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/gfp.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>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
47 #include <linux/swapops.h>
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
55 /* Incremented by the number of inactive pages that were scanned */
56 unsigned long nr_scanned
;
58 /* Number of pages freed so far during a call to shrink_zones() */
59 unsigned long nr_reclaimed
;
61 /* How many pages shrink_list() should reclaim */
62 unsigned long nr_to_reclaim
;
64 unsigned long hibernation_mode
;
66 /* This context's GFP mask */
71 /* Can mapped pages be reclaimed? */
74 /* Can pages be swapped as part of reclaim? */
82 * Intend to reclaim enough contenious memory rather than to reclaim
83 * enough amount memory. I.e, it's the mode for high order allocation.
85 bool lumpy_reclaim_mode
;
87 /* Which cgroup do we reclaim from */
88 struct mem_cgroup
*mem_cgroup
;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field) \
102 if ((_page)->lru.prev != _base) { \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
128 * From 0 .. 100. Higher means more swappy.
130 int vm_swappiness
= 60;
131 long vm_total_pages
; /* The total number of pages which the VM controls */
133 static LIST_HEAD(shrinker_list
);
134 static DECLARE_RWSEM(shrinker_rwsem
);
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
139 #define scanning_global_lru(sc) (1)
142 static struct zone_reclaim_stat
*get_reclaim_stat(struct zone
*zone
,
143 struct scan_control
*sc
)
145 if (!scanning_global_lru(sc
))
146 return mem_cgroup_get_reclaim_stat(sc
->mem_cgroup
, zone
);
148 return &zone
->reclaim_stat
;
151 static unsigned long zone_nr_lru_pages(struct zone
*zone
,
152 struct scan_control
*sc
, enum lru_list lru
)
154 if (!scanning_global_lru(sc
))
155 return mem_cgroup_zone_nr_pages(sc
->mem_cgroup
, zone
, lru
);
157 return zone_page_state(zone
, NR_LRU_BASE
+ lru
);
162 * Add a shrinker callback to be called from the vm
164 void register_shrinker(struct shrinker
*shrinker
)
167 down_write(&shrinker_rwsem
);
168 list_add_tail(&shrinker
->list
, &shrinker_list
);
169 up_write(&shrinker_rwsem
);
171 EXPORT_SYMBOL(register_shrinker
);
176 void unregister_shrinker(struct shrinker
*shrinker
)
178 down_write(&shrinker_rwsem
);
179 list_del(&shrinker
->list
);
180 up_write(&shrinker_rwsem
);
182 EXPORT_SYMBOL(unregister_shrinker
);
184 #define SHRINK_BATCH 128
186 * Call the shrink functions to age shrinkable caches
188 * Here we assume it costs one seek to replace a lru page and that it also
189 * takes a seek to recreate a cache object. With this in mind we age equal
190 * percentages of the lru and ageable caches. This should balance the seeks
191 * generated by these structures.
193 * If the vm encountered mapped pages on the LRU it increase the pressure on
194 * slab to avoid swapping.
196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
198 * `lru_pages' represents the number of on-LRU pages in all the zones which
199 * are eligible for the caller's allocation attempt. It is used for balancing
200 * slab reclaim versus page reclaim.
202 * Returns the number of slab objects which we shrunk.
204 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
205 unsigned long lru_pages
)
207 struct shrinker
*shrinker
;
208 unsigned long ret
= 0;
211 scanned
= SWAP_CLUSTER_MAX
;
213 if (!down_read_trylock(&shrinker_rwsem
))
214 return 1; /* Assume we'll be able to shrink next time */
216 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
217 unsigned long long delta
;
218 unsigned long total_scan
;
219 unsigned long max_pass
;
221 max_pass
= (*shrinker
->shrink
)(shrinker
, 0, gfp_mask
);
222 delta
= (4 * scanned
) / shrinker
->seeks
;
224 do_div(delta
, lru_pages
+ 1);
225 shrinker
->nr
+= delta
;
226 if (shrinker
->nr
< 0) {
227 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
229 shrinker
->shrink
, shrinker
->nr
);
230 shrinker
->nr
= max_pass
;
234 * Avoid risking looping forever due to too large nr value:
235 * never try to free more than twice the estimate number of
238 if (shrinker
->nr
> max_pass
* 2)
239 shrinker
->nr
= max_pass
* 2;
241 total_scan
= shrinker
->nr
;
244 while (total_scan
>= SHRINK_BATCH
) {
245 long this_scan
= SHRINK_BATCH
;
249 nr_before
= (*shrinker
->shrink
)(shrinker
, 0, gfp_mask
);
250 shrink_ret
= (*shrinker
->shrink
)(shrinker
, this_scan
,
252 if (shrink_ret
== -1)
254 if (shrink_ret
< nr_before
)
255 ret
+= nr_before
- shrink_ret
;
256 count_vm_events(SLABS_SCANNED
, this_scan
);
257 total_scan
-= this_scan
;
262 shrinker
->nr
+= total_scan
;
264 up_read(&shrinker_rwsem
);
268 static inline int is_page_cache_freeable(struct page
*page
)
271 * A freeable page cache page is referenced only by the caller
272 * that isolated the page, the page cache radix tree and
273 * optional buffer heads at page->private.
275 return page_count(page
) - page_has_private(page
) == 2;
278 static int may_write_to_queue(struct backing_dev_info
*bdi
)
280 if (current
->flags
& PF_SWAPWRITE
)
282 if (!bdi_write_congested(bdi
))
284 if (bdi
== current
->backing_dev_info
)
290 * We detected a synchronous write error writing a page out. Probably
291 * -ENOSPC. We need to propagate that into the address_space for a subsequent
292 * fsync(), msync() or close().
294 * The tricky part is that after writepage we cannot touch the mapping: nothing
295 * prevents it from being freed up. But we have a ref on the page and once
296 * that page is locked, the mapping is pinned.
298 * We're allowed to run sleeping lock_page() here because we know the caller has
301 static void handle_write_error(struct address_space
*mapping
,
302 struct page
*page
, int error
)
304 lock_page_nosync(page
);
305 if (page_mapping(page
) == mapping
)
306 mapping_set_error(mapping
, error
);
310 /* Request for sync pageout. */
316 /* possible outcome of pageout() */
318 /* failed to write page out, page is locked */
320 /* move page to the active list, page is locked */
322 /* page has been sent to the disk successfully, page is unlocked */
324 /* page is clean and locked */
329 * pageout is called by shrink_page_list() for each dirty page.
330 * Calls ->writepage().
332 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
333 enum pageout_io sync_writeback
)
336 * If the page is dirty, only perform writeback if that write
337 * will be non-blocking. To prevent this allocation from being
338 * stalled by pagecache activity. But note that there may be
339 * stalls if we need to run get_block(). We could test
340 * PagePrivate for that.
342 * If this process is currently in __generic_file_aio_write() against
343 * this page's queue, we can perform writeback even if that
346 * If the page is swapcache, write it back even if that would
347 * block, for some throttling. This happens by accident, because
348 * swap_backing_dev_info is bust: it doesn't reflect the
349 * congestion state of the swapdevs. Easy to fix, if needed.
351 if (!is_page_cache_freeable(page
))
355 * Some data journaling orphaned pages can have
356 * page->mapping == NULL while being dirty with clean buffers.
358 if (page_has_private(page
)) {
359 if (try_to_free_buffers(page
)) {
360 ClearPageDirty(page
);
361 printk("%s: orphaned page\n", __func__
);
367 if (mapping
->a_ops
->writepage
== NULL
)
368 return PAGE_ACTIVATE
;
369 if (!may_write_to_queue(mapping
->backing_dev_info
))
372 if (clear_page_dirty_for_io(page
)) {
374 struct writeback_control wbc
= {
375 .sync_mode
= WB_SYNC_NONE
,
376 .nr_to_write
= SWAP_CLUSTER_MAX
,
378 .range_end
= LLONG_MAX
,
383 SetPageReclaim(page
);
384 res
= mapping
->a_ops
->writepage(page
, &wbc
);
386 handle_write_error(mapping
, page
, res
);
387 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
388 ClearPageReclaim(page
);
389 return PAGE_ACTIVATE
;
393 * Wait on writeback if requested to. This happens when
394 * direct reclaiming a large contiguous area and the
395 * first attempt to free a range of pages fails.
397 if (PageWriteback(page
) && sync_writeback
== PAGEOUT_IO_SYNC
)
398 wait_on_page_writeback(page
);
400 if (!PageWriteback(page
)) {
401 /* synchronous write or broken a_ops? */
402 ClearPageReclaim(page
);
404 trace_mm_vmscan_writepage(page
,
405 trace_reclaim_flags(page
, sync_writeback
));
406 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
414 * Same as remove_mapping, but if the page is removed from the mapping, it
415 * gets returned with a refcount of 0.
417 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
419 BUG_ON(!PageLocked(page
));
420 BUG_ON(mapping
!= page_mapping(page
));
422 spin_lock_irq(&mapping
->tree_lock
);
424 * The non racy check for a busy page.
426 * Must be careful with the order of the tests. When someone has
427 * a ref to the page, it may be possible that they dirty it then
428 * drop the reference. So if PageDirty is tested before page_count
429 * here, then the following race may occur:
431 * get_user_pages(&page);
432 * [user mapping goes away]
434 * !PageDirty(page) [good]
435 * SetPageDirty(page);
437 * !page_count(page) [good, discard it]
439 * [oops, our write_to data is lost]
441 * Reversing the order of the tests ensures such a situation cannot
442 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
443 * load is not satisfied before that of page->_count.
445 * Note that if SetPageDirty is always performed via set_page_dirty,
446 * and thus under tree_lock, then this ordering is not required.
448 if (!page_freeze_refs(page
, 2))
450 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
451 if (unlikely(PageDirty(page
))) {
452 page_unfreeze_refs(page
, 2);
456 if (PageSwapCache(page
)) {
457 swp_entry_t swap
= { .val
= page_private(page
) };
458 __delete_from_swap_cache(page
);
459 spin_unlock_irq(&mapping
->tree_lock
);
460 swapcache_free(swap
, page
);
462 __remove_from_page_cache(page
);
463 spin_unlock_irq(&mapping
->tree_lock
);
464 mem_cgroup_uncharge_cache_page(page
);
470 spin_unlock_irq(&mapping
->tree_lock
);
475 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
476 * someone else has a ref on the page, abort and return 0. If it was
477 * successfully detached, return 1. Assumes the caller has a single ref on
480 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
482 if (__remove_mapping(mapping
, page
)) {
484 * Unfreezing the refcount with 1 rather than 2 effectively
485 * drops the pagecache ref for us without requiring another
488 page_unfreeze_refs(page
, 1);
495 * putback_lru_page - put previously isolated page onto appropriate LRU list
496 * @page: page to be put back to appropriate lru list
498 * Add previously isolated @page to appropriate LRU list.
499 * Page may still be unevictable for other reasons.
501 * lru_lock must not be held, interrupts must be enabled.
503 void putback_lru_page(struct page
*page
)
506 int active
= !!TestClearPageActive(page
);
507 int was_unevictable
= PageUnevictable(page
);
509 VM_BUG_ON(PageLRU(page
));
512 ClearPageUnevictable(page
);
514 if (page_evictable(page
, NULL
)) {
516 * For evictable pages, we can use the cache.
517 * In event of a race, worst case is we end up with an
518 * unevictable page on [in]active list.
519 * We know how to handle that.
521 lru
= active
+ page_lru_base_type(page
);
522 lru_cache_add_lru(page
, lru
);
525 * Put unevictable pages directly on zone's unevictable
528 lru
= LRU_UNEVICTABLE
;
529 add_page_to_unevictable_list(page
);
531 * When racing with an mlock clearing (page is
532 * unlocked), make sure that if the other thread does
533 * not observe our setting of PG_lru and fails
534 * isolation, we see PG_mlocked cleared below and move
535 * the page back to the evictable list.
537 * The other side is TestClearPageMlocked().
543 * page's status can change while we move it among lru. If an evictable
544 * page is on unevictable list, it never be freed. To avoid that,
545 * check after we added it to the list, again.
547 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
548 if (!isolate_lru_page(page
)) {
552 /* This means someone else dropped this page from LRU
553 * So, it will be freed or putback to LRU again. There is
554 * nothing to do here.
558 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
559 count_vm_event(UNEVICTABLE_PGRESCUED
);
560 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
561 count_vm_event(UNEVICTABLE_PGCULLED
);
563 put_page(page
); /* drop ref from isolate */
566 enum page_references
{
568 PAGEREF_RECLAIM_CLEAN
,
573 static enum page_references
page_check_references(struct page
*page
,
574 struct scan_control
*sc
)
576 int referenced_ptes
, referenced_page
;
577 unsigned long vm_flags
;
579 referenced_ptes
= page_referenced(page
, 1, sc
->mem_cgroup
, &vm_flags
);
580 referenced_page
= TestClearPageReferenced(page
);
582 /* Lumpy reclaim - ignore references */
583 if (sc
->lumpy_reclaim_mode
)
584 return PAGEREF_RECLAIM
;
587 * Mlock lost the isolation race with us. Let try_to_unmap()
588 * move the page to the unevictable list.
590 if (vm_flags
& VM_LOCKED
)
591 return PAGEREF_RECLAIM
;
593 if (referenced_ptes
) {
595 return PAGEREF_ACTIVATE
;
597 * All mapped pages start out with page table
598 * references from the instantiating fault, so we need
599 * to look twice if a mapped file page is used more
602 * Mark it and spare it for another trip around the
603 * inactive list. Another page table reference will
604 * lead to its activation.
606 * Note: the mark is set for activated pages as well
607 * so that recently deactivated but used pages are
610 SetPageReferenced(page
);
613 return PAGEREF_ACTIVATE
;
618 /* Reclaim if clean, defer dirty pages to writeback */
620 return PAGEREF_RECLAIM_CLEAN
;
622 return PAGEREF_RECLAIM
;
625 static noinline_for_stack
void free_page_list(struct list_head
*free_pages
)
627 struct pagevec freed_pvec
;
628 struct page
*page
, *tmp
;
630 pagevec_init(&freed_pvec
, 1);
632 list_for_each_entry_safe(page
, tmp
, free_pages
, lru
) {
633 list_del(&page
->lru
);
634 if (!pagevec_add(&freed_pvec
, page
)) {
635 __pagevec_free(&freed_pvec
);
636 pagevec_reinit(&freed_pvec
);
640 pagevec_free(&freed_pvec
);
644 * shrink_page_list() returns the number of reclaimed pages
646 static unsigned long shrink_page_list(struct list_head
*page_list
,
647 struct scan_control
*sc
,
648 enum pageout_io sync_writeback
)
650 LIST_HEAD(ret_pages
);
651 LIST_HEAD(free_pages
);
653 unsigned long nr_reclaimed
= 0;
657 while (!list_empty(page_list
)) {
658 enum page_references references
;
659 struct address_space
*mapping
;
665 page
= lru_to_page(page_list
);
666 list_del(&page
->lru
);
668 if (!trylock_page(page
))
671 VM_BUG_ON(PageActive(page
));
675 if (unlikely(!page_evictable(page
, NULL
)))
678 if (!sc
->may_unmap
&& page_mapped(page
))
681 /* Double the slab pressure for mapped and swapcache pages */
682 if (page_mapped(page
) || PageSwapCache(page
))
685 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
686 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
688 if (PageWriteback(page
)) {
690 * Synchronous reclaim is performed in two passes,
691 * first an asynchronous pass over the list to
692 * start parallel writeback, and a second synchronous
693 * pass to wait for the IO to complete. Wait here
694 * for any page for which writeback has already
697 if (sync_writeback
== PAGEOUT_IO_SYNC
&& may_enter_fs
)
698 wait_on_page_writeback(page
);
703 references
= page_check_references(page
, sc
);
704 switch (references
) {
705 case PAGEREF_ACTIVATE
:
706 goto activate_locked
;
709 case PAGEREF_RECLAIM
:
710 case PAGEREF_RECLAIM_CLEAN
:
711 ; /* try to reclaim the page below */
715 * Anonymous process memory has backing store?
716 * Try to allocate it some swap space here.
718 if (PageAnon(page
) && !PageSwapCache(page
)) {
719 if (!(sc
->gfp_mask
& __GFP_IO
))
721 if (!add_to_swap(page
))
722 goto activate_locked
;
726 mapping
= page_mapping(page
);
729 * The page is mapped into the page tables of one or more
730 * processes. Try to unmap it here.
732 if (page_mapped(page
) && mapping
) {
733 switch (try_to_unmap(page
, TTU_UNMAP
)) {
735 goto activate_locked
;
741 ; /* try to free the page below */
745 if (PageDirty(page
)) {
746 if (references
== PAGEREF_RECLAIM_CLEAN
)
750 if (!sc
->may_writepage
)
753 /* Page is dirty, try to write it out here */
754 switch (pageout(page
, mapping
, sync_writeback
)) {
758 goto activate_locked
;
760 if (PageWriteback(page
) || PageDirty(page
))
763 * A synchronous write - probably a ramdisk. Go
764 * ahead and try to reclaim the page.
766 if (!trylock_page(page
))
768 if (PageDirty(page
) || PageWriteback(page
))
770 mapping
= page_mapping(page
);
772 ; /* try to free the page below */
777 * If the page has buffers, try to free the buffer mappings
778 * associated with this page. If we succeed we try to free
781 * We do this even if the page is PageDirty().
782 * try_to_release_page() does not perform I/O, but it is
783 * possible for a page to have PageDirty set, but it is actually
784 * clean (all its buffers are clean). This happens if the
785 * buffers were written out directly, with submit_bh(). ext3
786 * will do this, as well as the blockdev mapping.
787 * try_to_release_page() will discover that cleanness and will
788 * drop the buffers and mark the page clean - it can be freed.
790 * Rarely, pages can have buffers and no ->mapping. These are
791 * the pages which were not successfully invalidated in
792 * truncate_complete_page(). We try to drop those buffers here
793 * and if that worked, and the page is no longer mapped into
794 * process address space (page_count == 1) it can be freed.
795 * Otherwise, leave the page on the LRU so it is swappable.
797 if (page_has_private(page
)) {
798 if (!try_to_release_page(page
, sc
->gfp_mask
))
799 goto activate_locked
;
800 if (!mapping
&& page_count(page
) == 1) {
802 if (put_page_testzero(page
))
806 * rare race with speculative reference.
807 * the speculative reference will free
808 * this page shortly, so we may
809 * increment nr_reclaimed here (and
810 * leave it off the LRU).
818 if (!mapping
|| !__remove_mapping(mapping
, page
))
822 * At this point, we have no other references and there is
823 * no way to pick any more up (removed from LRU, removed
824 * from pagecache). Can use non-atomic bitops now (and
825 * we obviously don't have to worry about waking up a process
826 * waiting on the page lock, because there are no references.
828 __clear_page_locked(page
);
833 * Is there need to periodically free_page_list? It would
834 * appear not as the counts should be low
836 list_add(&page
->lru
, &free_pages
);
840 if (PageSwapCache(page
))
841 try_to_free_swap(page
);
843 putback_lru_page(page
);
847 /* Not a candidate for swapping, so reclaim swap space. */
848 if (PageSwapCache(page
) && vm_swap_full())
849 try_to_free_swap(page
);
850 VM_BUG_ON(PageActive(page
));
856 list_add(&page
->lru
, &ret_pages
);
857 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
860 free_page_list(&free_pages
);
862 list_splice(&ret_pages
, page_list
);
863 count_vm_events(PGACTIVATE
, pgactivate
);
868 * Attempt to remove the specified page from its LRU. Only take this page
869 * if it is of the appropriate PageActive status. Pages which are being
870 * freed elsewhere are also ignored.
872 * page: page to consider
873 * mode: one of the LRU isolation modes defined above
875 * returns 0 on success, -ve errno on failure.
877 int __isolate_lru_page(struct page
*page
, int mode
, int file
)
881 /* Only take pages on the LRU. */
886 * When checking the active state, we need to be sure we are
887 * dealing with comparible boolean values. Take the logical not
890 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
893 if (mode
!= ISOLATE_BOTH
&& page_is_file_cache(page
) != file
)
897 * When this function is being called for lumpy reclaim, we
898 * initially look into all LRU pages, active, inactive and
899 * unevictable; only give shrink_page_list evictable pages.
901 if (PageUnevictable(page
))
906 if (likely(get_page_unless_zero(page
))) {
908 * Be careful not to clear PageLRU until after we're
909 * sure the page is not being freed elsewhere -- the
910 * page release code relies on it.
920 * zone->lru_lock is heavily contended. Some of the functions that
921 * shrink the lists perform better by taking out a batch of pages
922 * and working on them outside the LRU lock.
924 * For pagecache intensive workloads, this function is the hottest
925 * spot in the kernel (apart from copy_*_user functions).
927 * Appropriate locks must be held before calling this function.
929 * @nr_to_scan: The number of pages to look through on the list.
930 * @src: The LRU list to pull pages off.
931 * @dst: The temp list to put pages on to.
932 * @scanned: The number of pages that were scanned.
933 * @order: The caller's attempted allocation order
934 * @mode: One of the LRU isolation modes
935 * @file: True [1] if isolating file [!anon] pages
937 * returns how many pages were moved onto *@dst.
939 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
940 struct list_head
*src
, struct list_head
*dst
,
941 unsigned long *scanned
, int order
, int mode
, int file
)
943 unsigned long nr_taken
= 0;
944 unsigned long nr_lumpy_taken
= 0;
945 unsigned long nr_lumpy_dirty
= 0;
946 unsigned long nr_lumpy_failed
= 0;
949 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
952 unsigned long end_pfn
;
953 unsigned long page_pfn
;
956 page
= lru_to_page(src
);
957 prefetchw_prev_lru_page(page
, src
, flags
);
959 VM_BUG_ON(!PageLRU(page
));
961 switch (__isolate_lru_page(page
, mode
, file
)) {
963 list_move(&page
->lru
, dst
);
964 mem_cgroup_del_lru(page
);
969 /* else it is being freed elsewhere */
970 list_move(&page
->lru
, src
);
971 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
982 * Attempt to take all pages in the order aligned region
983 * surrounding the tag page. Only take those pages of
984 * the same active state as that tag page. We may safely
985 * round the target page pfn down to the requested order
986 * as the mem_map is guarenteed valid out to MAX_ORDER,
987 * where that page is in a different zone we will detect
988 * it from its zone id and abort this block scan.
990 zone_id
= page_zone_id(page
);
991 page_pfn
= page_to_pfn(page
);
992 pfn
= page_pfn
& ~((1 << order
) - 1);
993 end_pfn
= pfn
+ (1 << order
);
994 for (; pfn
< end_pfn
; pfn
++) {
995 struct page
*cursor_page
;
997 /* The target page is in the block, ignore it. */
998 if (unlikely(pfn
== page_pfn
))
1001 /* Avoid holes within the zone. */
1002 if (unlikely(!pfn_valid_within(pfn
)))
1005 cursor_page
= pfn_to_page(pfn
);
1007 /* Check that we have not crossed a zone boundary. */
1008 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
1012 * If we don't have enough swap space, reclaiming of
1013 * anon page which don't already have a swap slot is
1016 if (nr_swap_pages
<= 0 && PageAnon(cursor_page
) &&
1017 !PageSwapCache(cursor_page
))
1020 if (__isolate_lru_page(cursor_page
, mode
, file
) == 0) {
1021 list_move(&cursor_page
->lru
, dst
);
1022 mem_cgroup_del_lru(cursor_page
);
1025 if (PageDirty(cursor_page
))
1029 if (mode
== ISOLATE_BOTH
&&
1030 page_count(cursor_page
))
1038 trace_mm_vmscan_lru_isolate(order
,
1041 nr_lumpy_taken
, nr_lumpy_dirty
, nr_lumpy_failed
,
1046 static unsigned long isolate_pages_global(unsigned long nr
,
1047 struct list_head
*dst
,
1048 unsigned long *scanned
, int order
,
1049 int mode
, struct zone
*z
,
1050 int active
, int file
)
1057 return isolate_lru_pages(nr
, &z
->lru
[lru
].list
, dst
, scanned
, order
,
1062 * clear_active_flags() is a helper for shrink_active_list(), clearing
1063 * any active bits from the pages in the list.
1065 static unsigned long clear_active_flags(struct list_head
*page_list
,
1066 unsigned int *count
)
1072 list_for_each_entry(page
, page_list
, lru
) {
1073 lru
= page_lru_base_type(page
);
1074 if (PageActive(page
)) {
1076 ClearPageActive(page
);
1087 * isolate_lru_page - tries to isolate a page from its LRU list
1088 * @page: page to isolate from its LRU list
1090 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1091 * vmstat statistic corresponding to whatever LRU list the page was on.
1093 * Returns 0 if the page was removed from an LRU list.
1094 * Returns -EBUSY if the page was not on an LRU list.
1096 * The returned page will have PageLRU() cleared. If it was found on
1097 * the active list, it will have PageActive set. If it was found on
1098 * the unevictable list, it will have the PageUnevictable bit set. That flag
1099 * may need to be cleared by the caller before letting the page go.
1101 * The vmstat statistic corresponding to the list on which the page was
1102 * found will be decremented.
1105 * (1) Must be called with an elevated refcount on the page. This is a
1106 * fundamentnal difference from isolate_lru_pages (which is called
1107 * without a stable reference).
1108 * (2) the lru_lock must not be held.
1109 * (3) interrupts must be enabled.
1111 int isolate_lru_page(struct page
*page
)
1115 if (PageLRU(page
)) {
1116 struct zone
*zone
= page_zone(page
);
1118 spin_lock_irq(&zone
->lru_lock
);
1119 if (PageLRU(page
) && get_page_unless_zero(page
)) {
1120 int lru
= page_lru(page
);
1124 del_page_from_lru_list(zone
, page
, lru
);
1126 spin_unlock_irq(&zone
->lru_lock
);
1132 * Are there way too many processes in the direct reclaim path already?
1134 static int too_many_isolated(struct zone
*zone
, int file
,
1135 struct scan_control
*sc
)
1137 unsigned long inactive
, isolated
;
1139 if (current_is_kswapd())
1142 if (!scanning_global_lru(sc
))
1146 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1147 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1149 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1150 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1153 return isolated
> inactive
;
1157 * TODO: Try merging with migrations version of putback_lru_pages
1159 static noinline_for_stack
void
1160 putback_lru_pages(struct zone
*zone
, struct scan_control
*sc
,
1161 unsigned long nr_anon
, unsigned long nr_file
,
1162 struct list_head
*page_list
)
1165 struct pagevec pvec
;
1166 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1168 pagevec_init(&pvec
, 1);
1171 * Put back any unfreeable pages.
1173 spin_lock(&zone
->lru_lock
);
1174 while (!list_empty(page_list
)) {
1176 page
= lru_to_page(page_list
);
1177 VM_BUG_ON(PageLRU(page
));
1178 list_del(&page
->lru
);
1179 if (unlikely(!page_evictable(page
, NULL
))) {
1180 spin_unlock_irq(&zone
->lru_lock
);
1181 putback_lru_page(page
);
1182 spin_lock_irq(&zone
->lru_lock
);
1186 lru
= page_lru(page
);
1187 add_page_to_lru_list(zone
, page
, lru
);
1188 if (is_active_lru(lru
)) {
1189 int file
= is_file_lru(lru
);
1190 reclaim_stat
->recent_rotated
[file
]++;
1192 if (!pagevec_add(&pvec
, page
)) {
1193 spin_unlock_irq(&zone
->lru_lock
);
1194 __pagevec_release(&pvec
);
1195 spin_lock_irq(&zone
->lru_lock
);
1198 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1199 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1201 spin_unlock_irq(&zone
->lru_lock
);
1202 pagevec_release(&pvec
);
1205 static noinline_for_stack
void update_isolated_counts(struct zone
*zone
,
1206 struct scan_control
*sc
,
1207 unsigned long *nr_anon
,
1208 unsigned long *nr_file
,
1209 struct list_head
*isolated_list
)
1211 unsigned long nr_active
;
1212 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1213 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1215 nr_active
= clear_active_flags(isolated_list
, count
);
1216 __count_vm_events(PGDEACTIVATE
, nr_active
);
1218 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1219 -count
[LRU_ACTIVE_FILE
]);
1220 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1221 -count
[LRU_INACTIVE_FILE
]);
1222 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1223 -count
[LRU_ACTIVE_ANON
]);
1224 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1225 -count
[LRU_INACTIVE_ANON
]);
1227 *nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1228 *nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1229 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, *nr_anon
);
1230 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, *nr_file
);
1232 reclaim_stat
->recent_scanned
[0] += *nr_anon
;
1233 reclaim_stat
->recent_scanned
[1] += *nr_file
;
1237 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1238 * of reclaimed pages
1240 static noinline_for_stack
unsigned long
1241 shrink_inactive_list(unsigned long nr_to_scan
, struct zone
*zone
,
1242 struct scan_control
*sc
, int priority
, int file
)
1244 LIST_HEAD(page_list
);
1245 unsigned long nr_scanned
;
1246 unsigned long nr_reclaimed
= 0;
1247 unsigned long nr_taken
;
1248 unsigned long nr_active
;
1249 unsigned long nr_anon
;
1250 unsigned long nr_file
;
1252 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1253 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1255 /* We are about to die and free our memory. Return now. */
1256 if (fatal_signal_pending(current
))
1257 return SWAP_CLUSTER_MAX
;
1262 spin_lock_irq(&zone
->lru_lock
);
1264 if (scanning_global_lru(sc
)) {
1265 nr_taken
= isolate_pages_global(nr_to_scan
,
1266 &page_list
, &nr_scanned
, sc
->order
,
1267 sc
->lumpy_reclaim_mode
?
1268 ISOLATE_BOTH
: ISOLATE_INACTIVE
,
1270 zone
->pages_scanned
+= nr_scanned
;
1271 if (current_is_kswapd())
1272 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1275 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1278 nr_taken
= mem_cgroup_isolate_pages(nr_to_scan
,
1279 &page_list
, &nr_scanned
, sc
->order
,
1280 sc
->lumpy_reclaim_mode
?
1281 ISOLATE_BOTH
: ISOLATE_INACTIVE
,
1282 zone
, sc
->mem_cgroup
,
1285 * mem_cgroup_isolate_pages() keeps track of
1286 * scanned pages on its own.
1290 if (nr_taken
== 0) {
1291 spin_unlock_irq(&zone
->lru_lock
);
1295 update_isolated_counts(zone
, sc
, &nr_anon
, &nr_file
, &page_list
);
1297 spin_unlock_irq(&zone
->lru_lock
);
1299 nr_reclaimed
= shrink_page_list(&page_list
, sc
, PAGEOUT_IO_ASYNC
);
1302 * If we are direct reclaiming for contiguous pages and we do
1303 * not reclaim everything in the list, try again and wait
1304 * for IO to complete. This will stall high-order allocations
1305 * but that should be acceptable to the caller
1307 if (nr_reclaimed
< nr_taken
&& !current_is_kswapd() &&
1308 sc
->lumpy_reclaim_mode
) {
1309 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1312 * The attempt at page out may have made some
1313 * of the pages active, mark them inactive again.
1315 nr_active
= clear_active_flags(&page_list
, NULL
);
1316 count_vm_events(PGDEACTIVATE
, nr_active
);
1318 nr_reclaimed
+= shrink_page_list(&page_list
, sc
, PAGEOUT_IO_SYNC
);
1321 local_irq_disable();
1322 if (current_is_kswapd())
1323 __count_vm_events(KSWAPD_STEAL
, nr_reclaimed
);
1324 __count_zone_vm_events(PGSTEAL
, zone
, nr_reclaimed
);
1326 putback_lru_pages(zone
, sc
, nr_anon
, nr_file
, &page_list
);
1327 return nr_reclaimed
;
1331 * This moves pages from the active list to the inactive list.
1333 * We move them the other way if the page is referenced by one or more
1334 * processes, from rmap.
1336 * If the pages are mostly unmapped, the processing is fast and it is
1337 * appropriate to hold zone->lru_lock across the whole operation. But if
1338 * the pages are mapped, the processing is slow (page_referenced()) so we
1339 * should drop zone->lru_lock around each page. It's impossible to balance
1340 * this, so instead we remove the pages from the LRU while processing them.
1341 * It is safe to rely on PG_active against the non-LRU pages in here because
1342 * nobody will play with that bit on a non-LRU page.
1344 * The downside is that we have to touch page->_count against each page.
1345 * But we had to alter page->flags anyway.
1348 static void move_active_pages_to_lru(struct zone
*zone
,
1349 struct list_head
*list
,
1352 unsigned long pgmoved
= 0;
1353 struct pagevec pvec
;
1356 pagevec_init(&pvec
, 1);
1358 while (!list_empty(list
)) {
1359 page
= lru_to_page(list
);
1361 VM_BUG_ON(PageLRU(page
));
1364 list_move(&page
->lru
, &zone
->lru
[lru
].list
);
1365 mem_cgroup_add_lru_list(page
, lru
);
1368 if (!pagevec_add(&pvec
, page
) || list_empty(list
)) {
1369 spin_unlock_irq(&zone
->lru_lock
);
1370 if (buffer_heads_over_limit
)
1371 pagevec_strip(&pvec
);
1372 __pagevec_release(&pvec
);
1373 spin_lock_irq(&zone
->lru_lock
);
1376 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1377 if (!is_active_lru(lru
))
1378 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1381 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1382 struct scan_control
*sc
, int priority
, int file
)
1384 unsigned long nr_taken
;
1385 unsigned long pgscanned
;
1386 unsigned long vm_flags
;
1387 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1388 LIST_HEAD(l_active
);
1389 LIST_HEAD(l_inactive
);
1391 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1392 unsigned long nr_rotated
= 0;
1395 spin_lock_irq(&zone
->lru_lock
);
1396 if (scanning_global_lru(sc
)) {
1397 nr_taken
= isolate_pages_global(nr_pages
, &l_hold
,
1398 &pgscanned
, sc
->order
,
1399 ISOLATE_ACTIVE
, zone
,
1401 zone
->pages_scanned
+= pgscanned
;
1403 nr_taken
= mem_cgroup_isolate_pages(nr_pages
, &l_hold
,
1404 &pgscanned
, sc
->order
,
1405 ISOLATE_ACTIVE
, zone
,
1406 sc
->mem_cgroup
, 1, file
);
1408 * mem_cgroup_isolate_pages() keeps track of
1409 * scanned pages on its own.
1413 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1415 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1417 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1419 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1420 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1421 spin_unlock_irq(&zone
->lru_lock
);
1423 while (!list_empty(&l_hold
)) {
1425 page
= lru_to_page(&l_hold
);
1426 list_del(&page
->lru
);
1428 if (unlikely(!page_evictable(page
, NULL
))) {
1429 putback_lru_page(page
);
1433 if (page_referenced(page
, 0, sc
->mem_cgroup
, &vm_flags
)) {
1436 * Identify referenced, file-backed active pages and
1437 * give them one more trip around the active list. So
1438 * that executable code get better chances to stay in
1439 * memory under moderate memory pressure. Anon pages
1440 * are not likely to be evicted by use-once streaming
1441 * IO, plus JVM can create lots of anon VM_EXEC pages,
1442 * so we ignore them here.
1444 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1445 list_add(&page
->lru
, &l_active
);
1450 ClearPageActive(page
); /* we are de-activating */
1451 list_add(&page
->lru
, &l_inactive
);
1455 * Move pages back to the lru list.
1457 spin_lock_irq(&zone
->lru_lock
);
1459 * Count referenced pages from currently used mappings as rotated,
1460 * even though only some of them are actually re-activated. This
1461 * helps balance scan pressure between file and anonymous pages in
1464 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1466 move_active_pages_to_lru(zone
, &l_active
,
1467 LRU_ACTIVE
+ file
* LRU_FILE
);
1468 move_active_pages_to_lru(zone
, &l_inactive
,
1469 LRU_BASE
+ file
* LRU_FILE
);
1470 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1471 spin_unlock_irq(&zone
->lru_lock
);
1474 static int inactive_anon_is_low_global(struct zone
*zone
)
1476 unsigned long active
, inactive
;
1478 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1479 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1481 if (inactive
* zone
->inactive_ratio
< active
)
1488 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1489 * @zone: zone to check
1490 * @sc: scan control of this context
1492 * Returns true if the zone does not have enough inactive anon pages,
1493 * meaning some active anon pages need to be deactivated.
1495 static int inactive_anon_is_low(struct zone
*zone
, struct scan_control
*sc
)
1499 if (scanning_global_lru(sc
))
1500 low
= inactive_anon_is_low_global(zone
);
1502 low
= mem_cgroup_inactive_anon_is_low(sc
->mem_cgroup
);
1506 static int inactive_file_is_low_global(struct zone
*zone
)
1508 unsigned long active
, inactive
;
1510 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1511 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1513 return (active
> inactive
);
1517 * inactive_file_is_low - check if file pages need to be deactivated
1518 * @zone: zone to check
1519 * @sc: scan control of this context
1521 * When the system is doing streaming IO, memory pressure here
1522 * ensures that active file pages get deactivated, until more
1523 * than half of the file pages are on the inactive list.
1525 * Once we get to that situation, protect the system's working
1526 * set from being evicted by disabling active file page aging.
1528 * This uses a different ratio than the anonymous pages, because
1529 * the page cache uses a use-once replacement algorithm.
1531 static int inactive_file_is_low(struct zone
*zone
, struct scan_control
*sc
)
1535 if (scanning_global_lru(sc
))
1536 low
= inactive_file_is_low_global(zone
);
1538 low
= mem_cgroup_inactive_file_is_low(sc
->mem_cgroup
);
1542 static int inactive_list_is_low(struct zone
*zone
, struct scan_control
*sc
,
1546 return inactive_file_is_low(zone
, sc
);
1548 return inactive_anon_is_low(zone
, sc
);
1551 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1552 struct zone
*zone
, struct scan_control
*sc
, int priority
)
1554 int file
= is_file_lru(lru
);
1556 if (is_active_lru(lru
)) {
1557 if (inactive_list_is_low(zone
, sc
, file
))
1558 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1562 return shrink_inactive_list(nr_to_scan
, zone
, sc
, priority
, file
);
1566 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1567 * until we collected @swap_cluster_max pages to scan.
1569 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan
,
1570 unsigned long *nr_saved_scan
)
1574 *nr_saved_scan
+= nr_to_scan
;
1575 nr
= *nr_saved_scan
;
1577 if (nr
>= SWAP_CLUSTER_MAX
)
1586 * Determine how aggressively the anon and file LRU lists should be
1587 * scanned. The relative value of each set of LRU lists is determined
1588 * by looking at the fraction of the pages scanned we did rotate back
1589 * onto the active list instead of evict.
1591 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1593 static void get_scan_count(struct zone
*zone
, struct scan_control
*sc
,
1594 unsigned long *nr
, int priority
)
1596 unsigned long anon
, file
, free
;
1597 unsigned long anon_prio
, file_prio
;
1598 unsigned long ap
, fp
;
1599 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1600 u64 fraction
[2], denominator
;
1604 /* If we have no swap space, do not bother scanning anon pages. */
1605 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1613 anon
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_ANON
) +
1614 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_ANON
);
1615 file
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_FILE
) +
1616 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1618 if (scanning_global_lru(sc
)) {
1619 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1620 /* If we have very few page cache pages,
1621 force-scan anon pages. */
1622 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1631 * With swappiness at 100, anonymous and file have the same priority.
1632 * This scanning priority is essentially the inverse of IO cost.
1634 anon_prio
= sc
->swappiness
;
1635 file_prio
= 200 - sc
->swappiness
;
1638 * OK, so we have swap space and a fair amount of page cache
1639 * pages. We use the recently rotated / recently scanned
1640 * ratios to determine how valuable each cache is.
1642 * Because workloads change over time (and to avoid overflow)
1643 * we keep these statistics as a floating average, which ends
1644 * up weighing recent references more than old ones.
1646 * anon in [0], file in [1]
1648 spin_lock_irq(&zone
->lru_lock
);
1649 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1650 reclaim_stat
->recent_scanned
[0] /= 2;
1651 reclaim_stat
->recent_rotated
[0] /= 2;
1654 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1655 reclaim_stat
->recent_scanned
[1] /= 2;
1656 reclaim_stat
->recent_rotated
[1] /= 2;
1660 * The amount of pressure on anon vs file pages is inversely
1661 * proportional to the fraction of recently scanned pages on
1662 * each list that were recently referenced and in active use.
1664 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1665 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1667 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1668 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1669 spin_unlock_irq(&zone
->lru_lock
);
1673 denominator
= ap
+ fp
+ 1;
1675 for_each_evictable_lru(l
) {
1676 int file
= is_file_lru(l
);
1679 scan
= zone_nr_lru_pages(zone
, sc
, l
);
1680 if (priority
|| noswap
) {
1682 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1684 nr
[l
] = nr_scan_try_batch(scan
,
1685 &reclaim_stat
->nr_saved_scan
[l
]);
1689 static void set_lumpy_reclaim_mode(int priority
, struct scan_control
*sc
)
1692 * If we need a large contiguous chunk of memory, or have
1693 * trouble getting a small set of contiguous pages, we
1694 * will reclaim both active and inactive pages.
1696 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
1697 sc
->lumpy_reclaim_mode
= 1;
1698 else if (sc
->order
&& priority
< DEF_PRIORITY
- 2)
1699 sc
->lumpy_reclaim_mode
= 1;
1701 sc
->lumpy_reclaim_mode
= 0;
1705 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1707 static void shrink_zone(int priority
, struct zone
*zone
,
1708 struct scan_control
*sc
)
1710 unsigned long nr
[NR_LRU_LISTS
];
1711 unsigned long nr_to_scan
;
1713 unsigned long nr_reclaimed
= sc
->nr_reclaimed
;
1714 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1716 get_scan_count(zone
, sc
, nr
, priority
);
1718 set_lumpy_reclaim_mode(priority
, sc
);
1720 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1721 nr
[LRU_INACTIVE_FILE
]) {
1722 for_each_evictable_lru(l
) {
1724 nr_to_scan
= min_t(unsigned long,
1725 nr
[l
], SWAP_CLUSTER_MAX
);
1726 nr
[l
] -= nr_to_scan
;
1728 nr_reclaimed
+= shrink_list(l
, nr_to_scan
,
1729 zone
, sc
, priority
);
1733 * On large memory systems, scan >> priority can become
1734 * really large. This is fine for the starting priority;
1735 * we want to put equal scanning pressure on each zone.
1736 * However, if the VM has a harder time of freeing pages,
1737 * with multiple processes reclaiming pages, the total
1738 * freeing target can get unreasonably large.
1740 if (nr_reclaimed
>= nr_to_reclaim
&& priority
< DEF_PRIORITY
)
1744 sc
->nr_reclaimed
= nr_reclaimed
;
1747 * Even if we did not try to evict anon pages at all, we want to
1748 * rebalance the anon lru active/inactive ratio.
1750 if (inactive_anon_is_low(zone
, sc
) && nr_swap_pages
> 0)
1751 shrink_active_list(SWAP_CLUSTER_MAX
, zone
, sc
, priority
, 0);
1753 throttle_vm_writeout(sc
->gfp_mask
);
1757 * This is the direct reclaim path, for page-allocating processes. We only
1758 * try to reclaim pages from zones which will satisfy the caller's allocation
1761 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1763 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1765 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1766 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1767 * zone defense algorithm.
1769 * If a zone is deemed to be full of pinned pages then just give it a light
1770 * scan then give up on it.
1772 static bool shrink_zones(int priority
, struct zonelist
*zonelist
,
1773 struct scan_control
*sc
)
1777 bool all_unreclaimable
= true;
1779 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1780 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
1781 if (!populated_zone(zone
))
1784 * Take care memory controller reclaiming has small influence
1787 if (scanning_global_lru(sc
)) {
1788 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1790 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1791 continue; /* Let kswapd poll it */
1794 shrink_zone(priority
, zone
, sc
);
1795 all_unreclaimable
= false;
1797 return all_unreclaimable
;
1801 * This is the main entry point to direct page reclaim.
1803 * If a full scan of the inactive list fails to free enough memory then we
1804 * are "out of memory" and something needs to be killed.
1806 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1807 * high - the zone may be full of dirty or under-writeback pages, which this
1808 * caller can't do much about. We kick the writeback threads and take explicit
1809 * naps in the hope that some of these pages can be written. But if the
1810 * allocating task holds filesystem locks which prevent writeout this might not
1811 * work, and the allocation attempt will fail.
1813 * returns: 0, if no pages reclaimed
1814 * else, the number of pages reclaimed
1816 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
1817 struct scan_control
*sc
)
1820 bool all_unreclaimable
;
1821 unsigned long total_scanned
= 0;
1822 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1825 unsigned long writeback_threshold
;
1828 delayacct_freepages_start();
1830 if (scanning_global_lru(sc
))
1831 count_vm_event(ALLOCSTALL
);
1833 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1836 disable_swap_token();
1837 all_unreclaimable
= shrink_zones(priority
, zonelist
, sc
);
1839 * Don't shrink slabs when reclaiming memory from
1840 * over limit cgroups
1842 if (scanning_global_lru(sc
)) {
1843 unsigned long lru_pages
= 0;
1844 for_each_zone_zonelist(zone
, z
, zonelist
,
1845 gfp_zone(sc
->gfp_mask
)) {
1846 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1849 lru_pages
+= zone_reclaimable_pages(zone
);
1852 shrink_slab(sc
->nr_scanned
, sc
->gfp_mask
, lru_pages
);
1853 if (reclaim_state
) {
1854 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1855 reclaim_state
->reclaimed_slab
= 0;
1858 total_scanned
+= sc
->nr_scanned
;
1859 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
1863 * Try to write back as many pages as we just scanned. This
1864 * tends to cause slow streaming writers to write data to the
1865 * disk smoothly, at the dirtying rate, which is nice. But
1866 * that's undesirable in laptop mode, where we *want* lumpy
1867 * writeout. So in laptop mode, write out the whole world.
1869 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
1870 if (total_scanned
> writeback_threshold
) {
1871 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
);
1872 sc
->may_writepage
= 1;
1875 /* Take a nap, wait for some writeback to complete */
1876 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
1877 priority
< DEF_PRIORITY
- 2)
1878 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1883 * Now that we've scanned all the zones at this priority level, note
1884 * that level within the zone so that the next thread which performs
1885 * scanning of this zone will immediately start out at this priority
1886 * level. This affects only the decision whether or not to bring
1887 * mapped pages onto the inactive list.
1892 delayacct_freepages_end();
1895 if (sc
->nr_reclaimed
)
1896 return sc
->nr_reclaimed
;
1898 /* top priority shrink_zones still had more to do? don't OOM, then */
1899 if (scanning_global_lru(sc
) && !all_unreclaimable
)
1905 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
1906 gfp_t gfp_mask
, nodemask_t
*nodemask
)
1908 unsigned long nr_reclaimed
;
1909 struct scan_control sc
= {
1910 .gfp_mask
= gfp_mask
,
1911 .may_writepage
= !laptop_mode
,
1912 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
1915 .swappiness
= vm_swappiness
,
1918 .nodemask
= nodemask
,
1921 trace_mm_vmscan_direct_reclaim_begin(order
,
1925 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
1927 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
1929 return nr_reclaimed
;
1932 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1934 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*mem
,
1935 gfp_t gfp_mask
, bool noswap
,
1936 unsigned int swappiness
,
1937 struct zone
*zone
, int nid
)
1939 struct scan_control sc
= {
1940 .may_writepage
= !laptop_mode
,
1942 .may_swap
= !noswap
,
1943 .swappiness
= swappiness
,
1947 nodemask_t nm
= nodemask_of_node(nid
);
1949 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1950 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1952 sc
.nr_reclaimed
= 0;
1955 * NOTE: Although we can get the priority field, using it
1956 * here is not a good idea, since it limits the pages we can scan.
1957 * if we don't reclaim here, the shrink_zone from balance_pgdat
1958 * will pick up pages from other mem cgroup's as well. We hack
1959 * the priority and make it zero.
1961 shrink_zone(0, zone
, &sc
);
1962 return sc
.nr_reclaimed
;
1965 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*mem_cont
,
1968 unsigned int swappiness
)
1970 struct zonelist
*zonelist
;
1971 struct scan_control sc
= {
1972 .may_writepage
= !laptop_mode
,
1974 .may_swap
= !noswap
,
1975 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
1976 .swappiness
= swappiness
,
1978 .mem_cgroup
= mem_cont
,
1979 .nodemask
= NULL
, /* we don't care the placement */
1982 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1983 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1984 zonelist
= NODE_DATA(numa_node_id())->node_zonelists
;
1985 return do_try_to_free_pages(zonelist
, &sc
);
1989 /* is kswapd sleeping prematurely? */
1990 static int sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
)
1994 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1998 /* If after HZ/10, a zone is below the high mark, it's premature */
1999 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
2000 struct zone
*zone
= pgdat
->node_zones
+ i
;
2002 if (!populated_zone(zone
))
2005 if (zone
->all_unreclaimable
)
2008 if (!zone_watermark_ok(zone
, order
, high_wmark_pages(zone
),
2017 * For kswapd, balance_pgdat() will work across all this node's zones until
2018 * they are all at high_wmark_pages(zone).
2020 * Returns the number of pages which were actually freed.
2022 * There is special handling here for zones which are full of pinned pages.
2023 * This can happen if the pages are all mlocked, or if they are all used by
2024 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2025 * What we do is to detect the case where all pages in the zone have been
2026 * scanned twice and there has been zero successful reclaim. Mark the zone as
2027 * dead and from now on, only perform a short scan. Basically we're polling
2028 * the zone for when the problem goes away.
2030 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2031 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2032 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2033 * lower zones regardless of the number of free pages in the lower zones. This
2034 * interoperates with the page allocator fallback scheme to ensure that aging
2035 * of pages is balanced across the zones.
2037 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
2042 unsigned long total_scanned
;
2043 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2044 struct scan_control sc
= {
2045 .gfp_mask
= GFP_KERNEL
,
2049 * kswapd doesn't want to be bailed out while reclaim. because
2050 * we want to put equal scanning pressure on each zone.
2052 .nr_to_reclaim
= ULONG_MAX
,
2053 .swappiness
= vm_swappiness
,
2059 sc
.nr_reclaimed
= 0;
2060 sc
.may_writepage
= !laptop_mode
;
2061 count_vm_event(PAGEOUTRUN
);
2063 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2064 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2065 unsigned long lru_pages
= 0;
2066 int has_under_min_watermark_zone
= 0;
2068 /* The swap token gets in the way of swapout... */
2070 disable_swap_token();
2075 * Scan in the highmem->dma direction for the highest
2076 * zone which needs scanning
2078 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2079 struct zone
*zone
= pgdat
->node_zones
+ i
;
2081 if (!populated_zone(zone
))
2084 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2088 * Do some background aging of the anon list, to give
2089 * pages a chance to be referenced before reclaiming.
2091 if (inactive_anon_is_low(zone
, &sc
))
2092 shrink_active_list(SWAP_CLUSTER_MAX
, zone
,
2095 if (!zone_watermark_ok(zone
, order
,
2096 high_wmark_pages(zone
), 0, 0)) {
2104 for (i
= 0; i
<= end_zone
; i
++) {
2105 struct zone
*zone
= pgdat
->node_zones
+ i
;
2107 lru_pages
+= zone_reclaimable_pages(zone
);
2111 * Now scan the zone in the dma->highmem direction, stopping
2112 * at the last zone which needs scanning.
2114 * We do this because the page allocator works in the opposite
2115 * direction. This prevents the page allocator from allocating
2116 * pages behind kswapd's direction of progress, which would
2117 * cause too much scanning of the lower zones.
2119 for (i
= 0; i
<= end_zone
; i
++) {
2120 struct zone
*zone
= pgdat
->node_zones
+ i
;
2124 if (!populated_zone(zone
))
2127 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2132 nid
= pgdat
->node_id
;
2133 zid
= zone_idx(zone
);
2135 * Call soft limit reclaim before calling shrink_zone.
2136 * For now we ignore the return value
2138 mem_cgroup_soft_limit_reclaim(zone
, order
, sc
.gfp_mask
,
2141 * We put equal pressure on every zone, unless one
2142 * zone has way too many pages free already.
2144 if (!zone_watermark_ok(zone
, order
,
2145 8*high_wmark_pages(zone
), end_zone
, 0))
2146 shrink_zone(priority
, zone
, &sc
);
2147 reclaim_state
->reclaimed_slab
= 0;
2148 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
2150 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2151 total_scanned
+= sc
.nr_scanned
;
2152 if (zone
->all_unreclaimable
)
2155 zone
->pages_scanned
>= (zone_reclaimable_pages(zone
) * 6))
2156 zone
->all_unreclaimable
= 1;
2158 * If we've done a decent amount of scanning and
2159 * the reclaim ratio is low, start doing writepage
2160 * even in laptop mode
2162 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2163 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2164 sc
.may_writepage
= 1;
2166 if (!zone_watermark_ok(zone
, order
,
2167 high_wmark_pages(zone
), end_zone
, 0)) {
2170 * We are still under min water mark. This
2171 * means that we have a GFP_ATOMIC allocation
2172 * failure risk. Hurry up!
2174 if (!zone_watermark_ok(zone
, order
,
2175 min_wmark_pages(zone
), end_zone
, 0))
2176 has_under_min_watermark_zone
= 1;
2181 break; /* kswapd: all done */
2183 * OK, kswapd is getting into trouble. Take a nap, then take
2184 * another pass across the zones.
2186 if (total_scanned
&& (priority
< DEF_PRIORITY
- 2)) {
2187 if (has_under_min_watermark_zone
)
2188 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2190 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2194 * We do this so kswapd doesn't build up large priorities for
2195 * example when it is freeing in parallel with allocators. It
2196 * matches the direct reclaim path behaviour in terms of impact
2197 * on zone->*_priority.
2199 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2203 if (!all_zones_ok
) {
2209 * Fragmentation may mean that the system cannot be
2210 * rebalanced for high-order allocations in all zones.
2211 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2212 * it means the zones have been fully scanned and are still
2213 * not balanced. For high-order allocations, there is
2214 * little point trying all over again as kswapd may
2217 * Instead, recheck all watermarks at order-0 as they
2218 * are the most important. If watermarks are ok, kswapd will go
2219 * back to sleep. High-order users can still perform direct
2220 * reclaim if they wish.
2222 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2223 order
= sc
.order
= 0;
2228 return sc
.nr_reclaimed
;
2232 * The background pageout daemon, started as a kernel thread
2233 * from the init process.
2235 * This basically trickles out pages so that we have _some_
2236 * free memory available even if there is no other activity
2237 * that frees anything up. This is needed for things like routing
2238 * etc, where we otherwise might have all activity going on in
2239 * asynchronous contexts that cannot page things out.
2241 * If there are applications that are active memory-allocators
2242 * (most normal use), this basically shouldn't matter.
2244 static int kswapd(void *p
)
2246 unsigned long order
;
2247 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2248 struct task_struct
*tsk
= current
;
2250 struct reclaim_state reclaim_state
= {
2251 .reclaimed_slab
= 0,
2253 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2255 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2257 if (!cpumask_empty(cpumask
))
2258 set_cpus_allowed_ptr(tsk
, cpumask
);
2259 current
->reclaim_state
= &reclaim_state
;
2262 * Tell the memory management that we're a "memory allocator",
2263 * and that if we need more memory we should get access to it
2264 * regardless (see "__alloc_pages()"). "kswapd" should
2265 * never get caught in the normal page freeing logic.
2267 * (Kswapd normally doesn't need memory anyway, but sometimes
2268 * you need a small amount of memory in order to be able to
2269 * page out something else, and this flag essentially protects
2270 * us from recursively trying to free more memory as we're
2271 * trying to free the first piece of memory in the first place).
2273 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2278 unsigned long new_order
;
2281 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2282 new_order
= pgdat
->kswapd_max_order
;
2283 pgdat
->kswapd_max_order
= 0;
2284 if (order
< new_order
) {
2286 * Don't sleep if someone wants a larger 'order'
2291 if (!freezing(current
) && !kthread_should_stop()) {
2294 /* Try to sleep for a short interval */
2295 if (!sleeping_prematurely(pgdat
, order
, remaining
)) {
2296 remaining
= schedule_timeout(HZ
/10);
2297 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2298 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2302 * After a short sleep, check if it was a
2303 * premature sleep. If not, then go fully
2304 * to sleep until explicitly woken up
2306 if (!sleeping_prematurely(pgdat
, order
, remaining
)) {
2307 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2311 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2313 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2317 order
= pgdat
->kswapd_max_order
;
2319 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2321 ret
= try_to_freeze();
2322 if (kthread_should_stop())
2326 * We can speed up thawing tasks if we don't call balance_pgdat
2327 * after returning from the refrigerator
2330 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2331 balance_pgdat(pgdat
, order
);
2338 * A zone is low on free memory, so wake its kswapd task to service it.
2340 void wakeup_kswapd(struct zone
*zone
, int order
)
2344 if (!populated_zone(zone
))
2347 pgdat
= zone
->zone_pgdat
;
2348 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0, 0))
2350 if (pgdat
->kswapd_max_order
< order
)
2351 pgdat
->kswapd_max_order
= order
;
2352 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2353 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2355 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2357 wake_up_interruptible(&pgdat
->kswapd_wait
);
2361 * The reclaimable count would be mostly accurate.
2362 * The less reclaimable pages may be
2363 * - mlocked pages, which will be moved to unevictable list when encountered
2364 * - mapped pages, which may require several travels to be reclaimed
2365 * - dirty pages, which is not "instantly" reclaimable
2367 unsigned long global_reclaimable_pages(void)
2371 nr
= global_page_state(NR_ACTIVE_FILE
) +
2372 global_page_state(NR_INACTIVE_FILE
);
2374 if (nr_swap_pages
> 0)
2375 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2376 global_page_state(NR_INACTIVE_ANON
);
2381 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2385 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2386 zone_page_state(zone
, NR_INACTIVE_FILE
);
2388 if (nr_swap_pages
> 0)
2389 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2390 zone_page_state(zone
, NR_INACTIVE_ANON
);
2395 #ifdef CONFIG_HIBERNATION
2397 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2400 * Rather than trying to age LRUs the aim is to preserve the overall
2401 * LRU order by reclaiming preferentially
2402 * inactive > active > active referenced > active mapped
2404 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2406 struct reclaim_state reclaim_state
;
2407 struct scan_control sc
= {
2408 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2412 .nr_to_reclaim
= nr_to_reclaim
,
2413 .hibernation_mode
= 1,
2414 .swappiness
= vm_swappiness
,
2417 struct zonelist
* zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2418 struct task_struct
*p
= current
;
2419 unsigned long nr_reclaimed
;
2421 p
->flags
|= PF_MEMALLOC
;
2422 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2423 reclaim_state
.reclaimed_slab
= 0;
2424 p
->reclaim_state
= &reclaim_state
;
2426 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2428 p
->reclaim_state
= NULL
;
2429 lockdep_clear_current_reclaim_state();
2430 p
->flags
&= ~PF_MEMALLOC
;
2432 return nr_reclaimed
;
2434 #endif /* CONFIG_HIBERNATION */
2436 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2437 not required for correctness. So if the last cpu in a node goes
2438 away, we get changed to run anywhere: as the first one comes back,
2439 restore their cpu bindings. */
2440 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2441 unsigned long action
, void *hcpu
)
2445 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2446 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2447 pg_data_t
*pgdat
= NODE_DATA(nid
);
2448 const struct cpumask
*mask
;
2450 mask
= cpumask_of_node(pgdat
->node_id
);
2452 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2453 /* One of our CPUs online: restore mask */
2454 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2461 * This kswapd start function will be called by init and node-hot-add.
2462 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2464 int kswapd_run(int nid
)
2466 pg_data_t
*pgdat
= NODE_DATA(nid
);
2472 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2473 if (IS_ERR(pgdat
->kswapd
)) {
2474 /* failure at boot is fatal */
2475 BUG_ON(system_state
== SYSTEM_BOOTING
);
2476 printk("Failed to start kswapd on node %d\n",nid
);
2483 * Called by memory hotplug when all memory in a node is offlined.
2485 void kswapd_stop(int nid
)
2487 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
2490 kthread_stop(kswapd
);
2493 static int __init
kswapd_init(void)
2498 for_each_node_state(nid
, N_HIGH_MEMORY
)
2500 hotcpu_notifier(cpu_callback
, 0);
2504 module_init(kswapd_init
)
2510 * If non-zero call zone_reclaim when the number of free pages falls below
2513 int zone_reclaim_mode __read_mostly
;
2515 #define RECLAIM_OFF 0
2516 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2517 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2518 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2521 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2522 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2525 #define ZONE_RECLAIM_PRIORITY 4
2528 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2531 int sysctl_min_unmapped_ratio
= 1;
2534 * If the number of slab pages in a zone grows beyond this percentage then
2535 * slab reclaim needs to occur.
2537 int sysctl_min_slab_ratio
= 5;
2539 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
2541 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
2542 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
2543 zone_page_state(zone
, NR_ACTIVE_FILE
);
2546 * It's possible for there to be more file mapped pages than
2547 * accounted for by the pages on the file LRU lists because
2548 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2550 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
2553 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2554 static long zone_pagecache_reclaimable(struct zone
*zone
)
2556 long nr_pagecache_reclaimable
;
2560 * If RECLAIM_SWAP is set, then all file pages are considered
2561 * potentially reclaimable. Otherwise, we have to worry about
2562 * pages like swapcache and zone_unmapped_file_pages() provides
2565 if (zone_reclaim_mode
& RECLAIM_SWAP
)
2566 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
2568 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
2570 /* If we can't clean pages, remove dirty pages from consideration */
2571 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
2572 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
2574 /* Watch for any possible underflows due to delta */
2575 if (unlikely(delta
> nr_pagecache_reclaimable
))
2576 delta
= nr_pagecache_reclaimable
;
2578 return nr_pagecache_reclaimable
- delta
;
2582 * Try to free up some pages from this zone through reclaim.
2584 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2586 /* Minimum pages needed in order to stay on node */
2587 const unsigned long nr_pages
= 1 << order
;
2588 struct task_struct
*p
= current
;
2589 struct reclaim_state reclaim_state
;
2591 struct scan_control sc
= {
2592 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
2593 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
2595 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
2597 .gfp_mask
= gfp_mask
,
2598 .swappiness
= vm_swappiness
,
2601 unsigned long nr_slab_pages0
, nr_slab_pages1
;
2605 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2606 * and we also need to be able to write out pages for RECLAIM_WRITE
2609 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
2610 lockdep_set_current_reclaim_state(gfp_mask
);
2611 reclaim_state
.reclaimed_slab
= 0;
2612 p
->reclaim_state
= &reclaim_state
;
2614 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
2616 * Free memory by calling shrink zone with increasing
2617 * priorities until we have enough memory freed.
2619 priority
= ZONE_RECLAIM_PRIORITY
;
2621 shrink_zone(priority
, zone
, &sc
);
2623 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
2626 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2627 if (nr_slab_pages0
> zone
->min_slab_pages
) {
2629 * shrink_slab() does not currently allow us to determine how
2630 * many pages were freed in this zone. So we take the current
2631 * number of slab pages and shake the slab until it is reduced
2632 * by the same nr_pages that we used for reclaiming unmapped
2635 * Note that shrink_slab will free memory on all zones and may
2638 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
2639 (zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) + nr_pages
>
2644 * Update nr_reclaimed by the number of slab pages we
2645 * reclaimed from this zone.
2647 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2648 if (nr_slab_pages1
< nr_slab_pages0
)
2649 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
2652 p
->reclaim_state
= NULL
;
2653 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
2654 lockdep_clear_current_reclaim_state();
2655 return sc
.nr_reclaimed
>= nr_pages
;
2658 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2664 * Zone reclaim reclaims unmapped file backed pages and
2665 * slab pages if we are over the defined limits.
2667 * A small portion of unmapped file backed pages is needed for
2668 * file I/O otherwise pages read by file I/O will be immediately
2669 * thrown out if the zone is overallocated. So we do not reclaim
2670 * if less than a specified percentage of the zone is used by
2671 * unmapped file backed pages.
2673 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
2674 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
2675 return ZONE_RECLAIM_FULL
;
2677 if (zone
->all_unreclaimable
)
2678 return ZONE_RECLAIM_FULL
;
2681 * Do not scan if the allocation should not be delayed.
2683 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
2684 return ZONE_RECLAIM_NOSCAN
;
2687 * Only run zone reclaim on the local zone or on zones that do not
2688 * have associated processors. This will favor the local processor
2689 * over remote processors and spread off node memory allocations
2690 * as wide as possible.
2692 node_id
= zone_to_nid(zone
);
2693 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
2694 return ZONE_RECLAIM_NOSCAN
;
2696 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
2697 return ZONE_RECLAIM_NOSCAN
;
2699 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
2700 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
2703 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
2710 * page_evictable - test whether a page is evictable
2711 * @page: the page to test
2712 * @vma: the VMA in which the page is or will be mapped, may be NULL
2714 * Test whether page is evictable--i.e., should be placed on active/inactive
2715 * lists vs unevictable list. The vma argument is !NULL when called from the
2716 * fault path to determine how to instantate a new page.
2718 * Reasons page might not be evictable:
2719 * (1) page's mapping marked unevictable
2720 * (2) page is part of an mlocked VMA
2723 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
2726 if (mapping_unevictable(page_mapping(page
)))
2729 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
2736 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2737 * @page: page to check evictability and move to appropriate lru list
2738 * @zone: zone page is in
2740 * Checks a page for evictability and moves the page to the appropriate
2743 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2744 * have PageUnevictable set.
2746 static void check_move_unevictable_page(struct page
*page
, struct zone
*zone
)
2748 VM_BUG_ON(PageActive(page
));
2751 ClearPageUnevictable(page
);
2752 if (page_evictable(page
, NULL
)) {
2753 enum lru_list l
= page_lru_base_type(page
);
2755 __dec_zone_state(zone
, NR_UNEVICTABLE
);
2756 list_move(&page
->lru
, &zone
->lru
[l
].list
);
2757 mem_cgroup_move_lists(page
, LRU_UNEVICTABLE
, l
);
2758 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ l
);
2759 __count_vm_event(UNEVICTABLE_PGRESCUED
);
2762 * rotate unevictable list
2764 SetPageUnevictable(page
);
2765 list_move(&page
->lru
, &zone
->lru
[LRU_UNEVICTABLE
].list
);
2766 mem_cgroup_rotate_lru_list(page
, LRU_UNEVICTABLE
);
2767 if (page_evictable(page
, NULL
))
2773 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2774 * @mapping: struct address_space to scan for evictable pages
2776 * Scan all pages in mapping. Check unevictable pages for
2777 * evictability and move them to the appropriate zone lru list.
2779 void scan_mapping_unevictable_pages(struct address_space
*mapping
)
2782 pgoff_t end
= (i_size_read(mapping
->host
) + PAGE_CACHE_SIZE
- 1) >>
2785 struct pagevec pvec
;
2787 if (mapping
->nrpages
== 0)
2790 pagevec_init(&pvec
, 0);
2791 while (next
< end
&&
2792 pagevec_lookup(&pvec
, mapping
, next
, PAGEVEC_SIZE
)) {
2798 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
2799 struct page
*page
= pvec
.pages
[i
];
2800 pgoff_t page_index
= page
->index
;
2801 struct zone
*pagezone
= page_zone(page
);
2804 if (page_index
> next
)
2808 if (pagezone
!= zone
) {
2810 spin_unlock_irq(&zone
->lru_lock
);
2812 spin_lock_irq(&zone
->lru_lock
);
2815 if (PageLRU(page
) && PageUnevictable(page
))
2816 check_move_unevictable_page(page
, zone
);
2819 spin_unlock_irq(&zone
->lru_lock
);
2820 pagevec_release(&pvec
);
2822 count_vm_events(UNEVICTABLE_PGSCANNED
, pg_scanned
);
2828 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2829 * @zone - zone of which to scan the unevictable list
2831 * Scan @zone's unevictable LRU lists to check for pages that have become
2832 * evictable. Move those that have to @zone's inactive list where they
2833 * become candidates for reclaim, unless shrink_inactive_zone() decides
2834 * to reactivate them. Pages that are still unevictable are rotated
2835 * back onto @zone's unevictable list.
2837 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2838 static void scan_zone_unevictable_pages(struct zone
*zone
)
2840 struct list_head
*l_unevictable
= &zone
->lru
[LRU_UNEVICTABLE
].list
;
2842 unsigned long nr_to_scan
= zone_page_state(zone
, NR_UNEVICTABLE
);
2844 while (nr_to_scan
> 0) {
2845 unsigned long batch_size
= min(nr_to_scan
,
2846 SCAN_UNEVICTABLE_BATCH_SIZE
);
2848 spin_lock_irq(&zone
->lru_lock
);
2849 for (scan
= 0; scan
< batch_size
; scan
++) {
2850 struct page
*page
= lru_to_page(l_unevictable
);
2852 if (!trylock_page(page
))
2855 prefetchw_prev_lru_page(page
, l_unevictable
, flags
);
2857 if (likely(PageLRU(page
) && PageUnevictable(page
)))
2858 check_move_unevictable_page(page
, zone
);
2862 spin_unlock_irq(&zone
->lru_lock
);
2864 nr_to_scan
-= batch_size
;
2870 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2872 * A really big hammer: scan all zones' unevictable LRU lists to check for
2873 * pages that have become evictable. Move those back to the zones'
2874 * inactive list where they become candidates for reclaim.
2875 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2876 * and we add swap to the system. As such, it runs in the context of a task
2877 * that has possibly/probably made some previously unevictable pages
2880 static void scan_all_zones_unevictable_pages(void)
2884 for_each_zone(zone
) {
2885 scan_zone_unevictable_pages(zone
);
2890 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2891 * all nodes' unevictable lists for evictable pages
2893 unsigned long scan_unevictable_pages
;
2895 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
2896 void __user
*buffer
,
2897 size_t *length
, loff_t
*ppos
)
2899 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2901 if (write
&& *(unsigned long *)table
->data
)
2902 scan_all_zones_unevictable_pages();
2904 scan_unevictable_pages
= 0;
2909 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2910 * a specified node's per zone unevictable lists for evictable pages.
2913 static ssize_t
read_scan_unevictable_node(struct sys_device
*dev
,
2914 struct sysdev_attribute
*attr
,
2917 return sprintf(buf
, "0\n"); /* always zero; should fit... */
2920 static ssize_t
write_scan_unevictable_node(struct sys_device
*dev
,
2921 struct sysdev_attribute
*attr
,
2922 const char *buf
, size_t count
)
2924 struct zone
*node_zones
= NODE_DATA(dev
->id
)->node_zones
;
2927 unsigned long req
= strict_strtoul(buf
, 10, &res
);
2930 return 1; /* zero is no-op */
2932 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2933 if (!populated_zone(zone
))
2935 scan_zone_unevictable_pages(zone
);
2941 static SYSDEV_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
2942 read_scan_unevictable_node
,
2943 write_scan_unevictable_node
);
2945 int scan_unevictable_register_node(struct node
*node
)
2947 return sysdev_create_file(&node
->sysdev
, &attr_scan_unevictable_pages
);
2950 void scan_unevictable_unregister_node(struct node
*node
)
2952 sysdev_remove_file(&node
->sysdev
, &attr_scan_unevictable_pages
);