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? */
81 /* Scan (total_size >> priority) pages at once */
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
88 struct mem_cgroup
*target_mem_cgroup
;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 struct mem_cgroup_zone
{
98 struct mem_cgroup
*mem_cgroup
;
102 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104 #ifdef ARCH_HAS_PREFETCH
105 #define prefetch_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetch(&prev->_field); \
115 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
118 #ifdef ARCH_HAS_PREFETCHW
119 #define prefetchw_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetchw(&prev->_field); \
129 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
133 * From 0 .. 100. Higher means more swappy.
135 int vm_swappiness
= 60;
136 long vm_total_pages
; /* The total number of pages which the VM controls */
138 static LIST_HEAD(shrinker_list
);
139 static DECLARE_RWSEM(shrinker_rwsem
);
141 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
142 static bool global_reclaim(struct scan_control
*sc
)
144 return !sc
->target_mem_cgroup
;
147 static bool global_reclaim(struct scan_control
*sc
)
153 static struct zone_reclaim_stat
*get_reclaim_stat(struct mem_cgroup_zone
*mz
)
155 return &mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
)->reclaim_stat
;
158 static unsigned long get_lruvec_size(struct lruvec
*lruvec
, enum lru_list lru
)
160 if (!mem_cgroup_disabled())
161 return mem_cgroup_get_lruvec_size(lruvec
, lru
);
163 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
167 * Add a shrinker callback to be called from the vm
169 void register_shrinker(struct shrinker
*shrinker
)
171 atomic_long_set(&shrinker
->nr_in_batch
, 0);
172 down_write(&shrinker_rwsem
);
173 list_add_tail(&shrinker
->list
, &shrinker_list
);
174 up_write(&shrinker_rwsem
);
176 EXPORT_SYMBOL(register_shrinker
);
181 void unregister_shrinker(struct shrinker
*shrinker
)
183 down_write(&shrinker_rwsem
);
184 list_del(&shrinker
->list
);
185 up_write(&shrinker_rwsem
);
187 EXPORT_SYMBOL(unregister_shrinker
);
189 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
190 struct shrink_control
*sc
,
191 unsigned long nr_to_scan
)
193 sc
->nr_to_scan
= nr_to_scan
;
194 return (*shrinker
->shrink
)(shrinker
, sc
);
197 #define SHRINK_BATCH 128
199 * Call the shrink functions to age shrinkable caches
201 * Here we assume it costs one seek to replace a lru page and that it also
202 * takes a seek to recreate a cache object. With this in mind we age equal
203 * percentages of the lru and ageable caches. This should balance the seeks
204 * generated by these structures.
206 * If the vm encountered mapped pages on the LRU it increase the pressure on
207 * slab to avoid swapping.
209 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
211 * `lru_pages' represents the number of on-LRU pages in all the zones which
212 * are eligible for the caller's allocation attempt. It is used for balancing
213 * slab reclaim versus page reclaim.
215 * Returns the number of slab objects which we shrunk.
217 unsigned long shrink_slab(struct shrink_control
*shrink
,
218 unsigned long nr_pages_scanned
,
219 unsigned long lru_pages
)
221 struct shrinker
*shrinker
;
222 unsigned long ret
= 0;
224 if (nr_pages_scanned
== 0)
225 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
227 if (!down_read_trylock(&shrinker_rwsem
)) {
228 /* Assume we'll be able to shrink next time */
233 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
234 unsigned long long delta
;
240 long batch_size
= shrinker
->batch
? shrinker
->batch
243 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
248 * copy the current shrinker scan count into a local variable
249 * and zero it so that other concurrent shrinker invocations
250 * don't also do this scanning work.
252 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
255 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
257 do_div(delta
, lru_pages
+ 1);
259 if (total_scan
< 0) {
260 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
262 shrinker
->shrink
, total_scan
);
263 total_scan
= max_pass
;
267 * We need to avoid excessive windup on filesystem shrinkers
268 * due to large numbers of GFP_NOFS allocations causing the
269 * shrinkers to return -1 all the time. This results in a large
270 * nr being built up so when a shrink that can do some work
271 * comes along it empties the entire cache due to nr >>>
272 * max_pass. This is bad for sustaining a working set in
275 * Hence only allow the shrinker to scan the entire cache when
276 * a large delta change is calculated directly.
278 if (delta
< max_pass
/ 4)
279 total_scan
= min(total_scan
, max_pass
/ 2);
282 * Avoid risking looping forever due to too large nr value:
283 * never try to free more than twice the estimate number of
286 if (total_scan
> max_pass
* 2)
287 total_scan
= max_pass
* 2;
289 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
290 nr_pages_scanned
, lru_pages
,
291 max_pass
, delta
, total_scan
);
293 while (total_scan
>= batch_size
) {
296 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
297 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
299 if (shrink_ret
== -1)
301 if (shrink_ret
< nr_before
)
302 ret
+= nr_before
- shrink_ret
;
303 count_vm_events(SLABS_SCANNED
, batch_size
);
304 total_scan
-= batch_size
;
310 * move the unused scan count back into the shrinker in a
311 * manner that handles concurrent updates. If we exhausted the
312 * scan, there is no need to do an update.
315 new_nr
= atomic_long_add_return(total_scan
,
316 &shrinker
->nr_in_batch
);
318 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
320 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
322 up_read(&shrinker_rwsem
);
328 static inline int is_page_cache_freeable(struct page
*page
)
331 * A freeable page cache page is referenced only by the caller
332 * that isolated the page, the page cache radix tree and
333 * optional buffer heads at page->private.
335 return page_count(page
) - page_has_private(page
) == 2;
338 static int may_write_to_queue(struct backing_dev_info
*bdi
,
339 struct scan_control
*sc
)
341 if (current
->flags
& PF_SWAPWRITE
)
343 if (!bdi_write_congested(bdi
))
345 if (bdi
== current
->backing_dev_info
)
351 * We detected a synchronous write error writing a page out. Probably
352 * -ENOSPC. We need to propagate that into the address_space for a subsequent
353 * fsync(), msync() or close().
355 * The tricky part is that after writepage we cannot touch the mapping: nothing
356 * prevents it from being freed up. But we have a ref on the page and once
357 * that page is locked, the mapping is pinned.
359 * We're allowed to run sleeping lock_page() here because we know the caller has
362 static void handle_write_error(struct address_space
*mapping
,
363 struct page
*page
, int error
)
366 if (page_mapping(page
) == mapping
)
367 mapping_set_error(mapping
, error
);
371 /* possible outcome of pageout() */
373 /* failed to write page out, page is locked */
375 /* move page to the active list, page is locked */
377 /* page has been sent to the disk successfully, page is unlocked */
379 /* page is clean and locked */
384 * pageout is called by shrink_page_list() for each dirty page.
385 * Calls ->writepage().
387 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
388 struct scan_control
*sc
)
391 * If the page is dirty, only perform writeback if that write
392 * will be non-blocking. To prevent this allocation from being
393 * stalled by pagecache activity. But note that there may be
394 * stalls if we need to run get_block(). We could test
395 * PagePrivate for that.
397 * If this process is currently in __generic_file_aio_write() against
398 * this page's queue, we can perform writeback even if that
401 * If the page is swapcache, write it back even if that would
402 * block, for some throttling. This happens by accident, because
403 * swap_backing_dev_info is bust: it doesn't reflect the
404 * congestion state of the swapdevs. Easy to fix, if needed.
406 if (!is_page_cache_freeable(page
))
410 * Some data journaling orphaned pages can have
411 * page->mapping == NULL while being dirty with clean buffers.
413 if (page_has_private(page
)) {
414 if (try_to_free_buffers(page
)) {
415 ClearPageDirty(page
);
416 printk("%s: orphaned page\n", __func__
);
422 if (mapping
->a_ops
->writepage
== NULL
)
423 return PAGE_ACTIVATE
;
424 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
427 if (clear_page_dirty_for_io(page
)) {
429 struct writeback_control wbc
= {
430 .sync_mode
= WB_SYNC_NONE
,
431 .nr_to_write
= SWAP_CLUSTER_MAX
,
433 .range_end
= LLONG_MAX
,
437 SetPageReclaim(page
);
438 res
= mapping
->a_ops
->writepage(page
, &wbc
);
440 handle_write_error(mapping
, page
, res
);
441 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
442 ClearPageReclaim(page
);
443 return PAGE_ACTIVATE
;
446 if (!PageWriteback(page
)) {
447 /* synchronous write or broken a_ops? */
448 ClearPageReclaim(page
);
450 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
451 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
459 * Same as remove_mapping, but if the page is removed from the mapping, it
460 * gets returned with a refcount of 0.
462 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
464 BUG_ON(!PageLocked(page
));
465 BUG_ON(mapping
!= page_mapping(page
));
467 spin_lock_irq(&mapping
->tree_lock
);
469 * The non racy check for a busy page.
471 * Must be careful with the order of the tests. When someone has
472 * a ref to the page, it may be possible that they dirty it then
473 * drop the reference. So if PageDirty is tested before page_count
474 * here, then the following race may occur:
476 * get_user_pages(&page);
477 * [user mapping goes away]
479 * !PageDirty(page) [good]
480 * SetPageDirty(page);
482 * !page_count(page) [good, discard it]
484 * [oops, our write_to data is lost]
486 * Reversing the order of the tests ensures such a situation cannot
487 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
488 * load is not satisfied before that of page->_count.
490 * Note that if SetPageDirty is always performed via set_page_dirty,
491 * and thus under tree_lock, then this ordering is not required.
493 if (!page_freeze_refs(page
, 2))
495 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
496 if (unlikely(PageDirty(page
))) {
497 page_unfreeze_refs(page
, 2);
501 if (PageSwapCache(page
)) {
502 swp_entry_t swap
= { .val
= page_private(page
) };
503 __delete_from_swap_cache(page
);
504 spin_unlock_irq(&mapping
->tree_lock
);
505 swapcache_free(swap
, page
);
507 void (*freepage
)(struct page
*);
509 freepage
= mapping
->a_ops
->freepage
;
511 __delete_from_page_cache(page
);
512 spin_unlock_irq(&mapping
->tree_lock
);
513 mem_cgroup_uncharge_cache_page(page
);
515 if (freepage
!= NULL
)
522 spin_unlock_irq(&mapping
->tree_lock
);
527 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
528 * someone else has a ref on the page, abort and return 0. If it was
529 * successfully detached, return 1. Assumes the caller has a single ref on
532 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
534 if (__remove_mapping(mapping
, page
)) {
536 * Unfreezing the refcount with 1 rather than 2 effectively
537 * drops the pagecache ref for us without requiring another
540 page_unfreeze_refs(page
, 1);
547 * putback_lru_page - put previously isolated page onto appropriate LRU list
548 * @page: page to be put back to appropriate lru list
550 * Add previously isolated @page to appropriate LRU list.
551 * Page may still be unevictable for other reasons.
553 * lru_lock must not be held, interrupts must be enabled.
555 void putback_lru_page(struct page
*page
)
558 int active
= !!TestClearPageActive(page
);
559 int was_unevictable
= PageUnevictable(page
);
561 VM_BUG_ON(PageLRU(page
));
564 ClearPageUnevictable(page
);
566 if (page_evictable(page
, NULL
)) {
568 * For evictable pages, we can use the cache.
569 * In event of a race, worst case is we end up with an
570 * unevictable page on [in]active list.
571 * We know how to handle that.
573 lru
= active
+ page_lru_base_type(page
);
574 lru_cache_add_lru(page
, lru
);
577 * Put unevictable pages directly on zone's unevictable
580 lru
= LRU_UNEVICTABLE
;
581 add_page_to_unevictable_list(page
);
583 * When racing with an mlock or AS_UNEVICTABLE clearing
584 * (page is unlocked) make sure that if the other thread
585 * does not observe our setting of PG_lru and fails
586 * isolation/check_move_unevictable_pages,
587 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
588 * the page back to the evictable list.
590 * The other side is TestClearPageMlocked() or shmem_lock().
596 * page's status can change while we move it among lru. If an evictable
597 * page is on unevictable list, it never be freed. To avoid that,
598 * check after we added it to the list, again.
600 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
601 if (!isolate_lru_page(page
)) {
605 /* This means someone else dropped this page from LRU
606 * So, it will be freed or putback to LRU again. There is
607 * nothing to do here.
611 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
612 count_vm_event(UNEVICTABLE_PGRESCUED
);
613 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
614 count_vm_event(UNEVICTABLE_PGCULLED
);
616 put_page(page
); /* drop ref from isolate */
619 enum page_references
{
621 PAGEREF_RECLAIM_CLEAN
,
626 static enum page_references
page_check_references(struct page
*page
,
627 struct scan_control
*sc
)
629 int referenced_ptes
, referenced_page
;
630 unsigned long vm_flags
;
632 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
634 referenced_page
= TestClearPageReferenced(page
);
637 * Mlock lost the isolation race with us. Let try_to_unmap()
638 * move the page to the unevictable list.
640 if (vm_flags
& VM_LOCKED
)
641 return PAGEREF_RECLAIM
;
643 if (referenced_ptes
) {
644 if (PageSwapBacked(page
))
645 return PAGEREF_ACTIVATE
;
647 * All mapped pages start out with page table
648 * references from the instantiating fault, so we need
649 * to look twice if a mapped file page is used more
652 * Mark it and spare it for another trip around the
653 * inactive list. Another page table reference will
654 * lead to its activation.
656 * Note: the mark is set for activated pages as well
657 * so that recently deactivated but used pages are
660 SetPageReferenced(page
);
662 if (referenced_page
|| referenced_ptes
> 1)
663 return PAGEREF_ACTIVATE
;
666 * Activate file-backed executable pages after first usage.
668 if (vm_flags
& VM_EXEC
)
669 return PAGEREF_ACTIVATE
;
674 /* Reclaim if clean, defer dirty pages to writeback */
675 if (referenced_page
&& !PageSwapBacked(page
))
676 return PAGEREF_RECLAIM_CLEAN
;
678 return PAGEREF_RECLAIM
;
682 * shrink_page_list() returns the number of reclaimed pages
684 static unsigned long shrink_page_list(struct list_head
*page_list
,
686 struct scan_control
*sc
,
687 unsigned long *ret_nr_dirty
,
688 unsigned long *ret_nr_writeback
)
690 LIST_HEAD(ret_pages
);
691 LIST_HEAD(free_pages
);
693 unsigned long nr_dirty
= 0;
694 unsigned long nr_congested
= 0;
695 unsigned long nr_reclaimed
= 0;
696 unsigned long nr_writeback
= 0;
700 while (!list_empty(page_list
)) {
701 enum page_references references
;
702 struct address_space
*mapping
;
708 page
= lru_to_page(page_list
);
709 list_del(&page
->lru
);
711 if (!trylock_page(page
))
714 VM_BUG_ON(PageActive(page
));
715 VM_BUG_ON(page_zone(page
) != zone
);
719 if (unlikely(!page_evictable(page
, NULL
)))
722 if (!sc
->may_unmap
&& page_mapped(page
))
725 /* Double the slab pressure for mapped and swapcache pages */
726 if (page_mapped(page
) || PageSwapCache(page
))
729 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
730 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
732 if (PageWriteback(page
)) {
738 references
= page_check_references(page
, sc
);
739 switch (references
) {
740 case PAGEREF_ACTIVATE
:
741 goto activate_locked
;
744 case PAGEREF_RECLAIM
:
745 case PAGEREF_RECLAIM_CLEAN
:
746 ; /* try to reclaim the page below */
750 * Anonymous process memory has backing store?
751 * Try to allocate it some swap space here.
753 if (PageAnon(page
) && !PageSwapCache(page
)) {
754 if (!(sc
->gfp_mask
& __GFP_IO
))
756 if (!add_to_swap(page
))
757 goto activate_locked
;
761 mapping
= page_mapping(page
);
764 * The page is mapped into the page tables of one or more
765 * processes. Try to unmap it here.
767 if (page_mapped(page
) && mapping
) {
768 switch (try_to_unmap(page
, TTU_UNMAP
)) {
770 goto activate_locked
;
776 ; /* try to free the page below */
780 if (PageDirty(page
)) {
784 * Only kswapd can writeback filesystem pages to
785 * avoid risk of stack overflow but do not writeback
786 * unless under significant pressure.
788 if (page_is_file_cache(page
) &&
789 (!current_is_kswapd() ||
790 sc
->priority
>= DEF_PRIORITY
- 2)) {
792 * Immediately reclaim when written back.
793 * Similar in principal to deactivate_page()
794 * except we already have the page isolated
795 * and know it's dirty
797 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
798 SetPageReclaim(page
);
803 if (references
== PAGEREF_RECLAIM_CLEAN
)
807 if (!sc
->may_writepage
)
810 /* Page is dirty, try to write it out here */
811 switch (pageout(page
, mapping
, sc
)) {
816 goto activate_locked
;
818 if (PageWriteback(page
))
824 * A synchronous write - probably a ramdisk. Go
825 * ahead and try to reclaim the page.
827 if (!trylock_page(page
))
829 if (PageDirty(page
) || PageWriteback(page
))
831 mapping
= page_mapping(page
);
833 ; /* try to free the page below */
838 * If the page has buffers, try to free the buffer mappings
839 * associated with this page. If we succeed we try to free
842 * We do this even if the page is PageDirty().
843 * try_to_release_page() does not perform I/O, but it is
844 * possible for a page to have PageDirty set, but it is actually
845 * clean (all its buffers are clean). This happens if the
846 * buffers were written out directly, with submit_bh(). ext3
847 * will do this, as well as the blockdev mapping.
848 * try_to_release_page() will discover that cleanness and will
849 * drop the buffers and mark the page clean - it can be freed.
851 * Rarely, pages can have buffers and no ->mapping. These are
852 * the pages which were not successfully invalidated in
853 * truncate_complete_page(). We try to drop those buffers here
854 * and if that worked, and the page is no longer mapped into
855 * process address space (page_count == 1) it can be freed.
856 * Otherwise, leave the page on the LRU so it is swappable.
858 if (page_has_private(page
)) {
859 if (!try_to_release_page(page
, sc
->gfp_mask
))
860 goto activate_locked
;
861 if (!mapping
&& page_count(page
) == 1) {
863 if (put_page_testzero(page
))
867 * rare race with speculative reference.
868 * the speculative reference will free
869 * this page shortly, so we may
870 * increment nr_reclaimed here (and
871 * leave it off the LRU).
879 if (!mapping
|| !__remove_mapping(mapping
, page
))
883 * At this point, we have no other references and there is
884 * no way to pick any more up (removed from LRU, removed
885 * from pagecache). Can use non-atomic bitops now (and
886 * we obviously don't have to worry about waking up a process
887 * waiting on the page lock, because there are no references.
889 __clear_page_locked(page
);
894 * Is there need to periodically free_page_list? It would
895 * appear not as the counts should be low
897 list_add(&page
->lru
, &free_pages
);
901 if (PageSwapCache(page
))
902 try_to_free_swap(page
);
904 putback_lru_page(page
);
908 /* Not a candidate for swapping, so reclaim swap space. */
909 if (PageSwapCache(page
) && vm_swap_full())
910 try_to_free_swap(page
);
911 VM_BUG_ON(PageActive(page
));
917 list_add(&page
->lru
, &ret_pages
);
918 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
922 * Tag a zone as congested if all the dirty pages encountered were
923 * backed by a congested BDI. In this case, reclaimers should just
924 * back off and wait for congestion to clear because further reclaim
925 * will encounter the same problem
927 if (nr_dirty
&& nr_dirty
== nr_congested
&& global_reclaim(sc
))
928 zone_set_flag(zone
, ZONE_CONGESTED
);
930 free_hot_cold_page_list(&free_pages
, 1);
932 list_splice(&ret_pages
, page_list
);
933 count_vm_events(PGACTIVATE
, pgactivate
);
934 *ret_nr_dirty
+= nr_dirty
;
935 *ret_nr_writeback
+= nr_writeback
;
940 * Attempt to remove the specified page from its LRU. Only take this page
941 * if it is of the appropriate PageActive status. Pages which are being
942 * freed elsewhere are also ignored.
944 * page: page to consider
945 * mode: one of the LRU isolation modes defined above
947 * returns 0 on success, -ve errno on failure.
949 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
953 /* Only take pages on the LRU. */
957 /* Do not give back unevictable pages for compaction */
958 if (PageUnevictable(page
))
964 * To minimise LRU disruption, the caller can indicate that it only
965 * wants to isolate pages it will be able to operate on without
966 * blocking - clean pages for the most part.
968 * ISOLATE_CLEAN means that only clean pages should be isolated. This
969 * is used by reclaim when it is cannot write to backing storage
971 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
972 * that it is possible to migrate without blocking
974 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
975 /* All the caller can do on PageWriteback is block */
976 if (PageWriteback(page
))
979 if (PageDirty(page
)) {
980 struct address_space
*mapping
;
982 /* ISOLATE_CLEAN means only clean pages */
983 if (mode
& ISOLATE_CLEAN
)
987 * Only pages without mappings or that have a
988 * ->migratepage callback are possible to migrate
991 mapping
= page_mapping(page
);
992 if (mapping
&& !mapping
->a_ops
->migratepage
)
997 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1000 if (likely(get_page_unless_zero(page
))) {
1002 * Be careful not to clear PageLRU until after we're
1003 * sure the page is not being freed elsewhere -- the
1004 * page release code relies on it.
1014 * zone->lru_lock is heavily contended. Some of the functions that
1015 * shrink the lists perform better by taking out a batch of pages
1016 * and working on them outside the LRU lock.
1018 * For pagecache intensive workloads, this function is the hottest
1019 * spot in the kernel (apart from copy_*_user functions).
1021 * Appropriate locks must be held before calling this function.
1023 * @nr_to_scan: The number of pages to look through on the list.
1024 * @lruvec: The LRU vector to pull pages from.
1025 * @dst: The temp list to put pages on to.
1026 * @nr_scanned: The number of pages that were scanned.
1027 * @sc: The scan_control struct for this reclaim session
1028 * @mode: One of the LRU isolation modes
1029 * @lru: LRU list id for isolating
1031 * returns how many pages were moved onto *@dst.
1033 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1034 struct lruvec
*lruvec
, struct list_head
*dst
,
1035 unsigned long *nr_scanned
, struct scan_control
*sc
,
1036 isolate_mode_t mode
, enum lru_list lru
)
1038 struct list_head
*src
;
1039 unsigned long nr_taken
= 0;
1041 int file
= is_file_lru(lru
);
1043 src
= &lruvec
->lists
[lru
];
1045 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1048 page
= lru_to_page(src
);
1049 prefetchw_prev_lru_page(page
, src
, flags
);
1051 VM_BUG_ON(!PageLRU(page
));
1053 switch (__isolate_lru_page(page
, mode
)) {
1055 mem_cgroup_lru_del_list(page
, lru
);
1056 list_move(&page
->lru
, dst
);
1057 nr_taken
+= hpage_nr_pages(page
);
1061 /* else it is being freed elsewhere */
1062 list_move(&page
->lru
, src
);
1072 trace_mm_vmscan_lru_isolate(sc
->order
,
1080 * isolate_lru_page - tries to isolate a page from its LRU list
1081 * @page: page to isolate from its LRU list
1083 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1084 * vmstat statistic corresponding to whatever LRU list the page was on.
1086 * Returns 0 if the page was removed from an LRU list.
1087 * Returns -EBUSY if the page was not on an LRU list.
1089 * The returned page will have PageLRU() cleared. If it was found on
1090 * the active list, it will have PageActive set. If it was found on
1091 * the unevictable list, it will have the PageUnevictable bit set. That flag
1092 * may need to be cleared by the caller before letting the page go.
1094 * The vmstat statistic corresponding to the list on which the page was
1095 * found will be decremented.
1098 * (1) Must be called with an elevated refcount on the page. This is a
1099 * fundamentnal difference from isolate_lru_pages (which is called
1100 * without a stable reference).
1101 * (2) the lru_lock must not be held.
1102 * (3) interrupts must be enabled.
1104 int isolate_lru_page(struct page
*page
)
1108 VM_BUG_ON(!page_count(page
));
1110 if (PageLRU(page
)) {
1111 struct zone
*zone
= page_zone(page
);
1113 spin_lock_irq(&zone
->lru_lock
);
1114 if (PageLRU(page
)) {
1115 int lru
= page_lru(page
);
1120 del_page_from_lru_list(zone
, page
, lru
);
1122 spin_unlock_irq(&zone
->lru_lock
);
1128 * Are there way too many processes in the direct reclaim path already?
1130 static int too_many_isolated(struct zone
*zone
, int file
,
1131 struct scan_control
*sc
)
1133 unsigned long inactive
, isolated
;
1135 if (current_is_kswapd())
1138 if (!global_reclaim(sc
))
1142 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1143 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1145 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1146 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1149 return isolated
> inactive
;
1152 static noinline_for_stack
void
1153 putback_inactive_pages(struct lruvec
*lruvec
,
1154 struct list_head
*page_list
)
1156 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1157 struct zone
*zone
= lruvec_zone(lruvec
);
1158 LIST_HEAD(pages_to_free
);
1161 * Put back any unfreeable pages.
1163 while (!list_empty(page_list
)) {
1164 struct page
*page
= lru_to_page(page_list
);
1167 VM_BUG_ON(PageLRU(page
));
1168 list_del(&page
->lru
);
1169 if (unlikely(!page_evictable(page
, NULL
))) {
1170 spin_unlock_irq(&zone
->lru_lock
);
1171 putback_lru_page(page
);
1172 spin_lock_irq(&zone
->lru_lock
);
1176 lru
= page_lru(page
);
1177 add_page_to_lru_list(zone
, page
, lru
);
1178 if (is_active_lru(lru
)) {
1179 int file
= is_file_lru(lru
);
1180 int numpages
= hpage_nr_pages(page
);
1181 reclaim_stat
->recent_rotated
[file
] += numpages
;
1183 if (put_page_testzero(page
)) {
1184 __ClearPageLRU(page
);
1185 __ClearPageActive(page
);
1186 del_page_from_lru_list(zone
, page
, lru
);
1188 if (unlikely(PageCompound(page
))) {
1189 spin_unlock_irq(&zone
->lru_lock
);
1190 (*get_compound_page_dtor(page
))(page
);
1191 spin_lock_irq(&zone
->lru_lock
);
1193 list_add(&page
->lru
, &pages_to_free
);
1198 * To save our caller's stack, now use input list for pages to free.
1200 list_splice(&pages_to_free
, page_list
);
1204 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1205 * of reclaimed pages
1207 static noinline_for_stack
unsigned long
1208 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1209 struct scan_control
*sc
, enum lru_list lru
)
1211 LIST_HEAD(page_list
);
1212 unsigned long nr_scanned
;
1213 unsigned long nr_reclaimed
= 0;
1214 unsigned long nr_taken
;
1215 unsigned long nr_dirty
= 0;
1216 unsigned long nr_writeback
= 0;
1217 isolate_mode_t isolate_mode
= 0;
1218 int file
= is_file_lru(lru
);
1219 struct zone
*zone
= lruvec_zone(lruvec
);
1220 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1222 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1223 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1225 /* We are about to die and free our memory. Return now. */
1226 if (fatal_signal_pending(current
))
1227 return SWAP_CLUSTER_MAX
;
1233 isolate_mode
|= ISOLATE_UNMAPPED
;
1234 if (!sc
->may_writepage
)
1235 isolate_mode
|= ISOLATE_CLEAN
;
1237 spin_lock_irq(&zone
->lru_lock
);
1239 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1240 &nr_scanned
, sc
, isolate_mode
, lru
);
1242 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1243 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1245 if (global_reclaim(sc
)) {
1246 zone
->pages_scanned
+= nr_scanned
;
1247 if (current_is_kswapd())
1248 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1251 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1254 spin_unlock_irq(&zone
->lru_lock
);
1259 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
,
1260 &nr_dirty
, &nr_writeback
);
1262 spin_lock_irq(&zone
->lru_lock
);
1264 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1266 if (global_reclaim(sc
)) {
1267 if (current_is_kswapd())
1268 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1271 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1275 putback_inactive_pages(lruvec
, &page_list
);
1277 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1279 spin_unlock_irq(&zone
->lru_lock
);
1281 free_hot_cold_page_list(&page_list
, 1);
1284 * If reclaim is isolating dirty pages under writeback, it implies
1285 * that the long-lived page allocation rate is exceeding the page
1286 * laundering rate. Either the global limits are not being effective
1287 * at throttling processes due to the page distribution throughout
1288 * zones or there is heavy usage of a slow backing device. The
1289 * only option is to throttle from reclaim context which is not ideal
1290 * as there is no guarantee the dirtying process is throttled in the
1291 * same way balance_dirty_pages() manages.
1293 * This scales the number of dirty pages that must be under writeback
1294 * before throttling depending on priority. It is a simple backoff
1295 * function that has the most effect in the range DEF_PRIORITY to
1296 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1297 * in trouble and reclaim is considered to be in trouble.
1299 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1300 * DEF_PRIORITY-1 50% must be PageWriteback
1301 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1303 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1304 * isolated page is PageWriteback
1306 if (nr_writeback
&& nr_writeback
>=
1307 (nr_taken
>> (DEF_PRIORITY
- sc
->priority
)))
1308 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1310 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1312 nr_scanned
, nr_reclaimed
,
1314 trace_shrink_flags(file
));
1315 return nr_reclaimed
;
1319 * This moves pages from the active list to the inactive list.
1321 * We move them the other way if the page is referenced by one or more
1322 * processes, from rmap.
1324 * If the pages are mostly unmapped, the processing is fast and it is
1325 * appropriate to hold zone->lru_lock across the whole operation. But if
1326 * the pages are mapped, the processing is slow (page_referenced()) so we
1327 * should drop zone->lru_lock around each page. It's impossible to balance
1328 * this, so instead we remove the pages from the LRU while processing them.
1329 * It is safe to rely on PG_active against the non-LRU pages in here because
1330 * nobody will play with that bit on a non-LRU page.
1332 * The downside is that we have to touch page->_count against each page.
1333 * But we had to alter page->flags anyway.
1336 static void move_active_pages_to_lru(struct zone
*zone
,
1337 struct list_head
*list
,
1338 struct list_head
*pages_to_free
,
1341 unsigned long pgmoved
= 0;
1344 while (!list_empty(list
)) {
1345 struct lruvec
*lruvec
;
1347 page
= lru_to_page(list
);
1349 VM_BUG_ON(PageLRU(page
));
1352 lruvec
= mem_cgroup_lru_add_list(zone
, page
, lru
);
1353 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1354 pgmoved
+= hpage_nr_pages(page
);
1356 if (put_page_testzero(page
)) {
1357 __ClearPageLRU(page
);
1358 __ClearPageActive(page
);
1359 del_page_from_lru_list(zone
, page
, lru
);
1361 if (unlikely(PageCompound(page
))) {
1362 spin_unlock_irq(&zone
->lru_lock
);
1363 (*get_compound_page_dtor(page
))(page
);
1364 spin_lock_irq(&zone
->lru_lock
);
1366 list_add(&page
->lru
, pages_to_free
);
1369 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1370 if (!is_active_lru(lru
))
1371 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1374 static void shrink_active_list(unsigned long nr_to_scan
,
1375 struct lruvec
*lruvec
,
1376 struct scan_control
*sc
,
1379 unsigned long nr_taken
;
1380 unsigned long nr_scanned
;
1381 unsigned long vm_flags
;
1382 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1383 LIST_HEAD(l_active
);
1384 LIST_HEAD(l_inactive
);
1386 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1387 unsigned long nr_rotated
= 0;
1388 isolate_mode_t isolate_mode
= 0;
1389 int file
= is_file_lru(lru
);
1390 struct zone
*zone
= lruvec_zone(lruvec
);
1395 isolate_mode
|= ISOLATE_UNMAPPED
;
1396 if (!sc
->may_writepage
)
1397 isolate_mode
|= ISOLATE_CLEAN
;
1399 spin_lock_irq(&zone
->lru_lock
);
1401 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1402 &nr_scanned
, sc
, isolate_mode
, lru
);
1403 if (global_reclaim(sc
))
1404 zone
->pages_scanned
+= nr_scanned
;
1406 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1408 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1409 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1410 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1411 spin_unlock_irq(&zone
->lru_lock
);
1413 while (!list_empty(&l_hold
)) {
1415 page
= lru_to_page(&l_hold
);
1416 list_del(&page
->lru
);
1418 if (unlikely(!page_evictable(page
, NULL
))) {
1419 putback_lru_page(page
);
1423 if (unlikely(buffer_heads_over_limit
)) {
1424 if (page_has_private(page
) && trylock_page(page
)) {
1425 if (page_has_private(page
))
1426 try_to_release_page(page
, 0);
1431 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1433 nr_rotated
+= hpage_nr_pages(page
);
1435 * Identify referenced, file-backed active pages and
1436 * give them one more trip around the active list. So
1437 * that executable code get better chances to stay in
1438 * memory under moderate memory pressure. Anon pages
1439 * are not likely to be evicted by use-once streaming
1440 * IO, plus JVM can create lots of anon VM_EXEC pages,
1441 * so we ignore them here.
1443 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1444 list_add(&page
->lru
, &l_active
);
1449 ClearPageActive(page
); /* we are de-activating */
1450 list_add(&page
->lru
, &l_inactive
);
1454 * Move pages back to the lru list.
1456 spin_lock_irq(&zone
->lru_lock
);
1458 * Count referenced pages from currently used mappings as rotated,
1459 * even though only some of them are actually re-activated. This
1460 * helps balance scan pressure between file and anonymous pages in
1463 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1465 move_active_pages_to_lru(zone
, &l_active
, &l_hold
, lru
);
1466 move_active_pages_to_lru(zone
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1467 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1468 spin_unlock_irq(&zone
->lru_lock
);
1470 free_hot_cold_page_list(&l_hold
, 1);
1474 static int inactive_anon_is_low_global(struct zone
*zone
)
1476 unsigned long active
, inactive
;
1478 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1479 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1481 if (inactive
* zone
->inactive_ratio
< active
)
1488 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1489 * @lruvec: LRU vector to check
1491 * Returns true if the zone does not have enough inactive anon pages,
1492 * meaning some active anon pages need to be deactivated.
1494 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1497 * If we don't have swap space, anonymous page deactivation
1500 if (!total_swap_pages
)
1503 if (!mem_cgroup_disabled())
1504 return mem_cgroup_inactive_anon_is_low(lruvec
);
1506 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1509 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1515 static int inactive_file_is_low_global(struct zone
*zone
)
1517 unsigned long active
, inactive
;
1519 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1520 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1522 return (active
> inactive
);
1526 * inactive_file_is_low - check if file pages need to be deactivated
1527 * @lruvec: LRU vector to check
1529 * When the system is doing streaming IO, memory pressure here
1530 * ensures that active file pages get deactivated, until more
1531 * than half of the file pages are on the inactive list.
1533 * Once we get to that situation, protect the system's working
1534 * set from being evicted by disabling active file page aging.
1536 * This uses a different ratio than the anonymous pages, because
1537 * the page cache uses a use-once replacement algorithm.
1539 static int inactive_file_is_low(struct lruvec
*lruvec
)
1541 if (!mem_cgroup_disabled())
1542 return mem_cgroup_inactive_file_is_low(lruvec
);
1544 return inactive_file_is_low_global(lruvec_zone(lruvec
));
1547 static int inactive_list_is_low(struct lruvec
*lruvec
, int file
)
1550 return inactive_file_is_low(lruvec
);
1552 return inactive_anon_is_low(lruvec
);
1555 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1556 struct lruvec
*lruvec
, struct scan_control
*sc
)
1558 int file
= is_file_lru(lru
);
1560 if (is_active_lru(lru
)) {
1561 if (inactive_list_is_low(lruvec
, file
))
1562 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1566 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1569 static int vmscan_swappiness(struct scan_control
*sc
)
1571 if (global_reclaim(sc
))
1572 return vm_swappiness
;
1573 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1577 * Determine how aggressively the anon and file LRU lists should be
1578 * scanned. The relative value of each set of LRU lists is determined
1579 * by looking at the fraction of the pages scanned we did rotate back
1580 * onto the active list instead of evict.
1582 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1584 static void get_scan_count(struct mem_cgroup_zone
*mz
, struct scan_control
*sc
,
1587 unsigned long anon
, file
, free
;
1588 unsigned long anon_prio
, file_prio
;
1589 unsigned long ap
, fp
;
1590 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1591 u64 fraction
[2], denominator
;
1594 bool force_scan
= false;
1595 struct lruvec
*lruvec
;
1597 lruvec
= mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
);
1600 * If the zone or memcg is small, nr[l] can be 0. This
1601 * results in no scanning on this priority and a potential
1602 * priority drop. Global direct reclaim can go to the next
1603 * zone and tends to have no problems. Global kswapd is for
1604 * zone balancing and it needs to scan a minimum amount. When
1605 * reclaiming for a memcg, a priority drop can cause high
1606 * latencies, so it's better to scan a minimum amount there as
1609 if (current_is_kswapd() && mz
->zone
->all_unreclaimable
)
1611 if (!global_reclaim(sc
))
1614 /* If we have no swap space, do not bother scanning anon pages. */
1615 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1623 anon
= get_lruvec_size(lruvec
, LRU_ACTIVE_ANON
) +
1624 get_lruvec_size(lruvec
, LRU_INACTIVE_ANON
);
1625 file
= get_lruvec_size(lruvec
, LRU_ACTIVE_FILE
) +
1626 get_lruvec_size(lruvec
, LRU_INACTIVE_FILE
);
1628 if (global_reclaim(sc
)) {
1629 free
= zone_page_state(mz
->zone
, NR_FREE_PAGES
);
1630 /* If we have very few page cache pages,
1631 force-scan anon pages. */
1632 if (unlikely(file
+ free
<= high_wmark_pages(mz
->zone
))) {
1641 * With swappiness at 100, anonymous and file have the same priority.
1642 * This scanning priority is essentially the inverse of IO cost.
1644 anon_prio
= vmscan_swappiness(sc
);
1645 file_prio
= 200 - vmscan_swappiness(sc
);
1648 * OK, so we have swap space and a fair amount of page cache
1649 * pages. We use the recently rotated / recently scanned
1650 * ratios to determine how valuable each cache is.
1652 * Because workloads change over time (and to avoid overflow)
1653 * we keep these statistics as a floating average, which ends
1654 * up weighing recent references more than old ones.
1656 * anon in [0], file in [1]
1658 spin_lock_irq(&mz
->zone
->lru_lock
);
1659 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1660 reclaim_stat
->recent_scanned
[0] /= 2;
1661 reclaim_stat
->recent_rotated
[0] /= 2;
1664 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1665 reclaim_stat
->recent_scanned
[1] /= 2;
1666 reclaim_stat
->recent_rotated
[1] /= 2;
1670 * The amount of pressure on anon vs file pages is inversely
1671 * proportional to the fraction of recently scanned pages on
1672 * each list that were recently referenced and in active use.
1674 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1675 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1677 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1678 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1679 spin_unlock_irq(&mz
->zone
->lru_lock
);
1683 denominator
= ap
+ fp
+ 1;
1685 for_each_evictable_lru(lru
) {
1686 int file
= is_file_lru(lru
);
1689 scan
= get_lruvec_size(lruvec
, lru
);
1690 if (sc
->priority
|| noswap
|| !vmscan_swappiness(sc
)) {
1691 scan
>>= sc
->priority
;
1692 if (!scan
&& force_scan
)
1693 scan
= SWAP_CLUSTER_MAX
;
1694 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1700 /* Use reclaim/compaction for costly allocs or under memory pressure */
1701 static bool in_reclaim_compaction(struct scan_control
*sc
)
1703 if (COMPACTION_BUILD
&& sc
->order
&&
1704 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
1705 sc
->priority
< DEF_PRIORITY
- 2))
1712 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1713 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1714 * true if more pages should be reclaimed such that when the page allocator
1715 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1716 * It will give up earlier than that if there is difficulty reclaiming pages.
1718 static inline bool should_continue_reclaim(struct mem_cgroup_zone
*mz
,
1719 unsigned long nr_reclaimed
,
1720 unsigned long nr_scanned
,
1721 struct scan_control
*sc
)
1723 unsigned long pages_for_compaction
;
1724 unsigned long inactive_lru_pages
;
1725 struct lruvec
*lruvec
;
1727 /* If not in reclaim/compaction mode, stop */
1728 if (!in_reclaim_compaction(sc
))
1731 /* Consider stopping depending on scan and reclaim activity */
1732 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1734 * For __GFP_REPEAT allocations, stop reclaiming if the
1735 * full LRU list has been scanned and we are still failing
1736 * to reclaim pages. This full LRU scan is potentially
1737 * expensive but a __GFP_REPEAT caller really wants to succeed
1739 if (!nr_reclaimed
&& !nr_scanned
)
1743 * For non-__GFP_REPEAT allocations which can presumably
1744 * fail without consequence, stop if we failed to reclaim
1745 * any pages from the last SWAP_CLUSTER_MAX number of
1746 * pages that were scanned. This will return to the
1747 * caller faster at the risk reclaim/compaction and
1748 * the resulting allocation attempt fails
1755 * If we have not reclaimed enough pages for compaction and the
1756 * inactive lists are large enough, continue reclaiming
1758 lruvec
= mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
);
1759 pages_for_compaction
= (2UL << sc
->order
);
1760 inactive_lru_pages
= get_lruvec_size(lruvec
, LRU_INACTIVE_FILE
);
1761 if (nr_swap_pages
> 0)
1762 inactive_lru_pages
+= get_lruvec_size(lruvec
,
1764 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1765 inactive_lru_pages
> pages_for_compaction
)
1768 /* If compaction would go ahead or the allocation would succeed, stop */
1769 switch (compaction_suitable(mz
->zone
, sc
->order
)) {
1770 case COMPACT_PARTIAL
:
1771 case COMPACT_CONTINUE
:
1779 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1781 static void shrink_mem_cgroup_zone(struct mem_cgroup_zone
*mz
,
1782 struct scan_control
*sc
)
1784 unsigned long nr
[NR_LRU_LISTS
];
1785 unsigned long nr_to_scan
;
1787 unsigned long nr_reclaimed
, nr_scanned
;
1788 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1789 struct blk_plug plug
;
1790 struct lruvec
*lruvec
;
1792 lruvec
= mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
);
1796 nr_scanned
= sc
->nr_scanned
;
1797 get_scan_count(mz
, sc
, nr
);
1799 blk_start_plug(&plug
);
1800 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1801 nr
[LRU_INACTIVE_FILE
]) {
1802 for_each_evictable_lru(lru
) {
1804 nr_to_scan
= min_t(unsigned long,
1805 nr
[lru
], SWAP_CLUSTER_MAX
);
1806 nr
[lru
] -= nr_to_scan
;
1808 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1813 * On large memory systems, scan >> priority can become
1814 * really large. This is fine for the starting priority;
1815 * we want to put equal scanning pressure on each zone.
1816 * However, if the VM has a harder time of freeing pages,
1817 * with multiple processes reclaiming pages, the total
1818 * freeing target can get unreasonably large.
1820 if (nr_reclaimed
>= nr_to_reclaim
&&
1821 sc
->priority
< DEF_PRIORITY
)
1824 blk_finish_plug(&plug
);
1825 sc
->nr_reclaimed
+= nr_reclaimed
;
1828 * Even if we did not try to evict anon pages at all, we want to
1829 * rebalance the anon lru active/inactive ratio.
1831 if (inactive_anon_is_low(lruvec
))
1832 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
1833 sc
, LRU_ACTIVE_ANON
);
1835 /* reclaim/compaction might need reclaim to continue */
1836 if (should_continue_reclaim(mz
, nr_reclaimed
,
1837 sc
->nr_scanned
- nr_scanned
, sc
))
1840 throttle_vm_writeout(sc
->gfp_mask
);
1843 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
1845 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
1846 struct mem_cgroup_reclaim_cookie reclaim
= {
1848 .priority
= sc
->priority
,
1850 struct mem_cgroup
*memcg
;
1852 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
1854 struct mem_cgroup_zone mz
= {
1855 .mem_cgroup
= memcg
,
1859 shrink_mem_cgroup_zone(&mz
, sc
);
1861 * Limit reclaim has historically picked one memcg and
1862 * scanned it with decreasing priority levels until
1863 * nr_to_reclaim had been reclaimed. This priority
1864 * cycle is thus over after a single memcg.
1866 * Direct reclaim and kswapd, on the other hand, have
1867 * to scan all memory cgroups to fulfill the overall
1868 * scan target for the zone.
1870 if (!global_reclaim(sc
)) {
1871 mem_cgroup_iter_break(root
, memcg
);
1874 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
1878 /* Returns true if compaction should go ahead for a high-order request */
1879 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
1881 unsigned long balance_gap
, watermark
;
1884 /* Do not consider compaction for orders reclaim is meant to satisfy */
1885 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1889 * Compaction takes time to run and there are potentially other
1890 * callers using the pages just freed. Continue reclaiming until
1891 * there is a buffer of free pages available to give compaction
1892 * a reasonable chance of completing and allocating the page
1894 balance_gap
= min(low_wmark_pages(zone
),
1895 (zone
->present_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
1896 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
1897 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
1898 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
1901 * If compaction is deferred, reclaim up to a point where
1902 * compaction will have a chance of success when re-enabled
1904 if (compaction_deferred(zone
, sc
->order
))
1905 return watermark_ok
;
1907 /* If compaction is not ready to start, keep reclaiming */
1908 if (!compaction_suitable(zone
, sc
->order
))
1911 return watermark_ok
;
1915 * This is the direct reclaim path, for page-allocating processes. We only
1916 * try to reclaim pages from zones which will satisfy the caller's allocation
1919 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1921 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1923 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1924 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1925 * zone defense algorithm.
1927 * If a zone is deemed to be full of pinned pages then just give it a light
1928 * scan then give up on it.
1930 * This function returns true if a zone is being reclaimed for a costly
1931 * high-order allocation and compaction is ready to begin. This indicates to
1932 * the caller that it should consider retrying the allocation instead of
1935 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
1939 unsigned long nr_soft_reclaimed
;
1940 unsigned long nr_soft_scanned
;
1941 bool aborted_reclaim
= false;
1944 * If the number of buffer_heads in the machine exceeds the maximum
1945 * allowed level, force direct reclaim to scan the highmem zone as
1946 * highmem pages could be pinning lowmem pages storing buffer_heads
1948 if (buffer_heads_over_limit
)
1949 sc
->gfp_mask
|= __GFP_HIGHMEM
;
1951 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1952 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
1953 if (!populated_zone(zone
))
1956 * Take care memory controller reclaiming has small influence
1959 if (global_reclaim(sc
)) {
1960 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1962 if (zone
->all_unreclaimable
&&
1963 sc
->priority
!= DEF_PRIORITY
)
1964 continue; /* Let kswapd poll it */
1965 if (COMPACTION_BUILD
) {
1967 * If we already have plenty of memory free for
1968 * compaction in this zone, don't free any more.
1969 * Even though compaction is invoked for any
1970 * non-zero order, only frequent costly order
1971 * reclamation is disruptive enough to become a
1972 * noticeable problem, like transparent huge
1975 if (compaction_ready(zone
, sc
)) {
1976 aborted_reclaim
= true;
1981 * This steals pages from memory cgroups over softlimit
1982 * and returns the number of reclaimed pages and
1983 * scanned pages. This works for global memory pressure
1984 * and balancing, not for a memcg's limit.
1986 nr_soft_scanned
= 0;
1987 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
1988 sc
->order
, sc
->gfp_mask
,
1990 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
1991 sc
->nr_scanned
+= nr_soft_scanned
;
1992 /* need some check for avoid more shrink_zone() */
1995 shrink_zone(zone
, sc
);
1998 return aborted_reclaim
;
2001 static bool zone_reclaimable(struct zone
*zone
)
2003 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2006 /* All zones in zonelist are unreclaimable? */
2007 static bool all_unreclaimable(struct zonelist
*zonelist
,
2008 struct scan_control
*sc
)
2013 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2014 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2015 if (!populated_zone(zone
))
2017 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2019 if (!zone
->all_unreclaimable
)
2027 * This is the main entry point to direct page reclaim.
2029 * If a full scan of the inactive list fails to free enough memory then we
2030 * are "out of memory" and something needs to be killed.
2032 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2033 * high - the zone may be full of dirty or under-writeback pages, which this
2034 * caller can't do much about. We kick the writeback threads and take explicit
2035 * naps in the hope that some of these pages can be written. But if the
2036 * allocating task holds filesystem locks which prevent writeout this might not
2037 * work, and the allocation attempt will fail.
2039 * returns: 0, if no pages reclaimed
2040 * else, the number of pages reclaimed
2042 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2043 struct scan_control
*sc
,
2044 struct shrink_control
*shrink
)
2046 unsigned long total_scanned
= 0;
2047 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2050 unsigned long writeback_threshold
;
2051 bool aborted_reclaim
;
2053 delayacct_freepages_start();
2055 if (global_reclaim(sc
))
2056 count_vm_event(ALLOCSTALL
);
2060 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2063 * Don't shrink slabs when reclaiming memory from
2064 * over limit cgroups
2066 if (global_reclaim(sc
)) {
2067 unsigned long lru_pages
= 0;
2068 for_each_zone_zonelist(zone
, z
, zonelist
,
2069 gfp_zone(sc
->gfp_mask
)) {
2070 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2073 lru_pages
+= zone_reclaimable_pages(zone
);
2076 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2077 if (reclaim_state
) {
2078 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2079 reclaim_state
->reclaimed_slab
= 0;
2082 total_scanned
+= sc
->nr_scanned
;
2083 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2087 * Try to write back as many pages as we just scanned. This
2088 * tends to cause slow streaming writers to write data to the
2089 * disk smoothly, at the dirtying rate, which is nice. But
2090 * that's undesirable in laptop mode, where we *want* lumpy
2091 * writeout. So in laptop mode, write out the whole world.
2093 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2094 if (total_scanned
> writeback_threshold
) {
2095 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2096 WB_REASON_TRY_TO_FREE_PAGES
);
2097 sc
->may_writepage
= 1;
2100 /* Take a nap, wait for some writeback to complete */
2101 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2102 sc
->priority
< DEF_PRIORITY
- 2) {
2103 struct zone
*preferred_zone
;
2105 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2106 &cpuset_current_mems_allowed
,
2108 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2110 } while (--sc
->priority
>= 0);
2113 delayacct_freepages_end();
2115 if (sc
->nr_reclaimed
)
2116 return sc
->nr_reclaimed
;
2119 * As hibernation is going on, kswapd is freezed so that it can't mark
2120 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2123 if (oom_killer_disabled
)
2126 /* Aborted reclaim to try compaction? don't OOM, then */
2127 if (aborted_reclaim
)
2130 /* top priority shrink_zones still had more to do? don't OOM, then */
2131 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2137 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2138 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2140 unsigned long nr_reclaimed
;
2141 struct scan_control sc
= {
2142 .gfp_mask
= gfp_mask
,
2143 .may_writepage
= !laptop_mode
,
2144 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2148 .priority
= DEF_PRIORITY
,
2149 .target_mem_cgroup
= NULL
,
2150 .nodemask
= nodemask
,
2152 struct shrink_control shrink
= {
2153 .gfp_mask
= sc
.gfp_mask
,
2156 trace_mm_vmscan_direct_reclaim_begin(order
,
2160 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2162 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2164 return nr_reclaimed
;
2167 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2169 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2170 gfp_t gfp_mask
, bool noswap
,
2172 unsigned long *nr_scanned
)
2174 struct scan_control sc
= {
2176 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2177 .may_writepage
= !laptop_mode
,
2179 .may_swap
= !noswap
,
2182 .target_mem_cgroup
= memcg
,
2184 struct mem_cgroup_zone mz
= {
2185 .mem_cgroup
= memcg
,
2189 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2190 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2192 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2197 * NOTE: Although we can get the priority field, using it
2198 * here is not a good idea, since it limits the pages we can scan.
2199 * if we don't reclaim here, the shrink_zone from balance_pgdat
2200 * will pick up pages from other mem cgroup's as well. We hack
2201 * the priority and make it zero.
2203 shrink_mem_cgroup_zone(&mz
, &sc
);
2205 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2207 *nr_scanned
= sc
.nr_scanned
;
2208 return sc
.nr_reclaimed
;
2211 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2215 struct zonelist
*zonelist
;
2216 unsigned long nr_reclaimed
;
2218 struct scan_control sc
= {
2219 .may_writepage
= !laptop_mode
,
2221 .may_swap
= !noswap
,
2222 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2224 .priority
= DEF_PRIORITY
,
2225 .target_mem_cgroup
= memcg
,
2226 .nodemask
= NULL
, /* we don't care the placement */
2227 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2228 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2230 struct shrink_control shrink
= {
2231 .gfp_mask
= sc
.gfp_mask
,
2235 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2236 * take care of from where we get pages. So the node where we start the
2237 * scan does not need to be the current node.
2239 nid
= mem_cgroup_select_victim_node(memcg
);
2241 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2243 trace_mm_vmscan_memcg_reclaim_begin(0,
2247 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2249 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2251 return nr_reclaimed
;
2255 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2257 struct mem_cgroup
*memcg
;
2259 if (!total_swap_pages
)
2262 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2264 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2266 if (inactive_anon_is_low(lruvec
))
2267 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2268 sc
, LRU_ACTIVE_ANON
);
2270 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2275 * pgdat_balanced is used when checking if a node is balanced for high-order
2276 * allocations. Only zones that meet watermarks and are in a zone allowed
2277 * by the callers classzone_idx are added to balanced_pages. The total of
2278 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2279 * for the node to be considered balanced. Forcing all zones to be balanced
2280 * for high orders can cause excessive reclaim when there are imbalanced zones.
2281 * The choice of 25% is due to
2282 * o a 16M DMA zone that is balanced will not balance a zone on any
2283 * reasonable sized machine
2284 * o On all other machines, the top zone must be at least a reasonable
2285 * percentage of the middle zones. For example, on 32-bit x86, highmem
2286 * would need to be at least 256M for it to be balance a whole node.
2287 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2288 * to balance a node on its own. These seemed like reasonable ratios.
2290 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2293 unsigned long present_pages
= 0;
2296 for (i
= 0; i
<= classzone_idx
; i
++)
2297 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2299 /* A special case here: if zone has no page, we think it's balanced */
2300 return balanced_pages
>= (present_pages
>> 2);
2303 /* is kswapd sleeping prematurely? */
2304 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2308 unsigned long balanced
= 0;
2309 bool all_zones_ok
= true;
2311 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2315 /* Check the watermark levels */
2316 for (i
= 0; i
<= classzone_idx
; i
++) {
2317 struct zone
*zone
= pgdat
->node_zones
+ i
;
2319 if (!populated_zone(zone
))
2323 * balance_pgdat() skips over all_unreclaimable after
2324 * DEF_PRIORITY. Effectively, it considers them balanced so
2325 * they must be considered balanced here as well if kswapd
2328 if (zone
->all_unreclaimable
) {
2329 balanced
+= zone
->present_pages
;
2333 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2335 all_zones_ok
= false;
2337 balanced
+= zone
->present_pages
;
2341 * For high-order requests, the balanced zones must contain at least
2342 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2346 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2348 return !all_zones_ok
;
2352 * For kswapd, balance_pgdat() will work across all this node's zones until
2353 * they are all at high_wmark_pages(zone).
2355 * Returns the final order kswapd was reclaiming at
2357 * There is special handling here for zones which are full of pinned pages.
2358 * This can happen if the pages are all mlocked, or if they are all used by
2359 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2360 * What we do is to detect the case where all pages in the zone have been
2361 * scanned twice and there has been zero successful reclaim. Mark the zone as
2362 * dead and from now on, only perform a short scan. Basically we're polling
2363 * the zone for when the problem goes away.
2365 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2366 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2367 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2368 * lower zones regardless of the number of free pages in the lower zones. This
2369 * interoperates with the page allocator fallback scheme to ensure that aging
2370 * of pages is balanced across the zones.
2372 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2376 unsigned long balanced
;
2378 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2379 unsigned long total_scanned
;
2380 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2381 unsigned long nr_soft_reclaimed
;
2382 unsigned long nr_soft_scanned
;
2383 struct scan_control sc
= {
2384 .gfp_mask
= GFP_KERNEL
,
2388 * kswapd doesn't want to be bailed out while reclaim. because
2389 * we want to put equal scanning pressure on each zone.
2391 .nr_to_reclaim
= ULONG_MAX
,
2393 .target_mem_cgroup
= NULL
,
2395 struct shrink_control shrink
= {
2396 .gfp_mask
= sc
.gfp_mask
,
2400 sc
.priority
= DEF_PRIORITY
;
2401 sc
.nr_reclaimed
= 0;
2402 sc
.may_writepage
= !laptop_mode
;
2403 count_vm_event(PAGEOUTRUN
);
2406 unsigned long lru_pages
= 0;
2407 int has_under_min_watermark_zone
= 0;
2413 * Scan in the highmem->dma direction for the highest
2414 * zone which needs scanning
2416 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2417 struct zone
*zone
= pgdat
->node_zones
+ i
;
2419 if (!populated_zone(zone
))
2422 if (zone
->all_unreclaimable
&&
2423 sc
.priority
!= DEF_PRIORITY
)
2427 * Do some background aging of the anon list, to give
2428 * pages a chance to be referenced before reclaiming.
2430 age_active_anon(zone
, &sc
);
2433 * If the number of buffer_heads in the machine
2434 * exceeds the maximum allowed level and this node
2435 * has a highmem zone, force kswapd to reclaim from
2436 * it to relieve lowmem pressure.
2438 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2443 if (!zone_watermark_ok_safe(zone
, order
,
2444 high_wmark_pages(zone
), 0, 0)) {
2448 /* If balanced, clear the congested flag */
2449 zone_clear_flag(zone
, ZONE_CONGESTED
);
2455 for (i
= 0; i
<= end_zone
; i
++) {
2456 struct zone
*zone
= pgdat
->node_zones
+ i
;
2458 lru_pages
+= zone_reclaimable_pages(zone
);
2462 * Now scan the zone in the dma->highmem direction, stopping
2463 * at the last zone which needs scanning.
2465 * We do this because the page allocator works in the opposite
2466 * direction. This prevents the page allocator from allocating
2467 * pages behind kswapd's direction of progress, which would
2468 * cause too much scanning of the lower zones.
2470 for (i
= 0; i
<= end_zone
; i
++) {
2471 struct zone
*zone
= pgdat
->node_zones
+ i
;
2472 int nr_slab
, testorder
;
2473 unsigned long balance_gap
;
2475 if (!populated_zone(zone
))
2478 if (zone
->all_unreclaimable
&&
2479 sc
.priority
!= DEF_PRIORITY
)
2484 nr_soft_scanned
= 0;
2486 * Call soft limit reclaim before calling shrink_zone.
2488 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2491 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2492 total_scanned
+= nr_soft_scanned
;
2495 * We put equal pressure on every zone, unless
2496 * one zone has way too many pages free
2497 * already. The "too many pages" is defined
2498 * as the high wmark plus a "gap" where the
2499 * gap is either the low watermark or 1%
2500 * of the zone, whichever is smaller.
2502 balance_gap
= min(low_wmark_pages(zone
),
2503 (zone
->present_pages
+
2504 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2505 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2507 * Kswapd reclaims only single pages with compaction
2508 * enabled. Trying too hard to reclaim until contiguous
2509 * free pages have become available can hurt performance
2510 * by evicting too much useful data from memory.
2511 * Do not reclaim more than needed for compaction.
2514 if (COMPACTION_BUILD
&& order
&&
2515 compaction_suitable(zone
, order
) !=
2519 if ((buffer_heads_over_limit
&& is_highmem_idx(i
)) ||
2520 !zone_watermark_ok_safe(zone
, testorder
,
2521 high_wmark_pages(zone
) + balance_gap
,
2523 shrink_zone(zone
, &sc
);
2525 reclaim_state
->reclaimed_slab
= 0;
2526 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2527 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2528 total_scanned
+= sc
.nr_scanned
;
2530 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2531 zone
->all_unreclaimable
= 1;
2535 * If we've done a decent amount of scanning and
2536 * the reclaim ratio is low, start doing writepage
2537 * even in laptop mode
2539 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2540 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2541 sc
.may_writepage
= 1;
2543 if (zone
->all_unreclaimable
) {
2544 if (end_zone
&& end_zone
== i
)
2549 if (!zone_watermark_ok_safe(zone
, testorder
,
2550 high_wmark_pages(zone
), end_zone
, 0)) {
2553 * We are still under min water mark. This
2554 * means that we have a GFP_ATOMIC allocation
2555 * failure risk. Hurry up!
2557 if (!zone_watermark_ok_safe(zone
, order
,
2558 min_wmark_pages(zone
), end_zone
, 0))
2559 has_under_min_watermark_zone
= 1;
2562 * If a zone reaches its high watermark,
2563 * consider it to be no longer congested. It's
2564 * possible there are dirty pages backed by
2565 * congested BDIs but as pressure is relieved,
2566 * spectulatively avoid congestion waits
2568 zone_clear_flag(zone
, ZONE_CONGESTED
);
2569 if (i
<= *classzone_idx
)
2570 balanced
+= zone
->present_pages
;
2574 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2575 break; /* kswapd: all done */
2577 * OK, kswapd is getting into trouble. Take a nap, then take
2578 * another pass across the zones.
2580 if (total_scanned
&& (sc
.priority
< DEF_PRIORITY
- 2)) {
2581 if (has_under_min_watermark_zone
)
2582 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2584 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2588 * We do this so kswapd doesn't build up large priorities for
2589 * example when it is freeing in parallel with allocators. It
2590 * matches the direct reclaim path behaviour in terms of impact
2591 * on zone->*_priority.
2593 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2595 } while (--sc
.priority
>= 0);
2599 * order-0: All zones must meet high watermark for a balanced node
2600 * high-order: Balanced zones must make up at least 25% of the node
2601 * for the node to be balanced
2603 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2609 * Fragmentation may mean that the system cannot be
2610 * rebalanced for high-order allocations in all zones.
2611 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2612 * it means the zones have been fully scanned and are still
2613 * not balanced. For high-order allocations, there is
2614 * little point trying all over again as kswapd may
2617 * Instead, recheck all watermarks at order-0 as they
2618 * are the most important. If watermarks are ok, kswapd will go
2619 * back to sleep. High-order users can still perform direct
2620 * reclaim if they wish.
2622 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2623 order
= sc
.order
= 0;
2629 * If kswapd was reclaiming at a higher order, it has the option of
2630 * sleeping without all zones being balanced. Before it does, it must
2631 * ensure that the watermarks for order-0 on *all* zones are met and
2632 * that the congestion flags are cleared. The congestion flag must
2633 * be cleared as kswapd is the only mechanism that clears the flag
2634 * and it is potentially going to sleep here.
2637 int zones_need_compaction
= 1;
2639 for (i
= 0; i
<= end_zone
; i
++) {
2640 struct zone
*zone
= pgdat
->node_zones
+ i
;
2642 if (!populated_zone(zone
))
2645 if (zone
->all_unreclaimable
&&
2646 sc
.priority
!= DEF_PRIORITY
)
2649 /* Would compaction fail due to lack of free memory? */
2650 if (COMPACTION_BUILD
&&
2651 compaction_suitable(zone
, order
) == COMPACT_SKIPPED
)
2654 /* Confirm the zone is balanced for order-0 */
2655 if (!zone_watermark_ok(zone
, 0,
2656 high_wmark_pages(zone
), 0, 0)) {
2657 order
= sc
.order
= 0;
2661 /* Check if the memory needs to be defragmented. */
2662 if (zone_watermark_ok(zone
, order
,
2663 low_wmark_pages(zone
), *classzone_idx
, 0))
2664 zones_need_compaction
= 0;
2666 /* If balanced, clear the congested flag */
2667 zone_clear_flag(zone
, ZONE_CONGESTED
);
2670 if (zones_need_compaction
)
2671 compact_pgdat(pgdat
, order
);
2675 * Return the order we were reclaiming at so sleeping_prematurely()
2676 * makes a decision on the order we were last reclaiming at. However,
2677 * if another caller entered the allocator slow path while kswapd
2678 * was awake, order will remain at the higher level
2680 *classzone_idx
= end_zone
;
2684 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2689 if (freezing(current
) || kthread_should_stop())
2692 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2694 /* Try to sleep for a short interval */
2695 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2696 remaining
= schedule_timeout(HZ
/10);
2697 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2698 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2702 * After a short sleep, check if it was a premature sleep. If not, then
2703 * go fully to sleep until explicitly woken up.
2705 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2706 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2709 * vmstat counters are not perfectly accurate and the estimated
2710 * value for counters such as NR_FREE_PAGES can deviate from the
2711 * true value by nr_online_cpus * threshold. To avoid the zone
2712 * watermarks being breached while under pressure, we reduce the
2713 * per-cpu vmstat threshold while kswapd is awake and restore
2714 * them before going back to sleep.
2716 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2718 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2721 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2723 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2725 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2729 * The background pageout daemon, started as a kernel thread
2730 * from the init process.
2732 * This basically trickles out pages so that we have _some_
2733 * free memory available even if there is no other activity
2734 * that frees anything up. This is needed for things like routing
2735 * etc, where we otherwise might have all activity going on in
2736 * asynchronous contexts that cannot page things out.
2738 * If there are applications that are active memory-allocators
2739 * (most normal use), this basically shouldn't matter.
2741 static int kswapd(void *p
)
2743 unsigned long order
, new_order
;
2744 unsigned balanced_order
;
2745 int classzone_idx
, new_classzone_idx
;
2746 int balanced_classzone_idx
;
2747 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2748 struct task_struct
*tsk
= current
;
2750 struct reclaim_state reclaim_state
= {
2751 .reclaimed_slab
= 0,
2753 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2755 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2757 if (!cpumask_empty(cpumask
))
2758 set_cpus_allowed_ptr(tsk
, cpumask
);
2759 current
->reclaim_state
= &reclaim_state
;
2762 * Tell the memory management that we're a "memory allocator",
2763 * and that if we need more memory we should get access to it
2764 * regardless (see "__alloc_pages()"). "kswapd" should
2765 * never get caught in the normal page freeing logic.
2767 * (Kswapd normally doesn't need memory anyway, but sometimes
2768 * you need a small amount of memory in order to be able to
2769 * page out something else, and this flag essentially protects
2770 * us from recursively trying to free more memory as we're
2771 * trying to free the first piece of memory in the first place).
2773 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2776 order
= new_order
= 0;
2778 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
2779 balanced_classzone_idx
= classzone_idx
;
2784 * If the last balance_pgdat was unsuccessful it's unlikely a
2785 * new request of a similar or harder type will succeed soon
2786 * so consider going to sleep on the basis we reclaimed at
2788 if (balanced_classzone_idx
>= new_classzone_idx
&&
2789 balanced_order
== new_order
) {
2790 new_order
= pgdat
->kswapd_max_order
;
2791 new_classzone_idx
= pgdat
->classzone_idx
;
2792 pgdat
->kswapd_max_order
= 0;
2793 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2796 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2798 * Don't sleep if someone wants a larger 'order'
2799 * allocation or has tigher zone constraints
2802 classzone_idx
= new_classzone_idx
;
2804 kswapd_try_to_sleep(pgdat
, balanced_order
,
2805 balanced_classzone_idx
);
2806 order
= pgdat
->kswapd_max_order
;
2807 classzone_idx
= pgdat
->classzone_idx
;
2809 new_classzone_idx
= classzone_idx
;
2810 pgdat
->kswapd_max_order
= 0;
2811 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2814 ret
= try_to_freeze();
2815 if (kthread_should_stop())
2819 * We can speed up thawing tasks if we don't call balance_pgdat
2820 * after returning from the refrigerator
2823 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2824 balanced_classzone_idx
= classzone_idx
;
2825 balanced_order
= balance_pgdat(pgdat
, order
,
2826 &balanced_classzone_idx
);
2833 * A zone is low on free memory, so wake its kswapd task to service it.
2835 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2839 if (!populated_zone(zone
))
2842 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2844 pgdat
= zone
->zone_pgdat
;
2845 if (pgdat
->kswapd_max_order
< order
) {
2846 pgdat
->kswapd_max_order
= order
;
2847 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2849 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2851 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2854 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2855 wake_up_interruptible(&pgdat
->kswapd_wait
);
2859 * The reclaimable count would be mostly accurate.
2860 * The less reclaimable pages may be
2861 * - mlocked pages, which will be moved to unevictable list when encountered
2862 * - mapped pages, which may require several travels to be reclaimed
2863 * - dirty pages, which is not "instantly" reclaimable
2865 unsigned long global_reclaimable_pages(void)
2869 nr
= global_page_state(NR_ACTIVE_FILE
) +
2870 global_page_state(NR_INACTIVE_FILE
);
2872 if (nr_swap_pages
> 0)
2873 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2874 global_page_state(NR_INACTIVE_ANON
);
2879 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2883 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2884 zone_page_state(zone
, NR_INACTIVE_FILE
);
2886 if (nr_swap_pages
> 0)
2887 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2888 zone_page_state(zone
, NR_INACTIVE_ANON
);
2893 #ifdef CONFIG_HIBERNATION
2895 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2898 * Rather than trying to age LRUs the aim is to preserve the overall
2899 * LRU order by reclaiming preferentially
2900 * inactive > active > active referenced > active mapped
2902 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2904 struct reclaim_state reclaim_state
;
2905 struct scan_control sc
= {
2906 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2910 .nr_to_reclaim
= nr_to_reclaim
,
2911 .hibernation_mode
= 1,
2913 .priority
= DEF_PRIORITY
,
2915 struct shrink_control shrink
= {
2916 .gfp_mask
= sc
.gfp_mask
,
2918 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2919 struct task_struct
*p
= current
;
2920 unsigned long nr_reclaimed
;
2922 p
->flags
|= PF_MEMALLOC
;
2923 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2924 reclaim_state
.reclaimed_slab
= 0;
2925 p
->reclaim_state
= &reclaim_state
;
2927 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2929 p
->reclaim_state
= NULL
;
2930 lockdep_clear_current_reclaim_state();
2931 p
->flags
&= ~PF_MEMALLOC
;
2933 return nr_reclaimed
;
2935 #endif /* CONFIG_HIBERNATION */
2937 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2938 not required for correctness. So if the last cpu in a node goes
2939 away, we get changed to run anywhere: as the first one comes back,
2940 restore their cpu bindings. */
2941 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2942 unsigned long action
, void *hcpu
)
2946 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2947 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2948 pg_data_t
*pgdat
= NODE_DATA(nid
);
2949 const struct cpumask
*mask
;
2951 mask
= cpumask_of_node(pgdat
->node_id
);
2953 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2954 /* One of our CPUs online: restore mask */
2955 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2962 * This kswapd start function will be called by init and node-hot-add.
2963 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2965 int kswapd_run(int nid
)
2967 pg_data_t
*pgdat
= NODE_DATA(nid
);
2973 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2974 if (IS_ERR(pgdat
->kswapd
)) {
2975 /* failure at boot is fatal */
2976 BUG_ON(system_state
== SYSTEM_BOOTING
);
2977 printk("Failed to start kswapd on node %d\n",nid
);
2984 * Called by memory hotplug when all memory in a node is offlined.
2986 void kswapd_stop(int nid
)
2988 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
2991 kthread_stop(kswapd
);
2994 static int __init
kswapd_init(void)
2999 for_each_node_state(nid
, N_HIGH_MEMORY
)
3001 hotcpu_notifier(cpu_callback
, 0);
3005 module_init(kswapd_init
)
3011 * If non-zero call zone_reclaim when the number of free pages falls below
3014 int zone_reclaim_mode __read_mostly
;
3016 #define RECLAIM_OFF 0
3017 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3018 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3019 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3022 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3023 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3026 #define ZONE_RECLAIM_PRIORITY 4
3029 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3032 int sysctl_min_unmapped_ratio
= 1;
3035 * If the number of slab pages in a zone grows beyond this percentage then
3036 * slab reclaim needs to occur.
3038 int sysctl_min_slab_ratio
= 5;
3040 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3042 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3043 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3044 zone_page_state(zone
, NR_ACTIVE_FILE
);
3047 * It's possible for there to be more file mapped pages than
3048 * accounted for by the pages on the file LRU lists because
3049 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3051 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3054 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3055 static long zone_pagecache_reclaimable(struct zone
*zone
)
3057 long nr_pagecache_reclaimable
;
3061 * If RECLAIM_SWAP is set, then all file pages are considered
3062 * potentially reclaimable. Otherwise, we have to worry about
3063 * pages like swapcache and zone_unmapped_file_pages() provides
3066 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3067 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3069 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3071 /* If we can't clean pages, remove dirty pages from consideration */
3072 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3073 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3075 /* Watch for any possible underflows due to delta */
3076 if (unlikely(delta
> nr_pagecache_reclaimable
))
3077 delta
= nr_pagecache_reclaimable
;
3079 return nr_pagecache_reclaimable
- delta
;
3083 * Try to free up some pages from this zone through reclaim.
3085 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3087 /* Minimum pages needed in order to stay on node */
3088 const unsigned long nr_pages
= 1 << order
;
3089 struct task_struct
*p
= current
;
3090 struct reclaim_state reclaim_state
;
3091 struct scan_control sc
= {
3092 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3093 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3095 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3097 .gfp_mask
= gfp_mask
,
3099 .priority
= ZONE_RECLAIM_PRIORITY
,
3101 struct shrink_control shrink
= {
3102 .gfp_mask
= sc
.gfp_mask
,
3104 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3108 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3109 * and we also need to be able to write out pages for RECLAIM_WRITE
3112 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3113 lockdep_set_current_reclaim_state(gfp_mask
);
3114 reclaim_state
.reclaimed_slab
= 0;
3115 p
->reclaim_state
= &reclaim_state
;
3117 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3119 * Free memory by calling shrink zone with increasing
3120 * priorities until we have enough memory freed.
3123 shrink_zone(zone
, &sc
);
3124 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3127 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3128 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3130 * shrink_slab() does not currently allow us to determine how
3131 * many pages were freed in this zone. So we take the current
3132 * number of slab pages and shake the slab until it is reduced
3133 * by the same nr_pages that we used for reclaiming unmapped
3136 * Note that shrink_slab will free memory on all zones and may
3140 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3142 /* No reclaimable slab or very low memory pressure */
3143 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3146 /* Freed enough memory */
3147 nr_slab_pages1
= zone_page_state(zone
,
3148 NR_SLAB_RECLAIMABLE
);
3149 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3154 * Update nr_reclaimed by the number of slab pages we
3155 * reclaimed from this zone.
3157 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3158 if (nr_slab_pages1
< nr_slab_pages0
)
3159 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3162 p
->reclaim_state
= NULL
;
3163 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3164 lockdep_clear_current_reclaim_state();
3165 return sc
.nr_reclaimed
>= nr_pages
;
3168 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3174 * Zone reclaim reclaims unmapped file backed pages and
3175 * slab pages if we are over the defined limits.
3177 * A small portion of unmapped file backed pages is needed for
3178 * file I/O otherwise pages read by file I/O will be immediately
3179 * thrown out if the zone is overallocated. So we do not reclaim
3180 * if less than a specified percentage of the zone is used by
3181 * unmapped file backed pages.
3183 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3184 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3185 return ZONE_RECLAIM_FULL
;
3187 if (zone
->all_unreclaimable
)
3188 return ZONE_RECLAIM_FULL
;
3191 * Do not scan if the allocation should not be delayed.
3193 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3194 return ZONE_RECLAIM_NOSCAN
;
3197 * Only run zone reclaim on the local zone or on zones that do not
3198 * have associated processors. This will favor the local processor
3199 * over remote processors and spread off node memory allocations
3200 * as wide as possible.
3202 node_id
= zone_to_nid(zone
);
3203 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3204 return ZONE_RECLAIM_NOSCAN
;
3206 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3207 return ZONE_RECLAIM_NOSCAN
;
3209 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3210 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3213 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3220 * page_evictable - test whether a page is evictable
3221 * @page: the page to test
3222 * @vma: the VMA in which the page is or will be mapped, may be NULL
3224 * Test whether page is evictable--i.e., should be placed on active/inactive
3225 * lists vs unevictable list. The vma argument is !NULL when called from the
3226 * fault path to determine how to instantate a new page.
3228 * Reasons page might not be evictable:
3229 * (1) page's mapping marked unevictable
3230 * (2) page is part of an mlocked VMA
3233 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3236 if (mapping_unevictable(page_mapping(page
)))
3239 if (PageMlocked(page
) || (vma
&& mlocked_vma_newpage(vma
, page
)))
3247 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3248 * @pages: array of pages to check
3249 * @nr_pages: number of pages to check
3251 * Checks pages for evictability and moves them to the appropriate lru list.
3253 * This function is only used for SysV IPC SHM_UNLOCK.
3255 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3257 struct lruvec
*lruvec
;
3258 struct zone
*zone
= NULL
;
3263 for (i
= 0; i
< nr_pages
; i
++) {
3264 struct page
*page
= pages
[i
];
3265 struct zone
*pagezone
;
3268 pagezone
= page_zone(page
);
3269 if (pagezone
!= zone
) {
3271 spin_unlock_irq(&zone
->lru_lock
);
3273 spin_lock_irq(&zone
->lru_lock
);
3276 if (!PageLRU(page
) || !PageUnevictable(page
))
3279 if (page_evictable(page
, NULL
)) {
3280 enum lru_list lru
= page_lru_base_type(page
);
3282 VM_BUG_ON(PageActive(page
));
3283 ClearPageUnevictable(page
);
3284 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3285 lruvec
= mem_cgroup_lru_move_lists(zone
, page
,
3286 LRU_UNEVICTABLE
, lru
);
3287 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
3288 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ lru
);
3294 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3295 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3296 spin_unlock_irq(&zone
->lru_lock
);
3299 #endif /* CONFIG_SHMEM */
3301 static void warn_scan_unevictable_pages(void)
3303 printk_once(KERN_WARNING
3304 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3305 "disabled for lack of a legitimate use case. If you have "
3306 "one, please send an email to linux-mm@kvack.org.\n",
3311 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3312 * all nodes' unevictable lists for evictable pages
3314 unsigned long scan_unevictable_pages
;
3316 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3317 void __user
*buffer
,
3318 size_t *length
, loff_t
*ppos
)
3320 warn_scan_unevictable_pages();
3321 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3322 scan_unevictable_pages
= 0;
3328 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3329 * a specified node's per zone unevictable lists for evictable pages.
3332 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3333 struct device_attribute
*attr
,
3336 warn_scan_unevictable_pages();
3337 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3340 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3341 struct device_attribute
*attr
,
3342 const char *buf
, size_t count
)
3344 warn_scan_unevictable_pages();
3349 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3350 read_scan_unevictable_node
,
3351 write_scan_unevictable_node
);
3353 int scan_unevictable_register_node(struct node
*node
)
3355 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3358 void scan_unevictable_unregister_node(struct node
*node
)
3360 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);