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/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29 #include <linux/mm_inline.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 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned
;
61 /* Number of pages freed so far during a call to shrink_zones() */
62 unsigned long nr_reclaimed
;
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim
;
67 unsigned long hibernation_mode
;
69 /* This context's GFP mask */
74 /* Can mapped pages be reclaimed? */
77 /* Can pages be swapped as part of reclaim? */
82 /* Scan (total_size >> priority) pages at once */
86 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
89 struct mem_cgroup
*target_mem_cgroup
;
92 * Nodemask of nodes allowed by the caller. If NULL, all nodes
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
129 * From 0 .. 100. Higher means more swappy.
131 int vm_swappiness
= 60;
132 unsigned long vm_total_pages
; /* The total number of pages which the VM controls */
134 static LIST_HEAD(shrinker_list
);
135 static DECLARE_RWSEM(shrinker_rwsem
);
138 static bool global_reclaim(struct scan_control
*sc
)
140 return !sc
->target_mem_cgroup
;
143 static bool mem_cgroup_should_soft_reclaim(struct scan_control
*sc
)
145 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
146 return !mem_cgroup_disabled() &&
147 mem_cgroup_soft_reclaim_eligible(root
, root
) != SKIP_TREE
;
150 static bool global_reclaim(struct scan_control
*sc
)
155 static bool mem_cgroup_should_soft_reclaim(struct scan_control
*sc
)
161 unsigned long zone_reclaimable_pages(struct zone
*zone
)
165 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
166 zone_page_state(zone
, NR_INACTIVE_FILE
);
168 if (get_nr_swap_pages() > 0)
169 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
170 zone_page_state(zone
, NR_INACTIVE_ANON
);
175 bool zone_reclaimable(struct zone
*zone
)
177 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
180 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
182 if (!mem_cgroup_disabled())
183 return mem_cgroup_get_lru_size(lruvec
, lru
);
185 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
189 * Add a shrinker callback to be called from the vm
191 void register_shrinker(struct shrinker
*shrinker
)
193 atomic_long_set(&shrinker
->nr_in_batch
, 0);
194 down_write(&shrinker_rwsem
);
195 list_add_tail(&shrinker
->list
, &shrinker_list
);
196 up_write(&shrinker_rwsem
);
198 EXPORT_SYMBOL(register_shrinker
);
203 void unregister_shrinker(struct shrinker
*shrinker
)
205 down_write(&shrinker_rwsem
);
206 list_del(&shrinker
->list
);
207 up_write(&shrinker_rwsem
);
209 EXPORT_SYMBOL(unregister_shrinker
);
211 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
212 struct shrink_control
*sc
,
213 unsigned long nr_to_scan
)
215 sc
->nr_to_scan
= nr_to_scan
;
216 return (*shrinker
->shrink
)(shrinker
, sc
);
219 #define SHRINK_BATCH 128
221 * Call the shrink functions to age shrinkable caches
223 * Here we assume it costs one seek to replace a lru page and that it also
224 * takes a seek to recreate a cache object. With this in mind we age equal
225 * percentages of the lru and ageable caches. This should balance the seeks
226 * generated by these structures.
228 * If the vm encountered mapped pages on the LRU it increase the pressure on
229 * slab to avoid swapping.
231 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
233 * `lru_pages' represents the number of on-LRU pages in all the zones which
234 * are eligible for the caller's allocation attempt. It is used for balancing
235 * slab reclaim versus page reclaim.
237 * Returns the number of slab objects which we shrunk.
239 unsigned long shrink_slab(struct shrink_control
*shrink
,
240 unsigned long nr_pages_scanned
,
241 unsigned long lru_pages
)
243 struct shrinker
*shrinker
;
244 unsigned long ret
= 0;
246 if (nr_pages_scanned
== 0)
247 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
249 if (!down_read_trylock(&shrinker_rwsem
)) {
250 /* Assume we'll be able to shrink next time */
255 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
256 unsigned long long delta
;
262 long batch_size
= shrinker
->batch
? shrinker
->batch
265 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
270 * copy the current shrinker scan count into a local variable
271 * and zero it so that other concurrent shrinker invocations
272 * don't also do this scanning work.
274 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
277 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
279 do_div(delta
, lru_pages
+ 1);
281 if (total_scan
< 0) {
282 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
284 shrinker
->shrink
, total_scan
);
285 total_scan
= max_pass
;
289 * We need to avoid excessive windup on filesystem shrinkers
290 * due to large numbers of GFP_NOFS allocations causing the
291 * shrinkers to return -1 all the time. This results in a large
292 * nr being built up so when a shrink that can do some work
293 * comes along it empties the entire cache due to nr >>>
294 * max_pass. This is bad for sustaining a working set in
297 * Hence only allow the shrinker to scan the entire cache when
298 * a large delta change is calculated directly.
300 if (delta
< max_pass
/ 4)
301 total_scan
= min(total_scan
, max_pass
/ 2);
304 * Avoid risking looping forever due to too large nr value:
305 * never try to free more than twice the estimate number of
308 if (total_scan
> max_pass
* 2)
309 total_scan
= max_pass
* 2;
311 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
312 nr_pages_scanned
, lru_pages
,
313 max_pass
, delta
, total_scan
);
315 while (total_scan
>= batch_size
) {
318 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
319 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
321 if (shrink_ret
== -1)
323 if (shrink_ret
< nr_before
)
324 ret
+= nr_before
- shrink_ret
;
325 count_vm_events(SLABS_SCANNED
, batch_size
);
326 total_scan
-= batch_size
;
332 * move the unused scan count back into the shrinker in a
333 * manner that handles concurrent updates. If we exhausted the
334 * scan, there is no need to do an update.
337 new_nr
= atomic_long_add_return(total_scan
,
338 &shrinker
->nr_in_batch
);
340 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
342 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
344 up_read(&shrinker_rwsem
);
350 static inline int is_page_cache_freeable(struct page
*page
)
353 * A freeable page cache page is referenced only by the caller
354 * that isolated the page, the page cache radix tree and
355 * optional buffer heads at page->private.
357 return page_count(page
) - page_has_private(page
) == 2;
360 static int may_write_to_queue(struct backing_dev_info
*bdi
,
361 struct scan_control
*sc
)
363 if (current
->flags
& PF_SWAPWRITE
)
365 if (!bdi_write_congested(bdi
))
367 if (bdi
== current
->backing_dev_info
)
373 * We detected a synchronous write error writing a page out. Probably
374 * -ENOSPC. We need to propagate that into the address_space for a subsequent
375 * fsync(), msync() or close().
377 * The tricky part is that after writepage we cannot touch the mapping: nothing
378 * prevents it from being freed up. But we have a ref on the page and once
379 * that page is locked, the mapping is pinned.
381 * We're allowed to run sleeping lock_page() here because we know the caller has
384 static void handle_write_error(struct address_space
*mapping
,
385 struct page
*page
, int error
)
388 if (page_mapping(page
) == mapping
)
389 mapping_set_error(mapping
, error
);
393 /* possible outcome of pageout() */
395 /* failed to write page out, page is locked */
397 /* move page to the active list, page is locked */
399 /* page has been sent to the disk successfully, page is unlocked */
401 /* page is clean and locked */
406 * pageout is called by shrink_page_list() for each dirty page.
407 * Calls ->writepage().
409 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
410 struct scan_control
*sc
)
413 * If the page is dirty, only perform writeback if that write
414 * will be non-blocking. To prevent this allocation from being
415 * stalled by pagecache activity. But note that there may be
416 * stalls if we need to run get_block(). We could test
417 * PagePrivate for that.
419 * If this process is currently in __generic_file_aio_write() against
420 * this page's queue, we can perform writeback even if that
423 * If the page is swapcache, write it back even if that would
424 * block, for some throttling. This happens by accident, because
425 * swap_backing_dev_info is bust: it doesn't reflect the
426 * congestion state of the swapdevs. Easy to fix, if needed.
428 if (!is_page_cache_freeable(page
))
432 * Some data journaling orphaned pages can have
433 * page->mapping == NULL while being dirty with clean buffers.
435 if (page_has_private(page
)) {
436 if (try_to_free_buffers(page
)) {
437 ClearPageDirty(page
);
438 printk("%s: orphaned page\n", __func__
);
444 if (mapping
->a_ops
->writepage
== NULL
)
445 return PAGE_ACTIVATE
;
446 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
449 if (clear_page_dirty_for_io(page
)) {
451 struct writeback_control wbc
= {
452 .sync_mode
= WB_SYNC_NONE
,
453 .nr_to_write
= SWAP_CLUSTER_MAX
,
455 .range_end
= LLONG_MAX
,
459 SetPageReclaim(page
);
460 res
= mapping
->a_ops
->writepage(page
, &wbc
);
462 handle_write_error(mapping
, page
, res
);
463 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
464 ClearPageReclaim(page
);
465 return PAGE_ACTIVATE
;
468 if (!PageWriteback(page
)) {
469 /* synchronous write or broken a_ops? */
470 ClearPageReclaim(page
);
472 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
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 was_unevictable
= PageUnevictable(page
);
582 VM_BUG_ON(PageLRU(page
));
585 ClearPageUnevictable(page
);
587 if (page_evictable(page
)) {
589 * For evictable pages, we can use the cache.
590 * In event of a race, worst case is we end up with an
591 * unevictable page on [in]active list.
592 * We know how to handle that.
594 is_unevictable
= false;
598 * Put unevictable pages directly on zone's unevictable
601 is_unevictable
= true;
602 add_page_to_unevictable_list(page
);
604 * When racing with an mlock or AS_UNEVICTABLE clearing
605 * (page is unlocked) make sure that if the other thread
606 * does not observe our setting of PG_lru and fails
607 * isolation/check_move_unevictable_pages,
608 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
609 * the page back to the evictable list.
611 * The other side is TestClearPageMlocked() or shmem_lock().
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 (is_unevictable
&& page_evictable(page
)) {
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
&& !is_unevictable
)
633 count_vm_event(UNEVICTABLE_PGRESCUED
);
634 else if (!was_unevictable
&& is_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
->target_mem_cgroup
,
655 referenced_page
= TestClearPageReferenced(page
);
658 * Mlock lost the isolation race with us. Let try_to_unmap()
659 * move the page to the unevictable list.
661 if (vm_flags
& VM_LOCKED
)
662 return PAGEREF_RECLAIM
;
664 if (referenced_ptes
) {
665 if (PageSwapBacked(page
))
666 return PAGEREF_ACTIVATE
;
668 * All mapped pages start out with page table
669 * references from the instantiating fault, so we need
670 * to look twice if a mapped file page is used more
673 * Mark it and spare it for another trip around the
674 * inactive list. Another page table reference will
675 * lead to its activation.
677 * Note: the mark is set for activated pages as well
678 * so that recently deactivated but used pages are
681 SetPageReferenced(page
);
683 if (referenced_page
|| referenced_ptes
> 1)
684 return PAGEREF_ACTIVATE
;
687 * Activate file-backed executable pages after first usage.
689 if (vm_flags
& VM_EXEC
)
690 return PAGEREF_ACTIVATE
;
695 /* Reclaim if clean, defer dirty pages to writeback */
696 if (referenced_page
&& !PageSwapBacked(page
))
697 return PAGEREF_RECLAIM_CLEAN
;
699 return PAGEREF_RECLAIM
;
702 /* Check if a page is dirty or under writeback */
703 static void page_check_dirty_writeback(struct page
*page
,
704 bool *dirty
, bool *writeback
)
706 struct address_space
*mapping
;
709 * Anonymous pages are not handled by flushers and must be written
710 * from reclaim context. Do not stall reclaim based on them
712 if (!page_is_file_cache(page
)) {
718 /* By default assume that the page flags are accurate */
719 *dirty
= PageDirty(page
);
720 *writeback
= PageWriteback(page
);
722 /* Verify dirty/writeback state if the filesystem supports it */
723 if (!page_has_private(page
))
726 mapping
= page_mapping(page
);
727 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
728 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
732 * shrink_page_list() returns the number of reclaimed pages
734 static unsigned long shrink_page_list(struct list_head
*page_list
,
736 struct scan_control
*sc
,
737 enum ttu_flags ttu_flags
,
738 unsigned long *ret_nr_dirty
,
739 unsigned long *ret_nr_unqueued_dirty
,
740 unsigned long *ret_nr_congested
,
741 unsigned long *ret_nr_writeback
,
742 unsigned long *ret_nr_immediate
,
745 LIST_HEAD(ret_pages
);
746 LIST_HEAD(free_pages
);
748 unsigned long nr_unqueued_dirty
= 0;
749 unsigned long nr_dirty
= 0;
750 unsigned long nr_congested
= 0;
751 unsigned long nr_reclaimed
= 0;
752 unsigned long nr_writeback
= 0;
753 unsigned long nr_immediate
= 0;
757 mem_cgroup_uncharge_start();
758 while (!list_empty(page_list
)) {
759 struct address_space
*mapping
;
762 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
763 bool dirty
, writeback
;
767 page
= lru_to_page(page_list
);
768 list_del(&page
->lru
);
770 if (!trylock_page(page
))
773 VM_BUG_ON(PageActive(page
));
774 VM_BUG_ON(page_zone(page
) != zone
);
778 if (unlikely(!page_evictable(page
)))
781 if (!sc
->may_unmap
&& page_mapped(page
))
784 /* Double the slab pressure for mapped and swapcache pages */
785 if (page_mapped(page
) || PageSwapCache(page
))
788 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
789 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
792 * The number of dirty pages determines if a zone is marked
793 * reclaim_congested which affects wait_iff_congested. kswapd
794 * will stall and start writing pages if the tail of the LRU
795 * is all dirty unqueued pages.
797 page_check_dirty_writeback(page
, &dirty
, &writeback
);
798 if (dirty
|| writeback
)
801 if (dirty
&& !writeback
)
805 * Treat this page as congested if the underlying BDI is or if
806 * pages are cycling through the LRU so quickly that the
807 * pages marked for immediate reclaim are making it to the
808 * end of the LRU a second time.
810 mapping
= page_mapping(page
);
811 if ((mapping
&& bdi_write_congested(mapping
->backing_dev_info
)) ||
812 (writeback
&& PageReclaim(page
)))
816 * If a page at the tail of the LRU is under writeback, there
817 * are three cases to consider.
819 * 1) If reclaim is encountering an excessive number of pages
820 * under writeback and this page is both under writeback and
821 * PageReclaim then it indicates that pages are being queued
822 * for IO but are being recycled through the LRU before the
823 * IO can complete. Waiting on the page itself risks an
824 * indefinite stall if it is impossible to writeback the
825 * page due to IO error or disconnected storage so instead
826 * note that the LRU is being scanned too quickly and the
827 * caller can stall after page list has been processed.
829 * 2) Global reclaim encounters a page, memcg encounters a
830 * page that is not marked for immediate reclaim or
831 * the caller does not have __GFP_IO. In this case mark
832 * the page for immediate reclaim and continue scanning.
834 * __GFP_IO is checked because a loop driver thread might
835 * enter reclaim, and deadlock if it waits on a page for
836 * which it is needed to do the write (loop masks off
837 * __GFP_IO|__GFP_FS for this reason); but more thought
838 * would probably show more reasons.
840 * Don't require __GFP_FS, since we're not going into the
841 * FS, just waiting on its writeback completion. Worryingly,
842 * ext4 gfs2 and xfs allocate pages with
843 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
844 * may_enter_fs here is liable to OOM on them.
846 * 3) memcg encounters a page that is not already marked
847 * PageReclaim. memcg does not have any dirty pages
848 * throttling so we could easily OOM just because too many
849 * pages are in writeback and there is nothing else to
850 * reclaim. Wait for the writeback to complete.
852 if (PageWriteback(page
)) {
854 if (current_is_kswapd() &&
856 zone_is_reclaim_writeback(zone
)) {
861 } else if (global_reclaim(sc
) ||
862 !PageReclaim(page
) || !(sc
->gfp_mask
& __GFP_IO
)) {
864 * This is slightly racy - end_page_writeback()
865 * might have just cleared PageReclaim, then
866 * setting PageReclaim here end up interpreted
867 * as PageReadahead - but that does not matter
868 * enough to care. What we do want is for this
869 * page to have PageReclaim set next time memcg
870 * reclaim reaches the tests above, so it will
871 * then wait_on_page_writeback() to avoid OOM;
872 * and it's also appropriate in global reclaim.
874 SetPageReclaim(page
);
881 wait_on_page_writeback(page
);
886 references
= page_check_references(page
, sc
);
888 switch (references
) {
889 case PAGEREF_ACTIVATE
:
890 goto activate_locked
;
893 case PAGEREF_RECLAIM
:
894 case PAGEREF_RECLAIM_CLEAN
:
895 ; /* try to reclaim the page below */
899 * Anonymous process memory has backing store?
900 * Try to allocate it some swap space here.
902 if (PageAnon(page
) && !PageSwapCache(page
)) {
903 if (!(sc
->gfp_mask
& __GFP_IO
))
905 if (!add_to_swap(page
, page_list
))
906 goto activate_locked
;
909 /* Adding to swap updated mapping */
910 mapping
= page_mapping(page
);
914 * The page is mapped into the page tables of one or more
915 * processes. Try to unmap it here.
917 if (page_mapped(page
) && mapping
) {
918 switch (try_to_unmap(page
, ttu_flags
)) {
920 goto activate_locked
;
926 ; /* try to free the page below */
930 if (PageDirty(page
)) {
932 * Only kswapd can writeback filesystem pages to
933 * avoid risk of stack overflow but only writeback
934 * if many dirty pages have been encountered.
936 if (page_is_file_cache(page
) &&
937 (!current_is_kswapd() ||
938 !zone_is_reclaim_dirty(zone
))) {
940 * Immediately reclaim when written back.
941 * Similar in principal to deactivate_page()
942 * except we already have the page isolated
943 * and know it's dirty
945 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
946 SetPageReclaim(page
);
951 if (references
== PAGEREF_RECLAIM_CLEAN
)
955 if (!sc
->may_writepage
)
958 /* Page is dirty, try to write it out here */
959 switch (pageout(page
, mapping
, sc
)) {
963 goto activate_locked
;
965 if (PageWriteback(page
))
971 * A synchronous write - probably a ramdisk. Go
972 * ahead and try to reclaim the page.
974 if (!trylock_page(page
))
976 if (PageDirty(page
) || PageWriteback(page
))
978 mapping
= page_mapping(page
);
980 ; /* try to free the page below */
985 * If the page has buffers, try to free the buffer mappings
986 * associated with this page. If we succeed we try to free
989 * We do this even if the page is PageDirty().
990 * try_to_release_page() does not perform I/O, but it is
991 * possible for a page to have PageDirty set, but it is actually
992 * clean (all its buffers are clean). This happens if the
993 * buffers were written out directly, with submit_bh(). ext3
994 * will do this, as well as the blockdev mapping.
995 * try_to_release_page() will discover that cleanness and will
996 * drop the buffers and mark the page clean - it can be freed.
998 * Rarely, pages can have buffers and no ->mapping. These are
999 * the pages which were not successfully invalidated in
1000 * truncate_complete_page(). We try to drop those buffers here
1001 * and if that worked, and the page is no longer mapped into
1002 * process address space (page_count == 1) it can be freed.
1003 * Otherwise, leave the page on the LRU so it is swappable.
1005 if (page_has_private(page
)) {
1006 if (!try_to_release_page(page
, sc
->gfp_mask
))
1007 goto activate_locked
;
1008 if (!mapping
&& page_count(page
) == 1) {
1010 if (put_page_testzero(page
))
1014 * rare race with speculative reference.
1015 * the speculative reference will free
1016 * this page shortly, so we may
1017 * increment nr_reclaimed here (and
1018 * leave it off the LRU).
1026 if (!mapping
|| !__remove_mapping(mapping
, page
))
1030 * At this point, we have no other references and there is
1031 * no way to pick any more up (removed from LRU, removed
1032 * from pagecache). Can use non-atomic bitops now (and
1033 * we obviously don't have to worry about waking up a process
1034 * waiting on the page lock, because there are no references.
1036 __clear_page_locked(page
);
1041 * Is there need to periodically free_page_list? It would
1042 * appear not as the counts should be low
1044 list_add(&page
->lru
, &free_pages
);
1048 if (PageSwapCache(page
))
1049 try_to_free_swap(page
);
1051 putback_lru_page(page
);
1055 /* Not a candidate for swapping, so reclaim swap space. */
1056 if (PageSwapCache(page
) && vm_swap_full())
1057 try_to_free_swap(page
);
1058 VM_BUG_ON(PageActive(page
));
1059 SetPageActive(page
);
1064 list_add(&page
->lru
, &ret_pages
);
1065 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
1068 free_hot_cold_page_list(&free_pages
, 1);
1070 list_splice(&ret_pages
, page_list
);
1071 count_vm_events(PGACTIVATE
, pgactivate
);
1072 mem_cgroup_uncharge_end();
1073 *ret_nr_dirty
+= nr_dirty
;
1074 *ret_nr_congested
+= nr_congested
;
1075 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1076 *ret_nr_writeback
+= nr_writeback
;
1077 *ret_nr_immediate
+= nr_immediate
;
1078 return nr_reclaimed
;
1081 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1082 struct list_head
*page_list
)
1084 struct scan_control sc
= {
1085 .gfp_mask
= GFP_KERNEL
,
1086 .priority
= DEF_PRIORITY
,
1089 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1090 struct page
*page
, *next
;
1091 LIST_HEAD(clean_pages
);
1093 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1094 if (page_is_file_cache(page
) && !PageDirty(page
)) {
1095 ClearPageActive(page
);
1096 list_move(&page
->lru
, &clean_pages
);
1100 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1101 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1102 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1103 list_splice(&clean_pages
, page_list
);
1104 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1109 * Attempt to remove the specified page from its LRU. Only take this page
1110 * if it is of the appropriate PageActive status. Pages which are being
1111 * freed elsewhere are also ignored.
1113 * page: page to consider
1114 * mode: one of the LRU isolation modes defined above
1116 * returns 0 on success, -ve errno on failure.
1118 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1122 /* Only take pages on the LRU. */
1126 /* Compaction should not handle unevictable pages but CMA can do so */
1127 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1133 * To minimise LRU disruption, the caller can indicate that it only
1134 * wants to isolate pages it will be able to operate on without
1135 * blocking - clean pages for the most part.
1137 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1138 * is used by reclaim when it is cannot write to backing storage
1140 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1141 * that it is possible to migrate without blocking
1143 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1144 /* All the caller can do on PageWriteback is block */
1145 if (PageWriteback(page
))
1148 if (PageDirty(page
)) {
1149 struct address_space
*mapping
;
1151 /* ISOLATE_CLEAN means only clean pages */
1152 if (mode
& ISOLATE_CLEAN
)
1156 * Only pages without mappings or that have a
1157 * ->migratepage callback are possible to migrate
1160 mapping
= page_mapping(page
);
1161 if (mapping
&& !mapping
->a_ops
->migratepage
)
1166 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1169 if (likely(get_page_unless_zero(page
))) {
1171 * Be careful not to clear PageLRU until after we're
1172 * sure the page is not being freed elsewhere -- the
1173 * page release code relies on it.
1183 * zone->lru_lock is heavily contended. Some of the functions that
1184 * shrink the lists perform better by taking out a batch of pages
1185 * and working on them outside the LRU lock.
1187 * For pagecache intensive workloads, this function is the hottest
1188 * spot in the kernel (apart from copy_*_user functions).
1190 * Appropriate locks must be held before calling this function.
1192 * @nr_to_scan: The number of pages to look through on the list.
1193 * @lruvec: The LRU vector to pull pages from.
1194 * @dst: The temp list to put pages on to.
1195 * @nr_scanned: The number of pages that were scanned.
1196 * @sc: The scan_control struct for this reclaim session
1197 * @mode: One of the LRU isolation modes
1198 * @lru: LRU list id for isolating
1200 * returns how many pages were moved onto *@dst.
1202 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1203 struct lruvec
*lruvec
, struct list_head
*dst
,
1204 unsigned long *nr_scanned
, struct scan_control
*sc
,
1205 isolate_mode_t mode
, enum lru_list lru
)
1207 struct list_head
*src
= &lruvec
->lists
[lru
];
1208 unsigned long nr_taken
= 0;
1211 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1215 page
= lru_to_page(src
);
1216 prefetchw_prev_lru_page(page
, src
, flags
);
1218 VM_BUG_ON(!PageLRU(page
));
1220 switch (__isolate_lru_page(page
, mode
)) {
1222 nr_pages
= hpage_nr_pages(page
);
1223 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1224 list_move(&page
->lru
, dst
);
1225 nr_taken
+= nr_pages
;
1229 /* else it is being freed elsewhere */
1230 list_move(&page
->lru
, src
);
1239 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1240 nr_taken
, mode
, is_file_lru(lru
));
1245 * isolate_lru_page - tries to isolate a page from its LRU list
1246 * @page: page to isolate from its LRU list
1248 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1249 * vmstat statistic corresponding to whatever LRU list the page was on.
1251 * Returns 0 if the page was removed from an LRU list.
1252 * Returns -EBUSY if the page was not on an LRU list.
1254 * The returned page will have PageLRU() cleared. If it was found on
1255 * the active list, it will have PageActive set. If it was found on
1256 * the unevictable list, it will have the PageUnevictable bit set. That flag
1257 * may need to be cleared by the caller before letting the page go.
1259 * The vmstat statistic corresponding to the list on which the page was
1260 * found will be decremented.
1263 * (1) Must be called with an elevated refcount on the page. This is a
1264 * fundamentnal difference from isolate_lru_pages (which is called
1265 * without a stable reference).
1266 * (2) the lru_lock must not be held.
1267 * (3) interrupts must be enabled.
1269 int isolate_lru_page(struct page
*page
)
1273 VM_BUG_ON(!page_count(page
));
1275 if (PageLRU(page
)) {
1276 struct zone
*zone
= page_zone(page
);
1277 struct lruvec
*lruvec
;
1279 spin_lock_irq(&zone
->lru_lock
);
1280 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1281 if (PageLRU(page
)) {
1282 int lru
= page_lru(page
);
1285 del_page_from_lru_list(page
, lruvec
, lru
);
1288 spin_unlock_irq(&zone
->lru_lock
);
1294 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1295 * then get resheduled. When there are massive number of tasks doing page
1296 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1297 * the LRU list will go small and be scanned faster than necessary, leading to
1298 * unnecessary swapping, thrashing and OOM.
1300 static int too_many_isolated(struct zone
*zone
, int file
,
1301 struct scan_control
*sc
)
1303 unsigned long inactive
, isolated
;
1305 if (current_is_kswapd())
1308 if (!global_reclaim(sc
))
1312 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1313 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1315 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1316 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1320 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1321 * won't get blocked by normal direct-reclaimers, forming a circular
1324 if ((sc
->gfp_mask
& GFP_IOFS
) == GFP_IOFS
)
1327 return isolated
> inactive
;
1330 static noinline_for_stack
void
1331 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1333 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1334 struct zone
*zone
= lruvec_zone(lruvec
);
1335 LIST_HEAD(pages_to_free
);
1338 * Put back any unfreeable pages.
1340 while (!list_empty(page_list
)) {
1341 struct page
*page
= lru_to_page(page_list
);
1344 VM_BUG_ON(PageLRU(page
));
1345 list_del(&page
->lru
);
1346 if (unlikely(!page_evictable(page
))) {
1347 spin_unlock_irq(&zone
->lru_lock
);
1348 putback_lru_page(page
);
1349 spin_lock_irq(&zone
->lru_lock
);
1353 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1356 lru
= page_lru(page
);
1357 add_page_to_lru_list(page
, lruvec
, lru
);
1359 if (is_active_lru(lru
)) {
1360 int file
= is_file_lru(lru
);
1361 int numpages
= hpage_nr_pages(page
);
1362 reclaim_stat
->recent_rotated
[file
] += numpages
;
1364 if (put_page_testzero(page
)) {
1365 __ClearPageLRU(page
);
1366 __ClearPageActive(page
);
1367 del_page_from_lru_list(page
, lruvec
, lru
);
1369 if (unlikely(PageCompound(page
))) {
1370 spin_unlock_irq(&zone
->lru_lock
);
1371 (*get_compound_page_dtor(page
))(page
);
1372 spin_lock_irq(&zone
->lru_lock
);
1374 list_add(&page
->lru
, &pages_to_free
);
1379 * To save our caller's stack, now use input list for pages to free.
1381 list_splice(&pages_to_free
, page_list
);
1385 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1386 * of reclaimed pages
1388 static noinline_for_stack
unsigned long
1389 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1390 struct scan_control
*sc
, enum lru_list lru
)
1392 LIST_HEAD(page_list
);
1393 unsigned long nr_scanned
;
1394 unsigned long nr_reclaimed
= 0;
1395 unsigned long nr_taken
;
1396 unsigned long nr_dirty
= 0;
1397 unsigned long nr_congested
= 0;
1398 unsigned long nr_unqueued_dirty
= 0;
1399 unsigned long nr_writeback
= 0;
1400 unsigned long nr_immediate
= 0;
1401 isolate_mode_t isolate_mode
= 0;
1402 int file
= is_file_lru(lru
);
1403 struct zone
*zone
= lruvec_zone(lruvec
);
1404 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1406 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1407 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1409 /* We are about to die and free our memory. Return now. */
1410 if (fatal_signal_pending(current
))
1411 return SWAP_CLUSTER_MAX
;
1417 isolate_mode
|= ISOLATE_UNMAPPED
;
1418 if (!sc
->may_writepage
)
1419 isolate_mode
|= ISOLATE_CLEAN
;
1421 spin_lock_irq(&zone
->lru_lock
);
1423 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1424 &nr_scanned
, sc
, isolate_mode
, lru
);
1426 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1427 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1429 if (global_reclaim(sc
)) {
1430 zone
->pages_scanned
+= nr_scanned
;
1431 if (current_is_kswapd())
1432 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1434 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1436 spin_unlock_irq(&zone
->lru_lock
);
1441 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1442 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1443 &nr_writeback
, &nr_immediate
,
1446 spin_lock_irq(&zone
->lru_lock
);
1448 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1450 if (global_reclaim(sc
)) {
1451 if (current_is_kswapd())
1452 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1455 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1459 putback_inactive_pages(lruvec
, &page_list
);
1461 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1463 spin_unlock_irq(&zone
->lru_lock
);
1465 free_hot_cold_page_list(&page_list
, 1);
1468 * If reclaim is isolating dirty pages under writeback, it implies
1469 * that the long-lived page allocation rate is exceeding the page
1470 * laundering rate. Either the global limits are not being effective
1471 * at throttling processes due to the page distribution throughout
1472 * zones or there is heavy usage of a slow backing device. The
1473 * only option is to throttle from reclaim context which is not ideal
1474 * as there is no guarantee the dirtying process is throttled in the
1475 * same way balance_dirty_pages() manages.
1477 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1478 * of pages under pages flagged for immediate reclaim and stall if any
1479 * are encountered in the nr_immediate check below.
1481 if (nr_writeback
&& nr_writeback
== nr_taken
)
1482 zone_set_flag(zone
, ZONE_WRITEBACK
);
1485 * memcg will stall in page writeback so only consider forcibly
1486 * stalling for global reclaim
1488 if (global_reclaim(sc
)) {
1490 * Tag a zone as congested if all the dirty pages scanned were
1491 * backed by a congested BDI and wait_iff_congested will stall.
1493 if (nr_dirty
&& nr_dirty
== nr_congested
)
1494 zone_set_flag(zone
, ZONE_CONGESTED
);
1497 * If dirty pages are scanned that are not queued for IO, it
1498 * implies that flushers are not keeping up. In this case, flag
1499 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1500 * pages from reclaim context. It will forcibly stall in the
1503 if (nr_unqueued_dirty
== nr_taken
)
1504 zone_set_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
1507 * In addition, if kswapd scans pages marked marked for
1508 * immediate reclaim and under writeback (nr_immediate), it
1509 * implies that pages are cycling through the LRU faster than
1510 * they are written so also forcibly stall.
1512 if (nr_unqueued_dirty
== nr_taken
|| nr_immediate
)
1513 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1517 * Stall direct reclaim for IO completions if underlying BDIs or zone
1518 * is congested. Allow kswapd to continue until it starts encountering
1519 * unqueued dirty pages or cycling through the LRU too quickly.
1521 if (!sc
->hibernation_mode
&& !current_is_kswapd())
1522 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1524 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1526 nr_scanned
, nr_reclaimed
,
1528 trace_shrink_flags(file
));
1529 return nr_reclaimed
;
1533 * This moves pages from the active list to the inactive list.
1535 * We move them the other way if the page is referenced by one or more
1536 * processes, from rmap.
1538 * If the pages are mostly unmapped, the processing is fast and it is
1539 * appropriate to hold zone->lru_lock across the whole operation. But if
1540 * the pages are mapped, the processing is slow (page_referenced()) so we
1541 * should drop zone->lru_lock around each page. It's impossible to balance
1542 * this, so instead we remove the pages from the LRU while processing them.
1543 * It is safe to rely on PG_active against the non-LRU pages in here because
1544 * nobody will play with that bit on a non-LRU page.
1546 * The downside is that we have to touch page->_count against each page.
1547 * But we had to alter page->flags anyway.
1550 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1551 struct list_head
*list
,
1552 struct list_head
*pages_to_free
,
1555 struct zone
*zone
= lruvec_zone(lruvec
);
1556 unsigned long pgmoved
= 0;
1560 while (!list_empty(list
)) {
1561 page
= lru_to_page(list
);
1562 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1564 VM_BUG_ON(PageLRU(page
));
1567 nr_pages
= hpage_nr_pages(page
);
1568 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1569 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1570 pgmoved
+= nr_pages
;
1572 if (put_page_testzero(page
)) {
1573 __ClearPageLRU(page
);
1574 __ClearPageActive(page
);
1575 del_page_from_lru_list(page
, lruvec
, lru
);
1577 if (unlikely(PageCompound(page
))) {
1578 spin_unlock_irq(&zone
->lru_lock
);
1579 (*get_compound_page_dtor(page
))(page
);
1580 spin_lock_irq(&zone
->lru_lock
);
1582 list_add(&page
->lru
, pages_to_free
);
1585 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1586 if (!is_active_lru(lru
))
1587 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1590 static void shrink_active_list(unsigned long nr_to_scan
,
1591 struct lruvec
*lruvec
,
1592 struct scan_control
*sc
,
1595 unsigned long nr_taken
;
1596 unsigned long nr_scanned
;
1597 unsigned long vm_flags
;
1598 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1599 LIST_HEAD(l_active
);
1600 LIST_HEAD(l_inactive
);
1602 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1603 unsigned long nr_rotated
= 0;
1604 isolate_mode_t isolate_mode
= 0;
1605 int file
= is_file_lru(lru
);
1606 struct zone
*zone
= lruvec_zone(lruvec
);
1611 isolate_mode
|= ISOLATE_UNMAPPED
;
1612 if (!sc
->may_writepage
)
1613 isolate_mode
|= ISOLATE_CLEAN
;
1615 spin_lock_irq(&zone
->lru_lock
);
1617 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1618 &nr_scanned
, sc
, isolate_mode
, lru
);
1619 if (global_reclaim(sc
))
1620 zone
->pages_scanned
+= nr_scanned
;
1622 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1624 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1625 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1626 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1627 spin_unlock_irq(&zone
->lru_lock
);
1629 while (!list_empty(&l_hold
)) {
1631 page
= lru_to_page(&l_hold
);
1632 list_del(&page
->lru
);
1634 if (unlikely(!page_evictable(page
))) {
1635 putback_lru_page(page
);
1639 if (unlikely(buffer_heads_over_limit
)) {
1640 if (page_has_private(page
) && trylock_page(page
)) {
1641 if (page_has_private(page
))
1642 try_to_release_page(page
, 0);
1647 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1649 nr_rotated
+= hpage_nr_pages(page
);
1651 * Identify referenced, file-backed active pages and
1652 * give them one more trip around the active list. So
1653 * that executable code get better chances to stay in
1654 * memory under moderate memory pressure. Anon pages
1655 * are not likely to be evicted by use-once streaming
1656 * IO, plus JVM can create lots of anon VM_EXEC pages,
1657 * so we ignore them here.
1659 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1660 list_add(&page
->lru
, &l_active
);
1665 ClearPageActive(page
); /* we are de-activating */
1666 list_add(&page
->lru
, &l_inactive
);
1670 * Move pages back to the lru list.
1672 spin_lock_irq(&zone
->lru_lock
);
1674 * Count referenced pages from currently used mappings as rotated,
1675 * even though only some of them are actually re-activated. This
1676 * helps balance scan pressure between file and anonymous pages in
1679 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1681 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1682 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1683 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1684 spin_unlock_irq(&zone
->lru_lock
);
1686 free_hot_cold_page_list(&l_hold
, 1);
1690 static int inactive_anon_is_low_global(struct zone
*zone
)
1692 unsigned long active
, inactive
;
1694 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1695 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1697 if (inactive
* zone
->inactive_ratio
< active
)
1704 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1705 * @lruvec: LRU vector to check
1707 * Returns true if the zone does not have enough inactive anon pages,
1708 * meaning some active anon pages need to be deactivated.
1710 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1713 * If we don't have swap space, anonymous page deactivation
1716 if (!total_swap_pages
)
1719 if (!mem_cgroup_disabled())
1720 return mem_cgroup_inactive_anon_is_low(lruvec
);
1722 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1725 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1732 * inactive_file_is_low - check if file pages need to be deactivated
1733 * @lruvec: LRU vector to check
1735 * When the system is doing streaming IO, memory pressure here
1736 * ensures that active file pages get deactivated, until more
1737 * than half of the file pages are on the inactive list.
1739 * Once we get to that situation, protect the system's working
1740 * set from being evicted by disabling active file page aging.
1742 * This uses a different ratio than the anonymous pages, because
1743 * the page cache uses a use-once replacement algorithm.
1745 static int inactive_file_is_low(struct lruvec
*lruvec
)
1747 unsigned long inactive
;
1748 unsigned long active
;
1750 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1751 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1753 return active
> inactive
;
1756 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1758 if (is_file_lru(lru
))
1759 return inactive_file_is_low(lruvec
);
1761 return inactive_anon_is_low(lruvec
);
1764 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1765 struct lruvec
*lruvec
, struct scan_control
*sc
)
1767 if (is_active_lru(lru
)) {
1768 if (inactive_list_is_low(lruvec
, lru
))
1769 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1773 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1776 static int vmscan_swappiness(struct scan_control
*sc
)
1778 if (global_reclaim(sc
))
1779 return vm_swappiness
;
1780 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1791 * Determine how aggressively the anon and file LRU lists should be
1792 * scanned. The relative value of each set of LRU lists is determined
1793 * by looking at the fraction of the pages scanned we did rotate back
1794 * onto the active list instead of evict.
1796 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1797 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1799 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
1802 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1804 u64 denominator
= 0; /* gcc */
1805 struct zone
*zone
= lruvec_zone(lruvec
);
1806 unsigned long anon_prio
, file_prio
;
1807 enum scan_balance scan_balance
;
1808 unsigned long anon
, file
, free
;
1809 bool force_scan
= false;
1810 unsigned long ap
, fp
;
1814 * If the zone or memcg is small, nr[l] can be 0. This
1815 * results in no scanning on this priority and a potential
1816 * priority drop. Global direct reclaim can go to the next
1817 * zone and tends to have no problems. Global kswapd is for
1818 * zone balancing and it needs to scan a minimum amount. When
1819 * reclaiming for a memcg, a priority drop can cause high
1820 * latencies, so it's better to scan a minimum amount there as
1823 if (current_is_kswapd() && !zone_reclaimable(zone
))
1825 if (!global_reclaim(sc
))
1828 /* If we have no swap space, do not bother scanning anon pages. */
1829 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
1830 scan_balance
= SCAN_FILE
;
1835 * Global reclaim will swap to prevent OOM even with no
1836 * swappiness, but memcg users want to use this knob to
1837 * disable swapping for individual groups completely when
1838 * using the memory controller's swap limit feature would be
1841 if (!global_reclaim(sc
) && !vmscan_swappiness(sc
)) {
1842 scan_balance
= SCAN_FILE
;
1847 * Do not apply any pressure balancing cleverness when the
1848 * system is close to OOM, scan both anon and file equally
1849 * (unless the swappiness setting disagrees with swapping).
1851 if (!sc
->priority
&& vmscan_swappiness(sc
)) {
1852 scan_balance
= SCAN_EQUAL
;
1856 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1857 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1858 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1859 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1862 * If it's foreseeable that reclaiming the file cache won't be
1863 * enough to get the zone back into a desirable shape, we have
1864 * to swap. Better start now and leave the - probably heavily
1865 * thrashing - remaining file pages alone.
1867 if (global_reclaim(sc
)) {
1868 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1869 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1870 scan_balance
= SCAN_ANON
;
1876 * There is enough inactive page cache, do not reclaim
1877 * anything from the anonymous working set right now.
1879 if (!inactive_file_is_low(lruvec
)) {
1880 scan_balance
= SCAN_FILE
;
1884 scan_balance
= SCAN_FRACT
;
1887 * With swappiness at 100, anonymous and file have the same priority.
1888 * This scanning priority is essentially the inverse of IO cost.
1890 anon_prio
= vmscan_swappiness(sc
);
1891 file_prio
= 200 - anon_prio
;
1894 * OK, so we have swap space and a fair amount of page cache
1895 * pages. We use the recently rotated / recently scanned
1896 * ratios to determine how valuable each cache is.
1898 * Because workloads change over time (and to avoid overflow)
1899 * we keep these statistics as a floating average, which ends
1900 * up weighing recent references more than old ones.
1902 * anon in [0], file in [1]
1904 spin_lock_irq(&zone
->lru_lock
);
1905 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1906 reclaim_stat
->recent_scanned
[0] /= 2;
1907 reclaim_stat
->recent_rotated
[0] /= 2;
1910 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1911 reclaim_stat
->recent_scanned
[1] /= 2;
1912 reclaim_stat
->recent_rotated
[1] /= 2;
1916 * The amount of pressure on anon vs file pages is inversely
1917 * proportional to the fraction of recently scanned pages on
1918 * each list that were recently referenced and in active use.
1920 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1921 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1923 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1924 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1925 spin_unlock_irq(&zone
->lru_lock
);
1929 denominator
= ap
+ fp
+ 1;
1931 for_each_evictable_lru(lru
) {
1932 int file
= is_file_lru(lru
);
1936 size
= get_lru_size(lruvec
, lru
);
1937 scan
= size
>> sc
->priority
;
1939 if (!scan
&& force_scan
)
1940 scan
= min(size
, SWAP_CLUSTER_MAX
);
1942 switch (scan_balance
) {
1944 /* Scan lists relative to size */
1948 * Scan types proportional to swappiness and
1949 * their relative recent reclaim efficiency.
1951 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1955 /* Scan one type exclusively */
1956 if ((scan_balance
== SCAN_FILE
) != file
)
1960 /* Look ma, no brain */
1968 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1970 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
1972 unsigned long nr
[NR_LRU_LISTS
];
1973 unsigned long targets
[NR_LRU_LISTS
];
1974 unsigned long nr_to_scan
;
1976 unsigned long nr_reclaimed
= 0;
1977 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1978 struct blk_plug plug
;
1979 bool scan_adjusted
= false;
1981 get_scan_count(lruvec
, sc
, nr
);
1983 /* Record the original scan target for proportional adjustments later */
1984 memcpy(targets
, nr
, sizeof(nr
));
1986 blk_start_plug(&plug
);
1987 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1988 nr
[LRU_INACTIVE_FILE
]) {
1989 unsigned long nr_anon
, nr_file
, percentage
;
1990 unsigned long nr_scanned
;
1992 for_each_evictable_lru(lru
) {
1994 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
1995 nr
[lru
] -= nr_to_scan
;
1997 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2002 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2006 * For global direct reclaim, reclaim only the number of pages
2007 * requested. Less care is taken to scan proportionally as it
2008 * is more important to minimise direct reclaim stall latency
2009 * than it is to properly age the LRU lists.
2011 if (global_reclaim(sc
) && !current_is_kswapd())
2015 * For kswapd and memcg, reclaim at least the number of pages
2016 * requested. Ensure that the anon and file LRUs shrink
2017 * proportionally what was requested by get_scan_count(). We
2018 * stop reclaiming one LRU and reduce the amount scanning
2019 * proportional to the original scan target.
2021 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2022 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2024 if (nr_file
> nr_anon
) {
2025 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2026 targets
[LRU_ACTIVE_ANON
] + 1;
2028 percentage
= nr_anon
* 100 / scan_target
;
2030 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2031 targets
[LRU_ACTIVE_FILE
] + 1;
2033 percentage
= nr_file
* 100 / scan_target
;
2036 /* Stop scanning the smaller of the LRU */
2038 nr
[lru
+ LRU_ACTIVE
] = 0;
2041 * Recalculate the other LRU scan count based on its original
2042 * scan target and the percentage scanning already complete
2044 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2045 nr_scanned
= targets
[lru
] - nr
[lru
];
2046 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2047 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2050 nr_scanned
= targets
[lru
] - nr
[lru
];
2051 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2052 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2054 scan_adjusted
= true;
2056 blk_finish_plug(&plug
);
2057 sc
->nr_reclaimed
+= nr_reclaimed
;
2060 * Even if we did not try to evict anon pages at all, we want to
2061 * rebalance the anon lru active/inactive ratio.
2063 if (inactive_anon_is_low(lruvec
))
2064 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2065 sc
, LRU_ACTIVE_ANON
);
2067 throttle_vm_writeout(sc
->gfp_mask
);
2070 /* Use reclaim/compaction for costly allocs or under memory pressure */
2071 static bool in_reclaim_compaction(struct scan_control
*sc
)
2073 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2074 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2075 sc
->priority
< DEF_PRIORITY
- 2))
2082 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2083 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2084 * true if more pages should be reclaimed such that when the page allocator
2085 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2086 * It will give up earlier than that if there is difficulty reclaiming pages.
2088 static inline bool should_continue_reclaim(struct zone
*zone
,
2089 unsigned long nr_reclaimed
,
2090 unsigned long nr_scanned
,
2091 struct scan_control
*sc
)
2093 unsigned long pages_for_compaction
;
2094 unsigned long inactive_lru_pages
;
2096 /* If not in reclaim/compaction mode, stop */
2097 if (!in_reclaim_compaction(sc
))
2100 /* Consider stopping depending on scan and reclaim activity */
2101 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2103 * For __GFP_REPEAT allocations, stop reclaiming if the
2104 * full LRU list has been scanned and we are still failing
2105 * to reclaim pages. This full LRU scan is potentially
2106 * expensive but a __GFP_REPEAT caller really wants to succeed
2108 if (!nr_reclaimed
&& !nr_scanned
)
2112 * For non-__GFP_REPEAT allocations which can presumably
2113 * fail without consequence, stop if we failed to reclaim
2114 * any pages from the last SWAP_CLUSTER_MAX number of
2115 * pages that were scanned. This will return to the
2116 * caller faster at the risk reclaim/compaction and
2117 * the resulting allocation attempt fails
2124 * If we have not reclaimed enough pages for compaction and the
2125 * inactive lists are large enough, continue reclaiming
2127 pages_for_compaction
= (2UL << sc
->order
);
2128 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2129 if (get_nr_swap_pages() > 0)
2130 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2131 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2132 inactive_lru_pages
> pages_for_compaction
)
2135 /* If compaction would go ahead or the allocation would succeed, stop */
2136 switch (compaction_suitable(zone
, sc
->order
)) {
2137 case COMPACT_PARTIAL
:
2138 case COMPACT_CONTINUE
:
2146 __shrink_zone(struct zone
*zone
, struct scan_control
*sc
, bool soft_reclaim
)
2148 unsigned long nr_reclaimed
, nr_scanned
;
2149 int groups_scanned
= 0;
2152 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2153 struct mem_cgroup_reclaim_cookie reclaim
= {
2155 .priority
= sc
->priority
,
2157 struct mem_cgroup
*memcg
= NULL
;
2158 mem_cgroup_iter_filter filter
= (soft_reclaim
) ?
2159 mem_cgroup_soft_reclaim_eligible
: NULL
;
2161 nr_reclaimed
= sc
->nr_reclaimed
;
2162 nr_scanned
= sc
->nr_scanned
;
2164 while ((memcg
= mem_cgroup_iter_cond(root
, memcg
, &reclaim
, filter
))) {
2165 struct lruvec
*lruvec
;
2168 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2170 shrink_lruvec(lruvec
, sc
);
2173 * Direct reclaim and kswapd have to scan all memory
2174 * cgroups to fulfill the overall scan target for the
2177 * Limit reclaim, on the other hand, only cares about
2178 * nr_to_reclaim pages to be reclaimed and it will
2179 * retry with decreasing priority if one round over the
2180 * whole hierarchy is not sufficient.
2182 if (!global_reclaim(sc
) &&
2183 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2184 mem_cgroup_iter_break(root
, memcg
);
2189 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2190 sc
->nr_scanned
- nr_scanned
,
2191 sc
->nr_reclaimed
- nr_reclaimed
);
2193 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2194 sc
->nr_scanned
- nr_scanned
, sc
));
2196 return groups_scanned
;
2200 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
2202 bool do_soft_reclaim
= mem_cgroup_should_soft_reclaim(sc
);
2203 unsigned long nr_scanned
= sc
->nr_scanned
;
2206 scanned_groups
= __shrink_zone(zone
, sc
, do_soft_reclaim
);
2208 * memcg iterator might race with other reclaimer or start from
2209 * a incomplete tree walk so the tree walk in __shrink_zone
2210 * might have missed groups that are above the soft limit. Try
2211 * another loop to catch up with others. Do it just once to
2212 * prevent from reclaim latencies when other reclaimers always
2215 if (do_soft_reclaim
&& !scanned_groups
)
2216 __shrink_zone(zone
, sc
, do_soft_reclaim
);
2219 * No group is over the soft limit or those that are do not have
2220 * pages in the zone we are reclaiming so we have to reclaim everybody
2222 if (do_soft_reclaim
&& (sc
->nr_scanned
== nr_scanned
)) {
2223 __shrink_zone(zone
, sc
, false);
2228 /* Returns true if compaction should go ahead for a high-order request */
2229 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2231 unsigned long balance_gap
, watermark
;
2234 /* Do not consider compaction for orders reclaim is meant to satisfy */
2235 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
2239 * Compaction takes time to run and there are potentially other
2240 * callers using the pages just freed. Continue reclaiming until
2241 * there is a buffer of free pages available to give compaction
2242 * a reasonable chance of completing and allocating the page
2244 balance_gap
= min(low_wmark_pages(zone
),
2245 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2246 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2247 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
2248 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2251 * If compaction is deferred, reclaim up to a point where
2252 * compaction will have a chance of success when re-enabled
2254 if (compaction_deferred(zone
, sc
->order
))
2255 return watermark_ok
;
2257 /* If compaction is not ready to start, keep reclaiming */
2258 if (!compaction_suitable(zone
, sc
->order
))
2261 return watermark_ok
;
2265 * This is the direct reclaim path, for page-allocating processes. We only
2266 * try to reclaim pages from zones which will satisfy the caller's allocation
2269 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2271 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2273 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2274 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2275 * zone defense algorithm.
2277 * If a zone is deemed to be full of pinned pages then just give it a light
2278 * scan then give up on it.
2280 * This function returns true if a zone is being reclaimed for a costly
2281 * high-order allocation and compaction is ready to begin. This indicates to
2282 * the caller that it should consider retrying the allocation instead of
2285 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2289 bool aborted_reclaim
= false;
2292 * If the number of buffer_heads in the machine exceeds the maximum
2293 * allowed level, force direct reclaim to scan the highmem zone as
2294 * highmem pages could be pinning lowmem pages storing buffer_heads
2296 if (buffer_heads_over_limit
)
2297 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2299 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2300 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2301 if (!populated_zone(zone
))
2304 * Take care memory controller reclaiming has small influence
2307 if (global_reclaim(sc
)) {
2308 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2310 if (sc
->priority
!= DEF_PRIORITY
&&
2311 !zone_reclaimable(zone
))
2312 continue; /* Let kswapd poll it */
2313 if (IS_ENABLED(CONFIG_COMPACTION
)) {
2315 * If we already have plenty of memory free for
2316 * compaction in this zone, don't free any more.
2317 * Even though compaction is invoked for any
2318 * non-zero order, only frequent costly order
2319 * reclamation is disruptive enough to become a
2320 * noticeable problem, like transparent huge
2323 if (compaction_ready(zone
, sc
)) {
2324 aborted_reclaim
= true;
2328 /* need some check for avoid more shrink_zone() */
2331 shrink_zone(zone
, sc
);
2334 return aborted_reclaim
;
2337 /* All zones in zonelist are unreclaimable? */
2338 static bool all_unreclaimable(struct zonelist
*zonelist
,
2339 struct scan_control
*sc
)
2344 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2345 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2346 if (!populated_zone(zone
))
2348 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2350 if (zone_reclaimable(zone
))
2358 * This is the main entry point to direct page reclaim.
2360 * If a full scan of the inactive list fails to free enough memory then we
2361 * are "out of memory" and something needs to be killed.
2363 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2364 * high - the zone may be full of dirty or under-writeback pages, which this
2365 * caller can't do much about. We kick the writeback threads and take explicit
2366 * naps in the hope that some of these pages can be written. But if the
2367 * allocating task holds filesystem locks which prevent writeout this might not
2368 * work, and the allocation attempt will fail.
2370 * returns: 0, if no pages reclaimed
2371 * else, the number of pages reclaimed
2373 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2374 struct scan_control
*sc
,
2375 struct shrink_control
*shrink
)
2377 unsigned long total_scanned
= 0;
2378 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2381 unsigned long writeback_threshold
;
2382 bool aborted_reclaim
;
2384 delayacct_freepages_start();
2386 if (global_reclaim(sc
))
2387 count_vm_event(ALLOCSTALL
);
2390 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2393 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2396 * Don't shrink slabs when reclaiming memory from over limit
2397 * cgroups but do shrink slab at least once when aborting
2398 * reclaim for compaction to avoid unevenly scanning file/anon
2399 * LRU pages over slab pages.
2401 if (global_reclaim(sc
)) {
2402 unsigned long lru_pages
= 0;
2403 for_each_zone_zonelist(zone
, z
, zonelist
,
2404 gfp_zone(sc
->gfp_mask
)) {
2405 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2408 lru_pages
+= zone_reclaimable_pages(zone
);
2411 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2412 if (reclaim_state
) {
2413 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2414 reclaim_state
->reclaimed_slab
= 0;
2417 total_scanned
+= sc
->nr_scanned
;
2418 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2422 * If we're getting trouble reclaiming, start doing
2423 * writepage even in laptop mode.
2425 if (sc
->priority
< DEF_PRIORITY
- 2)
2426 sc
->may_writepage
= 1;
2429 * Try to write back as many pages as we just scanned. This
2430 * tends to cause slow streaming writers to write data to the
2431 * disk smoothly, at the dirtying rate, which is nice. But
2432 * that's undesirable in laptop mode, where we *want* lumpy
2433 * writeout. So in laptop mode, write out the whole world.
2435 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2436 if (total_scanned
> writeback_threshold
) {
2437 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2438 WB_REASON_TRY_TO_FREE_PAGES
);
2439 sc
->may_writepage
= 1;
2441 } while (--sc
->priority
>= 0 && !aborted_reclaim
);
2444 delayacct_freepages_end();
2446 if (sc
->nr_reclaimed
)
2447 return sc
->nr_reclaimed
;
2450 * As hibernation is going on, kswapd is freezed so that it can't mark
2451 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2454 if (oom_killer_disabled
)
2457 /* Aborted reclaim to try compaction? don't OOM, then */
2458 if (aborted_reclaim
)
2461 /* top priority shrink_zones still had more to do? don't OOM, then */
2462 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2468 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2471 unsigned long pfmemalloc_reserve
= 0;
2472 unsigned long free_pages
= 0;
2476 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2477 zone
= &pgdat
->node_zones
[i
];
2478 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2479 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2482 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2484 /* kswapd must be awake if processes are being throttled */
2485 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2486 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2487 (enum zone_type
)ZONE_NORMAL
);
2488 wake_up_interruptible(&pgdat
->kswapd_wait
);
2495 * Throttle direct reclaimers if backing storage is backed by the network
2496 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2497 * depleted. kswapd will continue to make progress and wake the processes
2498 * when the low watermark is reached.
2500 * Returns true if a fatal signal was delivered during throttling. If this
2501 * happens, the page allocator should not consider triggering the OOM killer.
2503 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2504 nodemask_t
*nodemask
)
2507 int high_zoneidx
= gfp_zone(gfp_mask
);
2511 * Kernel threads should not be throttled as they may be indirectly
2512 * responsible for cleaning pages necessary for reclaim to make forward
2513 * progress. kjournald for example may enter direct reclaim while
2514 * committing a transaction where throttling it could forcing other
2515 * processes to block on log_wait_commit().
2517 if (current
->flags
& PF_KTHREAD
)
2521 * If a fatal signal is pending, this process should not throttle.
2522 * It should return quickly so it can exit and free its memory
2524 if (fatal_signal_pending(current
))
2527 /* Check if the pfmemalloc reserves are ok */
2528 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
, &zone
);
2529 pgdat
= zone
->zone_pgdat
;
2530 if (pfmemalloc_watermark_ok(pgdat
))
2533 /* Account for the throttling */
2534 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2537 * If the caller cannot enter the filesystem, it's possible that it
2538 * is due to the caller holding an FS lock or performing a journal
2539 * transaction in the case of a filesystem like ext[3|4]. In this case,
2540 * it is not safe to block on pfmemalloc_wait as kswapd could be
2541 * blocked waiting on the same lock. Instead, throttle for up to a
2542 * second before continuing.
2544 if (!(gfp_mask
& __GFP_FS
)) {
2545 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2546 pfmemalloc_watermark_ok(pgdat
), HZ
);
2551 /* Throttle until kswapd wakes the process */
2552 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2553 pfmemalloc_watermark_ok(pgdat
));
2556 if (fatal_signal_pending(current
))
2563 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2564 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2566 unsigned long nr_reclaimed
;
2567 struct scan_control sc
= {
2568 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2569 .may_writepage
= !laptop_mode
,
2570 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2574 .priority
= DEF_PRIORITY
,
2575 .target_mem_cgroup
= NULL
,
2576 .nodemask
= nodemask
,
2578 struct shrink_control shrink
= {
2579 .gfp_mask
= sc
.gfp_mask
,
2583 * Do not enter reclaim if fatal signal was delivered while throttled.
2584 * 1 is returned so that the page allocator does not OOM kill at this
2587 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2590 trace_mm_vmscan_direct_reclaim_begin(order
,
2594 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2596 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2598 return nr_reclaimed
;
2603 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2604 gfp_t gfp_mask
, bool noswap
,
2606 unsigned long *nr_scanned
)
2608 struct scan_control sc
= {
2610 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2611 .may_writepage
= !laptop_mode
,
2613 .may_swap
= !noswap
,
2616 .target_mem_cgroup
= memcg
,
2618 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2620 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2621 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2623 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2628 * NOTE: Although we can get the priority field, using it
2629 * here is not a good idea, since it limits the pages we can scan.
2630 * if we don't reclaim here, the shrink_zone from balance_pgdat
2631 * will pick up pages from other mem cgroup's as well. We hack
2632 * the priority and make it zero.
2634 shrink_lruvec(lruvec
, &sc
);
2636 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2638 *nr_scanned
= sc
.nr_scanned
;
2639 return sc
.nr_reclaimed
;
2642 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2646 struct zonelist
*zonelist
;
2647 unsigned long nr_reclaimed
;
2649 struct scan_control sc
= {
2650 .may_writepage
= !laptop_mode
,
2652 .may_swap
= !noswap
,
2653 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2655 .priority
= DEF_PRIORITY
,
2656 .target_mem_cgroup
= memcg
,
2657 .nodemask
= NULL
, /* we don't care the placement */
2658 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2659 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2661 struct shrink_control shrink
= {
2662 .gfp_mask
= sc
.gfp_mask
,
2666 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2667 * take care of from where we get pages. So the node where we start the
2668 * scan does not need to be the current node.
2670 nid
= mem_cgroup_select_victim_node(memcg
);
2672 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2674 trace_mm_vmscan_memcg_reclaim_begin(0,
2678 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2680 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2682 return nr_reclaimed
;
2686 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2688 struct mem_cgroup
*memcg
;
2690 if (!total_swap_pages
)
2693 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2695 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2697 if (inactive_anon_is_low(lruvec
))
2698 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2699 sc
, LRU_ACTIVE_ANON
);
2701 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2705 static bool zone_balanced(struct zone
*zone
, int order
,
2706 unsigned long balance_gap
, int classzone_idx
)
2708 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2709 balance_gap
, classzone_idx
, 0))
2712 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&&
2713 !compaction_suitable(zone
, order
))
2720 * pgdat_balanced() is used when checking if a node is balanced.
2722 * For order-0, all zones must be balanced!
2724 * For high-order allocations only zones that meet watermarks and are in a
2725 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2726 * total of balanced pages must be at least 25% of the zones allowed by
2727 * classzone_idx for the node to be considered balanced. Forcing all zones to
2728 * be balanced for high orders can cause excessive reclaim when there are
2730 * The choice of 25% is due to
2731 * o a 16M DMA zone that is balanced will not balance a zone on any
2732 * reasonable sized machine
2733 * o On all other machines, the top zone must be at least a reasonable
2734 * percentage of the middle zones. For example, on 32-bit x86, highmem
2735 * would need to be at least 256M for it to be balance a whole node.
2736 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2737 * to balance a node on its own. These seemed like reasonable ratios.
2739 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2741 unsigned long managed_pages
= 0;
2742 unsigned long balanced_pages
= 0;
2745 /* Check the watermark levels */
2746 for (i
= 0; i
<= classzone_idx
; i
++) {
2747 struct zone
*zone
= pgdat
->node_zones
+ i
;
2749 if (!populated_zone(zone
))
2752 managed_pages
+= zone
->managed_pages
;
2755 * A special case here:
2757 * balance_pgdat() skips over all_unreclaimable after
2758 * DEF_PRIORITY. Effectively, it considers them balanced so
2759 * they must be considered balanced here as well!
2761 if (!zone_reclaimable(zone
)) {
2762 balanced_pages
+= zone
->managed_pages
;
2766 if (zone_balanced(zone
, order
, 0, i
))
2767 balanced_pages
+= zone
->managed_pages
;
2773 return balanced_pages
>= (managed_pages
>> 2);
2779 * Prepare kswapd for sleeping. This verifies that there are no processes
2780 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2782 * Returns true if kswapd is ready to sleep
2784 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
2787 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2792 * There is a potential race between when kswapd checks its watermarks
2793 * and a process gets throttled. There is also a potential race if
2794 * processes get throttled, kswapd wakes, a large process exits therby
2795 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2796 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2797 * so wake them now if necessary. If necessary, processes will wake
2798 * kswapd and get throttled again
2800 if (waitqueue_active(&pgdat
->pfmemalloc_wait
)) {
2801 wake_up(&pgdat
->pfmemalloc_wait
);
2805 return pgdat_balanced(pgdat
, order
, classzone_idx
);
2809 * kswapd shrinks the zone by the number of pages required to reach
2810 * the high watermark.
2812 * Returns true if kswapd scanned at least the requested number of pages to
2813 * reclaim or if the lack of progress was due to pages under writeback.
2814 * This is used to determine if the scanning priority needs to be raised.
2816 static bool kswapd_shrink_zone(struct zone
*zone
,
2818 struct scan_control
*sc
,
2819 unsigned long lru_pages
,
2820 unsigned long *nr_attempted
)
2822 int testorder
= sc
->order
;
2823 unsigned long balance_gap
;
2824 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2825 struct shrink_control shrink
= {
2826 .gfp_mask
= sc
->gfp_mask
,
2828 bool lowmem_pressure
;
2830 /* Reclaim above the high watermark. */
2831 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
2834 * Kswapd reclaims only single pages with compaction enabled. Trying
2835 * too hard to reclaim until contiguous free pages have become
2836 * available can hurt performance by evicting too much useful data
2837 * from memory. Do not reclaim more than needed for compaction.
2839 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2840 compaction_suitable(zone
, sc
->order
) !=
2845 * We put equal pressure on every zone, unless one zone has way too
2846 * many pages free already. The "too many pages" is defined as the
2847 * high wmark plus a "gap" where the gap is either the low
2848 * watermark or 1% of the zone, whichever is smaller.
2850 balance_gap
= min(low_wmark_pages(zone
),
2851 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2852 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2855 * If there is no low memory pressure or the zone is balanced then no
2856 * reclaim is necessary
2858 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
2859 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
2860 balance_gap
, classzone_idx
))
2863 shrink_zone(zone
, sc
);
2865 reclaim_state
->reclaimed_slab
= 0;
2866 shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2867 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2869 /* Account for the number of pages attempted to reclaim */
2870 *nr_attempted
+= sc
->nr_to_reclaim
;
2872 zone_clear_flag(zone
, ZONE_WRITEBACK
);
2875 * If a zone reaches its high watermark, consider it to be no longer
2876 * congested. It's possible there are dirty pages backed by congested
2877 * BDIs but as pressure is relieved, speculatively avoid congestion
2880 if (zone_reclaimable(zone
) &&
2881 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
2882 zone_clear_flag(zone
, ZONE_CONGESTED
);
2883 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2886 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
2890 * For kswapd, balance_pgdat() will work across all this node's zones until
2891 * they are all at high_wmark_pages(zone).
2893 * Returns the final order kswapd was reclaiming at
2895 * There is special handling here for zones which are full of pinned pages.
2896 * This can happen if the pages are all mlocked, or if they are all used by
2897 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2898 * What we do is to detect the case where all pages in the zone have been
2899 * scanned twice and there has been zero successful reclaim. Mark the zone as
2900 * dead and from now on, only perform a short scan. Basically we're polling
2901 * the zone for when the problem goes away.
2903 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2904 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2905 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2906 * lower zones regardless of the number of free pages in the lower zones. This
2907 * interoperates with the page allocator fallback scheme to ensure that aging
2908 * of pages is balanced across the zones.
2910 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2914 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2915 struct scan_control sc
= {
2916 .gfp_mask
= GFP_KERNEL
,
2917 .priority
= DEF_PRIORITY
,
2920 .may_writepage
= !laptop_mode
,
2922 .target_mem_cgroup
= NULL
,
2924 count_vm_event(PAGEOUTRUN
);
2927 unsigned long lru_pages
= 0;
2928 unsigned long nr_attempted
= 0;
2929 bool raise_priority
= true;
2930 bool pgdat_needs_compaction
= (order
> 0);
2932 sc
.nr_reclaimed
= 0;
2935 * Scan in the highmem->dma direction for the highest
2936 * zone which needs scanning
2938 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2939 struct zone
*zone
= pgdat
->node_zones
+ i
;
2941 if (!populated_zone(zone
))
2944 if (sc
.priority
!= DEF_PRIORITY
&&
2945 !zone_reclaimable(zone
))
2949 * Do some background aging of the anon list, to give
2950 * pages a chance to be referenced before reclaiming.
2952 age_active_anon(zone
, &sc
);
2955 * If the number of buffer_heads in the machine
2956 * exceeds the maximum allowed level and this node
2957 * has a highmem zone, force kswapd to reclaim from
2958 * it to relieve lowmem pressure.
2960 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2965 if (!zone_balanced(zone
, order
, 0, 0)) {
2970 * If balanced, clear the dirty and congested
2973 zone_clear_flag(zone
, ZONE_CONGESTED
);
2974 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2981 for (i
= 0; i
<= end_zone
; i
++) {
2982 struct zone
*zone
= pgdat
->node_zones
+ i
;
2984 if (!populated_zone(zone
))
2987 lru_pages
+= zone_reclaimable_pages(zone
);
2990 * If any zone is currently balanced then kswapd will
2991 * not call compaction as it is expected that the
2992 * necessary pages are already available.
2994 if (pgdat_needs_compaction
&&
2995 zone_watermark_ok(zone
, order
,
2996 low_wmark_pages(zone
),
2998 pgdat_needs_compaction
= false;
3002 * If we're getting trouble reclaiming, start doing writepage
3003 * even in laptop mode.
3005 if (sc
.priority
< DEF_PRIORITY
- 2)
3006 sc
.may_writepage
= 1;
3009 * Now scan the zone in the dma->highmem direction, stopping
3010 * at the last zone which needs scanning.
3012 * We do this because the page allocator works in the opposite
3013 * direction. This prevents the page allocator from allocating
3014 * pages behind kswapd's direction of progress, which would
3015 * cause too much scanning of the lower zones.
3017 for (i
= 0; i
<= end_zone
; i
++) {
3018 struct zone
*zone
= pgdat
->node_zones
+ i
;
3020 if (!populated_zone(zone
))
3023 if (sc
.priority
!= DEF_PRIORITY
&&
3024 !zone_reclaimable(zone
))
3030 * There should be no need to raise the scanning
3031 * priority if enough pages are already being scanned
3032 * that that high watermark would be met at 100%
3035 if (kswapd_shrink_zone(zone
, end_zone
, &sc
,
3036 lru_pages
, &nr_attempted
))
3037 raise_priority
= false;
3041 * If the low watermark is met there is no need for processes
3042 * to be throttled on pfmemalloc_wait as they should not be
3043 * able to safely make forward progress. Wake them
3045 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3046 pfmemalloc_watermark_ok(pgdat
))
3047 wake_up(&pgdat
->pfmemalloc_wait
);
3050 * Fragmentation may mean that the system cannot be rebalanced
3051 * for high-order allocations in all zones. If twice the
3052 * allocation size has been reclaimed and the zones are still
3053 * not balanced then recheck the watermarks at order-0 to
3054 * prevent kswapd reclaiming excessively. Assume that a
3055 * process requested a high-order can direct reclaim/compact.
3057 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3058 order
= sc
.order
= 0;
3060 /* Check if kswapd should be suspending */
3061 if (try_to_freeze() || kthread_should_stop())
3065 * Compact if necessary and kswapd is reclaiming at least the
3066 * high watermark number of pages as requsted
3068 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3069 compact_pgdat(pgdat
, order
);
3072 * Raise priority if scanning rate is too low or there was no
3073 * progress in reclaiming pages
3075 if (raise_priority
|| !sc
.nr_reclaimed
)
3077 } while (sc
.priority
>= 1 &&
3078 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3082 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3083 * makes a decision on the order we were last reclaiming at. However,
3084 * if another caller entered the allocator slow path while kswapd
3085 * was awake, order will remain at the higher level
3087 *classzone_idx
= end_zone
;
3091 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3096 if (freezing(current
) || kthread_should_stop())
3099 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3101 /* Try to sleep for a short interval */
3102 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3103 remaining
= schedule_timeout(HZ
/10);
3104 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3105 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3109 * After a short sleep, check if it was a premature sleep. If not, then
3110 * go fully to sleep until explicitly woken up.
3112 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3113 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3116 * vmstat counters are not perfectly accurate and the estimated
3117 * value for counters such as NR_FREE_PAGES can deviate from the
3118 * true value by nr_online_cpus * threshold. To avoid the zone
3119 * watermarks being breached while under pressure, we reduce the
3120 * per-cpu vmstat threshold while kswapd is awake and restore
3121 * them before going back to sleep.
3123 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3126 * Compaction records what page blocks it recently failed to
3127 * isolate pages from and skips them in the future scanning.
3128 * When kswapd is going to sleep, it is reasonable to assume
3129 * that pages and compaction may succeed so reset the cache.
3131 reset_isolation_suitable(pgdat
);
3133 if (!kthread_should_stop())
3136 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3139 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3141 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3143 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3147 * The background pageout daemon, started as a kernel thread
3148 * from the init process.
3150 * This basically trickles out pages so that we have _some_
3151 * free memory available even if there is no other activity
3152 * that frees anything up. This is needed for things like routing
3153 * etc, where we otherwise might have all activity going on in
3154 * asynchronous contexts that cannot page things out.
3156 * If there are applications that are active memory-allocators
3157 * (most normal use), this basically shouldn't matter.
3159 static int kswapd(void *p
)
3161 unsigned long order
, new_order
;
3162 unsigned balanced_order
;
3163 int classzone_idx
, new_classzone_idx
;
3164 int balanced_classzone_idx
;
3165 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3166 struct task_struct
*tsk
= current
;
3168 struct reclaim_state reclaim_state
= {
3169 .reclaimed_slab
= 0,
3171 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3173 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3175 if (!cpumask_empty(cpumask
))
3176 set_cpus_allowed_ptr(tsk
, cpumask
);
3177 current
->reclaim_state
= &reclaim_state
;
3180 * Tell the memory management that we're a "memory allocator",
3181 * and that if we need more memory we should get access to it
3182 * regardless (see "__alloc_pages()"). "kswapd" should
3183 * never get caught in the normal page freeing logic.
3185 * (Kswapd normally doesn't need memory anyway, but sometimes
3186 * you need a small amount of memory in order to be able to
3187 * page out something else, and this flag essentially protects
3188 * us from recursively trying to free more memory as we're
3189 * trying to free the first piece of memory in the first place).
3191 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3194 order
= new_order
= 0;
3196 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3197 balanced_classzone_idx
= classzone_idx
;
3202 * If the last balance_pgdat was unsuccessful it's unlikely a
3203 * new request of a similar or harder type will succeed soon
3204 * so consider going to sleep on the basis we reclaimed at
3206 if (balanced_classzone_idx
>= new_classzone_idx
&&
3207 balanced_order
== new_order
) {
3208 new_order
= pgdat
->kswapd_max_order
;
3209 new_classzone_idx
= pgdat
->classzone_idx
;
3210 pgdat
->kswapd_max_order
= 0;
3211 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3214 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3216 * Don't sleep if someone wants a larger 'order'
3217 * allocation or has tigher zone constraints
3220 classzone_idx
= new_classzone_idx
;
3222 kswapd_try_to_sleep(pgdat
, balanced_order
,
3223 balanced_classzone_idx
);
3224 order
= pgdat
->kswapd_max_order
;
3225 classzone_idx
= pgdat
->classzone_idx
;
3227 new_classzone_idx
= classzone_idx
;
3228 pgdat
->kswapd_max_order
= 0;
3229 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3232 ret
= try_to_freeze();
3233 if (kthread_should_stop())
3237 * We can speed up thawing tasks if we don't call balance_pgdat
3238 * after returning from the refrigerator
3241 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3242 balanced_classzone_idx
= classzone_idx
;
3243 balanced_order
= balance_pgdat(pgdat
, order
,
3244 &balanced_classzone_idx
);
3248 current
->reclaim_state
= NULL
;
3253 * A zone is low on free memory, so wake its kswapd task to service it.
3255 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3259 if (!populated_zone(zone
))
3262 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
3264 pgdat
= zone
->zone_pgdat
;
3265 if (pgdat
->kswapd_max_order
< order
) {
3266 pgdat
->kswapd_max_order
= order
;
3267 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3269 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3271 if (zone_balanced(zone
, order
, 0, 0))
3274 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3275 wake_up_interruptible(&pgdat
->kswapd_wait
);
3279 * The reclaimable count would be mostly accurate.
3280 * The less reclaimable pages may be
3281 * - mlocked pages, which will be moved to unevictable list when encountered
3282 * - mapped pages, which may require several travels to be reclaimed
3283 * - dirty pages, which is not "instantly" reclaimable
3285 unsigned long global_reclaimable_pages(void)
3289 nr
= global_page_state(NR_ACTIVE_FILE
) +
3290 global_page_state(NR_INACTIVE_FILE
);
3292 if (get_nr_swap_pages() > 0)
3293 nr
+= global_page_state(NR_ACTIVE_ANON
) +
3294 global_page_state(NR_INACTIVE_ANON
);
3299 #ifdef CONFIG_HIBERNATION
3301 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3304 * Rather than trying to age LRUs the aim is to preserve the overall
3305 * LRU order by reclaiming preferentially
3306 * inactive > active > active referenced > active mapped
3308 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3310 struct reclaim_state reclaim_state
;
3311 struct scan_control sc
= {
3312 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3316 .nr_to_reclaim
= nr_to_reclaim
,
3317 .hibernation_mode
= 1,
3319 .priority
= DEF_PRIORITY
,
3321 struct shrink_control shrink
= {
3322 .gfp_mask
= sc
.gfp_mask
,
3324 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3325 struct task_struct
*p
= current
;
3326 unsigned long nr_reclaimed
;
3328 p
->flags
|= PF_MEMALLOC
;
3329 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3330 reclaim_state
.reclaimed_slab
= 0;
3331 p
->reclaim_state
= &reclaim_state
;
3333 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3335 p
->reclaim_state
= NULL
;
3336 lockdep_clear_current_reclaim_state();
3337 p
->flags
&= ~PF_MEMALLOC
;
3339 return nr_reclaimed
;
3341 #endif /* CONFIG_HIBERNATION */
3343 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3344 not required for correctness. So if the last cpu in a node goes
3345 away, we get changed to run anywhere: as the first one comes back,
3346 restore their cpu bindings. */
3347 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3352 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3353 for_each_node_state(nid
, N_MEMORY
) {
3354 pg_data_t
*pgdat
= NODE_DATA(nid
);
3355 const struct cpumask
*mask
;
3357 mask
= cpumask_of_node(pgdat
->node_id
);
3359 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3360 /* One of our CPUs online: restore mask */
3361 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3368 * This kswapd start function will be called by init and node-hot-add.
3369 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3371 int kswapd_run(int nid
)
3373 pg_data_t
*pgdat
= NODE_DATA(nid
);
3379 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3380 if (IS_ERR(pgdat
->kswapd
)) {
3381 /* failure at boot is fatal */
3382 BUG_ON(system_state
== SYSTEM_BOOTING
);
3383 pr_err("Failed to start kswapd on node %d\n", nid
);
3384 ret
= PTR_ERR(pgdat
->kswapd
);
3385 pgdat
->kswapd
= NULL
;
3391 * Called by memory hotplug when all memory in a node is offlined. Caller must
3392 * hold lock_memory_hotplug().
3394 void kswapd_stop(int nid
)
3396 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3399 kthread_stop(kswapd
);
3400 NODE_DATA(nid
)->kswapd
= NULL
;
3404 static int __init
kswapd_init(void)
3409 for_each_node_state(nid
, N_MEMORY
)
3411 hotcpu_notifier(cpu_callback
, 0);
3415 module_init(kswapd_init
)
3421 * If non-zero call zone_reclaim when the number of free pages falls below
3424 int zone_reclaim_mode __read_mostly
;
3426 #define RECLAIM_OFF 0
3427 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3428 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3429 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3432 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3433 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3436 #define ZONE_RECLAIM_PRIORITY 4
3439 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3442 int sysctl_min_unmapped_ratio
= 1;
3445 * If the number of slab pages in a zone grows beyond this percentage then
3446 * slab reclaim needs to occur.
3448 int sysctl_min_slab_ratio
= 5;
3450 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3452 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3453 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3454 zone_page_state(zone
, NR_ACTIVE_FILE
);
3457 * It's possible for there to be more file mapped pages than
3458 * accounted for by the pages on the file LRU lists because
3459 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3461 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3464 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3465 static long zone_pagecache_reclaimable(struct zone
*zone
)
3467 long nr_pagecache_reclaimable
;
3471 * If RECLAIM_SWAP is set, then all file pages are considered
3472 * potentially reclaimable. Otherwise, we have to worry about
3473 * pages like swapcache and zone_unmapped_file_pages() provides
3476 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3477 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3479 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3481 /* If we can't clean pages, remove dirty pages from consideration */
3482 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3483 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3485 /* Watch for any possible underflows due to delta */
3486 if (unlikely(delta
> nr_pagecache_reclaimable
))
3487 delta
= nr_pagecache_reclaimable
;
3489 return nr_pagecache_reclaimable
- delta
;
3493 * Try to free up some pages from this zone through reclaim.
3495 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3497 /* Minimum pages needed in order to stay on node */
3498 const unsigned long nr_pages
= 1 << order
;
3499 struct task_struct
*p
= current
;
3500 struct reclaim_state reclaim_state
;
3501 struct scan_control sc
= {
3502 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3503 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3505 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3506 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3508 .priority
= ZONE_RECLAIM_PRIORITY
,
3510 struct shrink_control shrink
= {
3511 .gfp_mask
= sc
.gfp_mask
,
3513 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3517 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3518 * and we also need to be able to write out pages for RECLAIM_WRITE
3521 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3522 lockdep_set_current_reclaim_state(gfp_mask
);
3523 reclaim_state
.reclaimed_slab
= 0;
3524 p
->reclaim_state
= &reclaim_state
;
3526 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3528 * Free memory by calling shrink zone with increasing
3529 * priorities until we have enough memory freed.
3532 shrink_zone(zone
, &sc
);
3533 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3536 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3537 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3539 * shrink_slab() does not currently allow us to determine how
3540 * many pages were freed in this zone. So we take the current
3541 * number of slab pages and shake the slab until it is reduced
3542 * by the same nr_pages that we used for reclaiming unmapped
3545 * Note that shrink_slab will free memory on all zones and may
3549 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3551 /* No reclaimable slab or very low memory pressure */
3552 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3555 /* Freed enough memory */
3556 nr_slab_pages1
= zone_page_state(zone
,
3557 NR_SLAB_RECLAIMABLE
);
3558 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3563 * Update nr_reclaimed by the number of slab pages we
3564 * reclaimed from this zone.
3566 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3567 if (nr_slab_pages1
< nr_slab_pages0
)
3568 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3571 p
->reclaim_state
= NULL
;
3572 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3573 lockdep_clear_current_reclaim_state();
3574 return sc
.nr_reclaimed
>= nr_pages
;
3577 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3583 * Zone reclaim reclaims unmapped file backed pages and
3584 * slab pages if we are over the defined limits.
3586 * A small portion of unmapped file backed pages is needed for
3587 * file I/O otherwise pages read by file I/O will be immediately
3588 * thrown out if the zone is overallocated. So we do not reclaim
3589 * if less than a specified percentage of the zone is used by
3590 * unmapped file backed pages.
3592 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3593 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3594 return ZONE_RECLAIM_FULL
;
3596 if (!zone_reclaimable(zone
))
3597 return ZONE_RECLAIM_FULL
;
3600 * Do not scan if the allocation should not be delayed.
3602 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3603 return ZONE_RECLAIM_NOSCAN
;
3606 * Only run zone reclaim on the local zone or on zones that do not
3607 * have associated processors. This will favor the local processor
3608 * over remote processors and spread off node memory allocations
3609 * as wide as possible.
3611 node_id
= zone_to_nid(zone
);
3612 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3613 return ZONE_RECLAIM_NOSCAN
;
3615 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3616 return ZONE_RECLAIM_NOSCAN
;
3618 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3619 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3622 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3629 * page_evictable - test whether a page is evictable
3630 * @page: the page to test
3632 * Test whether page is evictable--i.e., should be placed on active/inactive
3633 * lists vs unevictable list.
3635 * Reasons page might not be evictable:
3636 * (1) page's mapping marked unevictable
3637 * (2) page is part of an mlocked VMA
3640 int page_evictable(struct page
*page
)
3642 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3647 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3648 * @pages: array of pages to check
3649 * @nr_pages: number of pages to check
3651 * Checks pages for evictability and moves them to the appropriate lru list.
3653 * This function is only used for SysV IPC SHM_UNLOCK.
3655 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3657 struct lruvec
*lruvec
;
3658 struct zone
*zone
= NULL
;
3663 for (i
= 0; i
< nr_pages
; i
++) {
3664 struct page
*page
= pages
[i
];
3665 struct zone
*pagezone
;
3668 pagezone
= page_zone(page
);
3669 if (pagezone
!= zone
) {
3671 spin_unlock_irq(&zone
->lru_lock
);
3673 spin_lock_irq(&zone
->lru_lock
);
3675 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3677 if (!PageLRU(page
) || !PageUnevictable(page
))
3680 if (page_evictable(page
)) {
3681 enum lru_list lru
= page_lru_base_type(page
);
3683 VM_BUG_ON(PageActive(page
));
3684 ClearPageUnevictable(page
);
3685 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3686 add_page_to_lru_list(page
, lruvec
, lru
);
3692 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3693 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3694 spin_unlock_irq(&zone
->lru_lock
);
3697 #endif /* CONFIG_SHMEM */
3699 static void warn_scan_unevictable_pages(void)
3701 printk_once(KERN_WARNING
3702 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3703 "disabled for lack of a legitimate use case. If you have "
3704 "one, please send an email to linux-mm@kvack.org.\n",
3709 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3710 * all nodes' unevictable lists for evictable pages
3712 unsigned long scan_unevictable_pages
;
3714 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3715 void __user
*buffer
,
3716 size_t *length
, loff_t
*ppos
)
3718 warn_scan_unevictable_pages();
3719 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3720 scan_unevictable_pages
= 0;
3726 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3727 * a specified node's per zone unevictable lists for evictable pages.
3730 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3731 struct device_attribute
*attr
,
3734 warn_scan_unevictable_pages();
3735 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3738 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3739 struct device_attribute
*attr
,
3740 const char *buf
, size_t count
)
3742 warn_scan_unevictable_pages();
3747 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3748 read_scan_unevictable_node
,
3749 write_scan_unevictable_node
);
3751 int scan_unevictable_register_node(struct node
*node
)
3753 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3756 void scan_unevictable_unregister_node(struct node
*node
)
3758 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);