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 /* Incremented by the number of inactive pages that were scanned */
56 unsigned long nr_scanned
;
58 unsigned long nr_mapped
; /* From page_state */
60 /* This context's GFP mask */
65 /* Can pages be swapped as part of reclaim? */
68 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
69 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
70 * In this context, it doesn't matter that we scan the
71 * whole list at once. */
76 * The list of shrinker callbacks used by to apply pressure to
81 struct list_head list
;
82 int seeks
; /* seeks to recreate an obj */
83 long nr
; /* objs pending delete */
86 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
88 #ifdef ARCH_HAS_PREFETCH
89 #define prefetch_prev_lru_page(_page, _base, _field) \
91 if ((_page)->lru.prev != _base) { \
94 prev = lru_to_page(&(_page->lru)); \
95 prefetch(&prev->_field); \
99 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
102 #ifdef ARCH_HAS_PREFETCHW
103 #define prefetchw_prev_lru_page(_page, _base, _field) \
105 if ((_page)->lru.prev != _base) { \
108 prev = lru_to_page(&(_page->lru)); \
109 prefetchw(&prev->_field); \
113 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
117 * From 0 .. 100. Higher means more swappy.
119 int vm_swappiness
= 60;
120 static long total_memory
;
122 static LIST_HEAD(shrinker_list
);
123 static DECLARE_RWSEM(shrinker_rwsem
);
126 * Add a shrinker callback to be called from the vm
128 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
130 struct shrinker
*shrinker
;
132 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
134 shrinker
->shrinker
= theshrinker
;
135 shrinker
->seeks
= seeks
;
137 down_write(&shrinker_rwsem
);
138 list_add_tail(&shrinker
->list
, &shrinker_list
);
139 up_write(&shrinker_rwsem
);
143 EXPORT_SYMBOL(set_shrinker
);
148 void remove_shrinker(struct shrinker
*shrinker
)
150 down_write(&shrinker_rwsem
);
151 list_del(&shrinker
->list
);
152 up_write(&shrinker_rwsem
);
155 EXPORT_SYMBOL(remove_shrinker
);
157 #define SHRINK_BATCH 128
159 * Call the shrink functions to age shrinkable caches
161 * Here we assume it costs one seek to replace a lru page and that it also
162 * takes a seek to recreate a cache object. With this in mind we age equal
163 * percentages of the lru and ageable caches. This should balance the seeks
164 * generated by these structures.
166 * If the vm encounted mapped pages on the LRU it increase the pressure on
167 * slab to avoid swapping.
169 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
171 * `lru_pages' represents the number of on-LRU pages in all the zones which
172 * are eligible for the caller's allocation attempt. It is used for balancing
173 * slab reclaim versus page reclaim.
175 * Returns the number of slab objects which we shrunk.
177 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
178 unsigned long lru_pages
)
180 struct shrinker
*shrinker
;
181 unsigned long ret
= 0;
184 scanned
= SWAP_CLUSTER_MAX
;
186 if (!down_read_trylock(&shrinker_rwsem
))
187 return 1; /* Assume we'll be able to shrink next time */
189 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
190 unsigned long long delta
;
191 unsigned long total_scan
;
192 unsigned long max_pass
= (*shrinker
->shrinker
)(0, gfp_mask
);
194 delta
= (4 * scanned
) / shrinker
->seeks
;
196 do_div(delta
, lru_pages
+ 1);
197 shrinker
->nr
+= delta
;
198 if (shrinker
->nr
< 0) {
199 printk(KERN_ERR
"%s: nr=%ld\n",
200 __FUNCTION__
, shrinker
->nr
);
201 shrinker
->nr
= max_pass
;
205 * Avoid risking looping forever due to too large nr value:
206 * never try to free more than twice the estimate number of
209 if (shrinker
->nr
> max_pass
* 2)
210 shrinker
->nr
= max_pass
* 2;
212 total_scan
= shrinker
->nr
;
215 while (total_scan
>= SHRINK_BATCH
) {
216 long this_scan
= SHRINK_BATCH
;
220 nr_before
= (*shrinker
->shrinker
)(0, gfp_mask
);
221 shrink_ret
= (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
222 if (shrink_ret
== -1)
224 if (shrink_ret
< nr_before
)
225 ret
+= nr_before
- shrink_ret
;
226 mod_page_state(slabs_scanned
, this_scan
);
227 total_scan
-= this_scan
;
232 shrinker
->nr
+= total_scan
;
234 up_read(&shrinker_rwsem
);
238 /* Called without lock on whether page is mapped, so answer is unstable */
239 static inline int page_mapping_inuse(struct page
*page
)
241 struct address_space
*mapping
;
243 /* Page is in somebody's page tables. */
244 if (page_mapped(page
))
247 /* Be more reluctant to reclaim swapcache than pagecache */
248 if (PageSwapCache(page
))
251 mapping
= page_mapping(page
);
255 /* File is mmap'd by somebody? */
256 return mapping_mapped(mapping
);
259 static inline int is_page_cache_freeable(struct page
*page
)
261 return page_count(page
) - !!PagePrivate(page
) == 2;
264 static int may_write_to_queue(struct backing_dev_info
*bdi
)
266 if (current
->flags
& PF_SWAPWRITE
)
268 if (!bdi_write_congested(bdi
))
270 if (bdi
== current
->backing_dev_info
)
276 * We detected a synchronous write error writing a page out. Probably
277 * -ENOSPC. We need to propagate that into the address_space for a subsequent
278 * fsync(), msync() or close().
280 * The tricky part is that after writepage we cannot touch the mapping: nothing
281 * prevents it from being freed up. But we have a ref on the page and once
282 * that page is locked, the mapping is pinned.
284 * We're allowed to run sleeping lock_page() here because we know the caller has
287 static void handle_write_error(struct address_space
*mapping
,
288 struct page
*page
, int error
)
291 if (page_mapping(page
) == mapping
) {
292 if (error
== -ENOSPC
)
293 set_bit(AS_ENOSPC
, &mapping
->flags
);
295 set_bit(AS_EIO
, &mapping
->flags
);
301 * pageout is called by shrink_page_list() for each dirty page.
302 * Calls ->writepage().
304 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
307 * If the page is dirty, only perform writeback if that write
308 * will be non-blocking. To prevent this allocation from being
309 * stalled by pagecache activity. But note that there may be
310 * stalls if we need to run get_block(). We could test
311 * PagePrivate for that.
313 * If this process is currently in generic_file_write() against
314 * this page's queue, we can perform writeback even if that
317 * If the page is swapcache, write it back even if that would
318 * block, for some throttling. This happens by accident, because
319 * swap_backing_dev_info is bust: it doesn't reflect the
320 * congestion state of the swapdevs. Easy to fix, if needed.
321 * See swapfile.c:page_queue_congested().
323 if (!is_page_cache_freeable(page
))
327 * Some data journaling orphaned pages can have
328 * page->mapping == NULL while being dirty with clean buffers.
330 if (PagePrivate(page
)) {
331 if (try_to_free_buffers(page
)) {
332 ClearPageDirty(page
);
333 printk("%s: orphaned page\n", __FUNCTION__
);
339 if (mapping
->a_ops
->writepage
== NULL
)
340 return PAGE_ACTIVATE
;
341 if (!may_write_to_queue(mapping
->backing_dev_info
))
344 if (clear_page_dirty_for_io(page
)) {
346 struct writeback_control wbc
= {
347 .sync_mode
= WB_SYNC_NONE
,
348 .nr_to_write
= SWAP_CLUSTER_MAX
,
353 SetPageReclaim(page
);
354 res
= mapping
->a_ops
->writepage(page
, &wbc
);
356 handle_write_error(mapping
, page
, res
);
357 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
358 ClearPageReclaim(page
);
359 return PAGE_ACTIVATE
;
361 if (!PageWriteback(page
)) {
362 /* synchronous write or broken a_ops? */
363 ClearPageReclaim(page
);
372 static int remove_mapping(struct address_space
*mapping
, struct page
*page
)
375 return 0; /* truncate got there first */
377 write_lock_irq(&mapping
->tree_lock
);
380 * The non-racy check for busy page. It is critical to check
381 * PageDirty _after_ making sure that the page is freeable and
382 * not in use by anybody. (pagecache + us == 2)
384 if (unlikely(page_count(page
) != 2))
387 if (unlikely(PageDirty(page
)))
390 if (PageSwapCache(page
)) {
391 swp_entry_t swap
= { .val
= page_private(page
) };
392 __delete_from_swap_cache(page
);
393 write_unlock_irq(&mapping
->tree_lock
);
395 __put_page(page
); /* The pagecache ref */
399 __remove_from_page_cache(page
);
400 write_unlock_irq(&mapping
->tree_lock
);
405 write_unlock_irq(&mapping
->tree_lock
);
410 * shrink_page_list() returns the number of reclaimed pages
412 static unsigned long shrink_page_list(struct list_head
*page_list
,
413 struct scan_control
*sc
)
415 LIST_HEAD(ret_pages
);
416 struct pagevec freed_pvec
;
418 unsigned long nr_reclaimed
= 0;
422 pagevec_init(&freed_pvec
, 1);
423 while (!list_empty(page_list
)) {
424 struct address_space
*mapping
;
431 page
= lru_to_page(page_list
);
432 list_del(&page
->lru
);
434 if (TestSetPageLocked(page
))
437 BUG_ON(PageActive(page
));
441 if (!sc
->may_swap
&& page_mapped(page
))
444 /* Double the slab pressure for mapped and swapcache pages */
445 if (page_mapped(page
) || PageSwapCache(page
))
448 if (PageWriteback(page
))
451 referenced
= page_referenced(page
, 1);
452 /* In active use or really unfreeable? Activate it. */
453 if (referenced
&& page_mapping_inuse(page
))
454 goto activate_locked
;
458 * Anonymous process memory has backing store?
459 * Try to allocate it some swap space here.
461 if (PageAnon(page
) && !PageSwapCache(page
)) {
464 if (!add_to_swap(page
, GFP_ATOMIC
))
465 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
) {
479 * No unmapping if we do not swap
484 switch (try_to_unmap(page
, 0)) {
486 goto activate_locked
;
490 ; /* try to free the page below */
494 if (PageDirty(page
)) {
499 if (!sc
->may_writepage
)
502 /* Page is dirty, try to write it out here */
503 switch(pageout(page
, mapping
)) {
507 goto activate_locked
;
509 if (PageWriteback(page
) || PageDirty(page
))
512 * A synchronous write - probably a ramdisk. Go
513 * ahead and try to reclaim the page.
515 if (TestSetPageLocked(page
))
517 if (PageDirty(page
) || PageWriteback(page
))
519 mapping
= page_mapping(page
);
521 ; /* try to free the page below */
526 * If the page has buffers, try to free the buffer mappings
527 * associated with this page. If we succeed we try to free
530 * We do this even if the page is PageDirty().
531 * try_to_release_page() does not perform I/O, but it is
532 * possible for a page to have PageDirty set, but it is actually
533 * clean (all its buffers are clean). This happens if the
534 * buffers were written out directly, with submit_bh(). ext3
535 * will do this, as well as the blockdev mapping.
536 * try_to_release_page() will discover that cleanness and will
537 * drop the buffers and mark the page clean - it can be freed.
539 * Rarely, pages can have buffers and no ->mapping. These are
540 * the pages which were not successfully invalidated in
541 * truncate_complete_page(). We try to drop those buffers here
542 * and if that worked, and the page is no longer mapped into
543 * process address space (page_count == 1) it can be freed.
544 * Otherwise, leave the page on the LRU so it is swappable.
546 if (PagePrivate(page
)) {
547 if (!try_to_release_page(page
, sc
->gfp_mask
))
548 goto activate_locked
;
549 if (!mapping
&& page_count(page
) == 1)
553 if (!remove_mapping(mapping
, page
))
559 if (!pagevec_add(&freed_pvec
, page
))
560 __pagevec_release_nonlru(&freed_pvec
);
569 list_add(&page
->lru
, &ret_pages
);
570 BUG_ON(PageLRU(page
));
572 list_splice(&ret_pages
, page_list
);
573 if (pagevec_count(&freed_pvec
))
574 __pagevec_release_nonlru(&freed_pvec
);
575 mod_page_state(pgactivate
, pgactivate
);
579 #ifdef CONFIG_MIGRATION
580 static inline void move_to_lru(struct page
*page
)
582 list_del(&page
->lru
);
583 if (PageActive(page
)) {
585 * lru_cache_add_active checks that
586 * the PG_active bit is off.
588 ClearPageActive(page
);
589 lru_cache_add_active(page
);
597 * Add isolated pages on the list back to the LRU.
599 * returns the number of pages put back.
601 unsigned long putback_lru_pages(struct list_head
*l
)
605 unsigned long count
= 0;
607 list_for_each_entry_safe(page
, page2
, l
, lru
) {
615 * Non migratable page
617 int fail_migrate_page(struct page
*newpage
, struct page
*page
)
621 EXPORT_SYMBOL(fail_migrate_page
);
624 * swapout a single page
625 * page is locked upon entry, unlocked on exit
627 static int swap_page(struct page
*page
)
629 struct address_space
*mapping
= page_mapping(page
);
631 if (page_mapped(page
) && mapping
)
632 if (try_to_unmap(page
, 1) != SWAP_SUCCESS
)
635 if (PageDirty(page
)) {
636 /* Page is dirty, try to write it out here */
637 switch(pageout(page
, mapping
)) {
646 ; /* try to free the page below */
650 if (PagePrivate(page
)) {
651 if (!try_to_release_page(page
, GFP_KERNEL
) ||
652 (!mapping
&& page_count(page
) == 1))
656 if (remove_mapping(mapping
, page
)) {
668 EXPORT_SYMBOL(swap_page
);
671 * Page migration was first developed in the context of the memory hotplug
672 * project. The main authors of the migration code are:
674 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
675 * Hirokazu Takahashi <taka@valinux.co.jp>
676 * Dave Hansen <haveblue@us.ibm.com>
677 * Christoph Lameter <clameter@sgi.com>
681 * Remove references for a page and establish the new page with the correct
682 * basic settings to be able to stop accesses to the page.
684 int migrate_page_remove_references(struct page
*newpage
,
685 struct page
*page
, int nr_refs
)
687 struct address_space
*mapping
= page_mapping(page
);
688 struct page
**radix_pointer
;
691 * Avoid doing any of the following work if the page count
692 * indicates that the page is in use or truncate has removed
695 if (!mapping
|| page_mapcount(page
) + nr_refs
!= page_count(page
))
699 * Establish swap ptes for anonymous pages or destroy pte
702 * In order to reestablish file backed mappings the fault handlers
703 * will take the radix tree_lock which may then be used to stop
704 * processses from accessing this page until the new page is ready.
706 * A process accessing via a swap pte (an anonymous page) will take a
707 * page_lock on the old page which will block the process until the
708 * migration attempt is complete. At that time the PageSwapCache bit
709 * will be examined. If the page was migrated then the PageSwapCache
710 * bit will be clear and the operation to retrieve the page will be
711 * retried which will find the new page in the radix tree. Then a new
712 * direct mapping may be generated based on the radix tree contents.
714 * If the page was not migrated then the PageSwapCache bit
715 * is still set and the operation may continue.
717 if (try_to_unmap(page
, 1) == SWAP_FAIL
)
718 /* A vma has VM_LOCKED set -> Permanent failure */
722 * Give up if we were unable to remove all mappings.
724 if (page_mapcount(page
))
727 write_lock_irq(&mapping
->tree_lock
);
729 radix_pointer
= (struct page
**)radix_tree_lookup_slot(
733 if (!page_mapping(page
) || page_count(page
) != nr_refs
||
734 *radix_pointer
!= page
) {
735 write_unlock_irq(&mapping
->tree_lock
);
740 * Now we know that no one else is looking at the page.
742 * Certain minimal information about a page must be available
743 * in order for other subsystems to properly handle the page if they
744 * find it through the radix tree update before we are finished
748 newpage
->index
= page
->index
;
749 newpage
->mapping
= page
->mapping
;
750 if (PageSwapCache(page
)) {
751 SetPageSwapCache(newpage
);
752 set_page_private(newpage
, page_private(page
));
755 *radix_pointer
= newpage
;
757 write_unlock_irq(&mapping
->tree_lock
);
761 EXPORT_SYMBOL(migrate_page_remove_references
);
764 * Copy the page to its new location
766 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
768 copy_highpage(newpage
, page
);
771 SetPageError(newpage
);
772 if (PageReferenced(page
))
773 SetPageReferenced(newpage
);
774 if (PageUptodate(page
))
775 SetPageUptodate(newpage
);
776 if (PageActive(page
))
777 SetPageActive(newpage
);
778 if (PageChecked(page
))
779 SetPageChecked(newpage
);
780 if (PageMappedToDisk(page
))
781 SetPageMappedToDisk(newpage
);
783 if (PageDirty(page
)) {
784 clear_page_dirty_for_io(page
);
785 set_page_dirty(newpage
);
788 ClearPageSwapCache(page
);
789 ClearPageActive(page
);
790 ClearPagePrivate(page
);
791 set_page_private(page
, 0);
792 page
->mapping
= NULL
;
795 * If any waiters have accumulated on the new page then
798 if (PageWriteback(newpage
))
799 end_page_writeback(newpage
);
801 EXPORT_SYMBOL(migrate_page_copy
);
804 * Common logic to directly migrate a single page suitable for
805 * pages that do not use PagePrivate.
807 * Pages are locked upon entry and exit.
809 int migrate_page(struct page
*newpage
, struct page
*page
)
813 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
815 rc
= migrate_page_remove_references(newpage
, page
, 2);
820 migrate_page_copy(newpage
, page
);
823 * Remove auxiliary swap entries and replace
824 * them with real ptes.
826 * Note that a real pte entry will allow processes that are not
827 * waiting on the page lock to use the new page via the page tables
828 * before the new page is unlocked.
830 remove_from_swap(newpage
);
833 EXPORT_SYMBOL(migrate_page
);
838 * Two lists are passed to this function. The first list
839 * contains the pages isolated from the LRU to be migrated.
840 * The second list contains new pages that the pages isolated
841 * can be moved to. If the second list is NULL then all
842 * pages are swapped out.
844 * The function returns after 10 attempts or if no pages
845 * are movable anymore because to has become empty
846 * or no retryable pages exist anymore.
848 * Return: Number of pages not migrated when "to" ran empty.
850 unsigned long migrate_pages(struct list_head
*from
, struct list_head
*to
,
851 struct list_head
*moved
, struct list_head
*failed
)
854 unsigned long nr_failed
= 0;
858 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
862 current
->flags
|= PF_SWAPWRITE
;
867 list_for_each_entry_safe(page
, page2
, from
, lru
) {
868 struct page
*newpage
= NULL
;
869 struct address_space
*mapping
;
874 if (page_count(page
) == 1)
875 /* page was freed from under us. So we are done. */
878 if (to
&& list_empty(to
))
882 * Skip locked pages during the first two passes to give the
883 * functions holding the lock time to release the page. Later we
884 * use lock_page() to have a higher chance of acquiring the
891 if (TestSetPageLocked(page
))
895 * Only wait on writeback if we have already done a pass where
896 * we we may have triggered writeouts for lots of pages.
899 wait_on_page_writeback(page
);
901 if (PageWriteback(page
))
906 * Anonymous pages must have swap cache references otherwise
907 * the information contained in the page maps cannot be
910 if (PageAnon(page
) && !PageSwapCache(page
)) {
911 if (!add_to_swap(page
, GFP_KERNEL
)) {
918 rc
= swap_page(page
);
922 newpage
= lru_to_page(to
);
926 * Pages are properly locked and writeback is complete.
927 * Try to migrate the page.
929 mapping
= page_mapping(page
);
933 if (mapping
->a_ops
->migratepage
) {
935 * Most pages have a mapping and most filesystems
936 * should provide a migration function. Anonymous
937 * pages are part of swap space which also has its
938 * own migration function. This is the most common
939 * path for page migration.
941 rc
= mapping
->a_ops
->migratepage(newpage
, page
);
946 * Default handling if a filesystem does not provide
947 * a migration function. We can only migrate clean
948 * pages so try to write out any dirty pages first.
950 if (PageDirty(page
)) {
951 switch (pageout(page
, mapping
)) {
957 unlock_page(newpage
);
961 ; /* try to migrate the page below */
966 * Buffers are managed in a filesystem specific way.
967 * We must have no buffers or drop them.
969 if (!page_has_buffers(page
) ||
970 try_to_release_page(page
, GFP_KERNEL
)) {
971 rc
= migrate_page(newpage
, page
);
976 * On early passes with mapped pages simply
977 * retry. There may be a lock held for some
978 * buffers that may go away. Later
983 * Persistently unable to drop buffers..... As a
984 * measure of last resort we fall back to
987 unlock_page(newpage
);
989 rc
= swap_page(page
);
994 unlock_page(newpage
);
1000 if (rc
== -EAGAIN
) {
1003 /* Permanent failure */
1004 list_move(&page
->lru
, failed
);
1008 /* Successful migration. Return page to LRU */
1009 move_to_lru(newpage
);
1011 list_move(&page
->lru
, moved
);
1014 if (retry
&& pass
++ < 10)
1018 current
->flags
&= ~PF_SWAPWRITE
;
1020 return nr_failed
+ retry
;
1024 * Isolate one page from the LRU lists and put it on the
1025 * indicated list with elevated refcount.
1028 * 0 = page not on LRU list
1029 * 1 = page removed from LRU list and added to the specified list.
1031 int isolate_lru_page(struct page
*page
)
1035 if (PageLRU(page
)) {
1036 struct zone
*zone
= page_zone(page
);
1037 spin_lock_irq(&zone
->lru_lock
);
1038 if (PageLRU(page
)) {
1042 if (PageActive(page
))
1043 del_page_from_active_list(zone
, page
);
1045 del_page_from_inactive_list(zone
, page
);
1047 spin_unlock_irq(&zone
->lru_lock
);
1055 * zone->lru_lock is heavily contended. Some of the functions that
1056 * shrink the lists perform better by taking out a batch of pages
1057 * and working on them outside the LRU lock.
1059 * For pagecache intensive workloads, this function is the hottest
1060 * spot in the kernel (apart from copy_*_user functions).
1062 * Appropriate locks must be held before calling this function.
1064 * @nr_to_scan: The number of pages to look through on the list.
1065 * @src: The LRU list to pull pages off.
1066 * @dst: The temp list to put pages on to.
1067 * @scanned: The number of pages that were scanned.
1069 * returns how many pages were moved onto *@dst.
1071 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1072 struct list_head
*src
, struct list_head
*dst
,
1073 unsigned long *scanned
)
1075 unsigned long nr_taken
= 0;
1079 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1080 struct list_head
*target
;
1081 page
= lru_to_page(src
);
1082 prefetchw_prev_lru_page(page
, src
, flags
);
1084 BUG_ON(!PageLRU(page
));
1086 list_del(&page
->lru
);
1088 if (likely(get_page_unless_zero(page
))) {
1090 * Be careful not to clear PageLRU until after we're
1091 * sure the page is not being freed elsewhere -- the
1092 * page release code relies on it.
1097 } /* else it is being freed elsewhere */
1099 list_add(&page
->lru
, target
);
1107 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1108 * of reclaimed pages
1110 static unsigned long shrink_inactive_list(unsigned long max_scan
,
1111 struct zone
*zone
, struct scan_control
*sc
)
1113 LIST_HEAD(page_list
);
1114 struct pagevec pvec
;
1115 unsigned long nr_scanned
= 0;
1116 unsigned long nr_reclaimed
= 0;
1118 pagevec_init(&pvec
, 1);
1121 spin_lock_irq(&zone
->lru_lock
);
1124 unsigned long nr_taken
;
1125 unsigned long nr_scan
;
1126 unsigned long nr_freed
;
1128 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
1129 &zone
->inactive_list
,
1130 &page_list
, &nr_scan
);
1131 zone
->nr_inactive
-= nr_taken
;
1132 zone
->pages_scanned
+= nr_scan
;
1133 spin_unlock_irq(&zone
->lru_lock
);
1135 nr_scanned
+= nr_scan
;
1136 nr_freed
= shrink_page_list(&page_list
, sc
);
1137 nr_reclaimed
+= nr_freed
;
1138 local_irq_disable();
1139 if (current_is_kswapd()) {
1140 __mod_page_state_zone(zone
, pgscan_kswapd
, nr_scan
);
1141 __mod_page_state(kswapd_steal
, nr_freed
);
1143 __mod_page_state_zone(zone
, pgscan_direct
, nr_scan
);
1144 __mod_page_state_zone(zone
, pgsteal
, nr_freed
);
1149 spin_lock(&zone
->lru_lock
);
1151 * Put back any unfreeable pages.
1153 while (!list_empty(&page_list
)) {
1154 page
= lru_to_page(&page_list
);
1155 BUG_ON(PageLRU(page
));
1157 list_del(&page
->lru
);
1158 if (PageActive(page
))
1159 add_page_to_active_list(zone
, page
);
1161 add_page_to_inactive_list(zone
, page
);
1162 if (!pagevec_add(&pvec
, page
)) {
1163 spin_unlock_irq(&zone
->lru_lock
);
1164 __pagevec_release(&pvec
);
1165 spin_lock_irq(&zone
->lru_lock
);
1168 } while (nr_scanned
< max_scan
);
1169 spin_unlock(&zone
->lru_lock
);
1172 pagevec_release(&pvec
);
1173 return nr_reclaimed
;
1177 * This moves pages from the active list to the inactive list.
1179 * We move them the other way if the page is referenced by one or more
1180 * processes, from rmap.
1182 * If the pages are mostly unmapped, the processing is fast and it is
1183 * appropriate to hold zone->lru_lock across the whole operation. But if
1184 * the pages are mapped, the processing is slow (page_referenced()) so we
1185 * should drop zone->lru_lock around each page. It's impossible to balance
1186 * this, so instead we remove the pages from the LRU while processing them.
1187 * It is safe to rely on PG_active against the non-LRU pages in here because
1188 * nobody will play with that bit on a non-LRU page.
1190 * The downside is that we have to touch page->_count against each page.
1191 * But we had to alter page->flags anyway.
1193 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1194 struct scan_control
*sc
)
1196 unsigned long pgmoved
;
1197 int pgdeactivate
= 0;
1198 unsigned long pgscanned
;
1199 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1200 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
1201 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
1203 struct pagevec pvec
;
1204 int reclaim_mapped
= 0;
1206 if (unlikely(sc
->may_swap
)) {
1212 * `distress' is a measure of how much trouble we're having
1213 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1215 distress
= 100 >> zone
->prev_priority
;
1218 * The point of this algorithm is to decide when to start
1219 * reclaiming mapped memory instead of just pagecache. Work out
1223 mapped_ratio
= (sc
->nr_mapped
* 100) / total_memory
;
1226 * Now decide how much we really want to unmap some pages. The
1227 * mapped ratio is downgraded - just because there's a lot of
1228 * mapped memory doesn't necessarily mean that page reclaim
1231 * The distress ratio is important - we don't want to start
1234 * A 100% value of vm_swappiness overrides this algorithm
1237 swap_tendency
= mapped_ratio
/ 2 + distress
+ vm_swappiness
;
1240 * Now use this metric to decide whether to start moving mapped
1241 * memory onto the inactive list.
1243 if (swap_tendency
>= 100)
1248 spin_lock_irq(&zone
->lru_lock
);
1249 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
1250 &l_hold
, &pgscanned
);
1251 zone
->pages_scanned
+= pgscanned
;
1252 zone
->nr_active
-= pgmoved
;
1253 spin_unlock_irq(&zone
->lru_lock
);
1255 while (!list_empty(&l_hold
)) {
1257 page
= lru_to_page(&l_hold
);
1258 list_del(&page
->lru
);
1259 if (page_mapped(page
)) {
1260 if (!reclaim_mapped
||
1261 (total_swap_pages
== 0 && PageAnon(page
)) ||
1262 page_referenced(page
, 0)) {
1263 list_add(&page
->lru
, &l_active
);
1267 list_add(&page
->lru
, &l_inactive
);
1270 pagevec_init(&pvec
, 1);
1272 spin_lock_irq(&zone
->lru_lock
);
1273 while (!list_empty(&l_inactive
)) {
1274 page
= lru_to_page(&l_inactive
);
1275 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
1276 BUG_ON(PageLRU(page
));
1278 BUG_ON(!PageActive(page
));
1279 ClearPageActive(page
);
1281 list_move(&page
->lru
, &zone
->inactive_list
);
1283 if (!pagevec_add(&pvec
, page
)) {
1284 zone
->nr_inactive
+= pgmoved
;
1285 spin_unlock_irq(&zone
->lru_lock
);
1286 pgdeactivate
+= pgmoved
;
1288 if (buffer_heads_over_limit
)
1289 pagevec_strip(&pvec
);
1290 __pagevec_release(&pvec
);
1291 spin_lock_irq(&zone
->lru_lock
);
1294 zone
->nr_inactive
+= pgmoved
;
1295 pgdeactivate
+= pgmoved
;
1296 if (buffer_heads_over_limit
) {
1297 spin_unlock_irq(&zone
->lru_lock
);
1298 pagevec_strip(&pvec
);
1299 spin_lock_irq(&zone
->lru_lock
);
1303 while (!list_empty(&l_active
)) {
1304 page
= lru_to_page(&l_active
);
1305 prefetchw_prev_lru_page(page
, &l_active
, flags
);
1306 BUG_ON(PageLRU(page
));
1308 BUG_ON(!PageActive(page
));
1309 list_move(&page
->lru
, &zone
->active_list
);
1311 if (!pagevec_add(&pvec
, page
)) {
1312 zone
->nr_active
+= pgmoved
;
1314 spin_unlock_irq(&zone
->lru_lock
);
1315 __pagevec_release(&pvec
);
1316 spin_lock_irq(&zone
->lru_lock
);
1319 zone
->nr_active
+= pgmoved
;
1320 spin_unlock(&zone
->lru_lock
);
1322 __mod_page_state_zone(zone
, pgrefill
, pgscanned
);
1323 __mod_page_state(pgdeactivate
, pgdeactivate
);
1326 pagevec_release(&pvec
);
1330 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1332 static unsigned long shrink_zone(int priority
, struct zone
*zone
,
1333 struct scan_control
*sc
)
1335 unsigned long nr_active
;
1336 unsigned long nr_inactive
;
1337 unsigned long nr_to_scan
;
1338 unsigned long nr_reclaimed
= 0;
1340 atomic_inc(&zone
->reclaim_in_progress
);
1343 * Add one to `nr_to_scan' just to make sure that the kernel will
1344 * slowly sift through the active list.
1346 zone
->nr_scan_active
+= (zone
->nr_active
>> priority
) + 1;
1347 nr_active
= zone
->nr_scan_active
;
1348 if (nr_active
>= sc
->swap_cluster_max
)
1349 zone
->nr_scan_active
= 0;
1353 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> priority
) + 1;
1354 nr_inactive
= zone
->nr_scan_inactive
;
1355 if (nr_inactive
>= sc
->swap_cluster_max
)
1356 zone
->nr_scan_inactive
= 0;
1360 while (nr_active
|| nr_inactive
) {
1362 nr_to_scan
= min(nr_active
,
1363 (unsigned long)sc
->swap_cluster_max
);
1364 nr_active
-= nr_to_scan
;
1365 shrink_active_list(nr_to_scan
, zone
, sc
);
1369 nr_to_scan
= min(nr_inactive
,
1370 (unsigned long)sc
->swap_cluster_max
);
1371 nr_inactive
-= nr_to_scan
;
1372 nr_reclaimed
+= shrink_inactive_list(nr_to_scan
, zone
,
1377 throttle_vm_writeout();
1379 atomic_dec(&zone
->reclaim_in_progress
);
1380 return nr_reclaimed
;
1384 * This is the direct reclaim path, for page-allocating processes. We only
1385 * try to reclaim pages from zones which will satisfy the caller's allocation
1388 * We reclaim from a zone even if that zone is over pages_high. Because:
1389 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1391 * b) The zones may be over pages_high but they must go *over* pages_high to
1392 * satisfy the `incremental min' zone defense algorithm.
1394 * Returns the number of reclaimed pages.
1396 * If a zone is deemed to be full of pinned pages then just give it a light
1397 * scan then give up on it.
1399 static unsigned long shrink_zones(int priority
, struct zone
**zones
,
1400 struct scan_control
*sc
)
1402 unsigned long nr_reclaimed
= 0;
1405 for (i
= 0; zones
[i
] != NULL
; i
++) {
1406 struct zone
*zone
= zones
[i
];
1408 if (!populated_zone(zone
))
1411 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1414 zone
->temp_priority
= priority
;
1415 if (zone
->prev_priority
> priority
)
1416 zone
->prev_priority
= priority
;
1418 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1419 continue; /* Let kswapd poll it */
1421 nr_reclaimed
+= shrink_zone(priority
, zone
, sc
);
1423 return nr_reclaimed
;
1427 * This is the main entry point to direct page reclaim.
1429 * If a full scan of the inactive list fails to free enough memory then we
1430 * are "out of memory" and something needs to be killed.
1432 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1433 * high - the zone may be full of dirty or under-writeback pages, which this
1434 * caller can't do much about. We kick pdflush and take explicit naps in the
1435 * hope that some of these pages can be written. But if the allocating task
1436 * holds filesystem locks which prevent writeout this might not work, and the
1437 * allocation attempt will fail.
1439 unsigned long try_to_free_pages(struct zone
**zones
, gfp_t gfp_mask
)
1443 unsigned long total_scanned
= 0;
1444 unsigned long nr_reclaimed
= 0;
1445 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1446 unsigned long lru_pages
= 0;
1448 struct scan_control sc
= {
1449 .gfp_mask
= gfp_mask
,
1450 .may_writepage
= !laptop_mode
,
1451 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1455 inc_page_state(allocstall
);
1457 for (i
= 0; zones
[i
] != NULL
; i
++) {
1458 struct zone
*zone
= zones
[i
];
1460 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1463 zone
->temp_priority
= DEF_PRIORITY
;
1464 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1467 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1468 sc
.nr_mapped
= read_page_state(nr_mapped
);
1471 disable_swap_token();
1472 nr_reclaimed
+= shrink_zones(priority
, zones
, &sc
);
1473 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1474 if (reclaim_state
) {
1475 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1476 reclaim_state
->reclaimed_slab
= 0;
1478 total_scanned
+= sc
.nr_scanned
;
1479 if (nr_reclaimed
>= sc
.swap_cluster_max
) {
1485 * Try to write back as many pages as we just scanned. This
1486 * tends to cause slow streaming writers to write data to the
1487 * disk smoothly, at the dirtying rate, which is nice. But
1488 * that's undesirable in laptop mode, where we *want* lumpy
1489 * writeout. So in laptop mode, write out the whole world.
1491 if (total_scanned
> sc
.swap_cluster_max
+
1492 sc
.swap_cluster_max
/ 2) {
1493 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1494 sc
.may_writepage
= 1;
1497 /* Take a nap, wait for some writeback to complete */
1498 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1499 blk_congestion_wait(WRITE
, HZ
/10);
1502 for (i
= 0; zones
[i
] != 0; i
++) {
1503 struct zone
*zone
= zones
[i
];
1505 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1508 zone
->prev_priority
= zone
->temp_priority
;
1514 * For kswapd, balance_pgdat() will work across all this node's zones until
1515 * they are all at pages_high.
1517 * If `nr_pages' is non-zero then it is the number of pages which are to be
1518 * reclaimed, regardless of the zone occupancies. This is a software suspend
1521 * Returns the number of pages which were actually freed.
1523 * There is special handling here for zones which are full of pinned pages.
1524 * This can happen if the pages are all mlocked, or if they are all used by
1525 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1526 * What we do is to detect the case where all pages in the zone have been
1527 * scanned twice and there has been zero successful reclaim. Mark the zone as
1528 * dead and from now on, only perform a short scan. Basically we're polling
1529 * the zone for when the problem goes away.
1531 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1532 * zones which have free_pages > pages_high, but once a zone is found to have
1533 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1534 * of the number of free pages in the lower zones. This interoperates with
1535 * the page allocator fallback scheme to ensure that aging of pages is balanced
1538 static unsigned long balance_pgdat(pg_data_t
*pgdat
, unsigned long nr_pages
,
1541 unsigned long to_free
= nr_pages
;
1545 unsigned long total_scanned
;
1546 unsigned long nr_reclaimed
;
1547 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1548 struct scan_control sc
= {
1549 .gfp_mask
= GFP_KERNEL
,
1551 .swap_cluster_max
= nr_pages
? nr_pages
: SWAP_CLUSTER_MAX
,
1557 sc
.may_writepage
= !laptop_mode
,
1558 sc
.nr_mapped
= read_page_state(nr_mapped
);
1560 inc_page_state(pageoutrun
);
1562 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1563 struct zone
*zone
= pgdat
->node_zones
+ i
;
1565 zone
->temp_priority
= DEF_PRIORITY
;
1568 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1569 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1570 unsigned long lru_pages
= 0;
1572 /* The swap token gets in the way of swapout... */
1574 disable_swap_token();
1578 if (nr_pages
== 0) {
1580 * Scan in the highmem->dma direction for the highest
1581 * zone which needs scanning
1583 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1584 struct zone
*zone
= pgdat
->node_zones
+ i
;
1586 if (!populated_zone(zone
))
1589 if (zone
->all_unreclaimable
&&
1590 priority
!= DEF_PRIORITY
)
1593 if (!zone_watermark_ok(zone
, order
,
1594 zone
->pages_high
, 0, 0)) {
1601 end_zone
= pgdat
->nr_zones
- 1;
1604 for (i
= 0; i
<= end_zone
; i
++) {
1605 struct zone
*zone
= pgdat
->node_zones
+ i
;
1607 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1611 * Now scan the zone in the dma->highmem direction, stopping
1612 * at the last zone which needs scanning.
1614 * We do this because the page allocator works in the opposite
1615 * direction. This prevents the page allocator from allocating
1616 * pages behind kswapd's direction of progress, which would
1617 * cause too much scanning of the lower zones.
1619 for (i
= 0; i
<= end_zone
; i
++) {
1620 struct zone
*zone
= pgdat
->node_zones
+ i
;
1623 if (!populated_zone(zone
))
1626 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1629 if (nr_pages
== 0) { /* Not software suspend */
1630 if (!zone_watermark_ok(zone
, order
,
1631 zone
->pages_high
, end_zone
, 0))
1634 zone
->temp_priority
= priority
;
1635 if (zone
->prev_priority
> priority
)
1636 zone
->prev_priority
= priority
;
1638 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1639 reclaim_state
->reclaimed_slab
= 0;
1640 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1642 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1643 total_scanned
+= sc
.nr_scanned
;
1644 if (zone
->all_unreclaimable
)
1646 if (nr_slab
== 0 && zone
->pages_scanned
>=
1647 (zone
->nr_active
+ zone
->nr_inactive
) * 4)
1648 zone
->all_unreclaimable
= 1;
1650 * If we've done a decent amount of scanning and
1651 * the reclaim ratio is low, start doing writepage
1652 * even in laptop mode
1654 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1655 total_scanned
> nr_reclaimed
+ nr_reclaimed
/ 2)
1656 sc
.may_writepage
= 1;
1658 if (nr_pages
&& to_free
> nr_reclaimed
)
1659 continue; /* swsusp: need to do more work */
1661 break; /* kswapd: all done */
1663 * OK, kswapd is getting into trouble. Take a nap, then take
1664 * another pass across the zones.
1666 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1667 blk_congestion_wait(WRITE
, HZ
/10);
1670 * We do this so kswapd doesn't build up large priorities for
1671 * example when it is freeing in parallel with allocators. It
1672 * matches the direct reclaim path behaviour in terms of impact
1673 * on zone->*_priority.
1675 if ((nr_reclaimed
>= SWAP_CLUSTER_MAX
) && !nr_pages
)
1679 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1680 struct zone
*zone
= pgdat
->node_zones
+ i
;
1682 zone
->prev_priority
= zone
->temp_priority
;
1684 if (!all_zones_ok
) {
1689 return nr_reclaimed
;
1693 * The background pageout daemon, started as a kernel thread
1694 * from the init process.
1696 * This basically trickles out pages so that we have _some_
1697 * free memory available even if there is no other activity
1698 * that frees anything up. This is needed for things like routing
1699 * etc, where we otherwise might have all activity going on in
1700 * asynchronous contexts that cannot page things out.
1702 * If there are applications that are active memory-allocators
1703 * (most normal use), this basically shouldn't matter.
1705 static int kswapd(void *p
)
1707 unsigned long order
;
1708 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1709 struct task_struct
*tsk
= current
;
1711 struct reclaim_state reclaim_state
= {
1712 .reclaimed_slab
= 0,
1716 daemonize("kswapd%d", pgdat
->node_id
);
1717 cpumask
= node_to_cpumask(pgdat
->node_id
);
1718 if (!cpus_empty(cpumask
))
1719 set_cpus_allowed(tsk
, cpumask
);
1720 current
->reclaim_state
= &reclaim_state
;
1723 * Tell the memory management that we're a "memory allocator",
1724 * and that if we need more memory we should get access to it
1725 * regardless (see "__alloc_pages()"). "kswapd" should
1726 * never get caught in the normal page freeing logic.
1728 * (Kswapd normally doesn't need memory anyway, but sometimes
1729 * you need a small amount of memory in order to be able to
1730 * page out something else, and this flag essentially protects
1731 * us from recursively trying to free more memory as we're
1732 * trying to free the first piece of memory in the first place).
1734 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1738 unsigned long new_order
;
1742 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1743 new_order
= pgdat
->kswapd_max_order
;
1744 pgdat
->kswapd_max_order
= 0;
1745 if (order
< new_order
) {
1747 * Don't sleep if someone wants a larger 'order'
1753 order
= pgdat
->kswapd_max_order
;
1755 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1757 balance_pgdat(pgdat
, 0, order
);
1763 * A zone is low on free memory, so wake its kswapd task to service it.
1765 void wakeup_kswapd(struct zone
*zone
, int order
)
1769 if (!populated_zone(zone
))
1772 pgdat
= zone
->zone_pgdat
;
1773 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1775 if (pgdat
->kswapd_max_order
< order
)
1776 pgdat
->kswapd_max_order
= order
;
1777 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1779 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1781 wake_up_interruptible(&pgdat
->kswapd_wait
);
1786 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1789 unsigned long shrink_all_memory(unsigned long nr_pages
)
1792 unsigned long nr_to_free
= nr_pages
;
1793 unsigned long ret
= 0;
1794 struct reclaim_state reclaim_state
= {
1795 .reclaimed_slab
= 0,
1798 current
->reclaim_state
= &reclaim_state
;
1799 for_each_pgdat(pgdat
) {
1800 unsigned long freed
;
1802 freed
= balance_pgdat(pgdat
, nr_to_free
, 0);
1804 nr_to_free
-= freed
;
1805 if ((long)nr_to_free
<= 0)
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
);