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.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim
;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup
*target_mem_cgroup
;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage
:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap
:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap
:1;
94 unsigned int hibernation_mode
:1;
96 /* One of the zones is ready for compaction */
97 unsigned int compaction_ready
:1;
99 /* Incremented by the number of inactive pages that were scanned */
100 unsigned long nr_scanned
;
102 /* Number of pages freed so far during a call to shrink_zones() */
103 unsigned long nr_reclaimed
;
106 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
108 #ifdef ARCH_HAS_PREFETCH
109 #define prefetch_prev_lru_page(_page, _base, _field) \
111 if ((_page)->lru.prev != _base) { \
114 prev = lru_to_page(&(_page->lru)); \
115 prefetch(&prev->_field); \
119 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122 #ifdef ARCH_HAS_PREFETCHW
123 #define prefetchw_prev_lru_page(_page, _base, _field) \
125 if ((_page)->lru.prev != _base) { \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetchw(&prev->_field); \
133 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137 * From 0 .. 100. Higher means more swappy.
139 int vm_swappiness
= 60;
141 * The total number of pages which are beyond the high watermark within all
144 unsigned long vm_total_pages
;
146 static LIST_HEAD(shrinker_list
);
147 static DECLARE_RWSEM(shrinker_rwsem
);
150 static bool global_reclaim(struct scan_control
*sc
)
152 return !sc
->target_mem_cgroup
;
155 static bool global_reclaim(struct scan_control
*sc
)
161 static unsigned long zone_reclaimable_pages(struct zone
*zone
)
165 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
166 zone_page_state(zone
, NR_INACTIVE_FILE
);
168 if (get_nr_swap_pages() > 0)
169 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
170 zone_page_state(zone
, NR_INACTIVE_ANON
);
175 bool zone_reclaimable(struct zone
*zone
)
177 return zone_page_state(zone
, NR_PAGES_SCANNED
) <
178 zone_reclaimable_pages(zone
) * 6;
181 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
183 if (!mem_cgroup_disabled())
184 return mem_cgroup_get_lru_size(lruvec
, lru
);
186 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
190 * Add a shrinker callback to be called from the vm.
192 int register_shrinker(struct shrinker
*shrinker
)
194 size_t size
= sizeof(*shrinker
->nr_deferred
);
197 * If we only have one possible node in the system anyway, save
198 * ourselves the trouble and disable NUMA aware behavior. This way we
199 * will save memory and some small loop time later.
201 if (nr_node_ids
== 1)
202 shrinker
->flags
&= ~SHRINKER_NUMA_AWARE
;
204 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
207 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
208 if (!shrinker
->nr_deferred
)
211 down_write(&shrinker_rwsem
);
212 list_add_tail(&shrinker
->list
, &shrinker_list
);
213 up_write(&shrinker_rwsem
);
216 EXPORT_SYMBOL(register_shrinker
);
221 void unregister_shrinker(struct shrinker
*shrinker
)
223 down_write(&shrinker_rwsem
);
224 list_del(&shrinker
->list
);
225 up_write(&shrinker_rwsem
);
226 kfree(shrinker
->nr_deferred
);
228 EXPORT_SYMBOL(unregister_shrinker
);
230 #define SHRINK_BATCH 128
233 shrink_slab_node(struct shrink_control
*shrinkctl
, struct shrinker
*shrinker
,
234 unsigned long nr_pages_scanned
, unsigned long lru_pages
)
236 unsigned long freed
= 0;
237 unsigned long long delta
;
242 int nid
= shrinkctl
->nid
;
243 long batch_size
= shrinker
->batch
? shrinker
->batch
246 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
251 * copy the current shrinker scan count into a local variable
252 * and zero it so that other concurrent shrinker invocations
253 * don't also do this scanning work.
255 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
258 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
260 do_div(delta
, lru_pages
+ 1);
262 if (total_scan
< 0) {
264 "shrink_slab: %pF negative objects to delete nr=%ld\n",
265 shrinker
->scan_objects
, total_scan
);
266 total_scan
= freeable
;
270 * We need to avoid excessive windup on filesystem shrinkers
271 * due to large numbers of GFP_NOFS allocations causing the
272 * shrinkers to return -1 all the time. This results in a large
273 * nr being built up so when a shrink that can do some work
274 * comes along it empties the entire cache due to nr >>>
275 * freeable. This is bad for sustaining a working set in
278 * Hence only allow the shrinker to scan the entire cache when
279 * a large delta change is calculated directly.
281 if (delta
< freeable
/ 4)
282 total_scan
= min(total_scan
, freeable
/ 2);
285 * Avoid risking looping forever due to too large nr value:
286 * never try to free more than twice the estimate number of
289 if (total_scan
> freeable
* 2)
290 total_scan
= freeable
* 2;
292 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
293 nr_pages_scanned
, lru_pages
,
294 freeable
, delta
, total_scan
);
297 * Normally, we should not scan less than batch_size objects in one
298 * pass to avoid too frequent shrinker calls, but if the slab has less
299 * than batch_size objects in total and we are really tight on memory,
300 * we will try to reclaim all available objects, otherwise we can end
301 * up failing allocations although there are plenty of reclaimable
302 * objects spread over several slabs with usage less than the
305 * We detect the "tight on memory" situations by looking at the total
306 * number of objects we want to scan (total_scan). If it is greater
307 * than the total number of objects on slab (freeable), we must be
308 * scanning at high prio and therefore should try to reclaim as much as
311 while (total_scan
>= batch_size
||
312 total_scan
>= freeable
) {
314 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
316 shrinkctl
->nr_to_scan
= nr_to_scan
;
317 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
318 if (ret
== SHRINK_STOP
)
322 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
323 total_scan
-= nr_to_scan
;
329 * move the unused scan count back into the shrinker in a
330 * manner that handles concurrent updates. If we exhausted the
331 * scan, there is no need to do an update.
334 new_nr
= atomic_long_add_return(total_scan
,
335 &shrinker
->nr_deferred
[nid
]);
337 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
339 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
344 * Call the shrink functions to age shrinkable caches
346 * Here we assume it costs one seek to replace a lru page and that it also
347 * takes a seek to recreate a cache object. With this in mind we age equal
348 * percentages of the lru and ageable caches. This should balance the seeks
349 * generated by these structures.
351 * If the vm encountered mapped pages on the LRU it increase the pressure on
352 * slab to avoid swapping.
354 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
356 * `lru_pages' represents the number of on-LRU pages in all the zones which
357 * are eligible for the caller's allocation attempt. It is used for balancing
358 * slab reclaim versus page reclaim.
360 * Returns the number of slab objects which we shrunk.
362 unsigned long shrink_slab(struct shrink_control
*shrinkctl
,
363 unsigned long nr_pages_scanned
,
364 unsigned long lru_pages
)
366 struct shrinker
*shrinker
;
367 unsigned long freed
= 0;
369 if (nr_pages_scanned
== 0)
370 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
372 if (!down_read_trylock(&shrinker_rwsem
)) {
374 * If we would return 0, our callers would understand that we
375 * have nothing else to shrink and give up trying. By returning
376 * 1 we keep it going and assume we'll be able to shrink next
383 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
384 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
)) {
386 freed
+= shrink_slab_node(shrinkctl
, shrinker
,
387 nr_pages_scanned
, lru_pages
);
391 for_each_node_mask(shrinkctl
->nid
, shrinkctl
->nodes_to_scan
) {
392 if (node_online(shrinkctl
->nid
))
393 freed
+= shrink_slab_node(shrinkctl
, shrinker
,
394 nr_pages_scanned
, lru_pages
);
398 up_read(&shrinker_rwsem
);
404 static inline int is_page_cache_freeable(struct page
*page
)
407 * A freeable page cache page is referenced only by the caller
408 * that isolated the page, the page cache radix tree and
409 * optional buffer heads at page->private.
411 return page_count(page
) - page_has_private(page
) == 2;
414 static int may_write_to_queue(struct backing_dev_info
*bdi
,
415 struct scan_control
*sc
)
417 if (current
->flags
& PF_SWAPWRITE
)
419 if (!bdi_write_congested(bdi
))
421 if (bdi
== current
->backing_dev_info
)
427 * We detected a synchronous write error writing a page out. Probably
428 * -ENOSPC. We need to propagate that into the address_space for a subsequent
429 * fsync(), msync() or close().
431 * The tricky part is that after writepage we cannot touch the mapping: nothing
432 * prevents it from being freed up. But we have a ref on the page and once
433 * that page is locked, the mapping is pinned.
435 * We're allowed to run sleeping lock_page() here because we know the caller has
438 static void handle_write_error(struct address_space
*mapping
,
439 struct page
*page
, int error
)
442 if (page_mapping(page
) == mapping
)
443 mapping_set_error(mapping
, error
);
447 /* possible outcome of pageout() */
449 /* failed to write page out, page is locked */
451 /* move page to the active list, page is locked */
453 /* page has been sent to the disk successfully, page is unlocked */
455 /* page is clean and locked */
460 * pageout is called by shrink_page_list() for each dirty page.
461 * Calls ->writepage().
463 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
464 struct scan_control
*sc
)
467 * If the page is dirty, only perform writeback if that write
468 * will be non-blocking. To prevent this allocation from being
469 * stalled by pagecache activity. But note that there may be
470 * stalls if we need to run get_block(). We could test
471 * PagePrivate for that.
473 * If this process is currently in __generic_file_write_iter() against
474 * this page's queue, we can perform writeback even if that
477 * If the page is swapcache, write it back even if that would
478 * block, for some throttling. This happens by accident, because
479 * swap_backing_dev_info is bust: it doesn't reflect the
480 * congestion state of the swapdevs. Easy to fix, if needed.
482 if (!is_page_cache_freeable(page
))
486 * Some data journaling orphaned pages can have
487 * page->mapping == NULL while being dirty with clean buffers.
489 if (page_has_private(page
)) {
490 if (try_to_free_buffers(page
)) {
491 ClearPageDirty(page
);
492 pr_info("%s: orphaned page\n", __func__
);
498 if (mapping
->a_ops
->writepage
== NULL
)
499 return PAGE_ACTIVATE
;
500 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
503 if (clear_page_dirty_for_io(page
)) {
505 struct writeback_control wbc
= {
506 .sync_mode
= WB_SYNC_NONE
,
507 .nr_to_write
= SWAP_CLUSTER_MAX
,
509 .range_end
= LLONG_MAX
,
513 SetPageReclaim(page
);
514 res
= mapping
->a_ops
->writepage(page
, &wbc
);
516 handle_write_error(mapping
, page
, res
);
517 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
518 ClearPageReclaim(page
);
519 return PAGE_ACTIVATE
;
522 if (!PageWriteback(page
)) {
523 /* synchronous write or broken a_ops? */
524 ClearPageReclaim(page
);
526 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
527 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
535 * Same as remove_mapping, but if the page is removed from the mapping, it
536 * gets returned with a refcount of 0.
538 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
541 BUG_ON(!PageLocked(page
));
542 BUG_ON(mapping
!= page_mapping(page
));
544 spin_lock_irq(&mapping
->tree_lock
);
546 * The non racy check for a busy page.
548 * Must be careful with the order of the tests. When someone has
549 * a ref to the page, it may be possible that they dirty it then
550 * drop the reference. So if PageDirty is tested before page_count
551 * here, then the following race may occur:
553 * get_user_pages(&page);
554 * [user mapping goes away]
556 * !PageDirty(page) [good]
557 * SetPageDirty(page);
559 * !page_count(page) [good, discard it]
561 * [oops, our write_to data is lost]
563 * Reversing the order of the tests ensures such a situation cannot
564 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
565 * load is not satisfied before that of page->_count.
567 * Note that if SetPageDirty is always performed via set_page_dirty,
568 * and thus under tree_lock, then this ordering is not required.
570 if (!page_freeze_refs(page
, 2))
572 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
573 if (unlikely(PageDirty(page
))) {
574 page_unfreeze_refs(page
, 2);
578 if (PageSwapCache(page
)) {
579 swp_entry_t swap
= { .val
= page_private(page
) };
580 __delete_from_swap_cache(page
);
581 spin_unlock_irq(&mapping
->tree_lock
);
582 swapcache_free(swap
, page
);
584 void (*freepage
)(struct page
*);
587 freepage
= mapping
->a_ops
->freepage
;
589 * Remember a shadow entry for reclaimed file cache in
590 * order to detect refaults, thus thrashing, later on.
592 * But don't store shadows in an address space that is
593 * already exiting. This is not just an optizimation,
594 * inode reclaim needs to empty out the radix tree or
595 * the nodes are lost. Don't plant shadows behind its
598 if (reclaimed
&& page_is_file_cache(page
) &&
599 !mapping_exiting(mapping
))
600 shadow
= workingset_eviction(mapping
, page
);
601 __delete_from_page_cache(page
, shadow
);
602 spin_unlock_irq(&mapping
->tree_lock
);
603 mem_cgroup_uncharge_cache_page(page
);
605 if (freepage
!= NULL
)
612 spin_unlock_irq(&mapping
->tree_lock
);
617 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
618 * someone else has a ref on the page, abort and return 0. If it was
619 * successfully detached, return 1. Assumes the caller has a single ref on
622 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
624 if (__remove_mapping(mapping
, page
, false)) {
626 * Unfreezing the refcount with 1 rather than 2 effectively
627 * drops the pagecache ref for us without requiring another
630 page_unfreeze_refs(page
, 1);
637 * putback_lru_page - put previously isolated page onto appropriate LRU list
638 * @page: page to be put back to appropriate lru list
640 * Add previously isolated @page to appropriate LRU list.
641 * Page may still be unevictable for other reasons.
643 * lru_lock must not be held, interrupts must be enabled.
645 void putback_lru_page(struct page
*page
)
648 int was_unevictable
= PageUnevictable(page
);
650 VM_BUG_ON_PAGE(PageLRU(page
), page
);
653 ClearPageUnevictable(page
);
655 if (page_evictable(page
)) {
657 * For evictable pages, we can use the cache.
658 * In event of a race, worst case is we end up with an
659 * unevictable page on [in]active list.
660 * We know how to handle that.
662 is_unevictable
= false;
666 * Put unevictable pages directly on zone's unevictable
669 is_unevictable
= true;
670 add_page_to_unevictable_list(page
);
672 * When racing with an mlock or AS_UNEVICTABLE clearing
673 * (page is unlocked) make sure that if the other thread
674 * does not observe our setting of PG_lru and fails
675 * isolation/check_move_unevictable_pages,
676 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
677 * the page back to the evictable list.
679 * The other side is TestClearPageMlocked() or shmem_lock().
685 * page's status can change while we move it among lru. If an evictable
686 * page is on unevictable list, it never be freed. To avoid that,
687 * check after we added it to the list, again.
689 if (is_unevictable
&& page_evictable(page
)) {
690 if (!isolate_lru_page(page
)) {
694 /* This means someone else dropped this page from LRU
695 * So, it will be freed or putback to LRU again. There is
696 * nothing to do here.
700 if (was_unevictable
&& !is_unevictable
)
701 count_vm_event(UNEVICTABLE_PGRESCUED
);
702 else if (!was_unevictable
&& is_unevictable
)
703 count_vm_event(UNEVICTABLE_PGCULLED
);
705 put_page(page
); /* drop ref from isolate */
708 enum page_references
{
710 PAGEREF_RECLAIM_CLEAN
,
715 static enum page_references
page_check_references(struct page
*page
,
716 struct scan_control
*sc
)
718 int referenced_ptes
, referenced_page
;
719 unsigned long vm_flags
;
721 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
723 referenced_page
= TestClearPageReferenced(page
);
726 * Mlock lost the isolation race with us. Let try_to_unmap()
727 * move the page to the unevictable list.
729 if (vm_flags
& VM_LOCKED
)
730 return PAGEREF_RECLAIM
;
732 if (referenced_ptes
) {
733 if (PageSwapBacked(page
))
734 return PAGEREF_ACTIVATE
;
736 * All mapped pages start out with page table
737 * references from the instantiating fault, so we need
738 * to look twice if a mapped file page is used more
741 * Mark it and spare it for another trip around the
742 * inactive list. Another page table reference will
743 * lead to its activation.
745 * Note: the mark is set for activated pages as well
746 * so that recently deactivated but used pages are
749 SetPageReferenced(page
);
751 if (referenced_page
|| referenced_ptes
> 1)
752 return PAGEREF_ACTIVATE
;
755 * Activate file-backed executable pages after first usage.
757 if (vm_flags
& VM_EXEC
)
758 return PAGEREF_ACTIVATE
;
763 /* Reclaim if clean, defer dirty pages to writeback */
764 if (referenced_page
&& !PageSwapBacked(page
))
765 return PAGEREF_RECLAIM_CLEAN
;
767 return PAGEREF_RECLAIM
;
770 /* Check if a page is dirty or under writeback */
771 static void page_check_dirty_writeback(struct page
*page
,
772 bool *dirty
, bool *writeback
)
774 struct address_space
*mapping
;
777 * Anonymous pages are not handled by flushers and must be written
778 * from reclaim context. Do not stall reclaim based on them
780 if (!page_is_file_cache(page
)) {
786 /* By default assume that the page flags are accurate */
787 *dirty
= PageDirty(page
);
788 *writeback
= PageWriteback(page
);
790 /* Verify dirty/writeback state if the filesystem supports it */
791 if (!page_has_private(page
))
794 mapping
= page_mapping(page
);
795 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
796 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
800 * shrink_page_list() returns the number of reclaimed pages
802 static unsigned long shrink_page_list(struct list_head
*page_list
,
804 struct scan_control
*sc
,
805 enum ttu_flags ttu_flags
,
806 unsigned long *ret_nr_dirty
,
807 unsigned long *ret_nr_unqueued_dirty
,
808 unsigned long *ret_nr_congested
,
809 unsigned long *ret_nr_writeback
,
810 unsigned long *ret_nr_immediate
,
813 LIST_HEAD(ret_pages
);
814 LIST_HEAD(free_pages
);
816 unsigned long nr_unqueued_dirty
= 0;
817 unsigned long nr_dirty
= 0;
818 unsigned long nr_congested
= 0;
819 unsigned long nr_reclaimed
= 0;
820 unsigned long nr_writeback
= 0;
821 unsigned long nr_immediate
= 0;
825 mem_cgroup_uncharge_start();
826 while (!list_empty(page_list
)) {
827 struct address_space
*mapping
;
830 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
831 bool dirty
, writeback
;
835 page
= lru_to_page(page_list
);
836 list_del(&page
->lru
);
838 if (!trylock_page(page
))
841 VM_BUG_ON_PAGE(PageActive(page
), page
);
842 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
846 if (unlikely(!page_evictable(page
)))
849 if (!sc
->may_unmap
&& page_mapped(page
))
852 /* Double the slab pressure for mapped and swapcache pages */
853 if (page_mapped(page
) || PageSwapCache(page
))
856 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
857 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
860 * The number of dirty pages determines if a zone is marked
861 * reclaim_congested which affects wait_iff_congested. kswapd
862 * will stall and start writing pages if the tail of the LRU
863 * is all dirty unqueued pages.
865 page_check_dirty_writeback(page
, &dirty
, &writeback
);
866 if (dirty
|| writeback
)
869 if (dirty
&& !writeback
)
873 * Treat this page as congested if the underlying BDI is or if
874 * pages are cycling through the LRU so quickly that the
875 * pages marked for immediate reclaim are making it to the
876 * end of the LRU a second time.
878 mapping
= page_mapping(page
);
879 if ((mapping
&& bdi_write_congested(mapping
->backing_dev_info
)) ||
880 (writeback
&& PageReclaim(page
)))
884 * If a page at the tail of the LRU is under writeback, there
885 * are three cases to consider.
887 * 1) If reclaim is encountering an excessive number of pages
888 * under writeback and this page is both under writeback and
889 * PageReclaim then it indicates that pages are being queued
890 * for IO but are being recycled through the LRU before the
891 * IO can complete. Waiting on the page itself risks an
892 * indefinite stall if it is impossible to writeback the
893 * page due to IO error or disconnected storage so instead
894 * note that the LRU is being scanned too quickly and the
895 * caller can stall after page list has been processed.
897 * 2) Global reclaim encounters a page, memcg encounters a
898 * page that is not marked for immediate reclaim or
899 * the caller does not have __GFP_IO. In this case mark
900 * the page for immediate reclaim and continue scanning.
902 * __GFP_IO is checked because a loop driver thread might
903 * enter reclaim, and deadlock if it waits on a page for
904 * which it is needed to do the write (loop masks off
905 * __GFP_IO|__GFP_FS for this reason); but more thought
906 * would probably show more reasons.
908 * Don't require __GFP_FS, since we're not going into the
909 * FS, just waiting on its writeback completion. Worryingly,
910 * ext4 gfs2 and xfs allocate pages with
911 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
912 * may_enter_fs here is liable to OOM on them.
914 * 3) memcg encounters a page that is not already marked
915 * PageReclaim. memcg does not have any dirty pages
916 * throttling so we could easily OOM just because too many
917 * pages are in writeback and there is nothing else to
918 * reclaim. Wait for the writeback to complete.
920 if (PageWriteback(page
)) {
922 if (current_is_kswapd() &&
924 zone_is_reclaim_writeback(zone
)) {
929 } else if (global_reclaim(sc
) ||
930 !PageReclaim(page
) || !(sc
->gfp_mask
& __GFP_IO
)) {
932 * This is slightly racy - end_page_writeback()
933 * might have just cleared PageReclaim, then
934 * setting PageReclaim here end up interpreted
935 * as PageReadahead - but that does not matter
936 * enough to care. What we do want is for this
937 * page to have PageReclaim set next time memcg
938 * reclaim reaches the tests above, so it will
939 * then wait_on_page_writeback() to avoid OOM;
940 * and it's also appropriate in global reclaim.
942 SetPageReclaim(page
);
949 wait_on_page_writeback(page
);
954 references
= page_check_references(page
, sc
);
956 switch (references
) {
957 case PAGEREF_ACTIVATE
:
958 goto activate_locked
;
961 case PAGEREF_RECLAIM
:
962 case PAGEREF_RECLAIM_CLEAN
:
963 ; /* try to reclaim the page below */
967 * Anonymous process memory has backing store?
968 * Try to allocate it some swap space here.
970 if (PageAnon(page
) && !PageSwapCache(page
)) {
971 if (!(sc
->gfp_mask
& __GFP_IO
))
973 if (!add_to_swap(page
, page_list
))
974 goto activate_locked
;
977 /* Adding to swap updated mapping */
978 mapping
= page_mapping(page
);
982 * The page is mapped into the page tables of one or more
983 * processes. Try to unmap it here.
985 if (page_mapped(page
) && mapping
) {
986 switch (try_to_unmap(page
, ttu_flags
)) {
988 goto activate_locked
;
994 ; /* try to free the page below */
998 if (PageDirty(page
)) {
1000 * Only kswapd can writeback filesystem pages to
1001 * avoid risk of stack overflow but only writeback
1002 * if many dirty pages have been encountered.
1004 if (page_is_file_cache(page
) &&
1005 (!current_is_kswapd() ||
1006 !zone_is_reclaim_dirty(zone
))) {
1008 * Immediately reclaim when written back.
1009 * Similar in principal to deactivate_page()
1010 * except we already have the page isolated
1011 * and know it's dirty
1013 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1014 SetPageReclaim(page
);
1019 if (references
== PAGEREF_RECLAIM_CLEAN
)
1023 if (!sc
->may_writepage
)
1026 /* Page is dirty, try to write it out here */
1027 switch (pageout(page
, mapping
, sc
)) {
1031 goto activate_locked
;
1033 if (PageWriteback(page
))
1035 if (PageDirty(page
))
1039 * A synchronous write - probably a ramdisk. Go
1040 * ahead and try to reclaim the page.
1042 if (!trylock_page(page
))
1044 if (PageDirty(page
) || PageWriteback(page
))
1046 mapping
= page_mapping(page
);
1048 ; /* try to free the page below */
1053 * If the page has buffers, try to free the buffer mappings
1054 * associated with this page. If we succeed we try to free
1057 * We do this even if the page is PageDirty().
1058 * try_to_release_page() does not perform I/O, but it is
1059 * possible for a page to have PageDirty set, but it is actually
1060 * clean (all its buffers are clean). This happens if the
1061 * buffers were written out directly, with submit_bh(). ext3
1062 * will do this, as well as the blockdev mapping.
1063 * try_to_release_page() will discover that cleanness and will
1064 * drop the buffers and mark the page clean - it can be freed.
1066 * Rarely, pages can have buffers and no ->mapping. These are
1067 * the pages which were not successfully invalidated in
1068 * truncate_complete_page(). We try to drop those buffers here
1069 * and if that worked, and the page is no longer mapped into
1070 * process address space (page_count == 1) it can be freed.
1071 * Otherwise, leave the page on the LRU so it is swappable.
1073 if (page_has_private(page
)) {
1074 if (!try_to_release_page(page
, sc
->gfp_mask
))
1075 goto activate_locked
;
1076 if (!mapping
&& page_count(page
) == 1) {
1078 if (put_page_testzero(page
))
1082 * rare race with speculative reference.
1083 * the speculative reference will free
1084 * this page shortly, so we may
1085 * increment nr_reclaimed here (and
1086 * leave it off the LRU).
1094 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1098 * At this point, we have no other references and there is
1099 * no way to pick any more up (removed from LRU, removed
1100 * from pagecache). Can use non-atomic bitops now (and
1101 * we obviously don't have to worry about waking up a process
1102 * waiting on the page lock, because there are no references.
1104 __clear_page_locked(page
);
1109 * Is there need to periodically free_page_list? It would
1110 * appear not as the counts should be low
1112 list_add(&page
->lru
, &free_pages
);
1116 if (PageSwapCache(page
))
1117 try_to_free_swap(page
);
1119 putback_lru_page(page
);
1123 /* Not a candidate for swapping, so reclaim swap space. */
1124 if (PageSwapCache(page
) && vm_swap_full())
1125 try_to_free_swap(page
);
1126 VM_BUG_ON_PAGE(PageActive(page
), page
);
1127 SetPageActive(page
);
1132 list_add(&page
->lru
, &ret_pages
);
1133 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1136 free_hot_cold_page_list(&free_pages
, true);
1138 list_splice(&ret_pages
, page_list
);
1139 count_vm_events(PGACTIVATE
, pgactivate
);
1140 mem_cgroup_uncharge_end();
1141 *ret_nr_dirty
+= nr_dirty
;
1142 *ret_nr_congested
+= nr_congested
;
1143 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1144 *ret_nr_writeback
+= nr_writeback
;
1145 *ret_nr_immediate
+= nr_immediate
;
1146 return nr_reclaimed
;
1149 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1150 struct list_head
*page_list
)
1152 struct scan_control sc
= {
1153 .gfp_mask
= GFP_KERNEL
,
1154 .priority
= DEF_PRIORITY
,
1157 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1158 struct page
*page
, *next
;
1159 LIST_HEAD(clean_pages
);
1161 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1162 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1163 !isolated_balloon_page(page
)) {
1164 ClearPageActive(page
);
1165 list_move(&page
->lru
, &clean_pages
);
1169 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1170 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1171 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1172 list_splice(&clean_pages
, page_list
);
1173 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1178 * Attempt to remove the specified page from its LRU. Only take this page
1179 * if it is of the appropriate PageActive status. Pages which are being
1180 * freed elsewhere are also ignored.
1182 * page: page to consider
1183 * mode: one of the LRU isolation modes defined above
1185 * returns 0 on success, -ve errno on failure.
1187 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1191 /* Only take pages on the LRU. */
1195 /* Compaction should not handle unevictable pages but CMA can do so */
1196 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1202 * To minimise LRU disruption, the caller can indicate that it only
1203 * wants to isolate pages it will be able to operate on without
1204 * blocking - clean pages for the most part.
1206 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1207 * is used by reclaim when it is cannot write to backing storage
1209 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1210 * that it is possible to migrate without blocking
1212 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1213 /* All the caller can do on PageWriteback is block */
1214 if (PageWriteback(page
))
1217 if (PageDirty(page
)) {
1218 struct address_space
*mapping
;
1220 /* ISOLATE_CLEAN means only clean pages */
1221 if (mode
& ISOLATE_CLEAN
)
1225 * Only pages without mappings or that have a
1226 * ->migratepage callback are possible to migrate
1229 mapping
= page_mapping(page
);
1230 if (mapping
&& !mapping
->a_ops
->migratepage
)
1235 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1238 if (likely(get_page_unless_zero(page
))) {
1240 * Be careful not to clear PageLRU until after we're
1241 * sure the page is not being freed elsewhere -- the
1242 * page release code relies on it.
1252 * zone->lru_lock is heavily contended. Some of the functions that
1253 * shrink the lists perform better by taking out a batch of pages
1254 * and working on them outside the LRU lock.
1256 * For pagecache intensive workloads, this function is the hottest
1257 * spot in the kernel (apart from copy_*_user functions).
1259 * Appropriate locks must be held before calling this function.
1261 * @nr_to_scan: The number of pages to look through on the list.
1262 * @lruvec: The LRU vector to pull pages from.
1263 * @dst: The temp list to put pages on to.
1264 * @nr_scanned: The number of pages that were scanned.
1265 * @sc: The scan_control struct for this reclaim session
1266 * @mode: One of the LRU isolation modes
1267 * @lru: LRU list id for isolating
1269 * returns how many pages were moved onto *@dst.
1271 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1272 struct lruvec
*lruvec
, struct list_head
*dst
,
1273 unsigned long *nr_scanned
, struct scan_control
*sc
,
1274 isolate_mode_t mode
, enum lru_list lru
)
1276 struct list_head
*src
= &lruvec
->lists
[lru
];
1277 unsigned long nr_taken
= 0;
1280 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1284 page
= lru_to_page(src
);
1285 prefetchw_prev_lru_page(page
, src
, flags
);
1287 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1289 switch (__isolate_lru_page(page
, mode
)) {
1291 nr_pages
= hpage_nr_pages(page
);
1292 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1293 list_move(&page
->lru
, dst
);
1294 nr_taken
+= nr_pages
;
1298 /* else it is being freed elsewhere */
1299 list_move(&page
->lru
, src
);
1308 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1309 nr_taken
, mode
, is_file_lru(lru
));
1314 * isolate_lru_page - tries to isolate a page from its LRU list
1315 * @page: page to isolate from its LRU list
1317 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1318 * vmstat statistic corresponding to whatever LRU list the page was on.
1320 * Returns 0 if the page was removed from an LRU list.
1321 * Returns -EBUSY if the page was not on an LRU list.
1323 * The returned page will have PageLRU() cleared. If it was found on
1324 * the active list, it will have PageActive set. If it was found on
1325 * the unevictable list, it will have the PageUnevictable bit set. That flag
1326 * may need to be cleared by the caller before letting the page go.
1328 * The vmstat statistic corresponding to the list on which the page was
1329 * found will be decremented.
1332 * (1) Must be called with an elevated refcount on the page. This is a
1333 * fundamentnal difference from isolate_lru_pages (which is called
1334 * without a stable reference).
1335 * (2) the lru_lock must not be held.
1336 * (3) interrupts must be enabled.
1338 int isolate_lru_page(struct page
*page
)
1342 VM_BUG_ON_PAGE(!page_count(page
), page
);
1344 if (PageLRU(page
)) {
1345 struct zone
*zone
= page_zone(page
);
1346 struct lruvec
*lruvec
;
1348 spin_lock_irq(&zone
->lru_lock
);
1349 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1350 if (PageLRU(page
)) {
1351 int lru
= page_lru(page
);
1354 del_page_from_lru_list(page
, lruvec
, lru
);
1357 spin_unlock_irq(&zone
->lru_lock
);
1363 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1364 * then get resheduled. When there are massive number of tasks doing page
1365 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1366 * the LRU list will go small and be scanned faster than necessary, leading to
1367 * unnecessary swapping, thrashing and OOM.
1369 static int too_many_isolated(struct zone
*zone
, int file
,
1370 struct scan_control
*sc
)
1372 unsigned long inactive
, isolated
;
1374 if (current_is_kswapd())
1377 if (!global_reclaim(sc
))
1381 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1382 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1384 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1385 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1389 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1390 * won't get blocked by normal direct-reclaimers, forming a circular
1393 if ((sc
->gfp_mask
& GFP_IOFS
) == GFP_IOFS
)
1396 return isolated
> inactive
;
1399 static noinline_for_stack
void
1400 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1402 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1403 struct zone
*zone
= lruvec_zone(lruvec
);
1404 LIST_HEAD(pages_to_free
);
1407 * Put back any unfreeable pages.
1409 while (!list_empty(page_list
)) {
1410 struct page
*page
= lru_to_page(page_list
);
1413 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1414 list_del(&page
->lru
);
1415 if (unlikely(!page_evictable(page
))) {
1416 spin_unlock_irq(&zone
->lru_lock
);
1417 putback_lru_page(page
);
1418 spin_lock_irq(&zone
->lru_lock
);
1422 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1425 lru
= page_lru(page
);
1426 add_page_to_lru_list(page
, lruvec
, lru
);
1428 if (is_active_lru(lru
)) {
1429 int file
= is_file_lru(lru
);
1430 int numpages
= hpage_nr_pages(page
);
1431 reclaim_stat
->recent_rotated
[file
] += numpages
;
1433 if (put_page_testzero(page
)) {
1434 __ClearPageLRU(page
);
1435 __ClearPageActive(page
);
1436 del_page_from_lru_list(page
, lruvec
, lru
);
1438 if (unlikely(PageCompound(page
))) {
1439 spin_unlock_irq(&zone
->lru_lock
);
1440 (*get_compound_page_dtor(page
))(page
);
1441 spin_lock_irq(&zone
->lru_lock
);
1443 list_add(&page
->lru
, &pages_to_free
);
1448 * To save our caller's stack, now use input list for pages to free.
1450 list_splice(&pages_to_free
, page_list
);
1454 * If a kernel thread (such as nfsd for loop-back mounts) services
1455 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1456 * In that case we should only throttle if the backing device it is
1457 * writing to is congested. In other cases it is safe to throttle.
1459 static int current_may_throttle(void)
1461 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1462 current
->backing_dev_info
== NULL
||
1463 bdi_write_congested(current
->backing_dev_info
);
1467 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1468 * of reclaimed pages
1470 static noinline_for_stack
unsigned long
1471 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1472 struct scan_control
*sc
, enum lru_list lru
)
1474 LIST_HEAD(page_list
);
1475 unsigned long nr_scanned
;
1476 unsigned long nr_reclaimed
= 0;
1477 unsigned long nr_taken
;
1478 unsigned long nr_dirty
= 0;
1479 unsigned long nr_congested
= 0;
1480 unsigned long nr_unqueued_dirty
= 0;
1481 unsigned long nr_writeback
= 0;
1482 unsigned long nr_immediate
= 0;
1483 isolate_mode_t isolate_mode
= 0;
1484 int file
= is_file_lru(lru
);
1485 struct zone
*zone
= lruvec_zone(lruvec
);
1486 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1488 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1489 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1491 /* We are about to die and free our memory. Return now. */
1492 if (fatal_signal_pending(current
))
1493 return SWAP_CLUSTER_MAX
;
1499 isolate_mode
|= ISOLATE_UNMAPPED
;
1500 if (!sc
->may_writepage
)
1501 isolate_mode
|= ISOLATE_CLEAN
;
1503 spin_lock_irq(&zone
->lru_lock
);
1505 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1506 &nr_scanned
, sc
, isolate_mode
, lru
);
1508 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1509 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1511 if (global_reclaim(sc
)) {
1512 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1513 if (current_is_kswapd())
1514 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1516 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1518 spin_unlock_irq(&zone
->lru_lock
);
1523 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1524 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1525 &nr_writeback
, &nr_immediate
,
1528 spin_lock_irq(&zone
->lru_lock
);
1530 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1532 if (global_reclaim(sc
)) {
1533 if (current_is_kswapd())
1534 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1537 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1541 putback_inactive_pages(lruvec
, &page_list
);
1543 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1545 spin_unlock_irq(&zone
->lru_lock
);
1547 free_hot_cold_page_list(&page_list
, true);
1550 * If reclaim is isolating dirty pages under writeback, it implies
1551 * that the long-lived page allocation rate is exceeding the page
1552 * laundering rate. Either the global limits are not being effective
1553 * at throttling processes due to the page distribution throughout
1554 * zones or there is heavy usage of a slow backing device. The
1555 * only option is to throttle from reclaim context which is not ideal
1556 * as there is no guarantee the dirtying process is throttled in the
1557 * same way balance_dirty_pages() manages.
1559 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1560 * of pages under pages flagged for immediate reclaim and stall if any
1561 * are encountered in the nr_immediate check below.
1563 if (nr_writeback
&& nr_writeback
== nr_taken
)
1564 zone_set_flag(zone
, ZONE_WRITEBACK
);
1567 * memcg will stall in page writeback so only consider forcibly
1568 * stalling for global reclaim
1570 if (global_reclaim(sc
)) {
1572 * Tag a zone as congested if all the dirty pages scanned were
1573 * backed by a congested BDI and wait_iff_congested will stall.
1575 if (nr_dirty
&& nr_dirty
== nr_congested
)
1576 zone_set_flag(zone
, ZONE_CONGESTED
);
1579 * If dirty pages are scanned that are not queued for IO, it
1580 * implies that flushers are not keeping up. In this case, flag
1581 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1582 * pages from reclaim context.
1584 if (nr_unqueued_dirty
== nr_taken
)
1585 zone_set_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
1588 * If kswapd scans pages marked marked for immediate
1589 * reclaim and under writeback (nr_immediate), it implies
1590 * that pages are cycling through the LRU faster than
1591 * they are written so also forcibly stall.
1593 if (nr_immediate
&& current_may_throttle())
1594 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1598 * Stall direct reclaim for IO completions if underlying BDIs or zone
1599 * is congested. Allow kswapd to continue until it starts encountering
1600 * unqueued dirty pages or cycling through the LRU too quickly.
1602 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1603 current_may_throttle())
1604 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1606 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1608 nr_scanned
, nr_reclaimed
,
1610 trace_shrink_flags(file
));
1611 return nr_reclaimed
;
1615 * This moves pages from the active list to the inactive list.
1617 * We move them the other way if the page is referenced by one or more
1618 * processes, from rmap.
1620 * If the pages are mostly unmapped, the processing is fast and it is
1621 * appropriate to hold zone->lru_lock across the whole operation. But if
1622 * the pages are mapped, the processing is slow (page_referenced()) so we
1623 * should drop zone->lru_lock around each page. It's impossible to balance
1624 * this, so instead we remove the pages from the LRU while processing them.
1625 * It is safe to rely on PG_active against the non-LRU pages in here because
1626 * nobody will play with that bit on a non-LRU page.
1628 * The downside is that we have to touch page->_count against each page.
1629 * But we had to alter page->flags anyway.
1632 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1633 struct list_head
*list
,
1634 struct list_head
*pages_to_free
,
1637 struct zone
*zone
= lruvec_zone(lruvec
);
1638 unsigned long pgmoved
= 0;
1642 while (!list_empty(list
)) {
1643 page
= lru_to_page(list
);
1644 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1646 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1649 nr_pages
= hpage_nr_pages(page
);
1650 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1651 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1652 pgmoved
+= nr_pages
;
1654 if (put_page_testzero(page
)) {
1655 __ClearPageLRU(page
);
1656 __ClearPageActive(page
);
1657 del_page_from_lru_list(page
, lruvec
, lru
);
1659 if (unlikely(PageCompound(page
))) {
1660 spin_unlock_irq(&zone
->lru_lock
);
1661 (*get_compound_page_dtor(page
))(page
);
1662 spin_lock_irq(&zone
->lru_lock
);
1664 list_add(&page
->lru
, pages_to_free
);
1667 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1668 if (!is_active_lru(lru
))
1669 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1672 static void shrink_active_list(unsigned long nr_to_scan
,
1673 struct lruvec
*lruvec
,
1674 struct scan_control
*sc
,
1677 unsigned long nr_taken
;
1678 unsigned long nr_scanned
;
1679 unsigned long vm_flags
;
1680 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1681 LIST_HEAD(l_active
);
1682 LIST_HEAD(l_inactive
);
1684 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1685 unsigned long nr_rotated
= 0;
1686 isolate_mode_t isolate_mode
= 0;
1687 int file
= is_file_lru(lru
);
1688 struct zone
*zone
= lruvec_zone(lruvec
);
1693 isolate_mode
|= ISOLATE_UNMAPPED
;
1694 if (!sc
->may_writepage
)
1695 isolate_mode
|= ISOLATE_CLEAN
;
1697 spin_lock_irq(&zone
->lru_lock
);
1699 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1700 &nr_scanned
, sc
, isolate_mode
, lru
);
1701 if (global_reclaim(sc
))
1702 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1704 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1706 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1707 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1708 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1709 spin_unlock_irq(&zone
->lru_lock
);
1711 while (!list_empty(&l_hold
)) {
1713 page
= lru_to_page(&l_hold
);
1714 list_del(&page
->lru
);
1716 if (unlikely(!page_evictable(page
))) {
1717 putback_lru_page(page
);
1721 if (unlikely(buffer_heads_over_limit
)) {
1722 if (page_has_private(page
) && trylock_page(page
)) {
1723 if (page_has_private(page
))
1724 try_to_release_page(page
, 0);
1729 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1731 nr_rotated
+= hpage_nr_pages(page
);
1733 * Identify referenced, file-backed active pages and
1734 * give them one more trip around the active list. So
1735 * that executable code get better chances to stay in
1736 * memory under moderate memory pressure. Anon pages
1737 * are not likely to be evicted by use-once streaming
1738 * IO, plus JVM can create lots of anon VM_EXEC pages,
1739 * so we ignore them here.
1741 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1742 list_add(&page
->lru
, &l_active
);
1747 ClearPageActive(page
); /* we are de-activating */
1748 list_add(&page
->lru
, &l_inactive
);
1752 * Move pages back to the lru list.
1754 spin_lock_irq(&zone
->lru_lock
);
1756 * Count referenced pages from currently used mappings as rotated,
1757 * even though only some of them are actually re-activated. This
1758 * helps balance scan pressure between file and anonymous pages in
1761 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1763 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1764 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1765 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1766 spin_unlock_irq(&zone
->lru_lock
);
1768 free_hot_cold_page_list(&l_hold
, true);
1772 static int inactive_anon_is_low_global(struct zone
*zone
)
1774 unsigned long active
, inactive
;
1776 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1777 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1779 if (inactive
* zone
->inactive_ratio
< active
)
1786 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1787 * @lruvec: LRU vector to check
1789 * Returns true if the zone does not have enough inactive anon pages,
1790 * meaning some active anon pages need to be deactivated.
1792 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1795 * If we don't have swap space, anonymous page deactivation
1798 if (!total_swap_pages
)
1801 if (!mem_cgroup_disabled())
1802 return mem_cgroup_inactive_anon_is_low(lruvec
);
1804 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1807 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1814 * inactive_file_is_low - check if file pages need to be deactivated
1815 * @lruvec: LRU vector to check
1817 * When the system is doing streaming IO, memory pressure here
1818 * ensures that active file pages get deactivated, until more
1819 * than half of the file pages are on the inactive list.
1821 * Once we get to that situation, protect the system's working
1822 * set from being evicted by disabling active file page aging.
1824 * This uses a different ratio than the anonymous pages, because
1825 * the page cache uses a use-once replacement algorithm.
1827 static int inactive_file_is_low(struct lruvec
*lruvec
)
1829 unsigned long inactive
;
1830 unsigned long active
;
1832 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1833 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1835 return active
> inactive
;
1838 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1840 if (is_file_lru(lru
))
1841 return inactive_file_is_low(lruvec
);
1843 return inactive_anon_is_low(lruvec
);
1846 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1847 struct lruvec
*lruvec
, struct scan_control
*sc
)
1849 if (is_active_lru(lru
)) {
1850 if (inactive_list_is_low(lruvec
, lru
))
1851 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1855 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1866 * Determine how aggressively the anon and file LRU lists should be
1867 * scanned. The relative value of each set of LRU lists is determined
1868 * by looking at the fraction of the pages scanned we did rotate back
1869 * onto the active list instead of evict.
1871 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1872 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1874 static void get_scan_count(struct lruvec
*lruvec
, int swappiness
,
1875 struct scan_control
*sc
, unsigned long *nr
)
1877 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1879 u64 denominator
= 0; /* gcc */
1880 struct zone
*zone
= lruvec_zone(lruvec
);
1881 unsigned long anon_prio
, file_prio
;
1882 enum scan_balance scan_balance
;
1883 unsigned long anon
, file
;
1884 bool force_scan
= false;
1885 unsigned long ap
, fp
;
1891 * If the zone or memcg is small, nr[l] can be 0. This
1892 * results in no scanning on this priority and a potential
1893 * priority drop. Global direct reclaim can go to the next
1894 * zone and tends to have no problems. Global kswapd is for
1895 * zone balancing and it needs to scan a minimum amount. When
1896 * reclaiming for a memcg, a priority drop can cause high
1897 * latencies, so it's better to scan a minimum amount there as
1900 if (current_is_kswapd() && !zone_reclaimable(zone
))
1902 if (!global_reclaim(sc
))
1905 /* If we have no swap space, do not bother scanning anon pages. */
1906 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
1907 scan_balance
= SCAN_FILE
;
1912 * Global reclaim will swap to prevent OOM even with no
1913 * swappiness, but memcg users want to use this knob to
1914 * disable swapping for individual groups completely when
1915 * using the memory controller's swap limit feature would be
1918 if (!global_reclaim(sc
) && !swappiness
) {
1919 scan_balance
= SCAN_FILE
;
1924 * Do not apply any pressure balancing cleverness when the
1925 * system is close to OOM, scan both anon and file equally
1926 * (unless the swappiness setting disagrees with swapping).
1928 if (!sc
->priority
&& swappiness
) {
1929 scan_balance
= SCAN_EQUAL
;
1934 * Prevent the reclaimer from falling into the cache trap: as
1935 * cache pages start out inactive, every cache fault will tip
1936 * the scan balance towards the file LRU. And as the file LRU
1937 * shrinks, so does the window for rotation from references.
1938 * This means we have a runaway feedback loop where a tiny
1939 * thrashing file LRU becomes infinitely more attractive than
1940 * anon pages. Try to detect this based on file LRU size.
1942 if (global_reclaim(sc
)) {
1943 unsigned long zonefile
;
1944 unsigned long zonefree
;
1946 zonefree
= zone_page_state(zone
, NR_FREE_PAGES
);
1947 zonefile
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
1948 zone_page_state(zone
, NR_INACTIVE_FILE
);
1950 if (unlikely(zonefile
+ zonefree
<= high_wmark_pages(zone
))) {
1951 scan_balance
= SCAN_ANON
;
1957 * There is enough inactive page cache, do not reclaim
1958 * anything from the anonymous working set right now.
1960 if (!inactive_file_is_low(lruvec
)) {
1961 scan_balance
= SCAN_FILE
;
1965 scan_balance
= SCAN_FRACT
;
1968 * With swappiness at 100, anonymous and file have the same priority.
1969 * This scanning priority is essentially the inverse of IO cost.
1971 anon_prio
= swappiness
;
1972 file_prio
= 200 - anon_prio
;
1975 * OK, so we have swap space and a fair amount of page cache
1976 * pages. We use the recently rotated / recently scanned
1977 * ratios to determine how valuable each cache is.
1979 * Because workloads change over time (and to avoid overflow)
1980 * we keep these statistics as a floating average, which ends
1981 * up weighing recent references more than old ones.
1983 * anon in [0], file in [1]
1986 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1987 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1988 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1989 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1991 spin_lock_irq(&zone
->lru_lock
);
1992 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1993 reclaim_stat
->recent_scanned
[0] /= 2;
1994 reclaim_stat
->recent_rotated
[0] /= 2;
1997 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1998 reclaim_stat
->recent_scanned
[1] /= 2;
1999 reclaim_stat
->recent_rotated
[1] /= 2;
2003 * The amount of pressure on anon vs file pages is inversely
2004 * proportional to the fraction of recently scanned pages on
2005 * each list that were recently referenced and in active use.
2007 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2008 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2010 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2011 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2012 spin_unlock_irq(&zone
->lru_lock
);
2016 denominator
= ap
+ fp
+ 1;
2018 some_scanned
= false;
2019 /* Only use force_scan on second pass. */
2020 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2021 for_each_evictable_lru(lru
) {
2022 int file
= is_file_lru(lru
);
2026 size
= get_lru_size(lruvec
, lru
);
2027 scan
= size
>> sc
->priority
;
2029 if (!scan
&& pass
&& force_scan
)
2030 scan
= min(size
, SWAP_CLUSTER_MAX
);
2032 switch (scan_balance
) {
2034 /* Scan lists relative to size */
2038 * Scan types proportional to swappiness and
2039 * their relative recent reclaim efficiency.
2041 scan
= div64_u64(scan
* fraction
[file
],
2046 /* Scan one type exclusively */
2047 if ((scan_balance
== SCAN_FILE
) != file
)
2051 /* Look ma, no brain */
2056 * Skip the second pass and don't force_scan,
2057 * if we found something to scan.
2059 some_scanned
|= !!scan
;
2065 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2067 static void shrink_lruvec(struct lruvec
*lruvec
, int swappiness
,
2068 struct scan_control
*sc
)
2070 unsigned long nr
[NR_LRU_LISTS
];
2071 unsigned long targets
[NR_LRU_LISTS
];
2072 unsigned long nr_to_scan
;
2074 unsigned long nr_reclaimed
= 0;
2075 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2076 struct blk_plug plug
;
2079 get_scan_count(lruvec
, swappiness
, sc
, nr
);
2081 /* Record the original scan target for proportional adjustments later */
2082 memcpy(targets
, nr
, sizeof(nr
));
2085 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2086 * event that can occur when there is little memory pressure e.g.
2087 * multiple streaming readers/writers. Hence, we do not abort scanning
2088 * when the requested number of pages are reclaimed when scanning at
2089 * DEF_PRIORITY on the assumption that the fact we are direct
2090 * reclaiming implies that kswapd is not keeping up and it is best to
2091 * do a batch of work at once. For memcg reclaim one check is made to
2092 * abort proportional reclaim if either the file or anon lru has already
2093 * dropped to zero at the first pass.
2095 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2096 sc
->priority
== DEF_PRIORITY
);
2098 blk_start_plug(&plug
);
2099 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2100 nr
[LRU_INACTIVE_FILE
]) {
2101 unsigned long nr_anon
, nr_file
, percentage
;
2102 unsigned long nr_scanned
;
2104 for_each_evictable_lru(lru
) {
2106 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2107 nr
[lru
] -= nr_to_scan
;
2109 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2114 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2118 * For kswapd and memcg, reclaim at least the number of pages
2119 * requested. Ensure that the anon and file LRUs are scanned
2120 * proportionally what was requested by get_scan_count(). We
2121 * stop reclaiming one LRU and reduce the amount scanning
2122 * proportional to the original scan target.
2124 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2125 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2128 * It's just vindictive to attack the larger once the smaller
2129 * has gone to zero. And given the way we stop scanning the
2130 * smaller below, this makes sure that we only make one nudge
2131 * towards proportionality once we've got nr_to_reclaim.
2133 if (!nr_file
|| !nr_anon
)
2136 if (nr_file
> nr_anon
) {
2137 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2138 targets
[LRU_ACTIVE_ANON
] + 1;
2140 percentage
= nr_anon
* 100 / scan_target
;
2142 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2143 targets
[LRU_ACTIVE_FILE
] + 1;
2145 percentage
= nr_file
* 100 / scan_target
;
2148 /* Stop scanning the smaller of the LRU */
2150 nr
[lru
+ LRU_ACTIVE
] = 0;
2153 * Recalculate the other LRU scan count based on its original
2154 * scan target and the percentage scanning already complete
2156 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2157 nr_scanned
= targets
[lru
] - nr
[lru
];
2158 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2159 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2162 nr_scanned
= targets
[lru
] - nr
[lru
];
2163 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2164 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2166 scan_adjusted
= true;
2168 blk_finish_plug(&plug
);
2169 sc
->nr_reclaimed
+= nr_reclaimed
;
2172 * Even if we did not try to evict anon pages at all, we want to
2173 * rebalance the anon lru active/inactive ratio.
2175 if (inactive_anon_is_low(lruvec
))
2176 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2177 sc
, LRU_ACTIVE_ANON
);
2179 throttle_vm_writeout(sc
->gfp_mask
);
2182 /* Use reclaim/compaction for costly allocs or under memory pressure */
2183 static bool in_reclaim_compaction(struct scan_control
*sc
)
2185 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2186 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2187 sc
->priority
< DEF_PRIORITY
- 2))
2194 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2195 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2196 * true if more pages should be reclaimed such that when the page allocator
2197 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2198 * It will give up earlier than that if there is difficulty reclaiming pages.
2200 static inline bool should_continue_reclaim(struct zone
*zone
,
2201 unsigned long nr_reclaimed
,
2202 unsigned long nr_scanned
,
2203 struct scan_control
*sc
)
2205 unsigned long pages_for_compaction
;
2206 unsigned long inactive_lru_pages
;
2208 /* If not in reclaim/compaction mode, stop */
2209 if (!in_reclaim_compaction(sc
))
2212 /* Consider stopping depending on scan and reclaim activity */
2213 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2215 * For __GFP_REPEAT allocations, stop reclaiming if the
2216 * full LRU list has been scanned and we are still failing
2217 * to reclaim pages. This full LRU scan is potentially
2218 * expensive but a __GFP_REPEAT caller really wants to succeed
2220 if (!nr_reclaimed
&& !nr_scanned
)
2224 * For non-__GFP_REPEAT allocations which can presumably
2225 * fail without consequence, stop if we failed to reclaim
2226 * any pages from the last SWAP_CLUSTER_MAX number of
2227 * pages that were scanned. This will return to the
2228 * caller faster at the risk reclaim/compaction and
2229 * the resulting allocation attempt fails
2236 * If we have not reclaimed enough pages for compaction and the
2237 * inactive lists are large enough, continue reclaiming
2239 pages_for_compaction
= (2UL << sc
->order
);
2240 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2241 if (get_nr_swap_pages() > 0)
2242 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2243 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2244 inactive_lru_pages
> pages_for_compaction
)
2247 /* If compaction would go ahead or the allocation would succeed, stop */
2248 switch (compaction_suitable(zone
, sc
->order
)) {
2249 case COMPACT_PARTIAL
:
2250 case COMPACT_CONTINUE
:
2257 static bool shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
2259 unsigned long nr_reclaimed
, nr_scanned
;
2260 bool reclaimable
= false;
2263 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2264 struct mem_cgroup_reclaim_cookie reclaim
= {
2266 .priority
= sc
->priority
,
2268 struct mem_cgroup
*memcg
;
2270 nr_reclaimed
= sc
->nr_reclaimed
;
2271 nr_scanned
= sc
->nr_scanned
;
2273 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2275 struct lruvec
*lruvec
;
2278 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2279 swappiness
= mem_cgroup_swappiness(memcg
);
2281 shrink_lruvec(lruvec
, swappiness
, sc
);
2284 * Direct reclaim and kswapd have to scan all memory
2285 * cgroups to fulfill the overall scan target for the
2288 * Limit reclaim, on the other hand, only cares about
2289 * nr_to_reclaim pages to be reclaimed and it will
2290 * retry with decreasing priority if one round over the
2291 * whole hierarchy is not sufficient.
2293 if (!global_reclaim(sc
) &&
2294 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2295 mem_cgroup_iter_break(root
, memcg
);
2298 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
2301 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2302 sc
->nr_scanned
- nr_scanned
,
2303 sc
->nr_reclaimed
- nr_reclaimed
);
2305 if (sc
->nr_reclaimed
- nr_reclaimed
)
2308 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2309 sc
->nr_scanned
- nr_scanned
, sc
));
2314 /* Returns true if compaction should go ahead for a high-order request */
2315 static inline bool compaction_ready(struct zone
*zone
, int order
)
2317 unsigned long balance_gap
, watermark
;
2321 * Compaction takes time to run and there are potentially other
2322 * callers using the pages just freed. Continue reclaiming until
2323 * there is a buffer of free pages available to give compaction
2324 * a reasonable chance of completing and allocating the page
2326 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2327 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2328 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << order
);
2329 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2332 * If compaction is deferred, reclaim up to a point where
2333 * compaction will have a chance of success when re-enabled
2335 if (compaction_deferred(zone
, order
))
2336 return watermark_ok
;
2338 /* If compaction is not ready to start, keep reclaiming */
2339 if (!compaction_suitable(zone
, order
))
2342 return watermark_ok
;
2346 * This is the direct reclaim path, for page-allocating processes. We only
2347 * try to reclaim pages from zones which will satisfy the caller's allocation
2350 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2352 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2354 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2355 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2356 * zone defense algorithm.
2358 * If a zone is deemed to be full of pinned pages then just give it a light
2359 * scan then give up on it.
2361 * Returns true if a zone was reclaimable.
2363 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2367 unsigned long nr_soft_reclaimed
;
2368 unsigned long nr_soft_scanned
;
2369 unsigned long lru_pages
= 0;
2370 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2372 struct shrink_control shrink
= {
2373 .gfp_mask
= sc
->gfp_mask
,
2375 enum zone_type requested_highidx
= gfp_zone(sc
->gfp_mask
);
2376 bool reclaimable
= false;
2379 * If the number of buffer_heads in the machine exceeds the maximum
2380 * allowed level, force direct reclaim to scan the highmem zone as
2381 * highmem pages could be pinning lowmem pages storing buffer_heads
2383 orig_mask
= sc
->gfp_mask
;
2384 if (buffer_heads_over_limit
)
2385 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2387 nodes_clear(shrink
.nodes_to_scan
);
2389 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2390 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2391 if (!populated_zone(zone
))
2394 * Take care memory controller reclaiming has small influence
2397 if (global_reclaim(sc
)) {
2398 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2401 lru_pages
+= zone_reclaimable_pages(zone
);
2402 node_set(zone_to_nid(zone
), shrink
.nodes_to_scan
);
2404 if (sc
->priority
!= DEF_PRIORITY
&&
2405 !zone_reclaimable(zone
))
2406 continue; /* Let kswapd poll it */
2409 * If we already have plenty of memory free for
2410 * compaction in this zone, don't free any more.
2411 * Even though compaction is invoked for any
2412 * non-zero order, only frequent costly order
2413 * reclamation is disruptive enough to become a
2414 * noticeable problem, like transparent huge
2417 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2418 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2419 zonelist_zone_idx(z
) <= requested_highidx
&&
2420 compaction_ready(zone
, sc
->order
)) {
2421 sc
->compaction_ready
= true;
2426 * This steals pages from memory cgroups over softlimit
2427 * and returns the number of reclaimed pages and
2428 * scanned pages. This works for global memory pressure
2429 * and balancing, not for a memcg's limit.
2431 nr_soft_scanned
= 0;
2432 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2433 sc
->order
, sc
->gfp_mask
,
2435 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2436 sc
->nr_scanned
+= nr_soft_scanned
;
2437 if (nr_soft_reclaimed
)
2439 /* need some check for avoid more shrink_zone() */
2442 if (shrink_zone(zone
, sc
))
2445 if (global_reclaim(sc
) &&
2446 !reclaimable
&& zone_reclaimable(zone
))
2451 * Don't shrink slabs when reclaiming memory from over limit cgroups
2452 * but do shrink slab at least once when aborting reclaim for
2453 * compaction to avoid unevenly scanning file/anon LRU pages over slab
2456 if (global_reclaim(sc
)) {
2457 shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2458 if (reclaim_state
) {
2459 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2460 reclaim_state
->reclaimed_slab
= 0;
2465 * Restore to original mask to avoid the impact on the caller if we
2466 * promoted it to __GFP_HIGHMEM.
2468 sc
->gfp_mask
= orig_mask
;
2474 * This is the main entry point to direct page reclaim.
2476 * If a full scan of the inactive list fails to free enough memory then we
2477 * are "out of memory" and something needs to be killed.
2479 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2480 * high - the zone may be full of dirty or under-writeback pages, which this
2481 * caller can't do much about. We kick the writeback threads and take explicit
2482 * naps in the hope that some of these pages can be written. But if the
2483 * allocating task holds filesystem locks which prevent writeout this might not
2484 * work, and the allocation attempt will fail.
2486 * returns: 0, if no pages reclaimed
2487 * else, the number of pages reclaimed
2489 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2490 struct scan_control
*sc
)
2492 unsigned long total_scanned
= 0;
2493 unsigned long writeback_threshold
;
2494 bool zones_reclaimable
;
2496 delayacct_freepages_start();
2498 if (global_reclaim(sc
))
2499 count_vm_event(ALLOCSTALL
);
2502 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2505 zones_reclaimable
= shrink_zones(zonelist
, sc
);
2507 total_scanned
+= sc
->nr_scanned
;
2508 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2511 if (sc
->compaction_ready
)
2515 * If we're getting trouble reclaiming, start doing
2516 * writepage even in laptop mode.
2518 if (sc
->priority
< DEF_PRIORITY
- 2)
2519 sc
->may_writepage
= 1;
2522 * Try to write back as many pages as we just scanned. This
2523 * tends to cause slow streaming writers to write data to the
2524 * disk smoothly, at the dirtying rate, which is nice. But
2525 * that's undesirable in laptop mode, where we *want* lumpy
2526 * writeout. So in laptop mode, write out the whole world.
2528 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2529 if (total_scanned
> writeback_threshold
) {
2530 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2531 WB_REASON_TRY_TO_FREE_PAGES
);
2532 sc
->may_writepage
= 1;
2534 } while (--sc
->priority
>= 0);
2536 delayacct_freepages_end();
2538 if (sc
->nr_reclaimed
)
2539 return sc
->nr_reclaimed
;
2541 /* Aborted reclaim to try compaction? don't OOM, then */
2542 if (sc
->compaction_ready
)
2545 /* Any of the zones still reclaimable? Don't OOM. */
2546 if (zones_reclaimable
)
2552 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2555 unsigned long pfmemalloc_reserve
= 0;
2556 unsigned long free_pages
= 0;
2560 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2561 zone
= &pgdat
->node_zones
[i
];
2562 if (!populated_zone(zone
))
2565 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2566 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2569 /* If there are no reserves (unexpected config) then do not throttle */
2570 if (!pfmemalloc_reserve
)
2573 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2575 /* kswapd must be awake if processes are being throttled */
2576 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2577 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2578 (enum zone_type
)ZONE_NORMAL
);
2579 wake_up_interruptible(&pgdat
->kswapd_wait
);
2586 * Throttle direct reclaimers if backing storage is backed by the network
2587 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2588 * depleted. kswapd will continue to make progress and wake the processes
2589 * when the low watermark is reached.
2591 * Returns true if a fatal signal was delivered during throttling. If this
2592 * happens, the page allocator should not consider triggering the OOM killer.
2594 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2595 nodemask_t
*nodemask
)
2599 pg_data_t
*pgdat
= NULL
;
2602 * Kernel threads should not be throttled as they may be indirectly
2603 * responsible for cleaning pages necessary for reclaim to make forward
2604 * progress. kjournald for example may enter direct reclaim while
2605 * committing a transaction where throttling it could forcing other
2606 * processes to block on log_wait_commit().
2608 if (current
->flags
& PF_KTHREAD
)
2612 * If a fatal signal is pending, this process should not throttle.
2613 * It should return quickly so it can exit and free its memory
2615 if (fatal_signal_pending(current
))
2619 * Check if the pfmemalloc reserves are ok by finding the first node
2620 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2621 * GFP_KERNEL will be required for allocating network buffers when
2622 * swapping over the network so ZONE_HIGHMEM is unusable.
2624 * Throttling is based on the first usable node and throttled processes
2625 * wait on a queue until kswapd makes progress and wakes them. There
2626 * is an affinity then between processes waking up and where reclaim
2627 * progress has been made assuming the process wakes on the same node.
2628 * More importantly, processes running on remote nodes will not compete
2629 * for remote pfmemalloc reserves and processes on different nodes
2630 * should make reasonable progress.
2632 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2633 gfp_mask
, nodemask
) {
2634 if (zone_idx(zone
) > ZONE_NORMAL
)
2637 /* Throttle based on the first usable node */
2638 pgdat
= zone
->zone_pgdat
;
2639 if (pfmemalloc_watermark_ok(pgdat
))
2644 /* If no zone was usable by the allocation flags then do not throttle */
2648 /* Account for the throttling */
2649 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2652 * If the caller cannot enter the filesystem, it's possible that it
2653 * is due to the caller holding an FS lock or performing a journal
2654 * transaction in the case of a filesystem like ext[3|4]. In this case,
2655 * it is not safe to block on pfmemalloc_wait as kswapd could be
2656 * blocked waiting on the same lock. Instead, throttle for up to a
2657 * second before continuing.
2659 if (!(gfp_mask
& __GFP_FS
)) {
2660 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2661 pfmemalloc_watermark_ok(pgdat
), HZ
);
2666 /* Throttle until kswapd wakes the process */
2667 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2668 pfmemalloc_watermark_ok(pgdat
));
2671 if (fatal_signal_pending(current
))
2678 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2679 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2681 unsigned long nr_reclaimed
;
2682 struct scan_control sc
= {
2683 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2684 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2686 .nodemask
= nodemask
,
2687 .priority
= DEF_PRIORITY
,
2688 .may_writepage
= !laptop_mode
,
2694 * Do not enter reclaim if fatal signal was delivered while throttled.
2695 * 1 is returned so that the page allocator does not OOM kill at this
2698 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2701 trace_mm_vmscan_direct_reclaim_begin(order
,
2705 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2707 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2709 return nr_reclaimed
;
2714 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2715 gfp_t gfp_mask
, bool noswap
,
2717 unsigned long *nr_scanned
)
2719 struct scan_control sc
= {
2720 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2721 .target_mem_cgroup
= memcg
,
2722 .may_writepage
= !laptop_mode
,
2724 .may_swap
= !noswap
,
2726 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2727 int swappiness
= mem_cgroup_swappiness(memcg
);
2729 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2730 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2732 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2737 * NOTE: Although we can get the priority field, using it
2738 * here is not a good idea, since it limits the pages we can scan.
2739 * if we don't reclaim here, the shrink_zone from balance_pgdat
2740 * will pick up pages from other mem cgroup's as well. We hack
2741 * the priority and make it zero.
2743 shrink_lruvec(lruvec
, swappiness
, &sc
);
2745 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2747 *nr_scanned
= sc
.nr_scanned
;
2748 return sc
.nr_reclaimed
;
2751 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2755 struct zonelist
*zonelist
;
2756 unsigned long nr_reclaimed
;
2758 struct scan_control sc
= {
2759 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2760 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2761 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2762 .target_mem_cgroup
= memcg
,
2763 .priority
= DEF_PRIORITY
,
2764 .may_writepage
= !laptop_mode
,
2766 .may_swap
= !noswap
,
2770 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2771 * take care of from where we get pages. So the node where we start the
2772 * scan does not need to be the current node.
2774 nid
= mem_cgroup_select_victim_node(memcg
);
2776 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2778 trace_mm_vmscan_memcg_reclaim_begin(0,
2782 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2784 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2786 return nr_reclaimed
;
2790 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2792 struct mem_cgroup
*memcg
;
2794 if (!total_swap_pages
)
2797 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2799 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2801 if (inactive_anon_is_low(lruvec
))
2802 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2803 sc
, LRU_ACTIVE_ANON
);
2805 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2809 static bool zone_balanced(struct zone
*zone
, int order
,
2810 unsigned long balance_gap
, int classzone_idx
)
2812 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2813 balance_gap
, classzone_idx
, 0))
2816 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&&
2817 !compaction_suitable(zone
, order
))
2824 * pgdat_balanced() is used when checking if a node is balanced.
2826 * For order-0, all zones must be balanced!
2828 * For high-order allocations only zones that meet watermarks and are in a
2829 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2830 * total of balanced pages must be at least 25% of the zones allowed by
2831 * classzone_idx for the node to be considered balanced. Forcing all zones to
2832 * be balanced for high orders can cause excessive reclaim when there are
2834 * The choice of 25% is due to
2835 * o a 16M DMA zone that is balanced will not balance a zone on any
2836 * reasonable sized machine
2837 * o On all other machines, the top zone must be at least a reasonable
2838 * percentage of the middle zones. For example, on 32-bit x86, highmem
2839 * would need to be at least 256M for it to be balance a whole node.
2840 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2841 * to balance a node on its own. These seemed like reasonable ratios.
2843 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2845 unsigned long managed_pages
= 0;
2846 unsigned long balanced_pages
= 0;
2849 /* Check the watermark levels */
2850 for (i
= 0; i
<= classzone_idx
; i
++) {
2851 struct zone
*zone
= pgdat
->node_zones
+ i
;
2853 if (!populated_zone(zone
))
2856 managed_pages
+= zone
->managed_pages
;
2859 * A special case here:
2861 * balance_pgdat() skips over all_unreclaimable after
2862 * DEF_PRIORITY. Effectively, it considers them balanced so
2863 * they must be considered balanced here as well!
2865 if (!zone_reclaimable(zone
)) {
2866 balanced_pages
+= zone
->managed_pages
;
2870 if (zone_balanced(zone
, order
, 0, i
))
2871 balanced_pages
+= zone
->managed_pages
;
2877 return balanced_pages
>= (managed_pages
>> 2);
2883 * Prepare kswapd for sleeping. This verifies that there are no processes
2884 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2886 * Returns true if kswapd is ready to sleep
2888 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
2891 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2896 * There is a potential race between when kswapd checks its watermarks
2897 * and a process gets throttled. There is also a potential race if
2898 * processes get throttled, kswapd wakes, a large process exits therby
2899 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2900 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2901 * so wake them now if necessary. If necessary, processes will wake
2902 * kswapd and get throttled again
2904 if (waitqueue_active(&pgdat
->pfmemalloc_wait
)) {
2905 wake_up(&pgdat
->pfmemalloc_wait
);
2909 return pgdat_balanced(pgdat
, order
, classzone_idx
);
2913 * kswapd shrinks the zone by the number of pages required to reach
2914 * the high watermark.
2916 * Returns true if kswapd scanned at least the requested number of pages to
2917 * reclaim or if the lack of progress was due to pages under writeback.
2918 * This is used to determine if the scanning priority needs to be raised.
2920 static bool kswapd_shrink_zone(struct zone
*zone
,
2922 struct scan_control
*sc
,
2923 unsigned long lru_pages
,
2924 unsigned long *nr_attempted
)
2926 int testorder
= sc
->order
;
2927 unsigned long balance_gap
;
2928 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2929 struct shrink_control shrink
= {
2930 .gfp_mask
= sc
->gfp_mask
,
2932 bool lowmem_pressure
;
2934 /* Reclaim above the high watermark. */
2935 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
2938 * Kswapd reclaims only single pages with compaction enabled. Trying
2939 * too hard to reclaim until contiguous free pages have become
2940 * available can hurt performance by evicting too much useful data
2941 * from memory. Do not reclaim more than needed for compaction.
2943 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2944 compaction_suitable(zone
, sc
->order
) !=
2949 * We put equal pressure on every zone, unless one zone has way too
2950 * many pages free already. The "too many pages" is defined as the
2951 * high wmark plus a "gap" where the gap is either the low
2952 * watermark or 1% of the zone, whichever is smaller.
2954 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2955 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2958 * If there is no low memory pressure or the zone is balanced then no
2959 * reclaim is necessary
2961 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
2962 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
2963 balance_gap
, classzone_idx
))
2966 shrink_zone(zone
, sc
);
2967 nodes_clear(shrink
.nodes_to_scan
);
2968 node_set(zone_to_nid(zone
), shrink
.nodes_to_scan
);
2970 reclaim_state
->reclaimed_slab
= 0;
2971 shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2972 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2974 /* Account for the number of pages attempted to reclaim */
2975 *nr_attempted
+= sc
->nr_to_reclaim
;
2977 zone_clear_flag(zone
, ZONE_WRITEBACK
);
2980 * If a zone reaches its high watermark, consider it to be no longer
2981 * congested. It's possible there are dirty pages backed by congested
2982 * BDIs but as pressure is relieved, speculatively avoid congestion
2985 if (zone_reclaimable(zone
) &&
2986 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
2987 zone_clear_flag(zone
, ZONE_CONGESTED
);
2988 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2991 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
2995 * For kswapd, balance_pgdat() will work across all this node's zones until
2996 * they are all at high_wmark_pages(zone).
2998 * Returns the final order kswapd was reclaiming at
3000 * There is special handling here for zones which are full of pinned pages.
3001 * This can happen if the pages are all mlocked, or if they are all used by
3002 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3003 * What we do is to detect the case where all pages in the zone have been
3004 * scanned twice and there has been zero successful reclaim. Mark the zone as
3005 * dead and from now on, only perform a short scan. Basically we're polling
3006 * the zone for when the problem goes away.
3008 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3009 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3010 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3011 * lower zones regardless of the number of free pages in the lower zones. This
3012 * interoperates with the page allocator fallback scheme to ensure that aging
3013 * of pages is balanced across the zones.
3015 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
3019 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
3020 unsigned long nr_soft_reclaimed
;
3021 unsigned long nr_soft_scanned
;
3022 struct scan_control sc
= {
3023 .gfp_mask
= GFP_KERNEL
,
3025 .priority
= DEF_PRIORITY
,
3026 .may_writepage
= !laptop_mode
,
3030 count_vm_event(PAGEOUTRUN
);
3033 unsigned long lru_pages
= 0;
3034 unsigned long nr_attempted
= 0;
3035 bool raise_priority
= true;
3036 bool pgdat_needs_compaction
= (order
> 0);
3038 sc
.nr_reclaimed
= 0;
3041 * Scan in the highmem->dma direction for the highest
3042 * zone which needs scanning
3044 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
3045 struct zone
*zone
= pgdat
->node_zones
+ i
;
3047 if (!populated_zone(zone
))
3050 if (sc
.priority
!= DEF_PRIORITY
&&
3051 !zone_reclaimable(zone
))
3055 * Do some background aging of the anon list, to give
3056 * pages a chance to be referenced before reclaiming.
3058 age_active_anon(zone
, &sc
);
3061 * If the number of buffer_heads in the machine
3062 * exceeds the maximum allowed level and this node
3063 * has a highmem zone, force kswapd to reclaim from
3064 * it to relieve lowmem pressure.
3066 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
3071 if (!zone_balanced(zone
, order
, 0, 0)) {
3076 * If balanced, clear the dirty and congested
3079 zone_clear_flag(zone
, ZONE_CONGESTED
);
3080 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
3087 for (i
= 0; i
<= end_zone
; i
++) {
3088 struct zone
*zone
= pgdat
->node_zones
+ i
;
3090 if (!populated_zone(zone
))
3093 lru_pages
+= zone_reclaimable_pages(zone
);
3096 * If any zone is currently balanced then kswapd will
3097 * not call compaction as it is expected that the
3098 * necessary pages are already available.
3100 if (pgdat_needs_compaction
&&
3101 zone_watermark_ok(zone
, order
,
3102 low_wmark_pages(zone
),
3104 pgdat_needs_compaction
= false;
3108 * If we're getting trouble reclaiming, start doing writepage
3109 * even in laptop mode.
3111 if (sc
.priority
< DEF_PRIORITY
- 2)
3112 sc
.may_writepage
= 1;
3115 * Now scan the zone in the dma->highmem direction, stopping
3116 * at the last zone which needs scanning.
3118 * We do this because the page allocator works in the opposite
3119 * direction. This prevents the page allocator from allocating
3120 * pages behind kswapd's direction of progress, which would
3121 * cause too much scanning of the lower zones.
3123 for (i
= 0; i
<= end_zone
; i
++) {
3124 struct zone
*zone
= pgdat
->node_zones
+ i
;
3126 if (!populated_zone(zone
))
3129 if (sc
.priority
!= DEF_PRIORITY
&&
3130 !zone_reclaimable(zone
))
3135 nr_soft_scanned
= 0;
3137 * Call soft limit reclaim before calling shrink_zone.
3139 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3142 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3145 * There should be no need to raise the scanning
3146 * priority if enough pages are already being scanned
3147 * that that high watermark would be met at 100%
3150 if (kswapd_shrink_zone(zone
, end_zone
, &sc
,
3151 lru_pages
, &nr_attempted
))
3152 raise_priority
= false;
3156 * If the low watermark is met there is no need for processes
3157 * to be throttled on pfmemalloc_wait as they should not be
3158 * able to safely make forward progress. Wake them
3160 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3161 pfmemalloc_watermark_ok(pgdat
))
3162 wake_up(&pgdat
->pfmemalloc_wait
);
3165 * Fragmentation may mean that the system cannot be rebalanced
3166 * for high-order allocations in all zones. If twice the
3167 * allocation size has been reclaimed and the zones are still
3168 * not balanced then recheck the watermarks at order-0 to
3169 * prevent kswapd reclaiming excessively. Assume that a
3170 * process requested a high-order can direct reclaim/compact.
3172 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3173 order
= sc
.order
= 0;
3175 /* Check if kswapd should be suspending */
3176 if (try_to_freeze() || kthread_should_stop())
3180 * Compact if necessary and kswapd is reclaiming at least the
3181 * high watermark number of pages as requsted
3183 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3184 compact_pgdat(pgdat
, order
);
3187 * Raise priority if scanning rate is too low or there was no
3188 * progress in reclaiming pages
3190 if (raise_priority
|| !sc
.nr_reclaimed
)
3192 } while (sc
.priority
>= 1 &&
3193 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3197 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3198 * makes a decision on the order we were last reclaiming at. However,
3199 * if another caller entered the allocator slow path while kswapd
3200 * was awake, order will remain at the higher level
3202 *classzone_idx
= end_zone
;
3206 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3211 if (freezing(current
) || kthread_should_stop())
3214 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3216 /* Try to sleep for a short interval */
3217 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3218 remaining
= schedule_timeout(HZ
/10);
3219 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3220 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3224 * After a short sleep, check if it was a premature sleep. If not, then
3225 * go fully to sleep until explicitly woken up.
3227 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3228 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3231 * vmstat counters are not perfectly accurate and the estimated
3232 * value for counters such as NR_FREE_PAGES can deviate from the
3233 * true value by nr_online_cpus * threshold. To avoid the zone
3234 * watermarks being breached while under pressure, we reduce the
3235 * per-cpu vmstat threshold while kswapd is awake and restore
3236 * them before going back to sleep.
3238 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3241 * Compaction records what page blocks it recently failed to
3242 * isolate pages from and skips them in the future scanning.
3243 * When kswapd is going to sleep, it is reasonable to assume
3244 * that pages and compaction may succeed so reset the cache.
3246 reset_isolation_suitable(pgdat
);
3248 if (!kthread_should_stop())
3251 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3254 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3256 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3258 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3262 * The background pageout daemon, started as a kernel thread
3263 * from the init process.
3265 * This basically trickles out pages so that we have _some_
3266 * free memory available even if there is no other activity
3267 * that frees anything up. This is needed for things like routing
3268 * etc, where we otherwise might have all activity going on in
3269 * asynchronous contexts that cannot page things out.
3271 * If there are applications that are active memory-allocators
3272 * (most normal use), this basically shouldn't matter.
3274 static int kswapd(void *p
)
3276 unsigned long order
, new_order
;
3277 unsigned balanced_order
;
3278 int classzone_idx
, new_classzone_idx
;
3279 int balanced_classzone_idx
;
3280 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3281 struct task_struct
*tsk
= current
;
3283 struct reclaim_state reclaim_state
= {
3284 .reclaimed_slab
= 0,
3286 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3288 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3290 if (!cpumask_empty(cpumask
))
3291 set_cpus_allowed_ptr(tsk
, cpumask
);
3292 current
->reclaim_state
= &reclaim_state
;
3295 * Tell the memory management that we're a "memory allocator",
3296 * and that if we need more memory we should get access to it
3297 * regardless (see "__alloc_pages()"). "kswapd" should
3298 * never get caught in the normal page freeing logic.
3300 * (Kswapd normally doesn't need memory anyway, but sometimes
3301 * you need a small amount of memory in order to be able to
3302 * page out something else, and this flag essentially protects
3303 * us from recursively trying to free more memory as we're
3304 * trying to free the first piece of memory in the first place).
3306 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3309 order
= new_order
= 0;
3311 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3312 balanced_classzone_idx
= classzone_idx
;
3317 * If the last balance_pgdat was unsuccessful it's unlikely a
3318 * new request of a similar or harder type will succeed soon
3319 * so consider going to sleep on the basis we reclaimed at
3321 if (balanced_classzone_idx
>= new_classzone_idx
&&
3322 balanced_order
== new_order
) {
3323 new_order
= pgdat
->kswapd_max_order
;
3324 new_classzone_idx
= pgdat
->classzone_idx
;
3325 pgdat
->kswapd_max_order
= 0;
3326 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3329 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3331 * Don't sleep if someone wants a larger 'order'
3332 * allocation or has tigher zone constraints
3335 classzone_idx
= new_classzone_idx
;
3337 kswapd_try_to_sleep(pgdat
, balanced_order
,
3338 balanced_classzone_idx
);
3339 order
= pgdat
->kswapd_max_order
;
3340 classzone_idx
= pgdat
->classzone_idx
;
3342 new_classzone_idx
= classzone_idx
;
3343 pgdat
->kswapd_max_order
= 0;
3344 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3347 ret
= try_to_freeze();
3348 if (kthread_should_stop())
3352 * We can speed up thawing tasks if we don't call balance_pgdat
3353 * after returning from the refrigerator
3356 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3357 balanced_classzone_idx
= classzone_idx
;
3358 balanced_order
= balance_pgdat(pgdat
, order
,
3359 &balanced_classzone_idx
);
3363 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3364 current
->reclaim_state
= NULL
;
3365 lockdep_clear_current_reclaim_state();
3371 * A zone is low on free memory, so wake its kswapd task to service it.
3373 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3377 if (!populated_zone(zone
))
3380 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
3382 pgdat
= zone
->zone_pgdat
;
3383 if (pgdat
->kswapd_max_order
< order
) {
3384 pgdat
->kswapd_max_order
= order
;
3385 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3387 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3389 if (zone_balanced(zone
, order
, 0, 0))
3392 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3393 wake_up_interruptible(&pgdat
->kswapd_wait
);
3396 #ifdef CONFIG_HIBERNATION
3398 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3401 * Rather than trying to age LRUs the aim is to preserve the overall
3402 * LRU order by reclaiming preferentially
3403 * inactive > active > active referenced > active mapped
3405 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3407 struct reclaim_state reclaim_state
;
3408 struct scan_control sc
= {
3409 .nr_to_reclaim
= nr_to_reclaim
,
3410 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3411 .priority
= DEF_PRIORITY
,
3415 .hibernation_mode
= 1,
3417 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3418 struct task_struct
*p
= current
;
3419 unsigned long nr_reclaimed
;
3421 p
->flags
|= PF_MEMALLOC
;
3422 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3423 reclaim_state
.reclaimed_slab
= 0;
3424 p
->reclaim_state
= &reclaim_state
;
3426 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3428 p
->reclaim_state
= NULL
;
3429 lockdep_clear_current_reclaim_state();
3430 p
->flags
&= ~PF_MEMALLOC
;
3432 return nr_reclaimed
;
3434 #endif /* CONFIG_HIBERNATION */
3436 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3437 not required for correctness. So if the last cpu in a node goes
3438 away, we get changed to run anywhere: as the first one comes back,
3439 restore their cpu bindings. */
3440 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3445 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3446 for_each_node_state(nid
, N_MEMORY
) {
3447 pg_data_t
*pgdat
= NODE_DATA(nid
);
3448 const struct cpumask
*mask
;
3450 mask
= cpumask_of_node(pgdat
->node_id
);
3452 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3453 /* One of our CPUs online: restore mask */
3454 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3461 * This kswapd start function will be called by init and node-hot-add.
3462 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3464 int kswapd_run(int nid
)
3466 pg_data_t
*pgdat
= NODE_DATA(nid
);
3472 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3473 if (IS_ERR(pgdat
->kswapd
)) {
3474 /* failure at boot is fatal */
3475 BUG_ON(system_state
== SYSTEM_BOOTING
);
3476 pr_err("Failed to start kswapd on node %d\n", nid
);
3477 ret
= PTR_ERR(pgdat
->kswapd
);
3478 pgdat
->kswapd
= NULL
;
3484 * Called by memory hotplug when all memory in a node is offlined. Caller must
3485 * hold mem_hotplug_begin/end().
3487 void kswapd_stop(int nid
)
3489 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3492 kthread_stop(kswapd
);
3493 NODE_DATA(nid
)->kswapd
= NULL
;
3497 static int __init
kswapd_init(void)
3502 for_each_node_state(nid
, N_MEMORY
)
3504 hotcpu_notifier(cpu_callback
, 0);
3508 module_init(kswapd_init
)
3514 * If non-zero call zone_reclaim when the number of free pages falls below
3517 int zone_reclaim_mode __read_mostly
;
3519 #define RECLAIM_OFF 0
3520 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3521 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3522 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3525 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3526 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3529 #define ZONE_RECLAIM_PRIORITY 4
3532 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3535 int sysctl_min_unmapped_ratio
= 1;
3538 * If the number of slab pages in a zone grows beyond this percentage then
3539 * slab reclaim needs to occur.
3541 int sysctl_min_slab_ratio
= 5;
3543 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3545 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3546 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3547 zone_page_state(zone
, NR_ACTIVE_FILE
);
3550 * It's possible for there to be more file mapped pages than
3551 * accounted for by the pages on the file LRU lists because
3552 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3554 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3557 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3558 static long zone_pagecache_reclaimable(struct zone
*zone
)
3560 long nr_pagecache_reclaimable
;
3564 * If RECLAIM_SWAP is set, then all file pages are considered
3565 * potentially reclaimable. Otherwise, we have to worry about
3566 * pages like swapcache and zone_unmapped_file_pages() provides
3569 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3570 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3572 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3574 /* If we can't clean pages, remove dirty pages from consideration */
3575 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3576 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3578 /* Watch for any possible underflows due to delta */
3579 if (unlikely(delta
> nr_pagecache_reclaimable
))
3580 delta
= nr_pagecache_reclaimable
;
3582 return nr_pagecache_reclaimable
- delta
;
3586 * Try to free up some pages from this zone through reclaim.
3588 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3590 /* Minimum pages needed in order to stay on node */
3591 const unsigned long nr_pages
= 1 << order
;
3592 struct task_struct
*p
= current
;
3593 struct reclaim_state reclaim_state
;
3594 struct scan_control sc
= {
3595 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3596 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3598 .priority
= ZONE_RECLAIM_PRIORITY
,
3599 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3600 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3603 struct shrink_control shrink
= {
3604 .gfp_mask
= sc
.gfp_mask
,
3606 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3610 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3611 * and we also need to be able to write out pages for RECLAIM_WRITE
3614 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3615 lockdep_set_current_reclaim_state(gfp_mask
);
3616 reclaim_state
.reclaimed_slab
= 0;
3617 p
->reclaim_state
= &reclaim_state
;
3619 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3621 * Free memory by calling shrink zone with increasing
3622 * priorities until we have enough memory freed.
3625 shrink_zone(zone
, &sc
);
3626 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3629 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3630 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3632 * shrink_slab() does not currently allow us to determine how
3633 * many pages were freed in this zone. So we take the current
3634 * number of slab pages and shake the slab until it is reduced
3635 * by the same nr_pages that we used for reclaiming unmapped
3638 nodes_clear(shrink
.nodes_to_scan
);
3639 node_set(zone_to_nid(zone
), shrink
.nodes_to_scan
);
3641 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3643 /* No reclaimable slab or very low memory pressure */
3644 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3647 /* Freed enough memory */
3648 nr_slab_pages1
= zone_page_state(zone
,
3649 NR_SLAB_RECLAIMABLE
);
3650 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3655 * Update nr_reclaimed by the number of slab pages we
3656 * reclaimed from this zone.
3658 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3659 if (nr_slab_pages1
< nr_slab_pages0
)
3660 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3663 p
->reclaim_state
= NULL
;
3664 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3665 lockdep_clear_current_reclaim_state();
3666 return sc
.nr_reclaimed
>= nr_pages
;
3669 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3675 * Zone reclaim reclaims unmapped file backed pages and
3676 * slab pages if we are over the defined limits.
3678 * A small portion of unmapped file backed pages is needed for
3679 * file I/O otherwise pages read by file I/O will be immediately
3680 * thrown out if the zone is overallocated. So we do not reclaim
3681 * if less than a specified percentage of the zone is used by
3682 * unmapped file backed pages.
3684 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3685 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3686 return ZONE_RECLAIM_FULL
;
3688 if (!zone_reclaimable(zone
))
3689 return ZONE_RECLAIM_FULL
;
3692 * Do not scan if the allocation should not be delayed.
3694 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3695 return ZONE_RECLAIM_NOSCAN
;
3698 * Only run zone reclaim on the local zone or on zones that do not
3699 * have associated processors. This will favor the local processor
3700 * over remote processors and spread off node memory allocations
3701 * as wide as possible.
3703 node_id
= zone_to_nid(zone
);
3704 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3705 return ZONE_RECLAIM_NOSCAN
;
3707 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3708 return ZONE_RECLAIM_NOSCAN
;
3710 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3711 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3714 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3721 * page_evictable - test whether a page is evictable
3722 * @page: the page to test
3724 * Test whether page is evictable--i.e., should be placed on active/inactive
3725 * lists vs unevictable list.
3727 * Reasons page might not be evictable:
3728 * (1) page's mapping marked unevictable
3729 * (2) page is part of an mlocked VMA
3732 int page_evictable(struct page
*page
)
3734 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3739 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3740 * @pages: array of pages to check
3741 * @nr_pages: number of pages to check
3743 * Checks pages for evictability and moves them to the appropriate lru list.
3745 * This function is only used for SysV IPC SHM_UNLOCK.
3747 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3749 struct lruvec
*lruvec
;
3750 struct zone
*zone
= NULL
;
3755 for (i
= 0; i
< nr_pages
; i
++) {
3756 struct page
*page
= pages
[i
];
3757 struct zone
*pagezone
;
3760 pagezone
= page_zone(page
);
3761 if (pagezone
!= zone
) {
3763 spin_unlock_irq(&zone
->lru_lock
);
3765 spin_lock_irq(&zone
->lru_lock
);
3767 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3769 if (!PageLRU(page
) || !PageUnevictable(page
))
3772 if (page_evictable(page
)) {
3773 enum lru_list lru
= page_lru_base_type(page
);
3775 VM_BUG_ON_PAGE(PageActive(page
), page
);
3776 ClearPageUnevictable(page
);
3777 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3778 add_page_to_lru_list(page
, lruvec
, lru
);
3784 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3785 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3786 spin_unlock_irq(&zone
->lru_lock
);
3789 #endif /* CONFIG_SHMEM */
3791 static void warn_scan_unevictable_pages(void)
3793 printk_once(KERN_WARNING
3794 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3795 "disabled for lack of a legitimate use case. If you have "
3796 "one, please send an email to linux-mm@kvack.org.\n",
3801 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3802 * all nodes' unevictable lists for evictable pages
3804 unsigned long scan_unevictable_pages
;
3806 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3807 void __user
*buffer
,
3808 size_t *length
, loff_t
*ppos
)
3810 warn_scan_unevictable_pages();
3811 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3812 scan_unevictable_pages
= 0;
3818 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3819 * a specified node's per zone unevictable lists for evictable pages.
3822 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3823 struct device_attribute
*attr
,
3826 warn_scan_unevictable_pages();
3827 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3830 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3831 struct device_attribute
*attr
,
3832 const char *buf
, size_t count
)
3834 warn_scan_unevictable_pages();
3839 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3840 read_scan_unevictable_node
,
3841 write_scan_unevictable_node
);
3843 int scan_unevictable_register_node(struct node
*node
)
3845 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3848 void scan_unevictable_unregister_node(struct node
*node
)
3850 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);