4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
36 #include <linux/delay.h>
38 #include <asm/tlbflush.h>
39 #include <asm/div64.h>
41 #include <linux/swapops.h>
45 /* possible outcome of pageout() */
47 /* failed to write page out, page is locked */
49 /* move page to the active list, page is locked */
51 /* page has been sent to the disk successfully, page is unlocked */
53 /* page is clean and locked */
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned
;
61 unsigned long nr_mapped
; /* From page_state */
63 /* This context's GFP mask */
68 /* Can pages be swapped as part of reclaim? */
71 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
72 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
73 * In this context, it doesn't matter that we scan the
74 * whole list at once. */
79 * The list of shrinker callbacks used by to apply pressure to
84 struct list_head list
;
85 int seeks
; /* seeks to recreate an obj */
86 long nr
; /* objs pending delete */
89 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
91 #ifdef ARCH_HAS_PREFETCH
92 #define prefetch_prev_lru_page(_page, _base, _field) \
94 if ((_page)->lru.prev != _base) { \
97 prev = lru_to_page(&(_page->lru)); \
98 prefetch(&prev->_field); \
102 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
105 #ifdef ARCH_HAS_PREFETCHW
106 #define prefetchw_prev_lru_page(_page, _base, _field) \
108 if ((_page)->lru.prev != _base) { \
111 prev = lru_to_page(&(_page->lru)); \
112 prefetchw(&prev->_field); \
116 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
120 * From 0 .. 100. Higher means more swappy.
122 int vm_swappiness
= 60;
123 static long total_memory
;
125 static LIST_HEAD(shrinker_list
);
126 static DECLARE_RWSEM(shrinker_rwsem
);
129 * Add a shrinker callback to be called from the vm
131 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
133 struct shrinker
*shrinker
;
135 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
137 shrinker
->shrinker
= theshrinker
;
138 shrinker
->seeks
= seeks
;
140 down_write(&shrinker_rwsem
);
141 list_add_tail(&shrinker
->list
, &shrinker_list
);
142 up_write(&shrinker_rwsem
);
146 EXPORT_SYMBOL(set_shrinker
);
151 void remove_shrinker(struct shrinker
*shrinker
)
153 down_write(&shrinker_rwsem
);
154 list_del(&shrinker
->list
);
155 up_write(&shrinker_rwsem
);
158 EXPORT_SYMBOL(remove_shrinker
);
160 #define SHRINK_BATCH 128
162 * Call the shrink functions to age shrinkable caches
164 * Here we assume it costs one seek to replace a lru page and that it also
165 * takes a seek to recreate a cache object. With this in mind we age equal
166 * percentages of the lru and ageable caches. This should balance the seeks
167 * generated by these structures.
169 * If the vm encounted mapped pages on the LRU it increase the pressure on
170 * slab to avoid swapping.
172 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
174 * `lru_pages' represents the number of on-LRU pages in all the zones which
175 * are eligible for the caller's allocation attempt. It is used for balancing
176 * slab reclaim versus page reclaim.
178 * Returns the number of slab objects which we shrunk.
180 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
181 unsigned long lru_pages
)
183 struct shrinker
*shrinker
;
184 unsigned long ret
= 0;
187 scanned
= SWAP_CLUSTER_MAX
;
189 if (!down_read_trylock(&shrinker_rwsem
))
190 return 1; /* Assume we'll be able to shrink next time */
192 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
193 unsigned long long delta
;
194 unsigned long total_scan
;
195 unsigned long max_pass
= (*shrinker
->shrinker
)(0, gfp_mask
);
197 delta
= (4 * scanned
) / shrinker
->seeks
;
199 do_div(delta
, lru_pages
+ 1);
200 shrinker
->nr
+= delta
;
201 if (shrinker
->nr
< 0) {
202 printk(KERN_ERR
"%s: nr=%ld\n",
203 __FUNCTION__
, shrinker
->nr
);
204 shrinker
->nr
= max_pass
;
208 * Avoid risking looping forever due to too large nr value:
209 * never try to free more than twice the estimate number of
212 if (shrinker
->nr
> max_pass
* 2)
213 shrinker
->nr
= max_pass
* 2;
215 total_scan
= shrinker
->nr
;
218 while (total_scan
>= SHRINK_BATCH
) {
219 long this_scan
= SHRINK_BATCH
;
223 nr_before
= (*shrinker
->shrinker
)(0, gfp_mask
);
224 shrink_ret
= (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
225 if (shrink_ret
== -1)
227 if (shrink_ret
< nr_before
)
228 ret
+= nr_before
- shrink_ret
;
229 mod_page_state(slabs_scanned
, this_scan
);
230 total_scan
-= this_scan
;
235 shrinker
->nr
+= total_scan
;
237 up_read(&shrinker_rwsem
);
241 /* Called without lock on whether page is mapped, so answer is unstable */
242 static inline int page_mapping_inuse(struct page
*page
)
244 struct address_space
*mapping
;
246 /* Page is in somebody's page tables. */
247 if (page_mapped(page
))
250 /* Be more reluctant to reclaim swapcache than pagecache */
251 if (PageSwapCache(page
))
254 mapping
= page_mapping(page
);
258 /* File is mmap'd by somebody? */
259 return mapping_mapped(mapping
);
262 static inline int is_page_cache_freeable(struct page
*page
)
264 return page_count(page
) - !!PagePrivate(page
) == 2;
267 static int may_write_to_queue(struct backing_dev_info
*bdi
)
269 if (current
->flags
& PF_SWAPWRITE
)
271 if (!bdi_write_congested(bdi
))
273 if (bdi
== current
->backing_dev_info
)
279 * We detected a synchronous write error writing a page out. Probably
280 * -ENOSPC. We need to propagate that into the address_space for a subsequent
281 * fsync(), msync() or close().
283 * The tricky part is that after writepage we cannot touch the mapping: nothing
284 * prevents it from being freed up. But we have a ref on the page and once
285 * that page is locked, the mapping is pinned.
287 * We're allowed to run sleeping lock_page() here because we know the caller has
290 static void handle_write_error(struct address_space
*mapping
,
291 struct page
*page
, int error
)
294 if (page_mapping(page
) == mapping
) {
295 if (error
== -ENOSPC
)
296 set_bit(AS_ENOSPC
, &mapping
->flags
);
298 set_bit(AS_EIO
, &mapping
->flags
);
304 * pageout is called by shrink_page_list() for each dirty page.
305 * Calls ->writepage().
307 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
310 * If the page is dirty, only perform writeback if that write
311 * will be non-blocking. To prevent this allocation from being
312 * stalled by pagecache activity. But note that there may be
313 * stalls if we need to run get_block(). We could test
314 * PagePrivate for that.
316 * If this process is currently in generic_file_write() against
317 * this page's queue, we can perform writeback even if that
320 * If the page is swapcache, write it back even if that would
321 * block, for some throttling. This happens by accident, because
322 * swap_backing_dev_info is bust: it doesn't reflect the
323 * congestion state of the swapdevs. Easy to fix, if needed.
324 * See swapfile.c:page_queue_congested().
326 if (!is_page_cache_freeable(page
))
330 * Some data journaling orphaned pages can have
331 * page->mapping == NULL while being dirty with clean buffers.
333 if (PagePrivate(page
)) {
334 if (try_to_free_buffers(page
)) {
335 ClearPageDirty(page
);
336 printk("%s: orphaned page\n", __FUNCTION__
);
342 if (mapping
->a_ops
->writepage
== NULL
)
343 return PAGE_ACTIVATE
;
344 if (!may_write_to_queue(mapping
->backing_dev_info
))
347 if (clear_page_dirty_for_io(page
)) {
349 struct writeback_control wbc
= {
350 .sync_mode
= WB_SYNC_NONE
,
351 .nr_to_write
= SWAP_CLUSTER_MAX
,
356 SetPageReclaim(page
);
357 res
= mapping
->a_ops
->writepage(page
, &wbc
);
359 handle_write_error(mapping
, page
, res
);
360 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
361 ClearPageReclaim(page
);
362 return PAGE_ACTIVATE
;
364 if (!PageWriteback(page
)) {
365 /* synchronous write or broken a_ops? */
366 ClearPageReclaim(page
);
375 static int remove_mapping(struct address_space
*mapping
, struct page
*page
)
378 return 0; /* truncate got there first */
380 write_lock_irq(&mapping
->tree_lock
);
383 * The non-racy check for busy page. It is critical to check
384 * PageDirty _after_ making sure that the page is freeable and
385 * not in use by anybody. (pagecache + us == 2)
387 if (unlikely(page_count(page
) != 2))
390 if (unlikely(PageDirty(page
)))
393 if (PageSwapCache(page
)) {
394 swp_entry_t swap
= { .val
= page_private(page
) };
395 __delete_from_swap_cache(page
);
396 write_unlock_irq(&mapping
->tree_lock
);
398 __put_page(page
); /* The pagecache ref */
402 __remove_from_page_cache(page
);
403 write_unlock_irq(&mapping
->tree_lock
);
408 write_unlock_irq(&mapping
->tree_lock
);
413 * shrink_page_list() returns the number of reclaimed pages
415 static unsigned long shrink_page_list(struct list_head
*page_list
,
416 struct scan_control
*sc
)
418 LIST_HEAD(ret_pages
);
419 struct pagevec freed_pvec
;
421 unsigned long nr_reclaimed
= 0;
425 pagevec_init(&freed_pvec
, 1);
426 while (!list_empty(page_list
)) {
427 struct address_space
*mapping
;
434 page
= lru_to_page(page_list
);
435 list_del(&page
->lru
);
437 if (TestSetPageLocked(page
))
440 BUG_ON(PageActive(page
));
444 if (!sc
->may_swap
&& page_mapped(page
))
447 /* Double the slab pressure for mapped and swapcache pages */
448 if (page_mapped(page
) || PageSwapCache(page
))
451 if (PageWriteback(page
))
454 referenced
= page_referenced(page
, 1);
455 /* In active use or really unfreeable? Activate it. */
456 if (referenced
&& page_mapping_inuse(page
))
457 goto activate_locked
;
461 * Anonymous process memory has backing store?
462 * Try to allocate it some swap space here.
464 if (PageAnon(page
) && !PageSwapCache(page
))
465 if (!add_to_swap(page
, GFP_ATOMIC
))
466 goto activate_locked
;
467 #endif /* CONFIG_SWAP */
469 mapping
= page_mapping(page
);
470 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
471 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
474 * The page is mapped into the page tables of one or more
475 * processes. Try to unmap it here.
477 if (page_mapped(page
) && mapping
) {
478 switch (try_to_unmap(page
, 0)) {
480 goto activate_locked
;
484 ; /* try to free the page below */
488 if (PageDirty(page
)) {
493 if (!sc
->may_writepage
)
496 /* Page is dirty, try to write it out here */
497 switch(pageout(page
, mapping
)) {
501 goto activate_locked
;
503 if (PageWriteback(page
) || PageDirty(page
))
506 * A synchronous write - probably a ramdisk. Go
507 * ahead and try to reclaim the page.
509 if (TestSetPageLocked(page
))
511 if (PageDirty(page
) || PageWriteback(page
))
513 mapping
= page_mapping(page
);
515 ; /* try to free the page below */
520 * If the page has buffers, try to free the buffer mappings
521 * associated with this page. If we succeed we try to free
524 * We do this even if the page is PageDirty().
525 * try_to_release_page() does not perform I/O, but it is
526 * possible for a page to have PageDirty set, but it is actually
527 * clean (all its buffers are clean). This happens if the
528 * buffers were written out directly, with submit_bh(). ext3
529 * will do this, as well as the blockdev mapping.
530 * try_to_release_page() will discover that cleanness and will
531 * drop the buffers and mark the page clean - it can be freed.
533 * Rarely, pages can have buffers and no ->mapping. These are
534 * the pages which were not successfully invalidated in
535 * truncate_complete_page(). We try to drop those buffers here
536 * and if that worked, and the page is no longer mapped into
537 * process address space (page_count == 1) it can be freed.
538 * Otherwise, leave the page on the LRU so it is swappable.
540 if (PagePrivate(page
)) {
541 if (!try_to_release_page(page
, sc
->gfp_mask
))
542 goto activate_locked
;
543 if (!mapping
&& page_count(page
) == 1)
547 if (!remove_mapping(mapping
, page
))
553 if (!pagevec_add(&freed_pvec
, page
))
554 __pagevec_release_nonlru(&freed_pvec
);
563 list_add(&page
->lru
, &ret_pages
);
564 BUG_ON(PageLRU(page
));
566 list_splice(&ret_pages
, page_list
);
567 if (pagevec_count(&freed_pvec
))
568 __pagevec_release_nonlru(&freed_pvec
);
569 mod_page_state(pgactivate
, pgactivate
);
573 #ifdef CONFIG_MIGRATION
574 static inline void move_to_lru(struct page
*page
)
576 list_del(&page
->lru
);
577 if (PageActive(page
)) {
579 * lru_cache_add_active checks that
580 * the PG_active bit is off.
582 ClearPageActive(page
);
583 lru_cache_add_active(page
);
591 * Add isolated pages on the list back to the LRU.
593 * returns the number of pages put back.
595 unsigned long putback_lru_pages(struct list_head
*l
)
599 unsigned long count
= 0;
601 list_for_each_entry_safe(page
, page2
, l
, lru
) {
609 * Non migratable page
611 int fail_migrate_page(struct page
*newpage
, struct page
*page
)
615 EXPORT_SYMBOL(fail_migrate_page
);
618 * swapout a single page
619 * page is locked upon entry, unlocked on exit
621 static int swap_page(struct page
*page
)
623 struct address_space
*mapping
= page_mapping(page
);
625 if (page_mapped(page
) && mapping
)
626 if (try_to_unmap(page
, 1) != SWAP_SUCCESS
)
629 if (PageDirty(page
)) {
630 /* Page is dirty, try to write it out here */
631 switch(pageout(page
, mapping
)) {
640 ; /* try to free the page below */
644 if (PagePrivate(page
)) {
645 if (!try_to_release_page(page
, GFP_KERNEL
) ||
646 (!mapping
&& page_count(page
) == 1))
650 if (remove_mapping(mapping
, page
)) {
662 EXPORT_SYMBOL(swap_page
);
665 * Page migration was first developed in the context of the memory hotplug
666 * project. The main authors of the migration code are:
668 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
669 * Hirokazu Takahashi <taka@valinux.co.jp>
670 * Dave Hansen <haveblue@us.ibm.com>
671 * Christoph Lameter <clameter@sgi.com>
675 * Remove references for a page and establish the new page with the correct
676 * basic settings to be able to stop accesses to the page.
678 int migrate_page_remove_references(struct page
*newpage
,
679 struct page
*page
, int nr_refs
)
681 struct address_space
*mapping
= page_mapping(page
);
682 struct page
**radix_pointer
;
685 * Avoid doing any of the following work if the page count
686 * indicates that the page is in use or truncate has removed
689 if (!mapping
|| page_mapcount(page
) + nr_refs
!= page_count(page
))
693 * Establish swap ptes for anonymous pages or destroy pte
696 * In order to reestablish file backed mappings the fault handlers
697 * will take the radix tree_lock which may then be used to stop
698 * processses from accessing this page until the new page is ready.
700 * A process accessing via a swap pte (an anonymous page) will take a
701 * page_lock on the old page which will block the process until the
702 * migration attempt is complete. At that time the PageSwapCache bit
703 * will be examined. If the page was migrated then the PageSwapCache
704 * bit will be clear and the operation to retrieve the page will be
705 * retried which will find the new page in the radix tree. Then a new
706 * direct mapping may be generated based on the radix tree contents.
708 * If the page was not migrated then the PageSwapCache bit
709 * is still set and the operation may continue.
711 if (try_to_unmap(page
, 1) == SWAP_FAIL
)
712 /* A vma has VM_LOCKED set -> Permanent failure */
716 * Give up if we were unable to remove all mappings.
718 if (page_mapcount(page
))
721 write_lock_irq(&mapping
->tree_lock
);
723 radix_pointer
= (struct page
**)radix_tree_lookup_slot(
727 if (!page_mapping(page
) || page_count(page
) != nr_refs
||
728 *radix_pointer
!= page
) {
729 write_unlock_irq(&mapping
->tree_lock
);
734 * Now we know that no one else is looking at the page.
736 * Certain minimal information about a page must be available
737 * in order for other subsystems to properly handle the page if they
738 * find it through the radix tree update before we are finished
742 newpage
->index
= page
->index
;
743 newpage
->mapping
= page
->mapping
;
744 if (PageSwapCache(page
)) {
745 SetPageSwapCache(newpage
);
746 set_page_private(newpage
, page_private(page
));
749 *radix_pointer
= newpage
;
751 write_unlock_irq(&mapping
->tree_lock
);
755 EXPORT_SYMBOL(migrate_page_remove_references
);
758 * Copy the page to its new location
760 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
762 copy_highpage(newpage
, page
);
765 SetPageError(newpage
);
766 if (PageReferenced(page
))
767 SetPageReferenced(newpage
);
768 if (PageUptodate(page
))
769 SetPageUptodate(newpage
);
770 if (PageActive(page
))
771 SetPageActive(newpage
);
772 if (PageChecked(page
))
773 SetPageChecked(newpage
);
774 if (PageMappedToDisk(page
))
775 SetPageMappedToDisk(newpage
);
777 if (PageDirty(page
)) {
778 clear_page_dirty_for_io(page
);
779 set_page_dirty(newpage
);
782 ClearPageSwapCache(page
);
783 ClearPageActive(page
);
784 ClearPagePrivate(page
);
785 set_page_private(page
, 0);
786 page
->mapping
= NULL
;
789 * If any waiters have accumulated on the new page then
792 if (PageWriteback(newpage
))
793 end_page_writeback(newpage
);
795 EXPORT_SYMBOL(migrate_page_copy
);
798 * Common logic to directly migrate a single page suitable for
799 * pages that do not use PagePrivate.
801 * Pages are locked upon entry and exit.
803 int migrate_page(struct page
*newpage
, struct page
*page
)
807 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
809 rc
= migrate_page_remove_references(newpage
, page
, 2);
814 migrate_page_copy(newpage
, page
);
817 * Remove auxiliary swap entries and replace
818 * them with real ptes.
820 * Note that a real pte entry will allow processes that are not
821 * waiting on the page lock to use the new page via the page tables
822 * before the new page is unlocked.
824 remove_from_swap(newpage
);
827 EXPORT_SYMBOL(migrate_page
);
832 * Two lists are passed to this function. The first list
833 * contains the pages isolated from the LRU to be migrated.
834 * The second list contains new pages that the pages isolated
835 * can be moved to. If the second list is NULL then all
836 * pages are swapped out.
838 * The function returns after 10 attempts or if no pages
839 * are movable anymore because to has become empty
840 * or no retryable pages exist anymore.
842 * Return: Number of pages not migrated when "to" ran empty.
844 unsigned long migrate_pages(struct list_head
*from
, struct list_head
*to
,
845 struct list_head
*moved
, struct list_head
*failed
)
848 unsigned long nr_failed
= 0;
852 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
856 current
->flags
|= PF_SWAPWRITE
;
861 list_for_each_entry_safe(page
, page2
, from
, lru
) {
862 struct page
*newpage
= NULL
;
863 struct address_space
*mapping
;
868 if (page_count(page
) == 1)
869 /* page was freed from under us. So we are done. */
872 if (to
&& list_empty(to
))
876 * Skip locked pages during the first two passes to give the
877 * functions holding the lock time to release the page. Later we
878 * use lock_page() to have a higher chance of acquiring the
885 if (TestSetPageLocked(page
))
889 * Only wait on writeback if we have already done a pass where
890 * we we may have triggered writeouts for lots of pages.
893 wait_on_page_writeback(page
);
895 if (PageWriteback(page
))
900 * Anonymous pages must have swap cache references otherwise
901 * the information contained in the page maps cannot be
904 if (PageAnon(page
) && !PageSwapCache(page
)) {
905 if (!add_to_swap(page
, GFP_KERNEL
)) {
912 rc
= swap_page(page
);
916 newpage
= lru_to_page(to
);
920 * Pages are properly locked and writeback is complete.
921 * Try to migrate the page.
923 mapping
= page_mapping(page
);
927 if (mapping
->a_ops
->migratepage
) {
929 * Most pages have a mapping and most filesystems
930 * should provide a migration function. Anonymous
931 * pages are part of swap space which also has its
932 * own migration function. This is the most common
933 * path for page migration.
935 rc
= mapping
->a_ops
->migratepage(newpage
, page
);
940 * Default handling if a filesystem does not provide
941 * a migration function. We can only migrate clean
942 * pages so try to write out any dirty pages first.
944 if (PageDirty(page
)) {
945 switch (pageout(page
, mapping
)) {
951 unlock_page(newpage
);
955 ; /* try to migrate the page below */
960 * Buffers are managed in a filesystem specific way.
961 * We must have no buffers or drop them.
963 if (!page_has_buffers(page
) ||
964 try_to_release_page(page
, GFP_KERNEL
)) {
965 rc
= migrate_page(newpage
, page
);
970 * On early passes with mapped pages simply
971 * retry. There may be a lock held for some
972 * buffers that may go away. Later
977 * Persistently unable to drop buffers..... As a
978 * measure of last resort we fall back to
981 unlock_page(newpage
);
983 rc
= swap_page(page
);
988 unlock_page(newpage
);
997 /* Permanent failure */
998 list_move(&page
->lru
, failed
);
1002 /* Successful migration. Return page to LRU */
1003 move_to_lru(newpage
);
1005 list_move(&page
->lru
, moved
);
1008 if (retry
&& pass
++ < 10)
1012 current
->flags
&= ~PF_SWAPWRITE
;
1014 return nr_failed
+ retry
;
1018 * Isolate one page from the LRU lists and put it on the
1019 * indicated list with elevated refcount.
1022 * 0 = page not on LRU list
1023 * 1 = page removed from LRU list and added to the specified list.
1025 int isolate_lru_page(struct page
*page
)
1029 if (PageLRU(page
)) {
1030 struct zone
*zone
= page_zone(page
);
1031 spin_lock_irq(&zone
->lru_lock
);
1032 if (PageLRU(page
)) {
1036 if (PageActive(page
))
1037 del_page_from_active_list(zone
, page
);
1039 del_page_from_inactive_list(zone
, page
);
1041 spin_unlock_irq(&zone
->lru_lock
);
1049 * zone->lru_lock is heavily contended. Some of the functions that
1050 * shrink the lists perform better by taking out a batch of pages
1051 * and working on them outside the LRU lock.
1053 * For pagecache intensive workloads, this function is the hottest
1054 * spot in the kernel (apart from copy_*_user functions).
1056 * Appropriate locks must be held before calling this function.
1058 * @nr_to_scan: The number of pages to look through on the list.
1059 * @src: The LRU list to pull pages off.
1060 * @dst: The temp list to put pages on to.
1061 * @scanned: The number of pages that were scanned.
1063 * returns how many pages were moved onto *@dst.
1065 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1066 struct list_head
*src
, struct list_head
*dst
,
1067 unsigned long *scanned
)
1069 unsigned long nr_taken
= 0;
1073 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1074 struct list_head
*target
;
1075 page
= lru_to_page(src
);
1076 prefetchw_prev_lru_page(page
, src
, flags
);
1078 BUG_ON(!PageLRU(page
));
1080 list_del(&page
->lru
);
1082 if (likely(get_page_unless_zero(page
))) {
1084 * Be careful not to clear PageLRU until after we're
1085 * sure the page is not being freed elsewhere -- the
1086 * page release code relies on it.
1091 } /* else it is being freed elsewhere */
1093 list_add(&page
->lru
, target
);
1101 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1102 * of reclaimed pages
1104 static unsigned long shrink_inactive_list(unsigned long max_scan
,
1105 struct zone
*zone
, struct scan_control
*sc
)
1107 LIST_HEAD(page_list
);
1108 struct pagevec pvec
;
1109 unsigned long nr_scanned
= 0;
1110 unsigned long nr_reclaimed
= 0;
1112 pagevec_init(&pvec
, 1);
1115 spin_lock_irq(&zone
->lru_lock
);
1118 unsigned long nr_taken
;
1119 unsigned long nr_scan
;
1120 unsigned long nr_freed
;
1122 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
1123 &zone
->inactive_list
,
1124 &page_list
, &nr_scan
);
1125 zone
->nr_inactive
-= nr_taken
;
1126 zone
->pages_scanned
+= nr_scan
;
1127 spin_unlock_irq(&zone
->lru_lock
);
1129 nr_scanned
+= nr_scan
;
1130 nr_freed
= shrink_page_list(&page_list
, sc
);
1131 nr_reclaimed
+= nr_freed
;
1132 local_irq_disable();
1133 if (current_is_kswapd()) {
1134 __mod_page_state_zone(zone
, pgscan_kswapd
, nr_scan
);
1135 __mod_page_state(kswapd_steal
, nr_freed
);
1137 __mod_page_state_zone(zone
, pgscan_direct
, nr_scan
);
1138 __mod_page_state_zone(zone
, pgsteal
, nr_freed
);
1143 spin_lock(&zone
->lru_lock
);
1145 * Put back any unfreeable pages.
1147 while (!list_empty(&page_list
)) {
1148 page
= lru_to_page(&page_list
);
1149 BUG_ON(PageLRU(page
));
1151 list_del(&page
->lru
);
1152 if (PageActive(page
))
1153 add_page_to_active_list(zone
, page
);
1155 add_page_to_inactive_list(zone
, page
);
1156 if (!pagevec_add(&pvec
, page
)) {
1157 spin_unlock_irq(&zone
->lru_lock
);
1158 __pagevec_release(&pvec
);
1159 spin_lock_irq(&zone
->lru_lock
);
1162 } while (nr_scanned
< max_scan
);
1163 spin_unlock(&zone
->lru_lock
);
1166 pagevec_release(&pvec
);
1167 return nr_reclaimed
;
1171 * This moves pages from the active list to the inactive list.
1173 * We move them the other way if the page is referenced by one or more
1174 * processes, from rmap.
1176 * If the pages are mostly unmapped, the processing is fast and it is
1177 * appropriate to hold zone->lru_lock across the whole operation. But if
1178 * the pages are mapped, the processing is slow (page_referenced()) so we
1179 * should drop zone->lru_lock around each page. It's impossible to balance
1180 * this, so instead we remove the pages from the LRU while processing them.
1181 * It is safe to rely on PG_active against the non-LRU pages in here because
1182 * nobody will play with that bit on a non-LRU page.
1184 * The downside is that we have to touch page->_count against each page.
1185 * But we had to alter page->flags anyway.
1187 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1188 struct scan_control
*sc
)
1190 unsigned long pgmoved
;
1191 int pgdeactivate
= 0;
1192 unsigned long pgscanned
;
1193 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1194 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
1195 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
1197 struct pagevec pvec
;
1198 int reclaim_mapped
= 0;
1206 * `distress' is a measure of how much trouble we're having
1207 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1209 distress
= 100 >> zone
->prev_priority
;
1212 * The point of this algorithm is to decide when to start
1213 * reclaiming mapped memory instead of just pagecache. Work out
1217 mapped_ratio
= (sc
->nr_mapped
* 100) / total_memory
;
1220 * Now decide how much we really want to unmap some pages. The
1221 * mapped ratio is downgraded - just because there's a lot of
1222 * mapped memory doesn't necessarily mean that page reclaim
1225 * The distress ratio is important - we don't want to start
1228 * A 100% value of vm_swappiness overrides this algorithm
1231 swap_tendency
= mapped_ratio
/ 2 + distress
+ vm_swappiness
;
1234 * Now use this metric to decide whether to start moving mapped
1235 * memory onto the inactive list.
1237 if (swap_tendency
>= 100)
1242 spin_lock_irq(&zone
->lru_lock
);
1243 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
1244 &l_hold
, &pgscanned
);
1245 zone
->pages_scanned
+= pgscanned
;
1246 zone
->nr_active
-= pgmoved
;
1247 spin_unlock_irq(&zone
->lru_lock
);
1249 while (!list_empty(&l_hold
)) {
1251 page
= lru_to_page(&l_hold
);
1252 list_del(&page
->lru
);
1253 if (page_mapped(page
)) {
1254 if (!reclaim_mapped
||
1255 (total_swap_pages
== 0 && PageAnon(page
)) ||
1256 page_referenced(page
, 0)) {
1257 list_add(&page
->lru
, &l_active
);
1261 list_add(&page
->lru
, &l_inactive
);
1264 pagevec_init(&pvec
, 1);
1266 spin_lock_irq(&zone
->lru_lock
);
1267 while (!list_empty(&l_inactive
)) {
1268 page
= lru_to_page(&l_inactive
);
1269 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
1270 BUG_ON(PageLRU(page
));
1272 BUG_ON(!PageActive(page
));
1273 ClearPageActive(page
);
1275 list_move(&page
->lru
, &zone
->inactive_list
);
1277 if (!pagevec_add(&pvec
, page
)) {
1278 zone
->nr_inactive
+= pgmoved
;
1279 spin_unlock_irq(&zone
->lru_lock
);
1280 pgdeactivate
+= pgmoved
;
1282 if (buffer_heads_over_limit
)
1283 pagevec_strip(&pvec
);
1284 __pagevec_release(&pvec
);
1285 spin_lock_irq(&zone
->lru_lock
);
1288 zone
->nr_inactive
+= pgmoved
;
1289 pgdeactivate
+= pgmoved
;
1290 if (buffer_heads_over_limit
) {
1291 spin_unlock_irq(&zone
->lru_lock
);
1292 pagevec_strip(&pvec
);
1293 spin_lock_irq(&zone
->lru_lock
);
1297 while (!list_empty(&l_active
)) {
1298 page
= lru_to_page(&l_active
);
1299 prefetchw_prev_lru_page(page
, &l_active
, flags
);
1300 BUG_ON(PageLRU(page
));
1302 BUG_ON(!PageActive(page
));
1303 list_move(&page
->lru
, &zone
->active_list
);
1305 if (!pagevec_add(&pvec
, page
)) {
1306 zone
->nr_active
+= pgmoved
;
1308 spin_unlock_irq(&zone
->lru_lock
);
1309 __pagevec_release(&pvec
);
1310 spin_lock_irq(&zone
->lru_lock
);
1313 zone
->nr_active
+= pgmoved
;
1314 spin_unlock(&zone
->lru_lock
);
1316 __mod_page_state_zone(zone
, pgrefill
, pgscanned
);
1317 __mod_page_state(pgdeactivate
, pgdeactivate
);
1320 pagevec_release(&pvec
);
1324 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1326 static unsigned long shrink_zone(int priority
, struct zone
*zone
,
1327 struct scan_control
*sc
)
1329 unsigned long nr_active
;
1330 unsigned long nr_inactive
;
1331 unsigned long nr_to_scan
;
1332 unsigned long nr_reclaimed
= 0;
1334 atomic_inc(&zone
->reclaim_in_progress
);
1337 * Add one to `nr_to_scan' just to make sure that the kernel will
1338 * slowly sift through the active list.
1340 zone
->nr_scan_active
+= (zone
->nr_active
>> priority
) + 1;
1341 nr_active
= zone
->nr_scan_active
;
1342 if (nr_active
>= sc
->swap_cluster_max
)
1343 zone
->nr_scan_active
= 0;
1347 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> priority
) + 1;
1348 nr_inactive
= zone
->nr_scan_inactive
;
1349 if (nr_inactive
>= sc
->swap_cluster_max
)
1350 zone
->nr_scan_inactive
= 0;
1354 while (nr_active
|| nr_inactive
) {
1356 nr_to_scan
= min(nr_active
,
1357 (unsigned long)sc
->swap_cluster_max
);
1358 nr_active
-= nr_to_scan
;
1359 shrink_active_list(nr_to_scan
, zone
, sc
);
1363 nr_to_scan
= min(nr_inactive
,
1364 (unsigned long)sc
->swap_cluster_max
);
1365 nr_inactive
-= nr_to_scan
;
1366 nr_reclaimed
+= shrink_inactive_list(nr_to_scan
, zone
,
1371 throttle_vm_writeout();
1373 atomic_dec(&zone
->reclaim_in_progress
);
1374 return nr_reclaimed
;
1378 * This is the direct reclaim path, for page-allocating processes. We only
1379 * try to reclaim pages from zones which will satisfy the caller's allocation
1382 * We reclaim from a zone even if that zone is over pages_high. Because:
1383 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1385 * b) The zones may be over pages_high but they must go *over* pages_high to
1386 * satisfy the `incremental min' zone defense algorithm.
1388 * Returns the number of reclaimed pages.
1390 * If a zone is deemed to be full of pinned pages then just give it a light
1391 * scan then give up on it.
1393 static unsigned long shrink_zones(int priority
, struct zone
**zones
,
1394 struct scan_control
*sc
)
1396 unsigned long nr_reclaimed
= 0;
1399 for (i
= 0; zones
[i
] != NULL
; i
++) {
1400 struct zone
*zone
= zones
[i
];
1402 if (!populated_zone(zone
))
1405 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1408 zone
->temp_priority
= priority
;
1409 if (zone
->prev_priority
> priority
)
1410 zone
->prev_priority
= priority
;
1412 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1413 continue; /* Let kswapd poll it */
1415 nr_reclaimed
+= shrink_zone(priority
, zone
, sc
);
1417 return nr_reclaimed
;
1421 * This is the main entry point to direct page reclaim.
1423 * If a full scan of the inactive list fails to free enough memory then we
1424 * are "out of memory" and something needs to be killed.
1426 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1427 * high - the zone may be full of dirty or under-writeback pages, which this
1428 * caller can't do much about. We kick pdflush and take explicit naps in the
1429 * hope that some of these pages can be written. But if the allocating task
1430 * holds filesystem locks which prevent writeout this might not work, and the
1431 * allocation attempt will fail.
1433 unsigned long try_to_free_pages(struct zone
**zones
, gfp_t gfp_mask
)
1437 unsigned long total_scanned
= 0;
1438 unsigned long nr_reclaimed
= 0;
1439 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1440 unsigned long lru_pages
= 0;
1442 struct scan_control sc
= {
1443 .gfp_mask
= gfp_mask
,
1444 .may_writepage
= !laptop_mode
,
1445 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1449 inc_page_state(allocstall
);
1451 for (i
= 0; zones
[i
] != NULL
; i
++) {
1452 struct zone
*zone
= zones
[i
];
1454 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1457 zone
->temp_priority
= DEF_PRIORITY
;
1458 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1461 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1462 sc
.nr_mapped
= read_page_state(nr_mapped
);
1465 disable_swap_token();
1466 nr_reclaimed
+= shrink_zones(priority
, zones
, &sc
);
1467 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1468 if (reclaim_state
) {
1469 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1470 reclaim_state
->reclaimed_slab
= 0;
1472 total_scanned
+= sc
.nr_scanned
;
1473 if (nr_reclaimed
>= sc
.swap_cluster_max
) {
1479 * Try to write back as many pages as we just scanned. This
1480 * tends to cause slow streaming writers to write data to the
1481 * disk smoothly, at the dirtying rate, which is nice. But
1482 * that's undesirable in laptop mode, where we *want* lumpy
1483 * writeout. So in laptop mode, write out the whole world.
1485 if (total_scanned
> sc
.swap_cluster_max
+
1486 sc
.swap_cluster_max
/ 2) {
1487 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1488 sc
.may_writepage
= 1;
1491 /* Take a nap, wait for some writeback to complete */
1492 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1493 blk_congestion_wait(WRITE
, HZ
/10);
1496 for (i
= 0; zones
[i
] != 0; i
++) {
1497 struct zone
*zone
= zones
[i
];
1499 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1502 zone
->prev_priority
= zone
->temp_priority
;
1508 * For kswapd, balance_pgdat() will work across all this node's zones until
1509 * they are all at pages_high.
1511 * If `nr_pages' is non-zero then it is the number of pages which are to be
1512 * reclaimed, regardless of the zone occupancies. This is a software suspend
1515 * Returns the number of pages which were actually freed.
1517 * There is special handling here for zones which are full of pinned pages.
1518 * This can happen if the pages are all mlocked, or if they are all used by
1519 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1520 * What we do is to detect the case where all pages in the zone have been
1521 * scanned twice and there has been zero successful reclaim. Mark the zone as
1522 * dead and from now on, only perform a short scan. Basically we're polling
1523 * the zone for when the problem goes away.
1525 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1526 * zones which have free_pages > pages_high, but once a zone is found to have
1527 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1528 * of the number of free pages in the lower zones. This interoperates with
1529 * the page allocator fallback scheme to ensure that aging of pages is balanced
1532 static unsigned long balance_pgdat(pg_data_t
*pgdat
, unsigned long nr_pages
,
1535 unsigned long to_free
= nr_pages
;
1539 unsigned long total_scanned
;
1540 unsigned long nr_reclaimed
;
1541 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1542 struct scan_control sc
= {
1543 .gfp_mask
= GFP_KERNEL
,
1545 .swap_cluster_max
= nr_pages
? nr_pages
: SWAP_CLUSTER_MAX
,
1551 sc
.may_writepage
= !laptop_mode
,
1552 sc
.nr_mapped
= read_page_state(nr_mapped
);
1554 inc_page_state(pageoutrun
);
1556 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1557 struct zone
*zone
= pgdat
->node_zones
+ i
;
1559 zone
->temp_priority
= DEF_PRIORITY
;
1562 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1563 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1564 unsigned long lru_pages
= 0;
1566 /* The swap token gets in the way of swapout... */
1568 disable_swap_token();
1572 if (nr_pages
== 0) {
1574 * Scan in the highmem->dma direction for the highest
1575 * zone which needs scanning
1577 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1578 struct zone
*zone
= pgdat
->node_zones
+ i
;
1580 if (!populated_zone(zone
))
1583 if (zone
->all_unreclaimable
&&
1584 priority
!= DEF_PRIORITY
)
1587 if (!zone_watermark_ok(zone
, order
,
1588 zone
->pages_high
, 0, 0)) {
1595 end_zone
= pgdat
->nr_zones
- 1;
1598 for (i
= 0; i
<= end_zone
; i
++) {
1599 struct zone
*zone
= pgdat
->node_zones
+ i
;
1601 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1605 * Now scan the zone in the dma->highmem direction, stopping
1606 * at the last zone which needs scanning.
1608 * We do this because the page allocator works in the opposite
1609 * direction. This prevents the page allocator from allocating
1610 * pages behind kswapd's direction of progress, which would
1611 * cause too much scanning of the lower zones.
1613 for (i
= 0; i
<= end_zone
; i
++) {
1614 struct zone
*zone
= pgdat
->node_zones
+ i
;
1617 if (!populated_zone(zone
))
1620 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1623 if (nr_pages
== 0) { /* Not software suspend */
1624 if (!zone_watermark_ok(zone
, order
,
1625 zone
->pages_high
, end_zone
, 0))
1628 zone
->temp_priority
= priority
;
1629 if (zone
->prev_priority
> priority
)
1630 zone
->prev_priority
= priority
;
1632 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1633 reclaim_state
->reclaimed_slab
= 0;
1634 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1636 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1637 total_scanned
+= sc
.nr_scanned
;
1638 if (zone
->all_unreclaimable
)
1640 if (nr_slab
== 0 && zone
->pages_scanned
>=
1641 (zone
->nr_active
+ zone
->nr_inactive
) * 4)
1642 zone
->all_unreclaimable
= 1;
1644 * If we've done a decent amount of scanning and
1645 * the reclaim ratio is low, start doing writepage
1646 * even in laptop mode
1648 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1649 total_scanned
> nr_reclaimed
+ nr_reclaimed
/ 2)
1650 sc
.may_writepage
= 1;
1652 if (nr_pages
&& to_free
> nr_reclaimed
)
1653 continue; /* swsusp: need to do more work */
1655 break; /* kswapd: all done */
1657 * OK, kswapd is getting into trouble. Take a nap, then take
1658 * another pass across the zones.
1660 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1661 blk_congestion_wait(WRITE
, HZ
/10);
1664 * We do this so kswapd doesn't build up large priorities for
1665 * example when it is freeing in parallel with allocators. It
1666 * matches the direct reclaim path behaviour in terms of impact
1667 * on zone->*_priority.
1669 if ((nr_reclaimed
>= SWAP_CLUSTER_MAX
) && !nr_pages
)
1673 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1674 struct zone
*zone
= pgdat
->node_zones
+ i
;
1676 zone
->prev_priority
= zone
->temp_priority
;
1678 if (!all_zones_ok
) {
1683 return nr_reclaimed
;
1687 * The background pageout daemon, started as a kernel thread
1688 * from the init process.
1690 * This basically trickles out pages so that we have _some_
1691 * free memory available even if there is no other activity
1692 * that frees anything up. This is needed for things like routing
1693 * etc, where we otherwise might have all activity going on in
1694 * asynchronous contexts that cannot page things out.
1696 * If there are applications that are active memory-allocators
1697 * (most normal use), this basically shouldn't matter.
1699 static int kswapd(void *p
)
1701 unsigned long order
;
1702 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1703 struct task_struct
*tsk
= current
;
1705 struct reclaim_state reclaim_state
= {
1706 .reclaimed_slab
= 0,
1710 daemonize("kswapd%d", pgdat
->node_id
);
1711 cpumask
= node_to_cpumask(pgdat
->node_id
);
1712 if (!cpus_empty(cpumask
))
1713 set_cpus_allowed(tsk
, cpumask
);
1714 current
->reclaim_state
= &reclaim_state
;
1717 * Tell the memory management that we're a "memory allocator",
1718 * and that if we need more memory we should get access to it
1719 * regardless (see "__alloc_pages()"). "kswapd" should
1720 * never get caught in the normal page freeing logic.
1722 * (Kswapd normally doesn't need memory anyway, but sometimes
1723 * you need a small amount of memory in order to be able to
1724 * page out something else, and this flag essentially protects
1725 * us from recursively trying to free more memory as we're
1726 * trying to free the first piece of memory in the first place).
1728 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1732 unsigned long new_order
;
1736 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1737 new_order
= pgdat
->kswapd_max_order
;
1738 pgdat
->kswapd_max_order
= 0;
1739 if (order
< new_order
) {
1741 * Don't sleep if someone wants a larger 'order'
1747 order
= pgdat
->kswapd_max_order
;
1749 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1751 balance_pgdat(pgdat
, 0, order
);
1757 * A zone is low on free memory, so wake its kswapd task to service it.
1759 void wakeup_kswapd(struct zone
*zone
, int order
)
1763 if (!populated_zone(zone
))
1766 pgdat
= zone
->zone_pgdat
;
1767 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1769 if (pgdat
->kswapd_max_order
< order
)
1770 pgdat
->kswapd_max_order
= order
;
1771 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1773 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1775 wake_up_interruptible(&pgdat
->kswapd_wait
);
1780 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1783 unsigned long shrink_all_memory(unsigned long nr_pages
)
1786 unsigned long nr_to_free
= nr_pages
;
1787 unsigned long ret
= 0;
1789 struct reclaim_state reclaim_state
= {
1790 .reclaimed_slab
= 0,
1793 current
->reclaim_state
= &reclaim_state
;
1795 for_each_pgdat(pgdat
) {
1796 unsigned long freed
;
1798 freed
= balance_pgdat(pgdat
, nr_to_free
, 0);
1800 nr_to_free
-= freed
;
1801 if ((long)nr_to_free
<= 0)
1804 if (retry
-- && ret
< nr_pages
) {
1805 blk_congestion_wait(WRITE
, HZ
/5);
1808 current
->reclaim_state
= NULL
;
1813 #ifdef CONFIG_HOTPLUG_CPU
1814 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1815 not required for correctness. So if the last cpu in a node goes
1816 away, we get changed to run anywhere: as the first one comes back,
1817 restore their cpu bindings. */
1818 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
1819 unsigned long action
, void *hcpu
)
1824 if (action
== CPU_ONLINE
) {
1825 for_each_pgdat(pgdat
) {
1826 mask
= node_to_cpumask(pgdat
->node_id
);
1827 if (any_online_cpu(mask
) != NR_CPUS
)
1828 /* One of our CPUs online: restore mask */
1829 set_cpus_allowed(pgdat
->kswapd
, mask
);
1834 #endif /* CONFIG_HOTPLUG_CPU */
1836 static int __init
kswapd_init(void)
1841 for_each_pgdat(pgdat
) {
1844 pid
= kernel_thread(kswapd
, pgdat
, CLONE_KERNEL
);
1846 pgdat
->kswapd
= find_task_by_pid(pid
);
1848 total_memory
= nr_free_pagecache_pages();
1849 hotcpu_notifier(cpu_callback
, 0);
1853 module_init(kswapd_init
)
1859 * If non-zero call zone_reclaim when the number of free pages falls below
1862 * In the future we may add flags to the mode. However, the page allocator
1863 * should only have to check that zone_reclaim_mode != 0 before calling
1866 int zone_reclaim_mode __read_mostly
;
1868 #define RECLAIM_OFF 0
1869 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1870 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1871 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1872 #define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
1875 * Mininum time between zone reclaim scans
1877 int zone_reclaim_interval __read_mostly
= 30*HZ
;
1880 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1881 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1884 #define ZONE_RECLAIM_PRIORITY 4
1887 * Try to free up some pages from this zone through reclaim.
1889 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1891 /* Minimum pages needed in order to stay on node */
1892 const unsigned long nr_pages
= 1 << order
;
1893 struct task_struct
*p
= current
;
1894 struct reclaim_state reclaim_state
;
1896 unsigned long nr_reclaimed
= 0;
1897 struct scan_control sc
= {
1898 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
1899 .may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
1900 .nr_mapped
= read_page_state(nr_mapped
),
1901 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
1903 .gfp_mask
= gfp_mask
,
1906 disable_swap_token();
1909 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1910 * and we also need to be able to write out pages for RECLAIM_WRITE
1913 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
1914 reclaim_state
.reclaimed_slab
= 0;
1915 p
->reclaim_state
= &reclaim_state
;
1918 * Free memory by calling shrink zone with increasing priorities
1919 * until we have enough memory freed.
1921 priority
= ZONE_RECLAIM_PRIORITY
;
1923 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1925 } while (priority
>= 0 && nr_reclaimed
< nr_pages
);
1927 if (nr_reclaimed
< nr_pages
&& (zone_reclaim_mode
& RECLAIM_SLAB
)) {
1929 * shrink_slab() does not currently allow us to determine how
1930 * many pages were freed in this zone. So we just shake the slab
1931 * a bit and then go off node for this particular allocation
1932 * despite possibly having freed enough memory to allocate in
1933 * this zone. If we freed local memory then the next
1934 * allocations will be local again.
1936 * shrink_slab will free memory on all zones and may take
1939 shrink_slab(sc
.nr_scanned
, gfp_mask
, order
);
1942 p
->reclaim_state
= NULL
;
1943 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
1945 if (nr_reclaimed
== 0) {
1947 * We were unable to reclaim enough pages to stay on node. We
1948 * now allow off node accesses for a certain time period before
1949 * trying again to reclaim pages from the local zone.
1951 zone
->last_unsuccessful_zone_reclaim
= jiffies
;
1954 return nr_reclaimed
>= nr_pages
;
1957 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1963 * Do not reclaim if there was a recent unsuccessful attempt at zone
1964 * reclaim. In that case we let allocations go off node for the
1965 * zone_reclaim_interval. Otherwise we would scan for each off-node
1968 if (time_before(jiffies
,
1969 zone
->last_unsuccessful_zone_reclaim
+ zone_reclaim_interval
))
1973 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1974 * not have reclaimable pages and if we should not delay the allocation
1977 if (!(gfp_mask
& __GFP_WAIT
) ||
1978 zone
->all_unreclaimable
||
1979 atomic_read(&zone
->reclaim_in_progress
) > 0 ||
1980 (current
->flags
& PF_MEMALLOC
))
1984 * Only run zone reclaim on the local zone or on zones that do not
1985 * have associated processors. This will favor the local processor
1986 * over remote processors and spread off node memory allocations
1987 * as wide as possible.
1989 node_id
= zone
->zone_pgdat
->node_id
;
1990 mask
= node_to_cpumask(node_id
);
1991 if (!cpus_empty(mask
) && node_id
!= numa_node_id())
1993 return __zone_reclaim(zone
, gfp_mask
, order
);