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/slab.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/init.h>
20 #include <linux/highmem.h>
21 #include <linux/file.h>
22 #include <linux/writeback.h>
23 #include <linux/suspend.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h> /* for try_to_release_page() */
26 #include <linux/mm_inline.h>
27 #include <linux/pagevec.h>
28 #include <linux/backing-dev.h>
29 #include <linux/rmap-locking.h>
31 #include <asm/pgalloc.h>
32 #include <asm/tlbflush.h>
33 #include <asm/topology.h>
34 #include <asm/div64.h>
36 #include <linux/swapops.h>
39 * The "priority" of VM scanning is how much of the queues we will scan in one
40 * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the
41 * queues ("queue_length >> 12") during an aging round.
43 #define DEF_PRIORITY 12
46 * From 0 .. 100. Higher means more swappy.
48 int vm_swappiness
= 60;
49 static long total_memory
;
51 #ifdef ARCH_HAS_PREFETCH
52 #define prefetch_prev_lru_page(_page, _base, _field) \
54 if ((_page)->lru.prev != _base) { \
57 prev = list_entry(_page->lru.prev, \
59 prefetch(&prev->_field); \
63 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
66 #ifdef ARCH_HAS_PREFETCHW
67 #define prefetchw_prev_lru_page(_page, _base, _field) \
69 if ((_page)->lru.prev != _base) { \
72 prev = list_entry(_page->lru.prev, \
74 prefetchw(&prev->_field); \
78 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
82 * The list of shrinker callbacks used by to apply pressure to
87 struct list_head list
;
88 int seeks
; /* seeks to recreate an obj */
89 long nr
; /* objs pending delete */
92 static LIST_HEAD(shrinker_list
);
93 static DECLARE_MUTEX(shrinker_sem
);
96 * Add a shrinker callback to be called from the vm
98 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
100 struct shrinker
*shrinker
;
102 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
104 shrinker
->shrinker
= theshrinker
;
105 shrinker
->seeks
= seeks
;
108 list_add(&shrinker
->list
, &shrinker_list
);
117 void remove_shrinker(struct shrinker
*shrinker
)
120 list_del(&shrinker
->list
);
125 #define SHRINK_BATCH 128
127 * Call the shrink functions to age shrinkable caches
129 * Here we assume it costs one seek to replace a lru page and that it also
130 * takes a seek to recreate a cache object. With this in mind we age equal
131 * percentages of the lru and ageable caches. This should balance the seeks
132 * generated by these structures.
134 * If the vm encounted mapped pages on the LRU it increase the pressure on
135 * slab to avoid swapping.
137 * FIXME: do not do for zone highmem
139 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
141 static int shrink_slab(long scanned
, unsigned int gfp_mask
)
143 struct shrinker
*shrinker
;
146 if (down_trylock(&shrinker_sem
))
149 pages
= nr_used_zone_pages();
150 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
153 delta
= scanned
* shrinker
->seeks
;
154 delta
*= (*shrinker
->shrinker
)(0, gfp_mask
);
155 do_div(delta
, pages
+ 1);
156 shrinker
->nr
+= delta
;
157 if (shrinker
->nr
> SHRINK_BATCH
) {
158 long nr_to_scan
= shrinker
->nr
;
162 long this_scan
= nr_to_scan
;
166 (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
167 nr_to_scan
-= this_scan
;
176 /* Must be called with page's pte_chain_lock held. */
177 static inline int page_mapping_inuse(struct page
*page
)
179 struct address_space
*mapping
= page
->mapping
;
181 /* Page is in somebody's page tables. */
182 if (page_mapped(page
))
185 /* XXX: does this happen ? */
189 /* Be more reluctant to reclaim swapcache than pagecache */
190 if (PageSwapCache(page
))
193 /* File is mmap'd by somebody. */
194 if (!list_empty(&mapping
->i_mmap
))
196 if (!list_empty(&mapping
->i_mmap_shared
))
202 static inline int is_page_cache_freeable(struct page
*page
)
204 return page_count(page
) - !!PagePrivate(page
) == 2;
207 static int may_write_to_queue(struct backing_dev_info
*bdi
)
209 if (current_is_kswapd())
211 if (current_is_pdflush()) /* This is unlikely, but why not... */
213 if (!bdi_write_congested(bdi
))
215 if (bdi
== current
->backing_dev_info
)
221 * shrink_list returns the number of reclaimed pages
224 shrink_list(struct list_head
*page_list
, unsigned int gfp_mask
,
225 int *max_scan
, int *nr_mapped
)
227 struct address_space
*mapping
;
228 LIST_HEAD(ret_pages
);
229 struct pagevec freed_pvec
;
235 pagevec_init(&freed_pvec
, 1);
236 while (!list_empty(page_list
)) {
240 page
= list_entry(page_list
->prev
, struct page
, lru
);
241 list_del(&page
->lru
);
243 if (TestSetPageLocked(page
))
246 /* Double the slab pressure for mapped and swapcache pages */
247 if (page_mapped(page
) || PageSwapCache(page
))
250 BUG_ON(PageActive(page
));
251 may_enter_fs
= (gfp_mask
& __GFP_FS
) ||
252 (PageSwapCache(page
) && (gfp_mask
& __GFP_IO
));
254 if (PageWriteback(page
))
257 pte_chain_lock(page
);
258 if (page_referenced(page
) && page_mapping_inuse(page
)) {
259 /* In active use or really unfreeable. Activate it. */
260 pte_chain_unlock(page
);
261 goto activate_locked
;
264 mapping
= page
->mapping
;
268 * Anonymous process memory without backing store. Try to
269 * allocate it some swap space here.
271 * XXX: implement swap clustering ?
273 if (page_mapped(page
) && !mapping
&& !PagePrivate(page
)) {
274 pte_chain_unlock(page
);
275 if (!add_to_swap(page
))
276 goto activate_locked
;
277 pte_chain_lock(page
);
278 mapping
= page
->mapping
;
282 * The page is mapped into the page tables of one or more
283 * processes. Try to unmap it here.
285 if (page_mapped(page
) && mapping
) {
286 switch (try_to_unmap(page
)) {
289 pte_chain_unlock(page
);
290 goto activate_locked
;
292 pte_chain_unlock(page
);
295 ; /* try to free the page below */
298 #endif /* CONFIG_SWAP */
299 pte_chain_unlock(page
);
302 * If the page is dirty, only perform writeback if that write
303 * will be non-blocking. To prevent this allocation from being
304 * stalled by pagecache activity. But note that there may be
305 * stalls if we need to run get_block(). We could test
306 * PagePrivate for that.
308 * If this process is currently in generic_file_write() against
309 * this page's queue, we can perform writeback even if that
312 * If the page is swapcache, write it back even if that would
313 * block, for some throttling. This happens by accident, because
314 * swap_backing_dev_info is bust: it doesn't reflect the
315 * congestion state of the swapdevs. Easy to fix, if needed.
316 * See swapfile.c:page_queue_congested().
318 if (PageDirty(page
)) {
319 if (!is_page_cache_freeable(page
))
323 if (mapping
->a_ops
->writepage
== NULL
)
324 goto activate_locked
;
327 if (!may_write_to_queue(mapping
->backing_dev_info
))
329 write_lock(&mapping
->page_lock
);
330 if (test_clear_page_dirty(page
)) {
332 struct writeback_control wbc
= {
333 .sync_mode
= WB_SYNC_NONE
,
334 .nr_to_write
= SWAP_CLUSTER_MAX
,
339 list_move(&page
->list
, &mapping
->locked_pages
);
340 write_unlock(&mapping
->page_lock
);
342 SetPageReclaim(page
);
343 res
= mapping
->a_ops
->writepage(page
, &wbc
);
345 if (res
== WRITEPAGE_ACTIVATE
) {
346 ClearPageReclaim(page
);
347 goto activate_locked
;
349 if (!PageWriteback(page
)) {
350 /* synchronous write or broken a_ops? */
351 ClearPageReclaim(page
);
355 write_unlock(&mapping
->page_lock
);
359 * If the page has buffers, try to free the buffer mappings
360 * associated with this page. If we succeed we try to free
363 * We do this even if the page is PageDirty().
364 * try_to_release_page() does not perform I/O, but it is
365 * possible for a page to have PageDirty set, but it is actually
366 * clean (all its buffers are clean). This happens if the
367 * buffers were written out directly, with submit_bh(). ext3
368 * will do this, as well as the blockdev mapping.
369 * try_to_release_page() will discover that cleanness and will
370 * drop the buffers and mark the page clean - it can be freed.
372 * Rarely, pages can have buffers and no ->mapping. These are
373 * the pages which were not successfully invalidated in
374 * truncate_complete_page(). We try to drop those buffers here
375 * and if that worked, and the page is no longer mapped into
376 * process address space (page_count == 0) it can be freed.
377 * Otherwise, leave the page on the LRU so it is swappable.
379 if (PagePrivate(page
)) {
380 if (!try_to_release_page(page
, gfp_mask
))
381 goto activate_locked
;
382 if (!mapping
&& page_count(page
) == 1)
387 goto keep_locked
; /* truncate got there first */
389 write_lock(&mapping
->page_lock
);
392 * The non-racy check for busy page. It is critical to check
393 * PageDirty _after_ making sure that the page is freeable and
394 * not in use by anybody. (pagecache + us == 2)
396 if (page_count(page
) != 2 || PageDirty(page
)) {
397 write_unlock(&mapping
->page_lock
);
402 if (PageSwapCache(page
)) {
403 swp_entry_t swap
= { .val
= page
->index
};
404 __delete_from_swap_cache(page
);
405 write_unlock(&mapping
->page_lock
);
407 __put_page(page
); /* The pagecache ref */
410 #endif /* CONFIG_SWAP */
412 __remove_from_page_cache(page
);
413 write_unlock(&mapping
->page_lock
);
419 if (!pagevec_add(&freed_pvec
, page
))
420 __pagevec_release_nonlru(&freed_pvec
);
429 list_add(&page
->lru
, &ret_pages
);
430 BUG_ON(PageLRU(page
));
432 list_splice(&ret_pages
, page_list
);
433 if (pagevec_count(&freed_pvec
))
434 __pagevec_release_nonlru(&freed_pvec
);
435 mod_page_state(pgsteal
, ret
);
436 if (current_is_kswapd())
437 mod_page_state(kswapd_steal
, ret
);
438 mod_page_state(pgactivate
, pgactivate
);
443 * zone->lru_lock is heavily contented. We relieve it by quickly privatising
444 * a batch of pages and working on them outside the lock. Any pages which were
445 * not freed will be added back to the LRU.
447 * shrink_cache() is passed the number of pages to try to free, and returns
448 * the number of pages which were reclaimed.
450 * For pagecache intensive workloads, the first loop here is the hottest spot
451 * in the kernel (apart from the copy_*_user functions).
454 shrink_cache(const int nr_pages
, struct zone
*zone
,
455 unsigned int gfp_mask
, int max_scan
, int *nr_mapped
)
457 LIST_HEAD(page_list
);
463 * Try to ensure that we free `nr_pages' pages in one pass of the loop.
465 nr_to_process
= nr_pages
;
466 if (nr_to_process
< SWAP_CLUSTER_MAX
)
467 nr_to_process
= SWAP_CLUSTER_MAX
;
469 pagevec_init(&pvec
, 1);
472 spin_lock_irq(&zone
->lru_lock
);
473 while (max_scan
> 0 && ret
< nr_pages
) {
479 while (nr_scan
++ < nr_to_process
&&
480 !list_empty(&zone
->inactive_list
)) {
481 page
= list_entry(zone
->inactive_list
.prev
,
484 prefetchw_prev_lru_page(page
,
485 &zone
->inactive_list
, flags
);
487 if (!TestClearPageLRU(page
))
489 list_del(&page
->lru
);
490 if (page_count(page
) == 0) {
491 /* It is currently in pagevec_release() */
493 list_add(&page
->lru
, &zone
->inactive_list
);
496 list_add(&page
->lru
, &page_list
);
497 page_cache_get(page
);
500 zone
->nr_inactive
-= nr_taken
;
501 zone
->pages_scanned
+= nr_taken
;
502 spin_unlock_irq(&zone
->lru_lock
);
508 mod_page_state(pgscan
, nr_scan
);
509 nr_freed
= shrink_list(&page_list
, gfp_mask
,
510 &max_scan
, nr_mapped
);
512 if (nr_freed
<= 0 && list_empty(&page_list
))
515 spin_lock_irq(&zone
->lru_lock
);
517 * Put back any unfreeable pages.
519 while (!list_empty(&page_list
)) {
520 page
= list_entry(page_list
.prev
, struct page
, lru
);
521 if (TestSetPageLRU(page
))
523 list_del(&page
->lru
);
524 if (PageActive(page
))
525 add_page_to_active_list(zone
, page
);
527 add_page_to_inactive_list(zone
, page
);
528 if (!pagevec_add(&pvec
, page
)) {
529 spin_unlock_irq(&zone
->lru_lock
);
530 __pagevec_release(&pvec
);
531 spin_lock_irq(&zone
->lru_lock
);
535 spin_unlock_irq(&zone
->lru_lock
);
537 pagevec_release(&pvec
);
542 * This moves pages from the active list to the inactive list.
544 * We move them the other way if the page is referenced by one or more
545 * processes, from rmap.
547 * If the pages are mostly unmapped, the processing is fast and it is
548 * appropriate to hold zone->lru_lock across the whole operation. But if
549 * the pages are mapped, the processing is slow (page_referenced()) so we
550 * should drop zone->lru_lock around each page. It's impossible to balance
551 * this, so instead we remove the pages from the LRU while processing them.
552 * It is safe to rely on PG_active against the non-LRU pages in here because
553 * nobody will play with that bit on a non-LRU page.
555 * The downside is that we have to touch page->count against each page.
556 * But we had to alter page->flags anyway.
559 refill_inactive_zone(struct zone
*zone
, const int nr_pages_in
,
560 struct page_state
*ps
, int priority
)
562 int pgdeactivate
= 0;
563 int nr_pages
= nr_pages_in
;
564 LIST_HEAD(l_hold
); /* The pages which were snipped off */
565 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
566 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
569 int reclaim_mapped
= 0;
575 spin_lock_irq(&zone
->lru_lock
);
576 while (nr_pages
&& !list_empty(&zone
->active_list
)) {
577 page
= list_entry(zone
->active_list
.prev
, struct page
, lru
);
578 prefetchw_prev_lru_page(page
, &zone
->active_list
, flags
);
579 if (!TestClearPageLRU(page
))
581 list_del(&page
->lru
);
582 if (page_count(page
) == 0) {
583 /* It is currently in pagevec_release() */
585 list_add(&page
->lru
, &zone
->active_list
);
587 page_cache_get(page
);
588 list_add(&page
->lru
, &l_hold
);
592 spin_unlock_irq(&zone
->lru_lock
);
595 * `distress' is a measure of how much trouble we're having reclaiming
596 * pages. 0 -> no problems. 100 -> great trouble.
598 distress
= 100 >> priority
;
601 * The point of this algorithm is to decide when to start reclaiming
602 * mapped memory instead of just pagecache. Work out how much memory
605 mapped_ratio
= (ps
->nr_mapped
* 100) / total_memory
;
608 * Now decide how much we really want to unmap some pages. The mapped
609 * ratio is downgraded - just because there's a lot of mapped memory
610 * doesn't necessarily mean that page reclaim isn't succeeding.
612 * The distress ratio is important - we don't want to start going oom.
614 * A 100% value of vm_swappiness overrides this algorithm altogether.
616 swap_tendency
= mapped_ratio
/ 2 + distress
+ vm_swappiness
;
619 * Now use this metric to decide whether to start moving mapped memory
620 * onto the inactive list.
622 if (swap_tendency
>= 100)
625 while (!list_empty(&l_hold
)) {
626 page
= list_entry(l_hold
.prev
, struct page
, lru
);
627 list_del(&page
->lru
);
628 if (page_mapped(page
)) {
629 pte_chain_lock(page
);
630 if (page_mapped(page
) && page_referenced(page
)) {
631 pte_chain_unlock(page
);
632 list_add(&page
->lru
, &l_active
);
635 pte_chain_unlock(page
);
636 if (!reclaim_mapped
) {
637 list_add(&page
->lru
, &l_active
);
642 * FIXME: need to consider page_count(page) here if/when we
643 * reap orphaned pages via the LRU (Daniel's locking stuff)
645 if (total_swap_pages
== 0 && !page
->mapping
&&
646 !PagePrivate(page
)) {
647 list_add(&page
->lru
, &l_active
);
650 list_add(&page
->lru
, &l_inactive
);
654 pagevec_init(&pvec
, 1);
655 spin_lock_irq(&zone
->lru_lock
);
656 while (!list_empty(&l_inactive
)) {
657 page
= list_entry(l_inactive
.prev
, struct page
, lru
);
658 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
659 if (TestSetPageLRU(page
))
661 if (!TestClearPageActive(page
))
663 list_move(&page
->lru
, &zone
->inactive_list
);
664 if (!pagevec_add(&pvec
, page
)) {
665 spin_unlock_irq(&zone
->lru_lock
);
666 if (buffer_heads_over_limit
)
667 pagevec_strip(&pvec
);
668 __pagevec_release(&pvec
);
669 spin_lock_irq(&zone
->lru_lock
);
672 if (buffer_heads_over_limit
) {
673 spin_unlock_irq(&zone
->lru_lock
);
674 pagevec_strip(&pvec
);
675 spin_lock_irq(&zone
->lru_lock
);
677 while (!list_empty(&l_active
)) {
678 page
= list_entry(l_active
.prev
, struct page
, lru
);
679 prefetchw_prev_lru_page(page
, &l_active
, flags
);
680 if (TestSetPageLRU(page
))
682 BUG_ON(!PageActive(page
));
683 list_move(&page
->lru
, &zone
->active_list
);
684 if (!pagevec_add(&pvec
, page
)) {
685 spin_unlock_irq(&zone
->lru_lock
);
686 __pagevec_release(&pvec
);
687 spin_lock_irq(&zone
->lru_lock
);
690 zone
->nr_active
-= pgdeactivate
;
691 zone
->nr_inactive
+= pgdeactivate
;
692 spin_unlock_irq(&zone
->lru_lock
);
693 pagevec_release(&pvec
);
695 mod_page_state(pgrefill
, nr_pages_in
- nr_pages
);
696 mod_page_state(pgdeactivate
, pgdeactivate
);
700 * Try to reclaim `nr_pages' from this zone. Returns the number of reclaimed
701 * pages. This is a basic per-zone page freer. Used by both kswapd and
705 shrink_zone(struct zone
*zone
, int max_scan
, unsigned int gfp_mask
,
706 const int nr_pages
, int *nr_mapped
, struct page_state
*ps
, int priority
)
711 * Try to keep the active list 2/3 of the size of the cache. And
712 * make sure that refill_inactive is given a decent number of pages.
714 * The "ratio+1" here is important. With pagecache-intensive workloads
715 * the inactive list is huge, and `ratio' evaluates to zero all the
716 * time. Which pins the active list memory. So we add one to `ratio'
717 * just to make sure that the kernel will slowly sift through the
720 ratio
= (unsigned long)nr_pages
* zone
->nr_active
/
721 ((zone
->nr_inactive
| 1) * 2);
722 atomic_add(ratio
+1, &zone
->refill_counter
);
723 if (atomic_read(&zone
->refill_counter
) > SWAP_CLUSTER_MAX
) {
727 * Don't try to bring down too many pages in one attempt.
728 * If this fails, the caller will increase `priority' and
729 * we'll try again, with an increased chance of reclaiming
732 count
= atomic_read(&zone
->refill_counter
);
733 if (count
> SWAP_CLUSTER_MAX
* 4)
734 count
= SWAP_CLUSTER_MAX
* 4;
735 atomic_sub(count
, &zone
->refill_counter
);
736 refill_inactive_zone(zone
, count
, ps
, priority
);
738 return shrink_cache(nr_pages
, zone
, gfp_mask
,
739 max_scan
, nr_mapped
);
743 * This is the direct reclaim path, for page-allocating processes. We only
744 * try to reclaim pages from zones which will satisfy the caller's allocation
747 * We reclaim from a zone even if that zone is over pages_high. Because:
748 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
750 * b) The zones may be over pages_high but they must go *over* pages_high to
751 * satisfy the `incremental min' zone defense algorithm.
753 * Returns the number of reclaimed pages.
755 * If a zone is deemed to be full of pinned pages then just give it a light
756 * scan then give up on it.
759 shrink_caches(struct zone
*classzone
, int priority
, int *total_scanned
,
760 int gfp_mask
, const int nr_pages
, struct page_state
*ps
)
762 struct zone
*first_classzone
;
766 first_classzone
= classzone
->zone_pgdat
->node_zones
;
767 for (zone
= classzone
; zone
>= first_classzone
; zone
--) {
768 int to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
);
772 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
773 continue; /* Let kswapd poll it */
776 * If we cannot reclaim `nr_pages' pages by scanning twice
777 * that many pages then fall back to the next zone.
779 max_scan
= zone
->nr_inactive
>> priority
;
780 if (max_scan
< to_reclaim
* 2)
781 max_scan
= to_reclaim
* 2;
782 ret
+= shrink_zone(zone
, max_scan
, gfp_mask
,
783 to_reclaim
, &nr_mapped
, ps
, priority
);
784 *total_scanned
+= max_scan
+ nr_mapped
;
792 * This is the main entry point to direct page reclaim.
794 * If a full scan of the inactive list fails to free enough memory then we
795 * are "out of memory" and something needs to be killed.
797 * If the caller is !__GFP_FS then the probability of a failure is reasonably
798 * high - the zone may be full of dirty or under-writeback pages, which this
799 * caller can't do much about. So for !__GFP_FS callers, we just perform a
800 * small LRU walk and if that didn't work out, fail the allocation back to the
801 * caller. GFP_NOFS allocators need to know how to deal with it. Kicking
802 * bdflush, waiting and retrying will work.
804 * This is a fairly lame algorithm - it can result in excessive CPU burning and
805 * excessive rotation of the inactive list, which is _supposed_ to be an LRU,
809 try_to_free_pages(struct zone
*classzone
,
810 unsigned int gfp_mask
, unsigned int order
)
813 const int nr_pages
= SWAP_CLUSTER_MAX
;
814 int nr_reclaimed
= 0;
816 inc_page_state(allocstall
);
818 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
819 int total_scanned
= 0;
820 struct page_state ps
;
823 nr_reclaimed
+= shrink_caches(classzone
, priority
,
824 &total_scanned
, gfp_mask
,
826 if (nr_reclaimed
>= nr_pages
)
828 if (!(gfp_mask
& __GFP_FS
))
829 break; /* Let the caller handle it */
831 * Try to write back as many pages as we just scanned. Not
832 * sure if that makes sense, but it's an attempt to avoid
833 * creating IO storms unnecessarily
835 wakeup_bdflush(total_scanned
);
837 /* Take a nap, wait for some writeback to complete */
838 blk_congestion_wait(WRITE
, HZ
/4);
839 shrink_slab(total_scanned
, gfp_mask
);
841 if (gfp_mask
& __GFP_FS
)
847 * For kswapd, balance_pgdat() will work across all this node's zones until
848 * they are all at pages_high.
850 * If `nr_pages' is non-zero then it is the number of pages which are to be
851 * reclaimed, regardless of the zone occupancies. This is a software suspend
854 * Returns the number of pages which were actually freed.
856 * There is special handling here for zones which are full of pinned pages.
857 * This can happen if the pages are all mlocked, or if they are all used by
858 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
859 * What we do is to detect the case where all pages in the zone have been
860 * scanned twice and there has been zero successful reclaim. Mark the zone as
861 * dead and from now on, only perform a short scan. Basically we're polling
862 * the zone for when the problem goes away.
864 static int balance_pgdat(pg_data_t
*pgdat
, int nr_pages
, struct page_state
*ps
)
866 int to_free
= nr_pages
;
870 inc_page_state(pageoutrun
);
872 for (priority
= DEF_PRIORITY
; priority
; priority
--) {
873 int all_zones_ok
= 1;
875 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
876 struct zone
*zone
= pgdat
->node_zones
+ i
;
881 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
884 if (nr_pages
&& to_free
> 0) { /* Software suspend */
885 to_reclaim
= min(to_free
, SWAP_CLUSTER_MAX
*8);
886 } else { /* Zone balancing */
887 to_reclaim
= zone
->pages_high
-zone
->free_pages
;
892 max_scan
= zone
->nr_inactive
>> priority
;
893 if (max_scan
< to_reclaim
* 2)
894 max_scan
= to_reclaim
* 2;
895 if (max_scan
< SWAP_CLUSTER_MAX
)
896 max_scan
= SWAP_CLUSTER_MAX
;
897 to_free
-= shrink_zone(zone
, max_scan
, GFP_KSWAPD
,
898 to_reclaim
, &nr_mapped
, ps
, priority
);
899 shrink_slab(max_scan
+ nr_mapped
, GFP_KSWAPD
);
900 if (zone
->all_unreclaimable
)
902 if (zone
->pages_scanned
> zone
->present_pages
* 2)
903 zone
->all_unreclaimable
= 1;
907 blk_congestion_wait(WRITE
, HZ
/4);
909 return nr_pages
- to_free
;
913 * The background pageout daemon, started as a kernel thread
914 * from the init process.
916 * This basically trickles out pages so that we have _some_
917 * free memory available even if there is no other activity
918 * that frees anything up. This is needed for things like routing
919 * etc, where we otherwise might have all activity going on in
920 * asynchronous contexts that cannot page things out.
922 * If there are applications that are active memory-allocators
923 * (most normal use), this basically shouldn't matter.
927 pg_data_t
*pgdat
= (pg_data_t
*)p
;
928 struct task_struct
*tsk
= current
;
932 set_cpus_allowed(tsk
, __node_to_cpu_mask(pgdat
->node_id
));
933 sprintf(tsk
->comm
, "kswapd%d", pgdat
->node_id
);
934 sigfillset(&tsk
->blocked
);
937 * Tell the memory management that we're a "memory allocator",
938 * and that if we need more memory we should get access to it
939 * regardless (see "__alloc_pages()"). "kswapd" should
940 * never get caught in the normal page freeing logic.
942 * (Kswapd normally doesn't need memory anyway, but sometimes
943 * you need a small amount of memory in order to be able to
944 * page out something else, and this flag essentially protects
945 * us from recursively trying to free more memory as we're
946 * trying to free the first piece of memory in the first place).
948 tsk
->flags
|= PF_MEMALLOC
|PF_KSWAPD
;
951 struct page_state ps
;
953 if (current
->flags
& PF_FREEZE
)
954 refrigerator(PF_IOTHREAD
);
955 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
957 finish_wait(&pgdat
->kswapd_wait
, &wait
);
959 balance_pgdat(pgdat
, 0, &ps
);
965 * A zone is low on free memory, so wake its kswapd task to service it.
967 void wakeup_kswapd(struct zone
*zone
)
969 if (zone
->free_pages
> zone
->pages_low
)
971 if (!waitqueue_active(&zone
->zone_pgdat
->kswapd_wait
))
973 wake_up_interruptible(&zone
->zone_pgdat
->kswapd_wait
);
976 #ifdef CONFIG_SOFTWARE_SUSPEND
978 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
981 int shrink_all_memory(int nr_pages
)
984 int nr_to_free
= nr_pages
;
987 for_each_pgdat(pgdat
) {
989 struct page_state ps
;
992 freed
= balance_pgdat(pgdat
, nr_to_free
, &ps
);
1002 static int __init
kswapd_init(void)
1006 for_each_pgdat(pgdat
)
1007 kernel_thread(kswapd
, pgdat
, CLONE_KERNEL
);
1008 total_memory
= nr_free_pagecache_pages();
1012 module_init(kswapd_init
)