1215 int node;
1216
1217 err = -EFAULT;
1218 if (get_user(p, pages + j + chunk_start))
1219 goto out_pm;
1220 pm[j].addr = (unsigned long) p;
1221
1222 if (get_user(node, nodes + j + chunk_start))
1223 goto out_pm;
1224
1225 err = -ENODEV;
1226 if (node < 0 || node >= MAX_NUMNODES)
1227 goto out_pm;
1228
1229 if (!node_state(node, N_HIGH_MEMORY))
1230 goto out_pm;
1231
1232 err = -EACCES;
1233 if (!node_isset(node, task_nodes))
1234 goto out_pm;
1235
1236 pm[j].node = node;
1237 }
1238
1239 /* End marker for this chunk */
1240 pm[chunk_nr_pages].node = MAX_NUMNODES;
1241
1242 /* Migrate this chunk */
1243 err = do_move_page_to_node_array(mm, pm,
1244 flags & MPOL_MF_MOVE_ALL);
1245 if (err < 0)
1246 goto out_pm;
1247
1248 /* Return status information */
1249 for (j = 0; j < chunk_nr_pages; j++)
1250 if (put_user(pm[j].status, status + j + chunk_start)) {
1251 err = -EFAULT;
1252 goto out_pm;
1253 }
1254 }
1255 err = 0;
1256
1257out_pm:
1258 free_page((unsigned long)pm);
1259out:
1260 return err;
1261}
1262
1263/*
1264 * Determine the nodes of an array of pages and store it in an array of status.
1265 */
1266static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1267 const void __user **pages, int *status)
1268{
1269 unsigned long i;
1270
1271 down_read(&mm->mmap_sem);
1272
1273 for (i = 0; i < nr_pages; i++) {
1274 unsigned long addr = (unsigned long)(*pages);
1275 struct vm_area_struct *vma;
1276 struct page *page;
1277 int err = -EFAULT;
1278
1279 vma = find_vma(mm, addr);
1280 if (!vma || addr < vma->vm_start)
1281 goto set_status;
1282
1283 page = follow_page(vma, addr, 0);
1284
1285 err = PTR_ERR(page);
1286 if (IS_ERR(page))
1287 goto set_status;
1288
1289 err = -ENOENT;
1290 /* Use PageReserved to check for zero page */
1291 if (!page || PageReserved(page) || PageKsm(page))
1292 goto set_status;
1293
1294 err = page_to_nid(page);
1295set_status:
1296 *status = err;
1297
1298 pages++;
1299 status++;
1300 }
1301
1302 up_read(&mm->mmap_sem);
1303}
1304
1305/*
1306 * Determine the nodes of a user array of pages and store it in
1307 * a user array of status.
1308 */
1309static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1310 const void __user * __user *pages,
1311 int __user *status)
1312{
1313#define DO_PAGES_STAT_CHUNK_NR 16
1314 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1315 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1316
1317 while (nr_pages) {
1318 unsigned long chunk_nr;
1319
1320 chunk_nr = nr_pages;
1321 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1322 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1323
1324 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1325 break;
1326
1327 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1328
1329 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1330 break;
1331
1332 pages += chunk_nr;
1333 status += chunk_nr;
1334 nr_pages -= chunk_nr;
1335 }
1336 return nr_pages ? -EFAULT : 0;
1337}
1338
1339/*
1340 * Move a list of pages in the address space of the currently executing
1341 * process.
1342 */
1343SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1344 const void __user * __user *, pages,
1345 const int __user *, nodes,
1346 int __user *, status, int, flags)
1347{
1348 const struct cred *cred = current_cred(), *tcred;
1349 struct task_struct *task;
1350 struct mm_struct *mm;
1351 int err;
1352 nodemask_t task_nodes;
1353
1354 /* Check flags */
1355 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1356 return -EINVAL;
1357
1358 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1359 return -EPERM;
1360
1361 /* Find the mm_struct */
1362 rcu_read_lock();
1363 task = pid ? find_task_by_vpid(pid) : current;
1364 if (!task) {
1365 rcu_read_unlock();
1366 return -ESRCH;
1367 }
1368 get_task_struct(task);
1369
1370 /*
1371 * Check if this process has the right to modify the specified
1372 * process. The right exists if the process has administrative
1373 * capabilities, superuser privileges or the same
1374 * userid as the target process.
1375 */
1376 tcred = __task_cred(task);
1377 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1378 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1379 !capable(CAP_SYS_NICE)) {
1380 rcu_read_unlock();
1381 err = -EPERM;
1382 goto out;
1383 }
1384 rcu_read_unlock();
1385
1386 err = security_task_movememory(task);
1387 if (err)
1388 goto out;
1389
1390 task_nodes = cpuset_mems_allowed(task);
1391 mm = get_task_mm(task);
1392 put_task_struct(task);
1393
1394 if (!mm)
1395 return -EINVAL;
1396
1397 if (nodes)
1398 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1399 nodes, status, flags);
1400 else
1401 err = do_pages_stat(mm, nr_pages, pages, status);
1402
1403 mmput(mm);
1404 return err;
1405
1406out:
1407 put_task_struct(task);
1408 return err;
1409}
1410
1411/*
1412 * Call migration functions in the vma_ops that may prepare
1413 * memory in a vm for migration. migration functions may perform
1414 * the migration for vmas that do not have an underlying page struct.
1415 */
1416int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1417 const nodemask_t *from, unsigned long flags)
1418{
1419 struct vm_area_struct *vma;
1420 int err = 0;
1421
1422 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1423 if (vma->vm_ops && vma->vm_ops->migrate) {
1424 err = vma->vm_ops->migrate(vma, to, from, flags);
1425 if (err)
1426 break;
1427 }
1428 }
1429 return err;
1430}
1431#endif