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/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
40 #include <asm/tlbflush.h>
41 #include <asm/div64.h>
43 #include <linux/swapops.h>
48 /* Incremented by the number of inactive pages that were scanned */
49 unsigned long nr_scanned
;
51 /* This context's GFP mask */
56 /* Can pages be swapped as part of reclaim? */
59 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
60 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
61 * In this context, it doesn't matter that we scan the
62 * whole list at once. */
67 int all_unreclaimable
;
71 * The list of shrinker callbacks used by to apply pressure to
76 struct list_head list
;
77 int seeks
; /* seeks to recreate an obj */
78 long nr
; /* objs pending delete */
81 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
83 #ifdef ARCH_HAS_PREFETCH
84 #define prefetch_prev_lru_page(_page, _base, _field) \
86 if ((_page)->lru.prev != _base) { \
89 prev = lru_to_page(&(_page->lru)); \
90 prefetch(&prev->_field); \
94 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
97 #ifdef ARCH_HAS_PREFETCHW
98 #define prefetchw_prev_lru_page(_page, _base, _field) \
100 if ((_page)->lru.prev != _base) { \
103 prev = lru_to_page(&(_page->lru)); \
104 prefetchw(&prev->_field); \
108 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
112 * From 0 .. 100. Higher means more swappy.
114 int vm_swappiness
= 60;
115 long vm_total_pages
; /* The total number of pages which the VM controls */
117 static LIST_HEAD(shrinker_list
);
118 static DECLARE_RWSEM(shrinker_rwsem
);
121 * Add a shrinker callback to be called from the vm
123 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
125 struct shrinker
*shrinker
;
127 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
129 shrinker
->shrinker
= theshrinker
;
130 shrinker
->seeks
= seeks
;
132 down_write(&shrinker_rwsem
);
133 list_add_tail(&shrinker
->list
, &shrinker_list
);
134 up_write(&shrinker_rwsem
);
138 EXPORT_SYMBOL(set_shrinker
);
143 void remove_shrinker(struct shrinker
*shrinker
)
145 down_write(&shrinker_rwsem
);
146 list_del(&shrinker
->list
);
147 up_write(&shrinker_rwsem
);
150 EXPORT_SYMBOL(remove_shrinker
);
152 #define SHRINK_BATCH 128
154 * Call the shrink functions to age shrinkable caches
156 * Here we assume it costs one seek to replace a lru page and that it also
157 * takes a seek to recreate a cache object. With this in mind we age equal
158 * percentages of the lru and ageable caches. This should balance the seeks
159 * generated by these structures.
161 * If the vm encounted mapped pages on the LRU it increase the pressure on
162 * slab to avoid swapping.
164 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
166 * `lru_pages' represents the number of on-LRU pages in all the zones which
167 * are eligible for the caller's allocation attempt. It is used for balancing
168 * slab reclaim versus page reclaim.
170 * Returns the number of slab objects which we shrunk.
172 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
173 unsigned long lru_pages
)
175 struct shrinker
*shrinker
;
176 unsigned long ret
= 0;
179 scanned
= SWAP_CLUSTER_MAX
;
181 if (!down_read_trylock(&shrinker_rwsem
))
182 return 1; /* Assume we'll be able to shrink next time */
184 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
185 unsigned long long delta
;
186 unsigned long total_scan
;
187 unsigned long max_pass
= (*shrinker
->shrinker
)(0, gfp_mask
);
189 delta
= (4 * scanned
) / shrinker
->seeks
;
191 do_div(delta
, lru_pages
+ 1);
192 shrinker
->nr
+= delta
;
193 if (shrinker
->nr
< 0) {
194 printk(KERN_ERR
"%s: nr=%ld\n",
195 __FUNCTION__
, shrinker
->nr
);
196 shrinker
->nr
= max_pass
;
200 * Avoid risking looping forever due to too large nr value:
201 * never try to free more than twice the estimate number of
204 if (shrinker
->nr
> max_pass
* 2)
205 shrinker
->nr
= max_pass
* 2;
207 total_scan
= shrinker
->nr
;
210 while (total_scan
>= SHRINK_BATCH
) {
211 long this_scan
= SHRINK_BATCH
;
215 nr_before
= (*shrinker
->shrinker
)(0, gfp_mask
);
216 shrink_ret
= (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
217 if (shrink_ret
== -1)
219 if (shrink_ret
< nr_before
)
220 ret
+= nr_before
- shrink_ret
;
221 count_vm_events(SLABS_SCANNED
, this_scan
);
222 total_scan
-= this_scan
;
227 shrinker
->nr
+= total_scan
;
229 up_read(&shrinker_rwsem
);
233 /* Called without lock on whether page is mapped, so answer is unstable */
234 static inline int page_mapping_inuse(struct page
*page
)
236 struct address_space
*mapping
;
238 /* Page is in somebody's page tables. */
239 if (page_mapped(page
))
242 /* Be more reluctant to reclaim swapcache than pagecache */
243 if (PageSwapCache(page
))
246 mapping
= page_mapping(page
);
250 /* File is mmap'd by somebody? */
251 return mapping_mapped(mapping
);
254 static inline int is_page_cache_freeable(struct page
*page
)
256 return page_count(page
) - !!PagePrivate(page
) == 2;
259 static int may_write_to_queue(struct backing_dev_info
*bdi
)
261 if (current
->flags
& PF_SWAPWRITE
)
263 if (!bdi_write_congested(bdi
))
265 if (bdi
== current
->backing_dev_info
)
271 * We detected a synchronous write error writing a page out. Probably
272 * -ENOSPC. We need to propagate that into the address_space for a subsequent
273 * fsync(), msync() or close().
275 * The tricky part is that after writepage we cannot touch the mapping: nothing
276 * prevents it from being freed up. But we have a ref on the page and once
277 * that page is locked, the mapping is pinned.
279 * We're allowed to run sleeping lock_page() here because we know the caller has
282 static void handle_write_error(struct address_space
*mapping
,
283 struct page
*page
, int error
)
286 if (page_mapping(page
) == mapping
) {
287 if (error
== -ENOSPC
)
288 set_bit(AS_ENOSPC
, &mapping
->flags
);
290 set_bit(AS_EIO
, &mapping
->flags
);
295 /* possible outcome of pageout() */
297 /* failed to write page out, page is locked */
299 /* move page to the active list, page is locked */
301 /* page has been sent to the disk successfully, page is unlocked */
303 /* page is clean and locked */
308 * pageout is called by shrink_page_list() for each dirty page.
309 * 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
,
357 .range_end
= LLONG_MAX
,
362 SetPageReclaim(page
);
363 res
= mapping
->a_ops
->writepage(page
, &wbc
);
365 handle_write_error(mapping
, page
, res
);
366 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
367 ClearPageReclaim(page
);
368 return PAGE_ACTIVATE
;
370 if (!PageWriteback(page
)) {
371 /* synchronous write or broken a_ops? */
372 ClearPageReclaim(page
);
374 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
382 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
383 * someone else has a ref on the page, abort and return 0. If it was
384 * successfully detached, return 1. Assumes the caller has a single ref on
387 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
389 BUG_ON(!PageLocked(page
));
390 BUG_ON(mapping
!= page_mapping(page
));
392 write_lock_irq(&mapping
->tree_lock
);
394 * The non racy check for a busy page.
396 * Must be careful with the order of the tests. When someone has
397 * a ref to the page, it may be possible that they dirty it then
398 * drop the reference. So if PageDirty is tested before page_count
399 * here, then the following race may occur:
401 * get_user_pages(&page);
402 * [user mapping goes away]
404 * !PageDirty(page) [good]
405 * SetPageDirty(page);
407 * !page_count(page) [good, discard it]
409 * [oops, our write_to data is lost]
411 * Reversing the order of the tests ensures such a situation cannot
412 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
413 * load is not satisfied before that of page->_count.
415 * Note that if SetPageDirty is always performed via set_page_dirty,
416 * and thus under tree_lock, then this ordering is not required.
418 if (unlikely(page_count(page
) != 2))
421 if (unlikely(PageDirty(page
)))
424 if (PageSwapCache(page
)) {
425 #if 0 // mask by Victor Yu. 03-15-2007
426 swp_entry_t swap
= { .val
= page_private(page
) };
429 swap
.val
= page_private(page
);
431 __delete_from_swap_cache(page
);
432 write_unlock_irq(&mapping
->tree_lock
);
434 __put_page(page
); /* The pagecache ref */
438 __remove_from_page_cache(page
);
439 write_unlock_irq(&mapping
->tree_lock
);
444 write_unlock_irq(&mapping
->tree_lock
);
449 * shrink_page_list() returns the number of reclaimed pages
451 static unsigned long shrink_page_list(struct list_head
*page_list
,
452 struct scan_control
*sc
)
454 LIST_HEAD(ret_pages
);
455 struct pagevec freed_pvec
;
457 unsigned long nr_reclaimed
= 0;
461 pagevec_init(&freed_pvec
, 1);
462 while (!list_empty(page_list
)) {
463 struct address_space
*mapping
;
470 page
= lru_to_page(page_list
);
471 list_del(&page
->lru
);
473 if (TestSetPageLocked(page
))
476 VM_BUG_ON(PageActive(page
));
480 if (!sc
->may_swap
&& page_mapped(page
))
483 /* Double the slab pressure for mapped and swapcache pages */
484 if (page_mapped(page
) || PageSwapCache(page
))
487 if (PageWriteback(page
))
490 referenced
= page_referenced(page
, 1);
491 /* In active use or really unfreeable? Activate it. */
492 if (referenced
&& page_mapping_inuse(page
))
493 goto activate_locked
;
497 * Anonymous process memory has backing store?
498 * Try to allocate it some swap space here.
500 if (PageAnon(page
) && !PageSwapCache(page
))
501 if (!add_to_swap(page
, GFP_ATOMIC
))
502 goto activate_locked
;
503 #endif /* CONFIG_SWAP */
505 mapping
= page_mapping(page
);
506 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
507 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
510 * The page is mapped into the page tables of one or more
511 * processes. Try to unmap it here.
513 if (page_mapped(page
) && mapping
) {
514 switch (try_to_unmap(page
, 0)) {
516 goto activate_locked
;
520 ; /* try to free the page below */
524 if (PageDirty(page
)) {
529 if (!sc
->may_writepage
)
532 /* Page is dirty, try to write it out here */
533 switch(pageout(page
, mapping
)) {
537 goto activate_locked
;
539 if (PageWriteback(page
) || PageDirty(page
))
542 * A synchronous write - probably a ramdisk. Go
543 * ahead and try to reclaim the page.
545 if (TestSetPageLocked(page
))
547 if (PageDirty(page
) || PageWriteback(page
))
549 mapping
= page_mapping(page
);
551 ; /* try to free the page below */
556 * If the page has buffers, try to free the buffer mappings
557 * associated with this page. If we succeed we try to free
560 * We do this even if the page is PageDirty().
561 * try_to_release_page() does not perform I/O, but it is
562 * possible for a page to have PageDirty set, but it is actually
563 * clean (all its buffers are clean). This happens if the
564 * buffers were written out directly, with submit_bh(). ext3
565 * will do this, as well as the blockdev mapping.
566 * try_to_release_page() will discover that cleanness and will
567 * drop the buffers and mark the page clean - it can be freed.
569 * Rarely, pages can have buffers and no ->mapping. These are
570 * the pages which were not successfully invalidated in
571 * truncate_complete_page(). We try to drop those buffers here
572 * and if that worked, and the page is no longer mapped into
573 * process address space (page_count == 1) it can be freed.
574 * Otherwise, leave the page on the LRU so it is swappable.
576 if (PagePrivate(page
)) {
577 if (!try_to_release_page(page
, sc
->gfp_mask
))
578 goto activate_locked
;
579 if (!mapping
&& page_count(page
) == 1)
583 if (!mapping
|| !remove_mapping(mapping
, page
))
589 if (!pagevec_add(&freed_pvec
, page
))
590 __pagevec_release_nonlru(&freed_pvec
);
599 list_add(&page
->lru
, &ret_pages
);
600 VM_BUG_ON(PageLRU(page
));
602 list_splice(&ret_pages
, page_list
);
603 if (pagevec_count(&freed_pvec
))
604 __pagevec_release_nonlru(&freed_pvec
);
605 count_vm_events(PGACTIVATE
, pgactivate
);
610 * zone->lru_lock is heavily contended. Some of the functions that
611 * shrink the lists perform better by taking out a batch of pages
612 * and working on them outside the LRU lock.
614 * For pagecache intensive workloads, this function is the hottest
615 * spot in the kernel (apart from copy_*_user functions).
617 * Appropriate locks must be held before calling this function.
619 * @nr_to_scan: The number of pages to look through on the list.
620 * @src: The LRU list to pull pages off.
621 * @dst: The temp list to put pages on to.
622 * @scanned: The number of pages that were scanned.
624 * returns how many pages were moved onto *@dst.
626 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
627 struct list_head
*src
, struct list_head
*dst
,
628 unsigned long *scanned
)
630 unsigned long nr_taken
= 0;
634 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
635 struct list_head
*target
;
636 page
= lru_to_page(src
);
637 prefetchw_prev_lru_page(page
, src
, flags
);
639 VM_BUG_ON(!PageLRU(page
));
641 list_del(&page
->lru
);
643 if (likely(get_page_unless_zero(page
))) {
645 * Be careful not to clear PageLRU until after we're
646 * sure the page is not being freed elsewhere -- the
647 * page release code relies on it.
652 } /* else it is being freed elsewhere */
654 list_add(&page
->lru
, target
);
662 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
665 static unsigned long shrink_inactive_list(unsigned long max_scan
,
666 struct zone
*zone
, struct scan_control
*sc
)
668 LIST_HEAD(page_list
);
670 unsigned long nr_scanned
= 0;
671 unsigned long nr_reclaimed
= 0;
673 pagevec_init(&pvec
, 1);
676 spin_lock_irq(&zone
->lru_lock
);
679 unsigned long nr_taken
;
680 unsigned long nr_scan
;
681 unsigned long nr_freed
;
683 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
684 &zone
->inactive_list
,
685 &page_list
, &nr_scan
);
686 zone
->nr_inactive
-= nr_taken
;
687 zone
->pages_scanned
+= nr_scan
;
688 spin_unlock_irq(&zone
->lru_lock
);
690 nr_scanned
+= nr_scan
;
691 nr_freed
= shrink_page_list(&page_list
, sc
);
692 nr_reclaimed
+= nr_freed
;
694 if (current_is_kswapd()) {
695 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scan
);
696 __count_vm_events(KSWAPD_STEAL
, nr_freed
);
698 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scan
);
699 __count_vm_events(PGACTIVATE
, nr_freed
);
704 spin_lock(&zone
->lru_lock
);
706 * Put back any unfreeable pages.
708 while (!list_empty(&page_list
)) {
709 page
= lru_to_page(&page_list
);
710 VM_BUG_ON(PageLRU(page
));
712 list_del(&page
->lru
);
713 if (PageActive(page
))
714 add_page_to_active_list(zone
, page
);
716 add_page_to_inactive_list(zone
, page
);
717 if (!pagevec_add(&pvec
, page
)) {
718 spin_unlock_irq(&zone
->lru_lock
);
719 __pagevec_release(&pvec
);
720 spin_lock_irq(&zone
->lru_lock
);
723 } while (nr_scanned
< max_scan
);
724 spin_unlock(&zone
->lru_lock
);
727 pagevec_release(&pvec
);
732 * We are about to scan this zone at a certain priority level. If that priority
733 * level is smaller (ie: more urgent) than the previous priority, then note
734 * that priority level within the zone. This is done so that when the next
735 * process comes in to scan this zone, it will immediately start out at this
736 * priority level rather than having to build up its own scanning priority.
737 * Here, this priority affects only the reclaim-mapped threshold.
739 static inline void note_zone_scanning_priority(struct zone
*zone
, int priority
)
741 if (priority
< zone
->prev_priority
)
742 zone
->prev_priority
= priority
;
745 static inline int zone_is_near_oom(struct zone
*zone
)
747 return zone
->pages_scanned
>= (zone
->nr_active
+ zone
->nr_inactive
)*3;
751 * This moves pages from the active list to the inactive list.
753 * We move them the other way if the page is referenced by one or more
754 * processes, from rmap.
756 * If the pages are mostly unmapped, the processing is fast and it is
757 * appropriate to hold zone->lru_lock across the whole operation. But if
758 * the pages are mapped, the processing is slow (page_referenced()) so we
759 * should drop zone->lru_lock around each page. It's impossible to balance
760 * this, so instead we remove the pages from the LRU while processing them.
761 * It is safe to rely on PG_active against the non-LRU pages in here because
762 * nobody will play with that bit on a non-LRU page.
764 * The downside is that we have to touch page->_count against each page.
765 * But we had to alter page->flags anyway.
767 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
768 struct scan_control
*sc
, int priority
)
770 unsigned long pgmoved
;
771 int pgdeactivate
= 0;
772 unsigned long pgscanned
;
773 LIST_HEAD(l_hold
); /* The pages which were snipped off */
774 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
775 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
778 int reclaim_mapped
= 0;
785 if (zone_is_near_oom(zone
))
786 goto force_reclaim_mapped
;
789 * `distress' is a measure of how much trouble we're having
790 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
792 distress
= 100 >> min(zone
->prev_priority
, priority
);
795 * The point of this algorithm is to decide when to start
796 * reclaiming mapped memory instead of just pagecache. Work out
800 mapped_ratio
= ((global_page_state(NR_FILE_MAPPED
) +
801 global_page_state(NR_ANON_PAGES
)) * 100) /
805 * Now decide how much we really want to unmap some pages. The
806 * mapped ratio is downgraded - just because there's a lot of
807 * mapped memory doesn't necessarily mean that page reclaim
810 * The distress ratio is important - we don't want to start
813 * A 100% value of vm_swappiness overrides this algorithm
816 swap_tendency
= mapped_ratio
/ 2 + distress
+ sc
->swappiness
;
819 * Now use this metric to decide whether to start moving mapped
820 * memory onto the inactive list.
822 if (swap_tendency
>= 100)
823 force_reclaim_mapped
:
828 spin_lock_irq(&zone
->lru_lock
);
829 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
830 &l_hold
, &pgscanned
);
831 zone
->pages_scanned
+= pgscanned
;
832 zone
->nr_active
-= pgmoved
;
833 spin_unlock_irq(&zone
->lru_lock
);
835 while (!list_empty(&l_hold
)) {
837 page
= lru_to_page(&l_hold
);
838 list_del(&page
->lru
);
839 if (page_mapped(page
)) {
840 if (!reclaim_mapped
||
841 (total_swap_pages
== 0 && PageAnon(page
)) ||
842 page_referenced(page
, 0)) {
843 list_add(&page
->lru
, &l_active
);
847 list_add(&page
->lru
, &l_inactive
);
850 pagevec_init(&pvec
, 1);
852 spin_lock_irq(&zone
->lru_lock
);
853 while (!list_empty(&l_inactive
)) {
854 page
= lru_to_page(&l_inactive
);
855 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
856 VM_BUG_ON(PageLRU(page
));
858 VM_BUG_ON(!PageActive(page
));
859 ClearPageActive(page
);
861 list_move(&page
->lru
, &zone
->inactive_list
);
863 if (!pagevec_add(&pvec
, page
)) {
864 zone
->nr_inactive
+= pgmoved
;
865 spin_unlock_irq(&zone
->lru_lock
);
866 pgdeactivate
+= pgmoved
;
868 if (buffer_heads_over_limit
)
869 pagevec_strip(&pvec
);
870 __pagevec_release(&pvec
);
871 spin_lock_irq(&zone
->lru_lock
);
874 zone
->nr_inactive
+= pgmoved
;
875 pgdeactivate
+= pgmoved
;
876 if (buffer_heads_over_limit
) {
877 spin_unlock_irq(&zone
->lru_lock
);
878 pagevec_strip(&pvec
);
879 spin_lock_irq(&zone
->lru_lock
);
883 while (!list_empty(&l_active
)) {
884 page
= lru_to_page(&l_active
);
885 prefetchw_prev_lru_page(page
, &l_active
, flags
);
886 VM_BUG_ON(PageLRU(page
));
888 VM_BUG_ON(!PageActive(page
));
889 list_move(&page
->lru
, &zone
->active_list
);
891 if (!pagevec_add(&pvec
, page
)) {
892 zone
->nr_active
+= pgmoved
;
894 spin_unlock_irq(&zone
->lru_lock
);
895 __pagevec_release(&pvec
);
896 spin_lock_irq(&zone
->lru_lock
);
899 zone
->nr_active
+= pgmoved
;
901 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
902 __count_vm_events(PGDEACTIVATE
, pgdeactivate
);
903 spin_unlock_irq(&zone
->lru_lock
);
905 pagevec_release(&pvec
);
909 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
911 static unsigned long shrink_zone(int priority
, struct zone
*zone
,
912 struct scan_control
*sc
)
914 unsigned long nr_active
;
915 unsigned long nr_inactive
;
916 unsigned long nr_to_scan
;
917 unsigned long nr_reclaimed
= 0;
919 atomic_inc(&zone
->reclaim_in_progress
);
922 * Add one to `nr_to_scan' just to make sure that the kernel will
923 * slowly sift through the active list.
925 zone
->nr_scan_active
+= (zone
->nr_active
>> priority
) + 1;
926 nr_active
= zone
->nr_scan_active
;
927 if (nr_active
>= sc
->swap_cluster_max
)
928 zone
->nr_scan_active
= 0;
932 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> priority
) + 1;
933 nr_inactive
= zone
->nr_scan_inactive
;
934 if (nr_inactive
>= sc
->swap_cluster_max
)
935 zone
->nr_scan_inactive
= 0;
939 while (nr_active
|| nr_inactive
) {
941 nr_to_scan
= min(nr_active
,
942 (unsigned long)sc
->swap_cluster_max
);
943 nr_active
-= nr_to_scan
;
944 shrink_active_list(nr_to_scan
, zone
, sc
, priority
);
948 nr_to_scan
= min(nr_inactive
,
949 (unsigned long)sc
->swap_cluster_max
);
950 nr_inactive
-= nr_to_scan
;
951 nr_reclaimed
+= shrink_inactive_list(nr_to_scan
, zone
,
956 throttle_vm_writeout();
958 atomic_dec(&zone
->reclaim_in_progress
);
963 * This is the direct reclaim path, for page-allocating processes. We only
964 * try to reclaim pages from zones which will satisfy the caller's allocation
967 * We reclaim from a zone even if that zone is over pages_high. Because:
968 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
970 * b) The zones may be over pages_high but they must go *over* pages_high to
971 * satisfy the `incremental min' zone defense algorithm.
973 * Returns the number of reclaimed pages.
975 * If a zone is deemed to be full of pinned pages then just give it a light
976 * scan then give up on it.
978 static unsigned long shrink_zones(int priority
, struct zone
**zones
,
979 struct scan_control
*sc
)
981 unsigned long nr_reclaimed
= 0;
984 sc
->all_unreclaimable
= 1;
985 for (i
= 0; zones
[i
] != NULL
; i
++) {
986 struct zone
*zone
= zones
[i
];
988 if (!populated_zone(zone
))
991 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
994 note_zone_scanning_priority(zone
, priority
);
996 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
997 continue; /* Let kswapd poll it */
999 sc
->all_unreclaimable
= 0;
1001 nr_reclaimed
+= shrink_zone(priority
, zone
, sc
);
1003 return nr_reclaimed
;
1007 * This is the main entry point to direct page reclaim.
1009 * If a full scan of the inactive list fails to free enough memory then we
1010 * are "out of memory" and something needs to be killed.
1012 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1013 * high - the zone may be full of dirty or under-writeback pages, which this
1014 * caller can't do much about. We kick pdflush and take explicit naps in the
1015 * hope that some of these pages can be written. But if the allocating task
1016 * holds filesystem locks which prevent writeout this might not work, and the
1017 * allocation attempt will fail.
1019 unsigned long try_to_free_pages(struct zone
**zones
, gfp_t gfp_mask
)
1023 unsigned long total_scanned
= 0;
1024 unsigned long nr_reclaimed
= 0;
1025 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1026 unsigned long lru_pages
= 0;
1028 #if 0 // mask by Victor Yu. 03-14-2007
1029 struct scan_control sc
= {
1030 .gfp_mask
= gfp_mask
,
1031 .may_writepage
= !laptop_mode
,
1032 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1034 .swappiness
= vm_swappiness
,
1037 struct scan_control sc
;
1038 sc
.gfp_mask
= gfp_mask
;
1039 sc
.may_writepage
= !laptop_mode
;
1040 sc
.swap_cluster_max
= SWAP_CLUSTER_MAX
;
1042 sc
.swappiness
= vm_swappiness
;
1045 count_vm_event(ALLOCSTALL
);
1047 for (i
= 0; zones
[i
] != NULL
; i
++) {
1048 struct zone
*zone
= zones
[i
];
1050 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1053 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1056 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1059 disable_swap_token();
1060 nr_reclaimed
+= shrink_zones(priority
, zones
, &sc
);
1061 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1062 if (reclaim_state
) {
1063 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1064 reclaim_state
->reclaimed_slab
= 0;
1066 total_scanned
+= sc
.nr_scanned
;
1067 if (nr_reclaimed
>= sc
.swap_cluster_max
) {
1073 * Try to write back as many pages as we just scanned. This
1074 * tends to cause slow streaming writers to write data to the
1075 * disk smoothly, at the dirtying rate, which is nice. But
1076 * that's undesirable in laptop mode, where we *want* lumpy
1077 * writeout. So in laptop mode, write out the whole world.
1079 if (total_scanned
> sc
.swap_cluster_max
+
1080 sc
.swap_cluster_max
/ 2) {
1081 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1082 sc
.may_writepage
= 1;
1085 /* Take a nap, wait for some writeback to complete */
1086 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1087 congestion_wait(WRITE
, HZ
/10);
1089 /* top priority shrink_caches still had more to do? don't OOM, then */
1090 if (!sc
.all_unreclaimable
)
1094 * Now that we've scanned all the zones at this priority level, note
1095 * that level within the zone so that the next thread which performs
1096 * scanning of this zone will immediately start out at this priority
1097 * level. This affects only the decision whether or not to bring
1098 * mapped pages onto the inactive list.
1102 for (i
= 0; zones
[i
] != 0; i
++) {
1103 struct zone
*zone
= zones
[i
];
1105 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1108 zone
->prev_priority
= priority
;
1114 * For kswapd, balance_pgdat() will work across all this node's zones until
1115 * they are all at pages_high.
1117 * Returns the number of pages which were actually freed.
1119 * There is special handling here for zones which are full of pinned pages.
1120 * This can happen if the pages are all mlocked, or if they are all used by
1121 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1122 * What we do is to detect the case where all pages in the zone have been
1123 * scanned twice and there has been zero successful reclaim. Mark the zone as
1124 * dead and from now on, only perform a short scan. Basically we're polling
1125 * the zone for when the problem goes away.
1127 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1128 * zones which have free_pages > pages_high, but once a zone is found to have
1129 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1130 * of the number of free pages in the lower zones. This interoperates with
1131 * the page allocator fallback scheme to ensure that aging of pages is balanced
1134 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
1139 unsigned long total_scanned
;
1140 unsigned long nr_reclaimed
;
1141 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1142 #if 0 // mask by Victor Yu. 03-14-2007
1143 struct scan_control sc
= {
1144 .gfp_mask
= GFP_KERNEL
,
1146 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1147 .swappiness
= vm_swappiness
,
1150 * temp_priority is used to remember the scanning priority at which
1151 * this zone was successfully refilled to free_pages == pages_high.
1153 int temp_priority
[MAX_NR_ZONES
];
1155 int temp_priority
[MAX_NR_ZONES
];
1156 struct scan_control sc
;
1157 sc
.gfp_mask
= GFP_KERNEL
;
1159 sc
.swap_cluster_max
= SWAP_CLUSTER_MAX
;
1160 sc
.swappiness
= vm_swappiness
;
1166 sc
.may_writepage
= !laptop_mode
;
1167 count_vm_event(PAGEOUTRUN
);
1169 for (i
= 0; i
< pgdat
->nr_zones
; i
++)
1170 temp_priority
[i
] = DEF_PRIORITY
;
1172 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1173 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1174 unsigned long lru_pages
= 0;
1176 /* The swap token gets in the way of swapout... */
1178 disable_swap_token();
1183 * Scan in the highmem->dma direction for the highest
1184 * zone which needs scanning
1186 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1187 struct zone
*zone
= pgdat
->node_zones
+ i
;
1189 if (!populated_zone(zone
))
1192 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1195 if (!zone_watermark_ok(zone
, order
, zone
->pages_high
,
1203 for (i
= 0; i
<= end_zone
; i
++) {
1204 struct zone
*zone
= pgdat
->node_zones
+ i
;
1206 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1210 * Now scan the zone in the dma->highmem direction, stopping
1211 * at the last zone which needs scanning.
1213 * We do this because the page allocator works in the opposite
1214 * direction. This prevents the page allocator from allocating
1215 * pages behind kswapd's direction of progress, which would
1216 * cause too much scanning of the lower zones.
1218 for (i
= 0; i
<= end_zone
; i
++) {
1219 struct zone
*zone
= pgdat
->node_zones
+ i
;
1222 if (!populated_zone(zone
))
1225 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1228 if (!zone_watermark_ok(zone
, order
, zone
->pages_high
,
1231 temp_priority
[i
] = priority
;
1233 note_zone_scanning_priority(zone
, priority
);
1234 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1235 reclaim_state
->reclaimed_slab
= 0;
1236 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1238 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1239 total_scanned
+= sc
.nr_scanned
;
1240 if (zone
->all_unreclaimable
)
1242 if (nr_slab
== 0 && zone
->pages_scanned
>=
1243 (zone
->nr_active
+ zone
->nr_inactive
) * 6)
1244 zone
->all_unreclaimable
= 1;
1246 * If we've done a decent amount of scanning and
1247 * the reclaim ratio is low, start doing writepage
1248 * even in laptop mode
1250 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1251 total_scanned
> nr_reclaimed
+ nr_reclaimed
/ 2)
1252 sc
.may_writepage
= 1;
1255 break; /* kswapd: all done */
1257 * OK, kswapd is getting into trouble. Take a nap, then take
1258 * another pass across the zones.
1260 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1261 congestion_wait(WRITE
, HZ
/10);
1264 * We do this so kswapd doesn't build up large priorities for
1265 * example when it is freeing in parallel with allocators. It
1266 * matches the direct reclaim path behaviour in terms of impact
1267 * on zone->*_priority.
1269 if (nr_reclaimed
>= SWAP_CLUSTER_MAX
)
1274 * Note within each zone the priority level at which this zone was
1275 * brought into a happy state. So that the next thread which scans this
1276 * zone will start out at that priority level.
1278 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1279 struct zone
*zone
= pgdat
->node_zones
+ i
;
1281 zone
->prev_priority
= temp_priority
[i
];
1283 if (!all_zones_ok
) {
1288 return nr_reclaimed
;
1292 * The background pageout daemon, started as a kernel thread
1293 * from the init process.
1295 * This basically trickles out pages so that we have _some_
1296 * free memory available even if there is no other activity
1297 * that frees anything up. This is needed for things like routing
1298 * etc, where we otherwise might have all activity going on in
1299 * asynchronous contexts that cannot page things out.
1301 * If there are applications that are active memory-allocators
1302 * (most normal use), this basically shouldn't matter.
1304 static int kswapd(void *p
)
1306 unsigned long order
;
1307 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1308 struct task_struct
*tsk
= current
;
1310 struct reclaim_state reclaim_state
= {
1311 .reclaimed_slab
= 0,
1315 cpumask
= node_to_cpumask(pgdat
->node_id
);
1316 if (!cpus_empty(cpumask
))
1317 set_cpus_allowed(tsk
, cpumask
);
1318 current
->reclaim_state
= &reclaim_state
;
1321 * Tell the memory management that we're a "memory allocator",
1322 * and that if we need more memory we should get access to it
1323 * regardless (see "__alloc_pages()"). "kswapd" should
1324 * never get caught in the normal page freeing logic.
1326 * (Kswapd normally doesn't need memory anyway, but sometimes
1327 * you need a small amount of memory in order to be able to
1328 * page out something else, and this flag essentially protects
1329 * us from recursively trying to free more memory as we're
1330 * trying to free the first piece of memory in the first place).
1332 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1336 unsigned long new_order
;
1340 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1341 new_order
= pgdat
->kswapd_max_order
;
1342 pgdat
->kswapd_max_order
= 0;
1343 if (order
< new_order
) {
1345 * Don't sleep if someone wants a larger 'order'
1351 order
= pgdat
->kswapd_max_order
;
1353 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1355 balance_pgdat(pgdat
, order
);
1361 * A zone is low on free memory, so wake its kswapd task to service it.
1363 void wakeup_kswapd(struct zone
*zone
, int order
)
1367 if (!populated_zone(zone
))
1370 pgdat
= zone
->zone_pgdat
;
1371 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1373 if (pgdat
->kswapd_max_order
< order
)
1374 pgdat
->kswapd_max_order
= order
;
1375 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1377 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1379 wake_up_interruptible(&pgdat
->kswapd_wait
);
1384 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1385 * from LRU lists system-wide, for given pass and priority, and returns the
1386 * number of reclaimed pages
1388 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1390 static unsigned long shrink_all_zones(unsigned long nr_pages
, int pass
,
1391 int prio
, struct scan_control
*sc
)
1394 unsigned long nr_to_scan
, ret
= 0;
1396 for_each_zone(zone
) {
1398 if (!populated_zone(zone
))
1401 if (zone
->all_unreclaimable
&& prio
!= DEF_PRIORITY
)
1404 /* For pass = 0 we don't shrink the active list */
1406 zone
->nr_scan_active
+= (zone
->nr_active
>> prio
) + 1;
1407 if (zone
->nr_scan_active
>= nr_pages
|| pass
> 3) {
1408 zone
->nr_scan_active
= 0;
1409 nr_to_scan
= min(nr_pages
, zone
->nr_active
);
1410 shrink_active_list(nr_to_scan
, zone
, sc
, prio
);
1414 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> prio
) + 1;
1415 if (zone
->nr_scan_inactive
>= nr_pages
|| pass
> 3) {
1416 zone
->nr_scan_inactive
= 0;
1417 nr_to_scan
= min(nr_pages
, zone
->nr_inactive
);
1418 ret
+= shrink_inactive_list(nr_to_scan
, zone
, sc
);
1419 if (ret
>= nr_pages
)
1428 * Try to free `nr_pages' of memory, system-wide, and return the number of
1431 * Rather than trying to age LRUs the aim is to preserve the overall
1432 * LRU order by reclaiming preferentially
1433 * inactive > active > active referenced > active mapped
1435 unsigned long shrink_all_memory(unsigned long nr_pages
)
1437 unsigned long lru_pages
, nr_slab
;
1438 unsigned long ret
= 0;
1440 struct reclaim_state reclaim_state
;
1442 #if 0 // mask by Victor Yu. 03-14-2007
1443 struct scan_control sc
= {
1444 .gfp_mask
= GFP_KERNEL
,
1446 .swap_cluster_max
= nr_pages
,
1448 .swappiness
= vm_swappiness
,
1451 struct scan_control sc
;
1452 sc
.gfp_mask
= GFP_KERNEL
;
1454 sc
.swap_cluster_max
= nr_pages
;
1455 sc
.may_writepage
= 1;
1456 sc
.swappiness
= vm_swappiness
;
1459 current
->reclaim_state
= &reclaim_state
;
1463 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1465 nr_slab
= global_page_state(NR_SLAB_RECLAIMABLE
);
1466 /* If slab caches are huge, it's better to hit them first */
1467 while (nr_slab
>= lru_pages
) {
1468 reclaim_state
.reclaimed_slab
= 0;
1469 shrink_slab(nr_pages
, sc
.gfp_mask
, lru_pages
);
1470 if (!reclaim_state
.reclaimed_slab
)
1473 ret
+= reclaim_state
.reclaimed_slab
;
1474 if (ret
>= nr_pages
)
1477 nr_slab
-= reclaim_state
.reclaimed_slab
;
1481 * We try to shrink LRUs in 5 passes:
1482 * 0 = Reclaim from inactive_list only
1483 * 1 = Reclaim from active list but don't reclaim mapped
1484 * 2 = 2nd pass of type 1
1485 * 3 = Reclaim mapped (normal reclaim)
1486 * 4 = 2nd pass of type 3
1488 for (pass
= 0; pass
< 5; pass
++) {
1491 /* Needed for shrinking slab caches later on */
1493 for_each_zone(zone
) {
1494 lru_pages
+= zone
->nr_active
;
1495 lru_pages
+= zone
->nr_inactive
;
1498 /* Force reclaiming mapped pages in the passes #3 and #4 */
1501 sc
.swappiness
= 100;
1504 for (prio
= DEF_PRIORITY
; prio
>= 0; prio
--) {
1505 unsigned long nr_to_scan
= nr_pages
- ret
;
1508 ret
+= shrink_all_zones(nr_to_scan
, prio
, pass
, &sc
);
1509 if (ret
>= nr_pages
)
1512 reclaim_state
.reclaimed_slab
= 0;
1513 shrink_slab(sc
.nr_scanned
, sc
.gfp_mask
, lru_pages
);
1514 ret
+= reclaim_state
.reclaimed_slab
;
1515 if (ret
>= nr_pages
)
1518 if (sc
.nr_scanned
&& prio
< DEF_PRIORITY
- 2)
1519 congestion_wait(WRITE
, HZ
/ 10);
1526 * If ret = 0, we could not shrink LRUs, but there may be something
1531 reclaim_state
.reclaimed_slab
= 0;
1532 shrink_slab(nr_pages
, sc
.gfp_mask
, lru_pages
);
1533 ret
+= reclaim_state
.reclaimed_slab
;
1534 } while (ret
< nr_pages
&& reclaim_state
.reclaimed_slab
> 0);
1537 current
->reclaim_state
= NULL
;
1543 #ifdef CONFIG_HOTPLUG_CPU
1544 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1545 not required for correctness. So if the last cpu in a node goes
1546 away, we get changed to run anywhere: as the first one comes back,
1547 restore their cpu bindings. */
1548 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
1549 unsigned long action
, void *hcpu
)
1554 if (action
== CPU_ONLINE
) {
1555 for_each_online_pgdat(pgdat
) {
1556 mask
= node_to_cpumask(pgdat
->node_id
);
1557 if (any_online_cpu(mask
) != NR_CPUS
)
1558 /* One of our CPUs online: restore mask */
1559 set_cpus_allowed(pgdat
->kswapd
, mask
);
1564 #endif /* CONFIG_HOTPLUG_CPU */
1567 * This kswapd start function will be called by init and node-hot-add.
1568 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1570 int kswapd_run(int nid
)
1572 pg_data_t
*pgdat
= NODE_DATA(nid
);
1578 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
1579 if (IS_ERR(pgdat
->kswapd
)) {
1580 /* failure at boot is fatal */
1581 BUG_ON(system_state
== SYSTEM_BOOTING
);
1582 printk("Failed to start kswapd on node %d\n",nid
);
1588 static int __init
kswapd_init(void)
1593 for_each_online_node(nid
)
1595 hotcpu_notifier(cpu_callback
, 0);
1599 module_init(kswapd_init
)
1605 * If non-zero call zone_reclaim when the number of free pages falls below
1608 int zone_reclaim_mode __read_mostly
;
1610 #define RECLAIM_OFF 0
1611 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1612 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1613 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1616 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1617 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1620 #define ZONE_RECLAIM_PRIORITY 4
1623 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1626 int sysctl_min_unmapped_ratio
= 1;
1629 * If the number of slab pages in a zone grows beyond this percentage then
1630 * slab reclaim needs to occur.
1632 int sysctl_min_slab_ratio
= 5;
1635 * Try to free up some pages from this zone through reclaim.
1637 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1639 /* Minimum pages needed in order to stay on node */
1640 const unsigned long nr_pages
= 1 << order
;
1641 struct task_struct
*p
= current
;
1642 struct reclaim_state reclaim_state
;
1644 unsigned long nr_reclaimed
= 0;
1645 #if 0 // mask by Victor Yu. 03-14-2007
1646 struct scan_control sc
= {
1647 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
1648 .may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
1649 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
1651 .gfp_mask
= gfp_mask
,
1652 .swappiness
= vm_swappiness
,
1654 unsigned long slab_reclaimable
;
1656 unsigned long slab_reclaimable
;
1657 struct scan_control sc
;
1658 sc
.may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
);
1659 sc
.may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
);
1660 sc
.swap_cluster_max
= max_t(unsigned long, nr_pages
, SWAP_CLUSTER_MAX
);
1661 sc
.gfp_mask
= gfp_mask
;
1662 sc
.swappiness
= vm_swappiness
;
1665 disable_swap_token();
1668 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1669 * and we also need to be able to write out pages for RECLAIM_WRITE
1672 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
1673 reclaim_state
.reclaimed_slab
= 0;
1674 p
->reclaim_state
= &reclaim_state
;
1676 if (zone_page_state(zone
, NR_FILE_PAGES
) -
1677 zone_page_state(zone
, NR_FILE_MAPPED
) >
1678 zone
->min_unmapped_pages
) {
1680 * Free memory by calling shrink zone with increasing
1681 * priorities until we have enough memory freed.
1683 priority
= ZONE_RECLAIM_PRIORITY
;
1685 note_zone_scanning_priority(zone
, priority
);
1686 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1688 } while (priority
>= 0 && nr_reclaimed
< nr_pages
);
1691 slab_reclaimable
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
1692 if (slab_reclaimable
> zone
->min_slab_pages
) {
1694 * shrink_slab() does not currently allow us to determine how
1695 * many pages were freed in this zone. So we take the current
1696 * number of slab pages and shake the slab until it is reduced
1697 * by the same nr_pages that we used for reclaiming unmapped
1700 * Note that shrink_slab will free memory on all zones and may
1703 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
1704 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) >
1705 slab_reclaimable
- nr_pages
)
1709 * Update nr_reclaimed by the number of slab pages we
1710 * reclaimed from this zone.
1712 nr_reclaimed
+= slab_reclaimable
-
1713 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
1716 p
->reclaim_state
= NULL
;
1717 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
1718 return nr_reclaimed
>= nr_pages
;
1721 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1727 * Zone reclaim reclaims unmapped file backed pages and
1728 * slab pages if we are over the defined limits.
1730 * A small portion of unmapped file backed pages is needed for
1731 * file I/O otherwise pages read by file I/O will be immediately
1732 * thrown out if the zone is overallocated. So we do not reclaim
1733 * if less than a specified percentage of the zone is used by
1734 * unmapped file backed pages.
1736 if (zone_page_state(zone
, NR_FILE_PAGES
) -
1737 zone_page_state(zone
, NR_FILE_MAPPED
) <= zone
->min_unmapped_pages
1738 && zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)
1739 <= zone
->min_slab_pages
)
1743 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1744 * not have reclaimable pages and if we should not delay the allocation
1747 if (!(gfp_mask
& __GFP_WAIT
) ||
1748 zone
->all_unreclaimable
||
1749 atomic_read(&zone
->reclaim_in_progress
) > 0 ||
1750 (current
->flags
& PF_MEMALLOC
))
1754 * Only run zone reclaim on the local zone or on zones that do not
1755 * have associated processors. This will favor the local processor
1756 * over remote processors and spread off node memory allocations
1757 * as wide as possible.
1759 node_id
= zone_to_nid(zone
);
1760 mask
= node_to_cpumask(node_id
);
1761 if (!cpus_empty(mask
) && node_id
!= numa_node_id())
1763 return __zone_reclaim(zone
, gfp_mask
, order
);