thp: memcg huge memory
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / migrate.c
blob1a531b760b3bedb3b3395f95e4d1f2d1fbf8b4b5
1 /*
2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter
15 #include <linux/migrate.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/gfp.h>
38 #include <asm/tlbflush.h>
40 #include "internal.h"
42 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
45 * migrate_prep() needs to be called before we start compiling a list of pages
46 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
47 * undesirable, use migrate_prep_local()
49 int migrate_prep(void)
52 * Clear the LRU lists so pages can be isolated.
53 * Note that pages may be moved off the LRU after we have
54 * drained them. Those pages will fail to migrate like other
55 * pages that may be busy.
57 lru_add_drain_all();
59 return 0;
62 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
63 int migrate_prep_local(void)
65 lru_add_drain();
67 return 0;
71 * Add isolated pages on the list back to the LRU under page lock
72 * to avoid leaking evictable pages back onto unevictable list.
74 void putback_lru_pages(struct list_head *l)
76 struct page *page;
77 struct page *page2;
79 list_for_each_entry_safe(page, page2, l, lru) {
80 list_del(&page->lru);
81 dec_zone_page_state(page, NR_ISOLATED_ANON +
82 page_is_file_cache(page));
83 putback_lru_page(page);
88 * Restore a potential migration pte to a working pte entry
90 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
91 unsigned long addr, void *old)
93 struct mm_struct *mm = vma->vm_mm;
94 swp_entry_t entry;
95 pgd_t *pgd;
96 pud_t *pud;
97 pmd_t *pmd;
98 pte_t *ptep, pte;
99 spinlock_t *ptl;
101 if (unlikely(PageHuge(new))) {
102 ptep = huge_pte_offset(mm, addr);
103 if (!ptep)
104 goto out;
105 ptl = &mm->page_table_lock;
106 } else {
107 pgd = pgd_offset(mm, addr);
108 if (!pgd_present(*pgd))
109 goto out;
111 pud = pud_offset(pgd, addr);
112 if (!pud_present(*pud))
113 goto out;
115 pmd = pmd_offset(pud, addr);
116 if (pmd_trans_huge(*pmd))
117 goto out;
118 if (!pmd_present(*pmd))
119 goto out;
121 ptep = pte_offset_map(pmd, addr);
123 if (!is_swap_pte(*ptep)) {
124 pte_unmap(ptep);
125 goto out;
128 ptl = pte_lockptr(mm, pmd);
131 spin_lock(ptl);
132 pte = *ptep;
133 if (!is_swap_pte(pte))
134 goto unlock;
136 entry = pte_to_swp_entry(pte);
138 if (!is_migration_entry(entry) ||
139 migration_entry_to_page(entry) != old)
140 goto unlock;
142 get_page(new);
143 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
144 if (is_write_migration_entry(entry))
145 pte = pte_mkwrite(pte);
146 #ifdef CONFIG_HUGETLB_PAGE
147 if (PageHuge(new))
148 pte = pte_mkhuge(pte);
149 #endif
150 flush_cache_page(vma, addr, pte_pfn(pte));
151 set_pte_at(mm, addr, ptep, pte);
153 if (PageHuge(new)) {
154 if (PageAnon(new))
155 hugepage_add_anon_rmap(new, vma, addr);
156 else
157 page_dup_rmap(new);
158 } else if (PageAnon(new))
159 page_add_anon_rmap(new, vma, addr);
160 else
161 page_add_file_rmap(new);
163 /* No need to invalidate - it was non-present before */
164 update_mmu_cache(vma, addr, ptep);
165 unlock:
166 pte_unmap_unlock(ptep, ptl);
167 out:
168 return SWAP_AGAIN;
172 * Get rid of all migration entries and replace them by
173 * references to the indicated page.
175 static void remove_migration_ptes(struct page *old, struct page *new)
177 rmap_walk(new, remove_migration_pte, old);
181 * Something used the pte of a page under migration. We need to
182 * get to the page and wait until migration is finished.
183 * When we return from this function the fault will be retried.
185 * This function is called from do_swap_page().
187 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
188 unsigned long address)
190 pte_t *ptep, pte;
191 spinlock_t *ptl;
192 swp_entry_t entry;
193 struct page *page;
195 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
196 pte = *ptep;
197 if (!is_swap_pte(pte))
198 goto out;
200 entry = pte_to_swp_entry(pte);
201 if (!is_migration_entry(entry))
202 goto out;
204 page = migration_entry_to_page(entry);
207 * Once radix-tree replacement of page migration started, page_count
208 * *must* be zero. And, we don't want to call wait_on_page_locked()
209 * against a page without get_page().
210 * So, we use get_page_unless_zero(), here. Even failed, page fault
211 * will occur again.
213 if (!get_page_unless_zero(page))
214 goto out;
215 pte_unmap_unlock(ptep, ptl);
216 wait_on_page_locked(page);
217 put_page(page);
218 return;
219 out:
220 pte_unmap_unlock(ptep, ptl);
224 * Replace the page in the mapping.
226 * The number of remaining references must be:
227 * 1 for anonymous pages without a mapping
228 * 2 for pages with a mapping
229 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
231 static int migrate_page_move_mapping(struct address_space *mapping,
232 struct page *newpage, struct page *page)
234 int expected_count;
235 void **pslot;
237 if (!mapping) {
238 /* Anonymous page without mapping */
239 if (page_count(page) != 1)
240 return -EAGAIN;
241 return 0;
244 spin_lock_irq(&mapping->tree_lock);
246 pslot = radix_tree_lookup_slot(&mapping->page_tree,
247 page_index(page));
249 expected_count = 2 + page_has_private(page);
250 if (page_count(page) != expected_count ||
251 (struct page *)radix_tree_deref_slot(pslot) != page) {
252 spin_unlock_irq(&mapping->tree_lock);
253 return -EAGAIN;
256 if (!page_freeze_refs(page, expected_count)) {
257 spin_unlock_irq(&mapping->tree_lock);
258 return -EAGAIN;
262 * Now we know that no one else is looking at the page.
264 get_page(newpage); /* add cache reference */
265 if (PageSwapCache(page)) {
266 SetPageSwapCache(newpage);
267 set_page_private(newpage, page_private(page));
270 radix_tree_replace_slot(pslot, newpage);
272 page_unfreeze_refs(page, expected_count);
274 * Drop cache reference from old page.
275 * We know this isn't the last reference.
277 __put_page(page);
280 * If moved to a different zone then also account
281 * the page for that zone. Other VM counters will be
282 * taken care of when we establish references to the
283 * new page and drop references to the old page.
285 * Note that anonymous pages are accounted for
286 * via NR_FILE_PAGES and NR_ANON_PAGES if they
287 * are mapped to swap space.
289 __dec_zone_page_state(page, NR_FILE_PAGES);
290 __inc_zone_page_state(newpage, NR_FILE_PAGES);
291 if (PageSwapBacked(page)) {
292 __dec_zone_page_state(page, NR_SHMEM);
293 __inc_zone_page_state(newpage, NR_SHMEM);
295 spin_unlock_irq(&mapping->tree_lock);
297 return 0;
301 * The expected number of remaining references is the same as that
302 * of migrate_page_move_mapping().
304 int migrate_huge_page_move_mapping(struct address_space *mapping,
305 struct page *newpage, struct page *page)
307 int expected_count;
308 void **pslot;
310 if (!mapping) {
311 if (page_count(page) != 1)
312 return -EAGAIN;
313 return 0;
316 spin_lock_irq(&mapping->tree_lock);
318 pslot = radix_tree_lookup_slot(&mapping->page_tree,
319 page_index(page));
321 expected_count = 2 + page_has_private(page);
322 if (page_count(page) != expected_count ||
323 (struct page *)radix_tree_deref_slot(pslot) != page) {
324 spin_unlock_irq(&mapping->tree_lock);
325 return -EAGAIN;
328 if (!page_freeze_refs(page, expected_count)) {
329 spin_unlock_irq(&mapping->tree_lock);
330 return -EAGAIN;
333 get_page(newpage);
335 radix_tree_replace_slot(pslot, newpage);
337 page_unfreeze_refs(page, expected_count);
339 __put_page(page);
341 spin_unlock_irq(&mapping->tree_lock);
342 return 0;
346 * Copy the page to its new location
348 void migrate_page_copy(struct page *newpage, struct page *page)
350 if (PageHuge(page))
351 copy_huge_page(newpage, page);
352 else
353 copy_highpage(newpage, page);
355 if (PageError(page))
356 SetPageError(newpage);
357 if (PageReferenced(page))
358 SetPageReferenced(newpage);
359 if (PageUptodate(page))
360 SetPageUptodate(newpage);
361 if (TestClearPageActive(page)) {
362 VM_BUG_ON(PageUnevictable(page));
363 SetPageActive(newpage);
364 } else if (TestClearPageUnevictable(page))
365 SetPageUnevictable(newpage);
366 if (PageChecked(page))
367 SetPageChecked(newpage);
368 if (PageMappedToDisk(page))
369 SetPageMappedToDisk(newpage);
371 if (PageDirty(page)) {
372 clear_page_dirty_for_io(page);
374 * Want to mark the page and the radix tree as dirty, and
375 * redo the accounting that clear_page_dirty_for_io undid,
376 * but we can't use set_page_dirty because that function
377 * is actually a signal that all of the page has become dirty.
378 * Wheras only part of our page may be dirty.
380 __set_page_dirty_nobuffers(newpage);
383 mlock_migrate_page(newpage, page);
384 ksm_migrate_page(newpage, page);
386 ClearPageSwapCache(page);
387 ClearPagePrivate(page);
388 set_page_private(page, 0);
389 page->mapping = NULL;
392 * If any waiters have accumulated on the new page then
393 * wake them up.
395 if (PageWriteback(newpage))
396 end_page_writeback(newpage);
399 /************************************************************
400 * Migration functions
401 ***********************************************************/
403 /* Always fail migration. Used for mappings that are not movable */
404 int fail_migrate_page(struct address_space *mapping,
405 struct page *newpage, struct page *page)
407 return -EIO;
409 EXPORT_SYMBOL(fail_migrate_page);
412 * Common logic to directly migrate a single page suitable for
413 * pages that do not use PagePrivate/PagePrivate2.
415 * Pages are locked upon entry and exit.
417 int migrate_page(struct address_space *mapping,
418 struct page *newpage, struct page *page)
420 int rc;
422 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
424 rc = migrate_page_move_mapping(mapping, newpage, page);
426 if (rc)
427 return rc;
429 migrate_page_copy(newpage, page);
430 return 0;
432 EXPORT_SYMBOL(migrate_page);
434 #ifdef CONFIG_BLOCK
436 * Migration function for pages with buffers. This function can only be used
437 * if the underlying filesystem guarantees that no other references to "page"
438 * exist.
440 int buffer_migrate_page(struct address_space *mapping,
441 struct page *newpage, struct page *page)
443 struct buffer_head *bh, *head;
444 int rc;
446 if (!page_has_buffers(page))
447 return migrate_page(mapping, newpage, page);
449 head = page_buffers(page);
451 rc = migrate_page_move_mapping(mapping, newpage, page);
453 if (rc)
454 return rc;
456 bh = head;
457 do {
458 get_bh(bh);
459 lock_buffer(bh);
460 bh = bh->b_this_page;
462 } while (bh != head);
464 ClearPagePrivate(page);
465 set_page_private(newpage, page_private(page));
466 set_page_private(page, 0);
467 put_page(page);
468 get_page(newpage);
470 bh = head;
471 do {
472 set_bh_page(bh, newpage, bh_offset(bh));
473 bh = bh->b_this_page;
475 } while (bh != head);
477 SetPagePrivate(newpage);
479 migrate_page_copy(newpage, page);
481 bh = head;
482 do {
483 unlock_buffer(bh);
484 put_bh(bh);
485 bh = bh->b_this_page;
487 } while (bh != head);
489 return 0;
491 EXPORT_SYMBOL(buffer_migrate_page);
492 #endif
495 * Writeback a page to clean the dirty state
497 static int writeout(struct address_space *mapping, struct page *page)
499 struct writeback_control wbc = {
500 .sync_mode = WB_SYNC_NONE,
501 .nr_to_write = 1,
502 .range_start = 0,
503 .range_end = LLONG_MAX,
504 .for_reclaim = 1
506 int rc;
508 if (!mapping->a_ops->writepage)
509 /* No write method for the address space */
510 return -EINVAL;
512 if (!clear_page_dirty_for_io(page))
513 /* Someone else already triggered a write */
514 return -EAGAIN;
517 * A dirty page may imply that the underlying filesystem has
518 * the page on some queue. So the page must be clean for
519 * migration. Writeout may mean we loose the lock and the
520 * page state is no longer what we checked for earlier.
521 * At this point we know that the migration attempt cannot
522 * be successful.
524 remove_migration_ptes(page, page);
526 rc = mapping->a_ops->writepage(page, &wbc);
528 if (rc != AOP_WRITEPAGE_ACTIVATE)
529 /* unlocked. Relock */
530 lock_page(page);
532 return (rc < 0) ? -EIO : -EAGAIN;
536 * Default handling if a filesystem does not provide a migration function.
538 static int fallback_migrate_page(struct address_space *mapping,
539 struct page *newpage, struct page *page)
541 if (PageDirty(page))
542 return writeout(mapping, page);
545 * Buffers may be managed in a filesystem specific way.
546 * We must have no buffers or drop them.
548 if (page_has_private(page) &&
549 !try_to_release_page(page, GFP_KERNEL))
550 return -EAGAIN;
552 return migrate_page(mapping, newpage, page);
556 * Move a page to a newly allocated page
557 * The page is locked and all ptes have been successfully removed.
559 * The new page will have replaced the old page if this function
560 * is successful.
562 * Return value:
563 * < 0 - error code
564 * == 0 - success
566 static int move_to_new_page(struct page *newpage, struct page *page,
567 int remap_swapcache)
569 struct address_space *mapping;
570 int rc;
573 * Block others from accessing the page when we get around to
574 * establishing additional references. We are the only one
575 * holding a reference to the new page at this point.
577 if (!trylock_page(newpage))
578 BUG();
580 /* Prepare mapping for the new page.*/
581 newpage->index = page->index;
582 newpage->mapping = page->mapping;
583 if (PageSwapBacked(page))
584 SetPageSwapBacked(newpage);
586 mapping = page_mapping(page);
587 if (!mapping)
588 rc = migrate_page(mapping, newpage, page);
589 else if (mapping->a_ops->migratepage)
591 * Most pages have a mapping and most filesystems
592 * should provide a migration function. Anonymous
593 * pages are part of swap space which also has its
594 * own migration function. This is the most common
595 * path for page migration.
597 rc = mapping->a_ops->migratepage(mapping,
598 newpage, page);
599 else
600 rc = fallback_migrate_page(mapping, newpage, page);
602 if (rc) {
603 newpage->mapping = NULL;
604 } else {
605 if (remap_swapcache)
606 remove_migration_ptes(page, newpage);
609 unlock_page(newpage);
611 return rc;
615 * Obtain the lock on page, remove all ptes and migrate the page
616 * to the newly allocated page in newpage.
618 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
619 struct page *page, int force, bool offlining, bool sync)
621 int rc = 0;
622 int *result = NULL;
623 struct page *newpage = get_new_page(page, private, &result);
624 int remap_swapcache = 1;
625 int rcu_locked = 0;
626 int charge = 0;
627 struct mem_cgroup *mem = NULL;
628 struct anon_vma *anon_vma = NULL;
630 if (!newpage)
631 return -ENOMEM;
633 if (page_count(page) == 1) {
634 /* page was freed from under us. So we are done. */
635 goto move_newpage;
637 if (unlikely(PageTransHuge(page)))
638 if (unlikely(split_huge_page(page)))
639 goto move_newpage;
641 /* prepare cgroup just returns 0 or -ENOMEM */
642 rc = -EAGAIN;
644 if (!trylock_page(page)) {
645 if (!force)
646 goto move_newpage;
649 * It's not safe for direct compaction to call lock_page.
650 * For example, during page readahead pages are added locked
651 * to the LRU. Later, when the IO completes the pages are
652 * marked uptodate and unlocked. However, the queueing
653 * could be merging multiple pages for one bio (e.g.
654 * mpage_readpages). If an allocation happens for the
655 * second or third page, the process can end up locking
656 * the same page twice and deadlocking. Rather than
657 * trying to be clever about what pages can be locked,
658 * avoid the use of lock_page for direct compaction
659 * altogether.
661 if (current->flags & PF_MEMALLOC)
662 goto move_newpage;
664 lock_page(page);
668 * Only memory hotplug's offline_pages() caller has locked out KSM,
669 * and can safely migrate a KSM page. The other cases have skipped
670 * PageKsm along with PageReserved - but it is only now when we have
671 * the page lock that we can be certain it will not go KSM beneath us
672 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
673 * its pagecount raised, but only here do we take the page lock which
674 * serializes that).
676 if (PageKsm(page) && !offlining) {
677 rc = -EBUSY;
678 goto unlock;
681 /* charge against new page */
682 charge = mem_cgroup_prepare_migration(page, newpage, &mem);
683 if (charge == -ENOMEM) {
684 rc = -ENOMEM;
685 goto unlock;
687 BUG_ON(charge);
689 if (PageWriteback(page)) {
690 if (!force || !sync)
691 goto uncharge;
692 wait_on_page_writeback(page);
695 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
696 * we cannot notice that anon_vma is freed while we migrates a page.
697 * This rcu_read_lock() delays freeing anon_vma pointer until the end
698 * of migration. File cache pages are no problem because of page_lock()
699 * File Caches may use write_page() or lock_page() in migration, then,
700 * just care Anon page here.
702 if (PageAnon(page)) {
703 rcu_read_lock();
704 rcu_locked = 1;
706 /* Determine how to safely use anon_vma */
707 if (!page_mapped(page)) {
708 if (!PageSwapCache(page))
709 goto rcu_unlock;
712 * We cannot be sure that the anon_vma of an unmapped
713 * swapcache page is safe to use because we don't
714 * know in advance if the VMA that this page belonged
715 * to still exists. If the VMA and others sharing the
716 * data have been freed, then the anon_vma could
717 * already be invalid.
719 * To avoid this possibility, swapcache pages get
720 * migrated but are not remapped when migration
721 * completes
723 remap_swapcache = 0;
724 } else {
726 * Take a reference count on the anon_vma if the
727 * page is mapped so that it is guaranteed to
728 * exist when the page is remapped later
730 anon_vma = page_anon_vma(page);
731 get_anon_vma(anon_vma);
736 * Corner case handling:
737 * 1. When a new swap-cache page is read into, it is added to the LRU
738 * and treated as swapcache but it has no rmap yet.
739 * Calling try_to_unmap() against a page->mapping==NULL page will
740 * trigger a BUG. So handle it here.
741 * 2. An orphaned page (see truncate_complete_page) might have
742 * fs-private metadata. The page can be picked up due to memory
743 * offlining. Everywhere else except page reclaim, the page is
744 * invisible to the vm, so the page can not be migrated. So try to
745 * free the metadata, so the page can be freed.
747 if (!page->mapping) {
748 if (!PageAnon(page) && page_has_private(page)) {
750 * Go direct to try_to_free_buffers() here because
751 * a) that's what try_to_release_page() would do anyway
752 * b) we may be under rcu_read_lock() here, so we can't
753 * use GFP_KERNEL which is what try_to_release_page()
754 * needs to be effective.
756 try_to_free_buffers(page);
757 goto rcu_unlock;
759 goto skip_unmap;
762 /* Establish migration ptes or remove ptes */
763 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
765 skip_unmap:
766 if (!page_mapped(page))
767 rc = move_to_new_page(newpage, page, remap_swapcache);
769 if (rc && remap_swapcache)
770 remove_migration_ptes(page, page);
771 rcu_unlock:
773 /* Drop an anon_vma reference if we took one */
774 if (anon_vma)
775 drop_anon_vma(anon_vma);
777 if (rcu_locked)
778 rcu_read_unlock();
779 uncharge:
780 if (!charge)
781 mem_cgroup_end_migration(mem, page, newpage);
782 unlock:
783 unlock_page(page);
785 if (rc != -EAGAIN) {
787 * A page that has been migrated has all references
788 * removed and will be freed. A page that has not been
789 * migrated will have kepts its references and be
790 * restored.
792 list_del(&page->lru);
793 dec_zone_page_state(page, NR_ISOLATED_ANON +
794 page_is_file_cache(page));
795 putback_lru_page(page);
798 move_newpage:
801 * Move the new page to the LRU. If migration was not successful
802 * then this will free the page.
804 putback_lru_page(newpage);
806 if (result) {
807 if (rc)
808 *result = rc;
809 else
810 *result = page_to_nid(newpage);
812 return rc;
816 * Counterpart of unmap_and_move_page() for hugepage migration.
818 * This function doesn't wait the completion of hugepage I/O
819 * because there is no race between I/O and migration for hugepage.
820 * Note that currently hugepage I/O occurs only in direct I/O
821 * where no lock is held and PG_writeback is irrelevant,
822 * and writeback status of all subpages are counted in the reference
823 * count of the head page (i.e. if all subpages of a 2MB hugepage are
824 * under direct I/O, the reference of the head page is 512 and a bit more.)
825 * This means that when we try to migrate hugepage whose subpages are
826 * doing direct I/O, some references remain after try_to_unmap() and
827 * hugepage migration fails without data corruption.
829 * There is also no race when direct I/O is issued on the page under migration,
830 * because then pte is replaced with migration swap entry and direct I/O code
831 * will wait in the page fault for migration to complete.
833 static int unmap_and_move_huge_page(new_page_t get_new_page,
834 unsigned long private, struct page *hpage,
835 int force, bool offlining, bool sync)
837 int rc = 0;
838 int *result = NULL;
839 struct page *new_hpage = get_new_page(hpage, private, &result);
840 int rcu_locked = 0;
841 struct anon_vma *anon_vma = NULL;
843 if (!new_hpage)
844 return -ENOMEM;
846 rc = -EAGAIN;
848 if (!trylock_page(hpage)) {
849 if (!force || !sync)
850 goto out;
851 lock_page(hpage);
854 if (PageAnon(hpage)) {
855 rcu_read_lock();
856 rcu_locked = 1;
858 if (page_mapped(hpage)) {
859 anon_vma = page_anon_vma(hpage);
860 atomic_inc(&anon_vma->external_refcount);
864 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
866 if (!page_mapped(hpage))
867 rc = move_to_new_page(new_hpage, hpage, 1);
869 if (rc)
870 remove_migration_ptes(hpage, hpage);
872 if (anon_vma && atomic_dec_and_lock(&anon_vma->external_refcount,
873 &anon_vma->lock)) {
874 int empty = list_empty(&anon_vma->head);
875 spin_unlock(&anon_vma->lock);
876 if (empty)
877 anon_vma_free(anon_vma);
880 if (rcu_locked)
881 rcu_read_unlock();
882 out:
883 unlock_page(hpage);
885 if (rc != -EAGAIN) {
886 list_del(&hpage->lru);
887 put_page(hpage);
890 put_page(new_hpage);
892 if (result) {
893 if (rc)
894 *result = rc;
895 else
896 *result = page_to_nid(new_hpage);
898 return rc;
902 * migrate_pages
904 * The function takes one list of pages to migrate and a function
905 * that determines from the page to be migrated and the private data
906 * the target of the move and allocates the page.
908 * The function returns after 10 attempts or if no pages
909 * are movable anymore because to has become empty
910 * or no retryable pages exist anymore.
911 * Caller should call putback_lru_pages to return pages to the LRU
912 * or free list.
914 * Return: Number of pages not migrated or error code.
916 int migrate_pages(struct list_head *from,
917 new_page_t get_new_page, unsigned long private, bool offlining,
918 bool sync)
920 int retry = 1;
921 int nr_failed = 0;
922 int pass = 0;
923 struct page *page;
924 struct page *page2;
925 int swapwrite = current->flags & PF_SWAPWRITE;
926 int rc;
928 if (!swapwrite)
929 current->flags |= PF_SWAPWRITE;
931 for(pass = 0; pass < 10 && retry; pass++) {
932 retry = 0;
934 list_for_each_entry_safe(page, page2, from, lru) {
935 cond_resched();
937 rc = unmap_and_move(get_new_page, private,
938 page, pass > 2, offlining,
939 sync);
941 switch(rc) {
942 case -ENOMEM:
943 goto out;
944 case -EAGAIN:
945 retry++;
946 break;
947 case 0:
948 break;
949 default:
950 /* Permanent failure */
951 nr_failed++;
952 break;
956 rc = 0;
957 out:
958 if (!swapwrite)
959 current->flags &= ~PF_SWAPWRITE;
961 if (rc)
962 return rc;
964 return nr_failed + retry;
967 int migrate_huge_pages(struct list_head *from,
968 new_page_t get_new_page, unsigned long private, bool offlining,
969 bool sync)
971 int retry = 1;
972 int nr_failed = 0;
973 int pass = 0;
974 struct page *page;
975 struct page *page2;
976 int rc;
978 for (pass = 0; pass < 10 && retry; pass++) {
979 retry = 0;
981 list_for_each_entry_safe(page, page2, from, lru) {
982 cond_resched();
984 rc = unmap_and_move_huge_page(get_new_page,
985 private, page, pass > 2, offlining,
986 sync);
988 switch(rc) {
989 case -ENOMEM:
990 goto out;
991 case -EAGAIN:
992 retry++;
993 break;
994 case 0:
995 break;
996 default:
997 /* Permanent failure */
998 nr_failed++;
999 break;
1003 rc = 0;
1004 out:
1006 list_for_each_entry_safe(page, page2, from, lru)
1007 put_page(page);
1009 if (rc)
1010 return rc;
1012 return nr_failed + retry;
1015 #ifdef CONFIG_NUMA
1017 * Move a list of individual pages
1019 struct page_to_node {
1020 unsigned long addr;
1021 struct page *page;
1022 int node;
1023 int status;
1026 static struct page *new_page_node(struct page *p, unsigned long private,
1027 int **result)
1029 struct page_to_node *pm = (struct page_to_node *)private;
1031 while (pm->node != MAX_NUMNODES && pm->page != p)
1032 pm++;
1034 if (pm->node == MAX_NUMNODES)
1035 return NULL;
1037 *result = &pm->status;
1039 return alloc_pages_exact_node(pm->node,
1040 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1044 * Move a set of pages as indicated in the pm array. The addr
1045 * field must be set to the virtual address of the page to be moved
1046 * and the node number must contain a valid target node.
1047 * The pm array ends with node = MAX_NUMNODES.
1049 static int do_move_page_to_node_array(struct mm_struct *mm,
1050 struct page_to_node *pm,
1051 int migrate_all)
1053 int err;
1054 struct page_to_node *pp;
1055 LIST_HEAD(pagelist);
1057 down_read(&mm->mmap_sem);
1060 * Build a list of pages to migrate
1062 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1063 struct vm_area_struct *vma;
1064 struct page *page;
1066 err = -EFAULT;
1067 vma = find_vma(mm, pp->addr);
1068 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1069 goto set_status;
1071 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1073 err = PTR_ERR(page);
1074 if (IS_ERR(page))
1075 goto set_status;
1077 err = -ENOENT;
1078 if (!page)
1079 goto set_status;
1081 /* Use PageReserved to check for zero page */
1082 if (PageReserved(page) || PageKsm(page))
1083 goto put_and_set;
1085 pp->page = page;
1086 err = page_to_nid(page);
1088 if (err == pp->node)
1090 * Node already in the right place
1092 goto put_and_set;
1094 err = -EACCES;
1095 if (page_mapcount(page) > 1 &&
1096 !migrate_all)
1097 goto put_and_set;
1099 err = isolate_lru_page(page);
1100 if (!err) {
1101 list_add_tail(&page->lru, &pagelist);
1102 inc_zone_page_state(page, NR_ISOLATED_ANON +
1103 page_is_file_cache(page));
1105 put_and_set:
1107 * Either remove the duplicate refcount from
1108 * isolate_lru_page() or drop the page ref if it was
1109 * not isolated.
1111 put_page(page);
1112 set_status:
1113 pp->status = err;
1116 err = 0;
1117 if (!list_empty(&pagelist)) {
1118 err = migrate_pages(&pagelist, new_page_node,
1119 (unsigned long)pm, 0, true);
1120 if (err)
1121 putback_lru_pages(&pagelist);
1124 up_read(&mm->mmap_sem);
1125 return err;
1129 * Migrate an array of page address onto an array of nodes and fill
1130 * the corresponding array of status.
1132 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1133 unsigned long nr_pages,
1134 const void __user * __user *pages,
1135 const int __user *nodes,
1136 int __user *status, int flags)
1138 struct page_to_node *pm;
1139 nodemask_t task_nodes;
1140 unsigned long chunk_nr_pages;
1141 unsigned long chunk_start;
1142 int err;
1144 task_nodes = cpuset_mems_allowed(task);
1146 err = -ENOMEM;
1147 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1148 if (!pm)
1149 goto out;
1151 migrate_prep();
1154 * Store a chunk of page_to_node array in a page,
1155 * but keep the last one as a marker
1157 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1159 for (chunk_start = 0;
1160 chunk_start < nr_pages;
1161 chunk_start += chunk_nr_pages) {
1162 int j;
1164 if (chunk_start + chunk_nr_pages > nr_pages)
1165 chunk_nr_pages = nr_pages - chunk_start;
1167 /* fill the chunk pm with addrs and nodes from user-space */
1168 for (j = 0; j < chunk_nr_pages; j++) {
1169 const void __user *p;
1170 int node;
1172 err = -EFAULT;
1173 if (get_user(p, pages + j + chunk_start))
1174 goto out_pm;
1175 pm[j].addr = (unsigned long) p;
1177 if (get_user(node, nodes + j + chunk_start))
1178 goto out_pm;
1180 err = -ENODEV;
1181 if (node < 0 || node >= MAX_NUMNODES)
1182 goto out_pm;
1184 if (!node_state(node, N_HIGH_MEMORY))
1185 goto out_pm;
1187 err = -EACCES;
1188 if (!node_isset(node, task_nodes))
1189 goto out_pm;
1191 pm[j].node = node;
1194 /* End marker for this chunk */
1195 pm[chunk_nr_pages].node = MAX_NUMNODES;
1197 /* Migrate this chunk */
1198 err = do_move_page_to_node_array(mm, pm,
1199 flags & MPOL_MF_MOVE_ALL);
1200 if (err < 0)
1201 goto out_pm;
1203 /* Return status information */
1204 for (j = 0; j < chunk_nr_pages; j++)
1205 if (put_user(pm[j].status, status + j + chunk_start)) {
1206 err = -EFAULT;
1207 goto out_pm;
1210 err = 0;
1212 out_pm:
1213 free_page((unsigned long)pm);
1214 out:
1215 return err;
1219 * Determine the nodes of an array of pages and store it in an array of status.
1221 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1222 const void __user **pages, int *status)
1224 unsigned long i;
1226 down_read(&mm->mmap_sem);
1228 for (i = 0; i < nr_pages; i++) {
1229 unsigned long addr = (unsigned long)(*pages);
1230 struct vm_area_struct *vma;
1231 struct page *page;
1232 int err = -EFAULT;
1234 vma = find_vma(mm, addr);
1235 if (!vma || addr < vma->vm_start)
1236 goto set_status;
1238 page = follow_page(vma, addr, 0);
1240 err = PTR_ERR(page);
1241 if (IS_ERR(page))
1242 goto set_status;
1244 err = -ENOENT;
1245 /* Use PageReserved to check for zero page */
1246 if (!page || PageReserved(page) || PageKsm(page))
1247 goto set_status;
1249 err = page_to_nid(page);
1250 set_status:
1251 *status = err;
1253 pages++;
1254 status++;
1257 up_read(&mm->mmap_sem);
1261 * Determine the nodes of a user array of pages and store it in
1262 * a user array of status.
1264 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1265 const void __user * __user *pages,
1266 int __user *status)
1268 #define DO_PAGES_STAT_CHUNK_NR 16
1269 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1270 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1272 while (nr_pages) {
1273 unsigned long chunk_nr;
1275 chunk_nr = nr_pages;
1276 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1277 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1279 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1280 break;
1282 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1284 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1285 break;
1287 pages += chunk_nr;
1288 status += chunk_nr;
1289 nr_pages -= chunk_nr;
1291 return nr_pages ? -EFAULT : 0;
1295 * Move a list of pages in the address space of the currently executing
1296 * process.
1298 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1299 const void __user * __user *, pages,
1300 const int __user *, nodes,
1301 int __user *, status, int, flags)
1303 const struct cred *cred = current_cred(), *tcred;
1304 struct task_struct *task;
1305 struct mm_struct *mm;
1306 int err;
1308 /* Check flags */
1309 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1310 return -EINVAL;
1312 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1313 return -EPERM;
1315 /* Find the mm_struct */
1316 read_lock(&tasklist_lock);
1317 task = pid ? find_task_by_vpid(pid) : current;
1318 if (!task) {
1319 read_unlock(&tasklist_lock);
1320 return -ESRCH;
1322 mm = get_task_mm(task);
1323 read_unlock(&tasklist_lock);
1325 if (!mm)
1326 return -EINVAL;
1329 * Check if this process has the right to modify the specified
1330 * process. The right exists if the process has administrative
1331 * capabilities, superuser privileges or the same
1332 * userid as the target process.
1334 rcu_read_lock();
1335 tcred = __task_cred(task);
1336 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1337 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1338 !capable(CAP_SYS_NICE)) {
1339 rcu_read_unlock();
1340 err = -EPERM;
1341 goto out;
1343 rcu_read_unlock();
1345 err = security_task_movememory(task);
1346 if (err)
1347 goto out;
1349 if (nodes) {
1350 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1351 flags);
1352 } else {
1353 err = do_pages_stat(mm, nr_pages, pages, status);
1356 out:
1357 mmput(mm);
1358 return err;
1362 * Call migration functions in the vma_ops that may prepare
1363 * memory in a vm for migration. migration functions may perform
1364 * the migration for vmas that do not have an underlying page struct.
1366 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1367 const nodemask_t *from, unsigned long flags)
1369 struct vm_area_struct *vma;
1370 int err = 0;
1372 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1373 if (vma->vm_ops && vma->vm_ops->migrate) {
1374 err = vma->vm_ops->migrate(vma, to, from, flags);
1375 if (err)
1376 break;
1379 return err;
1381 #endif