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>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
42 /* possible outcome of pageout() */
44 /* failed to write page out, page is locked */
46 /* move page to the active list, page is locked */
48 /* page has been sent to the disk successfully, page is unlocked */
50 /* page is clean and locked */
55 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
56 unsigned long nr_to_scan
;
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned
;
61 /* Incremented by the number of pages reclaimed */
62 unsigned long nr_reclaimed
;
64 unsigned long nr_mapped
; /* From page_state */
66 /* Ask shrink_caches, or shrink_zone to scan at this priority */
67 unsigned int priority
;
69 /* This context's GFP mask */
74 /* Can pages be swapped as part of reclaim? */
77 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
78 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
79 * In this context, it doesn't matter that we scan the
80 * whole list at once. */
85 * The list of shrinker callbacks used by to apply pressure to
90 struct list_head list
;
91 int seeks
; /* seeks to recreate an obj */
92 long nr
; /* objs pending delete */
95 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
97 #ifdef ARCH_HAS_PREFETCH
98 #define prefetch_prev_lru_page(_page, _base, _field) \
100 if ((_page)->lru.prev != _base) { \
103 prev = lru_to_page(&(_page->lru)); \
104 prefetch(&prev->_field); \
108 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #ifdef ARCH_HAS_PREFETCHW
112 #define prefetchw_prev_lru_page(_page, _base, _field) \
114 if ((_page)->lru.prev != _base) { \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetchw(&prev->_field); \
122 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
126 * From 0 .. 100. Higher means more swappy.
128 int vm_swappiness
= 60;
129 static long total_memory
;
131 static LIST_HEAD(shrinker_list
);
132 static DECLARE_RWSEM(shrinker_rwsem
);
135 * Add a shrinker callback to be called from the vm
137 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
139 struct shrinker
*shrinker
;
141 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
143 shrinker
->shrinker
= theshrinker
;
144 shrinker
->seeks
= seeks
;
146 down_write(&shrinker_rwsem
);
147 list_add_tail(&shrinker
->list
, &shrinker_list
);
148 up_write(&shrinker_rwsem
);
152 EXPORT_SYMBOL(set_shrinker
);
157 void remove_shrinker(struct shrinker
*shrinker
)
159 down_write(&shrinker_rwsem
);
160 list_del(&shrinker
->list
);
161 up_write(&shrinker_rwsem
);
164 EXPORT_SYMBOL(remove_shrinker
);
166 #define SHRINK_BATCH 128
168 * Call the shrink functions to age shrinkable caches
170 * Here we assume it costs one seek to replace a lru page and that it also
171 * takes a seek to recreate a cache object. With this in mind we age equal
172 * percentages of the lru and ageable caches. This should balance the seeks
173 * generated by these structures.
175 * If the vm encounted mapped pages on the LRU it increase the pressure on
176 * slab to avoid swapping.
178 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
180 * `lru_pages' represents the number of on-LRU pages in all the zones which
181 * are eligible for the caller's allocation attempt. It is used for balancing
182 * slab reclaim versus page reclaim.
184 * Returns the number of slab objects which we shrunk.
186 int shrink_slab(unsigned long scanned
, gfp_t gfp_mask
, unsigned long lru_pages
)
188 struct shrinker
*shrinker
;
192 scanned
= SWAP_CLUSTER_MAX
;
194 if (!down_read_trylock(&shrinker_rwsem
))
195 return 1; /* Assume we'll be able to shrink next time */
197 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
198 unsigned long long delta
;
199 unsigned long total_scan
;
200 unsigned long max_pass
= (*shrinker
->shrinker
)(0, gfp_mask
);
202 delta
= (4 * scanned
) / shrinker
->seeks
;
204 do_div(delta
, lru_pages
+ 1);
205 shrinker
->nr
+= delta
;
206 if (shrinker
->nr
< 0) {
207 printk(KERN_ERR
"%s: nr=%ld\n",
208 __FUNCTION__
, shrinker
->nr
);
209 shrinker
->nr
= max_pass
;
213 * Avoid risking looping forever due to too large nr value:
214 * never try to free more than twice the estimate number of
217 if (shrinker
->nr
> max_pass
* 2)
218 shrinker
->nr
= max_pass
* 2;
220 total_scan
= shrinker
->nr
;
223 while (total_scan
>= SHRINK_BATCH
) {
224 long this_scan
= SHRINK_BATCH
;
228 nr_before
= (*shrinker
->shrinker
)(0, gfp_mask
);
229 shrink_ret
= (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
230 if (shrink_ret
== -1)
232 if (shrink_ret
< nr_before
)
233 ret
+= nr_before
- shrink_ret
;
234 mod_page_state(slabs_scanned
, this_scan
);
235 total_scan
-= this_scan
;
240 shrinker
->nr
+= total_scan
;
242 up_read(&shrinker_rwsem
);
246 /* Called without lock on whether page is mapped, so answer is unstable */
247 static inline int page_mapping_inuse(struct page
*page
)
249 struct address_space
*mapping
;
251 /* Page is in somebody's page tables. */
252 if (page_mapped(page
))
255 /* Be more reluctant to reclaim swapcache than pagecache */
256 if (PageSwapCache(page
))
259 mapping
= page_mapping(page
);
263 /* File is mmap'd by somebody? */
264 return mapping_mapped(mapping
);
267 static inline int is_page_cache_freeable(struct page
*page
)
269 return page_count(page
) - !!PagePrivate(page
) == 2;
272 static int may_write_to_queue(struct backing_dev_info
*bdi
)
274 if (current
->flags
& PF_SWAPWRITE
)
276 if (!bdi_write_congested(bdi
))
278 if (bdi
== current
->backing_dev_info
)
284 * We detected a synchronous write error writing a page out. Probably
285 * -ENOSPC. We need to propagate that into the address_space for a subsequent
286 * fsync(), msync() or close().
288 * The tricky part is that after writepage we cannot touch the mapping: nothing
289 * prevents it from being freed up. But we have a ref on the page and once
290 * that page is locked, the mapping is pinned.
292 * We're allowed to run sleeping lock_page() here because we know the caller has
295 static void handle_write_error(struct address_space
*mapping
,
296 struct page
*page
, int error
)
299 if (page_mapping(page
) == mapping
) {
300 if (error
== -ENOSPC
)
301 set_bit(AS_ENOSPC
, &mapping
->flags
);
303 set_bit(AS_EIO
, &mapping
->flags
);
309 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
311 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
314 * If the page is dirty, only perform writeback if that write
315 * will be non-blocking. To prevent this allocation from being
316 * stalled by pagecache activity. But note that there may be
317 * stalls if we need to run get_block(). We could test
318 * PagePrivate for that.
320 * If this process is currently in generic_file_write() against
321 * this page's queue, we can perform writeback even if that
324 * If the page is swapcache, write it back even if that would
325 * block, for some throttling. This happens by accident, because
326 * swap_backing_dev_info is bust: it doesn't reflect the
327 * congestion state of the swapdevs. Easy to fix, if needed.
328 * See swapfile.c:page_queue_congested().
330 if (!is_page_cache_freeable(page
))
334 * Some data journaling orphaned pages can have
335 * page->mapping == NULL while being dirty with clean buffers.
337 if (PagePrivate(page
)) {
338 if (try_to_free_buffers(page
)) {
339 ClearPageDirty(page
);
340 printk("%s: orphaned page\n", __FUNCTION__
);
346 if (mapping
->a_ops
->writepage
== NULL
)
347 return PAGE_ACTIVATE
;
348 if (!may_write_to_queue(mapping
->backing_dev_info
))
351 if (clear_page_dirty_for_io(page
)) {
353 struct writeback_control wbc
= {
354 .sync_mode
= WB_SYNC_NONE
,
355 .nr_to_write
= SWAP_CLUSTER_MAX
,
360 SetPageReclaim(page
);
361 res
= mapping
->a_ops
->writepage(page
, &wbc
);
363 handle_write_error(mapping
, page
, res
);
364 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
365 ClearPageReclaim(page
);
366 return PAGE_ACTIVATE
;
368 if (!PageWriteback(page
)) {
369 /* synchronous write or broken a_ops? */
370 ClearPageReclaim(page
);
379 static int remove_mapping(struct address_space
*mapping
, struct page
*page
)
382 return 0; /* truncate got there first */
384 write_lock_irq(&mapping
->tree_lock
);
387 * The non-racy check for busy page. It is critical to check
388 * PageDirty _after_ making sure that the page is freeable and
389 * not in use by anybody. (pagecache + us == 2)
391 if (unlikely(page_count(page
) != 2))
394 if (unlikely(PageDirty(page
)))
397 if (PageSwapCache(page
)) {
398 swp_entry_t swap
= { .val
= page_private(page
) };
399 __delete_from_swap_cache(page
);
400 write_unlock_irq(&mapping
->tree_lock
);
402 __put_page(page
); /* The pagecache ref */
406 __remove_from_page_cache(page
);
407 write_unlock_irq(&mapping
->tree_lock
);
412 write_unlock_irq(&mapping
->tree_lock
);
417 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
419 static int shrink_list(struct list_head
*page_list
, struct scan_control
*sc
)
421 LIST_HEAD(ret_pages
);
422 struct pagevec freed_pvec
;
428 pagevec_init(&freed_pvec
, 1);
429 while (!list_empty(page_list
)) {
430 struct address_space
*mapping
;
437 page
= lru_to_page(page_list
);
438 list_del(&page
->lru
);
440 if (TestSetPageLocked(page
))
443 BUG_ON(PageActive(page
));
446 /* Double the slab pressure for mapped and swapcache pages */
447 if (page_mapped(page
) || PageSwapCache(page
))
450 if (PageWriteback(page
))
453 referenced
= page_referenced(page
, 1);
454 /* In active use or really unfreeable? Activate it. */
455 if (referenced
&& page_mapping_inuse(page
))
456 goto activate_locked
;
460 * Anonymous process memory has backing store?
461 * Try to allocate it some swap space here.
463 if (PageAnon(page
) && !PageSwapCache(page
)) {
466 if (!add_to_swap(page
, GFP_ATOMIC
))
467 goto activate_locked
;
469 #endif /* CONFIG_SWAP */
471 mapping
= page_mapping(page
);
472 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
473 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
476 * The page is mapped into the page tables of one or more
477 * processes. Try to unmap it here.
479 if (page_mapped(page
) && mapping
) {
481 * No unmapping if we do not swap
486 switch (try_to_unmap(page
, 0)) {
488 goto activate_locked
;
492 ; /* try to free the page below */
496 if (PageDirty(page
)) {
501 if (!sc
->may_writepage
)
504 /* Page is dirty, try to write it out here */
505 switch(pageout(page
, mapping
)) {
509 goto activate_locked
;
511 if (PageWriteback(page
) || PageDirty(page
))
514 * A synchronous write - probably a ramdisk. Go
515 * ahead and try to reclaim the page.
517 if (TestSetPageLocked(page
))
519 if (PageDirty(page
) || PageWriteback(page
))
521 mapping
= page_mapping(page
);
523 ; /* try to free the page below */
528 * If the page has buffers, try to free the buffer mappings
529 * associated with this page. If we succeed we try to free
532 * We do this even if the page is PageDirty().
533 * try_to_release_page() does not perform I/O, but it is
534 * possible for a page to have PageDirty set, but it is actually
535 * clean (all its buffers are clean). This happens if the
536 * buffers were written out directly, with submit_bh(). ext3
537 * will do this, as well as the blockdev mapping.
538 * try_to_release_page() will discover that cleanness and will
539 * drop the buffers and mark the page clean - it can be freed.
541 * Rarely, pages can have buffers and no ->mapping. These are
542 * the pages which were not successfully invalidated in
543 * truncate_complete_page(). We try to drop those buffers here
544 * and if that worked, and the page is no longer mapped into
545 * process address space (page_count == 1) it can be freed.
546 * Otherwise, leave the page on the LRU so it is swappable.
548 if (PagePrivate(page
)) {
549 if (!try_to_release_page(page
, sc
->gfp_mask
))
550 goto activate_locked
;
551 if (!mapping
&& page_count(page
) == 1)
555 if (!remove_mapping(mapping
, page
))
561 if (!pagevec_add(&freed_pvec
, page
))
562 __pagevec_release_nonlru(&freed_pvec
);
571 list_add(&page
->lru
, &ret_pages
);
572 BUG_ON(PageLRU(page
));
574 list_splice(&ret_pages
, page_list
);
575 if (pagevec_count(&freed_pvec
))
576 __pagevec_release_nonlru(&freed_pvec
);
577 mod_page_state(pgactivate
, pgactivate
);
578 sc
->nr_reclaimed
+= reclaimed
;
582 #ifdef CONFIG_MIGRATION
583 static inline void move_to_lru(struct page
*page
)
585 list_del(&page
->lru
);
586 if (PageActive(page
)) {
588 * lru_cache_add_active checks that
589 * the PG_active bit is off.
591 ClearPageActive(page
);
592 lru_cache_add_active(page
);
600 * Add isolated pages on the list back to the LRU.
602 * returns the number of pages put back.
604 int putback_lru_pages(struct list_head
*l
)
610 list_for_each_entry_safe(page
, page2
, l
, lru
) {
618 * Non migratable page
620 int fail_migrate_page(struct page
*newpage
, struct page
*page
)
624 EXPORT_SYMBOL(fail_migrate_page
);
627 * swapout a single page
628 * page is locked upon entry, unlocked on exit
630 static int swap_page(struct page
*page
)
632 struct address_space
*mapping
= page_mapping(page
);
634 if (page_mapped(page
) && mapping
)
635 if (try_to_unmap(page
, 0) != SWAP_SUCCESS
)
638 if (PageDirty(page
)) {
639 /* Page is dirty, try to write it out here */
640 switch(pageout(page
, mapping
)) {
649 ; /* try to free the page below */
653 if (PagePrivate(page
)) {
654 if (!try_to_release_page(page
, GFP_KERNEL
) ||
655 (!mapping
&& page_count(page
) == 1))
659 if (remove_mapping(mapping
, page
)) {
671 EXPORT_SYMBOL(swap_page
);
674 * Page migration was first developed in the context of the memory hotplug
675 * project. The main authors of the migration code are:
677 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
678 * Hirokazu Takahashi <taka@valinux.co.jp>
679 * Dave Hansen <haveblue@us.ibm.com>
680 * Christoph Lameter <clameter@sgi.com>
684 * Remove references for a page and establish the new page with the correct
685 * basic settings to be able to stop accesses to the page.
687 int migrate_page_remove_references(struct page
*newpage
,
688 struct page
*page
, int nr_refs
)
690 struct address_space
*mapping
= page_mapping(page
);
691 struct page
**radix_pointer
;
694 * Avoid doing any of the following work if the page count
695 * indicates that the page is in use or truncate has removed
698 if (!mapping
|| page_mapcount(page
) + nr_refs
!= page_count(page
))
702 * Establish swap ptes for anonymous pages or destroy pte
705 * In order to reestablish file backed mappings the fault handlers
706 * will take the radix tree_lock which may then be used to stop
707 * processses from accessing this page until the new page is ready.
709 * A process accessing via a swap pte (an anonymous page) will take a
710 * page_lock on the old page which will block the process until the
711 * migration attempt is complete. At that time the PageSwapCache bit
712 * will be examined. If the page was migrated then the PageSwapCache
713 * bit will be clear and the operation to retrieve the page will be
714 * retried which will find the new page in the radix tree. Then a new
715 * direct mapping may be generated based on the radix tree contents.
717 * If the page was not migrated then the PageSwapCache bit
718 * is still set and the operation may continue.
720 try_to_unmap(page
, 1);
723 * Give up if we were unable to remove all mappings.
725 if (page_mapcount(page
))
728 write_lock_irq(&mapping
->tree_lock
);
730 radix_pointer
= (struct page
**)radix_tree_lookup_slot(
734 if (!page_mapping(page
) || page_count(page
) != nr_refs
||
735 *radix_pointer
!= page
) {
736 write_unlock_irq(&mapping
->tree_lock
);
741 * Now we know that no one else is looking at the page.
743 * Certain minimal information about a page must be available
744 * in order for other subsystems to properly handle the page if they
745 * find it through the radix tree update before we are finished
749 newpage
->index
= page
->index
;
750 newpage
->mapping
= page
->mapping
;
751 if (PageSwapCache(page
)) {
752 SetPageSwapCache(newpage
);
753 set_page_private(newpage
, page_private(page
));
756 *radix_pointer
= newpage
;
758 write_unlock_irq(&mapping
->tree_lock
);
762 EXPORT_SYMBOL(migrate_page_remove_references
);
765 * Copy the page to its new location
767 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
769 copy_highpage(newpage
, page
);
772 SetPageError(newpage
);
773 if (PageReferenced(page
))
774 SetPageReferenced(newpage
);
775 if (PageUptodate(page
))
776 SetPageUptodate(newpage
);
777 if (PageActive(page
))
778 SetPageActive(newpage
);
779 if (PageChecked(page
))
780 SetPageChecked(newpage
);
781 if (PageMappedToDisk(page
))
782 SetPageMappedToDisk(newpage
);
784 if (PageDirty(page
)) {
785 clear_page_dirty_for_io(page
);
786 set_page_dirty(newpage
);
789 ClearPageSwapCache(page
);
790 ClearPageActive(page
);
791 ClearPagePrivate(page
);
792 set_page_private(page
, 0);
793 page
->mapping
= NULL
;
796 * If any waiters have accumulated on the new page then
799 if (PageWriteback(newpage
))
800 end_page_writeback(newpage
);
802 EXPORT_SYMBOL(migrate_page_copy
);
805 * Common logic to directly migrate a single page suitable for
806 * pages that do not use PagePrivate.
808 * Pages are locked upon entry and exit.
810 int migrate_page(struct page
*newpage
, struct page
*page
)
812 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
814 if (migrate_page_remove_references(newpage
, page
, 2))
817 migrate_page_copy(newpage
, page
);
820 * Remove auxiliary swap entries and replace
821 * them with real ptes.
823 * Note that a real pte entry will allow processes that are not
824 * waiting on the page lock to use the new page via the page tables
825 * before the new page is unlocked.
827 remove_from_swap(newpage
);
830 EXPORT_SYMBOL(migrate_page
);
835 * Two lists are passed to this function. The first list
836 * contains the pages isolated from the LRU to be migrated.
837 * The second list contains new pages that the pages isolated
838 * can be moved to. If the second list is NULL then all
839 * pages are swapped out.
841 * The function returns after 10 attempts or if no pages
842 * are movable anymore because t has become empty
843 * or no retryable pages exist anymore.
845 * Return: Number of pages not migrated when "to" ran empty.
847 int migrate_pages(struct list_head
*from
, struct list_head
*to
,
848 struct list_head
*moved
, struct list_head
*failed
)
855 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
859 current
->flags
|= PF_SWAPWRITE
;
864 list_for_each_entry_safe(page
, page2
, from
, lru
) {
865 struct page
*newpage
= NULL
;
866 struct address_space
*mapping
;
871 if (page_count(page
) == 1)
872 /* page was freed from under us. So we are done. */
875 if (to
&& list_empty(to
))
879 * Skip locked pages during the first two passes to give the
880 * functions holding the lock time to release the page. Later we
881 * use lock_page() to have a higher chance of acquiring the
888 if (TestSetPageLocked(page
))
892 * Only wait on writeback if we have already done a pass where
893 * we we may have triggered writeouts for lots of pages.
896 wait_on_page_writeback(page
);
898 if (PageWriteback(page
))
903 * Anonymous pages must have swap cache references otherwise
904 * the information contained in the page maps cannot be
907 if (PageAnon(page
) && !PageSwapCache(page
)) {
908 if (!add_to_swap(page
, GFP_KERNEL
)) {
915 rc
= swap_page(page
);
919 newpage
= lru_to_page(to
);
923 * Pages are properly locked and writeback is complete.
924 * Try to migrate the page.
926 mapping
= page_mapping(page
);
930 if (mapping
->a_ops
->migratepage
) {
931 rc
= mapping
->a_ops
->migratepage(newpage
, page
);
936 * Trigger writeout if page is dirty
938 if (PageDirty(page
)) {
939 switch (pageout(page
, mapping
)) {
945 unlock_page(newpage
);
949 ; /* try to migrate the page below */
953 * If we have no buffer or can release the buffer
954 * then do a simple migration.
956 if (!page_has_buffers(page
) ||
957 try_to_release_page(page
, GFP_KERNEL
)) {
958 rc
= migrate_page(newpage
, page
);
963 * On early passes with mapped pages simply
964 * retry. There may be a lock held for some
965 * buffers that may go away. Later
969 unlock_page(newpage
);
971 rc
= swap_page(page
);
976 unlock_page(newpage
);
985 /* Permanent failure */
986 list_move(&page
->lru
, failed
);
990 /* Successful migration. Return page to LRU */
991 move_to_lru(newpage
);
993 list_move(&page
->lru
, moved
);
996 if (retry
&& pass
++ < 10)
1000 current
->flags
&= ~PF_SWAPWRITE
;
1002 return nr_failed
+ retry
;
1006 * Isolate one page from the LRU lists and put it on the
1007 * indicated list with elevated refcount.
1010 * 0 = page not on LRU list
1011 * 1 = page removed from LRU list and added to the specified list.
1013 int isolate_lru_page(struct page
*page
)
1017 if (PageLRU(page
)) {
1018 struct zone
*zone
= page_zone(page
);
1019 spin_lock_irq(&zone
->lru_lock
);
1020 if (TestClearPageLRU(page
)) {
1023 if (PageActive(page
))
1024 del_page_from_active_list(zone
, page
);
1026 del_page_from_inactive_list(zone
, page
);
1028 spin_unlock_irq(&zone
->lru_lock
);
1036 * zone->lru_lock is heavily contended. Some of the functions that
1037 * shrink the lists perform better by taking out a batch of pages
1038 * and working on them outside the LRU lock.
1040 * For pagecache intensive workloads, this function is the hottest
1041 * spot in the kernel (apart from copy_*_user functions).
1043 * Appropriate locks must be held before calling this function.
1045 * @nr_to_scan: The number of pages to look through on the list.
1046 * @src: The LRU list to pull pages off.
1047 * @dst: The temp list to put pages on to.
1048 * @scanned: The number of pages that were scanned.
1050 * returns how many pages were moved onto *@dst.
1052 static int isolate_lru_pages(int nr_to_scan
, struct list_head
*src
,
1053 struct list_head
*dst
, int *scanned
)
1059 while (scan
++ < nr_to_scan
&& !list_empty(src
)) {
1060 page
= lru_to_page(src
);
1061 prefetchw_prev_lru_page(page
, src
, flags
);
1063 if (!TestClearPageLRU(page
))
1065 list_del(&page
->lru
);
1066 if (get_page_testone(page
)) {
1068 * It is being freed elsewhere
1072 list_add(&page
->lru
, src
);
1075 list_add(&page
->lru
, dst
);
1085 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1087 static void shrink_cache(struct zone
*zone
, struct scan_control
*sc
)
1089 LIST_HEAD(page_list
);
1090 struct pagevec pvec
;
1091 int max_scan
= sc
->nr_to_scan
;
1093 pagevec_init(&pvec
, 1);
1096 spin_lock_irq(&zone
->lru_lock
);
1097 while (max_scan
> 0) {
1103 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
1104 &zone
->inactive_list
,
1105 &page_list
, &nr_scan
);
1106 zone
->nr_inactive
-= nr_taken
;
1107 zone
->pages_scanned
+= nr_scan
;
1108 spin_unlock_irq(&zone
->lru_lock
);
1113 max_scan
-= nr_scan
;
1114 nr_freed
= shrink_list(&page_list
, sc
);
1116 local_irq_disable();
1117 if (current_is_kswapd()) {
1118 __mod_page_state_zone(zone
, pgscan_kswapd
, nr_scan
);
1119 __mod_page_state(kswapd_steal
, nr_freed
);
1121 __mod_page_state_zone(zone
, pgscan_direct
, nr_scan
);
1122 __mod_page_state_zone(zone
, pgsteal
, nr_freed
);
1124 spin_lock(&zone
->lru_lock
);
1126 * Put back any unfreeable pages.
1128 while (!list_empty(&page_list
)) {
1129 page
= lru_to_page(&page_list
);
1130 if (TestSetPageLRU(page
))
1132 list_del(&page
->lru
);
1133 if (PageActive(page
))
1134 add_page_to_active_list(zone
, page
);
1136 add_page_to_inactive_list(zone
, page
);
1137 if (!pagevec_add(&pvec
, page
)) {
1138 spin_unlock_irq(&zone
->lru_lock
);
1139 __pagevec_release(&pvec
);
1140 spin_lock_irq(&zone
->lru_lock
);
1144 spin_unlock_irq(&zone
->lru_lock
);
1146 pagevec_release(&pvec
);
1150 * This moves pages from the active list to the inactive list.
1152 * We move them the other way if the page is referenced by one or more
1153 * processes, from rmap.
1155 * If the pages are mostly unmapped, the processing is fast and it is
1156 * appropriate to hold zone->lru_lock across the whole operation. But if
1157 * the pages are mapped, the processing is slow (page_referenced()) so we
1158 * should drop zone->lru_lock around each page. It's impossible to balance
1159 * this, so instead we remove the pages from the LRU while processing them.
1160 * It is safe to rely on PG_active against the non-LRU pages in here because
1161 * nobody will play with that bit on a non-LRU page.
1163 * The downside is that we have to touch page->_count against each page.
1164 * But we had to alter page->flags anyway.
1167 refill_inactive_zone(struct zone
*zone
, struct scan_control
*sc
)
1170 int pgdeactivate
= 0;
1172 int nr_pages
= sc
->nr_to_scan
;
1173 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1174 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
1175 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
1177 struct pagevec pvec
;
1178 int reclaim_mapped
= 0;
1184 spin_lock_irq(&zone
->lru_lock
);
1185 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
1186 &l_hold
, &pgscanned
);
1187 zone
->pages_scanned
+= pgscanned
;
1188 zone
->nr_active
-= pgmoved
;
1189 spin_unlock_irq(&zone
->lru_lock
);
1192 * `distress' is a measure of how much trouble we're having reclaiming
1193 * pages. 0 -> no problems. 100 -> great trouble.
1195 distress
= 100 >> zone
->prev_priority
;
1198 * The point of this algorithm is to decide when to start reclaiming
1199 * mapped memory instead of just pagecache. Work out how much memory
1202 mapped_ratio
= (sc
->nr_mapped
* 100) / total_memory
;
1205 * Now decide how much we really want to unmap some pages. The mapped
1206 * ratio is downgraded - just because there's a lot of mapped memory
1207 * doesn't necessarily mean that page reclaim isn't succeeding.
1209 * The distress ratio is important - we don't want to start going oom.
1211 * A 100% value of vm_swappiness overrides this algorithm altogether.
1213 swap_tendency
= mapped_ratio
/ 2 + distress
+ vm_swappiness
;
1216 * Now use this metric to decide whether to start moving mapped memory
1217 * onto the inactive list.
1219 if (swap_tendency
>= 100)
1222 while (!list_empty(&l_hold
)) {
1224 page
= lru_to_page(&l_hold
);
1225 list_del(&page
->lru
);
1226 if (page_mapped(page
)) {
1227 if (!reclaim_mapped
||
1228 (total_swap_pages
== 0 && PageAnon(page
)) ||
1229 page_referenced(page
, 0)) {
1230 list_add(&page
->lru
, &l_active
);
1234 list_add(&page
->lru
, &l_inactive
);
1237 pagevec_init(&pvec
, 1);
1239 spin_lock_irq(&zone
->lru_lock
);
1240 while (!list_empty(&l_inactive
)) {
1241 page
= lru_to_page(&l_inactive
);
1242 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
1243 if (TestSetPageLRU(page
))
1245 if (!TestClearPageActive(page
))
1247 list_move(&page
->lru
, &zone
->inactive_list
);
1249 if (!pagevec_add(&pvec
, page
)) {
1250 zone
->nr_inactive
+= pgmoved
;
1251 spin_unlock_irq(&zone
->lru_lock
);
1252 pgdeactivate
+= pgmoved
;
1254 if (buffer_heads_over_limit
)
1255 pagevec_strip(&pvec
);
1256 __pagevec_release(&pvec
);
1257 spin_lock_irq(&zone
->lru_lock
);
1260 zone
->nr_inactive
+= pgmoved
;
1261 pgdeactivate
+= pgmoved
;
1262 if (buffer_heads_over_limit
) {
1263 spin_unlock_irq(&zone
->lru_lock
);
1264 pagevec_strip(&pvec
);
1265 spin_lock_irq(&zone
->lru_lock
);
1269 while (!list_empty(&l_active
)) {
1270 page
= lru_to_page(&l_active
);
1271 prefetchw_prev_lru_page(page
, &l_active
, flags
);
1272 if (TestSetPageLRU(page
))
1274 BUG_ON(!PageActive(page
));
1275 list_move(&page
->lru
, &zone
->active_list
);
1277 if (!pagevec_add(&pvec
, page
)) {
1278 zone
->nr_active
+= pgmoved
;
1280 spin_unlock_irq(&zone
->lru_lock
);
1281 __pagevec_release(&pvec
);
1282 spin_lock_irq(&zone
->lru_lock
);
1285 zone
->nr_active
+= pgmoved
;
1286 spin_unlock(&zone
->lru_lock
);
1288 __mod_page_state_zone(zone
, pgrefill
, pgscanned
);
1289 __mod_page_state(pgdeactivate
, pgdeactivate
);
1292 pagevec_release(&pvec
);
1296 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1299 shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
1301 unsigned long nr_active
;
1302 unsigned long nr_inactive
;
1304 atomic_inc(&zone
->reclaim_in_progress
);
1307 * Add one to `nr_to_scan' just to make sure that the kernel will
1308 * slowly sift through the active list.
1310 zone
->nr_scan_active
+= (zone
->nr_active
>> sc
->priority
) + 1;
1311 nr_active
= zone
->nr_scan_active
;
1312 if (nr_active
>= sc
->swap_cluster_max
)
1313 zone
->nr_scan_active
= 0;
1317 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> sc
->priority
) + 1;
1318 nr_inactive
= zone
->nr_scan_inactive
;
1319 if (nr_inactive
>= sc
->swap_cluster_max
)
1320 zone
->nr_scan_inactive
= 0;
1324 while (nr_active
|| nr_inactive
) {
1326 sc
->nr_to_scan
= min(nr_active
,
1327 (unsigned long)sc
->swap_cluster_max
);
1328 nr_active
-= sc
->nr_to_scan
;
1329 refill_inactive_zone(zone
, sc
);
1333 sc
->nr_to_scan
= min(nr_inactive
,
1334 (unsigned long)sc
->swap_cluster_max
);
1335 nr_inactive
-= sc
->nr_to_scan
;
1336 shrink_cache(zone
, sc
);
1340 throttle_vm_writeout();
1342 atomic_dec(&zone
->reclaim_in_progress
);
1346 * This is the direct reclaim path, for page-allocating processes. We only
1347 * try to reclaim pages from zones which will satisfy the caller's allocation
1350 * We reclaim from a zone even if that zone is over pages_high. Because:
1351 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1353 * b) The zones may be over pages_high but they must go *over* pages_high to
1354 * satisfy the `incremental min' zone defense algorithm.
1356 * Returns the number of reclaimed pages.
1358 * If a zone is deemed to be full of pinned pages then just give it a light
1359 * scan then give up on it.
1362 shrink_caches(struct zone
**zones
, struct scan_control
*sc
)
1366 for (i
= 0; zones
[i
] != NULL
; i
++) {
1367 struct zone
*zone
= zones
[i
];
1369 if (!populated_zone(zone
))
1372 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1375 zone
->temp_priority
= sc
->priority
;
1376 if (zone
->prev_priority
> sc
->priority
)
1377 zone
->prev_priority
= sc
->priority
;
1379 if (zone
->all_unreclaimable
&& sc
->priority
!= DEF_PRIORITY
)
1380 continue; /* Let kswapd poll it */
1382 shrink_zone(zone
, sc
);
1387 * This is the main entry point to direct page reclaim.
1389 * If a full scan of the inactive list fails to free enough memory then we
1390 * are "out of memory" and something needs to be killed.
1392 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1393 * high - the zone may be full of dirty or under-writeback pages, which this
1394 * caller can't do much about. We kick pdflush and take explicit naps in the
1395 * hope that some of these pages can be written. But if the allocating task
1396 * holds filesystem locks which prevent writeout this might not work, and the
1397 * allocation attempt will fail.
1399 int try_to_free_pages(struct zone
**zones
, gfp_t gfp_mask
)
1403 int total_scanned
= 0, total_reclaimed
= 0;
1404 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1405 struct scan_control sc
;
1406 unsigned long lru_pages
= 0;
1409 sc
.gfp_mask
= gfp_mask
;
1410 sc
.may_writepage
= !laptop_mode
;
1413 inc_page_state(allocstall
);
1415 for (i
= 0; zones
[i
] != NULL
; i
++) {
1416 struct zone
*zone
= zones
[i
];
1418 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1421 zone
->temp_priority
= DEF_PRIORITY
;
1422 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1425 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1426 sc
.nr_mapped
= read_page_state(nr_mapped
);
1428 sc
.nr_reclaimed
= 0;
1429 sc
.priority
= priority
;
1430 sc
.swap_cluster_max
= SWAP_CLUSTER_MAX
;
1432 disable_swap_token();
1433 shrink_caches(zones
, &sc
);
1434 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1435 if (reclaim_state
) {
1436 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1437 reclaim_state
->reclaimed_slab
= 0;
1439 total_scanned
+= sc
.nr_scanned
;
1440 total_reclaimed
+= sc
.nr_reclaimed
;
1441 if (total_reclaimed
>= sc
.swap_cluster_max
) {
1447 * Try to write back as many pages as we just scanned. This
1448 * tends to cause slow streaming writers to write data to the
1449 * disk smoothly, at the dirtying rate, which is nice. But
1450 * that's undesirable in laptop mode, where we *want* lumpy
1451 * writeout. So in laptop mode, write out the whole world.
1453 if (total_scanned
> sc
.swap_cluster_max
+ sc
.swap_cluster_max
/2) {
1454 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1455 sc
.may_writepage
= 1;
1458 /* Take a nap, wait for some writeback to complete */
1459 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1460 blk_congestion_wait(WRITE
, HZ
/10);
1463 for (i
= 0; zones
[i
] != 0; i
++) {
1464 struct zone
*zone
= zones
[i
];
1466 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1469 zone
->prev_priority
= zone
->temp_priority
;
1475 * For kswapd, balance_pgdat() will work across all this node's zones until
1476 * they are all at pages_high.
1478 * If `nr_pages' is non-zero then it is the number of pages which are to be
1479 * reclaimed, regardless of the zone occupancies. This is a software suspend
1482 * Returns the number of pages which were actually freed.
1484 * There is special handling here for zones which are full of pinned pages.
1485 * This can happen if the pages are all mlocked, or if they are all used by
1486 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1487 * What we do is to detect the case where all pages in the zone have been
1488 * scanned twice and there has been zero successful reclaim. Mark the zone as
1489 * dead and from now on, only perform a short scan. Basically we're polling
1490 * the zone for when the problem goes away.
1492 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1493 * zones which have free_pages > pages_high, but once a zone is found to have
1494 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1495 * of the number of free pages in the lower zones. This interoperates with
1496 * the page allocator fallback scheme to ensure that aging of pages is balanced
1499 static int balance_pgdat(pg_data_t
*pgdat
, int nr_pages
, int order
)
1501 int to_free
= nr_pages
;
1505 int total_scanned
, total_reclaimed
;
1506 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1507 struct scan_control sc
;
1511 total_reclaimed
= 0;
1512 sc
.gfp_mask
= GFP_KERNEL
;
1513 sc
.may_writepage
= !laptop_mode
;
1515 sc
.nr_mapped
= read_page_state(nr_mapped
);
1517 inc_page_state(pageoutrun
);
1519 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1520 struct zone
*zone
= pgdat
->node_zones
+ i
;
1522 zone
->temp_priority
= DEF_PRIORITY
;
1525 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1526 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1527 unsigned long lru_pages
= 0;
1529 /* The swap token gets in the way of swapout... */
1531 disable_swap_token();
1535 if (nr_pages
== 0) {
1537 * Scan in the highmem->dma direction for the highest
1538 * zone which needs scanning
1540 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1541 struct zone
*zone
= pgdat
->node_zones
+ i
;
1543 if (!populated_zone(zone
))
1546 if (zone
->all_unreclaimable
&&
1547 priority
!= DEF_PRIORITY
)
1550 if (!zone_watermark_ok(zone
, order
,
1551 zone
->pages_high
, 0, 0)) {
1558 end_zone
= pgdat
->nr_zones
- 1;
1561 for (i
= 0; i
<= end_zone
; i
++) {
1562 struct zone
*zone
= pgdat
->node_zones
+ i
;
1564 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1568 * Now scan the zone in the dma->highmem direction, stopping
1569 * at the last zone which needs scanning.
1571 * We do this because the page allocator works in the opposite
1572 * direction. This prevents the page allocator from allocating
1573 * pages behind kswapd's direction of progress, which would
1574 * cause too much scanning of the lower zones.
1576 for (i
= 0; i
<= end_zone
; i
++) {
1577 struct zone
*zone
= pgdat
->node_zones
+ i
;
1580 if (!populated_zone(zone
))
1583 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1586 if (nr_pages
== 0) { /* Not software suspend */
1587 if (!zone_watermark_ok(zone
, order
,
1588 zone
->pages_high
, end_zone
, 0))
1591 zone
->temp_priority
= priority
;
1592 if (zone
->prev_priority
> priority
)
1593 zone
->prev_priority
= priority
;
1595 sc
.nr_reclaimed
= 0;
1596 sc
.priority
= priority
;
1597 sc
.swap_cluster_max
= nr_pages
? nr_pages
: SWAP_CLUSTER_MAX
;
1598 atomic_inc(&zone
->reclaim_in_progress
);
1599 shrink_zone(zone
, &sc
);
1600 atomic_dec(&zone
->reclaim_in_progress
);
1601 reclaim_state
->reclaimed_slab
= 0;
1602 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1604 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1605 total_reclaimed
+= sc
.nr_reclaimed
;
1606 total_scanned
+= sc
.nr_scanned
;
1607 if (zone
->all_unreclaimable
)
1609 if (nr_slab
== 0 && zone
->pages_scanned
>=
1610 (zone
->nr_active
+ zone
->nr_inactive
) * 4)
1611 zone
->all_unreclaimable
= 1;
1613 * If we've done a decent amount of scanning and
1614 * the reclaim ratio is low, start doing writepage
1615 * even in laptop mode
1617 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1618 total_scanned
> total_reclaimed
+total_reclaimed
/2)
1619 sc
.may_writepage
= 1;
1621 if (nr_pages
&& to_free
> total_reclaimed
)
1622 continue; /* swsusp: need to do more work */
1624 break; /* kswapd: all done */
1626 * OK, kswapd is getting into trouble. Take a nap, then take
1627 * another pass across the zones.
1629 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1630 blk_congestion_wait(WRITE
, HZ
/10);
1633 * We do this so kswapd doesn't build up large priorities for
1634 * example when it is freeing in parallel with allocators. It
1635 * matches the direct reclaim path behaviour in terms of impact
1636 * on zone->*_priority.
1638 if ((total_reclaimed
>= SWAP_CLUSTER_MAX
) && (!nr_pages
))
1642 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1643 struct zone
*zone
= pgdat
->node_zones
+ i
;
1645 zone
->prev_priority
= zone
->temp_priority
;
1647 if (!all_zones_ok
) {
1652 return total_reclaimed
;
1656 * The background pageout daemon, started as a kernel thread
1657 * from the init process.
1659 * This basically trickles out pages so that we have _some_
1660 * free memory available even if there is no other activity
1661 * that frees anything up. This is needed for things like routing
1662 * etc, where we otherwise might have all activity going on in
1663 * asynchronous contexts that cannot page things out.
1665 * If there are applications that are active memory-allocators
1666 * (most normal use), this basically shouldn't matter.
1668 static int kswapd(void *p
)
1670 unsigned long order
;
1671 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1672 struct task_struct
*tsk
= current
;
1674 struct reclaim_state reclaim_state
= {
1675 .reclaimed_slab
= 0,
1679 daemonize("kswapd%d", pgdat
->node_id
);
1680 cpumask
= node_to_cpumask(pgdat
->node_id
);
1681 if (!cpus_empty(cpumask
))
1682 set_cpus_allowed(tsk
, cpumask
);
1683 current
->reclaim_state
= &reclaim_state
;
1686 * Tell the memory management that we're a "memory allocator",
1687 * and that if we need more memory we should get access to it
1688 * regardless (see "__alloc_pages()"). "kswapd" should
1689 * never get caught in the normal page freeing logic.
1691 * (Kswapd normally doesn't need memory anyway, but sometimes
1692 * you need a small amount of memory in order to be able to
1693 * page out something else, and this flag essentially protects
1694 * us from recursively trying to free more memory as we're
1695 * trying to free the first piece of memory in the first place).
1697 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1701 unsigned long new_order
;
1705 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1706 new_order
= pgdat
->kswapd_max_order
;
1707 pgdat
->kswapd_max_order
= 0;
1708 if (order
< new_order
) {
1710 * Don't sleep if someone wants a larger 'order'
1716 order
= pgdat
->kswapd_max_order
;
1718 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1720 balance_pgdat(pgdat
, 0, order
);
1726 * A zone is low on free memory, so wake its kswapd task to service it.
1728 void wakeup_kswapd(struct zone
*zone
, int order
)
1732 if (!populated_zone(zone
))
1735 pgdat
= zone
->zone_pgdat
;
1736 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1738 if (pgdat
->kswapd_max_order
< order
)
1739 pgdat
->kswapd_max_order
= order
;
1740 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1742 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1744 wake_up_interruptible(&pgdat
->kswapd_wait
);
1749 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1752 int shrink_all_memory(int nr_pages
)
1755 int nr_to_free
= nr_pages
;
1757 struct reclaim_state reclaim_state
= {
1758 .reclaimed_slab
= 0,
1761 current
->reclaim_state
= &reclaim_state
;
1762 for_each_pgdat(pgdat
) {
1764 freed
= balance_pgdat(pgdat
, nr_to_free
, 0);
1766 nr_to_free
-= freed
;
1767 if (nr_to_free
<= 0)
1770 current
->reclaim_state
= NULL
;
1775 #ifdef CONFIG_HOTPLUG_CPU
1776 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1777 not required for correctness. So if the last cpu in a node goes
1778 away, we get changed to run anywhere: as the first one comes back,
1779 restore their cpu bindings. */
1780 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
1781 unsigned long action
,
1787 if (action
== CPU_ONLINE
) {
1788 for_each_pgdat(pgdat
) {
1789 mask
= node_to_cpumask(pgdat
->node_id
);
1790 if (any_online_cpu(mask
) != NR_CPUS
)
1791 /* One of our CPUs online: restore mask */
1792 set_cpus_allowed(pgdat
->kswapd
, mask
);
1797 #endif /* CONFIG_HOTPLUG_CPU */
1799 static int __init
kswapd_init(void)
1803 for_each_pgdat(pgdat
)
1805 = find_task_by_pid(kernel_thread(kswapd
, pgdat
, CLONE_KERNEL
));
1806 total_memory
= nr_free_pagecache_pages();
1807 hotcpu_notifier(cpu_callback
, 0);
1811 module_init(kswapd_init
)
1817 * If non-zero call zone_reclaim when the number of free pages falls below
1820 * In the future we may add flags to the mode. However, the page allocator
1821 * should only have to check that zone_reclaim_mode != 0 before calling
1824 int zone_reclaim_mode __read_mostly
;
1826 #define RECLAIM_OFF 0
1827 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1828 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1829 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1830 #define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
1833 * Mininum time between zone reclaim scans
1835 int zone_reclaim_interval __read_mostly
= 30*HZ
;
1838 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1839 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1842 #define ZONE_RECLAIM_PRIORITY 4
1845 * Try to free up some pages from this zone through reclaim.
1847 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1850 struct task_struct
*p
= current
;
1851 struct reclaim_state reclaim_state
;
1852 struct scan_control sc
;
1856 if (time_before(jiffies
,
1857 zone
->last_unsuccessful_zone_reclaim
+ zone_reclaim_interval
))
1860 if (!(gfp_mask
& __GFP_WAIT
) ||
1861 zone
->all_unreclaimable
||
1862 atomic_read(&zone
->reclaim_in_progress
) > 0)
1865 node_id
= zone
->zone_pgdat
->node_id
;
1866 mask
= node_to_cpumask(node_id
);
1867 if (!cpus_empty(mask
) && node_id
!= numa_node_id())
1870 sc
.may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
);
1871 sc
.may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
);
1873 sc
.nr_reclaimed
= 0;
1874 sc
.priority
= ZONE_RECLAIM_PRIORITY
+ 1;
1875 sc
.nr_mapped
= read_page_state(nr_mapped
);
1876 sc
.gfp_mask
= gfp_mask
;
1878 disable_swap_token();
1880 nr_pages
= 1 << order
;
1881 if (nr_pages
> SWAP_CLUSTER_MAX
)
1882 sc
.swap_cluster_max
= nr_pages
;
1884 sc
.swap_cluster_max
= SWAP_CLUSTER_MAX
;
1887 p
->flags
|= PF_MEMALLOC
;
1888 reclaim_state
.reclaimed_slab
= 0;
1889 p
->reclaim_state
= &reclaim_state
;
1892 * Free memory by calling shrink zone with increasing priorities
1893 * until we have enough memory freed.
1897 shrink_zone(zone
, &sc
);
1899 } while (sc
.nr_reclaimed
< nr_pages
&& sc
.priority
> 0);
1901 if (sc
.nr_reclaimed
< nr_pages
&& (zone_reclaim_mode
& RECLAIM_SLAB
)) {
1903 * shrink_slab does not currently allow us to determine
1904 * how many pages were freed in the zone. So we just
1905 * shake the slab and then go offnode for a single allocation.
1907 * shrink_slab will free memory on all zones and may take
1910 shrink_slab(sc
.nr_scanned
, gfp_mask
, order
);
1911 sc
.nr_reclaimed
= 1; /* Avoid getting the off node timeout */
1914 p
->reclaim_state
= NULL
;
1915 current
->flags
&= ~PF_MEMALLOC
;
1917 if (sc
.nr_reclaimed
== 0)
1918 zone
->last_unsuccessful_zone_reclaim
= jiffies
;
1920 return sc
.nr_reclaimed
>= nr_pages
;