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 global_reclaim(struct scan_control
*sc
)
149 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
151 if (!mem_cgroup_disabled())
152 return mem_cgroup_get_lru_size(lruvec
, lru
);
154 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
158 * Add a shrinker callback to be called from the vm
160 void register_shrinker(struct shrinker
*shrinker
)
162 atomic_long_set(&shrinker
->nr_in_batch
, 0);
163 down_write(&shrinker_rwsem
);
164 list_add_tail(&shrinker
->list
, &shrinker_list
);
165 up_write(&shrinker_rwsem
);
167 EXPORT_SYMBOL(register_shrinker
);
172 void unregister_shrinker(struct shrinker
*shrinker
)
174 down_write(&shrinker_rwsem
);
175 list_del(&shrinker
->list
);
176 up_write(&shrinker_rwsem
);
178 EXPORT_SYMBOL(unregister_shrinker
);
180 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
181 struct shrink_control
*sc
,
182 unsigned long nr_to_scan
)
184 sc
->nr_to_scan
= nr_to_scan
;
185 return (*shrinker
->shrink
)(shrinker
, sc
);
188 #define SHRINK_BATCH 128
190 * Call the shrink functions to age shrinkable caches
192 * Here we assume it costs one seek to replace a lru page and that it also
193 * takes a seek to recreate a cache object. With this in mind we age equal
194 * percentages of the lru and ageable caches. This should balance the seeks
195 * generated by these structures.
197 * If the vm encountered mapped pages on the LRU it increase the pressure on
198 * slab to avoid swapping.
200 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
202 * `lru_pages' represents the number of on-LRU pages in all the zones which
203 * are eligible for the caller's allocation attempt. It is used for balancing
204 * slab reclaim versus page reclaim.
206 * Returns the number of slab objects which we shrunk.
208 unsigned long shrink_slab(struct shrink_control
*shrink
,
209 unsigned long nr_pages_scanned
,
210 unsigned long lru_pages
)
212 struct shrinker
*shrinker
;
213 unsigned long ret
= 0;
215 if (nr_pages_scanned
== 0)
216 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
218 if (!down_read_trylock(&shrinker_rwsem
)) {
219 /* Assume we'll be able to shrink next time */
224 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
225 unsigned long long delta
;
231 long batch_size
= shrinker
->batch
? shrinker
->batch
234 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
243 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
246 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
248 do_div(delta
, lru_pages
+ 1);
250 if (total_scan
< 0) {
251 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
253 shrinker
->shrink
, total_scan
);
254 total_scan
= max_pass
;
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * max_pass. This is bad for sustaining a working set in
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
269 if (delta
< max_pass
/ 4)
270 total_scan
= min(total_scan
, max_pass
/ 2);
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
277 if (total_scan
> max_pass
* 2)
278 total_scan
= max_pass
* 2;
280 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
281 nr_pages_scanned
, lru_pages
,
282 max_pass
, delta
, total_scan
);
284 while (total_scan
>= batch_size
) {
287 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
288 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
290 if (shrink_ret
== -1)
292 if (shrink_ret
< nr_before
)
293 ret
+= nr_before
- shrink_ret
;
294 count_vm_events(SLABS_SCANNED
, batch_size
);
295 total_scan
-= batch_size
;
301 * move the unused scan count back into the shrinker in a
302 * manner that handles concurrent updates. If we exhausted the
303 * scan, there is no need to do an update.
306 new_nr
= atomic_long_add_return(total_scan
,
307 &shrinker
->nr_in_batch
);
309 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
311 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
313 up_read(&shrinker_rwsem
);
319 static inline int is_page_cache_freeable(struct page
*page
)
322 * A freeable page cache page is referenced only by the caller
323 * that isolated the page, the page cache radix tree and
324 * optional buffer heads at page->private.
326 return page_count(page
) - page_has_private(page
) == 2;
329 static int may_write_to_queue(struct backing_dev_info
*bdi
,
330 struct scan_control
*sc
)
332 if (current
->flags
& PF_SWAPWRITE
)
334 if (!bdi_write_congested(bdi
))
336 if (bdi
== current
->backing_dev_info
)
342 * We detected a synchronous write error writing a page out. Probably
343 * -ENOSPC. We need to propagate that into the address_space for a subsequent
344 * fsync(), msync() or close().
346 * The tricky part is that after writepage we cannot touch the mapping: nothing
347 * prevents it from being freed up. But we have a ref on the page and once
348 * that page is locked, the mapping is pinned.
350 * We're allowed to run sleeping lock_page() here because we know the caller has
353 static void handle_write_error(struct address_space
*mapping
,
354 struct page
*page
, int error
)
357 if (page_mapping(page
) == mapping
)
358 mapping_set_error(mapping
, error
);
362 /* possible outcome of pageout() */
364 /* failed to write page out, page is locked */
366 /* move page to the active list, page is locked */
368 /* page has been sent to the disk successfully, page is unlocked */
370 /* page is clean and locked */
375 * pageout is called by shrink_page_list() for each dirty page.
376 * Calls ->writepage().
378 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
379 struct scan_control
*sc
)
382 * If the page is dirty, only perform writeback if that write
383 * will be non-blocking. To prevent this allocation from being
384 * stalled by pagecache activity. But note that there may be
385 * stalls if we need to run get_block(). We could test
386 * PagePrivate for that.
388 * If this process is currently in __generic_file_aio_write() against
389 * this page's queue, we can perform writeback even if that
392 * If the page is swapcache, write it back even if that would
393 * block, for some throttling. This happens by accident, because
394 * swap_backing_dev_info is bust: it doesn't reflect the
395 * congestion state of the swapdevs. Easy to fix, if needed.
397 if (!is_page_cache_freeable(page
))
401 * Some data journaling orphaned pages can have
402 * page->mapping == NULL while being dirty with clean buffers.
404 if (page_has_private(page
)) {
405 if (try_to_free_buffers(page
)) {
406 ClearPageDirty(page
);
407 printk("%s: orphaned page\n", __func__
);
413 if (mapping
->a_ops
->writepage
== NULL
)
414 return PAGE_ACTIVATE
;
415 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
418 if (clear_page_dirty_for_io(page
)) {
420 struct writeback_control wbc
= {
421 .sync_mode
= WB_SYNC_NONE
,
422 .nr_to_write
= SWAP_CLUSTER_MAX
,
424 .range_end
= LLONG_MAX
,
428 SetPageReclaim(page
);
429 res
= mapping
->a_ops
->writepage(page
, &wbc
);
431 handle_write_error(mapping
, page
, res
);
432 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
433 ClearPageReclaim(page
);
434 return PAGE_ACTIVATE
;
437 if (!PageWriteback(page
)) {
438 /* synchronous write or broken a_ops? */
439 ClearPageReclaim(page
);
441 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
442 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
450 * Same as remove_mapping, but if the page is removed from the mapping, it
451 * gets returned with a refcount of 0.
453 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
455 BUG_ON(!PageLocked(page
));
456 BUG_ON(mapping
!= page_mapping(page
));
458 spin_lock_irq(&mapping
->tree_lock
);
460 * The non racy check for a busy page.
462 * Must be careful with the order of the tests. When someone has
463 * a ref to the page, it may be possible that they dirty it then
464 * drop the reference. So if PageDirty is tested before page_count
465 * here, then the following race may occur:
467 * get_user_pages(&page);
468 * [user mapping goes away]
470 * !PageDirty(page) [good]
471 * SetPageDirty(page);
473 * !page_count(page) [good, discard it]
475 * [oops, our write_to data is lost]
477 * Reversing the order of the tests ensures such a situation cannot
478 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
479 * load is not satisfied before that of page->_count.
481 * Note that if SetPageDirty is always performed via set_page_dirty,
482 * and thus under tree_lock, then this ordering is not required.
484 if (!page_freeze_refs(page
, 2))
486 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
487 if (unlikely(PageDirty(page
))) {
488 page_unfreeze_refs(page
, 2);
492 if (PageSwapCache(page
)) {
493 swp_entry_t swap
= { .val
= page_private(page
) };
494 __delete_from_swap_cache(page
);
495 spin_unlock_irq(&mapping
->tree_lock
);
496 swapcache_free(swap
, page
);
498 void (*freepage
)(struct page
*);
500 freepage
= mapping
->a_ops
->freepage
;
502 __delete_from_page_cache(page
);
503 spin_unlock_irq(&mapping
->tree_lock
);
504 mem_cgroup_uncharge_cache_page(page
);
506 if (freepage
!= NULL
)
513 spin_unlock_irq(&mapping
->tree_lock
);
518 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
519 * someone else has a ref on the page, abort and return 0. If it was
520 * successfully detached, return 1. Assumes the caller has a single ref on
523 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
525 if (__remove_mapping(mapping
, page
)) {
527 * Unfreezing the refcount with 1 rather than 2 effectively
528 * drops the pagecache ref for us without requiring another
531 page_unfreeze_refs(page
, 1);
538 * putback_lru_page - put previously isolated page onto appropriate LRU list
539 * @page: page to be put back to appropriate lru list
541 * Add previously isolated @page to appropriate LRU list.
542 * Page may still be unevictable for other reasons.
544 * lru_lock must not be held, interrupts must be enabled.
546 void putback_lru_page(struct page
*page
)
549 int was_unevictable
= PageUnevictable(page
);
551 VM_BUG_ON(PageLRU(page
));
554 ClearPageUnevictable(page
);
556 if (page_evictable(page
)) {
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
563 lru
= page_lru_base_type(page
);
567 * Put unevictable pages directly on zone's unevictable
570 lru
= LRU_UNEVICTABLE
;
571 add_page_to_unevictable_list(page
);
573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
576 * isolation/check_move_unevictable_pages,
577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578 * the page back to the evictable list.
580 * The other side is TestClearPageMlocked() or shmem_lock().
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
590 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
)) {
591 if (!isolate_lru_page(page
)) {
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
601 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
602 count_vm_event(UNEVICTABLE_PGRESCUED
);
603 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
604 count_vm_event(UNEVICTABLE_PGCULLED
);
606 put_page(page
); /* drop ref from isolate */
609 enum page_references
{
611 PAGEREF_RECLAIM_CLEAN
,
616 static enum page_references
page_check_references(struct page
*page
,
617 struct scan_control
*sc
)
619 int referenced_ptes
, referenced_page
;
620 unsigned long vm_flags
;
622 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
624 referenced_page
= TestClearPageReferenced(page
);
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
630 if (vm_flags
& VM_LOCKED
)
631 return PAGEREF_RECLAIM
;
633 if (referenced_ptes
) {
634 if (PageSwapBacked(page
))
635 return PAGEREF_ACTIVATE
;
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
650 SetPageReferenced(page
);
652 if (referenced_page
|| referenced_ptes
> 1)
653 return PAGEREF_ACTIVATE
;
656 * Activate file-backed executable pages after first usage.
658 if (vm_flags
& VM_EXEC
)
659 return PAGEREF_ACTIVATE
;
664 /* Reclaim if clean, defer dirty pages to writeback */
665 if (referenced_page
&& !PageSwapBacked(page
))
666 return PAGEREF_RECLAIM_CLEAN
;
668 return PAGEREF_RECLAIM
;
671 /* Check if a page is dirty or under writeback */
672 static void page_check_dirty_writeback(struct page
*page
,
673 bool *dirty
, bool *writeback
)
675 struct address_space
*mapping
;
678 * Anonymous pages are not handled by flushers and must be written
679 * from reclaim context. Do not stall reclaim based on them
681 if (!page_is_file_cache(page
)) {
687 /* By default assume that the page flags are accurate */
688 *dirty
= PageDirty(page
);
689 *writeback
= PageWriteback(page
);
691 /* Verify dirty/writeback state if the filesystem supports it */
692 if (!page_has_private(page
))
695 mapping
= page_mapping(page
);
696 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
697 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
701 * shrink_page_list() returns the number of reclaimed pages
703 static unsigned long shrink_page_list(struct list_head
*page_list
,
705 struct scan_control
*sc
,
706 enum ttu_flags ttu_flags
,
707 unsigned long *ret_nr_dirty
,
708 unsigned long *ret_nr_unqueued_dirty
,
709 unsigned long *ret_nr_congested
,
710 unsigned long *ret_nr_writeback
,
711 unsigned long *ret_nr_immediate
,
714 LIST_HEAD(ret_pages
);
715 LIST_HEAD(free_pages
);
717 unsigned long nr_unqueued_dirty
= 0;
718 unsigned long nr_dirty
= 0;
719 unsigned long nr_congested
= 0;
720 unsigned long nr_reclaimed
= 0;
721 unsigned long nr_writeback
= 0;
722 unsigned long nr_immediate
= 0;
726 mem_cgroup_uncharge_start();
727 while (!list_empty(page_list
)) {
728 struct address_space
*mapping
;
731 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
732 bool dirty
, writeback
;
736 page
= lru_to_page(page_list
);
737 list_del(&page
->lru
);
739 if (!trylock_page(page
))
742 VM_BUG_ON(PageActive(page
));
743 VM_BUG_ON(page_zone(page
) != zone
);
747 if (unlikely(!page_evictable(page
)))
750 if (!sc
->may_unmap
&& page_mapped(page
))
753 /* Double the slab pressure for mapped and swapcache pages */
754 if (page_mapped(page
) || PageSwapCache(page
))
757 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
758 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
761 * The number of dirty pages determines if a zone is marked
762 * reclaim_congested which affects wait_iff_congested. kswapd
763 * will stall and start writing pages if the tail of the LRU
764 * is all dirty unqueued pages.
766 page_check_dirty_writeback(page
, &dirty
, &writeback
);
767 if (dirty
|| writeback
)
770 if (dirty
&& !writeback
)
774 * Treat this page as congested if the underlying BDI is or if
775 * pages are cycling through the LRU so quickly that the
776 * pages marked for immediate reclaim are making it to the
777 * end of the LRU a second time.
779 mapping
= page_mapping(page
);
780 if ((mapping
&& bdi_write_congested(mapping
->backing_dev_info
)) ||
781 (writeback
&& PageReclaim(page
)))
785 * If a page at the tail of the LRU is under writeback, there
786 * are three cases to consider.
788 * 1) If reclaim is encountering an excessive number of pages
789 * under writeback and this page is both under writeback and
790 * PageReclaim then it indicates that pages are being queued
791 * for IO but are being recycled through the LRU before the
792 * IO can complete. Waiting on the page itself risks an
793 * indefinite stall if it is impossible to writeback the
794 * page due to IO error or disconnected storage so instead
795 * note that the LRU is being scanned too quickly and the
796 * caller can stall after page list has been processed.
798 * 2) Global reclaim encounters a page, memcg encounters a
799 * page that is not marked for immediate reclaim or
800 * the caller does not have __GFP_IO. In this case mark
801 * the page for immediate reclaim and continue scanning.
803 * __GFP_IO is checked because a loop driver thread might
804 * enter reclaim, and deadlock if it waits on a page for
805 * which it is needed to do the write (loop masks off
806 * __GFP_IO|__GFP_FS for this reason); but more thought
807 * would probably show more reasons.
809 * Don't require __GFP_FS, since we're not going into the
810 * FS, just waiting on its writeback completion. Worryingly,
811 * ext4 gfs2 and xfs allocate pages with
812 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
813 * may_enter_fs here is liable to OOM on them.
815 * 3) memcg encounters a page that is not already marked
816 * PageReclaim. memcg does not have any dirty pages
817 * throttling so we could easily OOM just because too many
818 * pages are in writeback and there is nothing else to
819 * reclaim. Wait for the writeback to complete.
821 if (PageWriteback(page
)) {
823 if (current_is_kswapd() &&
825 zone_is_reclaim_writeback(zone
)) {
830 } else if (global_reclaim(sc
) ||
831 !PageReclaim(page
) || !(sc
->gfp_mask
& __GFP_IO
)) {
833 * This is slightly racy - end_page_writeback()
834 * might have just cleared PageReclaim, then
835 * setting PageReclaim here end up interpreted
836 * as PageReadahead - but that does not matter
837 * enough to care. What we do want is for this
838 * page to have PageReclaim set next time memcg
839 * reclaim reaches the tests above, so it will
840 * then wait_on_page_writeback() to avoid OOM;
841 * and it's also appropriate in global reclaim.
843 SetPageReclaim(page
);
850 wait_on_page_writeback(page
);
855 references
= page_check_references(page
, sc
);
857 switch (references
) {
858 case PAGEREF_ACTIVATE
:
859 goto activate_locked
;
862 case PAGEREF_RECLAIM
:
863 case PAGEREF_RECLAIM_CLEAN
:
864 ; /* try to reclaim the page below */
868 * Anonymous process memory has backing store?
869 * Try to allocate it some swap space here.
871 if (PageAnon(page
) && !PageSwapCache(page
)) {
872 if (!(sc
->gfp_mask
& __GFP_IO
))
874 if (!add_to_swap(page
, page_list
))
875 goto activate_locked
;
878 /* Adding to swap updated mapping */
879 mapping
= page_mapping(page
);
883 * The page is mapped into the page tables of one or more
884 * processes. Try to unmap it here.
886 if (page_mapped(page
) && mapping
) {
887 switch (try_to_unmap(page
, ttu_flags
)) {
889 goto activate_locked
;
895 ; /* try to free the page below */
899 if (PageDirty(page
)) {
901 * Only kswapd can writeback filesystem pages to
902 * avoid risk of stack overflow but only writeback
903 * if many dirty pages have been encountered.
905 if (page_is_file_cache(page
) &&
906 (!current_is_kswapd() ||
907 !zone_is_reclaim_dirty(zone
))) {
909 * Immediately reclaim when written back.
910 * Similar in principal to deactivate_page()
911 * except we already have the page isolated
912 * and know it's dirty
914 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
915 SetPageReclaim(page
);
920 if (references
== PAGEREF_RECLAIM_CLEAN
)
924 if (!sc
->may_writepage
)
927 /* Page is dirty, try to write it out here */
928 switch (pageout(page
, mapping
, sc
)) {
932 goto activate_locked
;
934 if (PageWriteback(page
))
940 * A synchronous write - probably a ramdisk. Go
941 * ahead and try to reclaim the page.
943 if (!trylock_page(page
))
945 if (PageDirty(page
) || PageWriteback(page
))
947 mapping
= page_mapping(page
);
949 ; /* try to free the page below */
954 * If the page has buffers, try to free the buffer mappings
955 * associated with this page. If we succeed we try to free
958 * We do this even if the page is PageDirty().
959 * try_to_release_page() does not perform I/O, but it is
960 * possible for a page to have PageDirty set, but it is actually
961 * clean (all its buffers are clean). This happens if the
962 * buffers were written out directly, with submit_bh(). ext3
963 * will do this, as well as the blockdev mapping.
964 * try_to_release_page() will discover that cleanness and will
965 * drop the buffers and mark the page clean - it can be freed.
967 * Rarely, pages can have buffers and no ->mapping. These are
968 * the pages which were not successfully invalidated in
969 * truncate_complete_page(). We try to drop those buffers here
970 * and if that worked, and the page is no longer mapped into
971 * process address space (page_count == 1) it can be freed.
972 * Otherwise, leave the page on the LRU so it is swappable.
974 if (page_has_private(page
)) {
975 if (!try_to_release_page(page
, sc
->gfp_mask
))
976 goto activate_locked
;
977 if (!mapping
&& page_count(page
) == 1) {
979 if (put_page_testzero(page
))
983 * rare race with speculative reference.
984 * the speculative reference will free
985 * this page shortly, so we may
986 * increment nr_reclaimed here (and
987 * leave it off the LRU).
995 if (!mapping
|| !__remove_mapping(mapping
, page
))
999 * At this point, we have no other references and there is
1000 * no way to pick any more up (removed from LRU, removed
1001 * from pagecache). Can use non-atomic bitops now (and
1002 * we obviously don't have to worry about waking up a process
1003 * waiting on the page lock, because there are no references.
1005 __clear_page_locked(page
);
1010 * Is there need to periodically free_page_list? It would
1011 * appear not as the counts should be low
1013 list_add(&page
->lru
, &free_pages
);
1017 if (PageSwapCache(page
))
1018 try_to_free_swap(page
);
1020 putback_lru_page(page
);
1024 /* Not a candidate for swapping, so reclaim swap space. */
1025 if (PageSwapCache(page
) && vm_swap_full())
1026 try_to_free_swap(page
);
1027 VM_BUG_ON(PageActive(page
));
1028 SetPageActive(page
);
1033 list_add(&page
->lru
, &ret_pages
);
1034 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
1037 free_hot_cold_page_list(&free_pages
, 1);
1039 list_splice(&ret_pages
, page_list
);
1040 count_vm_events(PGACTIVATE
, pgactivate
);
1041 mem_cgroup_uncharge_end();
1042 *ret_nr_dirty
+= nr_dirty
;
1043 *ret_nr_congested
+= nr_congested
;
1044 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1045 *ret_nr_writeback
+= nr_writeback
;
1046 *ret_nr_immediate
+= nr_immediate
;
1047 return nr_reclaimed
;
1050 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1051 struct list_head
*page_list
)
1053 struct scan_control sc
= {
1054 .gfp_mask
= GFP_KERNEL
,
1055 .priority
= DEF_PRIORITY
,
1058 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1059 struct page
*page
, *next
;
1060 LIST_HEAD(clean_pages
);
1062 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1063 if (page_is_file_cache(page
) && !PageDirty(page
)) {
1064 ClearPageActive(page
);
1065 list_move(&page
->lru
, &clean_pages
);
1069 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1070 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1071 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1072 list_splice(&clean_pages
, page_list
);
1073 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1078 * Attempt to remove the specified page from its LRU. Only take this page
1079 * if it is of the appropriate PageActive status. Pages which are being
1080 * freed elsewhere are also ignored.
1082 * page: page to consider
1083 * mode: one of the LRU isolation modes defined above
1085 * returns 0 on success, -ve errno on failure.
1087 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1091 /* Only take pages on the LRU. */
1095 /* Compaction should not handle unevictable pages but CMA can do so */
1096 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1102 * To minimise LRU disruption, the caller can indicate that it only
1103 * wants to isolate pages it will be able to operate on without
1104 * blocking - clean pages for the most part.
1106 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1107 * is used by reclaim when it is cannot write to backing storage
1109 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1110 * that it is possible to migrate without blocking
1112 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1113 /* All the caller can do on PageWriteback is block */
1114 if (PageWriteback(page
))
1117 if (PageDirty(page
)) {
1118 struct address_space
*mapping
;
1120 /* ISOLATE_CLEAN means only clean pages */
1121 if (mode
& ISOLATE_CLEAN
)
1125 * Only pages without mappings or that have a
1126 * ->migratepage callback are possible to migrate
1129 mapping
= page_mapping(page
);
1130 if (mapping
&& !mapping
->a_ops
->migratepage
)
1135 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1138 if (likely(get_page_unless_zero(page
))) {
1140 * Be careful not to clear PageLRU until after we're
1141 * sure the page is not being freed elsewhere -- the
1142 * page release code relies on it.
1152 * zone->lru_lock is heavily contended. Some of the functions that
1153 * shrink the lists perform better by taking out a batch of pages
1154 * and working on them outside the LRU lock.
1156 * For pagecache intensive workloads, this function is the hottest
1157 * spot in the kernel (apart from copy_*_user functions).
1159 * Appropriate locks must be held before calling this function.
1161 * @nr_to_scan: The number of pages to look through on the list.
1162 * @lruvec: The LRU vector to pull pages from.
1163 * @dst: The temp list to put pages on to.
1164 * @nr_scanned: The number of pages that were scanned.
1165 * @sc: The scan_control struct for this reclaim session
1166 * @mode: One of the LRU isolation modes
1167 * @lru: LRU list id for isolating
1169 * returns how many pages were moved onto *@dst.
1171 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1172 struct lruvec
*lruvec
, struct list_head
*dst
,
1173 unsigned long *nr_scanned
, struct scan_control
*sc
,
1174 isolate_mode_t mode
, enum lru_list lru
)
1176 struct list_head
*src
= &lruvec
->lists
[lru
];
1177 unsigned long nr_taken
= 0;
1180 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1184 page
= lru_to_page(src
);
1185 prefetchw_prev_lru_page(page
, src
, flags
);
1187 VM_BUG_ON(!PageLRU(page
));
1189 switch (__isolate_lru_page(page
, mode
)) {
1191 nr_pages
= hpage_nr_pages(page
);
1192 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1193 list_move(&page
->lru
, dst
);
1194 nr_taken
+= nr_pages
;
1198 /* else it is being freed elsewhere */
1199 list_move(&page
->lru
, src
);
1208 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1209 nr_taken
, mode
, is_file_lru(lru
));
1214 * isolate_lru_page - tries to isolate a page from its LRU list
1215 * @page: page to isolate from its LRU list
1217 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1218 * vmstat statistic corresponding to whatever LRU list the page was on.
1220 * Returns 0 if the page was removed from an LRU list.
1221 * Returns -EBUSY if the page was not on an LRU list.
1223 * The returned page will have PageLRU() cleared. If it was found on
1224 * the active list, it will have PageActive set. If it was found on
1225 * the unevictable list, it will have the PageUnevictable bit set. That flag
1226 * may need to be cleared by the caller before letting the page go.
1228 * The vmstat statistic corresponding to the list on which the page was
1229 * found will be decremented.
1232 * (1) Must be called with an elevated refcount on the page. This is a
1233 * fundamentnal difference from isolate_lru_pages (which is called
1234 * without a stable reference).
1235 * (2) the lru_lock must not be held.
1236 * (3) interrupts must be enabled.
1238 int isolate_lru_page(struct page
*page
)
1242 VM_BUG_ON(!page_count(page
));
1244 if (PageLRU(page
)) {
1245 struct zone
*zone
= page_zone(page
);
1246 struct lruvec
*lruvec
;
1248 spin_lock_irq(&zone
->lru_lock
);
1249 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1250 if (PageLRU(page
)) {
1251 int lru
= page_lru(page
);
1254 del_page_from_lru_list(page
, lruvec
, lru
);
1257 spin_unlock_irq(&zone
->lru_lock
);
1263 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1264 * then get resheduled. When there are massive number of tasks doing page
1265 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1266 * the LRU list will go small and be scanned faster than necessary, leading to
1267 * unnecessary swapping, thrashing and OOM.
1269 static int too_many_isolated(struct zone
*zone
, int file
,
1270 struct scan_control
*sc
)
1272 unsigned long inactive
, isolated
;
1274 if (current_is_kswapd())
1277 if (!global_reclaim(sc
))
1281 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1282 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1284 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1285 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1289 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1290 * won't get blocked by normal direct-reclaimers, forming a circular
1293 if ((sc
->gfp_mask
& GFP_IOFS
) == GFP_IOFS
)
1296 return isolated
> inactive
;
1299 static noinline_for_stack
void
1300 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1302 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1303 struct zone
*zone
= lruvec_zone(lruvec
);
1304 LIST_HEAD(pages_to_free
);
1307 * Put back any unfreeable pages.
1309 while (!list_empty(page_list
)) {
1310 struct page
*page
= lru_to_page(page_list
);
1313 VM_BUG_ON(PageLRU(page
));
1314 list_del(&page
->lru
);
1315 if (unlikely(!page_evictable(page
))) {
1316 spin_unlock_irq(&zone
->lru_lock
);
1317 putback_lru_page(page
);
1318 spin_lock_irq(&zone
->lru_lock
);
1322 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1325 lru
= page_lru(page
);
1326 add_page_to_lru_list(page
, lruvec
, lru
);
1328 if (is_active_lru(lru
)) {
1329 int file
= is_file_lru(lru
);
1330 int numpages
= hpage_nr_pages(page
);
1331 reclaim_stat
->recent_rotated
[file
] += numpages
;
1333 if (put_page_testzero(page
)) {
1334 __ClearPageLRU(page
);
1335 __ClearPageActive(page
);
1336 del_page_from_lru_list(page
, lruvec
, lru
);
1338 if (unlikely(PageCompound(page
))) {
1339 spin_unlock_irq(&zone
->lru_lock
);
1340 (*get_compound_page_dtor(page
))(page
);
1341 spin_lock_irq(&zone
->lru_lock
);
1343 list_add(&page
->lru
, &pages_to_free
);
1348 * To save our caller's stack, now use input list for pages to free.
1350 list_splice(&pages_to_free
, page_list
);
1354 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1355 * of reclaimed pages
1357 static noinline_for_stack
unsigned long
1358 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1359 struct scan_control
*sc
, enum lru_list lru
)
1361 LIST_HEAD(page_list
);
1362 unsigned long nr_scanned
;
1363 unsigned long nr_reclaimed
= 0;
1364 unsigned long nr_taken
;
1365 unsigned long nr_dirty
= 0;
1366 unsigned long nr_congested
= 0;
1367 unsigned long nr_unqueued_dirty
= 0;
1368 unsigned long nr_writeback
= 0;
1369 unsigned long nr_immediate
= 0;
1370 isolate_mode_t isolate_mode
= 0;
1371 int file
= is_file_lru(lru
);
1372 struct zone
*zone
= lruvec_zone(lruvec
);
1373 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1375 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1376 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1378 /* We are about to die and free our memory. Return now. */
1379 if (fatal_signal_pending(current
))
1380 return SWAP_CLUSTER_MAX
;
1386 isolate_mode
|= ISOLATE_UNMAPPED
;
1387 if (!sc
->may_writepage
)
1388 isolate_mode
|= ISOLATE_CLEAN
;
1390 spin_lock_irq(&zone
->lru_lock
);
1392 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1393 &nr_scanned
, sc
, isolate_mode
, lru
);
1395 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1396 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1398 if (global_reclaim(sc
)) {
1399 zone
->pages_scanned
+= nr_scanned
;
1400 if (current_is_kswapd())
1401 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1403 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1405 spin_unlock_irq(&zone
->lru_lock
);
1410 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1411 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1412 &nr_writeback
, &nr_immediate
,
1415 spin_lock_irq(&zone
->lru_lock
);
1417 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1419 if (global_reclaim(sc
)) {
1420 if (current_is_kswapd())
1421 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1424 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1428 putback_inactive_pages(lruvec
, &page_list
);
1430 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1432 spin_unlock_irq(&zone
->lru_lock
);
1434 free_hot_cold_page_list(&page_list
, 1);
1437 * If reclaim is isolating dirty pages under writeback, it implies
1438 * that the long-lived page allocation rate is exceeding the page
1439 * laundering rate. Either the global limits are not being effective
1440 * at throttling processes due to the page distribution throughout
1441 * zones or there is heavy usage of a slow backing device. The
1442 * only option is to throttle from reclaim context which is not ideal
1443 * as there is no guarantee the dirtying process is throttled in the
1444 * same way balance_dirty_pages() manages.
1446 * This scales the number of dirty pages that must be under writeback
1447 * before a zone gets flagged ZONE_WRITEBACK. It is a simple backoff
1448 * function that has the most effect in the range DEF_PRIORITY to
1449 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1450 * in trouble and reclaim is considered to be in trouble.
1452 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1453 * DEF_PRIORITY-1 50% must be PageWriteback
1454 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1456 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1457 * isolated page is PageWriteback
1459 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1460 * of pages under pages flagged for immediate reclaim and stall if any
1461 * are encountered in the nr_immediate check below.
1463 if (nr_writeback
&& nr_writeback
>=
1464 (nr_taken
>> (DEF_PRIORITY
- sc
->priority
)))
1465 zone_set_flag(zone
, ZONE_WRITEBACK
);
1468 * memcg will stall in page writeback so only consider forcibly
1469 * stalling for global reclaim
1471 if (global_reclaim(sc
)) {
1473 * Tag a zone as congested if all the dirty pages scanned were
1474 * backed by a congested BDI and wait_iff_congested will stall.
1476 if (nr_dirty
&& nr_dirty
== nr_congested
)
1477 zone_set_flag(zone
, ZONE_CONGESTED
);
1480 * If dirty pages are scanned that are not queued for IO, it
1481 * implies that flushers are not keeping up. In this case, flag
1482 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1483 * pages from reclaim context. It will forcibly stall in the
1486 if (nr_unqueued_dirty
== nr_taken
)
1487 zone_set_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
1490 * In addition, if kswapd scans pages marked marked for
1491 * immediate reclaim and under writeback (nr_immediate), it
1492 * implies that pages are cycling through the LRU faster than
1493 * they are written so also forcibly stall.
1495 if (nr_unqueued_dirty
== nr_taken
|| nr_immediate
)
1496 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1500 * Stall direct reclaim for IO completions if underlying BDIs or zone
1501 * is congested. Allow kswapd to continue until it starts encountering
1502 * unqueued dirty pages or cycling through the LRU too quickly.
1504 if (!sc
->hibernation_mode
&& !current_is_kswapd())
1505 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1507 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1509 nr_scanned
, nr_reclaimed
,
1511 trace_shrink_flags(file
));
1512 return nr_reclaimed
;
1516 * This moves pages from the active list to the inactive list.
1518 * We move them the other way if the page is referenced by one or more
1519 * processes, from rmap.
1521 * If the pages are mostly unmapped, the processing is fast and it is
1522 * appropriate to hold zone->lru_lock across the whole operation. But if
1523 * the pages are mapped, the processing is slow (page_referenced()) so we
1524 * should drop zone->lru_lock around each page. It's impossible to balance
1525 * this, so instead we remove the pages from the LRU while processing them.
1526 * It is safe to rely on PG_active against the non-LRU pages in here because
1527 * nobody will play with that bit on a non-LRU page.
1529 * The downside is that we have to touch page->_count against each page.
1530 * But we had to alter page->flags anyway.
1533 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1534 struct list_head
*list
,
1535 struct list_head
*pages_to_free
,
1538 struct zone
*zone
= lruvec_zone(lruvec
);
1539 unsigned long pgmoved
= 0;
1543 while (!list_empty(list
)) {
1544 page
= lru_to_page(list
);
1545 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1547 VM_BUG_ON(PageLRU(page
));
1550 nr_pages
= hpage_nr_pages(page
);
1551 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1552 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1553 pgmoved
+= nr_pages
;
1555 if (put_page_testzero(page
)) {
1556 __ClearPageLRU(page
);
1557 __ClearPageActive(page
);
1558 del_page_from_lru_list(page
, lruvec
, lru
);
1560 if (unlikely(PageCompound(page
))) {
1561 spin_unlock_irq(&zone
->lru_lock
);
1562 (*get_compound_page_dtor(page
))(page
);
1563 spin_lock_irq(&zone
->lru_lock
);
1565 list_add(&page
->lru
, pages_to_free
);
1568 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1569 if (!is_active_lru(lru
))
1570 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1573 static void shrink_active_list(unsigned long nr_to_scan
,
1574 struct lruvec
*lruvec
,
1575 struct scan_control
*sc
,
1578 unsigned long nr_taken
;
1579 unsigned long nr_scanned
;
1580 unsigned long vm_flags
;
1581 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1582 LIST_HEAD(l_active
);
1583 LIST_HEAD(l_inactive
);
1585 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1586 unsigned long nr_rotated
= 0;
1587 isolate_mode_t isolate_mode
= 0;
1588 int file
= is_file_lru(lru
);
1589 struct zone
*zone
= lruvec_zone(lruvec
);
1594 isolate_mode
|= ISOLATE_UNMAPPED
;
1595 if (!sc
->may_writepage
)
1596 isolate_mode
|= ISOLATE_CLEAN
;
1598 spin_lock_irq(&zone
->lru_lock
);
1600 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1601 &nr_scanned
, sc
, isolate_mode
, lru
);
1602 if (global_reclaim(sc
))
1603 zone
->pages_scanned
+= nr_scanned
;
1605 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1607 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1608 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1609 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1610 spin_unlock_irq(&zone
->lru_lock
);
1612 while (!list_empty(&l_hold
)) {
1614 page
= lru_to_page(&l_hold
);
1615 list_del(&page
->lru
);
1617 if (unlikely(!page_evictable(page
))) {
1618 putback_lru_page(page
);
1622 if (unlikely(buffer_heads_over_limit
)) {
1623 if (page_has_private(page
) && trylock_page(page
)) {
1624 if (page_has_private(page
))
1625 try_to_release_page(page
, 0);
1630 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1632 nr_rotated
+= hpage_nr_pages(page
);
1634 * Identify referenced, file-backed active pages and
1635 * give them one more trip around the active list. So
1636 * that executable code get better chances to stay in
1637 * memory under moderate memory pressure. Anon pages
1638 * are not likely to be evicted by use-once streaming
1639 * IO, plus JVM can create lots of anon VM_EXEC pages,
1640 * so we ignore them here.
1642 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1643 list_add(&page
->lru
, &l_active
);
1648 ClearPageActive(page
); /* we are de-activating */
1649 list_add(&page
->lru
, &l_inactive
);
1653 * Move pages back to the lru list.
1655 spin_lock_irq(&zone
->lru_lock
);
1657 * Count referenced pages from currently used mappings as rotated,
1658 * even though only some of them are actually re-activated. This
1659 * helps balance scan pressure between file and anonymous pages in
1662 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1664 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1665 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1666 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1667 spin_unlock_irq(&zone
->lru_lock
);
1669 free_hot_cold_page_list(&l_hold
, 1);
1673 static int inactive_anon_is_low_global(struct zone
*zone
)
1675 unsigned long active
, inactive
;
1677 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1678 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1680 if (inactive
* zone
->inactive_ratio
< active
)
1687 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1688 * @lruvec: LRU vector to check
1690 * Returns true if the zone does not have enough inactive anon pages,
1691 * meaning some active anon pages need to be deactivated.
1693 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1696 * If we don't have swap space, anonymous page deactivation
1699 if (!total_swap_pages
)
1702 if (!mem_cgroup_disabled())
1703 return mem_cgroup_inactive_anon_is_low(lruvec
);
1705 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1708 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1715 * inactive_file_is_low - check if file pages need to be deactivated
1716 * @lruvec: LRU vector to check
1718 * When the system is doing streaming IO, memory pressure here
1719 * ensures that active file pages get deactivated, until more
1720 * than half of the file pages are on the inactive list.
1722 * Once we get to that situation, protect the system's working
1723 * set from being evicted by disabling active file page aging.
1725 * This uses a different ratio than the anonymous pages, because
1726 * the page cache uses a use-once replacement algorithm.
1728 static int inactive_file_is_low(struct lruvec
*lruvec
)
1730 unsigned long inactive
;
1731 unsigned long active
;
1733 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1734 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1736 return active
> inactive
;
1739 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1741 if (is_file_lru(lru
))
1742 return inactive_file_is_low(lruvec
);
1744 return inactive_anon_is_low(lruvec
);
1747 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1748 struct lruvec
*lruvec
, struct scan_control
*sc
)
1750 if (is_active_lru(lru
)) {
1751 if (inactive_list_is_low(lruvec
, lru
))
1752 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1756 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1759 static int vmscan_swappiness(struct scan_control
*sc
)
1761 if (global_reclaim(sc
))
1762 return vm_swappiness
;
1763 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1774 * Determine how aggressively the anon and file LRU lists should be
1775 * scanned. The relative value of each set of LRU lists is determined
1776 * by looking at the fraction of the pages scanned we did rotate back
1777 * onto the active list instead of evict.
1779 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1780 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1782 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
1785 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1787 u64 denominator
= 0; /* gcc */
1788 struct zone
*zone
= lruvec_zone(lruvec
);
1789 unsigned long anon_prio
, file_prio
;
1790 enum scan_balance scan_balance
;
1791 unsigned long anon
, file
, free
;
1792 bool force_scan
= false;
1793 unsigned long ap
, fp
;
1797 * If the zone or memcg is small, nr[l] can be 0. This
1798 * results in no scanning on this priority and a potential
1799 * priority drop. Global direct reclaim can go to the next
1800 * zone and tends to have no problems. Global kswapd is for
1801 * zone balancing and it needs to scan a minimum amount. When
1802 * reclaiming for a memcg, a priority drop can cause high
1803 * latencies, so it's better to scan a minimum amount there as
1806 if (current_is_kswapd() && zone
->all_unreclaimable
)
1808 if (!global_reclaim(sc
))
1811 /* If we have no swap space, do not bother scanning anon pages. */
1812 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
1813 scan_balance
= SCAN_FILE
;
1818 * Global reclaim will swap to prevent OOM even with no
1819 * swappiness, but memcg users want to use this knob to
1820 * disable swapping for individual groups completely when
1821 * using the memory controller's swap limit feature would be
1824 if (!global_reclaim(sc
) && !vmscan_swappiness(sc
)) {
1825 scan_balance
= SCAN_FILE
;
1830 * Do not apply any pressure balancing cleverness when the
1831 * system is close to OOM, scan both anon and file equally
1832 * (unless the swappiness setting disagrees with swapping).
1834 if (!sc
->priority
&& vmscan_swappiness(sc
)) {
1835 scan_balance
= SCAN_EQUAL
;
1839 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1840 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1841 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1842 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1845 * If it's foreseeable that reclaiming the file cache won't be
1846 * enough to get the zone back into a desirable shape, we have
1847 * to swap. Better start now and leave the - probably heavily
1848 * thrashing - remaining file pages alone.
1850 if (global_reclaim(sc
)) {
1851 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1852 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1853 scan_balance
= SCAN_ANON
;
1859 * There is enough inactive page cache, do not reclaim
1860 * anything from the anonymous working set right now.
1862 if (!inactive_file_is_low(lruvec
)) {
1863 scan_balance
= SCAN_FILE
;
1867 scan_balance
= SCAN_FRACT
;
1870 * With swappiness at 100, anonymous and file have the same priority.
1871 * This scanning priority is essentially the inverse of IO cost.
1873 anon_prio
= vmscan_swappiness(sc
);
1874 file_prio
= 200 - anon_prio
;
1877 * OK, so we have swap space and a fair amount of page cache
1878 * pages. We use the recently rotated / recently scanned
1879 * ratios to determine how valuable each cache is.
1881 * Because workloads change over time (and to avoid overflow)
1882 * we keep these statistics as a floating average, which ends
1883 * up weighing recent references more than old ones.
1885 * anon in [0], file in [1]
1887 spin_lock_irq(&zone
->lru_lock
);
1888 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1889 reclaim_stat
->recent_scanned
[0] /= 2;
1890 reclaim_stat
->recent_rotated
[0] /= 2;
1893 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1894 reclaim_stat
->recent_scanned
[1] /= 2;
1895 reclaim_stat
->recent_rotated
[1] /= 2;
1899 * The amount of pressure on anon vs file pages is inversely
1900 * proportional to the fraction of recently scanned pages on
1901 * each list that were recently referenced and in active use.
1903 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1904 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1906 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1907 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1908 spin_unlock_irq(&zone
->lru_lock
);
1912 denominator
= ap
+ fp
+ 1;
1914 for_each_evictable_lru(lru
) {
1915 int file
= is_file_lru(lru
);
1919 size
= get_lru_size(lruvec
, lru
);
1920 scan
= size
>> sc
->priority
;
1922 if (!scan
&& force_scan
)
1923 scan
= min(size
, SWAP_CLUSTER_MAX
);
1925 switch (scan_balance
) {
1927 /* Scan lists relative to size */
1931 * Scan types proportional to swappiness and
1932 * their relative recent reclaim efficiency.
1934 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1938 /* Scan one type exclusively */
1939 if ((scan_balance
== SCAN_FILE
) != file
)
1943 /* Look ma, no brain */
1951 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1953 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
1955 unsigned long nr
[NR_LRU_LISTS
];
1956 unsigned long targets
[NR_LRU_LISTS
];
1957 unsigned long nr_to_scan
;
1959 unsigned long nr_reclaimed
= 0;
1960 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1961 struct blk_plug plug
;
1962 bool scan_adjusted
= false;
1964 get_scan_count(lruvec
, sc
, nr
);
1966 /* Record the original scan target for proportional adjustments later */
1967 memcpy(targets
, nr
, sizeof(nr
));
1969 blk_start_plug(&plug
);
1970 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1971 nr
[LRU_INACTIVE_FILE
]) {
1972 unsigned long nr_anon
, nr_file
, percentage
;
1973 unsigned long nr_scanned
;
1975 for_each_evictable_lru(lru
) {
1977 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
1978 nr
[lru
] -= nr_to_scan
;
1980 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1985 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
1989 * For global direct reclaim, reclaim only the number of pages
1990 * requested. Less care is taken to scan proportionally as it
1991 * is more important to minimise direct reclaim stall latency
1992 * than it is to properly age the LRU lists.
1994 if (global_reclaim(sc
) && !current_is_kswapd())
1998 * For kswapd and memcg, reclaim at least the number of pages
1999 * requested. Ensure that the anon and file LRUs shrink
2000 * proportionally what was requested by get_scan_count(). We
2001 * stop reclaiming one LRU and reduce the amount scanning
2002 * proportional to the original scan target.
2004 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2005 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2007 if (nr_file
> nr_anon
) {
2008 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2009 targets
[LRU_ACTIVE_ANON
] + 1;
2011 percentage
= nr_anon
* 100 / scan_target
;
2013 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2014 targets
[LRU_ACTIVE_FILE
] + 1;
2016 percentage
= nr_file
* 100 / scan_target
;
2019 /* Stop scanning the smaller of the LRU */
2021 nr
[lru
+ LRU_ACTIVE
] = 0;
2024 * Recalculate the other LRU scan count based on its original
2025 * scan target and the percentage scanning already complete
2027 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2028 nr_scanned
= targets
[lru
] - nr
[lru
];
2029 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2030 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2033 nr_scanned
= targets
[lru
] - nr
[lru
];
2034 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2035 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2037 scan_adjusted
= true;
2039 blk_finish_plug(&plug
);
2040 sc
->nr_reclaimed
+= nr_reclaimed
;
2043 * Even if we did not try to evict anon pages at all, we want to
2044 * rebalance the anon lru active/inactive ratio.
2046 if (inactive_anon_is_low(lruvec
))
2047 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2048 sc
, LRU_ACTIVE_ANON
);
2050 throttle_vm_writeout(sc
->gfp_mask
);
2053 /* Use reclaim/compaction for costly allocs or under memory pressure */
2054 static bool in_reclaim_compaction(struct scan_control
*sc
)
2056 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2057 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2058 sc
->priority
< DEF_PRIORITY
- 2))
2065 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2066 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2067 * true if more pages should be reclaimed such that when the page allocator
2068 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2069 * It will give up earlier than that if there is difficulty reclaiming pages.
2071 static inline bool should_continue_reclaim(struct zone
*zone
,
2072 unsigned long nr_reclaimed
,
2073 unsigned long nr_scanned
,
2074 struct scan_control
*sc
)
2076 unsigned long pages_for_compaction
;
2077 unsigned long inactive_lru_pages
;
2079 /* If not in reclaim/compaction mode, stop */
2080 if (!in_reclaim_compaction(sc
))
2083 /* Consider stopping depending on scan and reclaim activity */
2084 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2086 * For __GFP_REPEAT allocations, stop reclaiming if the
2087 * full LRU list has been scanned and we are still failing
2088 * to reclaim pages. This full LRU scan is potentially
2089 * expensive but a __GFP_REPEAT caller really wants to succeed
2091 if (!nr_reclaimed
&& !nr_scanned
)
2095 * For non-__GFP_REPEAT allocations which can presumably
2096 * fail without consequence, stop if we failed to reclaim
2097 * any pages from the last SWAP_CLUSTER_MAX number of
2098 * pages that were scanned. This will return to the
2099 * caller faster at the risk reclaim/compaction and
2100 * the resulting allocation attempt fails
2107 * If we have not reclaimed enough pages for compaction and the
2108 * inactive lists are large enough, continue reclaiming
2110 pages_for_compaction
= (2UL << sc
->order
);
2111 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2112 if (get_nr_swap_pages() > 0)
2113 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2114 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2115 inactive_lru_pages
> pages_for_compaction
)
2118 /* If compaction would go ahead or the allocation would succeed, stop */
2119 switch (compaction_suitable(zone
, sc
->order
)) {
2120 case COMPACT_PARTIAL
:
2121 case COMPACT_CONTINUE
:
2128 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
2130 unsigned long nr_reclaimed
, nr_scanned
;
2133 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2134 struct mem_cgroup_reclaim_cookie reclaim
= {
2136 .priority
= sc
->priority
,
2138 struct mem_cgroup
*memcg
;
2140 nr_reclaimed
= sc
->nr_reclaimed
;
2141 nr_scanned
= sc
->nr_scanned
;
2143 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2145 struct lruvec
*lruvec
;
2147 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2149 shrink_lruvec(lruvec
, sc
);
2152 * Direct reclaim and kswapd have to scan all memory
2153 * cgroups to fulfill the overall scan target for the
2156 * Limit reclaim, on the other hand, only cares about
2157 * nr_to_reclaim pages to be reclaimed and it will
2158 * retry with decreasing priority if one round over the
2159 * whole hierarchy is not sufficient.
2161 if (!global_reclaim(sc
) &&
2162 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2163 mem_cgroup_iter_break(root
, memcg
);
2166 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
2169 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2170 sc
->nr_scanned
- nr_scanned
,
2171 sc
->nr_reclaimed
- nr_reclaimed
);
2173 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2174 sc
->nr_scanned
- nr_scanned
, sc
));
2177 /* Returns true if compaction should go ahead for a high-order request */
2178 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2180 unsigned long balance_gap
, watermark
;
2183 /* Do not consider compaction for orders reclaim is meant to satisfy */
2184 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
2188 * Compaction takes time to run and there are potentially other
2189 * callers using the pages just freed. Continue reclaiming until
2190 * there is a buffer of free pages available to give compaction
2191 * a reasonable chance of completing and allocating the page
2193 balance_gap
= min(low_wmark_pages(zone
),
2194 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2195 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2196 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
2197 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2200 * If compaction is deferred, reclaim up to a point where
2201 * compaction will have a chance of success when re-enabled
2203 if (compaction_deferred(zone
, sc
->order
))
2204 return watermark_ok
;
2206 /* If compaction is not ready to start, keep reclaiming */
2207 if (!compaction_suitable(zone
, sc
->order
))
2210 return watermark_ok
;
2214 * This is the direct reclaim path, for page-allocating processes. We only
2215 * try to reclaim pages from zones which will satisfy the caller's allocation
2218 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2220 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2222 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2223 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2224 * zone defense algorithm.
2226 * If a zone is deemed to be full of pinned pages then just give it a light
2227 * scan then give up on it.
2229 * This function returns true if a zone is being reclaimed for a costly
2230 * high-order allocation and compaction is ready to begin. This indicates to
2231 * the caller that it should consider retrying the allocation instead of
2234 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2238 unsigned long nr_soft_reclaimed
;
2239 unsigned long nr_soft_scanned
;
2240 bool aborted_reclaim
= false;
2243 * If the number of buffer_heads in the machine exceeds the maximum
2244 * allowed level, force direct reclaim to scan the highmem zone as
2245 * highmem pages could be pinning lowmem pages storing buffer_heads
2247 if (buffer_heads_over_limit
)
2248 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2250 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2251 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2252 if (!populated_zone(zone
))
2255 * Take care memory controller reclaiming has small influence
2258 if (global_reclaim(sc
)) {
2259 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2261 if (zone
->all_unreclaimable
&&
2262 sc
->priority
!= DEF_PRIORITY
)
2263 continue; /* Let kswapd poll it */
2264 if (IS_ENABLED(CONFIG_COMPACTION
)) {
2266 * If we already have plenty of memory free for
2267 * compaction in this zone, don't free any more.
2268 * Even though compaction is invoked for any
2269 * non-zero order, only frequent costly order
2270 * reclamation is disruptive enough to become a
2271 * noticeable problem, like transparent huge
2274 if (compaction_ready(zone
, sc
)) {
2275 aborted_reclaim
= true;
2280 * This steals pages from memory cgroups over softlimit
2281 * and returns the number of reclaimed pages and
2282 * scanned pages. This works for global memory pressure
2283 * and balancing, not for a memcg's limit.
2285 nr_soft_scanned
= 0;
2286 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2287 sc
->order
, sc
->gfp_mask
,
2289 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2290 sc
->nr_scanned
+= nr_soft_scanned
;
2291 /* need some check for avoid more shrink_zone() */
2294 shrink_zone(zone
, sc
);
2297 return aborted_reclaim
;
2300 static bool zone_reclaimable(struct zone
*zone
)
2302 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2305 /* All zones in zonelist are unreclaimable? */
2306 static bool all_unreclaimable(struct zonelist
*zonelist
,
2307 struct scan_control
*sc
)
2312 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2313 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2314 if (!populated_zone(zone
))
2316 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2318 if (!zone
->all_unreclaimable
)
2326 * This is the main entry point to direct page reclaim.
2328 * If a full scan of the inactive list fails to free enough memory then we
2329 * are "out of memory" and something needs to be killed.
2331 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2332 * high - the zone may be full of dirty or under-writeback pages, which this
2333 * caller can't do much about. We kick the writeback threads and take explicit
2334 * naps in the hope that some of these pages can be written. But if the
2335 * allocating task holds filesystem locks which prevent writeout this might not
2336 * work, and the allocation attempt will fail.
2338 * returns: 0, if no pages reclaimed
2339 * else, the number of pages reclaimed
2341 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2342 struct scan_control
*sc
,
2343 struct shrink_control
*shrink
)
2345 unsigned long total_scanned
= 0;
2346 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2349 unsigned long writeback_threshold
;
2350 bool aborted_reclaim
;
2352 delayacct_freepages_start();
2354 if (global_reclaim(sc
))
2355 count_vm_event(ALLOCSTALL
);
2358 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2361 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2364 * Don't shrink slabs when reclaiming memory from over limit
2365 * cgroups but do shrink slab at least once when aborting
2366 * reclaim for compaction to avoid unevenly scanning file/anon
2367 * LRU pages over slab pages.
2369 if (global_reclaim(sc
)) {
2370 unsigned long lru_pages
= 0;
2371 for_each_zone_zonelist(zone
, z
, zonelist
,
2372 gfp_zone(sc
->gfp_mask
)) {
2373 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2376 lru_pages
+= zone_reclaimable_pages(zone
);
2379 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2380 if (reclaim_state
) {
2381 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2382 reclaim_state
->reclaimed_slab
= 0;
2385 total_scanned
+= sc
->nr_scanned
;
2386 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2390 * If we're getting trouble reclaiming, start doing
2391 * writepage even in laptop mode.
2393 if (sc
->priority
< DEF_PRIORITY
- 2)
2394 sc
->may_writepage
= 1;
2397 * Try to write back as many pages as we just scanned. This
2398 * tends to cause slow streaming writers to write data to the
2399 * disk smoothly, at the dirtying rate, which is nice. But
2400 * that's undesirable in laptop mode, where we *want* lumpy
2401 * writeout. So in laptop mode, write out the whole world.
2403 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2404 if (total_scanned
> writeback_threshold
) {
2405 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2406 WB_REASON_TRY_TO_FREE_PAGES
);
2407 sc
->may_writepage
= 1;
2409 } while (--sc
->priority
>= 0 && !aborted_reclaim
);
2412 delayacct_freepages_end();
2414 if (sc
->nr_reclaimed
)
2415 return sc
->nr_reclaimed
;
2418 * As hibernation is going on, kswapd is freezed so that it can't mark
2419 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2422 if (oom_killer_disabled
)
2425 /* Aborted reclaim to try compaction? don't OOM, then */
2426 if (aborted_reclaim
)
2429 /* top priority shrink_zones still had more to do? don't OOM, then */
2430 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2436 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2439 unsigned long pfmemalloc_reserve
= 0;
2440 unsigned long free_pages
= 0;
2444 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2445 zone
= &pgdat
->node_zones
[i
];
2446 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2447 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2450 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2452 /* kswapd must be awake if processes are being throttled */
2453 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2454 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2455 (enum zone_type
)ZONE_NORMAL
);
2456 wake_up_interruptible(&pgdat
->kswapd_wait
);
2463 * Throttle direct reclaimers if backing storage is backed by the network
2464 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2465 * depleted. kswapd will continue to make progress and wake the processes
2466 * when the low watermark is reached.
2468 * Returns true if a fatal signal was delivered during throttling. If this
2469 * happens, the page allocator should not consider triggering the OOM killer.
2471 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2472 nodemask_t
*nodemask
)
2475 int high_zoneidx
= gfp_zone(gfp_mask
);
2479 * Kernel threads should not be throttled as they may be indirectly
2480 * responsible for cleaning pages necessary for reclaim to make forward
2481 * progress. kjournald for example may enter direct reclaim while
2482 * committing a transaction where throttling it could forcing other
2483 * processes to block on log_wait_commit().
2485 if (current
->flags
& PF_KTHREAD
)
2489 * If a fatal signal is pending, this process should not throttle.
2490 * It should return quickly so it can exit and free its memory
2492 if (fatal_signal_pending(current
))
2495 /* Check if the pfmemalloc reserves are ok */
2496 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
, &zone
);
2497 pgdat
= zone
->zone_pgdat
;
2498 if (pfmemalloc_watermark_ok(pgdat
))
2501 /* Account for the throttling */
2502 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2505 * If the caller cannot enter the filesystem, it's possible that it
2506 * is due to the caller holding an FS lock or performing a journal
2507 * transaction in the case of a filesystem like ext[3|4]. In this case,
2508 * it is not safe to block on pfmemalloc_wait as kswapd could be
2509 * blocked waiting on the same lock. Instead, throttle for up to a
2510 * second before continuing.
2512 if (!(gfp_mask
& __GFP_FS
)) {
2513 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2514 pfmemalloc_watermark_ok(pgdat
), HZ
);
2519 /* Throttle until kswapd wakes the process */
2520 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2521 pfmemalloc_watermark_ok(pgdat
));
2524 if (fatal_signal_pending(current
))
2531 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2532 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2534 unsigned long nr_reclaimed
;
2535 struct scan_control sc
= {
2536 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2537 .may_writepage
= !laptop_mode
,
2538 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2542 .priority
= DEF_PRIORITY
,
2543 .target_mem_cgroup
= NULL
,
2544 .nodemask
= nodemask
,
2546 struct shrink_control shrink
= {
2547 .gfp_mask
= sc
.gfp_mask
,
2551 * Do not enter reclaim if fatal signal was delivered while throttled.
2552 * 1 is returned so that the page allocator does not OOM kill at this
2555 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2558 trace_mm_vmscan_direct_reclaim_begin(order
,
2562 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2564 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2566 return nr_reclaimed
;
2571 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2572 gfp_t gfp_mask
, bool noswap
,
2574 unsigned long *nr_scanned
)
2576 struct scan_control sc
= {
2578 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2579 .may_writepage
= !laptop_mode
,
2581 .may_swap
= !noswap
,
2584 .target_mem_cgroup
= memcg
,
2586 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2588 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2589 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2591 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2596 * NOTE: Although we can get the priority field, using it
2597 * here is not a good idea, since it limits the pages we can scan.
2598 * if we don't reclaim here, the shrink_zone from balance_pgdat
2599 * will pick up pages from other mem cgroup's as well. We hack
2600 * the priority and make it zero.
2602 shrink_lruvec(lruvec
, &sc
);
2604 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2606 *nr_scanned
= sc
.nr_scanned
;
2607 return sc
.nr_reclaimed
;
2610 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2614 struct zonelist
*zonelist
;
2615 unsigned long nr_reclaimed
;
2617 struct scan_control sc
= {
2618 .may_writepage
= !laptop_mode
,
2620 .may_swap
= !noswap
,
2621 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2623 .priority
= DEF_PRIORITY
,
2624 .target_mem_cgroup
= memcg
,
2625 .nodemask
= NULL
, /* we don't care the placement */
2626 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2627 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2629 struct shrink_control shrink
= {
2630 .gfp_mask
= sc
.gfp_mask
,
2634 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2635 * take care of from where we get pages. So the node where we start the
2636 * scan does not need to be the current node.
2638 nid
= mem_cgroup_select_victim_node(memcg
);
2640 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2642 trace_mm_vmscan_memcg_reclaim_begin(0,
2646 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2648 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2650 return nr_reclaimed
;
2654 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2656 struct mem_cgroup
*memcg
;
2658 if (!total_swap_pages
)
2661 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2663 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2665 if (inactive_anon_is_low(lruvec
))
2666 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2667 sc
, LRU_ACTIVE_ANON
);
2669 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2673 static bool zone_balanced(struct zone
*zone
, int order
,
2674 unsigned long balance_gap
, int classzone_idx
)
2676 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2677 balance_gap
, classzone_idx
, 0))
2680 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&&
2681 !compaction_suitable(zone
, order
))
2688 * pgdat_balanced() is used when checking if a node is balanced.
2690 * For order-0, all zones must be balanced!
2692 * For high-order allocations only zones that meet watermarks and are in a
2693 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2694 * total of balanced pages must be at least 25% of the zones allowed by
2695 * classzone_idx for the node to be considered balanced. Forcing all zones to
2696 * be balanced for high orders can cause excessive reclaim when there are
2698 * The choice of 25% is due to
2699 * o a 16M DMA zone that is balanced will not balance a zone on any
2700 * reasonable sized machine
2701 * o On all other machines, the top zone must be at least a reasonable
2702 * percentage of the middle zones. For example, on 32-bit x86, highmem
2703 * would need to be at least 256M for it to be balance a whole node.
2704 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2705 * to balance a node on its own. These seemed like reasonable ratios.
2707 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2709 unsigned long managed_pages
= 0;
2710 unsigned long balanced_pages
= 0;
2713 /* Check the watermark levels */
2714 for (i
= 0; i
<= classzone_idx
; i
++) {
2715 struct zone
*zone
= pgdat
->node_zones
+ i
;
2717 if (!populated_zone(zone
))
2720 managed_pages
+= zone
->managed_pages
;
2723 * A special case here:
2725 * balance_pgdat() skips over all_unreclaimable after
2726 * DEF_PRIORITY. Effectively, it considers them balanced so
2727 * they must be considered balanced here as well!
2729 if (zone
->all_unreclaimable
) {
2730 balanced_pages
+= zone
->managed_pages
;
2734 if (zone_balanced(zone
, order
, 0, i
))
2735 balanced_pages
+= zone
->managed_pages
;
2741 return balanced_pages
>= (managed_pages
>> 2);
2747 * Prepare kswapd for sleeping. This verifies that there are no processes
2748 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2750 * Returns true if kswapd is ready to sleep
2752 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
2755 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2760 * There is a potential race between when kswapd checks its watermarks
2761 * and a process gets throttled. There is also a potential race if
2762 * processes get throttled, kswapd wakes, a large process exits therby
2763 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2764 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2765 * so wake them now if necessary. If necessary, processes will wake
2766 * kswapd and get throttled again
2768 if (waitqueue_active(&pgdat
->pfmemalloc_wait
)) {
2769 wake_up(&pgdat
->pfmemalloc_wait
);
2773 return pgdat_balanced(pgdat
, order
, classzone_idx
);
2777 * kswapd shrinks the zone by the number of pages required to reach
2778 * the high watermark.
2780 * Returns true if kswapd scanned at least the requested number of pages to
2781 * reclaim or if the lack of progress was due to pages under writeback.
2782 * This is used to determine if the scanning priority needs to be raised.
2784 static bool kswapd_shrink_zone(struct zone
*zone
,
2786 struct scan_control
*sc
,
2787 unsigned long lru_pages
,
2788 unsigned long *nr_attempted
)
2790 unsigned long nr_slab
;
2791 int testorder
= sc
->order
;
2792 unsigned long balance_gap
;
2793 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2794 struct shrink_control shrink
= {
2795 .gfp_mask
= sc
->gfp_mask
,
2797 bool lowmem_pressure
;
2799 /* Reclaim above the high watermark. */
2800 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
2803 * Kswapd reclaims only single pages with compaction enabled. Trying
2804 * too hard to reclaim until contiguous free pages have become
2805 * available can hurt performance by evicting too much useful data
2806 * from memory. Do not reclaim more than needed for compaction.
2808 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2809 compaction_suitable(zone
, sc
->order
) !=
2814 * We put equal pressure on every zone, unless one zone has way too
2815 * many pages free already. The "too many pages" is defined as the
2816 * high wmark plus a "gap" where the gap is either the low
2817 * watermark or 1% of the zone, whichever is smaller.
2819 balance_gap
= min(low_wmark_pages(zone
),
2820 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2821 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2824 * If there is no low memory pressure or the zone is balanced then no
2825 * reclaim is necessary
2827 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
2828 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
2829 balance_gap
, classzone_idx
))
2832 shrink_zone(zone
, sc
);
2834 reclaim_state
->reclaimed_slab
= 0;
2835 nr_slab
= shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2836 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2838 /* Account for the number of pages attempted to reclaim */
2839 *nr_attempted
+= sc
->nr_to_reclaim
;
2841 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2842 zone
->all_unreclaimable
= 1;
2844 zone_clear_flag(zone
, ZONE_WRITEBACK
);
2847 * If a zone reaches its high watermark, consider it to be no longer
2848 * congested. It's possible there are dirty pages backed by congested
2849 * BDIs but as pressure is relieved, speculatively avoid congestion
2852 if (!zone
->all_unreclaimable
&&
2853 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
2854 zone_clear_flag(zone
, ZONE_CONGESTED
);
2855 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2858 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
2862 * For kswapd, balance_pgdat() will work across all this node's zones until
2863 * they are all at high_wmark_pages(zone).
2865 * Returns the final order kswapd was reclaiming at
2867 * There is special handling here for zones which are full of pinned pages.
2868 * This can happen if the pages are all mlocked, or if they are all used by
2869 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2870 * What we do is to detect the case where all pages in the zone have been
2871 * scanned twice and there has been zero successful reclaim. Mark the zone as
2872 * dead and from now on, only perform a short scan. Basically we're polling
2873 * the zone for when the problem goes away.
2875 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2876 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2877 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2878 * lower zones regardless of the number of free pages in the lower zones. This
2879 * interoperates with the page allocator fallback scheme to ensure that aging
2880 * of pages is balanced across the zones.
2882 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2886 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2887 unsigned long nr_soft_reclaimed
;
2888 unsigned long nr_soft_scanned
;
2889 struct scan_control sc
= {
2890 .gfp_mask
= GFP_KERNEL
,
2891 .priority
= DEF_PRIORITY
,
2894 .may_writepage
= !laptop_mode
,
2896 .target_mem_cgroup
= NULL
,
2898 count_vm_event(PAGEOUTRUN
);
2901 unsigned long lru_pages
= 0;
2902 unsigned long nr_attempted
= 0;
2903 bool raise_priority
= true;
2904 bool pgdat_needs_compaction
= (order
> 0);
2906 sc
.nr_reclaimed
= 0;
2909 * Scan in the highmem->dma direction for the highest
2910 * zone which needs scanning
2912 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2913 struct zone
*zone
= pgdat
->node_zones
+ i
;
2915 if (!populated_zone(zone
))
2918 if (zone
->all_unreclaimable
&&
2919 sc
.priority
!= DEF_PRIORITY
)
2923 * Do some background aging of the anon list, to give
2924 * pages a chance to be referenced before reclaiming.
2926 age_active_anon(zone
, &sc
);
2929 * If the number of buffer_heads in the machine
2930 * exceeds the maximum allowed level and this node
2931 * has a highmem zone, force kswapd to reclaim from
2932 * it to relieve lowmem pressure.
2934 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2939 if (!zone_balanced(zone
, order
, 0, 0)) {
2944 * If balanced, clear the dirty and congested
2947 zone_clear_flag(zone
, ZONE_CONGESTED
);
2948 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2955 for (i
= 0; i
<= end_zone
; i
++) {
2956 struct zone
*zone
= pgdat
->node_zones
+ i
;
2958 if (!populated_zone(zone
))
2961 lru_pages
+= zone_reclaimable_pages(zone
);
2964 * If any zone is currently balanced then kswapd will
2965 * not call compaction as it is expected that the
2966 * necessary pages are already available.
2968 if (pgdat_needs_compaction
&&
2969 zone_watermark_ok(zone
, order
,
2970 low_wmark_pages(zone
),
2972 pgdat_needs_compaction
= false;
2976 * If we're getting trouble reclaiming, start doing writepage
2977 * even in laptop mode.
2979 if (sc
.priority
< DEF_PRIORITY
- 2)
2980 sc
.may_writepage
= 1;
2983 * Now scan the zone in the dma->highmem direction, stopping
2984 * at the last zone which needs scanning.
2986 * We do this because the page allocator works in the opposite
2987 * direction. This prevents the page allocator from allocating
2988 * pages behind kswapd's direction of progress, which would
2989 * cause too much scanning of the lower zones.
2991 for (i
= 0; i
<= end_zone
; i
++) {
2992 struct zone
*zone
= pgdat
->node_zones
+ i
;
2994 if (!populated_zone(zone
))
2997 if (zone
->all_unreclaimable
&&
2998 sc
.priority
!= DEF_PRIORITY
)
3003 nr_soft_scanned
= 0;
3005 * Call soft limit reclaim before calling shrink_zone.
3007 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3010 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3013 * There should be no need to raise the scanning
3014 * priority if enough pages are already being scanned
3015 * that that high watermark would be met at 100%
3018 if (kswapd_shrink_zone(zone
, end_zone
, &sc
,
3019 lru_pages
, &nr_attempted
))
3020 raise_priority
= false;
3024 * If the low watermark is met there is no need for processes
3025 * to be throttled on pfmemalloc_wait as they should not be
3026 * able to safely make forward progress. Wake them
3028 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3029 pfmemalloc_watermark_ok(pgdat
))
3030 wake_up(&pgdat
->pfmemalloc_wait
);
3033 * Fragmentation may mean that the system cannot be rebalanced
3034 * for high-order allocations in all zones. If twice the
3035 * allocation size has been reclaimed and the zones are still
3036 * not balanced then recheck the watermarks at order-0 to
3037 * prevent kswapd reclaiming excessively. Assume that a
3038 * process requested a high-order can direct reclaim/compact.
3040 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3041 order
= sc
.order
= 0;
3043 /* Check if kswapd should be suspending */
3044 if (try_to_freeze() || kthread_should_stop())
3048 * Compact if necessary and kswapd is reclaiming at least the
3049 * high watermark number of pages as requsted
3051 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3052 compact_pgdat(pgdat
, order
);
3055 * Raise priority if scanning rate is too low or there was no
3056 * progress in reclaiming pages
3058 if (raise_priority
|| !sc
.nr_reclaimed
)
3060 } while (sc
.priority
>= 1 &&
3061 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3065 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3066 * makes a decision on the order we were last reclaiming at. However,
3067 * if another caller entered the allocator slow path while kswapd
3068 * was awake, order will remain at the higher level
3070 *classzone_idx
= end_zone
;
3074 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3079 if (freezing(current
) || kthread_should_stop())
3082 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3084 /* Try to sleep for a short interval */
3085 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3086 remaining
= schedule_timeout(HZ
/10);
3087 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3088 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3092 * After a short sleep, check if it was a premature sleep. If not, then
3093 * go fully to sleep until explicitly woken up.
3095 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3096 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3099 * vmstat counters are not perfectly accurate and the estimated
3100 * value for counters such as NR_FREE_PAGES can deviate from the
3101 * true value by nr_online_cpus * threshold. To avoid the zone
3102 * watermarks being breached while under pressure, we reduce the
3103 * per-cpu vmstat threshold while kswapd is awake and restore
3104 * them before going back to sleep.
3106 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3109 * Compaction records what page blocks it recently failed to
3110 * isolate pages from and skips them in the future scanning.
3111 * When kswapd is going to sleep, it is reasonable to assume
3112 * that pages and compaction may succeed so reset the cache.
3114 reset_isolation_suitable(pgdat
);
3116 if (!kthread_should_stop())
3119 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3122 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3124 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3126 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3130 * The background pageout daemon, started as a kernel thread
3131 * from the init process.
3133 * This basically trickles out pages so that we have _some_
3134 * free memory available even if there is no other activity
3135 * that frees anything up. This is needed for things like routing
3136 * etc, where we otherwise might have all activity going on in
3137 * asynchronous contexts that cannot page things out.
3139 * If there are applications that are active memory-allocators
3140 * (most normal use), this basically shouldn't matter.
3142 static int kswapd(void *p
)
3144 unsigned long order
, new_order
;
3145 unsigned balanced_order
;
3146 int classzone_idx
, new_classzone_idx
;
3147 int balanced_classzone_idx
;
3148 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3149 struct task_struct
*tsk
= current
;
3151 struct reclaim_state reclaim_state
= {
3152 .reclaimed_slab
= 0,
3154 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3156 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3158 if (!cpumask_empty(cpumask
))
3159 set_cpus_allowed_ptr(tsk
, cpumask
);
3160 current
->reclaim_state
= &reclaim_state
;
3163 * Tell the memory management that we're a "memory allocator",
3164 * and that if we need more memory we should get access to it
3165 * regardless (see "__alloc_pages()"). "kswapd" should
3166 * never get caught in the normal page freeing logic.
3168 * (Kswapd normally doesn't need memory anyway, but sometimes
3169 * you need a small amount of memory in order to be able to
3170 * page out something else, and this flag essentially protects
3171 * us from recursively trying to free more memory as we're
3172 * trying to free the first piece of memory in the first place).
3174 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3177 order
= new_order
= 0;
3179 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3180 balanced_classzone_idx
= classzone_idx
;
3185 * If the last balance_pgdat was unsuccessful it's unlikely a
3186 * new request of a similar or harder type will succeed soon
3187 * so consider going to sleep on the basis we reclaimed at
3189 if (balanced_classzone_idx
>= new_classzone_idx
&&
3190 balanced_order
== new_order
) {
3191 new_order
= pgdat
->kswapd_max_order
;
3192 new_classzone_idx
= pgdat
->classzone_idx
;
3193 pgdat
->kswapd_max_order
= 0;
3194 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3197 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3199 * Don't sleep if someone wants a larger 'order'
3200 * allocation or has tigher zone constraints
3203 classzone_idx
= new_classzone_idx
;
3205 kswapd_try_to_sleep(pgdat
, balanced_order
,
3206 balanced_classzone_idx
);
3207 order
= pgdat
->kswapd_max_order
;
3208 classzone_idx
= pgdat
->classzone_idx
;
3210 new_classzone_idx
= classzone_idx
;
3211 pgdat
->kswapd_max_order
= 0;
3212 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3215 ret
= try_to_freeze();
3216 if (kthread_should_stop())
3220 * We can speed up thawing tasks if we don't call balance_pgdat
3221 * after returning from the refrigerator
3224 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3225 balanced_classzone_idx
= classzone_idx
;
3226 balanced_order
= balance_pgdat(pgdat
, order
,
3227 &balanced_classzone_idx
);
3231 current
->reclaim_state
= NULL
;
3236 * A zone is low on free memory, so wake its kswapd task to service it.
3238 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3242 if (!populated_zone(zone
))
3245 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
3247 pgdat
= zone
->zone_pgdat
;
3248 if (pgdat
->kswapd_max_order
< order
) {
3249 pgdat
->kswapd_max_order
= order
;
3250 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3252 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3254 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
3257 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3258 wake_up_interruptible(&pgdat
->kswapd_wait
);
3262 * The reclaimable count would be mostly accurate.
3263 * The less reclaimable pages may be
3264 * - mlocked pages, which will be moved to unevictable list when encountered
3265 * - mapped pages, which may require several travels to be reclaimed
3266 * - dirty pages, which is not "instantly" reclaimable
3268 unsigned long global_reclaimable_pages(void)
3272 nr
= global_page_state(NR_ACTIVE_FILE
) +
3273 global_page_state(NR_INACTIVE_FILE
);
3275 if (get_nr_swap_pages() > 0)
3276 nr
+= global_page_state(NR_ACTIVE_ANON
) +
3277 global_page_state(NR_INACTIVE_ANON
);
3282 unsigned long zone_reclaimable_pages(struct zone
*zone
)
3286 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
3287 zone_page_state(zone
, NR_INACTIVE_FILE
);
3289 if (get_nr_swap_pages() > 0)
3290 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
3291 zone_page_state(zone
, NR_INACTIVE_ANON
);
3296 #ifdef CONFIG_HIBERNATION
3298 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3301 * Rather than trying to age LRUs the aim is to preserve the overall
3302 * LRU order by reclaiming preferentially
3303 * inactive > active > active referenced > active mapped
3305 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3307 struct reclaim_state reclaim_state
;
3308 struct scan_control sc
= {
3309 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3313 .nr_to_reclaim
= nr_to_reclaim
,
3314 .hibernation_mode
= 1,
3316 .priority
= DEF_PRIORITY
,
3318 struct shrink_control shrink
= {
3319 .gfp_mask
= sc
.gfp_mask
,
3321 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3322 struct task_struct
*p
= current
;
3323 unsigned long nr_reclaimed
;
3325 p
->flags
|= PF_MEMALLOC
;
3326 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3327 reclaim_state
.reclaimed_slab
= 0;
3328 p
->reclaim_state
= &reclaim_state
;
3330 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3332 p
->reclaim_state
= NULL
;
3333 lockdep_clear_current_reclaim_state();
3334 p
->flags
&= ~PF_MEMALLOC
;
3336 return nr_reclaimed
;
3338 #endif /* CONFIG_HIBERNATION */
3340 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3341 not required for correctness. So if the last cpu in a node goes
3342 away, we get changed to run anywhere: as the first one comes back,
3343 restore their cpu bindings. */
3344 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3349 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3350 for_each_node_state(nid
, N_MEMORY
) {
3351 pg_data_t
*pgdat
= NODE_DATA(nid
);
3352 const struct cpumask
*mask
;
3354 mask
= cpumask_of_node(pgdat
->node_id
);
3356 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3357 /* One of our CPUs online: restore mask */
3358 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3365 * This kswapd start function will be called by init and node-hot-add.
3366 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3368 int kswapd_run(int nid
)
3370 pg_data_t
*pgdat
= NODE_DATA(nid
);
3376 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3377 if (IS_ERR(pgdat
->kswapd
)) {
3378 /* failure at boot is fatal */
3379 BUG_ON(system_state
== SYSTEM_BOOTING
);
3380 pr_err("Failed to start kswapd on node %d\n", nid
);
3381 ret
= PTR_ERR(pgdat
->kswapd
);
3382 pgdat
->kswapd
= NULL
;
3388 * Called by memory hotplug when all memory in a node is offlined. Caller must
3389 * hold lock_memory_hotplug().
3391 void kswapd_stop(int nid
)
3393 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3396 kthread_stop(kswapd
);
3397 NODE_DATA(nid
)->kswapd
= NULL
;
3401 static int __init
kswapd_init(void)
3406 for_each_node_state(nid
, N_MEMORY
)
3408 hotcpu_notifier(cpu_callback
, 0);
3412 module_init(kswapd_init
)
3418 * If non-zero call zone_reclaim when the number of free pages falls below
3421 int zone_reclaim_mode __read_mostly
;
3423 #define RECLAIM_OFF 0
3424 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3425 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3426 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3429 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3430 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3433 #define ZONE_RECLAIM_PRIORITY 4
3436 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3439 int sysctl_min_unmapped_ratio
= 1;
3442 * If the number of slab pages in a zone grows beyond this percentage then
3443 * slab reclaim needs to occur.
3445 int sysctl_min_slab_ratio
= 5;
3447 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3449 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3450 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3451 zone_page_state(zone
, NR_ACTIVE_FILE
);
3454 * It's possible for there to be more file mapped pages than
3455 * accounted for by the pages on the file LRU lists because
3456 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3458 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3461 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3462 static long zone_pagecache_reclaimable(struct zone
*zone
)
3464 long nr_pagecache_reclaimable
;
3468 * If RECLAIM_SWAP is set, then all file pages are considered
3469 * potentially reclaimable. Otherwise, we have to worry about
3470 * pages like swapcache and zone_unmapped_file_pages() provides
3473 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3474 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3476 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3478 /* If we can't clean pages, remove dirty pages from consideration */
3479 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3480 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3482 /* Watch for any possible underflows due to delta */
3483 if (unlikely(delta
> nr_pagecache_reclaimable
))
3484 delta
= nr_pagecache_reclaimable
;
3486 return nr_pagecache_reclaimable
- delta
;
3490 * Try to free up some pages from this zone through reclaim.
3492 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3494 /* Minimum pages needed in order to stay on node */
3495 const unsigned long nr_pages
= 1 << order
;
3496 struct task_struct
*p
= current
;
3497 struct reclaim_state reclaim_state
;
3498 struct scan_control sc
= {
3499 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3500 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3502 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3503 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3505 .priority
= ZONE_RECLAIM_PRIORITY
,
3507 struct shrink_control shrink
= {
3508 .gfp_mask
= sc
.gfp_mask
,
3510 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3514 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3515 * and we also need to be able to write out pages for RECLAIM_WRITE
3518 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3519 lockdep_set_current_reclaim_state(gfp_mask
);
3520 reclaim_state
.reclaimed_slab
= 0;
3521 p
->reclaim_state
= &reclaim_state
;
3523 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3525 * Free memory by calling shrink zone with increasing
3526 * priorities until we have enough memory freed.
3529 shrink_zone(zone
, &sc
);
3530 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3533 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3534 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3536 * shrink_slab() does not currently allow us to determine how
3537 * many pages were freed in this zone. So we take the current
3538 * number of slab pages and shake the slab until it is reduced
3539 * by the same nr_pages that we used for reclaiming unmapped
3542 * Note that shrink_slab will free memory on all zones and may
3546 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3548 /* No reclaimable slab or very low memory pressure */
3549 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3552 /* Freed enough memory */
3553 nr_slab_pages1
= zone_page_state(zone
,
3554 NR_SLAB_RECLAIMABLE
);
3555 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3560 * Update nr_reclaimed by the number of slab pages we
3561 * reclaimed from this zone.
3563 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3564 if (nr_slab_pages1
< nr_slab_pages0
)
3565 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3568 p
->reclaim_state
= NULL
;
3569 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3570 lockdep_clear_current_reclaim_state();
3571 return sc
.nr_reclaimed
>= nr_pages
;
3574 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3580 * Zone reclaim reclaims unmapped file backed pages and
3581 * slab pages if we are over the defined limits.
3583 * A small portion of unmapped file backed pages is needed for
3584 * file I/O otherwise pages read by file I/O will be immediately
3585 * thrown out if the zone is overallocated. So we do not reclaim
3586 * if less than a specified percentage of the zone is used by
3587 * unmapped file backed pages.
3589 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3590 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3591 return ZONE_RECLAIM_FULL
;
3593 if (zone
->all_unreclaimable
)
3594 return ZONE_RECLAIM_FULL
;
3597 * Do not scan if the allocation should not be delayed.
3599 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3600 return ZONE_RECLAIM_NOSCAN
;
3603 * Only run zone reclaim on the local zone or on zones that do not
3604 * have associated processors. This will favor the local processor
3605 * over remote processors and spread off node memory allocations
3606 * as wide as possible.
3608 node_id
= zone_to_nid(zone
);
3609 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3610 return ZONE_RECLAIM_NOSCAN
;
3612 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3613 return ZONE_RECLAIM_NOSCAN
;
3615 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3616 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3619 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3626 * page_evictable - test whether a page is evictable
3627 * @page: the page to test
3629 * Test whether page is evictable--i.e., should be placed on active/inactive
3630 * lists vs unevictable list.
3632 * Reasons page might not be evictable:
3633 * (1) page's mapping marked unevictable
3634 * (2) page is part of an mlocked VMA
3637 int page_evictable(struct page
*page
)
3639 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3644 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3645 * @pages: array of pages to check
3646 * @nr_pages: number of pages to check
3648 * Checks pages for evictability and moves them to the appropriate lru list.
3650 * This function is only used for SysV IPC SHM_UNLOCK.
3652 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3654 struct lruvec
*lruvec
;
3655 struct zone
*zone
= NULL
;
3660 for (i
= 0; i
< nr_pages
; i
++) {
3661 struct page
*page
= pages
[i
];
3662 struct zone
*pagezone
;
3665 pagezone
= page_zone(page
);
3666 if (pagezone
!= zone
) {
3668 spin_unlock_irq(&zone
->lru_lock
);
3670 spin_lock_irq(&zone
->lru_lock
);
3672 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3674 if (!PageLRU(page
) || !PageUnevictable(page
))
3677 if (page_evictable(page
)) {
3678 enum lru_list lru
= page_lru_base_type(page
);
3680 VM_BUG_ON(PageActive(page
));
3681 ClearPageUnevictable(page
);
3682 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3683 add_page_to_lru_list(page
, lruvec
, lru
);
3689 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3690 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3691 spin_unlock_irq(&zone
->lru_lock
);
3694 #endif /* CONFIG_SHMEM */
3696 static void warn_scan_unevictable_pages(void)
3698 printk_once(KERN_WARNING
3699 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3700 "disabled for lack of a legitimate use case. If you have "
3701 "one, please send an email to linux-mm@kvack.org.\n",
3706 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3707 * all nodes' unevictable lists for evictable pages
3709 unsigned long scan_unevictable_pages
;
3711 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3712 void __user
*buffer
,
3713 size_t *length
, loff_t
*ppos
)
3715 warn_scan_unevictable_pages();
3716 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3717 scan_unevictable_pages
= 0;
3723 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3724 * a specified node's per zone unevictable lists for evictable pages.
3727 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3728 struct device_attribute
*attr
,
3731 warn_scan_unevictable_pages();
3732 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3735 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3736 struct device_attribute
*attr
,
3737 const char *buf
, size_t count
)
3739 warn_scan_unevictable_pages();
3744 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3745 read_scan_unevictable_node
,
3746 write_scan_unevictable_node
);
3748 int scan_unevictable_register_node(struct node
*node
)
3750 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3753 void scan_unevictable_unregister_node(struct node
*node
)
3755 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);