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>
39 #include <linux/freezer.h>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
44 #include <linux/swapops.h>
49 /* Incremented by the number of inactive pages that were scanned */
50 unsigned long nr_scanned
;
52 /* This context's GFP mask */
57 /* Can pages be swapped as part of reclaim? */
60 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
61 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
62 * In this context, it doesn't matter that we scan the
63 * whole list at once. */
68 int all_unreclaimable
;
73 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
75 #ifdef ARCH_HAS_PREFETCH
76 #define prefetch_prev_lru_page(_page, _base, _field) \
78 if ((_page)->lru.prev != _base) { \
81 prev = lru_to_page(&(_page->lru)); \
82 prefetch(&prev->_field); \
86 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
89 #ifdef ARCH_HAS_PREFETCHW
90 #define prefetchw_prev_lru_page(_page, _base, _field) \
92 if ((_page)->lru.prev != _base) { \
95 prev = lru_to_page(&(_page->lru)); \
96 prefetchw(&prev->_field); \
100 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
104 * From 0 .. 100. Higher means more swappy.
106 int vm_swappiness
= 60;
107 long vm_total_pages
; /* The total number of pages which the VM controls */
109 static LIST_HEAD(shrinker_list
);
110 static DECLARE_RWSEM(shrinker_rwsem
);
113 * Add a shrinker callback to be called from the vm
115 void register_shrinker(struct shrinker
*shrinker
)
118 down_write(&shrinker_rwsem
);
119 list_add_tail(&shrinker
->list
, &shrinker_list
);
120 up_write(&shrinker_rwsem
);
122 EXPORT_SYMBOL(register_shrinker
);
127 void unregister_shrinker(struct shrinker
*shrinker
)
129 down_write(&shrinker_rwsem
);
130 list_del(&shrinker
->list
);
131 up_write(&shrinker_rwsem
);
133 EXPORT_SYMBOL(unregister_shrinker
);
135 #define SHRINK_BATCH 128
137 * Call the shrink functions to age shrinkable caches
139 * Here we assume it costs one seek to replace a lru page and that it also
140 * takes a seek to recreate a cache object. With this in mind we age equal
141 * percentages of the lru and ageable caches. This should balance the seeks
142 * generated by these structures.
144 * If the vm encounted mapped pages on the LRU it increase the pressure on
145 * slab to avoid swapping.
147 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
149 * `lru_pages' represents the number of on-LRU pages in all the zones which
150 * are eligible for the caller's allocation attempt. It is used for balancing
151 * slab reclaim versus page reclaim.
153 * Returns the number of slab objects which we shrunk.
155 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
156 unsigned long lru_pages
)
158 struct shrinker
*shrinker
;
159 unsigned long ret
= 0;
162 scanned
= SWAP_CLUSTER_MAX
;
164 if (!down_read_trylock(&shrinker_rwsem
))
165 return 1; /* Assume we'll be able to shrink next time */
167 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
168 unsigned long long delta
;
169 unsigned long total_scan
;
170 unsigned long max_pass
= (*shrinker
->shrink
)(0, gfp_mask
);
172 delta
= (4 * scanned
) / shrinker
->seeks
;
174 do_div(delta
, lru_pages
+ 1);
175 shrinker
->nr
+= delta
;
176 if (shrinker
->nr
< 0) {
177 printk(KERN_ERR
"%s: nr=%ld\n",
178 __FUNCTION__
, shrinker
->nr
);
179 shrinker
->nr
= max_pass
;
183 * Avoid risking looping forever due to too large nr value:
184 * never try to free more than twice the estimate number of
187 if (shrinker
->nr
> max_pass
* 2)
188 shrinker
->nr
= max_pass
* 2;
190 total_scan
= shrinker
->nr
;
193 while (total_scan
>= SHRINK_BATCH
) {
194 long this_scan
= SHRINK_BATCH
;
198 nr_before
= (*shrinker
->shrink
)(0, gfp_mask
);
199 shrink_ret
= (*shrinker
->shrink
)(this_scan
, gfp_mask
);
200 if (shrink_ret
== -1)
202 if (shrink_ret
< nr_before
)
203 ret
+= nr_before
- shrink_ret
;
204 count_vm_events(SLABS_SCANNED
, this_scan
);
205 total_scan
-= this_scan
;
210 shrinker
->nr
+= total_scan
;
212 up_read(&shrinker_rwsem
);
216 /* Called without lock on whether page is mapped, so answer is unstable */
217 static inline int page_mapping_inuse(struct page
*page
)
219 struct address_space
*mapping
;
221 /* Page is in somebody's page tables. */
222 if (page_mapped(page
))
225 /* Be more reluctant to reclaim swapcache than pagecache */
226 if (PageSwapCache(page
))
229 mapping
= page_mapping(page
);
233 /* File is mmap'd by somebody? */
234 return mapping_mapped(mapping
);
237 static inline int is_page_cache_freeable(struct page
*page
)
239 return page_count(page
) - !!PagePrivate(page
) == 2;
242 static int may_write_to_queue(struct backing_dev_info
*bdi
)
244 if (current
->flags
& PF_SWAPWRITE
)
246 if (!bdi_write_congested(bdi
))
248 if (bdi
== current
->backing_dev_info
)
254 * We detected a synchronous write error writing a page out. Probably
255 * -ENOSPC. We need to propagate that into the address_space for a subsequent
256 * fsync(), msync() or close().
258 * The tricky part is that after writepage we cannot touch the mapping: nothing
259 * prevents it from being freed up. But we have a ref on the page and once
260 * that page is locked, the mapping is pinned.
262 * We're allowed to run sleeping lock_page() here because we know the caller has
265 static void handle_write_error(struct address_space
*mapping
,
266 struct page
*page
, int error
)
269 if (page_mapping(page
) == mapping
)
270 mapping_set_error(mapping
, error
);
274 /* possible outcome of pageout() */
276 /* failed to write page out, page is locked */
278 /* move page to the active list, page is locked */
280 /* page has been sent to the disk successfully, page is unlocked */
282 /* page is clean and locked */
287 * pageout is called by shrink_page_list() for each dirty page.
288 * Calls ->writepage().
290 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
293 * If the page is dirty, only perform writeback if that write
294 * will be non-blocking. To prevent this allocation from being
295 * stalled by pagecache activity. But note that there may be
296 * stalls if we need to run get_block(). We could test
297 * PagePrivate for that.
299 * If this process is currently in generic_file_write() against
300 * this page's queue, we can perform writeback even if that
303 * If the page is swapcache, write it back even if that would
304 * block, for some throttling. This happens by accident, because
305 * swap_backing_dev_info is bust: it doesn't reflect the
306 * congestion state of the swapdevs. Easy to fix, if needed.
307 * See swapfile.c:page_queue_congested().
309 if (!is_page_cache_freeable(page
))
313 * Some data journaling orphaned pages can have
314 * page->mapping == NULL while being dirty with clean buffers.
316 if (PagePrivate(page
)) {
317 if (try_to_free_buffers(page
)) {
318 ClearPageDirty(page
);
319 printk("%s: orphaned page\n", __FUNCTION__
);
325 if (mapping
->a_ops
->writepage
== NULL
)
326 return PAGE_ACTIVATE
;
327 if (!may_write_to_queue(mapping
->backing_dev_info
))
330 if (clear_page_dirty_for_io(page
)) {
332 struct writeback_control wbc
= {
333 .sync_mode
= WB_SYNC_NONE
,
334 .nr_to_write
= SWAP_CLUSTER_MAX
,
336 .range_end
= LLONG_MAX
,
341 SetPageReclaim(page
);
342 res
= mapping
->a_ops
->writepage(page
, &wbc
);
344 handle_write_error(mapping
, page
, res
);
345 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
346 ClearPageReclaim(page
);
347 return PAGE_ACTIVATE
;
349 if (!PageWriteback(page
)) {
350 /* synchronous write or broken a_ops? */
351 ClearPageReclaim(page
);
353 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
361 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
362 * someone else has a ref on the page, abort and return 0. If it was
363 * successfully detached, return 1. Assumes the caller has a single ref on
366 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
368 BUG_ON(!PageLocked(page
));
369 BUG_ON(mapping
!= page_mapping(page
));
371 write_lock_irq(&mapping
->tree_lock
);
373 * The non racy check for a busy page.
375 * Must be careful with the order of the tests. When someone has
376 * a ref to the page, it may be possible that they dirty it then
377 * drop the reference. So if PageDirty is tested before page_count
378 * here, then the following race may occur:
380 * get_user_pages(&page);
381 * [user mapping goes away]
383 * !PageDirty(page) [good]
384 * SetPageDirty(page);
386 * !page_count(page) [good, discard it]
388 * [oops, our write_to data is lost]
390 * Reversing the order of the tests ensures such a situation cannot
391 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
392 * load is not satisfied before that of page->_count.
394 * Note that if SetPageDirty is always performed via set_page_dirty,
395 * and thus under tree_lock, then this ordering is not required.
397 if (unlikely(page_count(page
) != 2))
400 if (unlikely(PageDirty(page
)))
403 if (PageSwapCache(page
)) {
404 swp_entry_t swap
= { .val
= page_private(page
) };
405 __delete_from_swap_cache(page
);
406 write_unlock_irq(&mapping
->tree_lock
);
408 __put_page(page
); /* The pagecache ref */
412 __remove_from_page_cache(page
);
413 write_unlock_irq(&mapping
->tree_lock
);
418 write_unlock_irq(&mapping
->tree_lock
);
423 * shrink_page_list() returns the number of reclaimed pages
425 static unsigned long shrink_page_list(struct list_head
*page_list
,
426 struct scan_control
*sc
)
428 LIST_HEAD(ret_pages
);
429 struct pagevec freed_pvec
;
431 unsigned long nr_reclaimed
= 0;
435 pagevec_init(&freed_pvec
, 1);
436 while (!list_empty(page_list
)) {
437 struct address_space
*mapping
;
444 page
= lru_to_page(page_list
);
445 list_del(&page
->lru
);
447 if (TestSetPageLocked(page
))
450 VM_BUG_ON(PageActive(page
));
454 if (!sc
->may_swap
&& page_mapped(page
))
457 /* Double the slab pressure for mapped and swapcache pages */
458 if (page_mapped(page
) || PageSwapCache(page
))
461 if (PageWriteback(page
))
464 referenced
= page_referenced(page
, 1);
465 /* In active use or really unfreeable? Activate it. */
466 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&&
467 referenced
&& page_mapping_inuse(page
))
468 goto activate_locked
;
472 * Anonymous process memory has backing store?
473 * Try to allocate it some swap space here.
475 if (PageAnon(page
) && !PageSwapCache(page
))
476 if (!add_to_swap(page
, GFP_ATOMIC
))
477 goto activate_locked
;
478 #endif /* CONFIG_SWAP */
480 mapping
= page_mapping(page
);
481 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
482 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
485 * The page is mapped into the page tables of one or more
486 * processes. Try to unmap it here.
488 if (page_mapped(page
) && mapping
) {
489 switch (try_to_unmap(page
, 0)) {
491 goto activate_locked
;
495 ; /* try to free the page below */
499 if (PageDirty(page
)) {
500 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&& referenced
)
504 if (!sc
->may_writepage
)
507 /* Page is dirty, try to write it out here */
508 switch(pageout(page
, mapping
)) {
512 goto activate_locked
;
514 if (PageWriteback(page
) || PageDirty(page
))
517 * A synchronous write - probably a ramdisk. Go
518 * ahead and try to reclaim the page.
520 if (TestSetPageLocked(page
))
522 if (PageDirty(page
) || PageWriteback(page
))
524 mapping
= page_mapping(page
);
526 ; /* try to free the page below */
531 * If the page has buffers, try to free the buffer mappings
532 * associated with this page. If we succeed we try to free
535 * We do this even if the page is PageDirty().
536 * try_to_release_page() does not perform I/O, but it is
537 * possible for a page to have PageDirty set, but it is actually
538 * clean (all its buffers are clean). This happens if the
539 * buffers were written out directly, with submit_bh(). ext3
540 * will do this, as well as the blockdev mapping.
541 * try_to_release_page() will discover that cleanness and will
542 * drop the buffers and mark the page clean - it can be freed.
544 * Rarely, pages can have buffers and no ->mapping. These are
545 * the pages which were not successfully invalidated in
546 * truncate_complete_page(). We try to drop those buffers here
547 * and if that worked, and the page is no longer mapped into
548 * process address space (page_count == 1) it can be freed.
549 * Otherwise, leave the page on the LRU so it is swappable.
551 if (PagePrivate(page
)) {
552 if (!try_to_release_page(page
, sc
->gfp_mask
))
553 goto activate_locked
;
554 if (!mapping
&& page_count(page
) == 1)
558 if (!mapping
|| !remove_mapping(mapping
, page
))
564 if (!pagevec_add(&freed_pvec
, page
))
565 __pagevec_release_nonlru(&freed_pvec
);
574 list_add(&page
->lru
, &ret_pages
);
575 VM_BUG_ON(PageLRU(page
));
577 list_splice(&ret_pages
, page_list
);
578 if (pagevec_count(&freed_pvec
))
579 __pagevec_release_nonlru(&freed_pvec
);
580 count_vm_events(PGACTIVATE
, pgactivate
);
584 /* LRU Isolation modes. */
585 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
586 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
587 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
590 * Attempt to remove the specified page from its LRU. Only take this page
591 * if it is of the appropriate PageActive status. Pages which are being
592 * freed elsewhere are also ignored.
594 * page: page to consider
595 * mode: one of the LRU isolation modes defined above
597 * returns 0 on success, -ve errno on failure.
599 static int __isolate_lru_page(struct page
*page
, int mode
)
603 /* Only take pages on the LRU. */
608 * When checking the active state, we need to be sure we are
609 * dealing with comparible boolean values. Take the logical not
612 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
616 if (likely(get_page_unless_zero(page
))) {
618 * Be careful not to clear PageLRU until after we're
619 * sure the page is not being freed elsewhere -- the
620 * page release code relies on it.
630 * zone->lru_lock is heavily contended. Some of the functions that
631 * shrink the lists perform better by taking out a batch of pages
632 * and working on them outside the LRU lock.
634 * For pagecache intensive workloads, this function is the hottest
635 * spot in the kernel (apart from copy_*_user functions).
637 * Appropriate locks must be held before calling this function.
639 * @nr_to_scan: The number of pages to look through on the list.
640 * @src: The LRU list to pull pages off.
641 * @dst: The temp list to put pages on to.
642 * @scanned: The number of pages that were scanned.
643 * @order: The caller's attempted allocation order
644 * @mode: One of the LRU isolation modes
646 * returns how many pages were moved onto *@dst.
648 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
649 struct list_head
*src
, struct list_head
*dst
,
650 unsigned long *scanned
, int order
, int mode
)
652 unsigned long nr_taken
= 0;
655 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
658 unsigned long end_pfn
;
659 unsigned long page_pfn
;
662 page
= lru_to_page(src
);
663 prefetchw_prev_lru_page(page
, src
, flags
);
665 VM_BUG_ON(!PageLRU(page
));
667 switch (__isolate_lru_page(page
, mode
)) {
669 list_move(&page
->lru
, dst
);
674 /* else it is being freed elsewhere */
675 list_move(&page
->lru
, src
);
686 * Attempt to take all pages in the order aligned region
687 * surrounding the tag page. Only take those pages of
688 * the same active state as that tag page. We may safely
689 * round the target page pfn down to the requested order
690 * as the mem_map is guarenteed valid out to MAX_ORDER,
691 * where that page is in a different zone we will detect
692 * it from its zone id and abort this block scan.
694 zone_id
= page_zone_id(page
);
695 page_pfn
= page_to_pfn(page
);
696 pfn
= page_pfn
& ~((1 << order
) - 1);
697 end_pfn
= pfn
+ (1 << order
);
698 for (; pfn
< end_pfn
; pfn
++) {
699 struct page
*cursor_page
;
701 /* The target page is in the block, ignore it. */
702 if (unlikely(pfn
== page_pfn
))
705 /* Avoid holes within the zone. */
706 if (unlikely(!pfn_valid_within(pfn
)))
709 cursor_page
= pfn_to_page(pfn
);
710 /* Check that we have not crossed a zone boundary. */
711 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
713 switch (__isolate_lru_page(cursor_page
, mode
)) {
715 list_move(&cursor_page
->lru
, dst
);
721 /* else it is being freed elsewhere */
722 list_move(&cursor_page
->lru
, src
);
734 * clear_active_flags() is a helper for shrink_active_list(), clearing
735 * any active bits from the pages in the list.
737 static unsigned long clear_active_flags(struct list_head
*page_list
)
742 list_for_each_entry(page
, page_list
, lru
)
743 if (PageActive(page
)) {
744 ClearPageActive(page
);
752 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
755 static unsigned long shrink_inactive_list(unsigned long max_scan
,
756 struct zone
*zone
, struct scan_control
*sc
)
758 LIST_HEAD(page_list
);
760 unsigned long nr_scanned
= 0;
761 unsigned long nr_reclaimed
= 0;
763 pagevec_init(&pvec
, 1);
766 spin_lock_irq(&zone
->lru_lock
);
769 unsigned long nr_taken
;
770 unsigned long nr_scan
;
771 unsigned long nr_freed
;
772 unsigned long nr_active
;
774 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
775 &zone
->inactive_list
,
776 &page_list
, &nr_scan
, sc
->order
,
777 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)?
778 ISOLATE_BOTH
: ISOLATE_INACTIVE
);
779 nr_active
= clear_active_flags(&page_list
);
780 __count_vm_events(PGDEACTIVATE
, nr_active
);
782 __mod_zone_page_state(zone
, NR_ACTIVE
, -nr_active
);
783 __mod_zone_page_state(zone
, NR_INACTIVE
,
784 -(nr_taken
- nr_active
));
785 zone
->pages_scanned
+= nr_scan
;
786 spin_unlock_irq(&zone
->lru_lock
);
788 nr_scanned
+= nr_scan
;
789 nr_freed
= shrink_page_list(&page_list
, sc
);
790 nr_reclaimed
+= nr_freed
;
792 if (current_is_kswapd()) {
793 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scan
);
794 __count_vm_events(KSWAPD_STEAL
, nr_freed
);
796 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scan
);
797 __count_zone_vm_events(PGSTEAL
, zone
, nr_freed
);
802 spin_lock(&zone
->lru_lock
);
804 * Put back any unfreeable pages.
806 while (!list_empty(&page_list
)) {
807 page
= lru_to_page(&page_list
);
808 VM_BUG_ON(PageLRU(page
));
810 list_del(&page
->lru
);
811 if (PageActive(page
))
812 add_page_to_active_list(zone
, page
);
814 add_page_to_inactive_list(zone
, page
);
815 if (!pagevec_add(&pvec
, page
)) {
816 spin_unlock_irq(&zone
->lru_lock
);
817 __pagevec_release(&pvec
);
818 spin_lock_irq(&zone
->lru_lock
);
821 } while (nr_scanned
< max_scan
);
822 spin_unlock(&zone
->lru_lock
);
825 pagevec_release(&pvec
);
830 * We are about to scan this zone at a certain priority level. If that priority
831 * level is smaller (ie: more urgent) than the previous priority, then note
832 * that priority level within the zone. This is done so that when the next
833 * process comes in to scan this zone, it will immediately start out at this
834 * priority level rather than having to build up its own scanning priority.
835 * Here, this priority affects only the reclaim-mapped threshold.
837 static inline void note_zone_scanning_priority(struct zone
*zone
, int priority
)
839 if (priority
< zone
->prev_priority
)
840 zone
->prev_priority
= priority
;
843 static inline int zone_is_near_oom(struct zone
*zone
)
845 return zone
->pages_scanned
>= (zone_page_state(zone
, NR_ACTIVE
)
846 + zone_page_state(zone
, NR_INACTIVE
))*3;
850 * This moves pages from the active list to the inactive list.
852 * We move them the other way if the page is referenced by one or more
853 * processes, from rmap.
855 * If the pages are mostly unmapped, the processing is fast and it is
856 * appropriate to hold zone->lru_lock across the whole operation. But if
857 * the pages are mapped, the processing is slow (page_referenced()) so we
858 * should drop zone->lru_lock around each page. It's impossible to balance
859 * this, so instead we remove the pages from the LRU while processing them.
860 * It is safe to rely on PG_active against the non-LRU pages in here because
861 * nobody will play with that bit on a non-LRU page.
863 * The downside is that we have to touch page->_count against each page.
864 * But we had to alter page->flags anyway.
866 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
867 struct scan_control
*sc
, int priority
)
869 unsigned long pgmoved
;
870 int pgdeactivate
= 0;
871 unsigned long pgscanned
;
872 LIST_HEAD(l_hold
); /* The pages which were snipped off */
873 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
874 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
877 int reclaim_mapped
= 0;
884 if (zone_is_near_oom(zone
))
885 goto force_reclaim_mapped
;
888 * `distress' is a measure of how much trouble we're having
889 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
891 distress
= 100 >> min(zone
->prev_priority
, priority
);
894 * The point of this algorithm is to decide when to start
895 * reclaiming mapped memory instead of just pagecache. Work out
899 mapped_ratio
= ((global_page_state(NR_FILE_MAPPED
) +
900 global_page_state(NR_ANON_PAGES
)) * 100) /
904 * Now decide how much we really want to unmap some pages. The
905 * mapped ratio is downgraded - just because there's a lot of
906 * mapped memory doesn't necessarily mean that page reclaim
909 * The distress ratio is important - we don't want to start
912 * A 100% value of vm_swappiness overrides this algorithm
915 swap_tendency
= mapped_ratio
/ 2 + distress
+ sc
->swappiness
;
918 * Now use this metric to decide whether to start moving mapped
919 * memory onto the inactive list.
921 if (swap_tendency
>= 100)
922 force_reclaim_mapped
:
927 spin_lock_irq(&zone
->lru_lock
);
928 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
929 &l_hold
, &pgscanned
, sc
->order
, ISOLATE_ACTIVE
);
930 zone
->pages_scanned
+= pgscanned
;
931 __mod_zone_page_state(zone
, NR_ACTIVE
, -pgmoved
);
932 spin_unlock_irq(&zone
->lru_lock
);
934 while (!list_empty(&l_hold
)) {
936 page
= lru_to_page(&l_hold
);
937 list_del(&page
->lru
);
938 if (page_mapped(page
)) {
939 if (!reclaim_mapped
||
940 (total_swap_pages
== 0 && PageAnon(page
)) ||
941 page_referenced(page
, 0)) {
942 list_add(&page
->lru
, &l_active
);
946 list_add(&page
->lru
, &l_inactive
);
949 pagevec_init(&pvec
, 1);
951 spin_lock_irq(&zone
->lru_lock
);
952 while (!list_empty(&l_inactive
)) {
953 page
= lru_to_page(&l_inactive
);
954 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
955 VM_BUG_ON(PageLRU(page
));
957 VM_BUG_ON(!PageActive(page
));
958 ClearPageActive(page
);
960 list_move(&page
->lru
, &zone
->inactive_list
);
962 if (!pagevec_add(&pvec
, page
)) {
963 __mod_zone_page_state(zone
, NR_INACTIVE
, pgmoved
);
964 spin_unlock_irq(&zone
->lru_lock
);
965 pgdeactivate
+= pgmoved
;
967 if (buffer_heads_over_limit
)
968 pagevec_strip(&pvec
);
969 __pagevec_release(&pvec
);
970 spin_lock_irq(&zone
->lru_lock
);
973 __mod_zone_page_state(zone
, NR_INACTIVE
, pgmoved
);
974 pgdeactivate
+= pgmoved
;
975 if (buffer_heads_over_limit
) {
976 spin_unlock_irq(&zone
->lru_lock
);
977 pagevec_strip(&pvec
);
978 spin_lock_irq(&zone
->lru_lock
);
982 while (!list_empty(&l_active
)) {
983 page
= lru_to_page(&l_active
);
984 prefetchw_prev_lru_page(page
, &l_active
, flags
);
985 VM_BUG_ON(PageLRU(page
));
987 VM_BUG_ON(!PageActive(page
));
988 list_move(&page
->lru
, &zone
->active_list
);
990 if (!pagevec_add(&pvec
, page
)) {
991 __mod_zone_page_state(zone
, NR_ACTIVE
, pgmoved
);
993 spin_unlock_irq(&zone
->lru_lock
);
994 __pagevec_release(&pvec
);
995 spin_lock_irq(&zone
->lru_lock
);
998 __mod_zone_page_state(zone
, NR_ACTIVE
, pgmoved
);
1000 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1001 __count_vm_events(PGDEACTIVATE
, pgdeactivate
);
1002 spin_unlock_irq(&zone
->lru_lock
);
1004 pagevec_release(&pvec
);
1008 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1010 static unsigned long shrink_zone(int priority
, struct zone
*zone
,
1011 struct scan_control
*sc
)
1013 unsigned long nr_active
;
1014 unsigned long nr_inactive
;
1015 unsigned long nr_to_scan
;
1016 unsigned long nr_reclaimed
= 0;
1018 atomic_inc(&zone
->reclaim_in_progress
);
1021 * Add one to `nr_to_scan' just to make sure that the kernel will
1022 * slowly sift through the active list.
1024 zone
->nr_scan_active
+=
1025 (zone_page_state(zone
, NR_ACTIVE
) >> priority
) + 1;
1026 nr_active
= zone
->nr_scan_active
;
1027 if (nr_active
>= sc
->swap_cluster_max
)
1028 zone
->nr_scan_active
= 0;
1032 zone
->nr_scan_inactive
+=
1033 (zone_page_state(zone
, NR_INACTIVE
) >> priority
) + 1;
1034 nr_inactive
= zone
->nr_scan_inactive
;
1035 if (nr_inactive
>= sc
->swap_cluster_max
)
1036 zone
->nr_scan_inactive
= 0;
1040 while (nr_active
|| nr_inactive
) {
1042 nr_to_scan
= min(nr_active
,
1043 (unsigned long)sc
->swap_cluster_max
);
1044 nr_active
-= nr_to_scan
;
1045 shrink_active_list(nr_to_scan
, zone
, sc
, priority
);
1049 nr_to_scan
= min(nr_inactive
,
1050 (unsigned long)sc
->swap_cluster_max
);
1051 nr_inactive
-= nr_to_scan
;
1052 nr_reclaimed
+= shrink_inactive_list(nr_to_scan
, zone
,
1057 throttle_vm_writeout(sc
->gfp_mask
);
1059 atomic_dec(&zone
->reclaim_in_progress
);
1060 return nr_reclaimed
;
1064 * This is the direct reclaim path, for page-allocating processes. We only
1065 * try to reclaim pages from zones which will satisfy the caller's allocation
1068 * We reclaim from a zone even if that zone is over pages_high. Because:
1069 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1071 * b) The zones may be over pages_high but they must go *over* pages_high to
1072 * satisfy the `incremental min' zone defense algorithm.
1074 * Returns the number of reclaimed pages.
1076 * If a zone is deemed to be full of pinned pages then just give it a light
1077 * scan then give up on it.
1079 static unsigned long shrink_zones(int priority
, struct zone
**zones
,
1080 struct scan_control
*sc
)
1082 unsigned long nr_reclaimed
= 0;
1085 sc
->all_unreclaimable
= 1;
1086 for (i
= 0; zones
[i
] != NULL
; i
++) {
1087 struct zone
*zone
= zones
[i
];
1089 if (!populated_zone(zone
))
1092 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1095 note_zone_scanning_priority(zone
, priority
);
1097 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1098 continue; /* Let kswapd poll it */
1100 sc
->all_unreclaimable
= 0;
1102 nr_reclaimed
+= shrink_zone(priority
, zone
, sc
);
1104 return nr_reclaimed
;
1108 * This is the main entry point to direct page reclaim.
1110 * If a full scan of the inactive list fails to free enough memory then we
1111 * are "out of memory" and something needs to be killed.
1113 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1114 * high - the zone may be full of dirty or under-writeback pages, which this
1115 * caller can't do much about. We kick pdflush and take explicit naps in the
1116 * hope that some of these pages can be written. But if the allocating task
1117 * holds filesystem locks which prevent writeout this might not work, and the
1118 * allocation attempt will fail.
1120 unsigned long try_to_free_pages(struct zone
**zones
, int order
, gfp_t gfp_mask
)
1124 unsigned long total_scanned
= 0;
1125 unsigned long nr_reclaimed
= 0;
1126 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1127 unsigned long lru_pages
= 0;
1129 struct scan_control sc
= {
1130 .gfp_mask
= gfp_mask
,
1131 .may_writepage
= !laptop_mode
,
1132 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1134 .swappiness
= vm_swappiness
,
1138 count_vm_event(ALLOCSTALL
);
1140 for (i
= 0; zones
[i
] != NULL
; i
++) {
1141 struct zone
*zone
= zones
[i
];
1143 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1146 lru_pages
+= zone_page_state(zone
, NR_ACTIVE
)
1147 + zone_page_state(zone
, NR_INACTIVE
);
1150 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1153 disable_swap_token();
1154 nr_reclaimed
+= shrink_zones(priority
, zones
, &sc
);
1155 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1156 if (reclaim_state
) {
1157 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1158 reclaim_state
->reclaimed_slab
= 0;
1160 total_scanned
+= sc
.nr_scanned
;
1161 if (nr_reclaimed
>= sc
.swap_cluster_max
) {
1167 * Try to write back as many pages as we just scanned. This
1168 * tends to cause slow streaming writers to write data to the
1169 * disk smoothly, at the dirtying rate, which is nice. But
1170 * that's undesirable in laptop mode, where we *want* lumpy
1171 * writeout. So in laptop mode, write out the whole world.
1173 if (total_scanned
> sc
.swap_cluster_max
+
1174 sc
.swap_cluster_max
/ 2) {
1175 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1176 sc
.may_writepage
= 1;
1179 /* Take a nap, wait for some writeback to complete */
1180 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1181 congestion_wait(WRITE
, HZ
/10);
1183 /* top priority shrink_caches still had more to do? don't OOM, then */
1184 if (!sc
.all_unreclaimable
)
1188 * Now that we've scanned all the zones at this priority level, note
1189 * that level within the zone so that the next thread which performs
1190 * scanning of this zone will immediately start out at this priority
1191 * level. This affects only the decision whether or not to bring
1192 * mapped pages onto the inactive list.
1196 for (i
= 0; zones
[i
] != 0; i
++) {
1197 struct zone
*zone
= zones
[i
];
1199 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1202 zone
->prev_priority
= priority
;
1208 * For kswapd, balance_pgdat() will work across all this node's zones until
1209 * they are all at pages_high.
1211 * Returns the number of pages which were actually freed.
1213 * There is special handling here for zones which are full of pinned pages.
1214 * This can happen if the pages are all mlocked, or if they are all used by
1215 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1216 * What we do is to detect the case where all pages in the zone have been
1217 * scanned twice and there has been zero successful reclaim. Mark the zone as
1218 * dead and from now on, only perform a short scan. Basically we're polling
1219 * the zone for when the problem goes away.
1221 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1222 * zones which have free_pages > pages_high, but once a zone is found to have
1223 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1224 * of the number of free pages in the lower zones. This interoperates with
1225 * the page allocator fallback scheme to ensure that aging of pages is balanced
1228 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
1233 unsigned long total_scanned
;
1234 unsigned long nr_reclaimed
;
1235 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1236 struct scan_control sc
= {
1237 .gfp_mask
= GFP_KERNEL
,
1239 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1240 .swappiness
= vm_swappiness
,
1244 * temp_priority is used to remember the scanning priority at which
1245 * this zone was successfully refilled to free_pages == pages_high.
1247 int temp_priority
[MAX_NR_ZONES
];
1252 sc
.may_writepage
= !laptop_mode
;
1253 count_vm_event(PAGEOUTRUN
);
1255 for (i
= 0; i
< pgdat
->nr_zones
; i
++)
1256 temp_priority
[i
] = DEF_PRIORITY
;
1258 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1259 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1260 unsigned long lru_pages
= 0;
1262 /* The swap token gets in the way of swapout... */
1264 disable_swap_token();
1269 * Scan in the highmem->dma direction for the highest
1270 * zone which needs scanning
1272 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1273 struct zone
*zone
= pgdat
->node_zones
+ i
;
1275 if (!populated_zone(zone
))
1278 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1281 if (!zone_watermark_ok(zone
, order
, zone
->pages_high
,
1290 for (i
= 0; i
<= end_zone
; i
++) {
1291 struct zone
*zone
= pgdat
->node_zones
+ i
;
1293 lru_pages
+= zone_page_state(zone
, NR_ACTIVE
)
1294 + zone_page_state(zone
, NR_INACTIVE
);
1298 * Now scan the zone in the dma->highmem direction, stopping
1299 * at the last zone which needs scanning.
1301 * We do this because the page allocator works in the opposite
1302 * direction. This prevents the page allocator from allocating
1303 * pages behind kswapd's direction of progress, which would
1304 * cause too much scanning of the lower zones.
1306 for (i
= 0; i
<= end_zone
; i
++) {
1307 struct zone
*zone
= pgdat
->node_zones
+ i
;
1310 if (!populated_zone(zone
))
1313 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1316 if (!zone_watermark_ok(zone
, order
, zone
->pages_high
,
1319 temp_priority
[i
] = priority
;
1321 note_zone_scanning_priority(zone
, priority
);
1322 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1323 reclaim_state
->reclaimed_slab
= 0;
1324 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1326 nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1327 total_scanned
+= sc
.nr_scanned
;
1328 if (zone
->all_unreclaimable
)
1330 if (nr_slab
== 0 && zone
->pages_scanned
>=
1331 (zone_page_state(zone
, NR_ACTIVE
)
1332 + zone_page_state(zone
, NR_INACTIVE
)) * 6)
1333 zone
->all_unreclaimable
= 1;
1335 * If we've done a decent amount of scanning and
1336 * the reclaim ratio is low, start doing writepage
1337 * even in laptop mode
1339 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1340 total_scanned
> nr_reclaimed
+ nr_reclaimed
/ 2)
1341 sc
.may_writepage
= 1;
1344 break; /* kswapd: all done */
1346 * OK, kswapd is getting into trouble. Take a nap, then take
1347 * another pass across the zones.
1349 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1350 congestion_wait(WRITE
, HZ
/10);
1353 * We do this so kswapd doesn't build up large priorities for
1354 * example when it is freeing in parallel with allocators. It
1355 * matches the direct reclaim path behaviour in terms of impact
1356 * on zone->*_priority.
1358 if (nr_reclaimed
>= SWAP_CLUSTER_MAX
)
1363 * Note within each zone the priority level at which this zone was
1364 * brought into a happy state. So that the next thread which scans this
1365 * zone will start out at that priority level.
1367 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1368 struct zone
*zone
= pgdat
->node_zones
+ i
;
1370 zone
->prev_priority
= temp_priority
[i
];
1372 if (!all_zones_ok
) {
1380 return nr_reclaimed
;
1384 * The background pageout daemon, started as a kernel thread
1385 * from the init process.
1387 * This basically trickles out pages so that we have _some_
1388 * free memory available even if there is no other activity
1389 * that frees anything up. This is needed for things like routing
1390 * etc, where we otherwise might have all activity going on in
1391 * asynchronous contexts that cannot page things out.
1393 * If there are applications that are active memory-allocators
1394 * (most normal use), this basically shouldn't matter.
1396 static int kswapd(void *p
)
1398 unsigned long order
;
1399 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1400 struct task_struct
*tsk
= current
;
1402 struct reclaim_state reclaim_state
= {
1403 .reclaimed_slab
= 0,
1407 cpumask
= node_to_cpumask(pgdat
->node_id
);
1408 if (!cpus_empty(cpumask
))
1409 set_cpus_allowed(tsk
, cpumask
);
1410 current
->reclaim_state
= &reclaim_state
;
1413 * Tell the memory management that we're a "memory allocator",
1414 * and that if we need more memory we should get access to it
1415 * regardless (see "__alloc_pages()"). "kswapd" should
1416 * never get caught in the normal page freeing logic.
1418 * (Kswapd normally doesn't need memory anyway, but sometimes
1419 * you need a small amount of memory in order to be able to
1420 * page out something else, and this flag essentially protects
1421 * us from recursively trying to free more memory as we're
1422 * trying to free the first piece of memory in the first place).
1424 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1429 unsigned long new_order
;
1431 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1432 new_order
= pgdat
->kswapd_max_order
;
1433 pgdat
->kswapd_max_order
= 0;
1434 if (order
< new_order
) {
1436 * Don't sleep if someone wants a larger 'order'
1441 if (!freezing(current
))
1444 order
= pgdat
->kswapd_max_order
;
1446 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1448 if (!try_to_freeze()) {
1449 /* We can speed up thawing tasks if we don't call
1450 * balance_pgdat after returning from the refrigerator
1452 balance_pgdat(pgdat
, order
);
1459 * A zone is low on free memory, so wake its kswapd task to service it.
1461 void wakeup_kswapd(struct zone
*zone
, int order
)
1465 if (!populated_zone(zone
))
1468 pgdat
= zone
->zone_pgdat
;
1469 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1471 if (pgdat
->kswapd_max_order
< order
)
1472 pgdat
->kswapd_max_order
= order
;
1473 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1475 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1477 wake_up_interruptible(&pgdat
->kswapd_wait
);
1482 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1483 * from LRU lists system-wide, for given pass and priority, and returns the
1484 * number of reclaimed pages
1486 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1488 static unsigned long shrink_all_zones(unsigned long nr_pages
, int prio
,
1489 int pass
, struct scan_control
*sc
)
1492 unsigned long nr_to_scan
, ret
= 0;
1494 for_each_zone(zone
) {
1496 if (!populated_zone(zone
))
1499 if (zone
->all_unreclaimable
&& prio
!= DEF_PRIORITY
)
1502 /* For pass = 0 we don't shrink the active list */
1504 zone
->nr_scan_active
+=
1505 (zone_page_state(zone
, NR_ACTIVE
) >> prio
) + 1;
1506 if (zone
->nr_scan_active
>= nr_pages
|| pass
> 3) {
1507 zone
->nr_scan_active
= 0;
1508 nr_to_scan
= min(nr_pages
,
1509 zone_page_state(zone
, NR_ACTIVE
));
1510 shrink_active_list(nr_to_scan
, zone
, sc
, prio
);
1514 zone
->nr_scan_inactive
+=
1515 (zone_page_state(zone
, NR_INACTIVE
) >> prio
) + 1;
1516 if (zone
->nr_scan_inactive
>= nr_pages
|| pass
> 3) {
1517 zone
->nr_scan_inactive
= 0;
1518 nr_to_scan
= min(nr_pages
,
1519 zone_page_state(zone
, NR_INACTIVE
));
1520 ret
+= shrink_inactive_list(nr_to_scan
, zone
, sc
);
1521 if (ret
>= nr_pages
)
1529 static unsigned long count_lru_pages(void)
1531 return global_page_state(NR_ACTIVE
) + global_page_state(NR_INACTIVE
);
1535 * Try to free `nr_pages' of memory, system-wide, and return the number of
1538 * Rather than trying to age LRUs the aim is to preserve the overall
1539 * LRU order by reclaiming preferentially
1540 * inactive > active > active referenced > active mapped
1542 unsigned long shrink_all_memory(unsigned long nr_pages
)
1544 unsigned long lru_pages
, nr_slab
;
1545 unsigned long ret
= 0;
1547 struct reclaim_state reclaim_state
;
1548 struct scan_control sc
= {
1549 .gfp_mask
= GFP_KERNEL
,
1551 .swap_cluster_max
= nr_pages
,
1553 .swappiness
= vm_swappiness
,
1556 current
->reclaim_state
= &reclaim_state
;
1558 lru_pages
= count_lru_pages();
1559 nr_slab
= global_page_state(NR_SLAB_RECLAIMABLE
);
1560 /* If slab caches are huge, it's better to hit them first */
1561 while (nr_slab
>= lru_pages
) {
1562 reclaim_state
.reclaimed_slab
= 0;
1563 shrink_slab(nr_pages
, sc
.gfp_mask
, lru_pages
);
1564 if (!reclaim_state
.reclaimed_slab
)
1567 ret
+= reclaim_state
.reclaimed_slab
;
1568 if (ret
>= nr_pages
)
1571 nr_slab
-= reclaim_state
.reclaimed_slab
;
1575 * We try to shrink LRUs in 5 passes:
1576 * 0 = Reclaim from inactive_list only
1577 * 1 = Reclaim from active list but don't reclaim mapped
1578 * 2 = 2nd pass of type 1
1579 * 3 = Reclaim mapped (normal reclaim)
1580 * 4 = 2nd pass of type 3
1582 for (pass
= 0; pass
< 5; pass
++) {
1585 /* Force reclaiming mapped pages in the passes #3 and #4 */
1588 sc
.swappiness
= 100;
1591 for (prio
= DEF_PRIORITY
; prio
>= 0; prio
--) {
1592 unsigned long nr_to_scan
= nr_pages
- ret
;
1595 ret
+= shrink_all_zones(nr_to_scan
, prio
, pass
, &sc
);
1596 if (ret
>= nr_pages
)
1599 reclaim_state
.reclaimed_slab
= 0;
1600 shrink_slab(sc
.nr_scanned
, sc
.gfp_mask
,
1602 ret
+= reclaim_state
.reclaimed_slab
;
1603 if (ret
>= nr_pages
)
1606 if (sc
.nr_scanned
&& prio
< DEF_PRIORITY
- 2)
1607 congestion_wait(WRITE
, HZ
/ 10);
1612 * If ret = 0, we could not shrink LRUs, but there may be something
1617 reclaim_state
.reclaimed_slab
= 0;
1618 shrink_slab(nr_pages
, sc
.gfp_mask
, count_lru_pages());
1619 ret
+= reclaim_state
.reclaimed_slab
;
1620 } while (ret
< nr_pages
&& reclaim_state
.reclaimed_slab
> 0);
1624 current
->reclaim_state
= NULL
;
1630 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1631 not required for correctness. So if the last cpu in a node goes
1632 away, we get changed to run anywhere: as the first one comes back,
1633 restore their cpu bindings. */
1634 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
1635 unsigned long action
, void *hcpu
)
1640 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
1641 for_each_online_pgdat(pgdat
) {
1642 mask
= node_to_cpumask(pgdat
->node_id
);
1643 if (any_online_cpu(mask
) != NR_CPUS
)
1644 /* One of our CPUs online: restore mask */
1645 set_cpus_allowed(pgdat
->kswapd
, mask
);
1652 * This kswapd start function will be called by init and node-hot-add.
1653 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1655 int kswapd_run(int nid
)
1657 pg_data_t
*pgdat
= NODE_DATA(nid
);
1663 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
1664 if (IS_ERR(pgdat
->kswapd
)) {
1665 /* failure at boot is fatal */
1666 BUG_ON(system_state
== SYSTEM_BOOTING
);
1667 printk("Failed to start kswapd on node %d\n",nid
);
1673 static int __init
kswapd_init(void)
1678 for_each_online_node(nid
)
1680 hotcpu_notifier(cpu_callback
, 0);
1684 module_init(kswapd_init
)
1690 * If non-zero call zone_reclaim when the number of free pages falls below
1693 int zone_reclaim_mode __read_mostly
;
1695 #define RECLAIM_OFF 0
1696 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1697 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1698 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1701 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1702 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1705 #define ZONE_RECLAIM_PRIORITY 4
1708 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1711 int sysctl_min_unmapped_ratio
= 1;
1714 * If the number of slab pages in a zone grows beyond this percentage then
1715 * slab reclaim needs to occur.
1717 int sysctl_min_slab_ratio
= 5;
1720 * Try to free up some pages from this zone through reclaim.
1722 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1724 /* Minimum pages needed in order to stay on node */
1725 const unsigned long nr_pages
= 1 << order
;
1726 struct task_struct
*p
= current
;
1727 struct reclaim_state reclaim_state
;
1729 unsigned long nr_reclaimed
= 0;
1730 struct scan_control sc
= {
1731 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
1732 .may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
1733 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
1735 .gfp_mask
= gfp_mask
,
1736 .swappiness
= vm_swappiness
,
1738 unsigned long slab_reclaimable
;
1740 disable_swap_token();
1743 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1744 * and we also need to be able to write out pages for RECLAIM_WRITE
1747 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
1748 reclaim_state
.reclaimed_slab
= 0;
1749 p
->reclaim_state
= &reclaim_state
;
1751 if (zone_page_state(zone
, NR_FILE_PAGES
) -
1752 zone_page_state(zone
, NR_FILE_MAPPED
) >
1753 zone
->min_unmapped_pages
) {
1755 * Free memory by calling shrink zone with increasing
1756 * priorities until we have enough memory freed.
1758 priority
= ZONE_RECLAIM_PRIORITY
;
1760 note_zone_scanning_priority(zone
, priority
);
1761 nr_reclaimed
+= shrink_zone(priority
, zone
, &sc
);
1763 } while (priority
>= 0 && nr_reclaimed
< nr_pages
);
1766 slab_reclaimable
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
1767 if (slab_reclaimable
> zone
->min_slab_pages
) {
1769 * shrink_slab() does not currently allow us to determine how
1770 * many pages were freed in this zone. So we take the current
1771 * number of slab pages and shake the slab until it is reduced
1772 * by the same nr_pages that we used for reclaiming unmapped
1775 * Note that shrink_slab will free memory on all zones and may
1778 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
1779 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) >
1780 slab_reclaimable
- nr_pages
)
1784 * Update nr_reclaimed by the number of slab pages we
1785 * reclaimed from this zone.
1787 nr_reclaimed
+= slab_reclaimable
-
1788 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
1791 p
->reclaim_state
= NULL
;
1792 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
1793 return nr_reclaimed
>= nr_pages
;
1796 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1802 * Zone reclaim reclaims unmapped file backed pages and
1803 * slab pages if we are over the defined limits.
1805 * A small portion of unmapped file backed pages is needed for
1806 * file I/O otherwise pages read by file I/O will be immediately
1807 * thrown out if the zone is overallocated. So we do not reclaim
1808 * if less than a specified percentage of the zone is used by
1809 * unmapped file backed pages.
1811 if (zone_page_state(zone
, NR_FILE_PAGES
) -
1812 zone_page_state(zone
, NR_FILE_MAPPED
) <= zone
->min_unmapped_pages
1813 && zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)
1814 <= zone
->min_slab_pages
)
1818 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1819 * not have reclaimable pages and if we should not delay the allocation
1822 if (!(gfp_mask
& __GFP_WAIT
) ||
1823 zone
->all_unreclaimable
||
1824 atomic_read(&zone
->reclaim_in_progress
) > 0 ||
1825 (current
->flags
& PF_MEMALLOC
))
1829 * Only run zone reclaim on the local zone or on zones that do not
1830 * have associated processors. This will favor the local processor
1831 * over remote processors and spread off node memory allocations
1832 * as wide as possible.
1834 node_id
= zone_to_nid(zone
);
1835 mask
= node_to_cpumask(node_id
);
1836 if (!cpus_empty(mask
) && node_id
!= numa_node_id())
1838 return __zone_reclaim(zone
, gfp_mask
, order
);