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/pfn_t.h>
42#include <linux/memremap.h>
43#include <linux/userfaultfd_k.h>
44#include <linux/balloon_compaction.h>
45#include <linux/mmu_notifier.h>
46#include <linux/page_idle.h>
47#include <linux/page_owner.h>
48#include <linux/sched/mm.h>
49#include <linux/ptrace.h>
50
51#include <asm/tlbflush.h>
52
53#define CREATE_TRACE_POINTS
54#include <trace/events/migrate.h>
55
56#include "internal.h"
57
58/*
59 * migrate_prep() needs to be called before we start compiling a list of pages
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
62 */
63int migrate_prep(void)
64{
65 /*
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
70 */
71 lru_add_drain_all();
72
73 return 0;
74}
75
76/* Do the necessary work of migrate_prep but not if it involves other CPUs */
77int migrate_prep_local(void)
78{
79 lru_add_drain();
80
81 return 0;
82}
83
84int isolate_movable_page(struct page *page, isolate_mode_t mode)
85{
86 struct address_space *mapping;
87
88 /*
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
91 *
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
96 */
97 if (unlikely(!get_page_unless_zero(page)))
98 goto out;
99
100 /*
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
104 */
105 if (unlikely(!__PageMovable(page)))
106 goto out_putpage;
107 /*
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
111 *
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
117 */
118 if (unlikely(!trylock_page(page)))
119 goto out_putpage;
120
121 if (!PageMovable(page) || PageIsolated(page))
122 goto out_no_isolated;
123
124 mapping = page_mapping(page);
125 VM_BUG_ON_PAGE(!mapping, page);
126
127 if (!mapping->a_ops->isolate_page(page, mode))
128 goto out_no_isolated;
129
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page));
132 __SetPageIsolated(page);
133 unlock_page(page);
134
135 return 0;
136
137out_no_isolated:
138 unlock_page(page);
139out_putpage:
140 put_page(page);
141out:
142 return -EBUSY;
143}
144
145/* It should be called on page which is PG_movable */
146void putback_movable_page(struct page *page)
147{
148 struct address_space *mapping;
149
150 VM_BUG_ON_PAGE(!PageLocked(page), page);
151 VM_BUG_ON_PAGE(!PageMovable(page), page);
152 VM_BUG_ON_PAGE(!PageIsolated(page), page);
153
154 mapping = page_mapping(page);
155 mapping->a_ops->putback_page(page);
156 __ClearPageIsolated(page);
157}
158
159/*
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
162 *
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
166 */
167void putback_movable_pages(struct list_head *l)
168{
169 struct page *page;
170 struct page *page2;
171
172 list_for_each_entry_safe(page, page2, l, lru) {
173 if (unlikely(PageHuge(page))) {
174 putback_active_hugepage(page);
175 continue;
176 }
177 list_del(&page->lru);
178 /*
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
182 */
183 if (unlikely(__PageMovable(page))) {
184 VM_BUG_ON_PAGE(!PageIsolated(page), page);
185 lock_page(page);
186 if (PageMovable(page))
187 putback_movable_page(page);
188 else
189 __ClearPageIsolated(page);
190 unlock_page(page);
191 put_page(page);
192 } else {
193 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194 page_is_file_cache(page), -hpage_nr_pages(page));
195 putback_lru_page(page);
196 }
197 }
198}
199
200/*
201 * Restore a potential migration pte to a working pte entry
202 */
203static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
204 unsigned long addr, void *old)
205{
206 struct page_vma_mapped_walk pvmw = {
207 .page = old,
208 .vma = vma,
209 .address = addr,
210 .flags = PVMW_SYNC | PVMW_MIGRATION,
211 };
212 struct page *new;
213 pte_t pte;
214 swp_entry_t entry;
215
216 VM_BUG_ON_PAGE(PageTail(page), page);
217 while (page_vma_mapped_walk(&pvmw)) {
218 if (PageKsm(page))
219 new = page;
220 else
221 new = page - pvmw.page->index +
222 linear_page_index(vma, pvmw.address);
223
224#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
226 if (!pvmw.pte) {
227 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228 remove_migration_pmd(&pvmw, new);
229 continue;
230 }
231#endif
232
233 get_page(new);
234 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235 if (pte_swp_soft_dirty(*pvmw.pte))
236 pte = pte_mksoft_dirty(pte);
237
238 /*
239 * Recheck VMA as permissions can change since migration started
240 */
241 entry = pte_to_swp_entry(*pvmw.pte);
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
244
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry = make_device_private_entry(new, pte_write(pte));
248 pte = swp_entry_to_pte(entry);
249 } else if (is_device_public_page(new)) {
250 pte = pte_mkdevmap(pte);
251 flush_dcache_page(new);
252 }
253 } else
254 flush_dcache_page(new);
255
256#ifdef CONFIG_HUGETLB_PAGE
257 if (PageHuge(new)) {
258 pte = pte_mkhuge(pte);
259 pte = arch_make_huge_pte(pte, vma, new, 0);
260 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
261 if (PageAnon(new))
262 hugepage_add_anon_rmap(new, vma, pvmw.address);
263 else
264 page_dup_rmap(new, true);
265 } else
266#endif
267 {
268 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269
270 if (PageAnon(new))
271 page_add_anon_rmap(new, vma, pvmw.address, false);
272 else
273 page_add_file_rmap(new, false);
274 }
275 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 mlock_vma_page(new);
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 radix-tree replacement of page migration started, page_count
327 * *must* be zero. And, we don't want to call wait_on_page_locked()
328 * against a page without get_page().
329 * So, we use get_page_unless_zero(), here. Even failed, page fault
330 * will occur again.
331 */
332 if (!get_page_unless_zero(page))
333 goto out;
334 pte_unmap_unlock(ptep, ptl);
335 wait_on_page_locked(page);
336 put_page(page);
337 return;
338out:
339 pte_unmap_unlock(ptep, ptl);
340}
341
342void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
343 unsigned long address)
344{
345 spinlock_t *ptl = pte_lockptr(mm, pmd);
346 pte_t *ptep = pte_offset_map(pmd, address);
347 __migration_entry_wait(mm, ptep, ptl);
348}
349
350void migration_entry_wait_huge(struct vm_area_struct *vma,
351 struct mm_struct *mm, pte_t *pte)
352{
353 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
354 __migration_entry_wait(mm, pte, ptl);
355}
356
357#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
358void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
359{
360 spinlock_t *ptl;
361 struct page *page;
362
363 ptl = pmd_lock(mm, pmd);
364 if (!is_pmd_migration_entry(*pmd))
365 goto unlock;
366 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
367 if (!get_page_unless_zero(page))
368 goto unlock;
369 spin_unlock(ptl);
370 wait_on_page_locked(page);
371 put_page(page);
372 return;
373unlock:
374 spin_unlock(ptl);
375}
376#endif
377
378#ifdef CONFIG_BLOCK
379/* Returns true if all buffers are successfully locked */
380static bool buffer_migrate_lock_buffers(struct buffer_head *head,
381 enum migrate_mode mode)
382{
383 struct buffer_head *bh = head;
384
385 /* Simple case, sync compaction */
386 if (mode != MIGRATE_ASYNC) {
387 do {
388 get_bh(bh);
389 lock_buffer(bh);
390 bh = bh->b_this_page;
391
392 } while (bh != head);
393
394 return true;
395 }
396
397 /* async case, we cannot block on lock_buffer so use trylock_buffer */
398 do {
399 get_bh(bh);
400 if (!trylock_buffer(bh)) {
401 /*
402 * We failed to lock the buffer and cannot stall in
403 * async migration. Release the taken locks
404 */
405 struct buffer_head *failed_bh = bh;
406 put_bh(failed_bh);
407 bh = head;
408 while (bh != failed_bh) {
409 unlock_buffer(bh);
410 put_bh(bh);
411 bh = bh->b_this_page;
412 }
413 return false;
414 }
415
416 bh = bh->b_this_page;
417 } while (bh != head);
418 return true;
419}
420#else
421static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
422 enum migrate_mode mode)
423{
424 return true;
425}
426#endif /* CONFIG_BLOCK */
427
428/*
429 * Replace the page in the mapping.
430 *
431 * The number of remaining references must be:
432 * 1 for anonymous pages without a mapping
433 * 2 for pages with a mapping
434 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
435 */
436int migrate_page_move_mapping(struct address_space *mapping,
437 struct page *newpage, struct page *page,
438 struct buffer_head *head, enum migrate_mode mode,
439 int extra_count)
440{
441 struct zone *oldzone, *newzone;
442 int dirty;
443 int expected_count = 1 + extra_count;
444 void **pslot;
445
446 /*
447 * Device public or private pages have an extra refcount as they are
448 * ZONE_DEVICE pages.
449 */
450 expected_count += is_device_private_page(page);
451 expected_count += is_device_public_page(page);
452
453 if (!mapping) {
454 /* Anonymous page without mapping */
455 if (page_count(page) != expected_count)
456 return -EAGAIN;
457
458 /* No turning back from here */
459 newpage->index = page->index;
460 newpage->mapping = page->mapping;
461 if (PageSwapBacked(page))
462 __SetPageSwapBacked(newpage);
463
464 return MIGRATEPAGE_SUCCESS;
465 }
466
467 oldzone = page_zone(page);
468 newzone = page_zone(newpage);
469
470 xa_lock_irq(&mapping->i_pages);
471
472 pslot = radix_tree_lookup_slot(&mapping->i_pages,
473 page_index(page));
474
475 expected_count += hpage_nr_pages(page) + page_has_private(page);
476 if (page_count(page) != expected_count ||
477 radix_tree_deref_slot_protected(pslot,
478 &mapping->i_pages.xa_lock) != page) {
479 xa_unlock_irq(&mapping->i_pages);
480 return -EAGAIN;
481 }
482
483 if (!page_ref_freeze(page, expected_count)) {
484 xa_unlock_irq(&mapping->i_pages);
485 return -EAGAIN;
486 }
487
488 /*
489 * In the async migration case of moving a page with buffers, lock the
490 * buffers using trylock before the mapping is moved. If the mapping
491 * was moved, we later failed to lock the buffers and could not move
492 * the mapping back due to an elevated page count, we would have to
493 * block waiting on other references to be dropped.
494 */
495 if (mode == MIGRATE_ASYNC && head &&
496 !buffer_migrate_lock_buffers(head, mode)) {
497 page_ref_unfreeze(page, expected_count);
498 xa_unlock_irq(&mapping->i_pages);
499 return -EAGAIN;
500 }
501
502 /*
503 * Now we know that no one else is looking at the page:
504 * no turning back from here.
505 */
506 newpage->index = page->index;
507 newpage->mapping = page->mapping;
508 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
509 if (PageSwapBacked(page)) {
510 __SetPageSwapBacked(newpage);
511 if (PageSwapCache(page)) {
512 SetPageSwapCache(newpage);
513 set_page_private(newpage, page_private(page));
514 }
515 } else {
516 VM_BUG_ON_PAGE(PageSwapCache(page), page);
517 }
518
519 /* Move dirty while page refs frozen and newpage not yet exposed */
520 dirty = PageDirty(page);
521 if (dirty) {
522 ClearPageDirty(page);
523 SetPageDirty(newpage);
524 }
525
526 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
527 if (PageTransHuge(page)) {
528 int i;
529 int index = page_index(page);
530
531 for (i = 1; i < HPAGE_PMD_NR; i++) {
532 pslot = radix_tree_lookup_slot(&mapping->i_pages,
533 index + i);
534 radix_tree_replace_slot(&mapping->i_pages, pslot,
535 newpage + i);
536 }
537 }
538
539 /*
540 * Drop cache reference from old page by unfreezing
541 * to one less reference.
542 * We know this isn't the last reference.
543 */
544 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
545
546 xa_unlock(&mapping->i_pages);
547 /* Leave irq disabled to prevent preemption while updating stats */
548
549 /*
550 * If moved to a different zone then also account
551 * the page for that zone. Other VM counters will be
552 * taken care of when we establish references to the
553 * new page and drop references to the old page.
554 *
555 * Note that anonymous pages are accounted for
556 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
557 * are mapped to swap space.
558 */
559 if (newzone != oldzone) {
560 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
561 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
562 if (PageSwapBacked(page) && !PageSwapCache(page)) {
563 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
564 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
565 }
566 if (dirty && mapping_cap_account_dirty(mapping)) {
567 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
568 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
569 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
570 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
571 }
572 }
573 local_irq_enable();
574
575 return MIGRATEPAGE_SUCCESS;
576}
577EXPORT_SYMBOL(migrate_page_move_mapping);
578
579/*
580 * The expected number of remaining references is the same as that
581 * of migrate_page_move_mapping().
582 */
583int migrate_huge_page_move_mapping(struct address_space *mapping,
584 struct page *newpage, struct page *page)
585{
586 int expected_count;
587 void **pslot;
588
589 xa_lock_irq(&mapping->i_pages);
590
591 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
592
593 expected_count = 2 + page_has_private(page);
594 if (page_count(page) != expected_count ||
595 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
596 xa_unlock_irq(&mapping->i_pages);
597 return -EAGAIN;
598 }
599
600 if (!page_ref_freeze(page, expected_count)) {
601 xa_unlock_irq(&mapping->i_pages);
602 return -EAGAIN;
603 }
604
605 newpage->index = page->index;
606 newpage->mapping = page->mapping;
607
608 get_page(newpage);
609
610 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
611
612 page_ref_unfreeze(page, expected_count - 1);
613
614 xa_unlock_irq(&mapping->i_pages);
615
616 return MIGRATEPAGE_SUCCESS;
617}
618
619/*
620 * Gigantic pages are so large that we do not guarantee that page++ pointer
621 * arithmetic will work across the entire page. We need something more
622 * specialized.
623 */
624static void __copy_gigantic_page(struct page *dst, struct page *src,
625 int nr_pages)
626{
627 int i;
628 struct page *dst_base = dst;
629 struct page *src_base = src;
630
631 for (i = 0; i < nr_pages; ) {
632 cond_resched();
633 copy_highpage(dst, src);
634
635 i++;
636 dst = mem_map_next(dst, dst_base, i);
637 src = mem_map_next(src, src_base, i);
638 }
639}
640
641static void copy_huge_page(struct page *dst, struct page *src)
642{
643 int i;
644 int nr_pages;
645
646 if (PageHuge(src)) {
647 /* hugetlbfs page */
648 struct hstate *h = page_hstate(src);
649 nr_pages = pages_per_huge_page(h);
650
651 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
652 __copy_gigantic_page(dst, src, nr_pages);
653 return;
654 }
655 } else {
656 /* thp page */
657 BUG_ON(!PageTransHuge(src));
658 nr_pages = hpage_nr_pages(src);
659 }
660
661 for (i = 0; i < nr_pages; i++) {
662 cond_resched();
663 copy_highpage(dst + i, src + i);
664 }
665}
666
667/*
668 * Copy the page to its new location
669 */
670void migrate_page_states(struct page *newpage, struct page *page)
671{
672 int cpupid;
673
674 if (PageError(page))
675 SetPageError(newpage);
676 if (PageReferenced(page))
677 SetPageReferenced(newpage);
678 if (PageUptodate(page))
679 SetPageUptodate(newpage);
680 if (TestClearPageActive(page)) {
681 VM_BUG_ON_PAGE(PageUnevictable(page), page);
682 SetPageActive(newpage);
683 } else if (TestClearPageUnevictable(page))
684 SetPageUnevictable(newpage);
685 if (PageChecked(page))
686 SetPageChecked(newpage);
687 if (PageMappedToDisk(page))
688 SetPageMappedToDisk(newpage);
689
690 /* Move dirty on pages not done by migrate_page_move_mapping() */
691 if (PageDirty(page))
692 SetPageDirty(newpage);
693
694 if (page_is_young(page))
695 set_page_young(newpage);
696 if (page_is_idle(page))
697 set_page_idle(newpage);
698
699 /*
700 * Copy NUMA information to the new page, to prevent over-eager
701 * future migrations of this same page.
702 */
703 cpupid = page_cpupid_xchg_last(page, -1);
704 page_cpupid_xchg_last(newpage, cpupid);
705
706 ksm_migrate_page(newpage, page);
707 /*
708 * Please do not reorder this without considering how mm/ksm.c's
709 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
710 */
711 if (PageSwapCache(page))
712 ClearPageSwapCache(page);
713 ClearPagePrivate(page);
714 set_page_private(page, 0);
715
716 /*
717 * If any waiters have accumulated on the new page then
718 * wake them up.
719 */
720 if (PageWriteback(newpage))
721 end_page_writeback(newpage);
722
723 copy_page_owner(page, newpage);
724
725 mem_cgroup_migrate(page, newpage);
726}
727EXPORT_SYMBOL(migrate_page_states);
728
729void migrate_page_copy(struct page *newpage, struct page *page)
730{
731 if (PageHuge(page) || PageTransHuge(page))
732 copy_huge_page(newpage, page);
733 else
734 copy_highpage(newpage, page);
735
736 migrate_page_states(newpage, page);
737}
738EXPORT_SYMBOL(migrate_page_copy);
739
740/************************************************************
741 * Migration functions
742 ***********************************************************/
743
744/*
745 * Common logic to directly migrate a single LRU page suitable for
746 * pages that do not use PagePrivate/PagePrivate2.
747 *
748 * Pages are locked upon entry and exit.
749 */
750int migrate_page(struct address_space *mapping,
751 struct page *newpage, struct page *page,
752 enum migrate_mode mode)
753{
754 int rc;
755
756 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
757
758 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
759
760 if (rc != MIGRATEPAGE_SUCCESS)
761 return rc;
762
763 if (mode != MIGRATE_SYNC_NO_COPY)
764 migrate_page_copy(newpage, page);
765 else
766 migrate_page_states(newpage, page);
767 return MIGRATEPAGE_SUCCESS;
768}
769EXPORT_SYMBOL(migrate_page);
770
771#ifdef CONFIG_BLOCK
772/*
773 * Migration function for pages with buffers. This function can only be used
774 * if the underlying filesystem guarantees that no other references to "page"
775 * exist.
776 */
777int buffer_migrate_page(struct address_space *mapping,
778 struct page *newpage, struct page *page, enum migrate_mode mode)
779{
780 struct buffer_head *bh, *head;
781 int rc;
782
783 if (!page_has_buffers(page))
784 return migrate_page(mapping, newpage, page, mode);
785
786 head = page_buffers(page);
787
788 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
789
790 if (rc != MIGRATEPAGE_SUCCESS)
791 return rc;
792
793 /*
794 * In the async case, migrate_page_move_mapping locked the buffers
795 * with an IRQ-safe spinlock held. In the sync case, the buffers
796 * need to be locked now
797 */
798 if (mode != MIGRATE_ASYNC)
799 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
800
801 ClearPagePrivate(page);
802 set_page_private(newpage, page_private(page));
803 set_page_private(page, 0);
804 put_page(page);
805 get_page(newpage);
806
807 bh = head;
808 do {
809 set_bh_page(bh, newpage, bh_offset(bh));
810 bh = bh->b_this_page;
811
812 } while (bh != head);
813
814 SetPagePrivate(newpage);
815
816 if (mode != MIGRATE_SYNC_NO_COPY)
817 migrate_page_copy(newpage, page);
818 else
819 migrate_page_states(newpage, page);
820
821 bh = head;
822 do {
823 unlock_buffer(bh);
824 put_bh(bh);
825 bh = bh->b_this_page;
826
827 } while (bh != head);
828
829 return MIGRATEPAGE_SUCCESS;
830}
831EXPORT_SYMBOL(buffer_migrate_page);
832#endif
833
834/*
835 * Writeback a page to clean the dirty state
836 */
837static int writeout(struct address_space *mapping, struct page *page)
838{
839 struct writeback_control wbc = {
840 .sync_mode = WB_SYNC_NONE,
841 .nr_to_write = 1,
842 .range_start = 0,
843 .range_end = LLONG_MAX,
844 .for_reclaim = 1
845 };
846 int rc;
847
848 if (!mapping->a_ops->writepage)
849 /* No write method for the address space */
850 return -EINVAL;
851
852 if (!clear_page_dirty_for_io(page))
853 /* Someone else already triggered a write */
854 return -EAGAIN;
855
856 /*
857 * A dirty page may imply that the underlying filesystem has
858 * the page on some queue. So the page must be clean for
859 * migration. Writeout may mean we loose the lock and the
860 * page state is no longer what we checked for earlier.
861 * At this point we know that the migration attempt cannot
862 * be successful.
863 */
864 remove_migration_ptes(page, page, false);
865
866 rc = mapping->a_ops->writepage(page, &wbc);
867
868 if (rc != AOP_WRITEPAGE_ACTIVATE)
869 /* unlocked. Relock */
870 lock_page(page);
871
872 return (rc < 0) ? -EIO : -EAGAIN;
873}
874
875/*
876 * Default handling if a filesystem does not provide a migration function.
877 */
878static int fallback_migrate_page(struct address_space *mapping,
879 struct page *newpage, struct page *page, enum migrate_mode mode)
880{
881 if (PageDirty(page)) {
882 /* Only writeback pages in full synchronous migration */
883 switch (mode) {
884 case MIGRATE_SYNC:
885 case MIGRATE_SYNC_NO_COPY:
886 break;
887 default:
888 return -EBUSY;
889 }
890 return writeout(mapping, page);
891 }
892
893 /*
894 * Buffers may be managed in a filesystem specific way.
895 * We must have no buffers or drop them.
896 */
897 if (page_has_private(page) &&
898 !try_to_release_page(page, GFP_KERNEL))
899 return -EAGAIN;
900
901 return migrate_page(mapping, newpage, page, mode);
902}
903
904/*
905 * Move a page to a newly allocated page
906 * The page is locked and all ptes have been successfully removed.
907 *
908 * The new page will have replaced the old page if this function
909 * is successful.
910 *
911 * Return value:
912 * < 0 - error code
913 * MIGRATEPAGE_SUCCESS - success
914 */
915static int move_to_new_page(struct page *newpage, struct page *page,
916 enum migrate_mode mode)
917{
918 struct address_space *mapping;
919 int rc = -EAGAIN;
920 bool is_lru = !__PageMovable(page);
921
922 VM_BUG_ON_PAGE(!PageLocked(page), page);
923 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
924
925 mapping = page_mapping(page);
926
927 if (likely(is_lru)) {
928 if (!mapping)
929 rc = migrate_page(mapping, newpage, page, mode);
930 else if (mapping->a_ops->migratepage)
931 /*
932 * Most pages have a mapping and most filesystems
933 * provide a migratepage callback. Anonymous pages
934 * are part of swap space which also has its own
935 * migratepage callback. This is the most common path
936 * for page migration.
937 */
938 rc = mapping->a_ops->migratepage(mapping, newpage,
939 page, mode);
940 else
941 rc = fallback_migrate_page(mapping, newpage,
942 page, mode);
943 } else {
944 /*
945 * In case of non-lru page, it could be released after
946 * isolation step. In that case, we shouldn't try migration.
947 */
948 VM_BUG_ON_PAGE(!PageIsolated(page), page);
949 if (!PageMovable(page)) {
950 rc = MIGRATEPAGE_SUCCESS;
951 __ClearPageIsolated(page);
952 goto out;
953 }
954
955 rc = mapping->a_ops->migratepage(mapping, newpage,
956 page, mode);
957 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
958 !PageIsolated(page));
959 }
960
961 /*
962 * When successful, old pagecache page->mapping must be cleared before
963 * page is freed; but stats require that PageAnon be left as PageAnon.
964 */
965 if (rc == MIGRATEPAGE_SUCCESS) {
966 if (__PageMovable(page)) {
967 VM_BUG_ON_PAGE(!PageIsolated(page), page);
968
969 /*
970 * We clear PG_movable under page_lock so any compactor
971 * cannot try to migrate this page.
972 */
973 __ClearPageIsolated(page);
974 }
975
976 /*
977 * Anonymous and movable page->mapping will be cleard by
978 * free_pages_prepare so don't reset it here for keeping
979 * the type to work PageAnon, for example.
980 */
981 if (!PageMappingFlags(page))
982 page->mapping = NULL;
983 }
984out:
985 return rc;
986}
987
988static int __unmap_and_move(struct page *page, struct page *newpage,
989 int force, enum migrate_mode mode)
990{
991 int rc = -EAGAIN;
992 int page_was_mapped = 0;
993 struct anon_vma *anon_vma = NULL;
994 bool is_lru = !__PageMovable(page);
995
996 if (!trylock_page(page)) {
997 if (!force || mode == MIGRATE_ASYNC)
998 goto out;
999
1000 /*
1001 * It's not safe for direct compaction to call lock_page.
1002 * For example, during page readahead pages are added locked
1003 * to the LRU. Later, when the IO completes the pages are
1004 * marked uptodate and unlocked. However, the queueing
1005 * could be merging multiple pages for one bio (e.g.
1006 * mpage_readpages). If an allocation happens for the
1007 * second or third page, the process can end up locking
1008 * the same page twice and deadlocking. Rather than
1009 * trying to be clever about what pages can be locked,
1010 * avoid the use of lock_page for direct compaction
1011 * altogether.
1012 */
1013 if (current->flags & PF_MEMALLOC)
1014 goto out;
1015
1016 lock_page(page);
1017 }
1018
1019 if (PageWriteback(page)) {
1020 /*
1021 * Only in the case of a full synchronous migration is it
1022 * necessary to wait for PageWriteback. In the async case,
1023 * the retry loop is too short and in the sync-light case,
1024 * the overhead of stalling is too much
1025 */
1026 switch (mode) {
1027 case MIGRATE_SYNC:
1028 case MIGRATE_SYNC_NO_COPY:
1029 break;
1030 default:
1031 rc = -EBUSY;
1032 goto out_unlock;
1033 }
1034 if (!force)
1035 goto out_unlock;
1036 wait_on_page_writeback(page);
1037 }
1038
1039 /*
1040 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1041 * we cannot notice that anon_vma is freed while we migrates a page.
1042 * This get_anon_vma() delays freeing anon_vma pointer until the end
1043 * of migration. File cache pages are no problem because of page_lock()
1044 * File Caches may use write_page() or lock_page() in migration, then,
1045 * just care Anon page here.
1046 *
1047 * Only page_get_anon_vma() understands the subtleties of
1048 * getting a hold on an anon_vma from outside one of its mms.
1049 * But if we cannot get anon_vma, then we won't need it anyway,
1050 * because that implies that the anon page is no longer mapped
1051 * (and cannot be remapped so long as we hold the page lock).
1052 */
1053 if (PageAnon(page) && !PageKsm(page))
1054 anon_vma = page_get_anon_vma(page);
1055
1056 /*
1057 * Block others from accessing the new page when we get around to
1058 * establishing additional references. We are usually the only one
1059 * holding a reference to newpage at this point. We used to have a BUG
1060 * here if trylock_page(newpage) fails, but would like to allow for
1061 * cases where there might be a race with the previous use of newpage.
1062 * This is much like races on refcount of oldpage: just don't BUG().
1063 */
1064 if (unlikely(!trylock_page(newpage)))
1065 goto out_unlock;
1066
1067 if (unlikely(!is_lru)) {
1068 rc = move_to_new_page(newpage, page, mode);
1069 goto out_unlock_both;
1070 }
1071
1072 /*
1073 * Corner case handling:
1074 * 1. When a new swap-cache page is read into, it is added to the LRU
1075 * and treated as swapcache but it has no rmap yet.
1076 * Calling try_to_unmap() against a page->mapping==NULL page will
1077 * trigger a BUG. So handle it here.
1078 * 2. An orphaned page (see truncate_complete_page) might have
1079 * fs-private metadata. The page can be picked up due to memory
1080 * offlining. Everywhere else except page reclaim, the page is
1081 * invisible to the vm, so the page can not be migrated. So try to
1082 * free the metadata, so the page can be freed.
1083 */
1084 if (!page->mapping) {
1085 VM_BUG_ON_PAGE(PageAnon(page), page);
1086 if (page_has_private(page)) {
1087 try_to_free_buffers(page);
1088 goto out_unlock_both;
1089 }
1090 } else if (page_mapped(page)) {
1091 /* Establish migration ptes */
1092 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1093 page);
1094 try_to_unmap(page,
1095 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1096 page_was_mapped = 1;
1097 }
1098
1099 if (!page_mapped(page))
1100 rc = move_to_new_page(newpage, page, mode);
1101
1102 if (page_was_mapped)
1103 remove_migration_ptes(page,
1104 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1105
1106out_unlock_both:
1107 unlock_page(newpage);
1108out_unlock:
1109 /* Drop an anon_vma reference if we took one */
1110 if (anon_vma)
1111 put_anon_vma(anon_vma);
1112 unlock_page(page);
1113out:
1114 /*
1115 * If migration is successful, decrease refcount of the newpage
1116 * which will not free the page because new page owner increased
1117 * refcounter. As well, if it is LRU page, add the page to LRU
1118 * list in here.
1119 */
1120 if (rc == MIGRATEPAGE_SUCCESS) {
1121 if (unlikely(__PageMovable(newpage)))
1122 put_page(newpage);
1123 else
1124 putback_lru_page(newpage);
1125 }
1126
1127 return rc;
1128}
1129
1130/*
1131 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1132 * around it.
1133 */
1134#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1135#define ICE_noinline noinline
1136#else
1137#define ICE_noinline
1138#endif
1139
1140/*
1141 * Obtain the lock on page, remove all ptes and migrate the page
1142 * to the newly allocated page in newpage.
1143 */
1144static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1145 free_page_t put_new_page,
1146 unsigned long private, struct page *page,
1147 int force, enum migrate_mode mode,
1148 enum migrate_reason reason)
1149{
1150 int rc = MIGRATEPAGE_SUCCESS;
1151 struct page *newpage;
1152
1153 if (!thp_migration_supported() && PageTransHuge(page))
1154 return -ENOMEM;
1155
1156 newpage = get_new_page(page, private);
1157 if (!newpage)
1158 return -ENOMEM;
1159
1160 if (page_count(page) == 1) {
1161 /* page was freed from under us. So we are done. */
1162 ClearPageActive(page);
1163 ClearPageUnevictable(page);
1164 if (unlikely(__PageMovable(page))) {
1165 lock_page(page);
1166 if (!PageMovable(page))
1167 __ClearPageIsolated(page);
1168 unlock_page(page);
1169 }
1170 if (put_new_page)
1171 put_new_page(newpage, private);
1172 else
1173 put_page(newpage);
1174 goto out;
1175 }
1176
1177 rc = __unmap_and_move(page, newpage, force, mode);
1178 if (rc == MIGRATEPAGE_SUCCESS)
1179 set_page_owner_migrate_reason(newpage, reason);
1180
1181out:
1182 if (rc != -EAGAIN) {
1183 /*
1184 * A page that has been migrated has all references
1185 * removed and will be freed. A page that has not been
1186 * migrated will have kepts its references and be
1187 * restored.
1188 */
1189 list_del(&page->lru);
1190
1191 /*
1192 * Compaction can migrate also non-LRU pages which are
1193 * not accounted to NR_ISOLATED_*. They can be recognized
1194 * as __PageMovable
1195 */
1196 if (likely(!__PageMovable(page)))
1197 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1198 page_is_file_cache(page), -hpage_nr_pages(page));
1199 }
1200
1201 /*
1202 * If migration is successful, releases reference grabbed during
1203 * isolation. Otherwise, restore the page to right list unless
1204 * we want to retry.
1205 */
1206 if (rc == MIGRATEPAGE_SUCCESS) {
1207 put_page(page);
1208 if (reason == MR_MEMORY_FAILURE) {
1209 /*
1210 * Set PG_HWPoison on just freed page
1211 * intentionally. Although it's rather weird,
1212 * it's how HWPoison flag works at the moment.
1213 */
1214 if (!test_set_page_hwpoison(page))
1215 num_poisoned_pages_inc();
1216 }
1217 } else {
1218 if (rc != -EAGAIN) {
1219 if (likely(!__PageMovable(page))) {
1220 putback_lru_page(page);
1221 goto put_new;
1222 }
1223
1224 lock_page(page);
1225 if (PageMovable(page))
1226 putback_movable_page(page);
1227 else
1228 __ClearPageIsolated(page);
1229 unlock_page(page);
1230 put_page(page);
1231 }
1232put_new:
1233 if (put_new_page)
1234 put_new_page(newpage, private);
1235 else
1236 put_page(newpage);
1237 }
1238
1239 return rc;
1240}
1241
1242/*
1243 * Counterpart of unmap_and_move_page() for hugepage migration.
1244 *
1245 * This function doesn't wait the completion of hugepage I/O
1246 * because there is no race between I/O and migration for hugepage.
1247 * Note that currently hugepage I/O occurs only in direct I/O
1248 * where no lock is held and PG_writeback is irrelevant,
1249 * and writeback status of all subpages are counted in the reference
1250 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1251 * under direct I/O, the reference of the head page is 512 and a bit more.)
1252 * This means that when we try to migrate hugepage whose subpages are
1253 * doing direct I/O, some references remain after try_to_unmap() and
1254 * hugepage migration fails without data corruption.
1255 *
1256 * There is also no race when direct I/O is issued on the page under migration,
1257 * because then pte is replaced with migration swap entry and direct I/O code
1258 * will wait in the page fault for migration to complete.
1259 */
1260static int unmap_and_move_huge_page(new_page_t get_new_page,
1261 free_page_t put_new_page, unsigned long private,
1262 struct page *hpage, int force,
1263 enum migrate_mode mode, int reason)
1264{
1265 int rc = -EAGAIN;
1266 int page_was_mapped = 0;
1267 struct page *new_hpage;
1268 struct anon_vma *anon_vma = NULL;
1269
1270 /*
1271 * Movability of hugepages depends on architectures and hugepage size.
1272 * This check is necessary because some callers of hugepage migration
1273 * like soft offline and memory hotremove don't walk through page
1274 * tables or check whether the hugepage is pmd-based or not before
1275 * kicking migration.
1276 */
1277 if (!hugepage_migration_supported(page_hstate(hpage))) {
1278 putback_active_hugepage(hpage);
1279 return -ENOSYS;
1280 }
1281
1282 new_hpage = get_new_page(hpage, private);
1283 if (!new_hpage)
1284 return -ENOMEM;
1285
1286 if (!trylock_page(hpage)) {
1287 if (!force)
1288 goto out;
1289 switch (mode) {
1290 case MIGRATE_SYNC:
1291 case MIGRATE_SYNC_NO_COPY:
1292 break;
1293 default:
1294 goto out;
1295 }
1296 lock_page(hpage);
1297 }
1298
1299 if (PageAnon(hpage))
1300 anon_vma = page_get_anon_vma(hpage);
1301
1302 if (unlikely(!trylock_page(new_hpage)))
1303 goto put_anon;
1304
1305 if (page_mapped(hpage)) {
1306 try_to_unmap(hpage,
1307 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1308 page_was_mapped = 1;
1309 }
1310
1311 if (!page_mapped(hpage))
1312 rc = move_to_new_page(new_hpage, hpage, mode);
1313
1314 if (page_was_mapped)
1315 remove_migration_ptes(hpage,
1316 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1317
1318 unlock_page(new_hpage);
1319
1320put_anon:
1321 if (anon_vma)
1322 put_anon_vma(anon_vma);
1323
1324 if (rc == MIGRATEPAGE_SUCCESS) {
1325 move_hugetlb_state(hpage, new_hpage, reason);
1326 put_new_page = NULL;
1327 }
1328
1329 unlock_page(hpage);
1330out:
1331 if (rc != -EAGAIN)
1332 putback_active_hugepage(hpage);
1333 if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage))
1334 num_poisoned_pages_inc();
1335
1336 /*
1337 * If migration was not successful and there's a freeing callback, use
1338 * it. Otherwise, put_page() will drop the reference grabbed during
1339 * isolation.
1340 */
1341 if (put_new_page)
1342 put_new_page(new_hpage, private);
1343 else
1344 putback_active_hugepage(new_hpage);
1345
1346 return rc;
1347}
1348
1349/*
1350 * migrate_pages - migrate the pages specified in a list, to the free pages
1351 * supplied as the target for the page migration
1352 *
1353 * @from: The list of pages to be migrated.
1354 * @get_new_page: The function used to allocate free pages to be used
1355 * as the target of the page migration.
1356 * @put_new_page: The function used to free target pages if migration
1357 * fails, or NULL if no special handling is necessary.
1358 * @private: Private data to be passed on to get_new_page()
1359 * @mode: The migration mode that specifies the constraints for
1360 * page migration, if any.
1361 * @reason: The reason for page migration.
1362 *
1363 * The function returns after 10 attempts or if no pages are movable any more
1364 * because the list has become empty or no retryable pages exist any more.
1365 * The caller should call putback_movable_pages() to return pages to the LRU
1366 * or free list only if ret != 0.
1367 *
1368 * Returns the number of pages that were not migrated, or an error code.
1369 */
1370int migrate_pages(struct list_head *from, new_page_t get_new_page,
1371 free_page_t put_new_page, unsigned long private,
1372 enum migrate_mode mode, int reason)
1373{
1374 int retry = 1;
1375 int nr_failed = 0;
1376 int nr_succeeded = 0;
1377 int pass = 0;
1378 struct page *page;
1379 struct page *page2;
1380 int swapwrite = current->flags & PF_SWAPWRITE;
1381 int rc;
1382
1383 if (!swapwrite)
1384 current->flags |= PF_SWAPWRITE;
1385
1386 for(pass = 0; pass < 10 && retry; pass++) {
1387 retry = 0;
1388
1389 list_for_each_entry_safe(page, page2, from, lru) {
1390retry:
1391 cond_resched();
1392
1393 if (PageHuge(page))
1394 rc = unmap_and_move_huge_page(get_new_page,
1395 put_new_page, private, page,
1396 pass > 2, mode, reason);
1397 else
1398 rc = unmap_and_move(get_new_page, put_new_page,
1399 private, page, pass > 2, mode,
1400 reason);
1401
1402 switch(rc) {
1403 case -ENOMEM:
1404 /*
1405 * THP migration might be unsupported or the
1406 * allocation could've failed so we should
1407 * retry on the same page with the THP split
1408 * to base pages.
1409 *
1410 * Head page is retried immediately and tail
1411 * pages are added to the tail of the list so
1412 * we encounter them after the rest of the list
1413 * is processed.
1414 */
1415 if (PageTransHuge(page)) {
1416 lock_page(page);
1417 rc = split_huge_page_to_list(page, from);
1418 unlock_page(page);
1419 if (!rc) {
1420 list_safe_reset_next(page, page2, lru);
1421 goto retry;
1422 }
1423 }
1424 nr_failed++;
1425 goto out;
1426 case -EAGAIN:
1427 retry++;
1428 break;
1429 case MIGRATEPAGE_SUCCESS:
1430 nr_succeeded++;
1431 break;
1432 default:
1433 /*
1434 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1435 * unlike -EAGAIN case, the failed page is
1436 * removed from migration page list and not
1437 * retried in the next outer loop.
1438 */
1439 nr_failed++;
1440 break;
1441 }
1442 }
1443 }
1444 nr_failed += retry;
1445 rc = nr_failed;
1446out:
1447 if (nr_succeeded)
1448 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1449 if (nr_failed)
1450 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1451 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1452
1453 if (!swapwrite)
1454 current->flags &= ~PF_SWAPWRITE;
1455
1456 return rc;
1457}
1458
1459#ifdef CONFIG_NUMA
1460
1461static int store_status(int __user *status, int start, int value, int nr)
1462{
1463 while (nr-- > 0) {
1464 if (put_user(value, status + start))
1465 return -EFAULT;
1466 start++;
1467 }
1468
1469 return 0;
1470}
1471
1472static int do_move_pages_to_node(struct mm_struct *mm,
1473 struct list_head *pagelist, int node)
1474{
1475 int err;
1476
1477 if (list_empty(pagelist))
1478 return 0;
1479
1480 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1481 MIGRATE_SYNC, MR_SYSCALL);
1482 if (err)
1483 putback_movable_pages(pagelist);
1484 return err;
1485}
1486
1487/*
1488 * Resolves the given address to a struct page, isolates it from the LRU and
1489 * puts it to the given pagelist.
1490 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1491 * queued or the page doesn't need to be migrated because it is already on
1492 * the target node
1493 */
1494static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1495 int node, struct list_head *pagelist, bool migrate_all)
1496{
1497 struct vm_area_struct *vma;
1498 struct page *page;
1499 unsigned int follflags;
1500 int err;
1501
1502 down_read(&mm->mmap_sem);
1503 err = -EFAULT;
1504 vma = find_vma(mm, addr);
1505 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1506 goto out;
1507
1508 /* FOLL_DUMP to ignore special (like zero) pages */
1509 follflags = FOLL_GET | FOLL_DUMP;
1510 page = follow_page(vma, addr, follflags);
1511
1512 err = PTR_ERR(page);
1513 if (IS_ERR(page))
1514 goto out;
1515
1516 err = -ENOENT;
1517 if (!page)
1518 goto out;
1519
1520 err = 0;
1521 if (page_to_nid(page) == node)
1522 goto out_putpage;
1523
1524 err = -EACCES;
1525 if (page_mapcount(page) > 1 && !migrate_all)
1526 goto out_putpage;
1527
1528 if (PageHuge(page)) {
1529 if (PageHead(page)) {
1530 isolate_huge_page(page, pagelist);
1531 err = 0;
1532 }
1533 } else {
1534 struct page *head;
1535
1536 head = compound_head(page);
1537 err = isolate_lru_page(head);
1538 if (err)
1539 goto out_putpage;
1540
1541 err = 0;
1542 list_add_tail(&head->lru, pagelist);
1543 mod_node_page_state(page_pgdat(head),
1544 NR_ISOLATED_ANON + page_is_file_cache(head),
1545 hpage_nr_pages(head));
1546 }
1547out_putpage:
1548 /*
1549 * Either remove the duplicate refcount from
1550 * isolate_lru_page() or drop the page ref if it was
1551 * not isolated.
1552 */
1553 put_page(page);
1554out:
1555 up_read(&mm->mmap_sem);
1556 return err;
1557}
1558
1559/*
1560 * Migrate an array of page address onto an array of nodes and fill
1561 * the corresponding array of status.
1562 */
1563static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1564 unsigned long nr_pages,
1565 const void __user * __user *pages,
1566 const int __user *nodes,
1567 int __user *status, int flags)
1568{
1569 int current_node = NUMA_NO_NODE;
1570 LIST_HEAD(pagelist);
1571 int start, i;
1572 int err = 0, err1;
1573
1574 migrate_prep();
1575
1576 for (i = start = 0; i < nr_pages; i++) {
1577 const void __user *p;
1578 unsigned long addr;
1579 int node;
1580
1581 err = -EFAULT;
1582 if (get_user(p, pages + i))
1583 goto out_flush;
1584 if (get_user(node, nodes + i))
1585 goto out_flush;
1586 addr = (unsigned long)p;
1587
1588 err = -ENODEV;
1589 if (node < 0 || node >= MAX_NUMNODES)
1590 goto out_flush;
1591 if (!node_state(node, N_MEMORY))
1592 goto out_flush;
1593
1594 err = -EACCES;
1595 if (!node_isset(node, task_nodes))
1596 goto out_flush;
1597
1598 if (current_node == NUMA_NO_NODE) {
1599 current_node = node;
1600 start = i;
1601 } else if (node != current_node) {
1602 err = do_move_pages_to_node(mm, &pagelist, current_node);
1603 if (err)
1604 goto out;
1605 err = store_status(status, start, current_node, i - start);
1606 if (err)
1607 goto out;
1608 start = i;
1609 current_node = node;
1610 }
1611
1612 /*
1613 * Errors in the page lookup or isolation are not fatal and we simply
1614 * report them via status
1615 */
1616 err = add_page_for_migration(mm, addr, current_node,
1617 &pagelist, flags & MPOL_MF_MOVE_ALL);
1618 if (!err)
1619 continue;
1620
1621 err = store_status(status, i, err, 1);
1622 if (err)
1623 goto out_flush;
1624
1625 err = do_move_pages_to_node(mm, &pagelist, current_node);
1626 if (err)
1627 goto out;
1628 if (i > start) {
1629 err = store_status(status, start, current_node, i - start);
1630 if (err)
1631 goto out;
1632 }
1633 current_node = NUMA_NO_NODE;
1634 }
1635out_flush:
1636 if (list_empty(&pagelist))
1637 return err;
1638
1639 /* Make sure we do not overwrite the existing error */
1640 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1641 if (!err1)
1642 err1 = store_status(status, start, current_node, i - start);
1643 if (!err)
1644 err = err1;
1645out:
1646 return err;
1647}
1648
1649/*
1650 * Determine the nodes of an array of pages and store it in an array of status.
1651 */
1652static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1653 const void __user **pages, int *status)
1654{
1655 unsigned long i;
1656
1657 down_read(&mm->mmap_sem);
1658
1659 for (i = 0; i < nr_pages; i++) {
1660 unsigned long addr = (unsigned long)(*pages);
1661 struct vm_area_struct *vma;
1662 struct page *page;
1663 int err = -EFAULT;
1664
1665 vma = find_vma(mm, addr);
1666 if (!vma || addr < vma->vm_start)
1667 goto set_status;
1668
1669 /* FOLL_DUMP to ignore special (like zero) pages */
1670 page = follow_page(vma, addr, FOLL_DUMP);
1671
1672 err = PTR_ERR(page);
1673 if (IS_ERR(page))
1674 goto set_status;
1675
1676 err = page ? page_to_nid(page) : -ENOENT;
1677set_status:
1678 *status = err;
1679
1680 pages++;
1681 status++;
1682 }
1683
1684 up_read(&mm->mmap_sem);
1685}
1686
1687/*
1688 * Determine the nodes of a user array of pages and store it in
1689 * a user array of status.
1690 */
1691static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1692 const void __user * __user *pages,
1693 int __user *status)
1694{
1695#define DO_PAGES_STAT_CHUNK_NR 16
1696 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1697 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1698
1699 while (nr_pages) {
1700 unsigned long chunk_nr;
1701
1702 chunk_nr = nr_pages;
1703 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1704 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1705
1706 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1707 break;
1708
1709 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1710
1711 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1712 break;
1713
1714 pages += chunk_nr;
1715 status += chunk_nr;
1716 nr_pages -= chunk_nr;
1717 }
1718 return nr_pages ? -EFAULT : 0;
1719}
1720
1721/*
1722 * Move a list of pages in the address space of the currently executing
1723 * process.
1724 */
1725static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1726 const void __user * __user *pages,
1727 const int __user *nodes,
1728 int __user *status, int flags)
1729{
1730 struct task_struct *task;
1731 struct mm_struct *mm;
1732 int err;
1733 nodemask_t task_nodes;
1734
1735 /* Check flags */
1736 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1737 return -EINVAL;
1738
1739 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1740 return -EPERM;
1741
1742 /* Find the mm_struct */
1743 rcu_read_lock();
1744 task = pid ? find_task_by_vpid(pid) : current;
1745 if (!task) {
1746 rcu_read_unlock();
1747 return -ESRCH;
1748 }
1749 get_task_struct(task);
1750
1751 /*
1752 * Check if this process has the right to modify the specified
1753 * process. Use the regular "ptrace_may_access()" checks.
1754 */
1755 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1756 rcu_read_unlock();
1757 err = -EPERM;
1758 goto out;
1759 }
1760 rcu_read_unlock();
1761
1762 err = security_task_movememory(task);
1763 if (err)
1764 goto out;
1765
1766 task_nodes = cpuset_mems_allowed(task);
1767 mm = get_task_mm(task);
1768 put_task_struct(task);
1769
1770 if (!mm)
1771 return -EINVAL;
1772
1773 if (nodes)
1774 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1775 nodes, status, flags);
1776 else
1777 err = do_pages_stat(mm, nr_pages, pages, status);
1778
1779 mmput(mm);
1780 return err;
1781
1782out:
1783 put_task_struct(task);
1784 return err;
1785}
1786
1787SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1788 const void __user * __user *, pages,
1789 const int __user *, nodes,
1790 int __user *, status, int, flags)
1791{
1792 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1793}
1794
1795#ifdef CONFIG_COMPAT
1796COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1797 compat_uptr_t __user *, pages32,
1798 const int __user *, nodes,
1799 int __user *, status,
1800 int, flags)
1801{
1802 const void __user * __user *pages;
1803 int i;
1804
1805 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1806 for (i = 0; i < nr_pages; i++) {
1807 compat_uptr_t p;
1808
1809 if (get_user(p, pages32 + i) ||
1810 put_user(compat_ptr(p), pages + i))
1811 return -EFAULT;
1812 }
1813 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1814}
1815#endif /* CONFIG_COMPAT */
1816
1817#ifdef CONFIG_NUMA_BALANCING
1818/*
1819 * Returns true if this is a safe migration target node for misplaced NUMA
1820 * pages. Currently it only checks the watermarks which crude
1821 */
1822static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1823 unsigned long nr_migrate_pages)
1824{
1825 int z;
1826
1827 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1828 struct zone *zone = pgdat->node_zones + z;
1829
1830 if (!populated_zone(zone))
1831 continue;
1832
1833 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1834 if (!zone_watermark_ok(zone, 0,
1835 high_wmark_pages(zone) +
1836 nr_migrate_pages,
1837 0, 0))
1838 continue;
1839 return true;
1840 }
1841 return false;
1842}
1843
1844static struct page *alloc_misplaced_dst_page(struct page *page,
1845 unsigned long data)
1846{
1847 int nid = (int) data;
1848 struct page *newpage;
1849
1850 newpage = __alloc_pages_node(nid,
1851 (GFP_HIGHUSER_MOVABLE |
1852 __GFP_THISNODE | __GFP_NOMEMALLOC |
1853 __GFP_NORETRY | __GFP_NOWARN) &
1854 ~__GFP_RECLAIM, 0);
1855
1856 return newpage;
1857}
1858
1859/*
1860 * page migration rate limiting control.
1861 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1862 * window of time. Default here says do not migrate more than 1280M per second.
1863 */
1864static unsigned int migrate_interval_millisecs __read_mostly = 100;
1865static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1866
1867/* Returns true if the node is migrate rate-limited after the update */
1868static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1869 unsigned long nr_pages)
1870{
1871 /*
1872 * Rate-limit the amount of data that is being migrated to a node.
1873 * Optimal placement is no good if the memory bus is saturated and
1874 * all the time is being spent migrating!
1875 */
1876 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1877 spin_lock(&pgdat->numabalancing_migrate_lock);
1878 pgdat->numabalancing_migrate_nr_pages = 0;
1879 pgdat->numabalancing_migrate_next_window = jiffies +
1880 msecs_to_jiffies(migrate_interval_millisecs);
1881 spin_unlock(&pgdat->numabalancing_migrate_lock);
1882 }
1883 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1884 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1885 nr_pages);
1886 return true;
1887 }
1888
1889 /*
1890 * This is an unlocked non-atomic update so errors are possible.
1891 * The consequences are failing to migrate when we potentiall should
1892 * have which is not severe enough to warrant locking. If it is ever
1893 * a problem, it can be converted to a per-cpu counter.
1894 */
1895 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1896 return false;
1897}
1898
1899static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1900{
1901 int page_lru;
1902
1903 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1904
1905 /* Avoid migrating to a node that is nearly full */
1906 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1907 return 0;
1908
1909 if (isolate_lru_page(page))
1910 return 0;
1911
1912 /*
1913 * migrate_misplaced_transhuge_page() skips page migration's usual
1914 * check on page_count(), so we must do it here, now that the page
1915 * has been isolated: a GUP pin, or any other pin, prevents migration.
1916 * The expected page count is 3: 1 for page's mapcount and 1 for the
1917 * caller's pin and 1 for the reference taken by isolate_lru_page().
1918 */
1919 if (PageTransHuge(page) && page_count(page) != 3) {
1920 putback_lru_page(page);
1921 return 0;
1922 }
1923
1924 page_lru = page_is_file_cache(page);
1925 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1926 hpage_nr_pages(page));
1927
1928 /*
1929 * Isolating the page has taken another reference, so the
1930 * caller's reference can be safely dropped without the page
1931 * disappearing underneath us during migration.
1932 */
1933 put_page(page);
1934 return 1;
1935}
1936
1937bool pmd_trans_migrating(pmd_t pmd)
1938{
1939 struct page *page = pmd_page(pmd);
1940 return PageLocked(page);
1941}
1942
1943/*
1944 * Attempt to migrate a misplaced page to the specified destination
1945 * node. Caller is expected to have an elevated reference count on
1946 * the page that will be dropped by this function before returning.
1947 */
1948int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1949 int node)
1950{
1951 pg_data_t *pgdat = NODE_DATA(node);
1952 int isolated;
1953 int nr_remaining;
1954 LIST_HEAD(migratepages);
1955
1956 /*
1957 * Don't migrate file pages that are mapped in multiple processes
1958 * with execute permissions as they are probably shared libraries.
1959 */
1960 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1961 (vma->vm_flags & VM_EXEC))
1962 goto out;
1963
1964 /*
1965 * Also do not migrate dirty pages as not all filesystems can move
1966 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1967 */
1968 if (page_is_file_cache(page) && PageDirty(page))
1969 goto out;
1970
1971 /*
1972 * Rate-limit the amount of data that is being migrated to a node.
1973 * Optimal placement is no good if the memory bus is saturated and
1974 * all the time is being spent migrating!
1975 */
1976 if (numamigrate_update_ratelimit(pgdat, 1))
1977 goto out;
1978
1979 isolated = numamigrate_isolate_page(pgdat, page);
1980 if (!isolated)
1981 goto out;
1982
1983 list_add(&page->lru, &migratepages);
1984 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1985 NULL, node, MIGRATE_ASYNC,
1986 MR_NUMA_MISPLACED);
1987 if (nr_remaining) {
1988 if (!list_empty(&migratepages)) {
1989 list_del(&page->lru);
1990 dec_node_page_state(page, NR_ISOLATED_ANON +
1991 page_is_file_cache(page));
1992 putback_lru_page(page);
1993 }
1994 isolated = 0;
1995 } else
1996 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1997 BUG_ON(!list_empty(&migratepages));
1998 return isolated;
1999
2000out:
2001 put_page(page);
2002 return 0;
2003}
2004#endif /* CONFIG_NUMA_BALANCING */
2005
2006#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2007/*
2008 * Migrates a THP to a given target node. page must be locked and is unlocked
2009 * before returning.
2010 */
2011int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2012 struct vm_area_struct *vma,
2013 pmd_t *pmd, pmd_t entry,
2014 unsigned long address,
2015 struct page *page, int node)
2016{
2017 spinlock_t *ptl;
2018 pg_data_t *pgdat = NODE_DATA(node);
2019 int isolated = 0;
2020 struct page *new_page = NULL;
2021 int page_lru = page_is_file_cache(page);
2022 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2023 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2024
2025 /*
2026 * Rate-limit the amount of data that is being migrated to a node.
2027 * Optimal placement is no good if the memory bus is saturated and
2028 * all the time is being spent migrating!
2029 */
2030 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
2031 goto out_dropref;
2032
2033 new_page = alloc_pages_node(node,
2034 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2035 HPAGE_PMD_ORDER);
2036 if (!new_page)
2037 goto out_fail;
2038 prep_transhuge_page(new_page);
2039
2040 isolated = numamigrate_isolate_page(pgdat, page);
2041 if (!isolated) {
2042 put_page(new_page);
2043 goto out_fail;
2044 }
2045
2046 /* Prepare a page as a migration target */
2047 __SetPageLocked(new_page);
2048 if (PageSwapBacked(page))
2049 __SetPageSwapBacked(new_page);
2050
2051 /* anon mapping, we can simply copy page->mapping to the new page: */
2052 new_page->mapping = page->mapping;
2053 new_page->index = page->index;
2054 migrate_page_copy(new_page, page);
2055 WARN_ON(PageLRU(new_page));
2056
2057 /* Recheck the target PMD */
2058 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2059 ptl = pmd_lock(mm, pmd);
2060 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2061 spin_unlock(ptl);
2062 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2063
2064 /* Reverse changes made by migrate_page_copy() */
2065 if (TestClearPageActive(new_page))
2066 SetPageActive(page);
2067 if (TestClearPageUnevictable(new_page))
2068 SetPageUnevictable(page);
2069
2070 unlock_page(new_page);
2071 put_page(new_page); /* Free it */
2072
2073 /* Retake the callers reference and putback on LRU */
2074 get_page(page);
2075 putback_lru_page(page);
2076 mod_node_page_state(page_pgdat(page),
2077 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2078
2079 goto out_unlock;
2080 }
2081
2082 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2083 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2084
2085 /*
2086 * Clear the old entry under pagetable lock and establish the new PTE.
2087 * Any parallel GUP will either observe the old page blocking on the
2088 * page lock, block on the page table lock or observe the new page.
2089 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2090 * guarantee the copy is visible before the pagetable update.
2091 */
2092 flush_cache_range(vma, mmun_start, mmun_end);
2093 page_add_anon_rmap(new_page, vma, mmun_start, true);
2094 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2095 set_pmd_at(mm, mmun_start, pmd, entry);
2096 update_mmu_cache_pmd(vma, address, &entry);
2097
2098 page_ref_unfreeze(page, 2);
2099 mlock_migrate_page(new_page, page);
2100 page_remove_rmap(page, true);
2101 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2102
2103 spin_unlock(ptl);
2104 /*
2105 * No need to double call mmu_notifier->invalidate_range() callback as
2106 * the above pmdp_huge_clear_flush_notify() did already call it.
2107 */
2108 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2109
2110 /* Take an "isolate" reference and put new page on the LRU. */
2111 get_page(new_page);
2112 putback_lru_page(new_page);
2113
2114 unlock_page(new_page);
2115 unlock_page(page);
2116 put_page(page); /* Drop the rmap reference */
2117 put_page(page); /* Drop the LRU isolation reference */
2118
2119 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2120 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2121
2122 mod_node_page_state(page_pgdat(page),
2123 NR_ISOLATED_ANON + page_lru,
2124 -HPAGE_PMD_NR);
2125 return isolated;
2126
2127out_fail:
2128 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2129out_dropref:
2130 ptl = pmd_lock(mm, pmd);
2131 if (pmd_same(*pmd, entry)) {
2132 entry = pmd_modify(entry, vma->vm_page_prot);
2133 set_pmd_at(mm, mmun_start, pmd, entry);
2134 update_mmu_cache_pmd(vma, address, &entry);
2135 }
2136 spin_unlock(ptl);
2137
2138out_unlock:
2139 unlock_page(page);
2140 put_page(page);
2141 return 0;
2142}
2143#endif /* CONFIG_NUMA_BALANCING */
2144
2145#endif /* CONFIG_NUMA */
2146
2147#if defined(CONFIG_MIGRATE_VMA_HELPER)
2148struct migrate_vma {
2149 struct vm_area_struct *vma;
2150 unsigned long *dst;
2151 unsigned long *src;
2152 unsigned long cpages;
2153 unsigned long npages;
2154 unsigned long start;
2155 unsigned long end;
2156};
2157
2158static int migrate_vma_collect_hole(unsigned long start,
2159 unsigned long end,
2160 struct mm_walk *walk)
2161{
2162 struct migrate_vma *migrate = walk->private;
2163 unsigned long addr;
2164
2165 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2166 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2167 migrate->dst[migrate->npages] = 0;
2168 migrate->npages++;
2169 migrate->cpages++;
2170 }
2171
2172 return 0;
2173}
2174
2175static int migrate_vma_collect_skip(unsigned long start,
2176 unsigned long end,
2177 struct mm_walk *walk)
2178{
2179 struct migrate_vma *migrate = walk->private;
2180 unsigned long addr;
2181
2182 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2183 migrate->dst[migrate->npages] = 0;
2184 migrate->src[migrate->npages++] = 0;
2185 }
2186
2187 return 0;
2188}
2189
2190static int migrate_vma_collect_pmd(pmd_t *pmdp,
2191 unsigned long start,
2192 unsigned long end,
2193 struct mm_walk *walk)
2194{
2195 struct migrate_vma *migrate = walk->private;
2196 struct vm_area_struct *vma = walk->vma;
2197 struct mm_struct *mm = vma->vm_mm;
2198 unsigned long addr = start, unmapped = 0;
2199 spinlock_t *ptl;
2200 pte_t *ptep;
2201
2202again:
2203 if (pmd_none(*pmdp))
2204 return migrate_vma_collect_hole(start, end, walk);
2205
2206 if (pmd_trans_huge(*pmdp)) {
2207 struct page *page;
2208
2209 ptl = pmd_lock(mm, pmdp);
2210 if (unlikely(!pmd_trans_huge(*pmdp))) {
2211 spin_unlock(ptl);
2212 goto again;
2213 }
2214
2215 page = pmd_page(*pmdp);
2216 if (is_huge_zero_page(page)) {
2217 spin_unlock(ptl);
2218 split_huge_pmd(vma, pmdp, addr);
2219 if (pmd_trans_unstable(pmdp))
2220 return migrate_vma_collect_skip(start, end,
2221 walk);
2222 } else {
2223 int ret;
2224
2225 get_page(page);
2226 spin_unlock(ptl);
2227 if (unlikely(!trylock_page(page)))
2228 return migrate_vma_collect_skip(start, end,
2229 walk);
2230 ret = split_huge_page(page);
2231 unlock_page(page);
2232 put_page(page);
2233 if (ret)
2234 return migrate_vma_collect_skip(start, end,
2235 walk);
2236 if (pmd_none(*pmdp))
2237 return migrate_vma_collect_hole(start, end,
2238 walk);
2239 }
2240 }
2241
2242 if (unlikely(pmd_bad(*pmdp)))
2243 return migrate_vma_collect_skip(start, end, walk);
2244
2245 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2246 arch_enter_lazy_mmu_mode();
2247
2248 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2249 unsigned long mpfn, pfn;
2250 struct page *page;
2251 swp_entry_t entry;
2252 pte_t pte;
2253
2254 pte = *ptep;
2255 pfn = pte_pfn(pte);
2256
2257 if (pte_none(pte)) {
2258 mpfn = MIGRATE_PFN_MIGRATE;
2259 migrate->cpages++;
2260 pfn = 0;
2261 goto next;
2262 }
2263
2264 if (!pte_present(pte)) {
2265 mpfn = pfn = 0;
2266
2267 /*
2268 * Only care about unaddressable device page special
2269 * page table entry. Other special swap entries are not
2270 * migratable, and we ignore regular swapped page.
2271 */
2272 entry = pte_to_swp_entry(pte);
2273 if (!is_device_private_entry(entry))
2274 goto next;
2275
2276 page = device_private_entry_to_page(entry);
2277 mpfn = migrate_pfn(page_to_pfn(page))|
2278 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2279 if (is_write_device_private_entry(entry))
2280 mpfn |= MIGRATE_PFN_WRITE;
2281 } else {
2282 if (is_zero_pfn(pfn)) {
2283 mpfn = MIGRATE_PFN_MIGRATE;
2284 migrate->cpages++;
2285 pfn = 0;
2286 goto next;
2287 }
2288 page = _vm_normal_page(migrate->vma, addr, pte, true);
2289 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2290 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2291 }
2292
2293 /* FIXME support THP */
2294 if (!page || !page->mapping || PageTransCompound(page)) {
2295 mpfn = pfn = 0;
2296 goto next;
2297 }
2298 pfn = page_to_pfn(page);
2299
2300 /*
2301 * By getting a reference on the page we pin it and that blocks
2302 * any kind of migration. Side effect is that it "freezes" the
2303 * pte.
2304 *
2305 * We drop this reference after isolating the page from the lru
2306 * for non device page (device page are not on the lru and thus
2307 * can't be dropped from it).
2308 */
2309 get_page(page);
2310 migrate->cpages++;
2311
2312 /*
2313 * Optimize for the common case where page is only mapped once
2314 * in one process. If we can lock the page, then we can safely
2315 * set up a special migration page table entry now.
2316 */
2317 if (trylock_page(page)) {
2318 pte_t swp_pte;
2319
2320 mpfn |= MIGRATE_PFN_LOCKED;
2321 ptep_get_and_clear(mm, addr, ptep);
2322
2323 /* Setup special migration page table entry */
2324 entry = make_migration_entry(page, mpfn &
2325 MIGRATE_PFN_WRITE);
2326 swp_pte = swp_entry_to_pte(entry);
2327 if (pte_soft_dirty(pte))
2328 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2329 set_pte_at(mm, addr, ptep, swp_pte);
2330
2331 /*
2332 * This is like regular unmap: we remove the rmap and
2333 * drop page refcount. Page won't be freed, as we took
2334 * a reference just above.
2335 */
2336 page_remove_rmap(page, false);
2337 put_page(page);
2338
2339 if (pte_present(pte))
2340 unmapped++;
2341 }
2342
2343next:
2344 migrate->dst[migrate->npages] = 0;
2345 migrate->src[migrate->npages++] = mpfn;
2346 }
2347 arch_leave_lazy_mmu_mode();
2348 pte_unmap_unlock(ptep - 1, ptl);
2349
2350 /* Only flush the TLB if we actually modified any entries */
2351 if (unmapped)
2352 flush_tlb_range(walk->vma, start, end);
2353
2354 return 0;
2355}
2356
2357/*
2358 * migrate_vma_collect() - collect pages over a range of virtual addresses
2359 * @migrate: migrate struct containing all migration information
2360 *
2361 * This will walk the CPU page table. For each virtual address backed by a
2362 * valid page, it updates the src array and takes a reference on the page, in
2363 * order to pin the page until we lock it and unmap it.
2364 */
2365static void migrate_vma_collect(struct migrate_vma *migrate)
2366{
2367 struct mm_walk mm_walk;
2368
2369 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2370 mm_walk.pte_entry = NULL;
2371 mm_walk.pte_hole = migrate_vma_collect_hole;
2372 mm_walk.hugetlb_entry = NULL;
2373 mm_walk.test_walk = NULL;
2374 mm_walk.vma = migrate->vma;
2375 mm_walk.mm = migrate->vma->vm_mm;
2376 mm_walk.private = migrate;
2377
2378 mmu_notifier_invalidate_range_start(mm_walk.mm,
2379 migrate->start,
2380 migrate->end);
2381 walk_page_range(migrate->start, migrate->end, &mm_walk);
2382 mmu_notifier_invalidate_range_end(mm_walk.mm,
2383 migrate->start,
2384 migrate->end);
2385
2386 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2387}
2388
2389/*
2390 * migrate_vma_check_page() - check if page is pinned or not
2391 * @page: struct page to check
2392 *
2393 * Pinned pages cannot be migrated. This is the same test as in
2394 * migrate_page_move_mapping(), except that here we allow migration of a
2395 * ZONE_DEVICE page.
2396 */
2397static bool migrate_vma_check_page(struct page *page)
2398{
2399 /*
2400 * One extra ref because caller holds an extra reference, either from
2401 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2402 * a device page.
2403 */
2404 int extra = 1;
2405
2406 /*
2407 * FIXME support THP (transparent huge page), it is bit more complex to
2408 * check them than regular pages, because they can be mapped with a pmd
2409 * or with a pte (split pte mapping).
2410 */
2411 if (PageCompound(page))
2412 return false;
2413
2414 /* Page from ZONE_DEVICE have one extra reference */
2415 if (is_zone_device_page(page)) {
2416 /*
2417 * Private page can never be pin as they have no valid pte and
2418 * GUP will fail for those. Yet if there is a pending migration
2419 * a thread might try to wait on the pte migration entry and
2420 * will bump the page reference count. Sadly there is no way to
2421 * differentiate a regular pin from migration wait. Hence to
2422 * avoid 2 racing thread trying to migrate back to CPU to enter
2423 * infinite loop (one stoping migration because the other is
2424 * waiting on pte migration entry). We always return true here.
2425 *
2426 * FIXME proper solution is to rework migration_entry_wait() so
2427 * it does not need to take a reference on page.
2428 */
2429 if (is_device_private_page(page))
2430 return true;
2431
2432 /*
2433 * Only allow device public page to be migrated and account for
2434 * the extra reference count imply by ZONE_DEVICE pages.
2435 */
2436 if (!is_device_public_page(page))
2437 return false;
2438 extra++;
2439 }
2440
2441 /* For file back page */
2442 if (page_mapping(page))
2443 extra += 1 + page_has_private(page);
2444
2445 if ((page_count(page) - extra) > page_mapcount(page))
2446 return false;
2447
2448 return true;
2449}
2450
2451/*
2452 * migrate_vma_prepare() - lock pages and isolate them from the lru
2453 * @migrate: migrate struct containing all migration information
2454 *
2455 * This locks pages that have been collected by migrate_vma_collect(). Once each
2456 * page is locked it is isolated from the lru (for non-device pages). Finally,
2457 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2458 * migrated by concurrent kernel threads.
2459 */
2460static void migrate_vma_prepare(struct migrate_vma *migrate)
2461{
2462 const unsigned long npages = migrate->npages;
2463 const unsigned long start = migrate->start;
2464 unsigned long addr, i, restore = 0;
2465 bool allow_drain = true;
2466
2467 lru_add_drain();
2468
2469 for (i = 0; (i < npages) && migrate->cpages; i++) {
2470 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2471 bool remap = true;
2472
2473 if (!page)
2474 continue;
2475
2476 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2477 /*
2478 * Because we are migrating several pages there can be
2479 * a deadlock between 2 concurrent migration where each
2480 * are waiting on each other page lock.
2481 *
2482 * Make migrate_vma() a best effort thing and backoff
2483 * for any page we can not lock right away.
2484 */
2485 if (!trylock_page(page)) {
2486 migrate->src[i] = 0;
2487 migrate->cpages--;
2488 put_page(page);
2489 continue;
2490 }
2491 remap = false;
2492 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2493 }
2494
2495 /* ZONE_DEVICE pages are not on LRU */
2496 if (!is_zone_device_page(page)) {
2497 if (!PageLRU(page) && allow_drain) {
2498 /* Drain CPU's pagevec */
2499 lru_add_drain_all();
2500 allow_drain = false;
2501 }
2502
2503 if (isolate_lru_page(page)) {
2504 if (remap) {
2505 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2506 migrate->cpages--;
2507 restore++;
2508 } else {
2509 migrate->src[i] = 0;
2510 unlock_page(page);
2511 migrate->cpages--;
2512 put_page(page);
2513 }
2514 continue;
2515 }
2516
2517 /* Drop the reference we took in collect */
2518 put_page(page);
2519 }
2520
2521 if (!migrate_vma_check_page(page)) {
2522 if (remap) {
2523 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2524 migrate->cpages--;
2525 restore++;
2526
2527 if (!is_zone_device_page(page)) {
2528 get_page(page);
2529 putback_lru_page(page);
2530 }
2531 } else {
2532 migrate->src[i] = 0;
2533 unlock_page(page);
2534 migrate->cpages--;
2535
2536 if (!is_zone_device_page(page))
2537 putback_lru_page(page);
2538 else
2539 put_page(page);
2540 }
2541 }
2542 }
2543
2544 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2545 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2546
2547 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2548 continue;
2549
2550 remove_migration_pte(page, migrate->vma, addr, page);
2551
2552 migrate->src[i] = 0;
2553 unlock_page(page);
2554 put_page(page);
2555 restore--;
2556 }
2557}
2558
2559/*
2560 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2561 * @migrate: migrate struct containing all migration information
2562 *
2563 * Replace page mapping (CPU page table pte) with a special migration pte entry
2564 * and check again if it has been pinned. Pinned pages are restored because we
2565 * cannot migrate them.
2566 *
2567 * This is the last step before we call the device driver callback to allocate
2568 * destination memory and copy contents of original page over to new page.
2569 */
2570static void migrate_vma_unmap(struct migrate_vma *migrate)
2571{
2572 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2573 const unsigned long npages = migrate->npages;
2574 const unsigned long start = migrate->start;
2575 unsigned long addr, i, restore = 0;
2576
2577 for (i = 0; i < npages; i++) {
2578 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2579
2580 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2581 continue;
2582
2583 if (page_mapped(page)) {
2584 try_to_unmap(page, flags);
2585 if (page_mapped(page))
2586 goto restore;
2587 }
2588
2589 if (migrate_vma_check_page(page))
2590 continue;
2591
2592restore:
2593 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2594 migrate->cpages--;
2595 restore++;
2596 }
2597
2598 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2599 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2600
2601 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2602 continue;
2603
2604 remove_migration_ptes(page, page, false);
2605
2606 migrate->src[i] = 0;
2607 unlock_page(page);
2608 restore--;
2609
2610 if (is_zone_device_page(page))
2611 put_page(page);
2612 else
2613 putback_lru_page(page);
2614 }
2615}
2616
2617static void migrate_vma_insert_page(struct migrate_vma *migrate,
2618 unsigned long addr,
2619 struct page *page,
2620 unsigned long *src,
2621 unsigned long *dst)
2622{
2623 struct vm_area_struct *vma = migrate->vma;
2624 struct mm_struct *mm = vma->vm_mm;
2625 struct mem_cgroup *memcg;
2626 bool flush = false;
2627 spinlock_t *ptl;
2628 pte_t entry;
2629 pgd_t *pgdp;
2630 p4d_t *p4dp;
2631 pud_t *pudp;
2632 pmd_t *pmdp;
2633 pte_t *ptep;
2634
2635 /* Only allow populating anonymous memory */
2636 if (!vma_is_anonymous(vma))
2637 goto abort;
2638
2639 pgdp = pgd_offset(mm, addr);
2640 p4dp = p4d_alloc(mm, pgdp, addr);
2641 if (!p4dp)
2642 goto abort;
2643 pudp = pud_alloc(mm, p4dp, addr);
2644 if (!pudp)
2645 goto abort;
2646 pmdp = pmd_alloc(mm, pudp, addr);
2647 if (!pmdp)
2648 goto abort;
2649
2650 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2651 goto abort;
2652
2653 /*
2654 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2655 * pte_offset_map() on pmds where a huge pmd might be created
2656 * from a different thread.
2657 *
2658 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2659 * parallel threads are excluded by other means.
2660 *
2661 * Here we only have down_read(mmap_sem).
2662 */
2663 if (pte_alloc(mm, pmdp, addr))
2664 goto abort;
2665
2666 /* See the comment in pte_alloc_one_map() */
2667 if (unlikely(pmd_trans_unstable(pmdp)))
2668 goto abort;
2669
2670 if (unlikely(anon_vma_prepare(vma)))
2671 goto abort;
2672 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2673 goto abort;
2674
2675 /*
2676 * The memory barrier inside __SetPageUptodate makes sure that
2677 * preceding stores to the page contents become visible before
2678 * the set_pte_at() write.
2679 */
2680 __SetPageUptodate(page);
2681
2682 if (is_zone_device_page(page)) {
2683 if (is_device_private_page(page)) {
2684 swp_entry_t swp_entry;
2685
2686 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2687 entry = swp_entry_to_pte(swp_entry);
2688 } else if (is_device_public_page(page)) {
2689 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2690 if (vma->vm_flags & VM_WRITE)
2691 entry = pte_mkwrite(pte_mkdirty(entry));
2692 entry = pte_mkdevmap(entry);
2693 }
2694 } else {
2695 entry = mk_pte(page, vma->vm_page_prot);
2696 if (vma->vm_flags & VM_WRITE)
2697 entry = pte_mkwrite(pte_mkdirty(entry));
2698 }
2699
2700 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2701
2702 if (pte_present(*ptep)) {
2703 unsigned long pfn = pte_pfn(*ptep);
2704
2705 if (!is_zero_pfn(pfn)) {
2706 pte_unmap_unlock(ptep, ptl);
2707 mem_cgroup_cancel_charge(page, memcg, false);
2708 goto abort;
2709 }
2710 flush = true;
2711 } else if (!pte_none(*ptep)) {
2712 pte_unmap_unlock(ptep, ptl);
2713 mem_cgroup_cancel_charge(page, memcg, false);
2714 goto abort;
2715 }
2716
2717 /*
2718 * Check for usefaultfd but do not deliver the fault. Instead,
2719 * just back off.
2720 */
2721 if (userfaultfd_missing(vma)) {
2722 pte_unmap_unlock(ptep, ptl);
2723 mem_cgroup_cancel_charge(page, memcg, false);
2724 goto abort;
2725 }
2726
2727 inc_mm_counter(mm, MM_ANONPAGES);
2728 page_add_new_anon_rmap(page, vma, addr, false);
2729 mem_cgroup_commit_charge(page, memcg, false, false);
2730 if (!is_zone_device_page(page))
2731 lru_cache_add_active_or_unevictable(page, vma);
2732 get_page(page);
2733
2734 if (flush) {
2735 flush_cache_page(vma, addr, pte_pfn(*ptep));
2736 ptep_clear_flush_notify(vma, addr, ptep);
2737 set_pte_at_notify(mm, addr, ptep, entry);
2738 update_mmu_cache(vma, addr, ptep);
2739 } else {
2740 /* No need to invalidate - it was non-present before */
2741 set_pte_at(mm, addr, ptep, entry);
2742 update_mmu_cache(vma, addr, ptep);
2743 }
2744
2745 pte_unmap_unlock(ptep, ptl);
2746 *src = MIGRATE_PFN_MIGRATE;
2747 return;
2748
2749abort:
2750 *src &= ~MIGRATE_PFN_MIGRATE;
2751}
2752
2753/*
2754 * migrate_vma_pages() - migrate meta-data from src page to dst page
2755 * @migrate: migrate struct containing all migration information
2756 *
2757 * This migrates struct page meta-data from source struct page to destination
2758 * struct page. This effectively finishes the migration from source page to the
2759 * destination page.
2760 */
2761static void migrate_vma_pages(struct migrate_vma *migrate)
2762{
2763 const unsigned long npages = migrate->npages;
2764 const unsigned long start = migrate->start;
2765 struct vm_area_struct *vma = migrate->vma;
2766 struct mm_struct *mm = vma->vm_mm;
2767 unsigned long addr, i, mmu_start;
2768 bool notified = false;
2769
2770 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2771 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2772 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2773 struct address_space *mapping;
2774 int r;
2775
2776 if (!newpage) {
2777 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2778 continue;
2779 }
2780
2781 if (!page) {
2782 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2783 continue;
2784 }
2785 if (!notified) {
2786 mmu_start = addr;
2787 notified = true;
2788 mmu_notifier_invalidate_range_start(mm,
2789 mmu_start,
2790 migrate->end);
2791 }
2792 migrate_vma_insert_page(migrate, addr, newpage,
2793 &migrate->src[i],
2794 &migrate->dst[i]);
2795 continue;
2796 }
2797
2798 mapping = page_mapping(page);
2799
2800 if (is_zone_device_page(newpage)) {
2801 if (is_device_private_page(newpage)) {
2802 /*
2803 * For now only support private anonymous when
2804 * migrating to un-addressable device memory.
2805 */
2806 if (mapping) {
2807 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2808 continue;
2809 }
2810 } else if (!is_device_public_page(newpage)) {
2811 /*
2812 * Other types of ZONE_DEVICE page are not
2813 * supported.
2814 */
2815 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2816 continue;
2817 }
2818 }
2819
2820 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2821 if (r != MIGRATEPAGE_SUCCESS)
2822 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2823 }
2824
2825 /*
2826 * No need to double call mmu_notifier->invalidate_range() callback as
2827 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2828 * did already call it.
2829 */
2830 if (notified)
2831 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2832 migrate->end);
2833}
2834
2835/*
2836 * migrate_vma_finalize() - restore CPU page table entry
2837 * @migrate: migrate struct containing all migration information
2838 *
2839 * This replaces the special migration pte entry with either a mapping to the
2840 * new page if migration was successful for that page, or to the original page
2841 * otherwise.
2842 *
2843 * This also unlocks the pages and puts them back on the lru, or drops the extra
2844 * refcount, for device pages.
2845 */
2846static void migrate_vma_finalize(struct migrate_vma *migrate)
2847{
2848 const unsigned long npages = migrate->npages;
2849 unsigned long i;
2850
2851 for (i = 0; i < npages; i++) {
2852 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2853 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2854
2855 if (!page) {
2856 if (newpage) {
2857 unlock_page(newpage);
2858 put_page(newpage);
2859 }
2860 continue;
2861 }
2862
2863 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2864 if (newpage) {
2865 unlock_page(newpage);
2866 put_page(newpage);
2867 }
2868 newpage = page;
2869 }
2870
2871 remove_migration_ptes(page, newpage, false);
2872 unlock_page(page);
2873 migrate->cpages--;
2874
2875 if (is_zone_device_page(page))
2876 put_page(page);
2877 else
2878 putback_lru_page(page);
2879
2880 if (newpage != page) {
2881 unlock_page(newpage);
2882 if (is_zone_device_page(newpage))
2883 put_page(newpage);
2884 else
2885 putback_lru_page(newpage);
2886 }
2887 }
2888}
2889
2890/*
2891 * migrate_vma() - migrate a range of memory inside vma
2892 *
2893 * @ops: migration callback for allocating destination memory and copying
2894 * @vma: virtual memory area containing the range to be migrated
2895 * @start: start address of the range to migrate (inclusive)
2896 * @end: end address of the range to migrate (exclusive)
2897 * @src: array of hmm_pfn_t containing source pfns
2898 * @dst: array of hmm_pfn_t containing destination pfns
2899 * @private: pointer passed back to each of the callback
2900 * Returns: 0 on success, error code otherwise
2901 *
2902 * This function tries to migrate a range of memory virtual address range, using
2903 * callbacks to allocate and copy memory from source to destination. First it
2904 * collects all the pages backing each virtual address in the range, saving this
2905 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2906 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2907 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2908 * in the corresponding src array entry. It then restores any pages that are
2909 * pinned, by remapping and unlocking those pages.
2910 *
2911 * At this point it calls the alloc_and_copy() callback. For documentation on
2912 * what is expected from that callback, see struct migrate_vma_ops comments in
2913 * include/linux/migrate.h
2914 *
2915 * After the alloc_and_copy() callback, this function goes over each entry in
2916 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2917 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2918 * then the function tries to migrate struct page information from the source
2919 * struct page to the destination struct page. If it fails to migrate the struct
2920 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2921 * array.
2922 *
2923 * At this point all successfully migrated pages have an entry in the src
2924 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2925 * array entry with MIGRATE_PFN_VALID flag set.
2926 *
2927 * It then calls the finalize_and_map() callback. See comments for "struct
2928 * migrate_vma_ops", in include/linux/migrate.h for details about
2929 * finalize_and_map() behavior.
2930 *
2931 * After the finalize_and_map() callback, for successfully migrated pages, this
2932 * function updates the CPU page table to point to new pages, otherwise it
2933 * restores the CPU page table to point to the original source pages.
2934 *
2935 * Function returns 0 after the above steps, even if no pages were migrated
2936 * (The function only returns an error if any of the arguments are invalid.)
2937 *
2938 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2939 * unsigned long entries.
2940 */
2941int migrate_vma(const struct migrate_vma_ops *ops,
2942 struct vm_area_struct *vma,
2943 unsigned long start,
2944 unsigned long end,
2945 unsigned long *src,
2946 unsigned long *dst,
2947 void *private)
2948{
2949 struct migrate_vma migrate;
2950
2951 /* Sanity check the arguments */
2952 start &= PAGE_MASK;
2953 end &= PAGE_MASK;
2954 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL))
2955 return -EINVAL;
2956 if (start < vma->vm_start || start >= vma->vm_end)
2957 return -EINVAL;
2958 if (end <= vma->vm_start || end > vma->vm_end)
2959 return -EINVAL;
2960 if (!ops || !src || !dst || start >= end)
2961 return -EINVAL;
2962
2963 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2964 migrate.src = src;
2965 migrate.dst = dst;
2966 migrate.start = start;
2967 migrate.npages = 0;
2968 migrate.cpages = 0;
2969 migrate.end = end;
2970 migrate.vma = vma;
2971
2972 /* Collect, and try to unmap source pages */
2973 migrate_vma_collect(&migrate);
2974 if (!migrate.cpages)
2975 return 0;
2976
2977 /* Lock and isolate page */
2978 migrate_vma_prepare(&migrate);
2979 if (!migrate.cpages)
2980 return 0;
2981
2982 /* Unmap pages */
2983 migrate_vma_unmap(&migrate);
2984 if (!migrate.cpages)
2985 return 0;
2986
2987 /*
2988 * At this point pages are locked and unmapped, and thus they have
2989 * stable content and can safely be copied to destination memory that
2990 * is allocated by the callback.
2991 *
2992 * Note that migration can fail in migrate_vma_struct_page() for each
2993 * individual page.
2994 */
2995 ops->alloc_and_copy(vma, src, dst, start, end, private);
2996
2997 /* This does the real migration of struct page */
2998 migrate_vma_pages(&migrate);
2999
3000 ops->finalize_and_map(vma, src, dst, start, end, private);
3001
3002 /* Unlock and remap pages */
3003 migrate_vma_finalize(&migrate);
3004
3005 return 0;
3006}
3007EXPORT_SYMBOL(migrate_vma);
3008#endif /* defined(MIGRATE_VMA_HELPER) */
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/hugetlb_cgroup.h>
37#include <linux/gfp.h>
38#include <linux/balloon_compaction.h>
39#include <linux/mmu_notifier.h>
40
41#include <asm/tlbflush.h>
42
43#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
46#include "internal.h"
47
48/*
49 * migrate_prep() needs to be called before we start compiling a list of pages
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
52 */
53int migrate_prep(void)
54{
55 /*
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
60 */
61 lru_add_drain_all();
62
63 return 0;
64}
65
66/* Do the necessary work of migrate_prep but not if it involves other CPUs */
67int migrate_prep_local(void)
68{
69 lru_add_drain();
70
71 return 0;
72}
73
74/*
75 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
78 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
81 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
87 list_for_each_entry_safe(page, page2, l, lru) {
88 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
92 list_del(&page->lru);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
95 if (unlikely(isolated_balloon_page(page)))
96 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
99 }
100}
101
102/*
103 * Restore a potential migration pte to a working pte entry
104 */
105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
119 } else {
120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
122 goto out;
123 if (pmd_trans_huge(*pmd))
124 goto out;
125
126 ptep = pte_offset_map(pmd, addr);
127
128 /*
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
131 */
132
133 ptl = pte_lockptr(mm, pmd);
134 }
135
136 spin_lock(ptl);
137 pte = *ptep;
138 if (!is_swap_pte(pte))
139 goto unlock;
140
141 entry = pte_to_swp_entry(pte);
142
143 if (!is_migration_entry(entry) ||
144 migration_entry_to_page(entry) != old)
145 goto unlock;
146
147 get_page(new);
148 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
149 if (pte_swp_soft_dirty(*ptep))
150 pte = pte_mksoft_dirty(pte);
151 if (is_write_migration_entry(entry))
152 pte = pte_mkwrite(pte);
153#ifdef CONFIG_HUGETLB_PAGE
154 if (PageHuge(new)) {
155 pte = pte_mkhuge(pte);
156 pte = arch_make_huge_pte(pte, vma, new, 0);
157 }
158#endif
159 flush_dcache_page(new);
160 set_pte_at(mm, addr, ptep, pte);
161
162 if (PageHuge(new)) {
163 if (PageAnon(new))
164 hugepage_add_anon_rmap(new, vma, addr);
165 else
166 page_dup_rmap(new);
167 } else if (PageAnon(new))
168 page_add_anon_rmap(new, vma, addr);
169 else
170 page_add_file_rmap(new);
171
172 /* No need to invalidate - it was non-present before */
173 update_mmu_cache(vma, addr, ptep);
174unlock:
175 pte_unmap_unlock(ptep, ptl);
176out:
177 return SWAP_AGAIN;
178}
179
180/*
181 * Congratulations to trinity for discovering this bug.
182 * mm/fremap.c's remap_file_pages() accepts any range within a single vma to
183 * convert that vma to VM_NONLINEAR; and generic_file_remap_pages() will then
184 * replace the specified range by file ptes throughout (maybe populated after).
185 * If page migration finds a page within that range, while it's still located
186 * by vma_interval_tree rather than lost to i_mmap_nonlinear list, no problem:
187 * zap_pte() clears the temporary migration entry before mmap_sem is dropped.
188 * But if the migrating page is in a part of the vma outside the range to be
189 * remapped, then it will not be cleared, and remove_migration_ptes() needs to
190 * deal with it. Fortunately, this part of the vma is of course still linear,
191 * so we just need to use linear location on the nonlinear list.
192 */
193static int remove_linear_migration_ptes_from_nonlinear(struct page *page,
194 struct address_space *mapping, void *arg)
195{
196 struct vm_area_struct *vma;
197 /* hugetlbfs does not support remap_pages, so no huge pgoff worries */
198 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
199 unsigned long addr;
200
201 list_for_each_entry(vma,
202 &mapping->i_mmap_nonlinear, shared.nonlinear) {
203
204 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
205 if (addr >= vma->vm_start && addr < vma->vm_end)
206 remove_migration_pte(page, vma, addr, arg);
207 }
208 return SWAP_AGAIN;
209}
210
211/*
212 * Get rid of all migration entries and replace them by
213 * references to the indicated page.
214 */
215static void remove_migration_ptes(struct page *old, struct page *new)
216{
217 struct rmap_walk_control rwc = {
218 .rmap_one = remove_migration_pte,
219 .arg = old,
220 .file_nonlinear = remove_linear_migration_ptes_from_nonlinear,
221 };
222
223 rmap_walk(new, &rwc);
224}
225
226/*
227 * Something used the pte of a page under migration. We need to
228 * get to the page and wait until migration is finished.
229 * When we return from this function the fault will be retried.
230 */
231static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
232 spinlock_t *ptl)
233{
234 pte_t pte;
235 swp_entry_t entry;
236 struct page *page;
237
238 spin_lock(ptl);
239 pte = *ptep;
240 if (!is_swap_pte(pte))
241 goto out;
242
243 entry = pte_to_swp_entry(pte);
244 if (!is_migration_entry(entry))
245 goto out;
246
247 page = migration_entry_to_page(entry);
248
249 /*
250 * Once radix-tree replacement of page migration started, page_count
251 * *must* be zero. And, we don't want to call wait_on_page_locked()
252 * against a page without get_page().
253 * So, we use get_page_unless_zero(), here. Even failed, page fault
254 * will occur again.
255 */
256 if (!get_page_unless_zero(page))
257 goto out;
258 pte_unmap_unlock(ptep, ptl);
259 wait_on_page_locked(page);
260 put_page(page);
261 return;
262out:
263 pte_unmap_unlock(ptep, ptl);
264}
265
266void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
267 unsigned long address)
268{
269 spinlock_t *ptl = pte_lockptr(mm, pmd);
270 pte_t *ptep = pte_offset_map(pmd, address);
271 __migration_entry_wait(mm, ptep, ptl);
272}
273
274void migration_entry_wait_huge(struct vm_area_struct *vma,
275 struct mm_struct *mm, pte_t *pte)
276{
277 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
278 __migration_entry_wait(mm, pte, ptl);
279}
280
281#ifdef CONFIG_BLOCK
282/* Returns true if all buffers are successfully locked */
283static bool buffer_migrate_lock_buffers(struct buffer_head *head,
284 enum migrate_mode mode)
285{
286 struct buffer_head *bh = head;
287
288 /* Simple case, sync compaction */
289 if (mode != MIGRATE_ASYNC) {
290 do {
291 get_bh(bh);
292 lock_buffer(bh);
293 bh = bh->b_this_page;
294
295 } while (bh != head);
296
297 return true;
298 }
299
300 /* async case, we cannot block on lock_buffer so use trylock_buffer */
301 do {
302 get_bh(bh);
303 if (!trylock_buffer(bh)) {
304 /*
305 * We failed to lock the buffer and cannot stall in
306 * async migration. Release the taken locks
307 */
308 struct buffer_head *failed_bh = bh;
309 put_bh(failed_bh);
310 bh = head;
311 while (bh != failed_bh) {
312 unlock_buffer(bh);
313 put_bh(bh);
314 bh = bh->b_this_page;
315 }
316 return false;
317 }
318
319 bh = bh->b_this_page;
320 } while (bh != head);
321 return true;
322}
323#else
324static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
325 enum migrate_mode mode)
326{
327 return true;
328}
329#endif /* CONFIG_BLOCK */
330
331/*
332 * Replace the page in the mapping.
333 *
334 * The number of remaining references must be:
335 * 1 for anonymous pages without a mapping
336 * 2 for pages with a mapping
337 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
338 */
339int migrate_page_move_mapping(struct address_space *mapping,
340 struct page *newpage, struct page *page,
341 struct buffer_head *head, enum migrate_mode mode,
342 int extra_count)
343{
344 int expected_count = 1 + extra_count;
345 void **pslot;
346
347 if (!mapping) {
348 /* Anonymous page without mapping */
349 if (page_count(page) != expected_count)
350 return -EAGAIN;
351 return MIGRATEPAGE_SUCCESS;
352 }
353
354 spin_lock_irq(&mapping->tree_lock);
355
356 pslot = radix_tree_lookup_slot(&mapping->page_tree,
357 page_index(page));
358
359 expected_count += 1 + page_has_private(page);
360 if (page_count(page) != expected_count ||
361 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
362 spin_unlock_irq(&mapping->tree_lock);
363 return -EAGAIN;
364 }
365
366 if (!page_freeze_refs(page, expected_count)) {
367 spin_unlock_irq(&mapping->tree_lock);
368 return -EAGAIN;
369 }
370
371 /*
372 * In the async migration case of moving a page with buffers, lock the
373 * buffers using trylock before the mapping is moved. If the mapping
374 * was moved, we later failed to lock the buffers and could not move
375 * the mapping back due to an elevated page count, we would have to
376 * block waiting on other references to be dropped.
377 */
378 if (mode == MIGRATE_ASYNC && head &&
379 !buffer_migrate_lock_buffers(head, mode)) {
380 page_unfreeze_refs(page, expected_count);
381 spin_unlock_irq(&mapping->tree_lock);
382 return -EAGAIN;
383 }
384
385 /*
386 * Now we know that no one else is looking at the page.
387 */
388 get_page(newpage); /* add cache reference */
389 if (PageSwapCache(page)) {
390 SetPageSwapCache(newpage);
391 set_page_private(newpage, page_private(page));
392 }
393
394 radix_tree_replace_slot(pslot, newpage);
395
396 /*
397 * Drop cache reference from old page by unfreezing
398 * to one less reference.
399 * We know this isn't the last reference.
400 */
401 page_unfreeze_refs(page, expected_count - 1);
402
403 /*
404 * If moved to a different zone then also account
405 * the page for that zone. Other VM counters will be
406 * taken care of when we establish references to the
407 * new page and drop references to the old page.
408 *
409 * Note that anonymous pages are accounted for
410 * via NR_FILE_PAGES and NR_ANON_PAGES if they
411 * are mapped to swap space.
412 */
413 __dec_zone_page_state(page, NR_FILE_PAGES);
414 __inc_zone_page_state(newpage, NR_FILE_PAGES);
415 if (!PageSwapCache(page) && PageSwapBacked(page)) {
416 __dec_zone_page_state(page, NR_SHMEM);
417 __inc_zone_page_state(newpage, NR_SHMEM);
418 }
419 spin_unlock_irq(&mapping->tree_lock);
420
421 return MIGRATEPAGE_SUCCESS;
422}
423
424/*
425 * The expected number of remaining references is the same as that
426 * of migrate_page_move_mapping().
427 */
428int migrate_huge_page_move_mapping(struct address_space *mapping,
429 struct page *newpage, struct page *page)
430{
431 int expected_count;
432 void **pslot;
433
434 if (!mapping) {
435 if (page_count(page) != 1)
436 return -EAGAIN;
437 return MIGRATEPAGE_SUCCESS;
438 }
439
440 spin_lock_irq(&mapping->tree_lock);
441
442 pslot = radix_tree_lookup_slot(&mapping->page_tree,
443 page_index(page));
444
445 expected_count = 2 + page_has_private(page);
446 if (page_count(page) != expected_count ||
447 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
448 spin_unlock_irq(&mapping->tree_lock);
449 return -EAGAIN;
450 }
451
452 if (!page_freeze_refs(page, expected_count)) {
453 spin_unlock_irq(&mapping->tree_lock);
454 return -EAGAIN;
455 }
456
457 get_page(newpage);
458
459 radix_tree_replace_slot(pslot, newpage);
460
461 page_unfreeze_refs(page, expected_count - 1);
462
463 spin_unlock_irq(&mapping->tree_lock);
464 return MIGRATEPAGE_SUCCESS;
465}
466
467/*
468 * Gigantic pages are so large that we do not guarantee that page++ pointer
469 * arithmetic will work across the entire page. We need something more
470 * specialized.
471 */
472static void __copy_gigantic_page(struct page *dst, struct page *src,
473 int nr_pages)
474{
475 int i;
476 struct page *dst_base = dst;
477 struct page *src_base = src;
478
479 for (i = 0; i < nr_pages; ) {
480 cond_resched();
481 copy_highpage(dst, src);
482
483 i++;
484 dst = mem_map_next(dst, dst_base, i);
485 src = mem_map_next(src, src_base, i);
486 }
487}
488
489static void copy_huge_page(struct page *dst, struct page *src)
490{
491 int i;
492 int nr_pages;
493
494 if (PageHuge(src)) {
495 /* hugetlbfs page */
496 struct hstate *h = page_hstate(src);
497 nr_pages = pages_per_huge_page(h);
498
499 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
500 __copy_gigantic_page(dst, src, nr_pages);
501 return;
502 }
503 } else {
504 /* thp page */
505 BUG_ON(!PageTransHuge(src));
506 nr_pages = hpage_nr_pages(src);
507 }
508
509 for (i = 0; i < nr_pages; i++) {
510 cond_resched();
511 copy_highpage(dst + i, src + i);
512 }
513}
514
515/*
516 * Copy the page to its new location
517 */
518void migrate_page_copy(struct page *newpage, struct page *page)
519{
520 int cpupid;
521
522 if (PageHuge(page) || PageTransHuge(page))
523 copy_huge_page(newpage, page);
524 else
525 copy_highpage(newpage, page);
526
527 if (PageError(page))
528 SetPageError(newpage);
529 if (PageReferenced(page))
530 SetPageReferenced(newpage);
531 if (PageUptodate(page))
532 SetPageUptodate(newpage);
533 if (TestClearPageActive(page)) {
534 VM_BUG_ON_PAGE(PageUnevictable(page), page);
535 SetPageActive(newpage);
536 } else if (TestClearPageUnevictable(page))
537 SetPageUnevictable(newpage);
538 if (PageChecked(page))
539 SetPageChecked(newpage);
540 if (PageMappedToDisk(page))
541 SetPageMappedToDisk(newpage);
542
543 if (PageDirty(page)) {
544 clear_page_dirty_for_io(page);
545 /*
546 * Want to mark the page and the radix tree as dirty, and
547 * redo the accounting that clear_page_dirty_for_io undid,
548 * but we can't use set_page_dirty because that function
549 * is actually a signal that all of the page has become dirty.
550 * Whereas only part of our page may be dirty.
551 */
552 if (PageSwapBacked(page))
553 SetPageDirty(newpage);
554 else
555 __set_page_dirty_nobuffers(newpage);
556 }
557
558 /*
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
561 */
562 cpupid = page_cpupid_xchg_last(page, -1);
563 page_cpupid_xchg_last(newpage, cpupid);
564
565 mlock_migrate_page(newpage, page);
566 ksm_migrate_page(newpage, page);
567 /*
568 * Please do not reorder this without considering how mm/ksm.c's
569 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
570 */
571 ClearPageSwapCache(page);
572 ClearPagePrivate(page);
573 set_page_private(page, 0);
574
575 /*
576 * If any waiters have accumulated on the new page then
577 * wake them up.
578 */
579 if (PageWriteback(newpage))
580 end_page_writeback(newpage);
581}
582
583/************************************************************
584 * Migration functions
585 ***********************************************************/
586
587/*
588 * Common logic to directly migrate a single page suitable for
589 * pages that do not use PagePrivate/PagePrivate2.
590 *
591 * Pages are locked upon entry and exit.
592 */
593int migrate_page(struct address_space *mapping,
594 struct page *newpage, struct page *page,
595 enum migrate_mode mode)
596{
597 int rc;
598
599 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
600
601 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
602
603 if (rc != MIGRATEPAGE_SUCCESS)
604 return rc;
605
606 migrate_page_copy(newpage, page);
607 return MIGRATEPAGE_SUCCESS;
608}
609EXPORT_SYMBOL(migrate_page);
610
611#ifdef CONFIG_BLOCK
612/*
613 * Migration function for pages with buffers. This function can only be used
614 * if the underlying filesystem guarantees that no other references to "page"
615 * exist.
616 */
617int buffer_migrate_page(struct address_space *mapping,
618 struct page *newpage, struct page *page, enum migrate_mode mode)
619{
620 struct buffer_head *bh, *head;
621 int rc;
622
623 if (!page_has_buffers(page))
624 return migrate_page(mapping, newpage, page, mode);
625
626 head = page_buffers(page);
627
628 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
629
630 if (rc != MIGRATEPAGE_SUCCESS)
631 return rc;
632
633 /*
634 * In the async case, migrate_page_move_mapping locked the buffers
635 * with an IRQ-safe spinlock held. In the sync case, the buffers
636 * need to be locked now
637 */
638 if (mode != MIGRATE_ASYNC)
639 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
640
641 ClearPagePrivate(page);
642 set_page_private(newpage, page_private(page));
643 set_page_private(page, 0);
644 put_page(page);
645 get_page(newpage);
646
647 bh = head;
648 do {
649 set_bh_page(bh, newpage, bh_offset(bh));
650 bh = bh->b_this_page;
651
652 } while (bh != head);
653
654 SetPagePrivate(newpage);
655
656 migrate_page_copy(newpage, page);
657
658 bh = head;
659 do {
660 unlock_buffer(bh);
661 put_bh(bh);
662 bh = bh->b_this_page;
663
664 } while (bh != head);
665
666 return MIGRATEPAGE_SUCCESS;
667}
668EXPORT_SYMBOL(buffer_migrate_page);
669#endif
670
671/*
672 * Writeback a page to clean the dirty state
673 */
674static int writeout(struct address_space *mapping, struct page *page)
675{
676 struct writeback_control wbc = {
677 .sync_mode = WB_SYNC_NONE,
678 .nr_to_write = 1,
679 .range_start = 0,
680 .range_end = LLONG_MAX,
681 .for_reclaim = 1
682 };
683 int rc;
684
685 if (!mapping->a_ops->writepage)
686 /* No write method for the address space */
687 return -EINVAL;
688
689 if (!clear_page_dirty_for_io(page))
690 /* Someone else already triggered a write */
691 return -EAGAIN;
692
693 /*
694 * A dirty page may imply that the underlying filesystem has
695 * the page on some queue. So the page must be clean for
696 * migration. Writeout may mean we loose the lock and the
697 * page state is no longer what we checked for earlier.
698 * At this point we know that the migration attempt cannot
699 * be successful.
700 */
701 remove_migration_ptes(page, page);
702
703 rc = mapping->a_ops->writepage(page, &wbc);
704
705 if (rc != AOP_WRITEPAGE_ACTIVATE)
706 /* unlocked. Relock */
707 lock_page(page);
708
709 return (rc < 0) ? -EIO : -EAGAIN;
710}
711
712/*
713 * Default handling if a filesystem does not provide a migration function.
714 */
715static int fallback_migrate_page(struct address_space *mapping,
716 struct page *newpage, struct page *page, enum migrate_mode mode)
717{
718 if (PageDirty(page)) {
719 /* Only writeback pages in full synchronous migration */
720 if (mode != MIGRATE_SYNC)
721 return -EBUSY;
722 return writeout(mapping, page);
723 }
724
725 /*
726 * Buffers may be managed in a filesystem specific way.
727 * We must have no buffers or drop them.
728 */
729 if (page_has_private(page) &&
730 !try_to_release_page(page, GFP_KERNEL))
731 return -EAGAIN;
732
733 return migrate_page(mapping, newpage, page, mode);
734}
735
736/*
737 * Move a page to a newly allocated page
738 * The page is locked and all ptes have been successfully removed.
739 *
740 * The new page will have replaced the old page if this function
741 * is successful.
742 *
743 * Return value:
744 * < 0 - error code
745 * MIGRATEPAGE_SUCCESS - success
746 */
747static int move_to_new_page(struct page *newpage, struct page *page,
748 int remap_swapcache, enum migrate_mode mode)
749{
750 struct address_space *mapping;
751 int rc;
752
753 /*
754 * Block others from accessing the page when we get around to
755 * establishing additional references. We are the only one
756 * holding a reference to the new page at this point.
757 */
758 if (!trylock_page(newpage))
759 BUG();
760
761 /* Prepare mapping for the new page.*/
762 newpage->index = page->index;
763 newpage->mapping = page->mapping;
764 if (PageSwapBacked(page))
765 SetPageSwapBacked(newpage);
766
767 mapping = page_mapping(page);
768 if (!mapping)
769 rc = migrate_page(mapping, newpage, page, mode);
770 else if (mapping->a_ops->migratepage)
771 /*
772 * Most pages have a mapping and most filesystems provide a
773 * migratepage callback. Anonymous pages are part of swap
774 * space which also has its own migratepage callback. This
775 * is the most common path for page migration.
776 */
777 rc = mapping->a_ops->migratepage(mapping,
778 newpage, page, mode);
779 else
780 rc = fallback_migrate_page(mapping, newpage, page, mode);
781
782 if (rc != MIGRATEPAGE_SUCCESS) {
783 newpage->mapping = NULL;
784 } else {
785 if (remap_swapcache)
786 remove_migration_ptes(page, newpage);
787 page->mapping = NULL;
788 }
789
790 unlock_page(newpage);
791
792 return rc;
793}
794
795static int __unmap_and_move(struct page *page, struct page *newpage,
796 int force, enum migrate_mode mode)
797{
798 int rc = -EAGAIN;
799 int remap_swapcache = 1;
800 struct mem_cgroup *mem;
801 struct anon_vma *anon_vma = NULL;
802
803 if (!trylock_page(page)) {
804 if (!force || mode == MIGRATE_ASYNC)
805 goto out;
806
807 /*
808 * It's not safe for direct compaction to call lock_page.
809 * For example, during page readahead pages are added locked
810 * to the LRU. Later, when the IO completes the pages are
811 * marked uptodate and unlocked. However, the queueing
812 * could be merging multiple pages for one bio (e.g.
813 * mpage_readpages). If an allocation happens for the
814 * second or third page, the process can end up locking
815 * the same page twice and deadlocking. Rather than
816 * trying to be clever about what pages can be locked,
817 * avoid the use of lock_page for direct compaction
818 * altogether.
819 */
820 if (current->flags & PF_MEMALLOC)
821 goto out;
822
823 lock_page(page);
824 }
825
826 /* charge against new page */
827 mem_cgroup_prepare_migration(page, newpage, &mem);
828
829 if (PageWriteback(page)) {
830 /*
831 * Only in the case of a full synchronous migration is it
832 * necessary to wait for PageWriteback. In the async case,
833 * the retry loop is too short and in the sync-light case,
834 * the overhead of stalling is too much
835 */
836 if (mode != MIGRATE_SYNC) {
837 rc = -EBUSY;
838 goto uncharge;
839 }
840 if (!force)
841 goto uncharge;
842 wait_on_page_writeback(page);
843 }
844 /*
845 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
846 * we cannot notice that anon_vma is freed while we migrates a page.
847 * This get_anon_vma() delays freeing anon_vma pointer until the end
848 * of migration. File cache pages are no problem because of page_lock()
849 * File Caches may use write_page() or lock_page() in migration, then,
850 * just care Anon page here.
851 */
852 if (PageAnon(page) && !PageKsm(page)) {
853 /*
854 * Only page_lock_anon_vma_read() understands the subtleties of
855 * getting a hold on an anon_vma from outside one of its mms.
856 */
857 anon_vma = page_get_anon_vma(page);
858 if (anon_vma) {
859 /*
860 * Anon page
861 */
862 } else if (PageSwapCache(page)) {
863 /*
864 * We cannot be sure that the anon_vma of an unmapped
865 * swapcache page is safe to use because we don't
866 * know in advance if the VMA that this page belonged
867 * to still exists. If the VMA and others sharing the
868 * data have been freed, then the anon_vma could
869 * already be invalid.
870 *
871 * To avoid this possibility, swapcache pages get
872 * migrated but are not remapped when migration
873 * completes
874 */
875 remap_swapcache = 0;
876 } else {
877 goto uncharge;
878 }
879 }
880
881 if (unlikely(balloon_page_movable(page))) {
882 /*
883 * A ballooned page does not need any special attention from
884 * physical to virtual reverse mapping procedures.
885 * Skip any attempt to unmap PTEs or to remap swap cache,
886 * in order to avoid burning cycles at rmap level, and perform
887 * the page migration right away (proteced by page lock).
888 */
889 rc = balloon_page_migrate(newpage, page, mode);
890 goto uncharge;
891 }
892
893 /*
894 * Corner case handling:
895 * 1. When a new swap-cache page is read into, it is added to the LRU
896 * and treated as swapcache but it has no rmap yet.
897 * Calling try_to_unmap() against a page->mapping==NULL page will
898 * trigger a BUG. So handle it here.
899 * 2. An orphaned page (see truncate_complete_page) might have
900 * fs-private metadata. The page can be picked up due to memory
901 * offlining. Everywhere else except page reclaim, the page is
902 * invisible to the vm, so the page can not be migrated. So try to
903 * free the metadata, so the page can be freed.
904 */
905 if (!page->mapping) {
906 VM_BUG_ON_PAGE(PageAnon(page), page);
907 if (page_has_private(page)) {
908 try_to_free_buffers(page);
909 goto uncharge;
910 }
911 goto skip_unmap;
912 }
913
914 /* Establish migration ptes or remove ptes */
915 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
916
917skip_unmap:
918 if (!page_mapped(page))
919 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
920
921 if (rc && remap_swapcache)
922 remove_migration_ptes(page, page);
923
924 /* Drop an anon_vma reference if we took one */
925 if (anon_vma)
926 put_anon_vma(anon_vma);
927
928uncharge:
929 mem_cgroup_end_migration(mem, page, newpage,
930 (rc == MIGRATEPAGE_SUCCESS ||
931 rc == MIGRATEPAGE_BALLOON_SUCCESS));
932 unlock_page(page);
933out:
934 return rc;
935}
936
937/*
938 * Obtain the lock on page, remove all ptes and migrate the page
939 * to the newly allocated page in newpage.
940 */
941static int unmap_and_move(new_page_t get_new_page, unsigned long private,
942 struct page *page, int force, enum migrate_mode mode)
943{
944 int rc = 0;
945 int *result = NULL;
946 struct page *newpage = get_new_page(page, private, &result);
947
948 if (!newpage)
949 return -ENOMEM;
950
951 if (page_count(page) == 1) {
952 /* page was freed from under us. So we are done. */
953 goto out;
954 }
955
956 if (unlikely(PageTransHuge(page)))
957 if (unlikely(split_huge_page(page)))
958 goto out;
959
960 rc = __unmap_and_move(page, newpage, force, mode);
961
962 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
963 /*
964 * A ballooned page has been migrated already.
965 * Now, it's the time to wrap-up counters,
966 * handle the page back to Buddy and return.
967 */
968 dec_zone_page_state(page, NR_ISOLATED_ANON +
969 page_is_file_cache(page));
970 balloon_page_free(page);
971 return MIGRATEPAGE_SUCCESS;
972 }
973out:
974 if (rc != -EAGAIN) {
975 /*
976 * A page that has been migrated has all references
977 * removed and will be freed. A page that has not been
978 * migrated will have kepts its references and be
979 * restored.
980 */
981 list_del(&page->lru);
982 dec_zone_page_state(page, NR_ISOLATED_ANON +
983 page_is_file_cache(page));
984 putback_lru_page(page);
985 }
986 /*
987 * Move the new page to the LRU. If migration was not successful
988 * then this will free the page.
989 */
990 putback_lru_page(newpage);
991 if (result) {
992 if (rc)
993 *result = rc;
994 else
995 *result = page_to_nid(newpage);
996 }
997 return rc;
998}
999
1000/*
1001 * Counterpart of unmap_and_move_page() for hugepage migration.
1002 *
1003 * This function doesn't wait the completion of hugepage I/O
1004 * because there is no race between I/O and migration for hugepage.
1005 * Note that currently hugepage I/O occurs only in direct I/O
1006 * where no lock is held and PG_writeback is irrelevant,
1007 * and writeback status of all subpages are counted in the reference
1008 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1009 * under direct I/O, the reference of the head page is 512 and a bit more.)
1010 * This means that when we try to migrate hugepage whose subpages are
1011 * doing direct I/O, some references remain after try_to_unmap() and
1012 * hugepage migration fails without data corruption.
1013 *
1014 * There is also no race when direct I/O is issued on the page under migration,
1015 * because then pte is replaced with migration swap entry and direct I/O code
1016 * will wait in the page fault for migration to complete.
1017 */
1018static int unmap_and_move_huge_page(new_page_t get_new_page,
1019 unsigned long private, struct page *hpage,
1020 int force, enum migrate_mode mode)
1021{
1022 int rc = 0;
1023 int *result = NULL;
1024 struct page *new_hpage;
1025 struct anon_vma *anon_vma = NULL;
1026
1027 /*
1028 * Movability of hugepages depends on architectures and hugepage size.
1029 * This check is necessary because some callers of hugepage migration
1030 * like soft offline and memory hotremove don't walk through page
1031 * tables or check whether the hugepage is pmd-based or not before
1032 * kicking migration.
1033 */
1034 if (!hugepage_migration_support(page_hstate(hpage))) {
1035 putback_active_hugepage(hpage);
1036 return -ENOSYS;
1037 }
1038
1039 new_hpage = get_new_page(hpage, private, &result);
1040 if (!new_hpage)
1041 return -ENOMEM;
1042
1043 rc = -EAGAIN;
1044
1045 if (!trylock_page(hpage)) {
1046 if (!force || mode != MIGRATE_SYNC)
1047 goto out;
1048 lock_page(hpage);
1049 }
1050
1051 if (PageAnon(hpage))
1052 anon_vma = page_get_anon_vma(hpage);
1053
1054 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1055
1056 if (!page_mapped(hpage))
1057 rc = move_to_new_page(new_hpage, hpage, 1, mode);
1058
1059 if (rc)
1060 remove_migration_ptes(hpage, hpage);
1061
1062 if (anon_vma)
1063 put_anon_vma(anon_vma);
1064
1065 if (!rc)
1066 hugetlb_cgroup_migrate(hpage, new_hpage);
1067
1068 unlock_page(hpage);
1069out:
1070 if (rc != -EAGAIN)
1071 putback_active_hugepage(hpage);
1072 put_page(new_hpage);
1073 if (result) {
1074 if (rc)
1075 *result = rc;
1076 else
1077 *result = page_to_nid(new_hpage);
1078 }
1079 return rc;
1080}
1081
1082/*
1083 * migrate_pages - migrate the pages specified in a list, to the free pages
1084 * supplied as the target for the page migration
1085 *
1086 * @from: The list of pages to be migrated.
1087 * @get_new_page: The function used to allocate free pages to be used
1088 * as the target of the page migration.
1089 * @private: Private data to be passed on to get_new_page()
1090 * @mode: The migration mode that specifies the constraints for
1091 * page migration, if any.
1092 * @reason: The reason for page migration.
1093 *
1094 * The function returns after 10 attempts or if no pages are movable any more
1095 * because the list has become empty or no retryable pages exist any more.
1096 * The caller should call putback_lru_pages() to return pages to the LRU
1097 * or free list only if ret != 0.
1098 *
1099 * Returns the number of pages that were not migrated, or an error code.
1100 */
1101int migrate_pages(struct list_head *from, new_page_t get_new_page,
1102 unsigned long private, enum migrate_mode mode, int reason)
1103{
1104 int retry = 1;
1105 int nr_failed = 0;
1106 int nr_succeeded = 0;
1107 int pass = 0;
1108 struct page *page;
1109 struct page *page2;
1110 int swapwrite = current->flags & PF_SWAPWRITE;
1111 int rc;
1112
1113 if (!swapwrite)
1114 current->flags |= PF_SWAPWRITE;
1115
1116 for(pass = 0; pass < 10 && retry; pass++) {
1117 retry = 0;
1118
1119 list_for_each_entry_safe(page, page2, from, lru) {
1120 cond_resched();
1121
1122 if (PageHuge(page))
1123 rc = unmap_and_move_huge_page(get_new_page,
1124 private, page, pass > 2, mode);
1125 else
1126 rc = unmap_and_move(get_new_page, private,
1127 page, pass > 2, mode);
1128
1129 switch(rc) {
1130 case -ENOMEM:
1131 goto out;
1132 case -EAGAIN:
1133 retry++;
1134 break;
1135 case MIGRATEPAGE_SUCCESS:
1136 nr_succeeded++;
1137 break;
1138 default:
1139 /*
1140 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1141 * unlike -EAGAIN case, the failed page is
1142 * removed from migration page list and not
1143 * retried in the next outer loop.
1144 */
1145 nr_failed++;
1146 break;
1147 }
1148 }
1149 }
1150 rc = nr_failed + retry;
1151out:
1152 if (nr_succeeded)
1153 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1154 if (nr_failed)
1155 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1156 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1157
1158 if (!swapwrite)
1159 current->flags &= ~PF_SWAPWRITE;
1160
1161 return rc;
1162}
1163
1164#ifdef CONFIG_NUMA
1165/*
1166 * Move a list of individual pages
1167 */
1168struct page_to_node {
1169 unsigned long addr;
1170 struct page *page;
1171 int node;
1172 int status;
1173};
1174
1175static struct page *new_page_node(struct page *p, unsigned long private,
1176 int **result)
1177{
1178 struct page_to_node *pm = (struct page_to_node *)private;
1179
1180 while (pm->node != MAX_NUMNODES && pm->page != p)
1181 pm++;
1182
1183 if (pm->node == MAX_NUMNODES)
1184 return NULL;
1185
1186 *result = &pm->status;
1187
1188 if (PageHuge(p))
1189 return alloc_huge_page_node(page_hstate(compound_head(p)),
1190 pm->node);
1191 else
1192 return alloc_pages_exact_node(pm->node,
1193 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1194}
1195
1196/*
1197 * Move a set of pages as indicated in the pm array. The addr
1198 * field must be set to the virtual address of the page to be moved
1199 * and the node number must contain a valid target node.
1200 * The pm array ends with node = MAX_NUMNODES.
1201 */
1202static int do_move_page_to_node_array(struct mm_struct *mm,
1203 struct page_to_node *pm,
1204 int migrate_all)
1205{
1206 int err;
1207 struct page_to_node *pp;
1208 LIST_HEAD(pagelist);
1209
1210 down_read(&mm->mmap_sem);
1211
1212 /*
1213 * Build a list of pages to migrate
1214 */
1215 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1216 struct vm_area_struct *vma;
1217 struct page *page;
1218
1219 err = -EFAULT;
1220 vma = find_vma(mm, pp->addr);
1221 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1222 goto set_status;
1223
1224 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1225
1226 err = PTR_ERR(page);
1227 if (IS_ERR(page))
1228 goto set_status;
1229
1230 err = -ENOENT;
1231 if (!page)
1232 goto set_status;
1233
1234 /* Use PageReserved to check for zero page */
1235 if (PageReserved(page))
1236 goto put_and_set;
1237
1238 pp->page = page;
1239 err = page_to_nid(page);
1240
1241 if (err == pp->node)
1242 /*
1243 * Node already in the right place
1244 */
1245 goto put_and_set;
1246
1247 err = -EACCES;
1248 if (page_mapcount(page) > 1 &&
1249 !migrate_all)
1250 goto put_and_set;
1251
1252 if (PageHuge(page)) {
1253 isolate_huge_page(page, &pagelist);
1254 goto put_and_set;
1255 }
1256
1257 err = isolate_lru_page(page);
1258 if (!err) {
1259 list_add_tail(&page->lru, &pagelist);
1260 inc_zone_page_state(page, NR_ISOLATED_ANON +
1261 page_is_file_cache(page));
1262 }
1263put_and_set:
1264 /*
1265 * Either remove the duplicate refcount from
1266 * isolate_lru_page() or drop the page ref if it was
1267 * not isolated.
1268 */
1269 put_page(page);
1270set_status:
1271 pp->status = err;
1272 }
1273
1274 err = 0;
1275 if (!list_empty(&pagelist)) {
1276 err = migrate_pages(&pagelist, new_page_node,
1277 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1278 if (err)
1279 putback_movable_pages(&pagelist);
1280 }
1281
1282 up_read(&mm->mmap_sem);
1283 return err;
1284}
1285
1286/*
1287 * Migrate an array of page address onto an array of nodes and fill
1288 * the corresponding array of status.
1289 */
1290static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1291 unsigned long nr_pages,
1292 const void __user * __user *pages,
1293 const int __user *nodes,
1294 int __user *status, int flags)
1295{
1296 struct page_to_node *pm;
1297 unsigned long chunk_nr_pages;
1298 unsigned long chunk_start;
1299 int err;
1300
1301 err = -ENOMEM;
1302 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1303 if (!pm)
1304 goto out;
1305
1306 migrate_prep();
1307
1308 /*
1309 * Store a chunk of page_to_node array in a page,
1310 * but keep the last one as a marker
1311 */
1312 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1313
1314 for (chunk_start = 0;
1315 chunk_start < nr_pages;
1316 chunk_start += chunk_nr_pages) {
1317 int j;
1318
1319 if (chunk_start + chunk_nr_pages > nr_pages)
1320 chunk_nr_pages = nr_pages - chunk_start;
1321
1322 /* fill the chunk pm with addrs and nodes from user-space */
1323 for (j = 0; j < chunk_nr_pages; j++) {
1324 const void __user *p;
1325 int node;
1326
1327 err = -EFAULT;
1328 if (get_user(p, pages + j + chunk_start))
1329 goto out_pm;
1330 pm[j].addr = (unsigned long) p;
1331
1332 if (get_user(node, nodes + j + chunk_start))
1333 goto out_pm;
1334
1335 err = -ENODEV;
1336 if (node < 0 || node >= MAX_NUMNODES)
1337 goto out_pm;
1338
1339 if (!node_state(node, N_MEMORY))
1340 goto out_pm;
1341
1342 err = -EACCES;
1343 if (!node_isset(node, task_nodes))
1344 goto out_pm;
1345
1346 pm[j].node = node;
1347 }
1348
1349 /* End marker for this chunk */
1350 pm[chunk_nr_pages].node = MAX_NUMNODES;
1351
1352 /* Migrate this chunk */
1353 err = do_move_page_to_node_array(mm, pm,
1354 flags & MPOL_MF_MOVE_ALL);
1355 if (err < 0)
1356 goto out_pm;
1357
1358 /* Return status information */
1359 for (j = 0; j < chunk_nr_pages; j++)
1360 if (put_user(pm[j].status, status + j + chunk_start)) {
1361 err = -EFAULT;
1362 goto out_pm;
1363 }
1364 }
1365 err = 0;
1366
1367out_pm:
1368 free_page((unsigned long)pm);
1369out:
1370 return err;
1371}
1372
1373/*
1374 * Determine the nodes of an array of pages and store it in an array of status.
1375 */
1376static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1377 const void __user **pages, int *status)
1378{
1379 unsigned long i;
1380
1381 down_read(&mm->mmap_sem);
1382
1383 for (i = 0; i < nr_pages; i++) {
1384 unsigned long addr = (unsigned long)(*pages);
1385 struct vm_area_struct *vma;
1386 struct page *page;
1387 int err = -EFAULT;
1388
1389 vma = find_vma(mm, addr);
1390 if (!vma || addr < vma->vm_start)
1391 goto set_status;
1392
1393 page = follow_page(vma, addr, 0);
1394
1395 err = PTR_ERR(page);
1396 if (IS_ERR(page))
1397 goto set_status;
1398
1399 err = -ENOENT;
1400 /* Use PageReserved to check for zero page */
1401 if (!page || PageReserved(page))
1402 goto set_status;
1403
1404 err = page_to_nid(page);
1405set_status:
1406 *status = err;
1407
1408 pages++;
1409 status++;
1410 }
1411
1412 up_read(&mm->mmap_sem);
1413}
1414
1415/*
1416 * Determine the nodes of a user array of pages and store it in
1417 * a user array of status.
1418 */
1419static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1420 const void __user * __user *pages,
1421 int __user *status)
1422{
1423#define DO_PAGES_STAT_CHUNK_NR 16
1424 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1425 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1426
1427 while (nr_pages) {
1428 unsigned long chunk_nr;
1429
1430 chunk_nr = nr_pages;
1431 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1432 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1433
1434 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1435 break;
1436
1437 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1438
1439 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1440 break;
1441
1442 pages += chunk_nr;
1443 status += chunk_nr;
1444 nr_pages -= chunk_nr;
1445 }
1446 return nr_pages ? -EFAULT : 0;
1447}
1448
1449/*
1450 * Move a list of pages in the address space of the currently executing
1451 * process.
1452 */
1453SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1454 const void __user * __user *, pages,
1455 const int __user *, nodes,
1456 int __user *, status, int, flags)
1457{
1458 const struct cred *cred = current_cred(), *tcred;
1459 struct task_struct *task;
1460 struct mm_struct *mm;
1461 int err;
1462 nodemask_t task_nodes;
1463
1464 /* Check flags */
1465 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1466 return -EINVAL;
1467
1468 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1469 return -EPERM;
1470
1471 /* Find the mm_struct */
1472 rcu_read_lock();
1473 task = pid ? find_task_by_vpid(pid) : current;
1474 if (!task) {
1475 rcu_read_unlock();
1476 return -ESRCH;
1477 }
1478 get_task_struct(task);
1479
1480 /*
1481 * Check if this process has the right to modify the specified
1482 * process. The right exists if the process has administrative
1483 * capabilities, superuser privileges or the same
1484 * userid as the target process.
1485 */
1486 tcred = __task_cred(task);
1487 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1488 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1489 !capable(CAP_SYS_NICE)) {
1490 rcu_read_unlock();
1491 err = -EPERM;
1492 goto out;
1493 }
1494 rcu_read_unlock();
1495
1496 err = security_task_movememory(task);
1497 if (err)
1498 goto out;
1499
1500 task_nodes = cpuset_mems_allowed(task);
1501 mm = get_task_mm(task);
1502 put_task_struct(task);
1503
1504 if (!mm)
1505 return -EINVAL;
1506
1507 if (nodes)
1508 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1509 nodes, status, flags);
1510 else
1511 err = do_pages_stat(mm, nr_pages, pages, status);
1512
1513 mmput(mm);
1514 return err;
1515
1516out:
1517 put_task_struct(task);
1518 return err;
1519}
1520
1521/*
1522 * Call migration functions in the vma_ops that may prepare
1523 * memory in a vm for migration. migration functions may perform
1524 * the migration for vmas that do not have an underlying page struct.
1525 */
1526int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1527 const nodemask_t *from, unsigned long flags)
1528{
1529 struct vm_area_struct *vma;
1530 int err = 0;
1531
1532 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1533 if (vma->vm_ops && vma->vm_ops->migrate) {
1534 err = vma->vm_ops->migrate(vma, to, from, flags);
1535 if (err)
1536 break;
1537 }
1538 }
1539 return err;
1540}
1541
1542#ifdef CONFIG_NUMA_BALANCING
1543/*
1544 * Returns true if this is a safe migration target node for misplaced NUMA
1545 * pages. Currently it only checks the watermarks which crude
1546 */
1547static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1548 unsigned long nr_migrate_pages)
1549{
1550 int z;
1551 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1552 struct zone *zone = pgdat->node_zones + z;
1553
1554 if (!populated_zone(zone))
1555 continue;
1556
1557 if (!zone_reclaimable(zone))
1558 continue;
1559
1560 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1561 if (!zone_watermark_ok(zone, 0,
1562 high_wmark_pages(zone) +
1563 nr_migrate_pages,
1564 0, 0))
1565 continue;
1566 return true;
1567 }
1568 return false;
1569}
1570
1571static struct page *alloc_misplaced_dst_page(struct page *page,
1572 unsigned long data,
1573 int **result)
1574{
1575 int nid = (int) data;
1576 struct page *newpage;
1577
1578 newpage = alloc_pages_exact_node(nid,
1579 (GFP_HIGHUSER_MOVABLE |
1580 __GFP_THISNODE | __GFP_NOMEMALLOC |
1581 __GFP_NORETRY | __GFP_NOWARN) &
1582 ~GFP_IOFS, 0);
1583
1584 return newpage;
1585}
1586
1587/*
1588 * page migration rate limiting control.
1589 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1590 * window of time. Default here says do not migrate more than 1280M per second.
1591 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1592 * as it is faults that reset the window, pte updates will happen unconditionally
1593 * if there has not been a fault since @pteupdate_interval_millisecs after the
1594 * throttle window closed.
1595 */
1596static unsigned int migrate_interval_millisecs __read_mostly = 100;
1597static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1598static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1599
1600/* Returns true if NUMA migration is currently rate limited */
1601bool migrate_ratelimited(int node)
1602{
1603 pg_data_t *pgdat = NODE_DATA(node);
1604
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1606 msecs_to_jiffies(pteupdate_interval_millisecs)))
1607 return false;
1608
1609 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1610 return false;
1611
1612 return true;
1613}
1614
1615/* Returns true if the node is migrate rate-limited after the update */
1616static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1617 unsigned long nr_pages)
1618{
1619 /*
1620 * Rate-limit the amount of data that is being migrated to a node.
1621 * Optimal placement is no good if the memory bus is saturated and
1622 * all the time is being spent migrating!
1623 */
1624 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1625 spin_lock(&pgdat->numabalancing_migrate_lock);
1626 pgdat->numabalancing_migrate_nr_pages = 0;
1627 pgdat->numabalancing_migrate_next_window = jiffies +
1628 msecs_to_jiffies(migrate_interval_millisecs);
1629 spin_unlock(&pgdat->numabalancing_migrate_lock);
1630 }
1631 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1632 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1633 nr_pages);
1634 return true;
1635 }
1636
1637 /*
1638 * This is an unlocked non-atomic update so errors are possible.
1639 * The consequences are failing to migrate when we potentiall should
1640 * have which is not severe enough to warrant locking. If it is ever
1641 * a problem, it can be converted to a per-cpu counter.
1642 */
1643 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1644 return false;
1645}
1646
1647static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1648{
1649 int page_lru;
1650
1651 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1652
1653 /* Avoid migrating to a node that is nearly full */
1654 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1655 return 0;
1656
1657 if (isolate_lru_page(page))
1658 return 0;
1659
1660 /*
1661 * migrate_misplaced_transhuge_page() skips page migration's usual
1662 * check on page_count(), so we must do it here, now that the page
1663 * has been isolated: a GUP pin, or any other pin, prevents migration.
1664 * The expected page count is 3: 1 for page's mapcount and 1 for the
1665 * caller's pin and 1 for the reference taken by isolate_lru_page().
1666 */
1667 if (PageTransHuge(page) && page_count(page) != 3) {
1668 putback_lru_page(page);
1669 return 0;
1670 }
1671
1672 page_lru = page_is_file_cache(page);
1673 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1674 hpage_nr_pages(page));
1675
1676 /*
1677 * Isolating the page has taken another reference, so the
1678 * caller's reference can be safely dropped without the page
1679 * disappearing underneath us during migration.
1680 */
1681 put_page(page);
1682 return 1;
1683}
1684
1685bool pmd_trans_migrating(pmd_t pmd)
1686{
1687 struct page *page = pmd_page(pmd);
1688 return PageLocked(page);
1689}
1690
1691void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1692{
1693 struct page *page = pmd_page(*pmd);
1694 wait_on_page_locked(page);
1695}
1696
1697/*
1698 * Attempt to migrate a misplaced page to the specified destination
1699 * node. Caller is expected to have an elevated reference count on
1700 * the page that will be dropped by this function before returning.
1701 */
1702int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1703 int node)
1704{
1705 pg_data_t *pgdat = NODE_DATA(node);
1706 int isolated;
1707 int nr_remaining;
1708 LIST_HEAD(migratepages);
1709
1710 /*
1711 * Don't migrate file pages that are mapped in multiple processes
1712 * with execute permissions as they are probably shared libraries.
1713 */
1714 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1715 (vma->vm_flags & VM_EXEC))
1716 goto out;
1717
1718 /*
1719 * Rate-limit the amount of data that is being migrated to a node.
1720 * Optimal placement is no good if the memory bus is saturated and
1721 * all the time is being spent migrating!
1722 */
1723 if (numamigrate_update_ratelimit(pgdat, 1))
1724 goto out;
1725
1726 isolated = numamigrate_isolate_page(pgdat, page);
1727 if (!isolated)
1728 goto out;
1729
1730 list_add(&page->lru, &migratepages);
1731 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1732 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1733 if (nr_remaining) {
1734 if (!list_empty(&migratepages)) {
1735 list_del(&page->lru);
1736 dec_zone_page_state(page, NR_ISOLATED_ANON +
1737 page_is_file_cache(page));
1738 putback_lru_page(page);
1739 }
1740 isolated = 0;
1741 } else
1742 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1743 BUG_ON(!list_empty(&migratepages));
1744 return isolated;
1745
1746out:
1747 put_page(page);
1748 return 0;
1749}
1750#endif /* CONFIG_NUMA_BALANCING */
1751
1752#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1753/*
1754 * Migrates a THP to a given target node. page must be locked and is unlocked
1755 * before returning.
1756 */
1757int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1758 struct vm_area_struct *vma,
1759 pmd_t *pmd, pmd_t entry,
1760 unsigned long address,
1761 struct page *page, int node)
1762{
1763 spinlock_t *ptl;
1764 pg_data_t *pgdat = NODE_DATA(node);
1765 int isolated = 0;
1766 struct page *new_page = NULL;
1767 struct mem_cgroup *memcg = NULL;
1768 int page_lru = page_is_file_cache(page);
1769 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1770 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1771 pmd_t orig_entry;
1772
1773 /*
1774 * Rate-limit the amount of data that is being migrated to a node.
1775 * Optimal placement is no good if the memory bus is saturated and
1776 * all the time is being spent migrating!
1777 */
1778 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1779 goto out_dropref;
1780
1781 new_page = alloc_pages_node(node,
1782 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1783 HPAGE_PMD_ORDER);
1784 if (!new_page)
1785 goto out_fail;
1786
1787 isolated = numamigrate_isolate_page(pgdat, page);
1788 if (!isolated) {
1789 put_page(new_page);
1790 goto out_fail;
1791 }
1792
1793 if (mm_tlb_flush_pending(mm))
1794 flush_tlb_range(vma, mmun_start, mmun_end);
1795
1796 /* Prepare a page as a migration target */
1797 __set_page_locked(new_page);
1798 SetPageSwapBacked(new_page);
1799
1800 /* anon mapping, we can simply copy page->mapping to the new page: */
1801 new_page->mapping = page->mapping;
1802 new_page->index = page->index;
1803 migrate_page_copy(new_page, page);
1804 WARN_ON(PageLRU(new_page));
1805
1806 /* Recheck the target PMD */
1807 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1808 ptl = pmd_lock(mm, pmd);
1809 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1810fail_putback:
1811 spin_unlock(ptl);
1812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1813
1814 /* Reverse changes made by migrate_page_copy() */
1815 if (TestClearPageActive(new_page))
1816 SetPageActive(page);
1817 if (TestClearPageUnevictable(new_page))
1818 SetPageUnevictable(page);
1819 mlock_migrate_page(page, new_page);
1820
1821 unlock_page(new_page);
1822 put_page(new_page); /* Free it */
1823
1824 /* Retake the callers reference and putback on LRU */
1825 get_page(page);
1826 putback_lru_page(page);
1827 mod_zone_page_state(page_zone(page),
1828 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1829
1830 goto out_unlock;
1831 }
1832
1833 /*
1834 * Traditional migration needs to prepare the memcg charge
1835 * transaction early to prevent the old page from being
1836 * uncharged when installing migration entries. Here we can
1837 * save the potential rollback and start the charge transfer
1838 * only when migration is already known to end successfully.
1839 */
1840 mem_cgroup_prepare_migration(page, new_page, &memcg);
1841
1842 orig_entry = *pmd;
1843 entry = mk_pmd(new_page, vma->vm_page_prot);
1844 entry = pmd_mkhuge(entry);
1845 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1846
1847 /*
1848 * Clear the old entry under pagetable lock and establish the new PTE.
1849 * Any parallel GUP will either observe the old page blocking on the
1850 * page lock, block on the page table lock or observe the new page.
1851 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1852 * guarantee the copy is visible before the pagetable update.
1853 */
1854 flush_cache_range(vma, mmun_start, mmun_end);
1855 page_add_new_anon_rmap(new_page, vma, mmun_start);
1856 pmdp_clear_flush(vma, mmun_start, pmd);
1857 set_pmd_at(mm, mmun_start, pmd, entry);
1858 flush_tlb_range(vma, mmun_start, mmun_end);
1859 update_mmu_cache_pmd(vma, address, &entry);
1860
1861 if (page_count(page) != 2) {
1862 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1863 flush_tlb_range(vma, mmun_start, mmun_end);
1864 update_mmu_cache_pmd(vma, address, &entry);
1865 page_remove_rmap(new_page);
1866 goto fail_putback;
1867 }
1868
1869 page_remove_rmap(page);
1870
1871 /*
1872 * Finish the charge transaction under the page table lock to
1873 * prevent split_huge_page() from dividing up the charge
1874 * before it's fully transferred to the new page.
1875 */
1876 mem_cgroup_end_migration(memcg, page, new_page, true);
1877 spin_unlock(ptl);
1878 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1879
1880 unlock_page(new_page);
1881 unlock_page(page);
1882 put_page(page); /* Drop the rmap reference */
1883 put_page(page); /* Drop the LRU isolation reference */
1884
1885 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1886 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1887
1888 mod_zone_page_state(page_zone(page),
1889 NR_ISOLATED_ANON + page_lru,
1890 -HPAGE_PMD_NR);
1891 return isolated;
1892
1893out_fail:
1894 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1895out_dropref:
1896 ptl = pmd_lock(mm, pmd);
1897 if (pmd_same(*pmd, entry)) {
1898 entry = pmd_mknonnuma(entry);
1899 set_pmd_at(mm, mmun_start, pmd, entry);
1900 update_mmu_cache_pmd(vma, address, &entry);
1901 }
1902 spin_unlock(ptl);
1903
1904out_unlock:
1905 unlock_page(page);
1906 put_page(page);
1907 return 0;
1908}
1909#endif /* CONFIG_NUMA_BALANCING */
1910
1911#endif /* CONFIG_NUMA */