neigh: Kill bogus SMP protected debugging message.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / migrate.c
blob33358f8781118ae9153ac4eb3e00530ca2f20e93
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);
124 * Peek to check is_swap_pte() before taking ptlock? No, we
125 * can race mremap's move_ptes(), which skips anon_vma lock.
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 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != 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 (!PageSwapCache(page) && 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 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != 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 * Whereas 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, bool sync)
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 {
591 * Do not writeback pages if !sync and migratepage is
592 * not pointing to migrate_page() which is nonblocking
593 * (swapcache/tmpfs uses migratepage = migrate_page).
595 if (PageDirty(page) && !sync &&
596 mapping->a_ops->migratepage != migrate_page)
597 rc = -EBUSY;
598 else if (mapping->a_ops->migratepage)
600 * Most pages have a mapping and most filesystems
601 * should provide a migration function. Anonymous
602 * pages are part of swap space which also has its
603 * own migration function. This is the most common
604 * path for page migration.
606 rc = mapping->a_ops->migratepage(mapping,
607 newpage, page);
608 else
609 rc = fallback_migrate_page(mapping, newpage, page);
612 if (rc) {
613 newpage->mapping = NULL;
614 } else {
615 if (remap_swapcache)
616 remove_migration_ptes(page, newpage);
619 unlock_page(newpage);
621 return rc;
624 static int __unmap_and_move(struct page *page, struct page *newpage,
625 int force, bool offlining, bool sync)
627 int rc = -EAGAIN;
628 int remap_swapcache = 1;
629 int charge = 0;
630 struct mem_cgroup *mem;
631 struct anon_vma *anon_vma = NULL;
633 if (!trylock_page(page)) {
634 if (!force || !sync)
635 goto out;
638 * It's not safe for direct compaction to call lock_page.
639 * For example, during page readahead pages are added locked
640 * to the LRU. Later, when the IO completes the pages are
641 * marked uptodate and unlocked. However, the queueing
642 * could be merging multiple pages for one bio (e.g.
643 * mpage_readpages). If an allocation happens for the
644 * second or third page, the process can end up locking
645 * the same page twice and deadlocking. Rather than
646 * trying to be clever about what pages can be locked,
647 * avoid the use of lock_page for direct compaction
648 * altogether.
650 if (current->flags & PF_MEMALLOC)
651 goto out;
653 lock_page(page);
657 * Only memory hotplug's offline_pages() caller has locked out KSM,
658 * and can safely migrate a KSM page. The other cases have skipped
659 * PageKsm along with PageReserved - but it is only now when we have
660 * the page lock that we can be certain it will not go KSM beneath us
661 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
662 * its pagecount raised, but only here do we take the page lock which
663 * serializes that).
665 if (PageKsm(page) && !offlining) {
666 rc = -EBUSY;
667 goto unlock;
670 /* charge against new page */
671 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
672 if (charge == -ENOMEM) {
673 rc = -ENOMEM;
674 goto unlock;
676 BUG_ON(charge);
678 if (PageWriteback(page)) {
680 * For !sync, there is no point retrying as the retry loop
681 * is expected to be too short for PageWriteback to be cleared
683 if (!sync) {
684 rc = -EBUSY;
685 goto uncharge;
687 if (!force)
688 goto uncharge;
689 wait_on_page_writeback(page);
692 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
693 * we cannot notice that anon_vma is freed while we migrates a page.
694 * This get_anon_vma() delays freeing anon_vma pointer until the end
695 * of migration. File cache pages are no problem because of page_lock()
696 * File Caches may use write_page() or lock_page() in migration, then,
697 * just care Anon page here.
699 if (PageAnon(page)) {
701 * Only page_lock_anon_vma() understands the subtleties of
702 * getting a hold on an anon_vma from outside one of its mms.
704 anon_vma = page_get_anon_vma(page);
705 if (anon_vma) {
707 * Anon page
709 } else if (PageSwapCache(page)) {
711 * We cannot be sure that the anon_vma of an unmapped
712 * swapcache page is safe to use because we don't
713 * know in advance if the VMA that this page belonged
714 * to still exists. If the VMA and others sharing the
715 * data have been freed, then the anon_vma could
716 * already be invalid.
718 * To avoid this possibility, swapcache pages get
719 * migrated but are not remapped when migration
720 * completes
722 remap_swapcache = 0;
723 } else {
724 goto uncharge;
729 * Corner case handling:
730 * 1. When a new swap-cache page is read into, it is added to the LRU
731 * and treated as swapcache but it has no rmap yet.
732 * Calling try_to_unmap() against a page->mapping==NULL page will
733 * trigger a BUG. So handle it here.
734 * 2. An orphaned page (see truncate_complete_page) might have
735 * fs-private metadata. The page can be picked up due to memory
736 * offlining. Everywhere else except page reclaim, the page is
737 * invisible to the vm, so the page can not be migrated. So try to
738 * free the metadata, so the page can be freed.
740 if (!page->mapping) {
741 VM_BUG_ON(PageAnon(page));
742 if (page_has_private(page)) {
743 try_to_free_buffers(page);
744 goto uncharge;
746 goto skip_unmap;
749 /* Establish migration ptes or remove ptes */
750 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
752 skip_unmap:
753 if (!page_mapped(page))
754 rc = move_to_new_page(newpage, page, remap_swapcache, sync);
756 if (rc && remap_swapcache)
757 remove_migration_ptes(page, page);
759 /* Drop an anon_vma reference if we took one */
760 if (anon_vma)
761 put_anon_vma(anon_vma);
763 uncharge:
764 if (!charge)
765 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
766 unlock:
767 unlock_page(page);
768 out:
769 return rc;
773 * Obtain the lock on page, remove all ptes and migrate the page
774 * to the newly allocated page in newpage.
776 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
777 struct page *page, int force, bool offlining, bool sync)
779 int rc = 0;
780 int *result = NULL;
781 struct page *newpage = get_new_page(page, private, &result);
783 if (!newpage)
784 return -ENOMEM;
786 if (page_count(page) == 1) {
787 /* page was freed from under us. So we are done. */
788 goto out;
791 if (unlikely(PageTransHuge(page)))
792 if (unlikely(split_huge_page(page)))
793 goto out;
795 rc = __unmap_and_move(page, newpage, force, offlining, sync);
796 out:
797 if (rc != -EAGAIN) {
799 * A page that has been migrated has all references
800 * removed and will be freed. A page that has not been
801 * migrated will have kepts its references and be
802 * restored.
804 list_del(&page->lru);
805 dec_zone_page_state(page, NR_ISOLATED_ANON +
806 page_is_file_cache(page));
807 putback_lru_page(page);
810 * Move the new page to the LRU. If migration was not successful
811 * then this will free the page.
813 putback_lru_page(newpage);
814 if (result) {
815 if (rc)
816 *result = rc;
817 else
818 *result = page_to_nid(newpage);
820 return rc;
824 * Counterpart of unmap_and_move_page() for hugepage migration.
826 * This function doesn't wait the completion of hugepage I/O
827 * because there is no race between I/O and migration for hugepage.
828 * Note that currently hugepage I/O occurs only in direct I/O
829 * where no lock is held and PG_writeback is irrelevant,
830 * and writeback status of all subpages are counted in the reference
831 * count of the head page (i.e. if all subpages of a 2MB hugepage are
832 * under direct I/O, the reference of the head page is 512 and a bit more.)
833 * This means that when we try to migrate hugepage whose subpages are
834 * doing direct I/O, some references remain after try_to_unmap() and
835 * hugepage migration fails without data corruption.
837 * There is also no race when direct I/O is issued on the page under migration,
838 * because then pte is replaced with migration swap entry and direct I/O code
839 * will wait in the page fault for migration to complete.
841 static int unmap_and_move_huge_page(new_page_t get_new_page,
842 unsigned long private, struct page *hpage,
843 int force, bool offlining, bool sync)
845 int rc = 0;
846 int *result = NULL;
847 struct page *new_hpage = get_new_page(hpage, private, &result);
848 struct anon_vma *anon_vma = NULL;
850 if (!new_hpage)
851 return -ENOMEM;
853 rc = -EAGAIN;
855 if (!trylock_page(hpage)) {
856 if (!force || !sync)
857 goto out;
858 lock_page(hpage);
861 if (PageAnon(hpage))
862 anon_vma = page_get_anon_vma(hpage);
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, sync);
869 if (rc)
870 remove_migration_ptes(hpage, hpage);
872 if (anon_vma)
873 put_anon_vma(anon_vma);
874 out:
875 unlock_page(hpage);
877 if (rc != -EAGAIN) {
878 list_del(&hpage->lru);
879 put_page(hpage);
882 put_page(new_hpage);
884 if (result) {
885 if (rc)
886 *result = rc;
887 else
888 *result = page_to_nid(new_hpage);
890 return rc;
894 * migrate_pages
896 * The function takes one list of pages to migrate and a function
897 * that determines from the page to be migrated and the private data
898 * the target of the move and allocates the page.
900 * The function returns after 10 attempts or if no pages
901 * are movable anymore because to has become empty
902 * or no retryable pages exist anymore.
903 * Caller should call putback_lru_pages to return pages to the LRU
904 * or free list only if ret != 0.
906 * Return: Number of pages not migrated or error code.
908 int migrate_pages(struct list_head *from,
909 new_page_t get_new_page, unsigned long private, bool offlining,
910 bool sync)
912 int retry = 1;
913 int nr_failed = 0;
914 int pass = 0;
915 struct page *page;
916 struct page *page2;
917 int swapwrite = current->flags & PF_SWAPWRITE;
918 int rc;
920 if (!swapwrite)
921 current->flags |= PF_SWAPWRITE;
923 for(pass = 0; pass < 10 && retry; pass++) {
924 retry = 0;
926 list_for_each_entry_safe(page, page2, from, lru) {
927 cond_resched();
929 rc = unmap_and_move(get_new_page, private,
930 page, pass > 2, offlining,
931 sync);
933 switch(rc) {
934 case -ENOMEM:
935 goto out;
936 case -EAGAIN:
937 retry++;
938 break;
939 case 0:
940 break;
941 default:
942 /* Permanent failure */
943 nr_failed++;
944 break;
948 rc = 0;
949 out:
950 if (!swapwrite)
951 current->flags &= ~PF_SWAPWRITE;
953 if (rc)
954 return rc;
956 return nr_failed + retry;
959 int migrate_huge_pages(struct list_head *from,
960 new_page_t get_new_page, unsigned long private, bool offlining,
961 bool sync)
963 int retry = 1;
964 int nr_failed = 0;
965 int pass = 0;
966 struct page *page;
967 struct page *page2;
968 int rc;
970 for (pass = 0; pass < 10 && retry; pass++) {
971 retry = 0;
973 list_for_each_entry_safe(page, page2, from, lru) {
974 cond_resched();
976 rc = unmap_and_move_huge_page(get_new_page,
977 private, page, pass > 2, offlining,
978 sync);
980 switch(rc) {
981 case -ENOMEM:
982 goto out;
983 case -EAGAIN:
984 retry++;
985 break;
986 case 0:
987 break;
988 default:
989 /* Permanent failure */
990 nr_failed++;
991 break;
995 rc = 0;
996 out:
997 if (rc)
998 return rc;
1000 return nr_failed + retry;
1003 #ifdef CONFIG_NUMA
1005 * Move a list of individual pages
1007 struct page_to_node {
1008 unsigned long addr;
1009 struct page *page;
1010 int node;
1011 int status;
1014 static struct page *new_page_node(struct page *p, unsigned long private,
1015 int **result)
1017 struct page_to_node *pm = (struct page_to_node *)private;
1019 while (pm->node != MAX_NUMNODES && pm->page != p)
1020 pm++;
1022 if (pm->node == MAX_NUMNODES)
1023 return NULL;
1025 *result = &pm->status;
1027 return alloc_pages_exact_node(pm->node,
1028 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1032 * Move a set of pages as indicated in the pm array. The addr
1033 * field must be set to the virtual address of the page to be moved
1034 * and the node number must contain a valid target node.
1035 * The pm array ends with node = MAX_NUMNODES.
1037 static int do_move_page_to_node_array(struct mm_struct *mm,
1038 struct page_to_node *pm,
1039 int migrate_all)
1041 int err;
1042 struct page_to_node *pp;
1043 LIST_HEAD(pagelist);
1045 down_read(&mm->mmap_sem);
1048 * Build a list of pages to migrate
1050 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1051 struct vm_area_struct *vma;
1052 struct page *page;
1054 err = -EFAULT;
1055 vma = find_vma(mm, pp->addr);
1056 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1057 goto set_status;
1059 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1061 err = PTR_ERR(page);
1062 if (IS_ERR(page))
1063 goto set_status;
1065 err = -ENOENT;
1066 if (!page)
1067 goto set_status;
1069 /* Use PageReserved to check for zero page */
1070 if (PageReserved(page) || PageKsm(page))
1071 goto put_and_set;
1073 pp->page = page;
1074 err = page_to_nid(page);
1076 if (err == pp->node)
1078 * Node already in the right place
1080 goto put_and_set;
1082 err = -EACCES;
1083 if (page_mapcount(page) > 1 &&
1084 !migrate_all)
1085 goto put_and_set;
1087 err = isolate_lru_page(page);
1088 if (!err) {
1089 list_add_tail(&page->lru, &pagelist);
1090 inc_zone_page_state(page, NR_ISOLATED_ANON +
1091 page_is_file_cache(page));
1093 put_and_set:
1095 * Either remove the duplicate refcount from
1096 * isolate_lru_page() or drop the page ref if it was
1097 * not isolated.
1099 put_page(page);
1100 set_status:
1101 pp->status = err;
1104 err = 0;
1105 if (!list_empty(&pagelist)) {
1106 err = migrate_pages(&pagelist, new_page_node,
1107 (unsigned long)pm, 0, true);
1108 if (err)
1109 putback_lru_pages(&pagelist);
1112 up_read(&mm->mmap_sem);
1113 return err;
1117 * Migrate an array of page address onto an array of nodes and fill
1118 * the corresponding array of status.
1120 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1121 unsigned long nr_pages,
1122 const void __user * __user *pages,
1123 const int __user *nodes,
1124 int __user *status, int flags)
1126 struct page_to_node *pm;
1127 nodemask_t task_nodes;
1128 unsigned long chunk_nr_pages;
1129 unsigned long chunk_start;
1130 int err;
1132 task_nodes = cpuset_mems_allowed(task);
1134 err = -ENOMEM;
1135 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1136 if (!pm)
1137 goto out;
1139 migrate_prep();
1142 * Store a chunk of page_to_node array in a page,
1143 * but keep the last one as a marker
1145 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1147 for (chunk_start = 0;
1148 chunk_start < nr_pages;
1149 chunk_start += chunk_nr_pages) {
1150 int j;
1152 if (chunk_start + chunk_nr_pages > nr_pages)
1153 chunk_nr_pages = nr_pages - chunk_start;
1155 /* fill the chunk pm with addrs and nodes from user-space */
1156 for (j = 0; j < chunk_nr_pages; j++) {
1157 const void __user *p;
1158 int node;
1160 err = -EFAULT;
1161 if (get_user(p, pages + j + chunk_start))
1162 goto out_pm;
1163 pm[j].addr = (unsigned long) p;
1165 if (get_user(node, nodes + j + chunk_start))
1166 goto out_pm;
1168 err = -ENODEV;
1169 if (node < 0 || node >= MAX_NUMNODES)
1170 goto out_pm;
1172 if (!node_state(node, N_HIGH_MEMORY))
1173 goto out_pm;
1175 err = -EACCES;
1176 if (!node_isset(node, task_nodes))
1177 goto out_pm;
1179 pm[j].node = node;
1182 /* End marker for this chunk */
1183 pm[chunk_nr_pages].node = MAX_NUMNODES;
1185 /* Migrate this chunk */
1186 err = do_move_page_to_node_array(mm, pm,
1187 flags & MPOL_MF_MOVE_ALL);
1188 if (err < 0)
1189 goto out_pm;
1191 /* Return status information */
1192 for (j = 0; j < chunk_nr_pages; j++)
1193 if (put_user(pm[j].status, status + j + chunk_start)) {
1194 err = -EFAULT;
1195 goto out_pm;
1198 err = 0;
1200 out_pm:
1201 free_page((unsigned long)pm);
1202 out:
1203 return err;
1207 * Determine the nodes of an array of pages and store it in an array of status.
1209 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1210 const void __user **pages, int *status)
1212 unsigned long i;
1214 down_read(&mm->mmap_sem);
1216 for (i = 0; i < nr_pages; i++) {
1217 unsigned long addr = (unsigned long)(*pages);
1218 struct vm_area_struct *vma;
1219 struct page *page;
1220 int err = -EFAULT;
1222 vma = find_vma(mm, addr);
1223 if (!vma || addr < vma->vm_start)
1224 goto set_status;
1226 page = follow_page(vma, addr, 0);
1228 err = PTR_ERR(page);
1229 if (IS_ERR(page))
1230 goto set_status;
1232 err = -ENOENT;
1233 /* Use PageReserved to check for zero page */
1234 if (!page || PageReserved(page) || PageKsm(page))
1235 goto set_status;
1237 err = page_to_nid(page);
1238 set_status:
1239 *status = err;
1241 pages++;
1242 status++;
1245 up_read(&mm->mmap_sem);
1249 * Determine the nodes of a user array of pages and store it in
1250 * a user array of status.
1252 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1253 const void __user * __user *pages,
1254 int __user *status)
1256 #define DO_PAGES_STAT_CHUNK_NR 16
1257 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1258 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1260 while (nr_pages) {
1261 unsigned long chunk_nr;
1263 chunk_nr = nr_pages;
1264 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1265 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1267 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1268 break;
1270 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1272 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1273 break;
1275 pages += chunk_nr;
1276 status += chunk_nr;
1277 nr_pages -= chunk_nr;
1279 return nr_pages ? -EFAULT : 0;
1283 * Move a list of pages in the address space of the currently executing
1284 * process.
1286 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1287 const void __user * __user *, pages,
1288 const int __user *, nodes,
1289 int __user *, status, int, flags)
1291 const struct cred *cred = current_cred(), *tcred;
1292 struct task_struct *task;
1293 struct mm_struct *mm;
1294 int err;
1296 /* Check flags */
1297 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1298 return -EINVAL;
1300 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1301 return -EPERM;
1303 /* Find the mm_struct */
1304 rcu_read_lock();
1305 task = pid ? find_task_by_vpid(pid) : current;
1306 if (!task) {
1307 rcu_read_unlock();
1308 return -ESRCH;
1310 mm = get_task_mm(task);
1311 rcu_read_unlock();
1313 if (!mm)
1314 return -EINVAL;
1317 * Check if this process has the right to modify the specified
1318 * process. The right exists if the process has administrative
1319 * capabilities, superuser privileges or the same
1320 * userid as the target process.
1322 rcu_read_lock();
1323 tcred = __task_cred(task);
1324 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1325 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1326 !capable(CAP_SYS_NICE)) {
1327 rcu_read_unlock();
1328 err = -EPERM;
1329 goto out;
1331 rcu_read_unlock();
1333 err = security_task_movememory(task);
1334 if (err)
1335 goto out;
1337 if (nodes) {
1338 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1339 flags);
1340 } else {
1341 err = do_pages_stat(mm, nr_pages, pages, status);
1344 out:
1345 mmput(mm);
1346 return err;
1350 * Call migration functions in the vma_ops that may prepare
1351 * memory in a vm for migration. migration functions may perform
1352 * the migration for vmas that do not have an underlying page struct.
1354 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1355 const nodemask_t *from, unsigned long flags)
1357 struct vm_area_struct *vma;
1358 int err = 0;
1360 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1361 if (vma->vm_ops && vma->vm_ops->migrate) {
1362 err = vma->vm_ops->migrate(vma, to, from, flags);
1363 if (err)
1364 break;
1367 return err;
1369 #endif