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/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.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>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned
;
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed
;
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 unsigned long hibernation_mode
;
68 /* This context's GFP mask */
73 /* Can mapped pages be reclaimed? */
76 /* Can pages be swapped as part of reclaim? */
82 * The memory cgroup that hit its limit and as a result is the
83 * primary target of this reclaim invocation.
85 struct mem_cgroup
*target_mem_cgroup
;
88 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 struct mem_cgroup_zone
{
95 struct mem_cgroup
*mem_cgroup
;
99 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
101 #ifdef ARCH_HAS_PREFETCH
102 #define prefetch_prev_lru_page(_page, _base, _field) \
104 if ((_page)->lru.prev != _base) { \
107 prev = lru_to_page(&(_page->lru)); \
108 prefetch(&prev->_field); \
112 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
115 #ifdef ARCH_HAS_PREFETCHW
116 #define prefetchw_prev_lru_page(_page, _base, _field) \
118 if ((_page)->lru.prev != _base) { \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetchw(&prev->_field); \
126 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
130 * From 0 .. 100. Higher means more swappy.
132 int vm_swappiness
= 60;
133 long vm_total_pages
; /* The total number of pages which the VM controls */
135 static LIST_HEAD(shrinker_list
);
136 static DECLARE_RWSEM(shrinker_rwsem
);
138 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
139 static bool global_reclaim(struct scan_control
*sc
)
141 return !sc
->target_mem_cgroup
;
144 static bool scanning_global_lru(struct mem_cgroup_zone
*mz
)
146 return !mz
->mem_cgroup
;
149 static bool global_reclaim(struct scan_control
*sc
)
154 static bool scanning_global_lru(struct mem_cgroup_zone
*mz
)
160 static struct zone_reclaim_stat
*get_reclaim_stat(struct mem_cgroup_zone
*mz
)
162 if (!scanning_global_lru(mz
))
163 return mem_cgroup_get_reclaim_stat(mz
->mem_cgroup
, mz
->zone
);
165 return &mz
->zone
->reclaim_stat
;
168 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone
*mz
,
171 if (!scanning_global_lru(mz
))
172 return mem_cgroup_zone_nr_lru_pages(mz
->mem_cgroup
,
173 zone_to_nid(mz
->zone
),
177 return zone_page_state(mz
->zone
, NR_LRU_BASE
+ lru
);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker
*shrinker
)
186 atomic_long_set(&shrinker
->nr_in_batch
, 0);
187 down_write(&shrinker_rwsem
);
188 list_add_tail(&shrinker
->list
, &shrinker_list
);
189 up_write(&shrinker_rwsem
);
191 EXPORT_SYMBOL(register_shrinker
);
196 void unregister_shrinker(struct shrinker
*shrinker
)
198 down_write(&shrinker_rwsem
);
199 list_del(&shrinker
->list
);
200 up_write(&shrinker_rwsem
);
202 EXPORT_SYMBOL(unregister_shrinker
);
204 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
205 struct shrink_control
*sc
,
206 unsigned long nr_to_scan
)
208 sc
->nr_to_scan
= nr_to_scan
;
209 return (*shrinker
->shrink
)(shrinker
, sc
);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control
*shrink
,
233 unsigned long nr_pages_scanned
,
234 unsigned long lru_pages
)
236 struct shrinker
*shrinker
;
237 unsigned long ret
= 0;
239 if (nr_pages_scanned
== 0)
240 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
242 if (!down_read_trylock(&shrinker_rwsem
)) {
243 /* Assume we'll be able to shrink next time */
248 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
249 unsigned long long delta
;
255 long batch_size
= shrinker
->batch
? shrinker
->batch
258 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
263 * copy the current shrinker scan count into a local variable
264 * and zero it so that other concurrent shrinker invocations
265 * don't also do this scanning work.
267 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
270 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
272 do_div(delta
, lru_pages
+ 1);
274 if (total_scan
< 0) {
275 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
277 shrinker
->shrink
, total_scan
);
278 total_scan
= max_pass
;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta
< max_pass
/ 4)
294 total_scan
= min(total_scan
, max_pass
/ 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan
> max_pass
* 2)
302 total_scan
= max_pass
* 2;
304 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
305 nr_pages_scanned
, lru_pages
,
306 max_pass
, delta
, total_scan
);
308 while (total_scan
>= batch_size
) {
311 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
312 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
314 if (shrink_ret
== -1)
316 if (shrink_ret
< nr_before
)
317 ret
+= nr_before
- shrink_ret
;
318 count_vm_events(SLABS_SCANNED
, batch_size
);
319 total_scan
-= batch_size
;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
330 new_nr
= atomic_long_add_return(total_scan
,
331 &shrinker
->nr_in_batch
);
333 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
335 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
337 up_read(&shrinker_rwsem
);
343 static inline int is_page_cache_freeable(struct page
*page
)
346 * A freeable page cache page is referenced only by the caller
347 * that isolated the page, the page cache radix tree and
348 * optional buffer heads at page->private.
350 return page_count(page
) - page_has_private(page
) == 2;
353 static int may_write_to_queue(struct backing_dev_info
*bdi
,
354 struct scan_control
*sc
)
356 if (current
->flags
& PF_SWAPWRITE
)
358 if (!bdi_write_congested(bdi
))
360 if (bdi
== current
->backing_dev_info
)
366 * We detected a synchronous write error writing a page out. Probably
367 * -ENOSPC. We need to propagate that into the address_space for a subsequent
368 * fsync(), msync() or close().
370 * The tricky part is that after writepage we cannot touch the mapping: nothing
371 * prevents it from being freed up. But we have a ref on the page and once
372 * that page is locked, the mapping is pinned.
374 * We're allowed to run sleeping lock_page() here because we know the caller has
377 static void handle_write_error(struct address_space
*mapping
,
378 struct page
*page
, int error
)
381 if (page_mapping(page
) == mapping
)
382 mapping_set_error(mapping
, error
);
386 /* possible outcome of pageout() */
388 /* failed to write page out, page is locked */
390 /* move page to the active list, page is locked */
392 /* page has been sent to the disk successfully, page is unlocked */
394 /* page is clean and locked */
399 * pageout is called by shrink_page_list() for each dirty page.
400 * Calls ->writepage().
402 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
403 struct scan_control
*sc
)
406 * If the page is dirty, only perform writeback if that write
407 * will be non-blocking. To prevent this allocation from being
408 * stalled by pagecache activity. But note that there may be
409 * stalls if we need to run get_block(). We could test
410 * PagePrivate for that.
412 * If this process is currently in __generic_file_aio_write() against
413 * this page's queue, we can perform writeback even if that
416 * If the page is swapcache, write it back even if that would
417 * block, for some throttling. This happens by accident, because
418 * swap_backing_dev_info is bust: it doesn't reflect the
419 * congestion state of the swapdevs. Easy to fix, if needed.
421 if (!is_page_cache_freeable(page
))
425 * Some data journaling orphaned pages can have
426 * page->mapping == NULL while being dirty with clean buffers.
428 if (page_has_private(page
)) {
429 if (try_to_free_buffers(page
)) {
430 ClearPageDirty(page
);
431 printk("%s: orphaned page\n", __func__
);
437 if (mapping
->a_ops
->writepage
== NULL
)
438 return PAGE_ACTIVATE
;
439 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
442 if (clear_page_dirty_for_io(page
)) {
444 struct writeback_control wbc
= {
445 .sync_mode
= WB_SYNC_NONE
,
446 .nr_to_write
= SWAP_CLUSTER_MAX
,
448 .range_end
= LLONG_MAX
,
452 SetPageReclaim(page
);
453 res
= mapping
->a_ops
->writepage(page
, &wbc
);
455 handle_write_error(mapping
, page
, res
);
456 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
457 ClearPageReclaim(page
);
458 return PAGE_ACTIVATE
;
461 if (!PageWriteback(page
)) {
462 /* synchronous write or broken a_ops? */
463 ClearPageReclaim(page
);
465 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
466 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
474 * Same as remove_mapping, but if the page is removed from the mapping, it
475 * gets returned with a refcount of 0.
477 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
479 BUG_ON(!PageLocked(page
));
480 BUG_ON(mapping
!= page_mapping(page
));
482 spin_lock_irq(&mapping
->tree_lock
);
484 * The non racy check for a busy page.
486 * Must be careful with the order of the tests. When someone has
487 * a ref to the page, it may be possible that they dirty it then
488 * drop the reference. So if PageDirty is tested before page_count
489 * here, then the following race may occur:
491 * get_user_pages(&page);
492 * [user mapping goes away]
494 * !PageDirty(page) [good]
495 * SetPageDirty(page);
497 * !page_count(page) [good, discard it]
499 * [oops, our write_to data is lost]
501 * Reversing the order of the tests ensures such a situation cannot
502 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
503 * load is not satisfied before that of page->_count.
505 * Note that if SetPageDirty is always performed via set_page_dirty,
506 * and thus under tree_lock, then this ordering is not required.
508 if (!page_freeze_refs(page
, 2))
510 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
511 if (unlikely(PageDirty(page
))) {
512 page_unfreeze_refs(page
, 2);
516 if (PageSwapCache(page
)) {
517 swp_entry_t swap
= { .val
= page_private(page
) };
518 __delete_from_swap_cache(page
);
519 spin_unlock_irq(&mapping
->tree_lock
);
520 swapcache_free(swap
, page
);
522 void (*freepage
)(struct page
*);
524 freepage
= mapping
->a_ops
->freepage
;
526 __delete_from_page_cache(page
);
527 spin_unlock_irq(&mapping
->tree_lock
);
528 mem_cgroup_uncharge_cache_page(page
);
530 if (freepage
!= NULL
)
537 spin_unlock_irq(&mapping
->tree_lock
);
542 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
543 * someone else has a ref on the page, abort and return 0. If it was
544 * successfully detached, return 1. Assumes the caller has a single ref on
547 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
549 if (__remove_mapping(mapping
, page
)) {
551 * Unfreezing the refcount with 1 rather than 2 effectively
552 * drops the pagecache ref for us without requiring another
555 page_unfreeze_refs(page
, 1);
562 * putback_lru_page - put previously isolated page onto appropriate LRU list
563 * @page: page to be put back to appropriate lru list
565 * Add previously isolated @page to appropriate LRU list.
566 * Page may still be unevictable for other reasons.
568 * lru_lock must not be held, interrupts must be enabled.
570 void putback_lru_page(struct page
*page
)
573 int active
= !!TestClearPageActive(page
);
574 int was_unevictable
= PageUnevictable(page
);
576 VM_BUG_ON(PageLRU(page
));
579 ClearPageUnevictable(page
);
581 if (page_evictable(page
, NULL
)) {
583 * For evictable pages, we can use the cache.
584 * In event of a race, worst case is we end up with an
585 * unevictable page on [in]active list.
586 * We know how to handle that.
588 lru
= active
+ page_lru_base_type(page
);
589 lru_cache_add_lru(page
, lru
);
592 * Put unevictable pages directly on zone's unevictable
595 lru
= LRU_UNEVICTABLE
;
596 add_page_to_unevictable_list(page
);
598 * When racing with an mlock or AS_UNEVICTABLE clearing
599 * (page is unlocked) make sure that if the other thread
600 * does not observe our setting of PG_lru and fails
601 * isolation/check_move_unevictable_pages,
602 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
603 * the page back to the evictable list.
605 * The other side is TestClearPageMlocked() or shmem_lock().
611 * page's status can change while we move it among lru. If an evictable
612 * page is on unevictable list, it never be freed. To avoid that,
613 * check after we added it to the list, again.
615 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
616 if (!isolate_lru_page(page
)) {
620 /* This means someone else dropped this page from LRU
621 * So, it will be freed or putback to LRU again. There is
622 * nothing to do here.
626 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
627 count_vm_event(UNEVICTABLE_PGRESCUED
);
628 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
629 count_vm_event(UNEVICTABLE_PGCULLED
);
631 put_page(page
); /* drop ref from isolate */
634 enum page_references
{
636 PAGEREF_RECLAIM_CLEAN
,
641 static enum page_references
page_check_references(struct page
*page
,
642 struct mem_cgroup_zone
*mz
,
643 struct scan_control
*sc
)
645 int referenced_ptes
, referenced_page
;
646 unsigned long vm_flags
;
648 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
650 referenced_page
= TestClearPageReferenced(page
);
653 * Mlock lost the isolation race with us. Let try_to_unmap()
654 * move the page to the unevictable list.
656 if (vm_flags
& VM_LOCKED
)
657 return PAGEREF_RECLAIM
;
659 if (referenced_ptes
) {
660 if (PageSwapBacked(page
))
661 return PAGEREF_ACTIVATE
;
663 * All mapped pages start out with page table
664 * references from the instantiating fault, so we need
665 * to look twice if a mapped file page is used more
668 * Mark it and spare it for another trip around the
669 * inactive list. Another page table reference will
670 * lead to its activation.
672 * Note: the mark is set for activated pages as well
673 * so that recently deactivated but used pages are
676 SetPageReferenced(page
);
678 if (referenced_page
|| referenced_ptes
> 1)
679 return PAGEREF_ACTIVATE
;
682 * Activate file-backed executable pages after first usage.
684 if (vm_flags
& VM_EXEC
)
685 return PAGEREF_ACTIVATE
;
690 /* Reclaim if clean, defer dirty pages to writeback */
691 if (referenced_page
&& !PageSwapBacked(page
))
692 return PAGEREF_RECLAIM_CLEAN
;
694 return PAGEREF_RECLAIM
;
698 * shrink_page_list() returns the number of reclaimed pages
700 static unsigned long shrink_page_list(struct list_head
*page_list
,
701 struct mem_cgroup_zone
*mz
,
702 struct scan_control
*sc
,
704 unsigned long *ret_nr_dirty
,
705 unsigned long *ret_nr_writeback
)
707 LIST_HEAD(ret_pages
);
708 LIST_HEAD(free_pages
);
710 unsigned long nr_dirty
= 0;
711 unsigned long nr_congested
= 0;
712 unsigned long nr_reclaimed
= 0;
713 unsigned long nr_writeback
= 0;
717 while (!list_empty(page_list
)) {
718 enum page_references references
;
719 struct address_space
*mapping
;
725 page
= lru_to_page(page_list
);
726 list_del(&page
->lru
);
728 if (!trylock_page(page
))
731 VM_BUG_ON(PageActive(page
));
732 VM_BUG_ON(page_zone(page
) != mz
->zone
);
736 if (unlikely(!page_evictable(page
, NULL
)))
739 if (!sc
->may_unmap
&& page_mapped(page
))
742 /* Double the slab pressure for mapped and swapcache pages */
743 if (page_mapped(page
) || PageSwapCache(page
))
746 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
747 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
749 if (PageWriteback(page
)) {
755 references
= page_check_references(page
, mz
, sc
);
756 switch (references
) {
757 case PAGEREF_ACTIVATE
:
758 goto activate_locked
;
761 case PAGEREF_RECLAIM
:
762 case PAGEREF_RECLAIM_CLEAN
:
763 ; /* try to reclaim the page below */
767 * Anonymous process memory has backing store?
768 * Try to allocate it some swap space here.
770 if (PageAnon(page
) && !PageSwapCache(page
)) {
771 if (!(sc
->gfp_mask
& __GFP_IO
))
773 if (!add_to_swap(page
))
774 goto activate_locked
;
778 mapping
= page_mapping(page
);
781 * The page is mapped into the page tables of one or more
782 * processes. Try to unmap it here.
784 if (page_mapped(page
) && mapping
) {
785 switch (try_to_unmap(page
, TTU_UNMAP
)) {
787 goto activate_locked
;
793 ; /* try to free the page below */
797 if (PageDirty(page
)) {
801 * Only kswapd can writeback filesystem pages to
802 * avoid risk of stack overflow but do not writeback
803 * unless under significant pressure.
805 if (page_is_file_cache(page
) &&
806 (!current_is_kswapd() || priority
>= DEF_PRIORITY
- 2)) {
808 * Immediately reclaim when written back.
809 * Similar in principal to deactivate_page()
810 * except we already have the page isolated
811 * and know it's dirty
813 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
814 SetPageReclaim(page
);
819 if (references
== PAGEREF_RECLAIM_CLEAN
)
823 if (!sc
->may_writepage
)
826 /* Page is dirty, try to write it out here */
827 switch (pageout(page
, mapping
, sc
)) {
832 goto activate_locked
;
834 if (PageWriteback(page
))
840 * A synchronous write - probably a ramdisk. Go
841 * ahead and try to reclaim the page.
843 if (!trylock_page(page
))
845 if (PageDirty(page
) || PageWriteback(page
))
847 mapping
= page_mapping(page
);
849 ; /* try to free the page below */
854 * If the page has buffers, try to free the buffer mappings
855 * associated with this page. If we succeed we try to free
858 * We do this even if the page is PageDirty().
859 * try_to_release_page() does not perform I/O, but it is
860 * possible for a page to have PageDirty set, but it is actually
861 * clean (all its buffers are clean). This happens if the
862 * buffers were written out directly, with submit_bh(). ext3
863 * will do this, as well as the blockdev mapping.
864 * try_to_release_page() will discover that cleanness and will
865 * drop the buffers and mark the page clean - it can be freed.
867 * Rarely, pages can have buffers and no ->mapping. These are
868 * the pages which were not successfully invalidated in
869 * truncate_complete_page(). We try to drop those buffers here
870 * and if that worked, and the page is no longer mapped into
871 * process address space (page_count == 1) it can be freed.
872 * Otherwise, leave the page on the LRU so it is swappable.
874 if (page_has_private(page
)) {
875 if (!try_to_release_page(page
, sc
->gfp_mask
))
876 goto activate_locked
;
877 if (!mapping
&& page_count(page
) == 1) {
879 if (put_page_testzero(page
))
883 * rare race with speculative reference.
884 * the speculative reference will free
885 * this page shortly, so we may
886 * increment nr_reclaimed here (and
887 * leave it off the LRU).
895 if (!mapping
|| !__remove_mapping(mapping
, page
))
899 * At this point, we have no other references and there is
900 * no way to pick any more up (removed from LRU, removed
901 * from pagecache). Can use non-atomic bitops now (and
902 * we obviously don't have to worry about waking up a process
903 * waiting on the page lock, because there are no references.
905 __clear_page_locked(page
);
910 * Is there need to periodically free_page_list? It would
911 * appear not as the counts should be low
913 list_add(&page
->lru
, &free_pages
);
917 if (PageSwapCache(page
))
918 try_to_free_swap(page
);
920 putback_lru_page(page
);
924 /* Not a candidate for swapping, so reclaim swap space. */
925 if (PageSwapCache(page
) && vm_swap_full())
926 try_to_free_swap(page
);
927 VM_BUG_ON(PageActive(page
));
933 list_add(&page
->lru
, &ret_pages
);
934 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
938 * Tag a zone as congested if all the dirty pages encountered were
939 * backed by a congested BDI. In this case, reclaimers should just
940 * back off and wait for congestion to clear because further reclaim
941 * will encounter the same problem
943 if (nr_dirty
&& nr_dirty
== nr_congested
&& global_reclaim(sc
))
944 zone_set_flag(mz
->zone
, ZONE_CONGESTED
);
946 free_hot_cold_page_list(&free_pages
, 1);
948 list_splice(&ret_pages
, page_list
);
949 count_vm_events(PGACTIVATE
, pgactivate
);
950 *ret_nr_dirty
+= nr_dirty
;
951 *ret_nr_writeback
+= nr_writeback
;
956 * Attempt to remove the specified page from its LRU. Only take this page
957 * if it is of the appropriate PageActive status. Pages which are being
958 * freed elsewhere are also ignored.
960 * page: page to consider
961 * mode: one of the LRU isolation modes defined above
963 * returns 0 on success, -ve errno on failure.
965 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
, int file
)
970 /* Only take pages on the LRU. */
974 all_lru_mode
= (mode
& (ISOLATE_ACTIVE
|ISOLATE_INACTIVE
)) ==
975 (ISOLATE_ACTIVE
|ISOLATE_INACTIVE
);
978 * When checking the active state, we need to be sure we are
979 * dealing with comparible boolean values. Take the logical not
982 if (!all_lru_mode
&& !PageActive(page
) != !(mode
& ISOLATE_ACTIVE
))
985 if (!all_lru_mode
&& !!page_is_file_cache(page
) != file
)
988 /* Do not give back unevictable pages for compaction */
989 if (PageUnevictable(page
))
995 * To minimise LRU disruption, the caller can indicate that it only
996 * wants to isolate pages it will be able to operate on without
997 * blocking - clean pages for the most part.
999 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1000 * is used by reclaim when it is cannot write to backing storage
1002 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1003 * that it is possible to migrate without blocking
1005 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1006 /* All the caller can do on PageWriteback is block */
1007 if (PageWriteback(page
))
1010 if (PageDirty(page
)) {
1011 struct address_space
*mapping
;
1013 /* ISOLATE_CLEAN means only clean pages */
1014 if (mode
& ISOLATE_CLEAN
)
1018 * Only pages without mappings or that have a
1019 * ->migratepage callback are possible to migrate
1022 mapping
= page_mapping(page
);
1023 if (mapping
&& !mapping
->a_ops
->migratepage
)
1028 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1031 if (likely(get_page_unless_zero(page
))) {
1033 * Be careful not to clear PageLRU until after we're
1034 * sure the page is not being freed elsewhere -- the
1035 * page release code relies on it.
1045 * zone->lru_lock is heavily contended. Some of the functions that
1046 * shrink the lists perform better by taking out a batch of pages
1047 * and working on them outside the LRU lock.
1049 * For pagecache intensive workloads, this function is the hottest
1050 * spot in the kernel (apart from copy_*_user functions).
1052 * Appropriate locks must be held before calling this function.
1054 * @nr_to_scan: The number of pages to look through on the list.
1055 * @mz: The mem_cgroup_zone to pull pages from.
1056 * @dst: The temp list to put pages on to.
1057 * @nr_scanned: The number of pages that were scanned.
1058 * @sc: The scan_control struct for this reclaim session
1059 * @mode: One of the LRU isolation modes
1060 * @active: True [1] if isolating active pages
1061 * @file: True [1] if isolating file [!anon] pages
1063 * returns how many pages were moved onto *@dst.
1065 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1066 struct mem_cgroup_zone
*mz
, struct list_head
*dst
,
1067 unsigned long *nr_scanned
, struct scan_control
*sc
,
1068 isolate_mode_t mode
, int active
, int file
)
1070 struct lruvec
*lruvec
;
1071 struct list_head
*src
;
1072 unsigned long nr_taken
= 0;
1076 lruvec
= mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
);
1081 src
= &lruvec
->lists
[lru
];
1083 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1086 page
= lru_to_page(src
);
1087 prefetchw_prev_lru_page(page
, src
, flags
);
1089 VM_BUG_ON(!PageLRU(page
));
1091 switch (__isolate_lru_page(page
, mode
, file
)) {
1093 mem_cgroup_lru_del(page
);
1094 list_move(&page
->lru
, dst
);
1095 nr_taken
+= hpage_nr_pages(page
);
1099 /* else it is being freed elsewhere */
1100 list_move(&page
->lru
, src
);
1110 trace_mm_vmscan_lru_isolate(sc
->order
,
1118 * isolate_lru_page - tries to isolate a page from its LRU list
1119 * @page: page to isolate from its LRU list
1121 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1122 * vmstat statistic corresponding to whatever LRU list the page was on.
1124 * Returns 0 if the page was removed from an LRU list.
1125 * Returns -EBUSY if the page was not on an LRU list.
1127 * The returned page will have PageLRU() cleared. If it was found on
1128 * the active list, it will have PageActive set. If it was found on
1129 * the unevictable list, it will have the PageUnevictable bit set. That flag
1130 * may need to be cleared by the caller before letting the page go.
1132 * The vmstat statistic corresponding to the list on which the page was
1133 * found will be decremented.
1136 * (1) Must be called with an elevated refcount on the page. This is a
1137 * fundamentnal difference from isolate_lru_pages (which is called
1138 * without a stable reference).
1139 * (2) the lru_lock must not be held.
1140 * (3) interrupts must be enabled.
1142 int isolate_lru_page(struct page
*page
)
1146 VM_BUG_ON(!page_count(page
));
1148 if (PageLRU(page
)) {
1149 struct zone
*zone
= page_zone(page
);
1151 spin_lock_irq(&zone
->lru_lock
);
1152 if (PageLRU(page
)) {
1153 int lru
= page_lru(page
);
1158 del_page_from_lru_list(zone
, page
, lru
);
1160 spin_unlock_irq(&zone
->lru_lock
);
1166 * Are there way too many processes in the direct reclaim path already?
1168 static int too_many_isolated(struct zone
*zone
, int file
,
1169 struct scan_control
*sc
)
1171 unsigned long inactive
, isolated
;
1173 if (current_is_kswapd())
1176 if (!global_reclaim(sc
))
1180 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1181 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1183 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1184 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1187 return isolated
> inactive
;
1190 static noinline_for_stack
void
1191 putback_inactive_pages(struct mem_cgroup_zone
*mz
,
1192 struct list_head
*page_list
)
1194 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1195 struct zone
*zone
= mz
->zone
;
1196 LIST_HEAD(pages_to_free
);
1199 * Put back any unfreeable pages.
1201 while (!list_empty(page_list
)) {
1202 struct page
*page
= lru_to_page(page_list
);
1205 VM_BUG_ON(PageLRU(page
));
1206 list_del(&page
->lru
);
1207 if (unlikely(!page_evictable(page
, NULL
))) {
1208 spin_unlock_irq(&zone
->lru_lock
);
1209 putback_lru_page(page
);
1210 spin_lock_irq(&zone
->lru_lock
);
1214 lru
= page_lru(page
);
1215 add_page_to_lru_list(zone
, page
, lru
);
1216 if (is_active_lru(lru
)) {
1217 int file
= is_file_lru(lru
);
1218 int numpages
= hpage_nr_pages(page
);
1219 reclaim_stat
->recent_rotated
[file
] += numpages
;
1221 if (put_page_testzero(page
)) {
1222 __ClearPageLRU(page
);
1223 __ClearPageActive(page
);
1224 del_page_from_lru_list(zone
, page
, lru
);
1226 if (unlikely(PageCompound(page
))) {
1227 spin_unlock_irq(&zone
->lru_lock
);
1228 (*get_compound_page_dtor(page
))(page
);
1229 spin_lock_irq(&zone
->lru_lock
);
1231 list_add(&page
->lru
, &pages_to_free
);
1236 * To save our caller's stack, now use input list for pages to free.
1238 list_splice(&pages_to_free
, page_list
);
1241 static noinline_for_stack
void
1242 update_isolated_counts(struct mem_cgroup_zone
*mz
,
1243 struct list_head
*page_list
,
1244 unsigned long *nr_anon
,
1245 unsigned long *nr_file
)
1247 struct zone
*zone
= mz
->zone
;
1248 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1249 unsigned long nr_active
= 0;
1254 * Count pages and clear active flags
1256 list_for_each_entry(page
, page_list
, lru
) {
1257 int numpages
= hpage_nr_pages(page
);
1258 lru
= page_lru_base_type(page
);
1259 if (PageActive(page
)) {
1261 ClearPageActive(page
);
1262 nr_active
+= numpages
;
1264 count
[lru
] += numpages
;
1268 __count_vm_events(PGDEACTIVATE
, nr_active
);
1270 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1271 -count
[LRU_ACTIVE_FILE
]);
1272 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1273 -count
[LRU_INACTIVE_FILE
]);
1274 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1275 -count
[LRU_ACTIVE_ANON
]);
1276 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1277 -count
[LRU_INACTIVE_ANON
]);
1279 *nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1280 *nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1282 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, *nr_anon
);
1283 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, *nr_file
);
1288 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1289 * of reclaimed pages
1291 static noinline_for_stack
unsigned long
1292 shrink_inactive_list(unsigned long nr_to_scan
, struct mem_cgroup_zone
*mz
,
1293 struct scan_control
*sc
, int priority
, int file
)
1295 LIST_HEAD(page_list
);
1296 unsigned long nr_scanned
;
1297 unsigned long nr_reclaimed
= 0;
1298 unsigned long nr_taken
;
1299 unsigned long nr_anon
;
1300 unsigned long nr_file
;
1301 unsigned long nr_dirty
= 0;
1302 unsigned long nr_writeback
= 0;
1303 isolate_mode_t isolate_mode
= ISOLATE_INACTIVE
;
1304 struct zone
*zone
= mz
->zone
;
1305 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1307 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1308 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1310 /* We are about to die and free our memory. Return now. */
1311 if (fatal_signal_pending(current
))
1312 return SWAP_CLUSTER_MAX
;
1318 isolate_mode
|= ISOLATE_UNMAPPED
;
1319 if (!sc
->may_writepage
)
1320 isolate_mode
|= ISOLATE_CLEAN
;
1322 spin_lock_irq(&zone
->lru_lock
);
1324 nr_taken
= isolate_lru_pages(nr_to_scan
, mz
, &page_list
, &nr_scanned
,
1325 sc
, isolate_mode
, 0, file
);
1326 if (global_reclaim(sc
)) {
1327 zone
->pages_scanned
+= nr_scanned
;
1328 if (current_is_kswapd())
1329 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1332 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1335 spin_unlock_irq(&zone
->lru_lock
);
1340 update_isolated_counts(mz
, &page_list
, &nr_anon
, &nr_file
);
1342 nr_reclaimed
= shrink_page_list(&page_list
, mz
, sc
, priority
,
1343 &nr_dirty
, &nr_writeback
);
1345 spin_lock_irq(&zone
->lru_lock
);
1347 reclaim_stat
->recent_scanned
[0] += nr_anon
;
1348 reclaim_stat
->recent_scanned
[1] += nr_file
;
1350 if (global_reclaim(sc
)) {
1351 if (current_is_kswapd())
1352 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1355 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1359 putback_inactive_pages(mz
, &page_list
);
1361 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1362 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1364 spin_unlock_irq(&zone
->lru_lock
);
1366 free_hot_cold_page_list(&page_list
, 1);
1369 * If reclaim is isolating dirty pages under writeback, it implies
1370 * that the long-lived page allocation rate is exceeding the page
1371 * laundering rate. Either the global limits are not being effective
1372 * at throttling processes due to the page distribution throughout
1373 * zones or there is heavy usage of a slow backing device. The
1374 * only option is to throttle from reclaim context which is not ideal
1375 * as there is no guarantee the dirtying process is throttled in the
1376 * same way balance_dirty_pages() manages.
1378 * This scales the number of dirty pages that must be under writeback
1379 * before throttling depending on priority. It is a simple backoff
1380 * function that has the most effect in the range DEF_PRIORITY to
1381 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1382 * in trouble and reclaim is considered to be in trouble.
1384 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1385 * DEF_PRIORITY-1 50% must be PageWriteback
1386 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1388 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1389 * isolated page is PageWriteback
1391 if (nr_writeback
&& nr_writeback
>= (nr_taken
>> (DEF_PRIORITY
-priority
)))
1392 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1394 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1396 nr_scanned
, nr_reclaimed
,
1398 trace_shrink_flags(file
));
1399 return nr_reclaimed
;
1403 * This moves pages from the active list to the inactive list.
1405 * We move them the other way if the page is referenced by one or more
1406 * processes, from rmap.
1408 * If the pages are mostly unmapped, the processing is fast and it is
1409 * appropriate to hold zone->lru_lock across the whole operation. But if
1410 * the pages are mapped, the processing is slow (page_referenced()) so we
1411 * should drop zone->lru_lock around each page. It's impossible to balance
1412 * this, so instead we remove the pages from the LRU while processing them.
1413 * It is safe to rely on PG_active against the non-LRU pages in here because
1414 * nobody will play with that bit on a non-LRU page.
1416 * The downside is that we have to touch page->_count against each page.
1417 * But we had to alter page->flags anyway.
1420 static void move_active_pages_to_lru(struct zone
*zone
,
1421 struct list_head
*list
,
1422 struct list_head
*pages_to_free
,
1425 unsigned long pgmoved
= 0;
1428 while (!list_empty(list
)) {
1429 struct lruvec
*lruvec
;
1431 page
= lru_to_page(list
);
1433 VM_BUG_ON(PageLRU(page
));
1436 lruvec
= mem_cgroup_lru_add_list(zone
, page
, lru
);
1437 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1438 pgmoved
+= hpage_nr_pages(page
);
1440 if (put_page_testzero(page
)) {
1441 __ClearPageLRU(page
);
1442 __ClearPageActive(page
);
1443 del_page_from_lru_list(zone
, page
, lru
);
1445 if (unlikely(PageCompound(page
))) {
1446 spin_unlock_irq(&zone
->lru_lock
);
1447 (*get_compound_page_dtor(page
))(page
);
1448 spin_lock_irq(&zone
->lru_lock
);
1450 list_add(&page
->lru
, pages_to_free
);
1453 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1454 if (!is_active_lru(lru
))
1455 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1458 static void shrink_active_list(unsigned long nr_to_scan
,
1459 struct mem_cgroup_zone
*mz
,
1460 struct scan_control
*sc
,
1461 int priority
, int file
)
1463 unsigned long nr_taken
;
1464 unsigned long nr_scanned
;
1465 unsigned long vm_flags
;
1466 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1467 LIST_HEAD(l_active
);
1468 LIST_HEAD(l_inactive
);
1470 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1471 unsigned long nr_rotated
= 0;
1472 isolate_mode_t isolate_mode
= ISOLATE_ACTIVE
;
1473 struct zone
*zone
= mz
->zone
;
1478 isolate_mode
|= ISOLATE_UNMAPPED
;
1479 if (!sc
->may_writepage
)
1480 isolate_mode
|= ISOLATE_CLEAN
;
1482 spin_lock_irq(&zone
->lru_lock
);
1484 nr_taken
= isolate_lru_pages(nr_to_scan
, mz
, &l_hold
, &nr_scanned
, sc
,
1485 isolate_mode
, 1, file
);
1486 if (global_reclaim(sc
))
1487 zone
->pages_scanned
+= nr_scanned
;
1489 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1491 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1493 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1495 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1496 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1497 spin_unlock_irq(&zone
->lru_lock
);
1499 while (!list_empty(&l_hold
)) {
1501 page
= lru_to_page(&l_hold
);
1502 list_del(&page
->lru
);
1504 if (unlikely(!page_evictable(page
, NULL
))) {
1505 putback_lru_page(page
);
1509 if (unlikely(buffer_heads_over_limit
)) {
1510 if (page_has_private(page
) && trylock_page(page
)) {
1511 if (page_has_private(page
))
1512 try_to_release_page(page
, 0);
1517 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1519 nr_rotated
+= hpage_nr_pages(page
);
1521 * Identify referenced, file-backed active pages and
1522 * give them one more trip around the active list. So
1523 * that executable code get better chances to stay in
1524 * memory under moderate memory pressure. Anon pages
1525 * are not likely to be evicted by use-once streaming
1526 * IO, plus JVM can create lots of anon VM_EXEC pages,
1527 * so we ignore them here.
1529 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1530 list_add(&page
->lru
, &l_active
);
1535 ClearPageActive(page
); /* we are de-activating */
1536 list_add(&page
->lru
, &l_inactive
);
1540 * Move pages back to the lru list.
1542 spin_lock_irq(&zone
->lru_lock
);
1544 * Count referenced pages from currently used mappings as rotated,
1545 * even though only some of them are actually re-activated. This
1546 * helps balance scan pressure between file and anonymous pages in
1549 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1551 move_active_pages_to_lru(zone
, &l_active
, &l_hold
,
1552 LRU_ACTIVE
+ file
* LRU_FILE
);
1553 move_active_pages_to_lru(zone
, &l_inactive
, &l_hold
,
1554 LRU_BASE
+ file
* LRU_FILE
);
1555 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1556 spin_unlock_irq(&zone
->lru_lock
);
1558 free_hot_cold_page_list(&l_hold
, 1);
1562 static int inactive_anon_is_low_global(struct zone
*zone
)
1564 unsigned long active
, inactive
;
1566 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1567 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1569 if (inactive
* zone
->inactive_ratio
< active
)
1576 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1577 * @zone: zone to check
1578 * @sc: scan control of this context
1580 * Returns true if the zone does not have enough inactive anon pages,
1581 * meaning some active anon pages need to be deactivated.
1583 static int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1586 * If we don't have swap space, anonymous page deactivation
1589 if (!total_swap_pages
)
1592 if (!scanning_global_lru(mz
))
1593 return mem_cgroup_inactive_anon_is_low(mz
->mem_cgroup
,
1596 return inactive_anon_is_low_global(mz
->zone
);
1599 static inline int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1605 static int inactive_file_is_low_global(struct zone
*zone
)
1607 unsigned long active
, inactive
;
1609 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1610 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1612 return (active
> inactive
);
1616 * inactive_file_is_low - check if file pages need to be deactivated
1617 * @mz: memory cgroup and zone to check
1619 * When the system is doing streaming IO, memory pressure here
1620 * ensures that active file pages get deactivated, until more
1621 * than half of the file pages are on the inactive list.
1623 * Once we get to that situation, protect the system's working
1624 * set from being evicted by disabling active file page aging.
1626 * This uses a different ratio than the anonymous pages, because
1627 * the page cache uses a use-once replacement algorithm.
1629 static int inactive_file_is_low(struct mem_cgroup_zone
*mz
)
1631 if (!scanning_global_lru(mz
))
1632 return mem_cgroup_inactive_file_is_low(mz
->mem_cgroup
,
1635 return inactive_file_is_low_global(mz
->zone
);
1638 static int inactive_list_is_low(struct mem_cgroup_zone
*mz
, int file
)
1641 return inactive_file_is_low(mz
);
1643 return inactive_anon_is_low(mz
);
1646 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1647 struct mem_cgroup_zone
*mz
,
1648 struct scan_control
*sc
, int priority
)
1650 int file
= is_file_lru(lru
);
1652 if (is_active_lru(lru
)) {
1653 if (inactive_list_is_low(mz
, file
))
1654 shrink_active_list(nr_to_scan
, mz
, sc
, priority
, file
);
1658 return shrink_inactive_list(nr_to_scan
, mz
, sc
, priority
, file
);
1661 static int vmscan_swappiness(struct mem_cgroup_zone
*mz
,
1662 struct scan_control
*sc
)
1664 if (global_reclaim(sc
))
1665 return vm_swappiness
;
1666 return mem_cgroup_swappiness(mz
->mem_cgroup
);
1670 * Determine how aggressively the anon and file LRU lists should be
1671 * scanned. The relative value of each set of LRU lists is determined
1672 * by looking at the fraction of the pages scanned we did rotate back
1673 * onto the active list instead of evict.
1675 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1677 static void get_scan_count(struct mem_cgroup_zone
*mz
, struct scan_control
*sc
,
1678 unsigned long *nr
, int priority
)
1680 unsigned long anon
, file
, free
;
1681 unsigned long anon_prio
, file_prio
;
1682 unsigned long ap
, fp
;
1683 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1684 u64 fraction
[2], denominator
;
1687 bool force_scan
= false;
1690 * If the zone or memcg is small, nr[l] can be 0. This
1691 * results in no scanning on this priority and a potential
1692 * priority drop. Global direct reclaim can go to the next
1693 * zone and tends to have no problems. Global kswapd is for
1694 * zone balancing and it needs to scan a minimum amount. When
1695 * reclaiming for a memcg, a priority drop can cause high
1696 * latencies, so it's better to scan a minimum amount there as
1699 if (current_is_kswapd() && mz
->zone
->all_unreclaimable
)
1701 if (!global_reclaim(sc
))
1704 /* If we have no swap space, do not bother scanning anon pages. */
1705 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1713 anon
= zone_nr_lru_pages(mz
, LRU_ACTIVE_ANON
) +
1714 zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1715 file
= zone_nr_lru_pages(mz
, LRU_ACTIVE_FILE
) +
1716 zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1718 if (global_reclaim(sc
)) {
1719 free
= zone_page_state(mz
->zone
, NR_FREE_PAGES
);
1720 /* If we have very few page cache pages,
1721 force-scan anon pages. */
1722 if (unlikely(file
+ free
<= high_wmark_pages(mz
->zone
))) {
1731 * With swappiness at 100, anonymous and file have the same priority.
1732 * This scanning priority is essentially the inverse of IO cost.
1734 anon_prio
= vmscan_swappiness(mz
, sc
);
1735 file_prio
= 200 - vmscan_swappiness(mz
, sc
);
1738 * OK, so we have swap space and a fair amount of page cache
1739 * pages. We use the recently rotated / recently scanned
1740 * ratios to determine how valuable each cache is.
1742 * Because workloads change over time (and to avoid overflow)
1743 * we keep these statistics as a floating average, which ends
1744 * up weighing recent references more than old ones.
1746 * anon in [0], file in [1]
1748 spin_lock_irq(&mz
->zone
->lru_lock
);
1749 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1750 reclaim_stat
->recent_scanned
[0] /= 2;
1751 reclaim_stat
->recent_rotated
[0] /= 2;
1754 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1755 reclaim_stat
->recent_scanned
[1] /= 2;
1756 reclaim_stat
->recent_rotated
[1] /= 2;
1760 * The amount of pressure on anon vs file pages is inversely
1761 * proportional to the fraction of recently scanned pages on
1762 * each list that were recently referenced and in active use.
1764 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1765 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1767 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1768 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1769 spin_unlock_irq(&mz
->zone
->lru_lock
);
1773 denominator
= ap
+ fp
+ 1;
1775 for_each_evictable_lru(lru
) {
1776 int file
= is_file_lru(lru
);
1779 scan
= zone_nr_lru_pages(mz
, lru
);
1780 if (priority
|| noswap
) {
1782 if (!scan
&& force_scan
)
1783 scan
= SWAP_CLUSTER_MAX
;
1784 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1790 /* Use reclaim/compaction for costly allocs or under memory pressure */
1791 static bool in_reclaim_compaction(int priority
, struct scan_control
*sc
)
1793 if (COMPACTION_BUILD
&& sc
->order
&&
1794 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
1795 priority
< DEF_PRIORITY
- 2))
1802 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1803 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1804 * true if more pages should be reclaimed such that when the page allocator
1805 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1806 * It will give up earlier than that if there is difficulty reclaiming pages.
1808 static inline bool should_continue_reclaim(struct mem_cgroup_zone
*mz
,
1809 unsigned long nr_reclaimed
,
1810 unsigned long nr_scanned
,
1812 struct scan_control
*sc
)
1814 unsigned long pages_for_compaction
;
1815 unsigned long inactive_lru_pages
;
1817 /* If not in reclaim/compaction mode, stop */
1818 if (!in_reclaim_compaction(priority
, sc
))
1821 /* Consider stopping depending on scan and reclaim activity */
1822 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1824 * For __GFP_REPEAT allocations, stop reclaiming if the
1825 * full LRU list has been scanned and we are still failing
1826 * to reclaim pages. This full LRU scan is potentially
1827 * expensive but a __GFP_REPEAT caller really wants to succeed
1829 if (!nr_reclaimed
&& !nr_scanned
)
1833 * For non-__GFP_REPEAT allocations which can presumably
1834 * fail without consequence, stop if we failed to reclaim
1835 * any pages from the last SWAP_CLUSTER_MAX number of
1836 * pages that were scanned. This will return to the
1837 * caller faster at the risk reclaim/compaction and
1838 * the resulting allocation attempt fails
1845 * If we have not reclaimed enough pages for compaction and the
1846 * inactive lists are large enough, continue reclaiming
1848 pages_for_compaction
= (2UL << sc
->order
);
1849 inactive_lru_pages
= zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1850 if (nr_swap_pages
> 0)
1851 inactive_lru_pages
+= zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1852 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1853 inactive_lru_pages
> pages_for_compaction
)
1856 /* If compaction would go ahead or the allocation would succeed, stop */
1857 switch (compaction_suitable(mz
->zone
, sc
->order
)) {
1858 case COMPACT_PARTIAL
:
1859 case COMPACT_CONTINUE
:
1867 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1869 static void shrink_mem_cgroup_zone(int priority
, struct mem_cgroup_zone
*mz
,
1870 struct scan_control
*sc
)
1872 unsigned long nr
[NR_LRU_LISTS
];
1873 unsigned long nr_to_scan
;
1875 unsigned long nr_reclaimed
, nr_scanned
;
1876 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1877 struct blk_plug plug
;
1881 nr_scanned
= sc
->nr_scanned
;
1882 get_scan_count(mz
, sc
, nr
, priority
);
1884 blk_start_plug(&plug
);
1885 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1886 nr
[LRU_INACTIVE_FILE
]) {
1887 for_each_evictable_lru(lru
) {
1889 nr_to_scan
= min_t(unsigned long,
1890 nr
[lru
], SWAP_CLUSTER_MAX
);
1891 nr
[lru
] -= nr_to_scan
;
1893 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1898 * On large memory systems, scan >> priority can become
1899 * really large. This is fine for the starting priority;
1900 * we want to put equal scanning pressure on each zone.
1901 * However, if the VM has a harder time of freeing pages,
1902 * with multiple processes reclaiming pages, the total
1903 * freeing target can get unreasonably large.
1905 if (nr_reclaimed
>= nr_to_reclaim
&& priority
< DEF_PRIORITY
)
1908 blk_finish_plug(&plug
);
1909 sc
->nr_reclaimed
+= nr_reclaimed
;
1912 * Even if we did not try to evict anon pages at all, we want to
1913 * rebalance the anon lru active/inactive ratio.
1915 if (inactive_anon_is_low(mz
))
1916 shrink_active_list(SWAP_CLUSTER_MAX
, mz
, sc
, priority
, 0);
1918 /* reclaim/compaction might need reclaim to continue */
1919 if (should_continue_reclaim(mz
, nr_reclaimed
,
1920 sc
->nr_scanned
- nr_scanned
,
1924 throttle_vm_writeout(sc
->gfp_mask
);
1927 static void shrink_zone(int priority
, struct zone
*zone
,
1928 struct scan_control
*sc
)
1930 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
1931 struct mem_cgroup_reclaim_cookie reclaim
= {
1933 .priority
= priority
,
1935 struct mem_cgroup
*memcg
;
1937 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
1939 struct mem_cgroup_zone mz
= {
1940 .mem_cgroup
= memcg
,
1944 shrink_mem_cgroup_zone(priority
, &mz
, sc
);
1946 * Limit reclaim has historically picked one memcg and
1947 * scanned it with decreasing priority levels until
1948 * nr_to_reclaim had been reclaimed. This priority
1949 * cycle is thus over after a single memcg.
1951 * Direct reclaim and kswapd, on the other hand, have
1952 * to scan all memory cgroups to fulfill the overall
1953 * scan target for the zone.
1955 if (!global_reclaim(sc
)) {
1956 mem_cgroup_iter_break(root
, memcg
);
1959 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
1963 /* Returns true if compaction should go ahead for a high-order request */
1964 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
1966 unsigned long balance_gap
, watermark
;
1969 /* Do not consider compaction for orders reclaim is meant to satisfy */
1970 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1974 * Compaction takes time to run and there are potentially other
1975 * callers using the pages just freed. Continue reclaiming until
1976 * there is a buffer of free pages available to give compaction
1977 * a reasonable chance of completing and allocating the page
1979 balance_gap
= min(low_wmark_pages(zone
),
1980 (zone
->present_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
1981 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
1982 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
1983 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
1986 * If compaction is deferred, reclaim up to a point where
1987 * compaction will have a chance of success when re-enabled
1989 if (compaction_deferred(zone
, sc
->order
))
1990 return watermark_ok
;
1992 /* If compaction is not ready to start, keep reclaiming */
1993 if (!compaction_suitable(zone
, sc
->order
))
1996 return watermark_ok
;
2000 * This is the direct reclaim path, for page-allocating processes. We only
2001 * try to reclaim pages from zones which will satisfy the caller's allocation
2004 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2006 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2008 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2009 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2010 * zone defense algorithm.
2012 * If a zone is deemed to be full of pinned pages then just give it a light
2013 * scan then give up on it.
2015 * This function returns true if a zone is being reclaimed for a costly
2016 * high-order allocation and compaction is ready to begin. This indicates to
2017 * the caller that it should consider retrying the allocation instead of
2020 static bool shrink_zones(int priority
, struct zonelist
*zonelist
,
2021 struct scan_control
*sc
)
2025 unsigned long nr_soft_reclaimed
;
2026 unsigned long nr_soft_scanned
;
2027 bool aborted_reclaim
= false;
2030 * If the number of buffer_heads in the machine exceeds the maximum
2031 * allowed level, force direct reclaim to scan the highmem zone as
2032 * highmem pages could be pinning lowmem pages storing buffer_heads
2034 if (buffer_heads_over_limit
)
2035 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2037 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2038 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2039 if (!populated_zone(zone
))
2042 * Take care memory controller reclaiming has small influence
2045 if (global_reclaim(sc
)) {
2046 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2048 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2049 continue; /* Let kswapd poll it */
2050 if (COMPACTION_BUILD
) {
2052 * If we already have plenty of memory free for
2053 * compaction in this zone, don't free any more.
2054 * Even though compaction is invoked for any
2055 * non-zero order, only frequent costly order
2056 * reclamation is disruptive enough to become a
2057 * noticeable problem, like transparent huge
2060 if (compaction_ready(zone
, sc
)) {
2061 aborted_reclaim
= true;
2066 * This steals pages from memory cgroups over softlimit
2067 * and returns the number of reclaimed pages and
2068 * scanned pages. This works for global memory pressure
2069 * and balancing, not for a memcg's limit.
2071 nr_soft_scanned
= 0;
2072 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2073 sc
->order
, sc
->gfp_mask
,
2075 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2076 sc
->nr_scanned
+= nr_soft_scanned
;
2077 /* need some check for avoid more shrink_zone() */
2080 shrink_zone(priority
, zone
, sc
);
2083 return aborted_reclaim
;
2086 static bool zone_reclaimable(struct zone
*zone
)
2088 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2091 /* All zones in zonelist are unreclaimable? */
2092 static bool all_unreclaimable(struct zonelist
*zonelist
,
2093 struct scan_control
*sc
)
2098 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2099 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2100 if (!populated_zone(zone
))
2102 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2104 if (!zone
->all_unreclaimable
)
2112 * This is the main entry point to direct page reclaim.
2114 * If a full scan of the inactive list fails to free enough memory then we
2115 * are "out of memory" and something needs to be killed.
2117 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2118 * high - the zone may be full of dirty or under-writeback pages, which this
2119 * caller can't do much about. We kick the writeback threads and take explicit
2120 * naps in the hope that some of these pages can be written. But if the
2121 * allocating task holds filesystem locks which prevent writeout this might not
2122 * work, and the allocation attempt will fail.
2124 * returns: 0, if no pages reclaimed
2125 * else, the number of pages reclaimed
2127 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2128 struct scan_control
*sc
,
2129 struct shrink_control
*shrink
)
2132 unsigned long total_scanned
= 0;
2133 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2136 unsigned long writeback_threshold
;
2137 bool aborted_reclaim
;
2139 delayacct_freepages_start();
2141 if (global_reclaim(sc
))
2142 count_vm_event(ALLOCSTALL
);
2144 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2146 aborted_reclaim
= shrink_zones(priority
, zonelist
, sc
);
2149 * Don't shrink slabs when reclaiming memory from
2150 * over limit cgroups
2152 if (global_reclaim(sc
)) {
2153 unsigned long lru_pages
= 0;
2154 for_each_zone_zonelist(zone
, z
, zonelist
,
2155 gfp_zone(sc
->gfp_mask
)) {
2156 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2159 lru_pages
+= zone_reclaimable_pages(zone
);
2162 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2163 if (reclaim_state
) {
2164 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2165 reclaim_state
->reclaimed_slab
= 0;
2168 total_scanned
+= sc
->nr_scanned
;
2169 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2173 * Try to write back as many pages as we just scanned. This
2174 * tends to cause slow streaming writers to write data to the
2175 * disk smoothly, at the dirtying rate, which is nice. But
2176 * that's undesirable in laptop mode, where we *want* lumpy
2177 * writeout. So in laptop mode, write out the whole world.
2179 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2180 if (total_scanned
> writeback_threshold
) {
2181 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2182 WB_REASON_TRY_TO_FREE_PAGES
);
2183 sc
->may_writepage
= 1;
2186 /* Take a nap, wait for some writeback to complete */
2187 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2188 priority
< DEF_PRIORITY
- 2) {
2189 struct zone
*preferred_zone
;
2191 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2192 &cpuset_current_mems_allowed
,
2194 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2199 delayacct_freepages_end();
2201 if (sc
->nr_reclaimed
)
2202 return sc
->nr_reclaimed
;
2205 * As hibernation is going on, kswapd is freezed so that it can't mark
2206 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2209 if (oom_killer_disabled
)
2212 /* Aborted reclaim to try compaction? don't OOM, then */
2213 if (aborted_reclaim
)
2216 /* top priority shrink_zones still had more to do? don't OOM, then */
2217 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2223 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2224 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2226 unsigned long nr_reclaimed
;
2227 struct scan_control sc
= {
2228 .gfp_mask
= gfp_mask
,
2229 .may_writepage
= !laptop_mode
,
2230 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2234 .target_mem_cgroup
= NULL
,
2235 .nodemask
= nodemask
,
2237 struct shrink_control shrink
= {
2238 .gfp_mask
= sc
.gfp_mask
,
2241 trace_mm_vmscan_direct_reclaim_begin(order
,
2245 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2247 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2249 return nr_reclaimed
;
2252 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2254 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2255 gfp_t gfp_mask
, bool noswap
,
2257 unsigned long *nr_scanned
)
2259 struct scan_control sc
= {
2261 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2262 .may_writepage
= !laptop_mode
,
2264 .may_swap
= !noswap
,
2266 .target_mem_cgroup
= memcg
,
2268 struct mem_cgroup_zone mz
= {
2269 .mem_cgroup
= memcg
,
2273 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2274 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2276 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2281 * NOTE: Although we can get the priority field, using it
2282 * here is not a good idea, since it limits the pages we can scan.
2283 * if we don't reclaim here, the shrink_zone from balance_pgdat
2284 * will pick up pages from other mem cgroup's as well. We hack
2285 * the priority and make it zero.
2287 shrink_mem_cgroup_zone(0, &mz
, &sc
);
2289 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2291 *nr_scanned
= sc
.nr_scanned
;
2292 return sc
.nr_reclaimed
;
2295 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2299 struct zonelist
*zonelist
;
2300 unsigned long nr_reclaimed
;
2302 struct scan_control sc
= {
2303 .may_writepage
= !laptop_mode
,
2305 .may_swap
= !noswap
,
2306 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2308 .target_mem_cgroup
= memcg
,
2309 .nodemask
= NULL
, /* we don't care the placement */
2310 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2311 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2313 struct shrink_control shrink
= {
2314 .gfp_mask
= sc
.gfp_mask
,
2318 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2319 * take care of from where we get pages. So the node where we start the
2320 * scan does not need to be the current node.
2322 nid
= mem_cgroup_select_victim_node(memcg
);
2324 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2326 trace_mm_vmscan_memcg_reclaim_begin(0,
2330 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2332 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2334 return nr_reclaimed
;
2338 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
,
2341 struct mem_cgroup
*memcg
;
2343 if (!total_swap_pages
)
2346 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2348 struct mem_cgroup_zone mz
= {
2349 .mem_cgroup
= memcg
,
2353 if (inactive_anon_is_low(&mz
))
2354 shrink_active_list(SWAP_CLUSTER_MAX
, &mz
,
2357 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2362 * pgdat_balanced is used when checking if a node is balanced for high-order
2363 * allocations. Only zones that meet watermarks and are in a zone allowed
2364 * by the callers classzone_idx are added to balanced_pages. The total of
2365 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2366 * for the node to be considered balanced. Forcing all zones to be balanced
2367 * for high orders can cause excessive reclaim when there are imbalanced zones.
2368 * The choice of 25% is due to
2369 * o a 16M DMA zone that is balanced will not balance a zone on any
2370 * reasonable sized machine
2371 * o On all other machines, the top zone must be at least a reasonable
2372 * percentage of the middle zones. For example, on 32-bit x86, highmem
2373 * would need to be at least 256M for it to be balance a whole node.
2374 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2375 * to balance a node on its own. These seemed like reasonable ratios.
2377 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2380 unsigned long present_pages
= 0;
2383 for (i
= 0; i
<= classzone_idx
; i
++)
2384 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2386 /* A special case here: if zone has no page, we think it's balanced */
2387 return balanced_pages
>= (present_pages
>> 2);
2390 /* is kswapd sleeping prematurely? */
2391 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2395 unsigned long balanced
= 0;
2396 bool all_zones_ok
= true;
2398 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2402 /* Check the watermark levels */
2403 for (i
= 0; i
<= classzone_idx
; i
++) {
2404 struct zone
*zone
= pgdat
->node_zones
+ i
;
2406 if (!populated_zone(zone
))
2410 * balance_pgdat() skips over all_unreclaimable after
2411 * DEF_PRIORITY. Effectively, it considers them balanced so
2412 * they must be considered balanced here as well if kswapd
2415 if (zone
->all_unreclaimable
) {
2416 balanced
+= zone
->present_pages
;
2420 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2422 all_zones_ok
= false;
2424 balanced
+= zone
->present_pages
;
2428 * For high-order requests, the balanced zones must contain at least
2429 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2433 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2435 return !all_zones_ok
;
2439 * For kswapd, balance_pgdat() will work across all this node's zones until
2440 * they are all at high_wmark_pages(zone).
2442 * Returns the final order kswapd was reclaiming at
2444 * There is special handling here for zones which are full of pinned pages.
2445 * This can happen if the pages are all mlocked, or if they are all used by
2446 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2447 * What we do is to detect the case where all pages in the zone have been
2448 * scanned twice and there has been zero successful reclaim. Mark the zone as
2449 * dead and from now on, only perform a short scan. Basically we're polling
2450 * the zone for when the problem goes away.
2452 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2453 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2454 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2455 * lower zones regardless of the number of free pages in the lower zones. This
2456 * interoperates with the page allocator fallback scheme to ensure that aging
2457 * of pages is balanced across the zones.
2459 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2463 unsigned long balanced
;
2466 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2467 unsigned long total_scanned
;
2468 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2469 unsigned long nr_soft_reclaimed
;
2470 unsigned long nr_soft_scanned
;
2471 struct scan_control sc
= {
2472 .gfp_mask
= GFP_KERNEL
,
2476 * kswapd doesn't want to be bailed out while reclaim. because
2477 * we want to put equal scanning pressure on each zone.
2479 .nr_to_reclaim
= ULONG_MAX
,
2481 .target_mem_cgroup
= NULL
,
2483 struct shrink_control shrink
= {
2484 .gfp_mask
= sc
.gfp_mask
,
2488 sc
.nr_reclaimed
= 0;
2489 sc
.may_writepage
= !laptop_mode
;
2490 count_vm_event(PAGEOUTRUN
);
2492 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2493 unsigned long lru_pages
= 0;
2494 int has_under_min_watermark_zone
= 0;
2500 * Scan in the highmem->dma direction for the highest
2501 * zone which needs scanning
2503 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2504 struct zone
*zone
= pgdat
->node_zones
+ i
;
2506 if (!populated_zone(zone
))
2509 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2513 * Do some background aging of the anon list, to give
2514 * pages a chance to be referenced before reclaiming.
2516 age_active_anon(zone
, &sc
, priority
);
2519 * If the number of buffer_heads in the machine
2520 * exceeds the maximum allowed level and this node
2521 * has a highmem zone, force kswapd to reclaim from
2522 * it to relieve lowmem pressure.
2524 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2529 if (!zone_watermark_ok_safe(zone
, order
,
2530 high_wmark_pages(zone
), 0, 0)) {
2534 /* If balanced, clear the congested flag */
2535 zone_clear_flag(zone
, ZONE_CONGESTED
);
2541 for (i
= 0; i
<= end_zone
; i
++) {
2542 struct zone
*zone
= pgdat
->node_zones
+ i
;
2544 lru_pages
+= zone_reclaimable_pages(zone
);
2548 * Now scan the zone in the dma->highmem direction, stopping
2549 * at the last zone which needs scanning.
2551 * We do this because the page allocator works in the opposite
2552 * direction. This prevents the page allocator from allocating
2553 * pages behind kswapd's direction of progress, which would
2554 * cause too much scanning of the lower zones.
2556 for (i
= 0; i
<= end_zone
; i
++) {
2557 struct zone
*zone
= pgdat
->node_zones
+ i
;
2558 int nr_slab
, testorder
;
2559 unsigned long balance_gap
;
2561 if (!populated_zone(zone
))
2564 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2569 nr_soft_scanned
= 0;
2571 * Call soft limit reclaim before calling shrink_zone.
2573 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2576 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2577 total_scanned
+= nr_soft_scanned
;
2580 * We put equal pressure on every zone, unless
2581 * one zone has way too many pages free
2582 * already. The "too many pages" is defined
2583 * as the high wmark plus a "gap" where the
2584 * gap is either the low watermark or 1%
2585 * of the zone, whichever is smaller.
2587 balance_gap
= min(low_wmark_pages(zone
),
2588 (zone
->present_pages
+
2589 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2590 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2592 * Kswapd reclaims only single pages with compaction
2593 * enabled. Trying too hard to reclaim until contiguous
2594 * free pages have become available can hurt performance
2595 * by evicting too much useful data from memory.
2596 * Do not reclaim more than needed for compaction.
2599 if (COMPACTION_BUILD
&& order
&&
2600 compaction_suitable(zone
, order
) !=
2604 if ((buffer_heads_over_limit
&& is_highmem_idx(i
)) ||
2605 !zone_watermark_ok_safe(zone
, testorder
,
2606 high_wmark_pages(zone
) + balance_gap
,
2608 shrink_zone(priority
, zone
, &sc
);
2610 reclaim_state
->reclaimed_slab
= 0;
2611 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2612 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2613 total_scanned
+= sc
.nr_scanned
;
2615 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2616 zone
->all_unreclaimable
= 1;
2620 * If we've done a decent amount of scanning and
2621 * the reclaim ratio is low, start doing writepage
2622 * even in laptop mode
2624 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2625 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2626 sc
.may_writepage
= 1;
2628 if (zone
->all_unreclaimable
) {
2629 if (end_zone
&& end_zone
== i
)
2634 if (!zone_watermark_ok_safe(zone
, testorder
,
2635 high_wmark_pages(zone
), end_zone
, 0)) {
2638 * We are still under min water mark. This
2639 * means that we have a GFP_ATOMIC allocation
2640 * failure risk. Hurry up!
2642 if (!zone_watermark_ok_safe(zone
, order
,
2643 min_wmark_pages(zone
), end_zone
, 0))
2644 has_under_min_watermark_zone
= 1;
2647 * If a zone reaches its high watermark,
2648 * consider it to be no longer congested. It's
2649 * possible there are dirty pages backed by
2650 * congested BDIs but as pressure is relieved,
2651 * spectulatively avoid congestion waits
2653 zone_clear_flag(zone
, ZONE_CONGESTED
);
2654 if (i
<= *classzone_idx
)
2655 balanced
+= zone
->present_pages
;
2659 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2660 break; /* kswapd: all done */
2662 * OK, kswapd is getting into trouble. Take a nap, then take
2663 * another pass across the zones.
2665 if (total_scanned
&& (priority
< DEF_PRIORITY
- 2)) {
2666 if (has_under_min_watermark_zone
)
2667 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2669 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2673 * We do this so kswapd doesn't build up large priorities for
2674 * example when it is freeing in parallel with allocators. It
2675 * matches the direct reclaim path behaviour in terms of impact
2676 * on zone->*_priority.
2678 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2684 * order-0: All zones must meet high watermark for a balanced node
2685 * high-order: Balanced zones must make up at least 25% of the node
2686 * for the node to be balanced
2688 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2694 * Fragmentation may mean that the system cannot be
2695 * rebalanced for high-order allocations in all zones.
2696 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2697 * it means the zones have been fully scanned and are still
2698 * not balanced. For high-order allocations, there is
2699 * little point trying all over again as kswapd may
2702 * Instead, recheck all watermarks at order-0 as they
2703 * are the most important. If watermarks are ok, kswapd will go
2704 * back to sleep. High-order users can still perform direct
2705 * reclaim if they wish.
2707 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2708 order
= sc
.order
= 0;
2714 * If kswapd was reclaiming at a higher order, it has the option of
2715 * sleeping without all zones being balanced. Before it does, it must
2716 * ensure that the watermarks for order-0 on *all* zones are met and
2717 * that the congestion flags are cleared. The congestion flag must
2718 * be cleared as kswapd is the only mechanism that clears the flag
2719 * and it is potentially going to sleep here.
2722 int zones_need_compaction
= 1;
2724 for (i
= 0; i
<= end_zone
; i
++) {
2725 struct zone
*zone
= pgdat
->node_zones
+ i
;
2727 if (!populated_zone(zone
))
2730 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2733 /* Would compaction fail due to lack of free memory? */
2734 if (COMPACTION_BUILD
&&
2735 compaction_suitable(zone
, order
) == COMPACT_SKIPPED
)
2738 /* Confirm the zone is balanced for order-0 */
2739 if (!zone_watermark_ok(zone
, 0,
2740 high_wmark_pages(zone
), 0, 0)) {
2741 order
= sc
.order
= 0;
2745 /* Check if the memory needs to be defragmented. */
2746 if (zone_watermark_ok(zone
, order
,
2747 low_wmark_pages(zone
), *classzone_idx
, 0))
2748 zones_need_compaction
= 0;
2750 /* If balanced, clear the congested flag */
2751 zone_clear_flag(zone
, ZONE_CONGESTED
);
2754 if (zones_need_compaction
)
2755 compact_pgdat(pgdat
, order
);
2759 * Return the order we were reclaiming at so sleeping_prematurely()
2760 * makes a decision on the order we were last reclaiming at. However,
2761 * if another caller entered the allocator slow path while kswapd
2762 * was awake, order will remain at the higher level
2764 *classzone_idx
= end_zone
;
2768 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2773 if (freezing(current
) || kthread_should_stop())
2776 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2778 /* Try to sleep for a short interval */
2779 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2780 remaining
= schedule_timeout(HZ
/10);
2781 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2782 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2786 * After a short sleep, check if it was a premature sleep. If not, then
2787 * go fully to sleep until explicitly woken up.
2789 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2790 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2793 * vmstat counters are not perfectly accurate and the estimated
2794 * value for counters such as NR_FREE_PAGES can deviate from the
2795 * true value by nr_online_cpus * threshold. To avoid the zone
2796 * watermarks being breached while under pressure, we reduce the
2797 * per-cpu vmstat threshold while kswapd is awake and restore
2798 * them before going back to sleep.
2800 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2802 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2805 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2807 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2809 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2813 * The background pageout daemon, started as a kernel thread
2814 * from the init process.
2816 * This basically trickles out pages so that we have _some_
2817 * free memory available even if there is no other activity
2818 * that frees anything up. This is needed for things like routing
2819 * etc, where we otherwise might have all activity going on in
2820 * asynchronous contexts that cannot page things out.
2822 * If there are applications that are active memory-allocators
2823 * (most normal use), this basically shouldn't matter.
2825 static int kswapd(void *p
)
2827 unsigned long order
, new_order
;
2828 unsigned balanced_order
;
2829 int classzone_idx
, new_classzone_idx
;
2830 int balanced_classzone_idx
;
2831 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2832 struct task_struct
*tsk
= current
;
2834 struct reclaim_state reclaim_state
= {
2835 .reclaimed_slab
= 0,
2837 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2839 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2841 if (!cpumask_empty(cpumask
))
2842 set_cpus_allowed_ptr(tsk
, cpumask
);
2843 current
->reclaim_state
= &reclaim_state
;
2846 * Tell the memory management that we're a "memory allocator",
2847 * and that if we need more memory we should get access to it
2848 * regardless (see "__alloc_pages()"). "kswapd" should
2849 * never get caught in the normal page freeing logic.
2851 * (Kswapd normally doesn't need memory anyway, but sometimes
2852 * you need a small amount of memory in order to be able to
2853 * page out something else, and this flag essentially protects
2854 * us from recursively trying to free more memory as we're
2855 * trying to free the first piece of memory in the first place).
2857 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2860 order
= new_order
= 0;
2862 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
2863 balanced_classzone_idx
= classzone_idx
;
2868 * If the last balance_pgdat was unsuccessful it's unlikely a
2869 * new request of a similar or harder type will succeed soon
2870 * so consider going to sleep on the basis we reclaimed at
2872 if (balanced_classzone_idx
>= new_classzone_idx
&&
2873 balanced_order
== new_order
) {
2874 new_order
= pgdat
->kswapd_max_order
;
2875 new_classzone_idx
= pgdat
->classzone_idx
;
2876 pgdat
->kswapd_max_order
= 0;
2877 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2880 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2882 * Don't sleep if someone wants a larger 'order'
2883 * allocation or has tigher zone constraints
2886 classzone_idx
= new_classzone_idx
;
2888 kswapd_try_to_sleep(pgdat
, balanced_order
,
2889 balanced_classzone_idx
);
2890 order
= pgdat
->kswapd_max_order
;
2891 classzone_idx
= pgdat
->classzone_idx
;
2893 new_classzone_idx
= classzone_idx
;
2894 pgdat
->kswapd_max_order
= 0;
2895 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2898 ret
= try_to_freeze();
2899 if (kthread_should_stop())
2903 * We can speed up thawing tasks if we don't call balance_pgdat
2904 * after returning from the refrigerator
2907 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2908 balanced_classzone_idx
= classzone_idx
;
2909 balanced_order
= balance_pgdat(pgdat
, order
,
2910 &balanced_classzone_idx
);
2917 * A zone is low on free memory, so wake its kswapd task to service it.
2919 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2923 if (!populated_zone(zone
))
2926 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2928 pgdat
= zone
->zone_pgdat
;
2929 if (pgdat
->kswapd_max_order
< order
) {
2930 pgdat
->kswapd_max_order
= order
;
2931 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2933 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2935 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2938 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2939 wake_up_interruptible(&pgdat
->kswapd_wait
);
2943 * The reclaimable count would be mostly accurate.
2944 * The less reclaimable pages may be
2945 * - mlocked pages, which will be moved to unevictable list when encountered
2946 * - mapped pages, which may require several travels to be reclaimed
2947 * - dirty pages, which is not "instantly" reclaimable
2949 unsigned long global_reclaimable_pages(void)
2953 nr
= global_page_state(NR_ACTIVE_FILE
) +
2954 global_page_state(NR_INACTIVE_FILE
);
2956 if (nr_swap_pages
> 0)
2957 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2958 global_page_state(NR_INACTIVE_ANON
);
2963 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2967 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2968 zone_page_state(zone
, NR_INACTIVE_FILE
);
2970 if (nr_swap_pages
> 0)
2971 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2972 zone_page_state(zone
, NR_INACTIVE_ANON
);
2977 #ifdef CONFIG_HIBERNATION
2979 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2982 * Rather than trying to age LRUs the aim is to preserve the overall
2983 * LRU order by reclaiming preferentially
2984 * inactive > active > active referenced > active mapped
2986 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2988 struct reclaim_state reclaim_state
;
2989 struct scan_control sc
= {
2990 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2994 .nr_to_reclaim
= nr_to_reclaim
,
2995 .hibernation_mode
= 1,
2998 struct shrink_control shrink
= {
2999 .gfp_mask
= sc
.gfp_mask
,
3001 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3002 struct task_struct
*p
= current
;
3003 unsigned long nr_reclaimed
;
3005 p
->flags
|= PF_MEMALLOC
;
3006 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3007 reclaim_state
.reclaimed_slab
= 0;
3008 p
->reclaim_state
= &reclaim_state
;
3010 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3012 p
->reclaim_state
= NULL
;
3013 lockdep_clear_current_reclaim_state();
3014 p
->flags
&= ~PF_MEMALLOC
;
3016 return nr_reclaimed
;
3018 #endif /* CONFIG_HIBERNATION */
3020 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3021 not required for correctness. So if the last cpu in a node goes
3022 away, we get changed to run anywhere: as the first one comes back,
3023 restore their cpu bindings. */
3024 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
3025 unsigned long action
, void *hcpu
)
3029 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3030 for_each_node_state(nid
, N_HIGH_MEMORY
) {
3031 pg_data_t
*pgdat
= NODE_DATA(nid
);
3032 const struct cpumask
*mask
;
3034 mask
= cpumask_of_node(pgdat
->node_id
);
3036 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3037 /* One of our CPUs online: restore mask */
3038 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3045 * This kswapd start function will be called by init and node-hot-add.
3046 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3048 int kswapd_run(int nid
)
3050 pg_data_t
*pgdat
= NODE_DATA(nid
);
3056 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3057 if (IS_ERR(pgdat
->kswapd
)) {
3058 /* failure at boot is fatal */
3059 BUG_ON(system_state
== SYSTEM_BOOTING
);
3060 printk("Failed to start kswapd on node %d\n",nid
);
3067 * Called by memory hotplug when all memory in a node is offlined.
3069 void kswapd_stop(int nid
)
3071 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3074 kthread_stop(kswapd
);
3077 static int __init
kswapd_init(void)
3082 for_each_node_state(nid
, N_HIGH_MEMORY
)
3084 hotcpu_notifier(cpu_callback
, 0);
3088 module_init(kswapd_init
)
3094 * If non-zero call zone_reclaim when the number of free pages falls below
3097 int zone_reclaim_mode __read_mostly
;
3099 #define RECLAIM_OFF 0
3100 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3101 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3102 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3105 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3106 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3109 #define ZONE_RECLAIM_PRIORITY 4
3112 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3115 int sysctl_min_unmapped_ratio
= 1;
3118 * If the number of slab pages in a zone grows beyond this percentage then
3119 * slab reclaim needs to occur.
3121 int sysctl_min_slab_ratio
= 5;
3123 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3125 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3126 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3127 zone_page_state(zone
, NR_ACTIVE_FILE
);
3130 * It's possible for there to be more file mapped pages than
3131 * accounted for by the pages on the file LRU lists because
3132 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3134 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3137 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3138 static long zone_pagecache_reclaimable(struct zone
*zone
)
3140 long nr_pagecache_reclaimable
;
3144 * If RECLAIM_SWAP is set, then all file pages are considered
3145 * potentially reclaimable. Otherwise, we have to worry about
3146 * pages like swapcache and zone_unmapped_file_pages() provides
3149 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3150 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3152 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3154 /* If we can't clean pages, remove dirty pages from consideration */
3155 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3156 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3158 /* Watch for any possible underflows due to delta */
3159 if (unlikely(delta
> nr_pagecache_reclaimable
))
3160 delta
= nr_pagecache_reclaimable
;
3162 return nr_pagecache_reclaimable
- delta
;
3166 * Try to free up some pages from this zone through reclaim.
3168 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3170 /* Minimum pages needed in order to stay on node */
3171 const unsigned long nr_pages
= 1 << order
;
3172 struct task_struct
*p
= current
;
3173 struct reclaim_state reclaim_state
;
3175 struct scan_control sc
= {
3176 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3177 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3179 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3181 .gfp_mask
= gfp_mask
,
3184 struct shrink_control shrink
= {
3185 .gfp_mask
= sc
.gfp_mask
,
3187 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3191 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3192 * and we also need to be able to write out pages for RECLAIM_WRITE
3195 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3196 lockdep_set_current_reclaim_state(gfp_mask
);
3197 reclaim_state
.reclaimed_slab
= 0;
3198 p
->reclaim_state
= &reclaim_state
;
3200 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3202 * Free memory by calling shrink zone with increasing
3203 * priorities until we have enough memory freed.
3205 priority
= ZONE_RECLAIM_PRIORITY
;
3207 shrink_zone(priority
, zone
, &sc
);
3209 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
3212 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3213 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3215 * shrink_slab() does not currently allow us to determine how
3216 * many pages were freed in this zone. So we take the current
3217 * number of slab pages and shake the slab until it is reduced
3218 * by the same nr_pages that we used for reclaiming unmapped
3221 * Note that shrink_slab will free memory on all zones and may
3225 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3227 /* No reclaimable slab or very low memory pressure */
3228 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3231 /* Freed enough memory */
3232 nr_slab_pages1
= zone_page_state(zone
,
3233 NR_SLAB_RECLAIMABLE
);
3234 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3239 * Update nr_reclaimed by the number of slab pages we
3240 * reclaimed from this zone.
3242 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3243 if (nr_slab_pages1
< nr_slab_pages0
)
3244 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3247 p
->reclaim_state
= NULL
;
3248 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3249 lockdep_clear_current_reclaim_state();
3250 return sc
.nr_reclaimed
>= nr_pages
;
3253 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3259 * Zone reclaim reclaims unmapped file backed pages and
3260 * slab pages if we are over the defined limits.
3262 * A small portion of unmapped file backed pages is needed for
3263 * file I/O otherwise pages read by file I/O will be immediately
3264 * thrown out if the zone is overallocated. So we do not reclaim
3265 * if less than a specified percentage of the zone is used by
3266 * unmapped file backed pages.
3268 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3269 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3270 return ZONE_RECLAIM_FULL
;
3272 if (zone
->all_unreclaimable
)
3273 return ZONE_RECLAIM_FULL
;
3276 * Do not scan if the allocation should not be delayed.
3278 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3279 return ZONE_RECLAIM_NOSCAN
;
3282 * Only run zone reclaim on the local zone or on zones that do not
3283 * have associated processors. This will favor the local processor
3284 * over remote processors and spread off node memory allocations
3285 * as wide as possible.
3287 node_id
= zone_to_nid(zone
);
3288 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3289 return ZONE_RECLAIM_NOSCAN
;
3291 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3292 return ZONE_RECLAIM_NOSCAN
;
3294 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3295 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3298 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3305 * page_evictable - test whether a page is evictable
3306 * @page: the page to test
3307 * @vma: the VMA in which the page is or will be mapped, may be NULL
3309 * Test whether page is evictable--i.e., should be placed on active/inactive
3310 * lists vs unevictable list. The vma argument is !NULL when called from the
3311 * fault path to determine how to instantate a new page.
3313 * Reasons page might not be evictable:
3314 * (1) page's mapping marked unevictable
3315 * (2) page is part of an mlocked VMA
3318 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3321 if (mapping_unevictable(page_mapping(page
)))
3324 if (PageMlocked(page
) || (vma
&& mlocked_vma_newpage(vma
, page
)))
3332 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3333 * @pages: array of pages to check
3334 * @nr_pages: number of pages to check
3336 * Checks pages for evictability and moves them to the appropriate lru list.
3338 * This function is only used for SysV IPC SHM_UNLOCK.
3340 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3342 struct lruvec
*lruvec
;
3343 struct zone
*zone
= NULL
;
3348 for (i
= 0; i
< nr_pages
; i
++) {
3349 struct page
*page
= pages
[i
];
3350 struct zone
*pagezone
;
3353 pagezone
= page_zone(page
);
3354 if (pagezone
!= zone
) {
3356 spin_unlock_irq(&zone
->lru_lock
);
3358 spin_lock_irq(&zone
->lru_lock
);
3361 if (!PageLRU(page
) || !PageUnevictable(page
))
3364 if (page_evictable(page
, NULL
)) {
3365 enum lru_list lru
= page_lru_base_type(page
);
3367 VM_BUG_ON(PageActive(page
));
3368 ClearPageUnevictable(page
);
3369 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3370 lruvec
= mem_cgroup_lru_move_lists(zone
, page
,
3371 LRU_UNEVICTABLE
, lru
);
3372 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
3373 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ lru
);
3379 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3380 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3381 spin_unlock_irq(&zone
->lru_lock
);
3384 #endif /* CONFIG_SHMEM */
3386 static void warn_scan_unevictable_pages(void)
3388 printk_once(KERN_WARNING
3389 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3390 "disabled for lack of a legitimate use case. If you have "
3391 "one, please send an email to linux-mm@kvack.org.\n",
3396 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3397 * all nodes' unevictable lists for evictable pages
3399 unsigned long scan_unevictable_pages
;
3401 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3402 void __user
*buffer
,
3403 size_t *length
, loff_t
*ppos
)
3405 warn_scan_unevictable_pages();
3406 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3407 scan_unevictable_pages
= 0;
3413 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3414 * a specified node's per zone unevictable lists for evictable pages.
3417 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3418 struct device_attribute
*attr
,
3421 warn_scan_unevictable_pages();
3422 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3425 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3426 struct device_attribute
*attr
,
3427 const char *buf
, size_t count
)
3429 warn_scan_unevictable_pages();
3434 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3435 read_scan_unevictable_node
,
3436 write_scan_unevictable_node
);
3438 int scan_unevictable_register_node(struct node
*node
)
3440 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3443 void scan_unevictable_unregister_node(struct node
*node
)
3445 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);