4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66 * order-0 pages and then compact the zone
68 typedef unsigned __bitwise__ reclaim_mode_t
;
69 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
76 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned
;
79 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed
;
82 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim
;
85 unsigned long hibernation_mode
;
87 /* This context's GFP mask */
92 /* Can mapped pages be reclaimed? */
95 /* Can pages be swapped as part of reclaim? */
103 * Intend to reclaim enough continuous memory rather than reclaim
104 * enough amount of memory. i.e, mode for high order allocation.
106 reclaim_mode_t reclaim_mode
;
108 /* Which cgroup do we reclaim from */
109 struct mem_cgroup
*mem_cgroup
;
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
115 nodemask_t
*nodemask
;
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness
= 60;
152 long vm_total_pages
; /* The total number of pages which the VM controls */
154 static LIST_HEAD(shrinker_list
);
155 static DECLARE_RWSEM(shrinker_rwsem
);
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
160 #define scanning_global_lru(sc) (1)
163 static struct zone_reclaim_stat
*get_reclaim_stat(struct zone
*zone
,
164 struct scan_control
*sc
)
166 if (!scanning_global_lru(sc
))
167 return mem_cgroup_get_reclaim_stat(sc
->mem_cgroup
, zone
);
169 return &zone
->reclaim_stat
;
172 static unsigned long zone_nr_lru_pages(struct zone
*zone
,
173 struct scan_control
*sc
, enum lru_list lru
)
175 if (!scanning_global_lru(sc
))
176 return mem_cgroup_zone_nr_pages(sc
->mem_cgroup
, zone
, lru
);
178 return zone_page_state(zone
, NR_LRU_BASE
+ lru
);
183 * Add a shrinker callback to be called from the vm
185 void register_shrinker(struct shrinker
*shrinker
)
188 down_write(&shrinker_rwsem
);
189 list_add_tail(&shrinker
->list
, &shrinker_list
);
190 up_write(&shrinker_rwsem
);
192 EXPORT_SYMBOL(register_shrinker
);
197 void unregister_shrinker(struct shrinker
*shrinker
)
199 down_write(&shrinker_rwsem
);
200 list_del(&shrinker
->list
);
201 up_write(&shrinker_rwsem
);
203 EXPORT_SYMBOL(unregister_shrinker
);
205 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
206 struct shrink_control
*sc
,
207 unsigned long nr_to_scan
)
209 sc
->nr_to_scan
= nr_to_scan
;
210 return (*shrinker
->shrink
)(shrinker
, sc
);
213 #define SHRINK_BATCH 128
215 * Call the shrink functions to age shrinkable caches
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
222 * If the vm encountered mapped pages on the LRU it increase the pressure on
223 * slab to avoid swapping.
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
231 * Returns the number of slab objects which we shrunk.
233 unsigned long shrink_slab(struct shrink_control
*shrink
,
234 unsigned long nr_pages_scanned
,
235 unsigned long lru_pages
)
237 struct shrinker
*shrinker
;
238 unsigned long ret
= 0;
240 if (nr_pages_scanned
== 0)
241 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
243 if (!down_read_trylock(&shrinker_rwsem
)) {
244 /* Assume we'll be able to shrink next time */
249 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
250 unsigned long long delta
;
251 unsigned long total_scan
;
252 unsigned long max_pass
;
254 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
255 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
257 do_div(delta
, lru_pages
+ 1);
258 shrinker
->nr
+= delta
;
259 if (shrinker
->nr
< 0) {
260 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
262 shrinker
->shrink
, shrinker
->nr
);
263 shrinker
->nr
= max_pass
;
267 * Avoid risking looping forever due to too large nr value:
268 * never try to free more than twice the estimate number of
271 if (shrinker
->nr
> max_pass
* 2)
272 shrinker
->nr
= max_pass
* 2;
274 total_scan
= shrinker
->nr
;
277 while (total_scan
>= SHRINK_BATCH
) {
278 long this_scan
= SHRINK_BATCH
;
282 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
283 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
285 if (shrink_ret
== -1)
287 if (shrink_ret
< nr_before
)
288 ret
+= nr_before
- shrink_ret
;
289 count_vm_events(SLABS_SCANNED
, this_scan
);
290 total_scan
-= this_scan
;
295 shrinker
->nr
+= total_scan
;
297 up_read(&shrinker_rwsem
);
303 static void set_reclaim_mode(int priority
, struct scan_control
*sc
,
306 reclaim_mode_t syncmode
= sync
? RECLAIM_MODE_SYNC
: RECLAIM_MODE_ASYNC
;
309 * Initially assume we are entering either lumpy reclaim or
310 * reclaim/compaction.Depending on the order, we will either set the
311 * sync mode or just reclaim order-0 pages later.
313 if (COMPACTION_BUILD
)
314 sc
->reclaim_mode
= RECLAIM_MODE_COMPACTION
;
316 sc
->reclaim_mode
= RECLAIM_MODE_LUMPYRECLAIM
;
319 * Avoid using lumpy reclaim or reclaim/compaction if possible by
320 * restricting when its set to either costly allocations or when
321 * under memory pressure
323 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
324 sc
->reclaim_mode
|= syncmode
;
325 else if (sc
->order
&& priority
< DEF_PRIORITY
- 2)
326 sc
->reclaim_mode
|= syncmode
;
328 sc
->reclaim_mode
= RECLAIM_MODE_SINGLE
| RECLAIM_MODE_ASYNC
;
331 static void reset_reclaim_mode(struct scan_control
*sc
)
333 sc
->reclaim_mode
= RECLAIM_MODE_SINGLE
| RECLAIM_MODE_ASYNC
;
336 static inline int is_page_cache_freeable(struct page
*page
)
339 * A freeable page cache page is referenced only by the caller
340 * that isolated the page, the page cache radix tree and
341 * optional buffer heads at page->private.
343 return page_count(page
) - page_has_private(page
) == 2;
346 static int may_write_to_queue(struct backing_dev_info
*bdi
,
347 struct scan_control
*sc
)
349 if (current
->flags
& PF_SWAPWRITE
)
351 if (!bdi_write_congested(bdi
))
353 if (bdi
== current
->backing_dev_info
)
356 /* lumpy reclaim for hugepage often need a lot of write */
357 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
363 * We detected a synchronous write error writing a page out. Probably
364 * -ENOSPC. We need to propagate that into the address_space for a subsequent
365 * fsync(), msync() or close().
367 * The tricky part is that after writepage we cannot touch the mapping: nothing
368 * prevents it from being freed up. But we have a ref on the page and once
369 * that page is locked, the mapping is pinned.
371 * We're allowed to run sleeping lock_page() here because we know the caller has
374 static void handle_write_error(struct address_space
*mapping
,
375 struct page
*page
, int error
)
378 if (page_mapping(page
) == mapping
)
379 mapping_set_error(mapping
, error
);
383 /* possible outcome of pageout() */
385 /* failed to write page out, page is locked */
387 /* move page to the active list, page is locked */
389 /* page has been sent to the disk successfully, page is unlocked */
391 /* page is clean and locked */
396 * pageout is called by shrink_page_list() for each dirty page.
397 * Calls ->writepage().
399 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
400 struct scan_control
*sc
)
403 * If the page is dirty, only perform writeback if that write
404 * will be non-blocking. To prevent this allocation from being
405 * stalled by pagecache activity. But note that there may be
406 * stalls if we need to run get_block(). We could test
407 * PagePrivate for that.
409 * If this process is currently in __generic_file_aio_write() against
410 * this page's queue, we can perform writeback even if that
413 * If the page is swapcache, write it back even if that would
414 * block, for some throttling. This happens by accident, because
415 * swap_backing_dev_info is bust: it doesn't reflect the
416 * congestion state of the swapdevs. Easy to fix, if needed.
418 if (!is_page_cache_freeable(page
))
422 * Some data journaling orphaned pages can have
423 * page->mapping == NULL while being dirty with clean buffers.
425 if (page_has_private(page
)) {
426 if (try_to_free_buffers(page
)) {
427 ClearPageDirty(page
);
428 printk("%s: orphaned page\n", __func__
);
434 if (mapping
->a_ops
->writepage
== NULL
)
435 return PAGE_ACTIVATE
;
436 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
439 if (clear_page_dirty_for_io(page
)) {
441 struct writeback_control wbc
= {
442 .sync_mode
= WB_SYNC_NONE
,
443 .nr_to_write
= SWAP_CLUSTER_MAX
,
445 .range_end
= LLONG_MAX
,
449 SetPageReclaim(page
);
450 res
= mapping
->a_ops
->writepage(page
, &wbc
);
452 handle_write_error(mapping
, page
, res
);
453 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
454 ClearPageReclaim(page
);
455 return PAGE_ACTIVATE
;
459 * Wait on writeback if requested to. This happens when
460 * direct reclaiming a large contiguous area and the
461 * first attempt to free a range of pages fails.
463 if (PageWriteback(page
) &&
464 (sc
->reclaim_mode
& RECLAIM_MODE_SYNC
))
465 wait_on_page_writeback(page
);
467 if (!PageWriteback(page
)) {
468 /* synchronous write or broken a_ops? */
469 ClearPageReclaim(page
);
471 trace_mm_vmscan_writepage(page
,
472 trace_reclaim_flags(page
, sc
->reclaim_mode
));
473 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
481 * Same as remove_mapping, but if the page is removed from the mapping, it
482 * gets returned with a refcount of 0.
484 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
486 BUG_ON(!PageLocked(page
));
487 BUG_ON(mapping
!= page_mapping(page
));
489 spin_lock_irq(&mapping
->tree_lock
);
491 * The non racy check for a busy page.
493 * Must be careful with the order of the tests. When someone has
494 * a ref to the page, it may be possible that they dirty it then
495 * drop the reference. So if PageDirty is tested before page_count
496 * here, then the following race may occur:
498 * get_user_pages(&page);
499 * [user mapping goes away]
501 * !PageDirty(page) [good]
502 * SetPageDirty(page);
504 * !page_count(page) [good, discard it]
506 * [oops, our write_to data is lost]
508 * Reversing the order of the tests ensures such a situation cannot
509 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
510 * load is not satisfied before that of page->_count.
512 * Note that if SetPageDirty is always performed via set_page_dirty,
513 * and thus under tree_lock, then this ordering is not required.
515 if (!page_freeze_refs(page
, 2))
517 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
518 if (unlikely(PageDirty(page
))) {
519 page_unfreeze_refs(page
, 2);
523 if (PageSwapCache(page
)) {
524 swp_entry_t swap
= { .val
= page_private(page
) };
525 __delete_from_swap_cache(page
);
526 spin_unlock_irq(&mapping
->tree_lock
);
527 swapcache_free(swap
, page
);
529 void (*freepage
)(struct page
*);
531 freepage
= mapping
->a_ops
->freepage
;
533 __delete_from_page_cache(page
);
534 spin_unlock_irq(&mapping
->tree_lock
);
535 mem_cgroup_uncharge_cache_page(page
);
537 if (freepage
!= NULL
)
544 spin_unlock_irq(&mapping
->tree_lock
);
549 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
550 * someone else has a ref on the page, abort and return 0. If it was
551 * successfully detached, return 1. Assumes the caller has a single ref on
554 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
556 if (__remove_mapping(mapping
, page
)) {
558 * Unfreezing the refcount with 1 rather than 2 effectively
559 * drops the pagecache ref for us without requiring another
562 page_unfreeze_refs(page
, 1);
569 * putback_lru_page - put previously isolated page onto appropriate LRU list
570 * @page: page to be put back to appropriate lru list
572 * Add previously isolated @page to appropriate LRU list.
573 * Page may still be unevictable for other reasons.
575 * lru_lock must not be held, interrupts must be enabled.
577 void putback_lru_page(struct page
*page
)
580 int active
= !!TestClearPageActive(page
);
581 int was_unevictable
= PageUnevictable(page
);
583 VM_BUG_ON(PageLRU(page
));
586 ClearPageUnevictable(page
);
588 if (page_evictable(page
, NULL
)) {
590 * For evictable pages, we can use the cache.
591 * In event of a race, worst case is we end up with an
592 * unevictable page on [in]active list.
593 * We know how to handle that.
595 lru
= active
+ page_lru_base_type(page
);
596 lru_cache_add_lru(page
, lru
);
599 * Put unevictable pages directly on zone's unevictable
602 lru
= LRU_UNEVICTABLE
;
603 add_page_to_unevictable_list(page
);
605 * When racing with an mlock clearing (page is
606 * unlocked), make sure that if the other thread does
607 * not observe our setting of PG_lru and fails
608 * isolation, we see PG_mlocked cleared below and move
609 * the page back to the evictable list.
611 * The other side is TestClearPageMlocked().
617 * page's status can change while we move it among lru. If an evictable
618 * page is on unevictable list, it never be freed. To avoid that,
619 * check after we added it to the list, again.
621 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
622 if (!isolate_lru_page(page
)) {
626 /* This means someone else dropped this page from LRU
627 * So, it will be freed or putback to LRU again. There is
628 * nothing to do here.
632 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
633 count_vm_event(UNEVICTABLE_PGRESCUED
);
634 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
635 count_vm_event(UNEVICTABLE_PGCULLED
);
637 put_page(page
); /* drop ref from isolate */
640 enum page_references
{
642 PAGEREF_RECLAIM_CLEAN
,
647 static enum page_references
page_check_references(struct page
*page
,
648 struct scan_control
*sc
)
650 int referenced_ptes
, referenced_page
;
651 unsigned long vm_flags
;
653 referenced_ptes
= page_referenced(page
, 1, sc
->mem_cgroup
, &vm_flags
);
654 referenced_page
= TestClearPageReferenced(page
);
656 /* Lumpy reclaim - ignore references */
657 if (sc
->reclaim_mode
& RECLAIM_MODE_LUMPYRECLAIM
)
658 return PAGEREF_RECLAIM
;
661 * Mlock lost the isolation race with us. Let try_to_unmap()
662 * move the page to the unevictable list.
664 if (vm_flags
& VM_LOCKED
)
665 return PAGEREF_RECLAIM
;
667 if (referenced_ptes
) {
669 return PAGEREF_ACTIVATE
;
671 * All mapped pages start out with page table
672 * references from the instantiating fault, so we need
673 * to look twice if a mapped file page is used more
676 * Mark it and spare it for another trip around the
677 * inactive list. Another page table reference will
678 * lead to its activation.
680 * Note: the mark is set for activated pages as well
681 * so that recently deactivated but used pages are
684 SetPageReferenced(page
);
687 return PAGEREF_ACTIVATE
;
692 /* Reclaim if clean, defer dirty pages to writeback */
693 if (referenced_page
&& !PageSwapBacked(page
))
694 return PAGEREF_RECLAIM_CLEAN
;
696 return PAGEREF_RECLAIM
;
699 static noinline_for_stack
void free_page_list(struct list_head
*free_pages
)
701 struct pagevec freed_pvec
;
702 struct page
*page
, *tmp
;
704 pagevec_init(&freed_pvec
, 1);
706 list_for_each_entry_safe(page
, tmp
, free_pages
, lru
) {
707 list_del(&page
->lru
);
708 if (!pagevec_add(&freed_pvec
, page
)) {
709 __pagevec_free(&freed_pvec
);
710 pagevec_reinit(&freed_pvec
);
714 pagevec_free(&freed_pvec
);
718 * shrink_page_list() returns the number of reclaimed pages
720 static unsigned long shrink_page_list(struct list_head
*page_list
,
722 struct scan_control
*sc
)
724 LIST_HEAD(ret_pages
);
725 LIST_HEAD(free_pages
);
727 unsigned long nr_dirty
= 0;
728 unsigned long nr_congested
= 0;
729 unsigned long nr_reclaimed
= 0;
733 while (!list_empty(page_list
)) {
734 enum page_references references
;
735 struct address_space
*mapping
;
741 page
= lru_to_page(page_list
);
742 list_del(&page
->lru
);
744 if (!trylock_page(page
))
747 VM_BUG_ON(PageActive(page
));
748 VM_BUG_ON(page_zone(page
) != zone
);
752 if (unlikely(!page_evictable(page
, NULL
)))
755 if (!sc
->may_unmap
&& page_mapped(page
))
758 /* Double the slab pressure for mapped and swapcache pages */
759 if (page_mapped(page
) || PageSwapCache(page
))
762 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
763 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
765 if (PageWriteback(page
)) {
767 * Synchronous reclaim is performed in two passes,
768 * first an asynchronous pass over the list to
769 * start parallel writeback, and a second synchronous
770 * pass to wait for the IO to complete. Wait here
771 * for any page for which writeback has already
774 if ((sc
->reclaim_mode
& RECLAIM_MODE_SYNC
) &&
776 wait_on_page_writeback(page
);
783 references
= page_check_references(page
, sc
);
784 switch (references
) {
785 case PAGEREF_ACTIVATE
:
786 goto activate_locked
;
789 case PAGEREF_RECLAIM
:
790 case PAGEREF_RECLAIM_CLEAN
:
791 ; /* try to reclaim the page below */
795 * Anonymous process memory has backing store?
796 * Try to allocate it some swap space here.
798 if (PageAnon(page
) && !PageSwapCache(page
)) {
799 if (!(sc
->gfp_mask
& __GFP_IO
))
801 if (!add_to_swap(page
))
802 goto activate_locked
;
806 mapping
= page_mapping(page
);
809 * The page is mapped into the page tables of one or more
810 * processes. Try to unmap it here.
812 if (page_mapped(page
) && mapping
) {
813 switch (try_to_unmap(page
, TTU_UNMAP
)) {
815 goto activate_locked
;
821 ; /* try to free the page below */
825 if (PageDirty(page
)) {
828 if (references
== PAGEREF_RECLAIM_CLEAN
)
832 if (!sc
->may_writepage
)
835 /* Page is dirty, try to write it out here */
836 switch (pageout(page
, mapping
, sc
)) {
841 goto activate_locked
;
843 if (PageWriteback(page
))
849 * A synchronous write - probably a ramdisk. Go
850 * ahead and try to reclaim the page.
852 if (!trylock_page(page
))
854 if (PageDirty(page
) || PageWriteback(page
))
856 mapping
= page_mapping(page
);
858 ; /* try to free the page below */
863 * If the page has buffers, try to free the buffer mappings
864 * associated with this page. If we succeed we try to free
867 * We do this even if the page is PageDirty().
868 * try_to_release_page() does not perform I/O, but it is
869 * possible for a page to have PageDirty set, but it is actually
870 * clean (all its buffers are clean). This happens if the
871 * buffers were written out directly, with submit_bh(). ext3
872 * will do this, as well as the blockdev mapping.
873 * try_to_release_page() will discover that cleanness and will
874 * drop the buffers and mark the page clean - it can be freed.
876 * Rarely, pages can have buffers and no ->mapping. These are
877 * the pages which were not successfully invalidated in
878 * truncate_complete_page(). We try to drop those buffers here
879 * and if that worked, and the page is no longer mapped into
880 * process address space (page_count == 1) it can be freed.
881 * Otherwise, leave the page on the LRU so it is swappable.
883 if (page_has_private(page
)) {
884 if (!try_to_release_page(page
, sc
->gfp_mask
))
885 goto activate_locked
;
886 if (!mapping
&& page_count(page
) == 1) {
888 if (put_page_testzero(page
))
892 * rare race with speculative reference.
893 * the speculative reference will free
894 * this page shortly, so we may
895 * increment nr_reclaimed here (and
896 * leave it off the LRU).
904 if (!mapping
|| !__remove_mapping(mapping
, page
))
908 * At this point, we have no other references and there is
909 * no way to pick any more up (removed from LRU, removed
910 * from pagecache). Can use non-atomic bitops now (and
911 * we obviously don't have to worry about waking up a process
912 * waiting on the page lock, because there are no references.
914 __clear_page_locked(page
);
919 * Is there need to periodically free_page_list? It would
920 * appear not as the counts should be low
922 list_add(&page
->lru
, &free_pages
);
926 if (PageSwapCache(page
))
927 try_to_free_swap(page
);
929 putback_lru_page(page
);
930 reset_reclaim_mode(sc
);
934 /* Not a candidate for swapping, so reclaim swap space. */
935 if (PageSwapCache(page
) && vm_swap_full())
936 try_to_free_swap(page
);
937 VM_BUG_ON(PageActive(page
));
943 reset_reclaim_mode(sc
);
945 list_add(&page
->lru
, &ret_pages
);
946 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
950 * Tag a zone as congested if all the dirty pages encountered were
951 * backed by a congested BDI. In this case, reclaimers should just
952 * back off and wait for congestion to clear because further reclaim
953 * will encounter the same problem
955 if (nr_dirty
&& nr_dirty
== nr_congested
&& scanning_global_lru(sc
))
956 zone_set_flag(zone
, ZONE_CONGESTED
);
958 free_page_list(&free_pages
);
960 list_splice(&ret_pages
, page_list
);
961 count_vm_events(PGACTIVATE
, pgactivate
);
966 * Attempt to remove the specified page from its LRU. Only take this page
967 * if it is of the appropriate PageActive status. Pages which are being
968 * freed elsewhere are also ignored.
970 * page: page to consider
971 * mode: one of the LRU isolation modes defined above
973 * returns 0 on success, -ve errno on failure.
975 int __isolate_lru_page(struct page
*page
, int mode
, int file
)
979 /* Only take pages on the LRU. */
984 * When checking the active state, we need to be sure we are
985 * dealing with comparible boolean values. Take the logical not
988 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
991 if (mode
!= ISOLATE_BOTH
&& page_is_file_cache(page
) != file
)
995 * When this function is being called for lumpy reclaim, we
996 * initially look into all LRU pages, active, inactive and
997 * unevictable; only give shrink_page_list evictable pages.
999 if (PageUnevictable(page
))
1004 if (likely(get_page_unless_zero(page
))) {
1006 * Be careful not to clear PageLRU until after we're
1007 * sure the page is not being freed elsewhere -- the
1008 * page release code relies on it.
1018 * zone->lru_lock is heavily contended. Some of the functions that
1019 * shrink the lists perform better by taking out a batch of pages
1020 * and working on them outside the LRU lock.
1022 * For pagecache intensive workloads, this function is the hottest
1023 * spot in the kernel (apart from copy_*_user functions).
1025 * Appropriate locks must be held before calling this function.
1027 * @nr_to_scan: The number of pages to look through on the list.
1028 * @src: The LRU list to pull pages off.
1029 * @dst: The temp list to put pages on to.
1030 * @scanned: The number of pages that were scanned.
1031 * @order: The caller's attempted allocation order
1032 * @mode: One of the LRU isolation modes
1033 * @file: True [1] if isolating file [!anon] pages
1035 * returns how many pages were moved onto *@dst.
1037 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1038 struct list_head
*src
, struct list_head
*dst
,
1039 unsigned long *scanned
, int order
, int mode
, int file
)
1041 unsigned long nr_taken
= 0;
1042 unsigned long nr_lumpy_taken
= 0;
1043 unsigned long nr_lumpy_dirty
= 0;
1044 unsigned long nr_lumpy_failed
= 0;
1047 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1050 unsigned long end_pfn
;
1051 unsigned long page_pfn
;
1054 page
= lru_to_page(src
);
1055 prefetchw_prev_lru_page(page
, src
, flags
);
1057 VM_BUG_ON(!PageLRU(page
));
1059 switch (__isolate_lru_page(page
, mode
, file
)) {
1061 list_move(&page
->lru
, dst
);
1062 mem_cgroup_del_lru(page
);
1063 nr_taken
+= hpage_nr_pages(page
);
1067 /* else it is being freed elsewhere */
1068 list_move(&page
->lru
, src
);
1069 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
1080 * Attempt to take all pages in the order aligned region
1081 * surrounding the tag page. Only take those pages of
1082 * the same active state as that tag page. We may safely
1083 * round the target page pfn down to the requested order
1084 * as the mem_map is guaranteed valid out to MAX_ORDER,
1085 * where that page is in a different zone we will detect
1086 * it from its zone id and abort this block scan.
1088 zone_id
= page_zone_id(page
);
1089 page_pfn
= page_to_pfn(page
);
1090 pfn
= page_pfn
& ~((1 << order
) - 1);
1091 end_pfn
= pfn
+ (1 << order
);
1092 for (; pfn
< end_pfn
; pfn
++) {
1093 struct page
*cursor_page
;
1095 /* The target page is in the block, ignore it. */
1096 if (unlikely(pfn
== page_pfn
))
1099 /* Avoid holes within the zone. */
1100 if (unlikely(!pfn_valid_within(pfn
)))
1103 cursor_page
= pfn_to_page(pfn
);
1105 /* Check that we have not crossed a zone boundary. */
1106 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
1110 * If we don't have enough swap space, reclaiming of
1111 * anon page which don't already have a swap slot is
1114 if (nr_swap_pages
<= 0 && PageAnon(cursor_page
) &&
1115 !PageSwapCache(cursor_page
))
1118 if (__isolate_lru_page(cursor_page
, mode
, file
) == 0) {
1119 list_move(&cursor_page
->lru
, dst
);
1120 mem_cgroup_del_lru(cursor_page
);
1121 nr_taken
+= hpage_nr_pages(page
);
1123 if (PageDirty(cursor_page
))
1127 /* the page is freed already. */
1128 if (!page_count(cursor_page
))
1134 /* If we break out of the loop above, lumpy reclaim failed */
1141 trace_mm_vmscan_lru_isolate(order
,
1144 nr_lumpy_taken
, nr_lumpy_dirty
, nr_lumpy_failed
,
1149 static unsigned long isolate_pages_global(unsigned long nr
,
1150 struct list_head
*dst
,
1151 unsigned long *scanned
, int order
,
1152 int mode
, struct zone
*z
,
1153 int active
, int file
)
1160 return isolate_lru_pages(nr
, &z
->lru
[lru
].list
, dst
, scanned
, order
,
1165 * clear_active_flags() is a helper for shrink_active_list(), clearing
1166 * any active bits from the pages in the list.
1168 static unsigned long clear_active_flags(struct list_head
*page_list
,
1169 unsigned int *count
)
1175 list_for_each_entry(page
, page_list
, lru
) {
1176 int numpages
= hpage_nr_pages(page
);
1177 lru
= page_lru_base_type(page
);
1178 if (PageActive(page
)) {
1180 ClearPageActive(page
);
1181 nr_active
+= numpages
;
1184 count
[lru
] += numpages
;
1191 * isolate_lru_page - tries to isolate a page from its LRU list
1192 * @page: page to isolate from its LRU list
1194 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1195 * vmstat statistic corresponding to whatever LRU list the page was on.
1197 * Returns 0 if the page was removed from an LRU list.
1198 * Returns -EBUSY if the page was not on an LRU list.
1200 * The returned page will have PageLRU() cleared. If it was found on
1201 * the active list, it will have PageActive set. If it was found on
1202 * the unevictable list, it will have the PageUnevictable bit set. That flag
1203 * may need to be cleared by the caller before letting the page go.
1205 * The vmstat statistic corresponding to the list on which the page was
1206 * found will be decremented.
1209 * (1) Must be called with an elevated refcount on the page. This is a
1210 * fundamentnal difference from isolate_lru_pages (which is called
1211 * without a stable reference).
1212 * (2) the lru_lock must not be held.
1213 * (3) interrupts must be enabled.
1215 int isolate_lru_page(struct page
*page
)
1219 VM_BUG_ON(!page_count(page
));
1221 if (PageLRU(page
)) {
1222 struct zone
*zone
= page_zone(page
);
1224 spin_lock_irq(&zone
->lru_lock
);
1225 if (PageLRU(page
)) {
1226 int lru
= page_lru(page
);
1231 del_page_from_lru_list(zone
, page
, lru
);
1233 spin_unlock_irq(&zone
->lru_lock
);
1239 * Are there way too many processes in the direct reclaim path already?
1241 static int too_many_isolated(struct zone
*zone
, int file
,
1242 struct scan_control
*sc
)
1244 unsigned long inactive
, isolated
;
1246 if (current_is_kswapd())
1249 if (!scanning_global_lru(sc
))
1253 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1254 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1256 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1257 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1260 return isolated
> inactive
;
1264 * TODO: Try merging with migrations version of putback_lru_pages
1266 static noinline_for_stack
void
1267 putback_lru_pages(struct zone
*zone
, struct scan_control
*sc
,
1268 unsigned long nr_anon
, unsigned long nr_file
,
1269 struct list_head
*page_list
)
1272 struct pagevec pvec
;
1273 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1275 pagevec_init(&pvec
, 1);
1278 * Put back any unfreeable pages.
1280 spin_lock(&zone
->lru_lock
);
1281 while (!list_empty(page_list
)) {
1283 page
= lru_to_page(page_list
);
1284 VM_BUG_ON(PageLRU(page
));
1285 list_del(&page
->lru
);
1286 if (unlikely(!page_evictable(page
, NULL
))) {
1287 spin_unlock_irq(&zone
->lru_lock
);
1288 putback_lru_page(page
);
1289 spin_lock_irq(&zone
->lru_lock
);
1293 lru
= page_lru(page
);
1294 add_page_to_lru_list(zone
, page
, lru
);
1295 if (is_active_lru(lru
)) {
1296 int file
= is_file_lru(lru
);
1297 int numpages
= hpage_nr_pages(page
);
1298 reclaim_stat
->recent_rotated
[file
] += numpages
;
1300 if (!pagevec_add(&pvec
, page
)) {
1301 spin_unlock_irq(&zone
->lru_lock
);
1302 __pagevec_release(&pvec
);
1303 spin_lock_irq(&zone
->lru_lock
);
1306 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1307 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1309 spin_unlock_irq(&zone
->lru_lock
);
1310 pagevec_release(&pvec
);
1313 static noinline_for_stack
void update_isolated_counts(struct zone
*zone
,
1314 struct scan_control
*sc
,
1315 unsigned long *nr_anon
,
1316 unsigned long *nr_file
,
1317 struct list_head
*isolated_list
)
1319 unsigned long nr_active
;
1320 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1321 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1323 nr_active
= clear_active_flags(isolated_list
, count
);
1324 __count_vm_events(PGDEACTIVATE
, nr_active
);
1326 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1327 -count
[LRU_ACTIVE_FILE
]);
1328 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1329 -count
[LRU_INACTIVE_FILE
]);
1330 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1331 -count
[LRU_ACTIVE_ANON
]);
1332 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1333 -count
[LRU_INACTIVE_ANON
]);
1335 *nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1336 *nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1337 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, *nr_anon
);
1338 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, *nr_file
);
1340 reclaim_stat
->recent_scanned
[0] += *nr_anon
;
1341 reclaim_stat
->recent_scanned
[1] += *nr_file
;
1345 * Returns true if the caller should wait to clean dirty/writeback pages.
1347 * If we are direct reclaiming for contiguous pages and we do not reclaim
1348 * everything in the list, try again and wait for writeback IO to complete.
1349 * This will stall high-order allocations noticeably. Only do that when really
1350 * need to free the pages under high memory pressure.
1352 static inline bool should_reclaim_stall(unsigned long nr_taken
,
1353 unsigned long nr_freed
,
1355 struct scan_control
*sc
)
1357 int lumpy_stall_priority
;
1359 /* kswapd should not stall on sync IO */
1360 if (current_is_kswapd())
1363 /* Only stall on lumpy reclaim */
1364 if (sc
->reclaim_mode
& RECLAIM_MODE_SINGLE
)
1367 /* If we have relaimed everything on the isolated list, no stall */
1368 if (nr_freed
== nr_taken
)
1372 * For high-order allocations, there are two stall thresholds.
1373 * High-cost allocations stall immediately where as lower
1374 * order allocations such as stacks require the scanning
1375 * priority to be much higher before stalling.
1377 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
1378 lumpy_stall_priority
= DEF_PRIORITY
;
1380 lumpy_stall_priority
= DEF_PRIORITY
/ 3;
1382 return priority
<= lumpy_stall_priority
;
1386 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1387 * of reclaimed pages
1389 static noinline_for_stack
unsigned long
1390 shrink_inactive_list(unsigned long nr_to_scan
, struct zone
*zone
,
1391 struct scan_control
*sc
, int priority
, int file
)
1393 LIST_HEAD(page_list
);
1394 unsigned long nr_scanned
;
1395 unsigned long nr_reclaimed
= 0;
1396 unsigned long nr_taken
;
1397 unsigned long nr_anon
;
1398 unsigned long nr_file
;
1400 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1401 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1403 /* We are about to die and free our memory. Return now. */
1404 if (fatal_signal_pending(current
))
1405 return SWAP_CLUSTER_MAX
;
1408 set_reclaim_mode(priority
, sc
, false);
1410 spin_lock_irq(&zone
->lru_lock
);
1412 if (scanning_global_lru(sc
)) {
1413 nr_taken
= isolate_pages_global(nr_to_scan
,
1414 &page_list
, &nr_scanned
, sc
->order
,
1415 sc
->reclaim_mode
& RECLAIM_MODE_LUMPYRECLAIM
?
1416 ISOLATE_BOTH
: ISOLATE_INACTIVE
,
1418 zone
->pages_scanned
+= nr_scanned
;
1419 if (current_is_kswapd())
1420 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1423 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1426 nr_taken
= mem_cgroup_isolate_pages(nr_to_scan
,
1427 &page_list
, &nr_scanned
, sc
->order
,
1428 sc
->reclaim_mode
& RECLAIM_MODE_LUMPYRECLAIM
?
1429 ISOLATE_BOTH
: ISOLATE_INACTIVE
,
1430 zone
, sc
->mem_cgroup
,
1433 * mem_cgroup_isolate_pages() keeps track of
1434 * scanned pages on its own.
1438 if (nr_taken
== 0) {
1439 spin_unlock_irq(&zone
->lru_lock
);
1443 update_isolated_counts(zone
, sc
, &nr_anon
, &nr_file
, &page_list
);
1445 spin_unlock_irq(&zone
->lru_lock
);
1447 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
);
1449 /* Check if we should syncronously wait for writeback */
1450 if (should_reclaim_stall(nr_taken
, nr_reclaimed
, priority
, sc
)) {
1451 set_reclaim_mode(priority
, sc
, true);
1452 nr_reclaimed
+= shrink_page_list(&page_list
, zone
, sc
);
1455 local_irq_disable();
1456 if (current_is_kswapd())
1457 __count_vm_events(KSWAPD_STEAL
, nr_reclaimed
);
1458 __count_zone_vm_events(PGSTEAL
, zone
, nr_reclaimed
);
1460 putback_lru_pages(zone
, sc
, nr_anon
, nr_file
, &page_list
);
1462 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1464 nr_scanned
, nr_reclaimed
,
1466 trace_shrink_flags(file
, sc
->reclaim_mode
));
1467 return nr_reclaimed
;
1471 * This moves pages from the active list to the inactive list.
1473 * We move them the other way if the page is referenced by one or more
1474 * processes, from rmap.
1476 * If the pages are mostly unmapped, the processing is fast and it is
1477 * appropriate to hold zone->lru_lock across the whole operation. But if
1478 * the pages are mapped, the processing is slow (page_referenced()) so we
1479 * should drop zone->lru_lock around each page. It's impossible to balance
1480 * this, so instead we remove the pages from the LRU while processing them.
1481 * It is safe to rely on PG_active against the non-LRU pages in here because
1482 * nobody will play with that bit on a non-LRU page.
1484 * The downside is that we have to touch page->_count against each page.
1485 * But we had to alter page->flags anyway.
1488 static void move_active_pages_to_lru(struct zone
*zone
,
1489 struct list_head
*list
,
1492 unsigned long pgmoved
= 0;
1493 struct pagevec pvec
;
1496 pagevec_init(&pvec
, 1);
1498 while (!list_empty(list
)) {
1499 page
= lru_to_page(list
);
1501 VM_BUG_ON(PageLRU(page
));
1504 list_move(&page
->lru
, &zone
->lru
[lru
].list
);
1505 mem_cgroup_add_lru_list(page
, lru
);
1506 pgmoved
+= hpage_nr_pages(page
);
1508 if (!pagevec_add(&pvec
, page
) || list_empty(list
)) {
1509 spin_unlock_irq(&zone
->lru_lock
);
1510 if (buffer_heads_over_limit
)
1511 pagevec_strip(&pvec
);
1512 __pagevec_release(&pvec
);
1513 spin_lock_irq(&zone
->lru_lock
);
1516 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1517 if (!is_active_lru(lru
))
1518 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1521 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1522 struct scan_control
*sc
, int priority
, int file
)
1524 unsigned long nr_taken
;
1525 unsigned long pgscanned
;
1526 unsigned long vm_flags
;
1527 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1528 LIST_HEAD(l_active
);
1529 LIST_HEAD(l_inactive
);
1531 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1532 unsigned long nr_rotated
= 0;
1535 spin_lock_irq(&zone
->lru_lock
);
1536 if (scanning_global_lru(sc
)) {
1537 nr_taken
= isolate_pages_global(nr_pages
, &l_hold
,
1538 &pgscanned
, sc
->order
,
1539 ISOLATE_ACTIVE
, zone
,
1541 zone
->pages_scanned
+= pgscanned
;
1543 nr_taken
= mem_cgroup_isolate_pages(nr_pages
, &l_hold
,
1544 &pgscanned
, sc
->order
,
1545 ISOLATE_ACTIVE
, zone
,
1546 sc
->mem_cgroup
, 1, file
);
1548 * mem_cgroup_isolate_pages() keeps track of
1549 * scanned pages on its own.
1553 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1555 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1557 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1559 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1560 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1561 spin_unlock_irq(&zone
->lru_lock
);
1563 while (!list_empty(&l_hold
)) {
1565 page
= lru_to_page(&l_hold
);
1566 list_del(&page
->lru
);
1568 if (unlikely(!page_evictable(page
, NULL
))) {
1569 putback_lru_page(page
);
1573 if (page_referenced(page
, 0, sc
->mem_cgroup
, &vm_flags
)) {
1574 nr_rotated
+= hpage_nr_pages(page
);
1576 * Identify referenced, file-backed active pages and
1577 * give them one more trip around the active list. So
1578 * that executable code get better chances to stay in
1579 * memory under moderate memory pressure. Anon pages
1580 * are not likely to be evicted by use-once streaming
1581 * IO, plus JVM can create lots of anon VM_EXEC pages,
1582 * so we ignore them here.
1584 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1585 list_add(&page
->lru
, &l_active
);
1590 ClearPageActive(page
); /* we are de-activating */
1591 list_add(&page
->lru
, &l_inactive
);
1595 * Move pages back to the lru list.
1597 spin_lock_irq(&zone
->lru_lock
);
1599 * Count referenced pages from currently used mappings as rotated,
1600 * even though only some of them are actually re-activated. This
1601 * helps balance scan pressure between file and anonymous pages in
1604 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1606 move_active_pages_to_lru(zone
, &l_active
,
1607 LRU_ACTIVE
+ file
* LRU_FILE
);
1608 move_active_pages_to_lru(zone
, &l_inactive
,
1609 LRU_BASE
+ file
* LRU_FILE
);
1610 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1611 spin_unlock_irq(&zone
->lru_lock
);
1615 static int inactive_anon_is_low_global(struct zone
*zone
)
1617 unsigned long active
, inactive
;
1619 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1620 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1622 if (inactive
* zone
->inactive_ratio
< active
)
1629 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1630 * @zone: zone to check
1631 * @sc: scan control of this context
1633 * Returns true if the zone does not have enough inactive anon pages,
1634 * meaning some active anon pages need to be deactivated.
1636 static int inactive_anon_is_low(struct zone
*zone
, struct scan_control
*sc
)
1641 * If we don't have swap space, anonymous page deactivation
1644 if (!total_swap_pages
)
1647 if (scanning_global_lru(sc
))
1648 low
= inactive_anon_is_low_global(zone
);
1650 low
= mem_cgroup_inactive_anon_is_low(sc
->mem_cgroup
);
1654 static inline int inactive_anon_is_low(struct zone
*zone
,
1655 struct scan_control
*sc
)
1661 static int inactive_file_is_low_global(struct zone
*zone
)
1663 unsigned long active
, inactive
;
1665 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1666 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1668 return (active
> inactive
);
1672 * inactive_file_is_low - check if file pages need to be deactivated
1673 * @zone: zone to check
1674 * @sc: scan control of this context
1676 * When the system is doing streaming IO, memory pressure here
1677 * ensures that active file pages get deactivated, until more
1678 * than half of the file pages are on the inactive list.
1680 * Once we get to that situation, protect the system's working
1681 * set from being evicted by disabling active file page aging.
1683 * This uses a different ratio than the anonymous pages, because
1684 * the page cache uses a use-once replacement algorithm.
1686 static int inactive_file_is_low(struct zone
*zone
, struct scan_control
*sc
)
1690 if (scanning_global_lru(sc
))
1691 low
= inactive_file_is_low_global(zone
);
1693 low
= mem_cgroup_inactive_file_is_low(sc
->mem_cgroup
);
1697 static int inactive_list_is_low(struct zone
*zone
, struct scan_control
*sc
,
1701 return inactive_file_is_low(zone
, sc
);
1703 return inactive_anon_is_low(zone
, sc
);
1706 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1707 struct zone
*zone
, struct scan_control
*sc
, int priority
)
1709 int file
= is_file_lru(lru
);
1711 if (is_active_lru(lru
)) {
1712 if (inactive_list_is_low(zone
, sc
, file
))
1713 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1717 return shrink_inactive_list(nr_to_scan
, zone
, sc
, priority
, file
);
1721 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1722 * until we collected @swap_cluster_max pages to scan.
1724 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan
,
1725 unsigned long *nr_saved_scan
)
1729 *nr_saved_scan
+= nr_to_scan
;
1730 nr
= *nr_saved_scan
;
1732 if (nr
>= SWAP_CLUSTER_MAX
)
1741 * Determine how aggressively the anon and file LRU lists should be
1742 * scanned. The relative value of each set of LRU lists is determined
1743 * by looking at the fraction of the pages scanned we did rotate back
1744 * onto the active list instead of evict.
1746 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1748 static void get_scan_count(struct zone
*zone
, struct scan_control
*sc
,
1749 unsigned long *nr
, int priority
)
1751 unsigned long anon
, file
, free
;
1752 unsigned long anon_prio
, file_prio
;
1753 unsigned long ap
, fp
;
1754 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1755 u64 fraction
[2], denominator
;
1759 /* If we have no swap space, do not bother scanning anon pages. */
1760 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1768 anon
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_ANON
) +
1769 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_ANON
);
1770 file
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_FILE
) +
1771 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1773 if (scanning_global_lru(sc
)) {
1774 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1775 /* If we have very few page cache pages,
1776 force-scan anon pages. */
1777 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1786 * With swappiness at 100, anonymous and file have the same priority.
1787 * This scanning priority is essentially the inverse of IO cost.
1789 anon_prio
= sc
->swappiness
;
1790 file_prio
= 200 - sc
->swappiness
;
1793 * OK, so we have swap space and a fair amount of page cache
1794 * pages. We use the recently rotated / recently scanned
1795 * ratios to determine how valuable each cache is.
1797 * Because workloads change over time (and to avoid overflow)
1798 * we keep these statistics as a floating average, which ends
1799 * up weighing recent references more than old ones.
1801 * anon in [0], file in [1]
1803 spin_lock_irq(&zone
->lru_lock
);
1804 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1805 reclaim_stat
->recent_scanned
[0] /= 2;
1806 reclaim_stat
->recent_rotated
[0] /= 2;
1809 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1810 reclaim_stat
->recent_scanned
[1] /= 2;
1811 reclaim_stat
->recent_rotated
[1] /= 2;
1815 * The amount of pressure on anon vs file pages is inversely
1816 * proportional to the fraction of recently scanned pages on
1817 * each list that were recently referenced and in active use.
1819 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1820 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1822 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1823 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1824 spin_unlock_irq(&zone
->lru_lock
);
1828 denominator
= ap
+ fp
+ 1;
1830 for_each_evictable_lru(l
) {
1831 int file
= is_file_lru(l
);
1834 scan
= zone_nr_lru_pages(zone
, sc
, l
);
1835 if (priority
|| noswap
) {
1837 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1839 nr
[l
] = nr_scan_try_batch(scan
,
1840 &reclaim_stat
->nr_saved_scan
[l
]);
1845 * Reclaim/compaction depends on a number of pages being freed. To avoid
1846 * disruption to the system, a small number of order-0 pages continue to be
1847 * rotated and reclaimed in the normal fashion. However, by the time we get
1848 * back to the allocator and call try_to_compact_zone(), we ensure that
1849 * there are enough free pages for it to be likely successful
1851 static inline bool should_continue_reclaim(struct zone
*zone
,
1852 unsigned long nr_reclaimed
,
1853 unsigned long nr_scanned
,
1854 struct scan_control
*sc
)
1856 unsigned long pages_for_compaction
;
1857 unsigned long inactive_lru_pages
;
1859 /* If not in reclaim/compaction mode, stop */
1860 if (!(sc
->reclaim_mode
& RECLAIM_MODE_COMPACTION
))
1863 /* Consider stopping depending on scan and reclaim activity */
1864 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1866 * For __GFP_REPEAT allocations, stop reclaiming if the
1867 * full LRU list has been scanned and we are still failing
1868 * to reclaim pages. This full LRU scan is potentially
1869 * expensive but a __GFP_REPEAT caller really wants to succeed
1871 if (!nr_reclaimed
&& !nr_scanned
)
1875 * For non-__GFP_REPEAT allocations which can presumably
1876 * fail without consequence, stop if we failed to reclaim
1877 * any pages from the last SWAP_CLUSTER_MAX number of
1878 * pages that were scanned. This will return to the
1879 * caller faster at the risk reclaim/compaction and
1880 * the resulting allocation attempt fails
1887 * If we have not reclaimed enough pages for compaction and the
1888 * inactive lists are large enough, continue reclaiming
1890 pages_for_compaction
= (2UL << sc
->order
);
1891 inactive_lru_pages
= zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_ANON
) +
1892 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1893 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1894 inactive_lru_pages
> pages_for_compaction
)
1897 /* If compaction would go ahead or the allocation would succeed, stop */
1898 switch (compaction_suitable(zone
, sc
->order
)) {
1899 case COMPACT_PARTIAL
:
1900 case COMPACT_CONTINUE
:
1908 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1910 static void shrink_zone(int priority
, struct zone
*zone
,
1911 struct scan_control
*sc
)
1913 unsigned long nr
[NR_LRU_LISTS
];
1914 unsigned long nr_to_scan
;
1916 unsigned long nr_reclaimed
, nr_scanned
;
1917 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1921 nr_scanned
= sc
->nr_scanned
;
1922 get_scan_count(zone
, sc
, nr
, priority
);
1924 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1925 nr
[LRU_INACTIVE_FILE
]) {
1926 for_each_evictable_lru(l
) {
1928 nr_to_scan
= min_t(unsigned long,
1929 nr
[l
], SWAP_CLUSTER_MAX
);
1930 nr
[l
] -= nr_to_scan
;
1932 nr_reclaimed
+= shrink_list(l
, nr_to_scan
,
1933 zone
, sc
, priority
);
1937 * On large memory systems, scan >> priority can become
1938 * really large. This is fine for the starting priority;
1939 * we want to put equal scanning pressure on each zone.
1940 * However, if the VM has a harder time of freeing pages,
1941 * with multiple processes reclaiming pages, the total
1942 * freeing target can get unreasonably large.
1944 if (nr_reclaimed
>= nr_to_reclaim
&& priority
< DEF_PRIORITY
)
1947 sc
->nr_reclaimed
+= nr_reclaimed
;
1950 * Even if we did not try to evict anon pages at all, we want to
1951 * rebalance the anon lru active/inactive ratio.
1953 if (inactive_anon_is_low(zone
, sc
))
1954 shrink_active_list(SWAP_CLUSTER_MAX
, zone
, sc
, priority
, 0);
1956 /* reclaim/compaction might need reclaim to continue */
1957 if (should_continue_reclaim(zone
, nr_reclaimed
,
1958 sc
->nr_scanned
- nr_scanned
, sc
))
1961 throttle_vm_writeout(sc
->gfp_mask
);
1965 * This is the direct reclaim path, for page-allocating processes. We only
1966 * try to reclaim pages from zones which will satisfy the caller's allocation
1969 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1971 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1973 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1974 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1975 * zone defense algorithm.
1977 * If a zone is deemed to be full of pinned pages then just give it a light
1978 * scan then give up on it.
1980 static unsigned long shrink_zones(int priority
, struct zonelist
*zonelist
,
1981 struct scan_control
*sc
)
1985 unsigned long nr_soft_reclaimed
;
1986 unsigned long nr_soft_scanned
;
1987 unsigned long total_scanned
= 0;
1989 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1990 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
1991 if (!populated_zone(zone
))
1994 * Take care memory controller reclaiming has small influence
1997 if (scanning_global_lru(sc
)) {
1998 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2000 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2001 continue; /* Let kswapd poll it */
2004 nr_soft_scanned
= 0;
2005 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2006 sc
->order
, sc
->gfp_mask
,
2008 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2009 total_scanned
+= nr_soft_scanned
;
2011 shrink_zone(priority
, zone
, sc
);
2014 return total_scanned
;
2017 static bool zone_reclaimable(struct zone
*zone
)
2019 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2022 /* All zones in zonelist are unreclaimable? */
2023 static bool all_unreclaimable(struct zonelist
*zonelist
,
2024 struct scan_control
*sc
)
2029 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2030 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2031 if (!populated_zone(zone
))
2033 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2035 if (!zone
->all_unreclaimable
)
2043 * This is the main entry point to direct page reclaim.
2045 * If a full scan of the inactive list fails to free enough memory then we
2046 * are "out of memory" and something needs to be killed.
2048 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2049 * high - the zone may be full of dirty or under-writeback pages, which this
2050 * caller can't do much about. We kick the writeback threads and take explicit
2051 * naps in the hope that some of these pages can be written. But if the
2052 * allocating task holds filesystem locks which prevent writeout this might not
2053 * work, and the allocation attempt will fail.
2055 * returns: 0, if no pages reclaimed
2056 * else, the number of pages reclaimed
2058 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2059 struct scan_control
*sc
,
2060 struct shrink_control
*shrink
)
2063 unsigned long total_scanned
= 0;
2064 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2067 unsigned long writeback_threshold
;
2070 delayacct_freepages_start();
2072 if (scanning_global_lru(sc
))
2073 count_vm_event(ALLOCSTALL
);
2075 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2078 disable_swap_token();
2079 total_scanned
+= shrink_zones(priority
, zonelist
, sc
);
2081 * Don't shrink slabs when reclaiming memory from
2082 * over limit cgroups
2084 if (scanning_global_lru(sc
)) {
2085 unsigned long lru_pages
= 0;
2086 for_each_zone_zonelist(zone
, z
, zonelist
,
2087 gfp_zone(sc
->gfp_mask
)) {
2088 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2091 lru_pages
+= zone_reclaimable_pages(zone
);
2094 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2095 if (reclaim_state
) {
2096 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2097 reclaim_state
->reclaimed_slab
= 0;
2100 total_scanned
+= sc
->nr_scanned
;
2101 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2105 * Try to write back as many pages as we just scanned. This
2106 * tends to cause slow streaming writers to write data to the
2107 * disk smoothly, at the dirtying rate, which is nice. But
2108 * that's undesirable in laptop mode, where we *want* lumpy
2109 * writeout. So in laptop mode, write out the whole world.
2111 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2112 if (total_scanned
> writeback_threshold
) {
2113 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
);
2114 sc
->may_writepage
= 1;
2117 /* Take a nap, wait for some writeback to complete */
2118 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2119 priority
< DEF_PRIORITY
- 2) {
2120 struct zone
*preferred_zone
;
2122 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2123 &cpuset_current_mems_allowed
,
2125 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2130 delayacct_freepages_end();
2133 if (sc
->nr_reclaimed
)
2134 return sc
->nr_reclaimed
;
2137 * As hibernation is going on, kswapd is freezed so that it can't mark
2138 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2141 if (oom_killer_disabled
)
2144 /* top priority shrink_zones still had more to do? don't OOM, then */
2145 if (scanning_global_lru(sc
) && !all_unreclaimable(zonelist
, sc
))
2151 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2152 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2154 unsigned long nr_reclaimed
;
2155 struct scan_control sc
= {
2156 .gfp_mask
= gfp_mask
,
2157 .may_writepage
= !laptop_mode
,
2158 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2161 .swappiness
= vm_swappiness
,
2164 .nodemask
= nodemask
,
2166 struct shrink_control shrink
= {
2167 .gfp_mask
= sc
.gfp_mask
,
2170 trace_mm_vmscan_direct_reclaim_begin(order
,
2174 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2176 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2178 return nr_reclaimed
;
2181 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2183 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*mem
,
2184 gfp_t gfp_mask
, bool noswap
,
2185 unsigned int swappiness
,
2187 unsigned long *nr_scanned
)
2189 struct scan_control sc
= {
2191 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2192 .may_writepage
= !laptop_mode
,
2194 .may_swap
= !noswap
,
2195 .swappiness
= swappiness
,
2200 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2201 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2203 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2208 * NOTE: Although we can get the priority field, using it
2209 * here is not a good idea, since it limits the pages we can scan.
2210 * if we don't reclaim here, the shrink_zone from balance_pgdat
2211 * will pick up pages from other mem cgroup's as well. We hack
2212 * the priority and make it zero.
2214 shrink_zone(0, zone
, &sc
);
2216 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2218 *nr_scanned
= sc
.nr_scanned
;
2219 return sc
.nr_reclaimed
;
2222 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*mem_cont
,
2225 unsigned int swappiness
)
2227 struct zonelist
*zonelist
;
2228 unsigned long nr_reclaimed
;
2229 struct scan_control sc
= {
2230 .may_writepage
= !laptop_mode
,
2232 .may_swap
= !noswap
,
2233 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2234 .swappiness
= swappiness
,
2236 .mem_cgroup
= mem_cont
,
2237 .nodemask
= NULL
, /* we don't care the placement */
2238 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2239 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2241 struct shrink_control shrink
= {
2242 .gfp_mask
= sc
.gfp_mask
,
2245 zonelist
= NODE_DATA(numa_node_id())->node_zonelists
;
2247 trace_mm_vmscan_memcg_reclaim_begin(0,
2251 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2253 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2255 return nr_reclaimed
;
2260 * pgdat_balanced is used when checking if a node is balanced for high-order
2261 * allocations. Only zones that meet watermarks and are in a zone allowed
2262 * by the callers classzone_idx are added to balanced_pages. The total of
2263 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2264 * for the node to be considered balanced. Forcing all zones to be balanced
2265 * for high orders can cause excessive reclaim when there are imbalanced zones.
2266 * The choice of 25% is due to
2267 * o a 16M DMA zone that is balanced will not balance a zone on any
2268 * reasonable sized machine
2269 * o On all other machines, the top zone must be at least a reasonable
2270 * percentage of the middle zones. For example, on 32-bit x86, highmem
2271 * would need to be at least 256M for it to be balance a whole node.
2272 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2273 * to balance a node on its own. These seemed like reasonable ratios.
2275 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2278 unsigned long present_pages
= 0;
2281 for (i
= 0; i
<= classzone_idx
; i
++)
2282 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2284 return balanced_pages
> (present_pages
>> 2);
2287 /* is kswapd sleeping prematurely? */
2288 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2292 unsigned long balanced
= 0;
2293 bool all_zones_ok
= true;
2295 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2299 /* Check the watermark levels */
2300 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
2301 struct zone
*zone
= pgdat
->node_zones
+ i
;
2303 if (!populated_zone(zone
))
2307 * balance_pgdat() skips over all_unreclaimable after
2308 * DEF_PRIORITY. Effectively, it considers them balanced so
2309 * they must be considered balanced here as well if kswapd
2312 if (zone
->all_unreclaimable
) {
2313 balanced
+= zone
->present_pages
;
2317 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2319 all_zones_ok
= false;
2321 balanced
+= zone
->present_pages
;
2325 * For high-order requests, the balanced zones must contain at least
2326 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2330 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2332 return !all_zones_ok
;
2336 * For kswapd, balance_pgdat() will work across all this node's zones until
2337 * they are all at high_wmark_pages(zone).
2339 * Returns the final order kswapd was reclaiming at
2341 * There is special handling here for zones which are full of pinned pages.
2342 * This can happen if the pages are all mlocked, or if they are all used by
2343 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2344 * What we do is to detect the case where all pages in the zone have been
2345 * scanned twice and there has been zero successful reclaim. Mark the zone as
2346 * dead and from now on, only perform a short scan. Basically we're polling
2347 * the zone for when the problem goes away.
2349 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2350 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2351 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2352 * lower zones regardless of the number of free pages in the lower zones. This
2353 * interoperates with the page allocator fallback scheme to ensure that aging
2354 * of pages is balanced across the zones.
2356 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2360 unsigned long balanced
;
2363 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2364 unsigned long total_scanned
;
2365 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2366 unsigned long nr_soft_reclaimed
;
2367 unsigned long nr_soft_scanned
;
2368 struct scan_control sc
= {
2369 .gfp_mask
= GFP_KERNEL
,
2373 * kswapd doesn't want to be bailed out while reclaim. because
2374 * we want to put equal scanning pressure on each zone.
2376 .nr_to_reclaim
= ULONG_MAX
,
2377 .swappiness
= vm_swappiness
,
2381 struct shrink_control shrink
= {
2382 .gfp_mask
= sc
.gfp_mask
,
2386 sc
.nr_reclaimed
= 0;
2387 sc
.may_writepage
= !laptop_mode
;
2388 count_vm_event(PAGEOUTRUN
);
2390 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2391 unsigned long lru_pages
= 0;
2392 int has_under_min_watermark_zone
= 0;
2394 /* The swap token gets in the way of swapout... */
2396 disable_swap_token();
2402 * Scan in the highmem->dma direction for the highest
2403 * zone which needs scanning
2405 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2406 struct zone
*zone
= pgdat
->node_zones
+ i
;
2408 if (!populated_zone(zone
))
2411 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2415 * Do some background aging of the anon list, to give
2416 * pages a chance to be referenced before reclaiming.
2418 if (inactive_anon_is_low(zone
, &sc
))
2419 shrink_active_list(SWAP_CLUSTER_MAX
, zone
,
2422 if (!zone_watermark_ok_safe(zone
, order
,
2423 high_wmark_pages(zone
), 0, 0)) {
2432 for (i
= 0; i
<= end_zone
; i
++) {
2433 struct zone
*zone
= pgdat
->node_zones
+ i
;
2435 lru_pages
+= zone_reclaimable_pages(zone
);
2439 * Now scan the zone in the dma->highmem direction, stopping
2440 * at the last zone which needs scanning.
2442 * We do this because the page allocator works in the opposite
2443 * direction. This prevents the page allocator from allocating
2444 * pages behind kswapd's direction of progress, which would
2445 * cause too much scanning of the lower zones.
2447 for (i
= 0; i
<= end_zone
; i
++) {
2448 struct zone
*zone
= pgdat
->node_zones
+ i
;
2450 unsigned long balance_gap
;
2452 if (!populated_zone(zone
))
2455 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2460 nr_soft_scanned
= 0;
2462 * Call soft limit reclaim before calling shrink_zone.
2464 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2467 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2468 total_scanned
+= nr_soft_scanned
;
2471 * We put equal pressure on every zone, unless
2472 * one zone has way too many pages free
2473 * already. The "too many pages" is defined
2474 * as the high wmark plus a "gap" where the
2475 * gap is either the low watermark or 1%
2476 * of the zone, whichever is smaller.
2478 balance_gap
= min(low_wmark_pages(zone
),
2479 (zone
->present_pages
+
2480 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2481 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2482 if (!zone_watermark_ok_safe(zone
, order
,
2483 high_wmark_pages(zone
) + balance_gap
,
2485 shrink_zone(priority
, zone
, &sc
);
2486 reclaim_state
->reclaimed_slab
= 0;
2487 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2488 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2489 total_scanned
+= sc
.nr_scanned
;
2491 if (zone
->all_unreclaimable
)
2494 !zone_reclaimable(zone
))
2495 zone
->all_unreclaimable
= 1;
2497 * If we've done a decent amount of scanning and
2498 * the reclaim ratio is low, start doing writepage
2499 * even in laptop mode
2501 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2502 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2503 sc
.may_writepage
= 1;
2505 if (!zone_watermark_ok_safe(zone
, order
,
2506 high_wmark_pages(zone
), end_zone
, 0)) {
2509 * We are still under min water mark. This
2510 * means that we have a GFP_ATOMIC allocation
2511 * failure risk. Hurry up!
2513 if (!zone_watermark_ok_safe(zone
, order
,
2514 min_wmark_pages(zone
), end_zone
, 0))
2515 has_under_min_watermark_zone
= 1;
2518 * If a zone reaches its high watermark,
2519 * consider it to be no longer congested. It's
2520 * possible there are dirty pages backed by
2521 * congested BDIs but as pressure is relieved,
2522 * spectulatively avoid congestion waits
2524 zone_clear_flag(zone
, ZONE_CONGESTED
);
2525 if (i
<= *classzone_idx
)
2526 balanced
+= zone
->present_pages
;
2530 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2531 break; /* kswapd: all done */
2533 * OK, kswapd is getting into trouble. Take a nap, then take
2534 * another pass across the zones.
2536 if (total_scanned
&& (priority
< DEF_PRIORITY
- 2)) {
2537 if (has_under_min_watermark_zone
)
2538 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2540 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2544 * We do this so kswapd doesn't build up large priorities for
2545 * example when it is freeing in parallel with allocators. It
2546 * matches the direct reclaim path behaviour in terms of impact
2547 * on zone->*_priority.
2549 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2555 * order-0: All zones must meet high watermark for a balanced node
2556 * high-order: Balanced zones must make up at least 25% of the node
2557 * for the node to be balanced
2559 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2565 * Fragmentation may mean that the system cannot be
2566 * rebalanced for high-order allocations in all zones.
2567 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2568 * it means the zones have been fully scanned and are still
2569 * not balanced. For high-order allocations, there is
2570 * little point trying all over again as kswapd may
2573 * Instead, recheck all watermarks at order-0 as they
2574 * are the most important. If watermarks are ok, kswapd will go
2575 * back to sleep. High-order users can still perform direct
2576 * reclaim if they wish.
2578 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2579 order
= sc
.order
= 0;
2585 * If kswapd was reclaiming at a higher order, it has the option of
2586 * sleeping without all zones being balanced. Before it does, it must
2587 * ensure that the watermarks for order-0 on *all* zones are met and
2588 * that the congestion flags are cleared. The congestion flag must
2589 * be cleared as kswapd is the only mechanism that clears the flag
2590 * and it is potentially going to sleep here.
2593 for (i
= 0; i
<= end_zone
; i
++) {
2594 struct zone
*zone
= pgdat
->node_zones
+ i
;
2596 if (!populated_zone(zone
))
2599 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2602 /* Confirm the zone is balanced for order-0 */
2603 if (!zone_watermark_ok(zone
, 0,
2604 high_wmark_pages(zone
), 0, 0)) {
2605 order
= sc
.order
= 0;
2609 /* If balanced, clear the congested flag */
2610 zone_clear_flag(zone
, ZONE_CONGESTED
);
2615 * Return the order we were reclaiming at so sleeping_prematurely()
2616 * makes a decision on the order we were last reclaiming at. However,
2617 * if another caller entered the allocator slow path while kswapd
2618 * was awake, order will remain at the higher level
2620 *classzone_idx
= end_zone
;
2624 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2629 if (freezing(current
) || kthread_should_stop())
2632 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2634 /* Try to sleep for a short interval */
2635 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2636 remaining
= schedule_timeout(HZ
/10);
2637 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2638 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2642 * After a short sleep, check if it was a premature sleep. If not, then
2643 * go fully to sleep until explicitly woken up.
2645 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2646 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2649 * vmstat counters are not perfectly accurate and the estimated
2650 * value for counters such as NR_FREE_PAGES can deviate from the
2651 * true value by nr_online_cpus * threshold. To avoid the zone
2652 * watermarks being breached while under pressure, we reduce the
2653 * per-cpu vmstat threshold while kswapd is awake and restore
2654 * them before going back to sleep.
2656 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2658 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2661 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2663 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2665 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2669 * The background pageout daemon, started as a kernel thread
2670 * from the init process.
2672 * This basically trickles out pages so that we have _some_
2673 * free memory available even if there is no other activity
2674 * that frees anything up. This is needed for things like routing
2675 * etc, where we otherwise might have all activity going on in
2676 * asynchronous contexts that cannot page things out.
2678 * If there are applications that are active memory-allocators
2679 * (most normal use), this basically shouldn't matter.
2681 static int kswapd(void *p
)
2683 unsigned long order
;
2685 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2686 struct task_struct
*tsk
= current
;
2688 struct reclaim_state reclaim_state
= {
2689 .reclaimed_slab
= 0,
2691 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2693 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2695 if (!cpumask_empty(cpumask
))
2696 set_cpus_allowed_ptr(tsk
, cpumask
);
2697 current
->reclaim_state
= &reclaim_state
;
2700 * Tell the memory management that we're a "memory allocator",
2701 * and that if we need more memory we should get access to it
2702 * regardless (see "__alloc_pages()"). "kswapd" should
2703 * never get caught in the normal page freeing logic.
2705 * (Kswapd normally doesn't need memory anyway, but sometimes
2706 * you need a small amount of memory in order to be able to
2707 * page out something else, and this flag essentially protects
2708 * us from recursively trying to free more memory as we're
2709 * trying to free the first piece of memory in the first place).
2711 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2715 classzone_idx
= MAX_NR_ZONES
- 1;
2717 unsigned long new_order
;
2718 int new_classzone_idx
;
2721 new_order
= pgdat
->kswapd_max_order
;
2722 new_classzone_idx
= pgdat
->classzone_idx
;
2723 pgdat
->kswapd_max_order
= 0;
2724 pgdat
->classzone_idx
= MAX_NR_ZONES
- 1;
2725 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2727 * Don't sleep if someone wants a larger 'order'
2728 * allocation or has tigher zone constraints
2731 classzone_idx
= new_classzone_idx
;
2733 kswapd_try_to_sleep(pgdat
, order
, classzone_idx
);
2734 order
= pgdat
->kswapd_max_order
;
2735 classzone_idx
= pgdat
->classzone_idx
;
2736 pgdat
->kswapd_max_order
= 0;
2737 pgdat
->classzone_idx
= MAX_NR_ZONES
- 1;
2740 ret
= try_to_freeze();
2741 if (kthread_should_stop())
2745 * We can speed up thawing tasks if we don't call balance_pgdat
2746 * after returning from the refrigerator
2749 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2750 order
= balance_pgdat(pgdat
, order
, &classzone_idx
);
2757 * A zone is low on free memory, so wake its kswapd task to service it.
2759 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2763 if (!populated_zone(zone
))
2766 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2768 pgdat
= zone
->zone_pgdat
;
2769 if (pgdat
->kswapd_max_order
< order
) {
2770 pgdat
->kswapd_max_order
= order
;
2771 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2773 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2775 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2778 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2779 wake_up_interruptible(&pgdat
->kswapd_wait
);
2783 * The reclaimable count would be mostly accurate.
2784 * The less reclaimable pages may be
2785 * - mlocked pages, which will be moved to unevictable list when encountered
2786 * - mapped pages, which may require several travels to be reclaimed
2787 * - dirty pages, which is not "instantly" reclaimable
2789 unsigned long global_reclaimable_pages(void)
2793 nr
= global_page_state(NR_ACTIVE_FILE
) +
2794 global_page_state(NR_INACTIVE_FILE
);
2796 if (nr_swap_pages
> 0)
2797 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2798 global_page_state(NR_INACTIVE_ANON
);
2803 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2807 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2808 zone_page_state(zone
, NR_INACTIVE_FILE
);
2810 if (nr_swap_pages
> 0)
2811 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2812 zone_page_state(zone
, NR_INACTIVE_ANON
);
2817 #ifdef CONFIG_HIBERNATION
2819 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2822 * Rather than trying to age LRUs the aim is to preserve the overall
2823 * LRU order by reclaiming preferentially
2824 * inactive > active > active referenced > active mapped
2826 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2828 struct reclaim_state reclaim_state
;
2829 struct scan_control sc
= {
2830 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2834 .nr_to_reclaim
= nr_to_reclaim
,
2835 .hibernation_mode
= 1,
2836 .swappiness
= vm_swappiness
,
2839 struct shrink_control shrink
= {
2840 .gfp_mask
= sc
.gfp_mask
,
2842 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2843 struct task_struct
*p
= current
;
2844 unsigned long nr_reclaimed
;
2846 p
->flags
|= PF_MEMALLOC
;
2847 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2848 reclaim_state
.reclaimed_slab
= 0;
2849 p
->reclaim_state
= &reclaim_state
;
2851 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2853 p
->reclaim_state
= NULL
;
2854 lockdep_clear_current_reclaim_state();
2855 p
->flags
&= ~PF_MEMALLOC
;
2857 return nr_reclaimed
;
2859 #endif /* CONFIG_HIBERNATION */
2861 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2862 not required for correctness. So if the last cpu in a node goes
2863 away, we get changed to run anywhere: as the first one comes back,
2864 restore their cpu bindings. */
2865 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2866 unsigned long action
, void *hcpu
)
2870 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2871 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2872 pg_data_t
*pgdat
= NODE_DATA(nid
);
2873 const struct cpumask
*mask
;
2875 mask
= cpumask_of_node(pgdat
->node_id
);
2877 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2878 /* One of our CPUs online: restore mask */
2879 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2886 * This kswapd start function will be called by init and node-hot-add.
2887 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2889 int kswapd_run(int nid
)
2891 pg_data_t
*pgdat
= NODE_DATA(nid
);
2897 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2898 if (IS_ERR(pgdat
->kswapd
)) {
2899 /* failure at boot is fatal */
2900 BUG_ON(system_state
== SYSTEM_BOOTING
);
2901 printk("Failed to start kswapd on node %d\n",nid
);
2908 * Called by memory hotplug when all memory in a node is offlined.
2910 void kswapd_stop(int nid
)
2912 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
2915 kthread_stop(kswapd
);
2918 static int __init
kswapd_init(void)
2923 for_each_node_state(nid
, N_HIGH_MEMORY
)
2925 hotcpu_notifier(cpu_callback
, 0);
2929 module_init(kswapd_init
)
2935 * If non-zero call zone_reclaim when the number of free pages falls below
2938 int zone_reclaim_mode __read_mostly
;
2940 #define RECLAIM_OFF 0
2941 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2942 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2943 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2946 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2947 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2950 #define ZONE_RECLAIM_PRIORITY 4
2953 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2956 int sysctl_min_unmapped_ratio
= 1;
2959 * If the number of slab pages in a zone grows beyond this percentage then
2960 * slab reclaim needs to occur.
2962 int sysctl_min_slab_ratio
= 5;
2964 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
2966 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
2967 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
2968 zone_page_state(zone
, NR_ACTIVE_FILE
);
2971 * It's possible for there to be more file mapped pages than
2972 * accounted for by the pages on the file LRU lists because
2973 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2975 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
2978 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2979 static long zone_pagecache_reclaimable(struct zone
*zone
)
2981 long nr_pagecache_reclaimable
;
2985 * If RECLAIM_SWAP is set, then all file pages are considered
2986 * potentially reclaimable. Otherwise, we have to worry about
2987 * pages like swapcache and zone_unmapped_file_pages() provides
2990 if (zone_reclaim_mode
& RECLAIM_SWAP
)
2991 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
2993 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
2995 /* If we can't clean pages, remove dirty pages from consideration */
2996 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
2997 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
2999 /* Watch for any possible underflows due to delta */
3000 if (unlikely(delta
> nr_pagecache_reclaimable
))
3001 delta
= nr_pagecache_reclaimable
;
3003 return nr_pagecache_reclaimable
- delta
;
3007 * Try to free up some pages from this zone through reclaim.
3009 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3011 /* Minimum pages needed in order to stay on node */
3012 const unsigned long nr_pages
= 1 << order
;
3013 struct task_struct
*p
= current
;
3014 struct reclaim_state reclaim_state
;
3016 struct scan_control sc
= {
3017 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3018 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3020 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3022 .gfp_mask
= gfp_mask
,
3023 .swappiness
= vm_swappiness
,
3026 struct shrink_control shrink
= {
3027 .gfp_mask
= sc
.gfp_mask
,
3029 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3033 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3034 * and we also need to be able to write out pages for RECLAIM_WRITE
3037 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3038 lockdep_set_current_reclaim_state(gfp_mask
);
3039 reclaim_state
.reclaimed_slab
= 0;
3040 p
->reclaim_state
= &reclaim_state
;
3042 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3044 * Free memory by calling shrink zone with increasing
3045 * priorities until we have enough memory freed.
3047 priority
= ZONE_RECLAIM_PRIORITY
;
3049 shrink_zone(priority
, zone
, &sc
);
3051 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
3054 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3055 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3057 * shrink_slab() does not currently allow us to determine how
3058 * many pages were freed in this zone. So we take the current
3059 * number of slab pages and shake the slab until it is reduced
3060 * by the same nr_pages that we used for reclaiming unmapped
3063 * Note that shrink_slab will free memory on all zones and may
3067 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3069 /* No reclaimable slab or very low memory pressure */
3070 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3073 /* Freed enough memory */
3074 nr_slab_pages1
= zone_page_state(zone
,
3075 NR_SLAB_RECLAIMABLE
);
3076 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3081 * Update nr_reclaimed by the number of slab pages we
3082 * reclaimed from this zone.
3084 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3085 if (nr_slab_pages1
< nr_slab_pages0
)
3086 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3089 p
->reclaim_state
= NULL
;
3090 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3091 lockdep_clear_current_reclaim_state();
3092 return sc
.nr_reclaimed
>= nr_pages
;
3095 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3101 * Zone reclaim reclaims unmapped file backed pages and
3102 * slab pages if we are over the defined limits.
3104 * A small portion of unmapped file backed pages is needed for
3105 * file I/O otherwise pages read by file I/O will be immediately
3106 * thrown out if the zone is overallocated. So we do not reclaim
3107 * if less than a specified percentage of the zone is used by
3108 * unmapped file backed pages.
3110 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3111 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3112 return ZONE_RECLAIM_FULL
;
3114 if (zone
->all_unreclaimable
)
3115 return ZONE_RECLAIM_FULL
;
3118 * Do not scan if the allocation should not be delayed.
3120 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3121 return ZONE_RECLAIM_NOSCAN
;
3124 * Only run zone reclaim on the local zone or on zones that do not
3125 * have associated processors. This will favor the local processor
3126 * over remote processors and spread off node memory allocations
3127 * as wide as possible.
3129 node_id
= zone_to_nid(zone
);
3130 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3131 return ZONE_RECLAIM_NOSCAN
;
3133 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3134 return ZONE_RECLAIM_NOSCAN
;
3136 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3137 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3140 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3147 * page_evictable - test whether a page is evictable
3148 * @page: the page to test
3149 * @vma: the VMA in which the page is or will be mapped, may be NULL
3151 * Test whether page is evictable--i.e., should be placed on active/inactive
3152 * lists vs unevictable list. The vma argument is !NULL when called from the
3153 * fault path to determine how to instantate a new page.
3155 * Reasons page might not be evictable:
3156 * (1) page's mapping marked unevictable
3157 * (2) page is part of an mlocked VMA
3160 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3163 if (mapping_unevictable(page_mapping(page
)))
3166 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
3173 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3174 * @page: page to check evictability and move to appropriate lru list
3175 * @zone: zone page is in
3177 * Checks a page for evictability and moves the page to the appropriate
3180 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3181 * have PageUnevictable set.
3183 static void check_move_unevictable_page(struct page
*page
, struct zone
*zone
)
3185 VM_BUG_ON(PageActive(page
));
3188 ClearPageUnevictable(page
);
3189 if (page_evictable(page
, NULL
)) {
3190 enum lru_list l
= page_lru_base_type(page
);
3192 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3193 list_move(&page
->lru
, &zone
->lru
[l
].list
);
3194 mem_cgroup_move_lists(page
, LRU_UNEVICTABLE
, l
);
3195 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ l
);
3196 __count_vm_event(UNEVICTABLE_PGRESCUED
);
3199 * rotate unevictable list
3201 SetPageUnevictable(page
);
3202 list_move(&page
->lru
, &zone
->lru
[LRU_UNEVICTABLE
].list
);
3203 mem_cgroup_rotate_lru_list(page
, LRU_UNEVICTABLE
);
3204 if (page_evictable(page
, NULL
))
3210 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3211 * @mapping: struct address_space to scan for evictable pages
3213 * Scan all pages in mapping. Check unevictable pages for
3214 * evictability and move them to the appropriate zone lru list.
3216 void scan_mapping_unevictable_pages(struct address_space
*mapping
)
3219 pgoff_t end
= (i_size_read(mapping
->host
) + PAGE_CACHE_SIZE
- 1) >>
3222 struct pagevec pvec
;
3224 if (mapping
->nrpages
== 0)
3227 pagevec_init(&pvec
, 0);
3228 while (next
< end
&&
3229 pagevec_lookup(&pvec
, mapping
, next
, PAGEVEC_SIZE
)) {
3235 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
3236 struct page
*page
= pvec
.pages
[i
];
3237 pgoff_t page_index
= page
->index
;
3238 struct zone
*pagezone
= page_zone(page
);
3241 if (page_index
> next
)
3245 if (pagezone
!= zone
) {
3247 spin_unlock_irq(&zone
->lru_lock
);
3249 spin_lock_irq(&zone
->lru_lock
);
3252 if (PageLRU(page
) && PageUnevictable(page
))
3253 check_move_unevictable_page(page
, zone
);
3256 spin_unlock_irq(&zone
->lru_lock
);
3257 pagevec_release(&pvec
);
3259 count_vm_events(UNEVICTABLE_PGSCANNED
, pg_scanned
);
3265 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3266 * @zone - zone of which to scan the unevictable list
3268 * Scan @zone's unevictable LRU lists to check for pages that have become
3269 * evictable. Move those that have to @zone's inactive list where they
3270 * become candidates for reclaim, unless shrink_inactive_zone() decides
3271 * to reactivate them. Pages that are still unevictable are rotated
3272 * back onto @zone's unevictable list.
3274 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3275 static void scan_zone_unevictable_pages(struct zone
*zone
)
3277 struct list_head
*l_unevictable
= &zone
->lru
[LRU_UNEVICTABLE
].list
;
3279 unsigned long nr_to_scan
= zone_page_state(zone
, NR_UNEVICTABLE
);
3281 while (nr_to_scan
> 0) {
3282 unsigned long batch_size
= min(nr_to_scan
,
3283 SCAN_UNEVICTABLE_BATCH_SIZE
);
3285 spin_lock_irq(&zone
->lru_lock
);
3286 for (scan
= 0; scan
< batch_size
; scan
++) {
3287 struct page
*page
= lru_to_page(l_unevictable
);
3289 if (!trylock_page(page
))
3292 prefetchw_prev_lru_page(page
, l_unevictable
, flags
);
3294 if (likely(PageLRU(page
) && PageUnevictable(page
)))
3295 check_move_unevictable_page(page
, zone
);
3299 spin_unlock_irq(&zone
->lru_lock
);
3301 nr_to_scan
-= batch_size
;
3307 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3309 * A really big hammer: scan all zones' unevictable LRU lists to check for
3310 * pages that have become evictable. Move those back to the zones'
3311 * inactive list where they become candidates for reclaim.
3312 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3313 * and we add swap to the system. As such, it runs in the context of a task
3314 * that has possibly/probably made some previously unevictable pages
3317 static void scan_all_zones_unevictable_pages(void)
3321 for_each_zone(zone
) {
3322 scan_zone_unevictable_pages(zone
);
3327 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3328 * all nodes' unevictable lists for evictable pages
3330 unsigned long scan_unevictable_pages
;
3332 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3333 void __user
*buffer
,
3334 size_t *length
, loff_t
*ppos
)
3336 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3338 if (write
&& *(unsigned long *)table
->data
)
3339 scan_all_zones_unevictable_pages();
3341 scan_unevictable_pages
= 0;
3347 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3348 * a specified node's per zone unevictable lists for evictable pages.
3351 static ssize_t
read_scan_unevictable_node(struct sys_device
*dev
,
3352 struct sysdev_attribute
*attr
,
3355 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3358 static ssize_t
write_scan_unevictable_node(struct sys_device
*dev
,
3359 struct sysdev_attribute
*attr
,
3360 const char *buf
, size_t count
)
3362 struct zone
*node_zones
= NODE_DATA(dev
->id
)->node_zones
;
3365 unsigned long req
= strict_strtoul(buf
, 10, &res
);
3368 return 1; /* zero is no-op */
3370 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
3371 if (!populated_zone(zone
))
3373 scan_zone_unevictable_pages(zone
);
3379 static SYSDEV_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3380 read_scan_unevictable_node
,
3381 write_scan_unevictable_node
);
3383 int scan_unevictable_register_node(struct node
*node
)
3385 return sysdev_create_file(&node
->sysdev
, &attr_scan_unevictable_pages
);
3388 void scan_unevictable_unregister_node(struct node
*node
)
3390 sysdev_remove_file(&node
->sysdev
, &attr_scan_unevictable_pages
);