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/sched/mm.h>
18 #include <linux/module.h>
19 #include <linux/gfp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/vmpressure.h>
26 #include <linux/vmstat.h>
27 #include <linux/file.h>
28 #include <linux/writeback.h>
29 #include <linux/blkdev.h>
30 #include <linux/buffer_head.h> /* for try_to_release_page(),
31 buffer_heads_over_limit */
32 #include <linux/mm_inline.h>
33 #include <linux/backing-dev.h>
34 #include <linux/rmap.h>
35 #include <linux/topology.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/compaction.h>
39 #include <linux/notifier.h>
40 #include <linux/rwsem.h>
41 #include <linux/delay.h>
42 #include <linux/kthread.h>
43 #include <linux/freezer.h>
44 #include <linux/memcontrol.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
55 #include <linux/swapops.h>
56 #include <linux/balloon_compaction.h>
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/vmscan.h>
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim
;
67 /* This context's GFP mask */
70 /* Allocation order */
74 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * The memory cgroup that hit its limit and as a result is the
81 * primary target of this reclaim invocation.
83 struct mem_cgroup
*target_mem_cgroup
;
85 /* Scan (total_size >> priority) pages at once */
88 /* The highest zone to isolate pages for reclaim from */
89 enum zone_type reclaim_idx
;
91 /* Writepage batching in laptop mode; RECLAIM_WRITE */
92 unsigned int may_writepage
:1;
94 /* Can mapped pages be reclaimed? */
95 unsigned int may_unmap
:1;
97 /* Can pages be swapped as part of reclaim? */
98 unsigned int may_swap
:1;
100 /* Can cgroups be reclaimed below their normal consumption range? */
101 unsigned int may_thrash
:1;
103 unsigned int hibernation_mode
:1;
105 /* One of the zones is ready for compaction */
106 unsigned int compaction_ready
:1;
108 /* Incremented by the number of inactive pages that were scanned */
109 unsigned long nr_scanned
;
111 /* Number of pages freed so far during a call to shrink_zones() */
112 unsigned long nr_reclaimed
;
115 #ifdef ARCH_HAS_PREFETCH
116 #define prefetch_prev_lru_page(_page, _base, _field) \
118 if ((_page)->lru.prev != _base) { \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetch(&prev->_field); \
126 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
129 #ifdef ARCH_HAS_PREFETCHW
130 #define prefetchw_prev_lru_page(_page, _base, _field) \
132 if ((_page)->lru.prev != _base) { \
135 prev = lru_to_page(&(_page->lru)); \
136 prefetchw(&prev->_field); \
140 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 * From 0 .. 100. Higher means more swappy.
146 int vm_swappiness
= 60;
148 * The total number of pages which are beyond the high watermark within all
151 unsigned long vm_total_pages
;
153 static LIST_HEAD(shrinker_list
);
154 static DECLARE_RWSEM(shrinker_rwsem
);
157 static bool global_reclaim(struct scan_control
*sc
)
159 return !sc
->target_mem_cgroup
;
163 * sane_reclaim - is the usual dirty throttling mechanism operational?
164 * @sc: scan_control in question
166 * The normal page dirty throttling mechanism in balance_dirty_pages() is
167 * completely broken with the legacy memcg and direct stalling in
168 * shrink_page_list() is used for throttling instead, which lacks all the
169 * niceties such as fairness, adaptive pausing, bandwidth proportional
170 * allocation and configurability.
172 * This function tests whether the vmscan currently in progress can assume
173 * that the normal dirty throttling mechanism is operational.
175 static bool sane_reclaim(struct scan_control
*sc
)
177 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
181 #ifdef CONFIG_CGROUP_WRITEBACK
182 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
188 static bool global_reclaim(struct scan_control
*sc
)
193 static bool sane_reclaim(struct scan_control
*sc
)
200 * This misses isolated pages which are not accounted for to save counters.
201 * As the data only determines if reclaim or compaction continues, it is
202 * not expected that isolated pages will be a dominating factor.
204 unsigned long zone_reclaimable_pages(struct zone
*zone
)
208 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
209 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
210 if (get_nr_swap_pages() > 0)
211 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
212 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
217 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
221 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
222 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
223 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
225 if (get_nr_swap_pages() > 0)
226 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
227 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
228 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
234 * lruvec_lru_size - Returns the number of pages on the given LRU list.
235 * @lruvec: lru vector
237 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
239 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
241 unsigned long lru_size
;
244 if (!mem_cgroup_disabled())
245 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
247 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
249 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
250 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
253 if (!managed_zone(zone
))
256 if (!mem_cgroup_disabled())
257 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
259 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
260 NR_ZONE_LRU_BASE
+ lru
);
261 lru_size
-= min(size
, lru_size
);
269 * Add a shrinker callback to be called from the vm.
271 int register_shrinker(struct shrinker
*shrinker
)
273 size_t size
= sizeof(*shrinker
->nr_deferred
);
275 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
278 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
279 if (!shrinker
->nr_deferred
)
282 down_write(&shrinker_rwsem
);
283 list_add_tail(&shrinker
->list
, &shrinker_list
);
284 up_write(&shrinker_rwsem
);
287 EXPORT_SYMBOL(register_shrinker
);
292 void unregister_shrinker(struct shrinker
*shrinker
)
294 down_write(&shrinker_rwsem
);
295 list_del(&shrinker
->list
);
296 up_write(&shrinker_rwsem
);
297 kfree(shrinker
->nr_deferred
);
299 EXPORT_SYMBOL(unregister_shrinker
);
301 #define SHRINK_BATCH 128
303 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
304 struct shrinker
*shrinker
,
305 unsigned long nr_scanned
,
306 unsigned long nr_eligible
)
308 unsigned long freed
= 0;
309 unsigned long long delta
;
314 int nid
= shrinkctl
->nid
;
315 long batch_size
= shrinker
->batch
? shrinker
->batch
317 long scanned
= 0, next_deferred
;
319 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
324 * copy the current shrinker scan count into a local variable
325 * and zero it so that other concurrent shrinker invocations
326 * don't also do this scanning work.
328 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
331 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
333 do_div(delta
, nr_eligible
+ 1);
335 if (total_scan
< 0) {
336 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
337 shrinker
->scan_objects
, total_scan
);
338 total_scan
= freeable
;
341 next_deferred
= total_scan
;
344 * We need to avoid excessive windup on filesystem shrinkers
345 * due to large numbers of GFP_NOFS allocations causing the
346 * shrinkers to return -1 all the time. This results in a large
347 * nr being built up so when a shrink that can do some work
348 * comes along it empties the entire cache due to nr >>>
349 * freeable. This is bad for sustaining a working set in
352 * Hence only allow the shrinker to scan the entire cache when
353 * a large delta change is calculated directly.
355 if (delta
< freeable
/ 4)
356 total_scan
= min(total_scan
, freeable
/ 2);
359 * Avoid risking looping forever due to too large nr value:
360 * never try to free more than twice the estimate number of
363 if (total_scan
> freeable
* 2)
364 total_scan
= freeable
* 2;
366 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
367 nr_scanned
, nr_eligible
,
368 freeable
, delta
, total_scan
);
371 * Normally, we should not scan less than batch_size objects in one
372 * pass to avoid too frequent shrinker calls, but if the slab has less
373 * than batch_size objects in total and we are really tight on memory,
374 * we will try to reclaim all available objects, otherwise we can end
375 * up failing allocations although there are plenty of reclaimable
376 * objects spread over several slabs with usage less than the
379 * We detect the "tight on memory" situations by looking at the total
380 * number of objects we want to scan (total_scan). If it is greater
381 * than the total number of objects on slab (freeable), we must be
382 * scanning at high prio and therefore should try to reclaim as much as
385 while (total_scan
>= batch_size
||
386 total_scan
>= freeable
) {
388 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
390 shrinkctl
->nr_to_scan
= nr_to_scan
;
391 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
392 if (ret
== SHRINK_STOP
)
396 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
397 total_scan
-= nr_to_scan
;
398 scanned
+= nr_to_scan
;
403 if (next_deferred
>= scanned
)
404 next_deferred
-= scanned
;
408 * move the unused scan count back into the shrinker in a
409 * manner that handles concurrent updates. If we exhausted the
410 * scan, there is no need to do an update.
412 if (next_deferred
> 0)
413 new_nr
= atomic_long_add_return(next_deferred
,
414 &shrinker
->nr_deferred
[nid
]);
416 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
418 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
423 * shrink_slab - shrink slab caches
424 * @gfp_mask: allocation context
425 * @nid: node whose slab caches to target
426 * @memcg: memory cgroup whose slab caches to target
427 * @nr_scanned: pressure numerator
428 * @nr_eligible: pressure denominator
430 * Call the shrink functions to age shrinkable caches.
432 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
433 * unaware shrinkers will receive a node id of 0 instead.
435 * @memcg specifies the memory cgroup to target. If it is not NULL,
436 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
437 * objects from the memory cgroup specified. Otherwise, only unaware
438 * shrinkers are called.
440 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
441 * the available objects should be scanned. Page reclaim for example
442 * passes the number of pages scanned and the number of pages on the
443 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
444 * when it encountered mapped pages. The ratio is further biased by
445 * the ->seeks setting of the shrink function, which indicates the
446 * cost to recreate an object relative to that of an LRU page.
448 * Returns the number of reclaimed slab objects.
450 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
451 struct mem_cgroup
*memcg
,
452 unsigned long nr_scanned
,
453 unsigned long nr_eligible
)
455 struct shrinker
*shrinker
;
456 unsigned long freed
= 0;
458 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
462 nr_scanned
= SWAP_CLUSTER_MAX
;
464 if (!down_read_trylock(&shrinker_rwsem
)) {
466 * If we would return 0, our callers would understand that we
467 * have nothing else to shrink and give up trying. By returning
468 * 1 we keep it going and assume we'll be able to shrink next
475 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
476 struct shrink_control sc
= {
477 .gfp_mask
= gfp_mask
,
483 * If kernel memory accounting is disabled, we ignore
484 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
485 * passing NULL for memcg.
487 if (memcg_kmem_enabled() &&
488 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
491 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
494 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
497 up_read(&shrinker_rwsem
);
503 void drop_slab_node(int nid
)
508 struct mem_cgroup
*memcg
= NULL
;
512 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
514 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
515 } while (freed
> 10);
522 for_each_online_node(nid
)
526 static inline int is_page_cache_freeable(struct page
*page
)
529 * A freeable page cache page is referenced only by the caller
530 * that isolated the page, the page cache radix tree and
531 * optional buffer heads at page->private.
533 return page_count(page
) - page_has_private(page
) == 2;
536 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
538 if (current
->flags
& PF_SWAPWRITE
)
540 if (!inode_write_congested(inode
))
542 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
548 * We detected a synchronous write error writing a page out. Probably
549 * -ENOSPC. We need to propagate that into the address_space for a subsequent
550 * fsync(), msync() or close().
552 * The tricky part is that after writepage we cannot touch the mapping: nothing
553 * prevents it from being freed up. But we have a ref on the page and once
554 * that page is locked, the mapping is pinned.
556 * We're allowed to run sleeping lock_page() here because we know the caller has
559 static void handle_write_error(struct address_space
*mapping
,
560 struct page
*page
, int error
)
563 if (page_mapping(page
) == mapping
)
564 mapping_set_error(mapping
, error
);
568 /* possible outcome of pageout() */
570 /* failed to write page out, page is locked */
572 /* move page to the active list, page is locked */
574 /* page has been sent to the disk successfully, page is unlocked */
576 /* page is clean and locked */
581 * pageout is called by shrink_page_list() for each dirty page.
582 * Calls ->writepage().
584 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
585 struct scan_control
*sc
)
588 * If the page is dirty, only perform writeback if that write
589 * will be non-blocking. To prevent this allocation from being
590 * stalled by pagecache activity. But note that there may be
591 * stalls if we need to run get_block(). We could test
592 * PagePrivate for that.
594 * If this process is currently in __generic_file_write_iter() against
595 * this page's queue, we can perform writeback even if that
598 * If the page is swapcache, write it back even if that would
599 * block, for some throttling. This happens by accident, because
600 * swap_backing_dev_info is bust: it doesn't reflect the
601 * congestion state of the swapdevs. Easy to fix, if needed.
603 if (!is_page_cache_freeable(page
))
607 * Some data journaling orphaned pages can have
608 * page->mapping == NULL while being dirty with clean buffers.
610 if (page_has_private(page
)) {
611 if (try_to_free_buffers(page
)) {
612 ClearPageDirty(page
);
613 pr_info("%s: orphaned page\n", __func__
);
619 if (mapping
->a_ops
->writepage
== NULL
)
620 return PAGE_ACTIVATE
;
621 if (!may_write_to_inode(mapping
->host
, sc
))
624 if (clear_page_dirty_for_io(page
)) {
626 struct writeback_control wbc
= {
627 .sync_mode
= WB_SYNC_NONE
,
628 .nr_to_write
= SWAP_CLUSTER_MAX
,
630 .range_end
= LLONG_MAX
,
634 SetPageReclaim(page
);
635 res
= mapping
->a_ops
->writepage(page
, &wbc
);
637 handle_write_error(mapping
, page
, res
);
638 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
639 ClearPageReclaim(page
);
640 return PAGE_ACTIVATE
;
643 if (!PageWriteback(page
)) {
644 /* synchronous write or broken a_ops? */
645 ClearPageReclaim(page
);
647 trace_mm_vmscan_writepage(page
);
648 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
656 * Same as remove_mapping, but if the page is removed from the mapping, it
657 * gets returned with a refcount of 0.
659 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
664 BUG_ON(!PageLocked(page
));
665 BUG_ON(mapping
!= page_mapping(page
));
667 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
669 * The non racy check for a busy page.
671 * Must be careful with the order of the tests. When someone has
672 * a ref to the page, it may be possible that they dirty it then
673 * drop the reference. So if PageDirty is tested before page_count
674 * here, then the following race may occur:
676 * get_user_pages(&page);
677 * [user mapping goes away]
679 * !PageDirty(page) [good]
680 * SetPageDirty(page);
682 * !page_count(page) [good, discard it]
684 * [oops, our write_to data is lost]
686 * Reversing the order of the tests ensures such a situation cannot
687 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
688 * load is not satisfied before that of page->_refcount.
690 * Note that if SetPageDirty is always performed via set_page_dirty,
691 * and thus under tree_lock, then this ordering is not required.
693 if (!page_ref_freeze(page
, 2))
695 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
696 if (unlikely(PageDirty(page
))) {
697 page_ref_unfreeze(page
, 2);
701 if (PageSwapCache(page
)) {
702 swp_entry_t swap
= { .val
= page_private(page
) };
703 mem_cgroup_swapout(page
, swap
);
704 __delete_from_swap_cache(page
);
705 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
706 swapcache_free(swap
);
708 void (*freepage
)(struct page
*);
711 freepage
= mapping
->a_ops
->freepage
;
713 * Remember a shadow entry for reclaimed file cache in
714 * order to detect refaults, thus thrashing, later on.
716 * But don't store shadows in an address space that is
717 * already exiting. This is not just an optizimation,
718 * inode reclaim needs to empty out the radix tree or
719 * the nodes are lost. Don't plant shadows behind its
722 * We also don't store shadows for DAX mappings because the
723 * only page cache pages found in these are zero pages
724 * covering holes, and because we don't want to mix DAX
725 * exceptional entries and shadow exceptional entries in the
728 if (reclaimed
&& page_is_file_cache(page
) &&
729 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
730 shadow
= workingset_eviction(mapping
, page
);
731 __delete_from_page_cache(page
, shadow
);
732 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
734 if (freepage
!= NULL
)
741 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
746 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
747 * someone else has a ref on the page, abort and return 0. If it was
748 * successfully detached, return 1. Assumes the caller has a single ref on
751 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
753 if (__remove_mapping(mapping
, page
, false)) {
755 * Unfreezing the refcount with 1 rather than 2 effectively
756 * drops the pagecache ref for us without requiring another
759 page_ref_unfreeze(page
, 1);
766 * putback_lru_page - put previously isolated page onto appropriate LRU list
767 * @page: page to be put back to appropriate lru list
769 * Add previously isolated @page to appropriate LRU list.
770 * Page may still be unevictable for other reasons.
772 * lru_lock must not be held, interrupts must be enabled.
774 void putback_lru_page(struct page
*page
)
777 int was_unevictable
= PageUnevictable(page
);
779 VM_BUG_ON_PAGE(PageLRU(page
), page
);
782 ClearPageUnevictable(page
);
784 if (page_evictable(page
)) {
786 * For evictable pages, we can use the cache.
787 * In event of a race, worst case is we end up with an
788 * unevictable page on [in]active list.
789 * We know how to handle that.
791 is_unevictable
= false;
795 * Put unevictable pages directly on zone's unevictable
798 is_unevictable
= true;
799 add_page_to_unevictable_list(page
);
801 * When racing with an mlock or AS_UNEVICTABLE clearing
802 * (page is unlocked) make sure that if the other thread
803 * does not observe our setting of PG_lru and fails
804 * isolation/check_move_unevictable_pages,
805 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
806 * the page back to the evictable list.
808 * The other side is TestClearPageMlocked() or shmem_lock().
814 * page's status can change while we move it among lru. If an evictable
815 * page is on unevictable list, it never be freed. To avoid that,
816 * check after we added it to the list, again.
818 if (is_unevictable
&& page_evictable(page
)) {
819 if (!isolate_lru_page(page
)) {
823 /* This means someone else dropped this page from LRU
824 * So, it will be freed or putback to LRU again. There is
825 * nothing to do here.
829 if (was_unevictable
&& !is_unevictable
)
830 count_vm_event(UNEVICTABLE_PGRESCUED
);
831 else if (!was_unevictable
&& is_unevictable
)
832 count_vm_event(UNEVICTABLE_PGCULLED
);
834 put_page(page
); /* drop ref from isolate */
837 enum page_references
{
839 PAGEREF_RECLAIM_CLEAN
,
844 static enum page_references
page_check_references(struct page
*page
,
845 struct scan_control
*sc
)
847 int referenced_ptes
, referenced_page
;
848 unsigned long vm_flags
;
850 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
852 referenced_page
= TestClearPageReferenced(page
);
855 * Mlock lost the isolation race with us. Let try_to_unmap()
856 * move the page to the unevictable list.
858 if (vm_flags
& VM_LOCKED
)
859 return PAGEREF_RECLAIM
;
861 if (referenced_ptes
) {
862 if (PageSwapBacked(page
))
863 return PAGEREF_ACTIVATE
;
865 * All mapped pages start out with page table
866 * references from the instantiating fault, so we need
867 * to look twice if a mapped file page is used more
870 * Mark it and spare it for another trip around the
871 * inactive list. Another page table reference will
872 * lead to its activation.
874 * Note: the mark is set for activated pages as well
875 * so that recently deactivated but used pages are
878 SetPageReferenced(page
);
880 if (referenced_page
|| referenced_ptes
> 1)
881 return PAGEREF_ACTIVATE
;
884 * Activate file-backed executable pages after first usage.
886 if (vm_flags
& VM_EXEC
)
887 return PAGEREF_ACTIVATE
;
892 /* Reclaim if clean, defer dirty pages to writeback */
893 if (referenced_page
&& !PageSwapBacked(page
))
894 return PAGEREF_RECLAIM_CLEAN
;
896 return PAGEREF_RECLAIM
;
899 /* Check if a page is dirty or under writeback */
900 static void page_check_dirty_writeback(struct page
*page
,
901 bool *dirty
, bool *writeback
)
903 struct address_space
*mapping
;
906 * Anonymous pages are not handled by flushers and must be written
907 * from reclaim context. Do not stall reclaim based on them
909 if (!page_is_file_cache(page
)) {
915 /* By default assume that the page flags are accurate */
916 *dirty
= PageDirty(page
);
917 *writeback
= PageWriteback(page
);
919 /* Verify dirty/writeback state if the filesystem supports it */
920 if (!page_has_private(page
))
923 mapping
= page_mapping(page
);
924 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
925 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
928 struct reclaim_stat
{
930 unsigned nr_unqueued_dirty
;
931 unsigned nr_congested
;
932 unsigned nr_writeback
;
933 unsigned nr_immediate
;
934 unsigned nr_activate
;
935 unsigned nr_ref_keep
;
936 unsigned nr_unmap_fail
;
940 * shrink_page_list() returns the number of reclaimed pages
942 static unsigned long shrink_page_list(struct list_head
*page_list
,
943 struct pglist_data
*pgdat
,
944 struct scan_control
*sc
,
945 enum ttu_flags ttu_flags
,
946 struct reclaim_stat
*stat
,
949 LIST_HEAD(ret_pages
);
950 LIST_HEAD(free_pages
);
952 unsigned nr_unqueued_dirty
= 0;
953 unsigned nr_dirty
= 0;
954 unsigned nr_congested
= 0;
955 unsigned nr_reclaimed
= 0;
956 unsigned nr_writeback
= 0;
957 unsigned nr_immediate
= 0;
958 unsigned nr_ref_keep
= 0;
959 unsigned nr_unmap_fail
= 0;
963 while (!list_empty(page_list
)) {
964 struct address_space
*mapping
;
967 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
968 bool dirty
, writeback
;
969 int ret
= SWAP_SUCCESS
;
973 page
= lru_to_page(page_list
);
974 list_del(&page
->lru
);
976 if (!trylock_page(page
))
979 VM_BUG_ON_PAGE(PageActive(page
), page
);
983 if (unlikely(!page_evictable(page
)))
986 if (!sc
->may_unmap
&& page_mapped(page
))
989 /* Double the slab pressure for mapped and swapcache pages */
990 if (page_mapped(page
) || PageSwapCache(page
))
993 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
994 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
997 * The number of dirty pages determines if a zone is marked
998 * reclaim_congested which affects wait_iff_congested. kswapd
999 * will stall and start writing pages if the tail of the LRU
1000 * is all dirty unqueued pages.
1002 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1003 if (dirty
|| writeback
)
1006 if (dirty
&& !writeback
)
1007 nr_unqueued_dirty
++;
1010 * Treat this page as congested if the underlying BDI is or if
1011 * pages are cycling through the LRU so quickly that the
1012 * pages marked for immediate reclaim are making it to the
1013 * end of the LRU a second time.
1015 mapping
= page_mapping(page
);
1016 if (((dirty
|| writeback
) && mapping
&&
1017 inode_write_congested(mapping
->host
)) ||
1018 (writeback
&& PageReclaim(page
)))
1022 * If a page at the tail of the LRU is under writeback, there
1023 * are three cases to consider.
1025 * 1) If reclaim is encountering an excessive number of pages
1026 * under writeback and this page is both under writeback and
1027 * PageReclaim then it indicates that pages are being queued
1028 * for IO but are being recycled through the LRU before the
1029 * IO can complete. Waiting on the page itself risks an
1030 * indefinite stall if it is impossible to writeback the
1031 * page due to IO error or disconnected storage so instead
1032 * note that the LRU is being scanned too quickly and the
1033 * caller can stall after page list has been processed.
1035 * 2) Global or new memcg reclaim encounters a page that is
1036 * not marked for immediate reclaim, or the caller does not
1037 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1038 * not to fs). In this case mark the page for immediate
1039 * reclaim and continue scanning.
1041 * Require may_enter_fs because we would wait on fs, which
1042 * may not have submitted IO yet. And the loop driver might
1043 * enter reclaim, and deadlock if it waits on a page for
1044 * which it is needed to do the write (loop masks off
1045 * __GFP_IO|__GFP_FS for this reason); but more thought
1046 * would probably show more reasons.
1048 * 3) Legacy memcg encounters a page that is already marked
1049 * PageReclaim. memcg does not have any dirty pages
1050 * throttling so we could easily OOM just because too many
1051 * pages are in writeback and there is nothing else to
1052 * reclaim. Wait for the writeback to complete.
1054 * In cases 1) and 2) we activate the pages to get them out of
1055 * the way while we continue scanning for clean pages on the
1056 * inactive list and refilling from the active list. The
1057 * observation here is that waiting for disk writes is more
1058 * expensive than potentially causing reloads down the line.
1059 * Since they're marked for immediate reclaim, they won't put
1060 * memory pressure on the cache working set any longer than it
1061 * takes to write them to disk.
1063 if (PageWriteback(page
)) {
1065 if (current_is_kswapd() &&
1066 PageReclaim(page
) &&
1067 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1069 goto activate_locked
;
1072 } else if (sane_reclaim(sc
) ||
1073 !PageReclaim(page
) || !may_enter_fs
) {
1075 * This is slightly racy - end_page_writeback()
1076 * might have just cleared PageReclaim, then
1077 * setting PageReclaim here end up interpreted
1078 * as PageReadahead - but that does not matter
1079 * enough to care. What we do want is for this
1080 * page to have PageReclaim set next time memcg
1081 * reclaim reaches the tests above, so it will
1082 * then wait_on_page_writeback() to avoid OOM;
1083 * and it's also appropriate in global reclaim.
1085 SetPageReclaim(page
);
1087 goto activate_locked
;
1092 wait_on_page_writeback(page
);
1093 /* then go back and try same page again */
1094 list_add_tail(&page
->lru
, page_list
);
1100 references
= page_check_references(page
, sc
);
1102 switch (references
) {
1103 case PAGEREF_ACTIVATE
:
1104 goto activate_locked
;
1108 case PAGEREF_RECLAIM
:
1109 case PAGEREF_RECLAIM_CLEAN
:
1110 ; /* try to reclaim the page below */
1114 * Anonymous process memory has backing store?
1115 * Try to allocate it some swap space here.
1117 if (PageAnon(page
) && !PageSwapCache(page
)) {
1118 if (!(sc
->gfp_mask
& __GFP_IO
))
1120 if (!add_to_swap(page
, page_list
))
1121 goto activate_locked
;
1124 /* Adding to swap updated mapping */
1125 mapping
= page_mapping(page
);
1126 } else if (unlikely(PageTransHuge(page
))) {
1127 /* Split file THP */
1128 if (split_huge_page_to_list(page
, page_list
))
1132 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1135 * The page is mapped into the page tables of one or more
1136 * processes. Try to unmap it here.
1138 if (page_mapped(page
) && mapping
) {
1139 switch (ret
= try_to_unmap(page
,
1140 ttu_flags
| TTU_BATCH_FLUSH
)) {
1143 goto activate_locked
;
1151 ; /* try to free the page below */
1155 if (PageDirty(page
)) {
1157 * Only kswapd can writeback filesystem pages
1158 * to avoid risk of stack overflow. But avoid
1159 * injecting inefficient single-page IO into
1160 * flusher writeback as much as possible: only
1161 * write pages when we've encountered many
1162 * dirty pages, and when we've already scanned
1163 * the rest of the LRU for clean pages and see
1164 * the same dirty pages again (PageReclaim).
1166 if (page_is_file_cache(page
) &&
1167 (!current_is_kswapd() || !PageReclaim(page
) ||
1168 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1170 * Immediately reclaim when written back.
1171 * Similar in principal to deactivate_page()
1172 * except we already have the page isolated
1173 * and know it's dirty
1175 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1176 SetPageReclaim(page
);
1178 goto activate_locked
;
1181 if (references
== PAGEREF_RECLAIM_CLEAN
)
1185 if (!sc
->may_writepage
)
1189 * Page is dirty. Flush the TLB if a writable entry
1190 * potentially exists to avoid CPU writes after IO
1191 * starts and then write it out here.
1193 try_to_unmap_flush_dirty();
1194 switch (pageout(page
, mapping
, sc
)) {
1198 goto activate_locked
;
1200 if (PageWriteback(page
))
1202 if (PageDirty(page
))
1206 * A synchronous write - probably a ramdisk. Go
1207 * ahead and try to reclaim the page.
1209 if (!trylock_page(page
))
1211 if (PageDirty(page
) || PageWriteback(page
))
1213 mapping
= page_mapping(page
);
1215 ; /* try to free the page below */
1220 * If the page has buffers, try to free the buffer mappings
1221 * associated with this page. If we succeed we try to free
1224 * We do this even if the page is PageDirty().
1225 * try_to_release_page() does not perform I/O, but it is
1226 * possible for a page to have PageDirty set, but it is actually
1227 * clean (all its buffers are clean). This happens if the
1228 * buffers were written out directly, with submit_bh(). ext3
1229 * will do this, as well as the blockdev mapping.
1230 * try_to_release_page() will discover that cleanness and will
1231 * drop the buffers and mark the page clean - it can be freed.
1233 * Rarely, pages can have buffers and no ->mapping. These are
1234 * the pages which were not successfully invalidated in
1235 * truncate_complete_page(). We try to drop those buffers here
1236 * and if that worked, and the page is no longer mapped into
1237 * process address space (page_count == 1) it can be freed.
1238 * Otherwise, leave the page on the LRU so it is swappable.
1240 if (page_has_private(page
)) {
1241 if (!try_to_release_page(page
, sc
->gfp_mask
))
1242 goto activate_locked
;
1243 if (!mapping
&& page_count(page
) == 1) {
1245 if (put_page_testzero(page
))
1249 * rare race with speculative reference.
1250 * the speculative reference will free
1251 * this page shortly, so we may
1252 * increment nr_reclaimed here (and
1253 * leave it off the LRU).
1262 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1266 * At this point, we have no other references and there is
1267 * no way to pick any more up (removed from LRU, removed
1268 * from pagecache). Can use non-atomic bitops now (and
1269 * we obviously don't have to worry about waking up a process
1270 * waiting on the page lock, because there are no references.
1272 __ClearPageLocked(page
);
1274 if (ret
== SWAP_LZFREE
)
1275 count_vm_event(PGLAZYFREED
);
1280 * Is there need to periodically free_page_list? It would
1281 * appear not as the counts should be low
1283 list_add(&page
->lru
, &free_pages
);
1287 if (PageSwapCache(page
))
1288 try_to_free_swap(page
);
1290 list_add(&page
->lru
, &ret_pages
);
1294 /* Not a candidate for swapping, so reclaim swap space. */
1295 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1296 try_to_free_swap(page
);
1297 VM_BUG_ON_PAGE(PageActive(page
), page
);
1298 SetPageActive(page
);
1303 list_add(&page
->lru
, &ret_pages
);
1304 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1307 mem_cgroup_uncharge_list(&free_pages
);
1308 try_to_unmap_flush();
1309 free_hot_cold_page_list(&free_pages
, true);
1311 list_splice(&ret_pages
, page_list
);
1312 count_vm_events(PGACTIVATE
, pgactivate
);
1315 stat
->nr_dirty
= nr_dirty
;
1316 stat
->nr_congested
= nr_congested
;
1317 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1318 stat
->nr_writeback
= nr_writeback
;
1319 stat
->nr_immediate
= nr_immediate
;
1320 stat
->nr_activate
= pgactivate
;
1321 stat
->nr_ref_keep
= nr_ref_keep
;
1322 stat
->nr_unmap_fail
= nr_unmap_fail
;
1324 return nr_reclaimed
;
1327 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1328 struct list_head
*page_list
)
1330 struct scan_control sc
= {
1331 .gfp_mask
= GFP_KERNEL
,
1332 .priority
= DEF_PRIORITY
,
1336 struct page
*page
, *next
;
1337 LIST_HEAD(clean_pages
);
1339 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1340 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1341 !__PageMovable(page
)) {
1342 ClearPageActive(page
);
1343 list_move(&page
->lru
, &clean_pages
);
1347 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1348 TTU_IGNORE_ACCESS
, NULL
, true);
1349 list_splice(&clean_pages
, page_list
);
1350 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1355 * Attempt to remove the specified page from its LRU. Only take this page
1356 * if it is of the appropriate PageActive status. Pages which are being
1357 * freed elsewhere are also ignored.
1359 * page: page to consider
1360 * mode: one of the LRU isolation modes defined above
1362 * returns 0 on success, -ve errno on failure.
1364 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1368 /* Only take pages on the LRU. */
1372 /* Compaction should not handle unevictable pages but CMA can do so */
1373 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1379 * To minimise LRU disruption, the caller can indicate that it only
1380 * wants to isolate pages it will be able to operate on without
1381 * blocking - clean pages for the most part.
1383 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1384 * that it is possible to migrate without blocking
1386 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1387 /* All the caller can do on PageWriteback is block */
1388 if (PageWriteback(page
))
1391 if (PageDirty(page
)) {
1392 struct address_space
*mapping
;
1395 * Only pages without mappings or that have a
1396 * ->migratepage callback are possible to migrate
1399 mapping
= page_mapping(page
);
1400 if (mapping
&& !mapping
->a_ops
->migratepage
)
1405 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1408 if (likely(get_page_unless_zero(page
))) {
1410 * Be careful not to clear PageLRU until after we're
1411 * sure the page is not being freed elsewhere -- the
1412 * page release code relies on it.
1423 * Update LRU sizes after isolating pages. The LRU size updates must
1424 * be complete before mem_cgroup_update_lru_size due to a santity check.
1426 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1427 enum lru_list lru
, unsigned long *nr_zone_taken
)
1431 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1432 if (!nr_zone_taken
[zid
])
1435 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1437 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1444 * zone_lru_lock is heavily contended. Some of the functions that
1445 * shrink the lists perform better by taking out a batch of pages
1446 * and working on them outside the LRU lock.
1448 * For pagecache intensive workloads, this function is the hottest
1449 * spot in the kernel (apart from copy_*_user functions).
1451 * Appropriate locks must be held before calling this function.
1453 * @nr_to_scan: The number of pages to look through on the list.
1454 * @lruvec: The LRU vector to pull pages from.
1455 * @dst: The temp list to put pages on to.
1456 * @nr_scanned: The number of pages that were scanned.
1457 * @sc: The scan_control struct for this reclaim session
1458 * @mode: One of the LRU isolation modes
1459 * @lru: LRU list id for isolating
1461 * returns how many pages were moved onto *@dst.
1463 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1464 struct lruvec
*lruvec
, struct list_head
*dst
,
1465 unsigned long *nr_scanned
, struct scan_control
*sc
,
1466 isolate_mode_t mode
, enum lru_list lru
)
1468 struct list_head
*src
= &lruvec
->lists
[lru
];
1469 unsigned long nr_taken
= 0;
1470 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1471 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1472 unsigned long skipped
= 0;
1473 unsigned long scan
, nr_pages
;
1474 LIST_HEAD(pages_skipped
);
1476 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1477 !list_empty(src
); scan
++) {
1480 page
= lru_to_page(src
);
1481 prefetchw_prev_lru_page(page
, src
, flags
);
1483 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1485 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1486 list_move(&page
->lru
, &pages_skipped
);
1487 nr_skipped
[page_zonenum(page
)]++;
1491 switch (__isolate_lru_page(page
, mode
)) {
1493 nr_pages
= hpage_nr_pages(page
);
1494 nr_taken
+= nr_pages
;
1495 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1496 list_move(&page
->lru
, dst
);
1500 /* else it is being freed elsewhere */
1501 list_move(&page
->lru
, src
);
1510 * Splice any skipped pages to the start of the LRU list. Note that
1511 * this disrupts the LRU order when reclaiming for lower zones but
1512 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1513 * scanning would soon rescan the same pages to skip and put the
1514 * system at risk of premature OOM.
1516 if (!list_empty(&pages_skipped
)) {
1519 list_splice(&pages_skipped
, src
);
1520 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1521 if (!nr_skipped
[zid
])
1524 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1525 skipped
+= nr_skipped
[zid
];
1529 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1530 scan
, skipped
, nr_taken
, mode
, lru
);
1531 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1536 * isolate_lru_page - tries to isolate a page from its LRU list
1537 * @page: page to isolate from its LRU list
1539 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1540 * vmstat statistic corresponding to whatever LRU list the page was on.
1542 * Returns 0 if the page was removed from an LRU list.
1543 * Returns -EBUSY if the page was not on an LRU list.
1545 * The returned page will have PageLRU() cleared. If it was found on
1546 * the active list, it will have PageActive set. If it was found on
1547 * the unevictable list, it will have the PageUnevictable bit set. That flag
1548 * may need to be cleared by the caller before letting the page go.
1550 * The vmstat statistic corresponding to the list on which the page was
1551 * found will be decremented.
1554 * (1) Must be called with an elevated refcount on the page. This is a
1555 * fundamentnal difference from isolate_lru_pages (which is called
1556 * without a stable reference).
1557 * (2) the lru_lock must not be held.
1558 * (3) interrupts must be enabled.
1560 int isolate_lru_page(struct page
*page
)
1564 VM_BUG_ON_PAGE(!page_count(page
), page
);
1565 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1567 if (PageLRU(page
)) {
1568 struct zone
*zone
= page_zone(page
);
1569 struct lruvec
*lruvec
;
1571 spin_lock_irq(zone_lru_lock(zone
));
1572 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1573 if (PageLRU(page
)) {
1574 int lru
= page_lru(page
);
1577 del_page_from_lru_list(page
, lruvec
, lru
);
1580 spin_unlock_irq(zone_lru_lock(zone
));
1586 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1587 * then get resheduled. When there are massive number of tasks doing page
1588 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1589 * the LRU list will go small and be scanned faster than necessary, leading to
1590 * unnecessary swapping, thrashing and OOM.
1592 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1593 struct scan_control
*sc
)
1595 unsigned long inactive
, isolated
;
1597 if (current_is_kswapd())
1600 if (!sane_reclaim(sc
))
1604 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1605 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1607 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1608 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1612 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1613 * won't get blocked by normal direct-reclaimers, forming a circular
1616 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1619 return isolated
> inactive
;
1622 static noinline_for_stack
void
1623 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1625 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1626 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1627 LIST_HEAD(pages_to_free
);
1630 * Put back any unfreeable pages.
1632 while (!list_empty(page_list
)) {
1633 struct page
*page
= lru_to_page(page_list
);
1636 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1637 list_del(&page
->lru
);
1638 if (unlikely(!page_evictable(page
))) {
1639 spin_unlock_irq(&pgdat
->lru_lock
);
1640 putback_lru_page(page
);
1641 spin_lock_irq(&pgdat
->lru_lock
);
1645 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1648 lru
= page_lru(page
);
1649 add_page_to_lru_list(page
, lruvec
, lru
);
1651 if (is_active_lru(lru
)) {
1652 int file
= is_file_lru(lru
);
1653 int numpages
= hpage_nr_pages(page
);
1654 reclaim_stat
->recent_rotated
[file
] += numpages
;
1656 if (put_page_testzero(page
)) {
1657 __ClearPageLRU(page
);
1658 __ClearPageActive(page
);
1659 del_page_from_lru_list(page
, lruvec
, lru
);
1661 if (unlikely(PageCompound(page
))) {
1662 spin_unlock_irq(&pgdat
->lru_lock
);
1663 mem_cgroup_uncharge(page
);
1664 (*get_compound_page_dtor(page
))(page
);
1665 spin_lock_irq(&pgdat
->lru_lock
);
1667 list_add(&page
->lru
, &pages_to_free
);
1672 * To save our caller's stack, now use input list for pages to free.
1674 list_splice(&pages_to_free
, page_list
);
1678 * If a kernel thread (such as nfsd for loop-back mounts) services
1679 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1680 * In that case we should only throttle if the backing device it is
1681 * writing to is congested. In other cases it is safe to throttle.
1683 static int current_may_throttle(void)
1685 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1686 current
->backing_dev_info
== NULL
||
1687 bdi_write_congested(current
->backing_dev_info
);
1691 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1692 * of reclaimed pages
1694 static noinline_for_stack
unsigned long
1695 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1696 struct scan_control
*sc
, enum lru_list lru
)
1698 LIST_HEAD(page_list
);
1699 unsigned long nr_scanned
;
1700 unsigned long nr_reclaimed
= 0;
1701 unsigned long nr_taken
;
1702 struct reclaim_stat stat
= {};
1703 isolate_mode_t isolate_mode
= 0;
1704 int file
= is_file_lru(lru
);
1705 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1706 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1708 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1709 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1711 /* We are about to die and free our memory. Return now. */
1712 if (fatal_signal_pending(current
))
1713 return SWAP_CLUSTER_MAX
;
1719 isolate_mode
|= ISOLATE_UNMAPPED
;
1721 spin_lock_irq(&pgdat
->lru_lock
);
1723 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1724 &nr_scanned
, sc
, isolate_mode
, lru
);
1726 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1727 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1729 if (global_reclaim(sc
)) {
1730 if (current_is_kswapd())
1731 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1733 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1735 spin_unlock_irq(&pgdat
->lru_lock
);
1740 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1743 spin_lock_irq(&pgdat
->lru_lock
);
1745 if (global_reclaim(sc
)) {
1746 if (current_is_kswapd())
1747 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1749 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1752 putback_inactive_pages(lruvec
, &page_list
);
1754 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1756 spin_unlock_irq(&pgdat
->lru_lock
);
1758 mem_cgroup_uncharge_list(&page_list
);
1759 free_hot_cold_page_list(&page_list
, true);
1762 * If reclaim is isolating dirty pages under writeback, it implies
1763 * that the long-lived page allocation rate is exceeding the page
1764 * laundering rate. Either the global limits are not being effective
1765 * at throttling processes due to the page distribution throughout
1766 * zones or there is heavy usage of a slow backing device. The
1767 * only option is to throttle from reclaim context which is not ideal
1768 * as there is no guarantee the dirtying process is throttled in the
1769 * same way balance_dirty_pages() manages.
1771 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1772 * of pages under pages flagged for immediate reclaim and stall if any
1773 * are encountered in the nr_immediate check below.
1775 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1776 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1779 * Legacy memcg will stall in page writeback so avoid forcibly
1782 if (sane_reclaim(sc
)) {
1784 * Tag a zone as congested if all the dirty pages scanned were
1785 * backed by a congested BDI and wait_iff_congested will stall.
1787 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1788 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1791 * If dirty pages are scanned that are not queued for IO, it
1792 * implies that flushers are not doing their job. This can
1793 * happen when memory pressure pushes dirty pages to the end of
1794 * the LRU before the dirty limits are breached and the dirty
1795 * data has expired. It can also happen when the proportion of
1796 * dirty pages grows not through writes but through memory
1797 * pressure reclaiming all the clean cache. And in some cases,
1798 * the flushers simply cannot keep up with the allocation
1799 * rate. Nudge the flusher threads in case they are asleep, but
1800 * also allow kswapd to start writing pages during reclaim.
1802 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1803 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1804 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1808 * If kswapd scans pages marked marked for immediate
1809 * reclaim and under writeback (nr_immediate), it implies
1810 * that pages are cycling through the LRU faster than
1811 * they are written so also forcibly stall.
1813 if (stat
.nr_immediate
&& current_may_throttle())
1814 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1818 * Stall direct reclaim for IO completions if underlying BDIs or zone
1819 * is congested. Allow kswapd to continue until it starts encountering
1820 * unqueued dirty pages or cycling through the LRU too quickly.
1822 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1823 current_may_throttle())
1824 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1826 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1827 nr_scanned
, nr_reclaimed
,
1828 stat
.nr_dirty
, stat
.nr_writeback
,
1829 stat
.nr_congested
, stat
.nr_immediate
,
1830 stat
.nr_activate
, stat
.nr_ref_keep
,
1832 sc
->priority
, file
);
1833 return nr_reclaimed
;
1837 * This moves pages from the active list to the inactive list.
1839 * We move them the other way if the page is referenced by one or more
1840 * processes, from rmap.
1842 * If the pages are mostly unmapped, the processing is fast and it is
1843 * appropriate to hold zone_lru_lock across the whole operation. But if
1844 * the pages are mapped, the processing is slow (page_referenced()) so we
1845 * should drop zone_lru_lock around each page. It's impossible to balance
1846 * this, so instead we remove the pages from the LRU while processing them.
1847 * It is safe to rely on PG_active against the non-LRU pages in here because
1848 * nobody will play with that bit on a non-LRU page.
1850 * The downside is that we have to touch page->_refcount against each page.
1851 * But we had to alter page->flags anyway.
1853 * Returns the number of pages moved to the given lru.
1856 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1857 struct list_head
*list
,
1858 struct list_head
*pages_to_free
,
1861 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1866 while (!list_empty(list
)) {
1867 page
= lru_to_page(list
);
1868 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1870 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1873 nr_pages
= hpage_nr_pages(page
);
1874 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1875 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1877 if (put_page_testzero(page
)) {
1878 __ClearPageLRU(page
);
1879 __ClearPageActive(page
);
1880 del_page_from_lru_list(page
, lruvec
, lru
);
1882 if (unlikely(PageCompound(page
))) {
1883 spin_unlock_irq(&pgdat
->lru_lock
);
1884 mem_cgroup_uncharge(page
);
1885 (*get_compound_page_dtor(page
))(page
);
1886 spin_lock_irq(&pgdat
->lru_lock
);
1888 list_add(&page
->lru
, pages_to_free
);
1890 nr_moved
+= nr_pages
;
1894 if (!is_active_lru(lru
))
1895 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1900 static void shrink_active_list(unsigned long nr_to_scan
,
1901 struct lruvec
*lruvec
,
1902 struct scan_control
*sc
,
1905 unsigned long nr_taken
;
1906 unsigned long nr_scanned
;
1907 unsigned long vm_flags
;
1908 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1909 LIST_HEAD(l_active
);
1910 LIST_HEAD(l_inactive
);
1912 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1913 unsigned nr_deactivate
, nr_activate
;
1914 unsigned nr_rotated
= 0;
1915 isolate_mode_t isolate_mode
= 0;
1916 int file
= is_file_lru(lru
);
1917 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1922 isolate_mode
|= ISOLATE_UNMAPPED
;
1924 spin_lock_irq(&pgdat
->lru_lock
);
1926 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1927 &nr_scanned
, sc
, isolate_mode
, lru
);
1929 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1930 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1932 __count_vm_events(PGREFILL
, nr_scanned
);
1934 spin_unlock_irq(&pgdat
->lru_lock
);
1936 while (!list_empty(&l_hold
)) {
1938 page
= lru_to_page(&l_hold
);
1939 list_del(&page
->lru
);
1941 if (unlikely(!page_evictable(page
))) {
1942 putback_lru_page(page
);
1946 if (unlikely(buffer_heads_over_limit
)) {
1947 if (page_has_private(page
) && trylock_page(page
)) {
1948 if (page_has_private(page
))
1949 try_to_release_page(page
, 0);
1954 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1956 nr_rotated
+= hpage_nr_pages(page
);
1958 * Identify referenced, file-backed active pages and
1959 * give them one more trip around the active list. So
1960 * that executable code get better chances to stay in
1961 * memory under moderate memory pressure. Anon pages
1962 * are not likely to be evicted by use-once streaming
1963 * IO, plus JVM can create lots of anon VM_EXEC pages,
1964 * so we ignore them here.
1966 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1967 list_add(&page
->lru
, &l_active
);
1972 ClearPageActive(page
); /* we are de-activating */
1973 list_add(&page
->lru
, &l_inactive
);
1977 * Move pages back to the lru list.
1979 spin_lock_irq(&pgdat
->lru_lock
);
1981 * Count referenced pages from currently used mappings as rotated,
1982 * even though only some of them are actually re-activated. This
1983 * helps balance scan pressure between file and anonymous pages in
1986 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1988 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1989 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1990 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1991 spin_unlock_irq(&pgdat
->lru_lock
);
1993 mem_cgroup_uncharge_list(&l_hold
);
1994 free_hot_cold_page_list(&l_hold
, true);
1995 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
1996 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2000 * The inactive anon list should be small enough that the VM never has
2001 * to do too much work.
2003 * The inactive file list should be small enough to leave most memory
2004 * to the established workingset on the scan-resistant active list,
2005 * but large enough to avoid thrashing the aggregate readahead window.
2007 * Both inactive lists should also be large enough that each inactive
2008 * page has a chance to be referenced again before it is reclaimed.
2010 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2011 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2012 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2015 * memory ratio inactive
2016 * -------------------------------------
2025 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2026 struct scan_control
*sc
, bool trace
)
2028 unsigned long inactive_ratio
;
2029 unsigned long inactive
, active
;
2030 enum lru_list inactive_lru
= file
* LRU_FILE
;
2031 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2035 * If we don't have swap space, anonymous page deactivation
2038 if (!file
&& !total_swap_pages
)
2041 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2042 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2044 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2046 inactive_ratio
= int_sqrt(10 * gb
);
2051 trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec
)->node_id
,
2053 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2054 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2055 inactive_ratio
, file
);
2057 return inactive
* inactive_ratio
< active
;
2060 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2061 struct lruvec
*lruvec
, struct scan_control
*sc
)
2063 if (is_active_lru(lru
)) {
2064 if (inactive_list_is_low(lruvec
, is_file_lru(lru
), sc
, true))
2065 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2069 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2080 * Determine how aggressively the anon and file LRU lists should be
2081 * scanned. The relative value of each set of LRU lists is determined
2082 * by looking at the fraction of the pages scanned we did rotate back
2083 * onto the active list instead of evict.
2085 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2086 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2088 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2089 struct scan_control
*sc
, unsigned long *nr
,
2090 unsigned long *lru_pages
)
2092 int swappiness
= mem_cgroup_swappiness(memcg
);
2093 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2095 u64 denominator
= 0; /* gcc */
2096 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2097 unsigned long anon_prio
, file_prio
;
2098 enum scan_balance scan_balance
;
2099 unsigned long anon
, file
;
2100 unsigned long ap
, fp
;
2103 /* If we have no swap space, do not bother scanning anon pages. */
2104 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2105 scan_balance
= SCAN_FILE
;
2110 * Global reclaim will swap to prevent OOM even with no
2111 * swappiness, but memcg users want to use this knob to
2112 * disable swapping for individual groups completely when
2113 * using the memory controller's swap limit feature would be
2116 if (!global_reclaim(sc
) && !swappiness
) {
2117 scan_balance
= SCAN_FILE
;
2122 * Do not apply any pressure balancing cleverness when the
2123 * system is close to OOM, scan both anon and file equally
2124 * (unless the swappiness setting disagrees with swapping).
2126 if (!sc
->priority
&& swappiness
) {
2127 scan_balance
= SCAN_EQUAL
;
2132 * Prevent the reclaimer from falling into the cache trap: as
2133 * cache pages start out inactive, every cache fault will tip
2134 * the scan balance towards the file LRU. And as the file LRU
2135 * shrinks, so does the window for rotation from references.
2136 * This means we have a runaway feedback loop where a tiny
2137 * thrashing file LRU becomes infinitely more attractive than
2138 * anon pages. Try to detect this based on file LRU size.
2140 if (global_reclaim(sc
)) {
2141 unsigned long pgdatfile
;
2142 unsigned long pgdatfree
;
2144 unsigned long total_high_wmark
= 0;
2146 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2147 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2148 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2150 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2151 struct zone
*zone
= &pgdat
->node_zones
[z
];
2152 if (!managed_zone(zone
))
2155 total_high_wmark
+= high_wmark_pages(zone
);
2158 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2159 scan_balance
= SCAN_ANON
;
2165 * If there is enough inactive page cache, i.e. if the size of the
2166 * inactive list is greater than that of the active list *and* the
2167 * inactive list actually has some pages to scan on this priority, we
2168 * do not reclaim anything from the anonymous working set right now.
2169 * Without the second condition we could end up never scanning an
2170 * lruvec even if it has plenty of old anonymous pages unless the
2171 * system is under heavy pressure.
2173 if (!inactive_list_is_low(lruvec
, true, sc
, false) &&
2174 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2175 scan_balance
= SCAN_FILE
;
2179 scan_balance
= SCAN_FRACT
;
2182 * With swappiness at 100, anonymous and file have the same priority.
2183 * This scanning priority is essentially the inverse of IO cost.
2185 anon_prio
= swappiness
;
2186 file_prio
= 200 - anon_prio
;
2189 * OK, so we have swap space and a fair amount of page cache
2190 * pages. We use the recently rotated / recently scanned
2191 * ratios to determine how valuable each cache is.
2193 * Because workloads change over time (and to avoid overflow)
2194 * we keep these statistics as a floating average, which ends
2195 * up weighing recent references more than old ones.
2197 * anon in [0], file in [1]
2200 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2201 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2202 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2203 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2205 spin_lock_irq(&pgdat
->lru_lock
);
2206 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2207 reclaim_stat
->recent_scanned
[0] /= 2;
2208 reclaim_stat
->recent_rotated
[0] /= 2;
2211 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2212 reclaim_stat
->recent_scanned
[1] /= 2;
2213 reclaim_stat
->recent_rotated
[1] /= 2;
2217 * The amount of pressure on anon vs file pages is inversely
2218 * proportional to the fraction of recently scanned pages on
2219 * each list that were recently referenced and in active use.
2221 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2222 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2224 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2225 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2226 spin_unlock_irq(&pgdat
->lru_lock
);
2230 denominator
= ap
+ fp
+ 1;
2233 for_each_evictable_lru(lru
) {
2234 int file
= is_file_lru(lru
);
2238 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2239 scan
= size
>> sc
->priority
;
2241 * If the cgroup's already been deleted, make sure to
2242 * scrape out the remaining cache.
2244 if (!scan
&& !mem_cgroup_online(memcg
))
2245 scan
= min(size
, SWAP_CLUSTER_MAX
);
2247 switch (scan_balance
) {
2249 /* Scan lists relative to size */
2253 * Scan types proportional to swappiness and
2254 * their relative recent reclaim efficiency.
2256 scan
= div64_u64(scan
* fraction
[file
],
2261 /* Scan one type exclusively */
2262 if ((scan_balance
== SCAN_FILE
) != file
) {
2268 /* Look ma, no brain */
2278 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2280 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2281 struct scan_control
*sc
, unsigned long *lru_pages
)
2283 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2284 unsigned long nr
[NR_LRU_LISTS
];
2285 unsigned long targets
[NR_LRU_LISTS
];
2286 unsigned long nr_to_scan
;
2288 unsigned long nr_reclaimed
= 0;
2289 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2290 struct blk_plug plug
;
2293 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2295 /* Record the original scan target for proportional adjustments later */
2296 memcpy(targets
, nr
, sizeof(nr
));
2299 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2300 * event that can occur when there is little memory pressure e.g.
2301 * multiple streaming readers/writers. Hence, we do not abort scanning
2302 * when the requested number of pages are reclaimed when scanning at
2303 * DEF_PRIORITY on the assumption that the fact we are direct
2304 * reclaiming implies that kswapd is not keeping up and it is best to
2305 * do a batch of work at once. For memcg reclaim one check is made to
2306 * abort proportional reclaim if either the file or anon lru has already
2307 * dropped to zero at the first pass.
2309 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2310 sc
->priority
== DEF_PRIORITY
);
2312 blk_start_plug(&plug
);
2313 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2314 nr
[LRU_INACTIVE_FILE
]) {
2315 unsigned long nr_anon
, nr_file
, percentage
;
2316 unsigned long nr_scanned
;
2318 for_each_evictable_lru(lru
) {
2320 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2321 nr
[lru
] -= nr_to_scan
;
2323 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2330 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2334 * For kswapd and memcg, reclaim at least the number of pages
2335 * requested. Ensure that the anon and file LRUs are scanned
2336 * proportionally what was requested by get_scan_count(). We
2337 * stop reclaiming one LRU and reduce the amount scanning
2338 * proportional to the original scan target.
2340 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2341 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2344 * It's just vindictive to attack the larger once the smaller
2345 * has gone to zero. And given the way we stop scanning the
2346 * smaller below, this makes sure that we only make one nudge
2347 * towards proportionality once we've got nr_to_reclaim.
2349 if (!nr_file
|| !nr_anon
)
2352 if (nr_file
> nr_anon
) {
2353 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2354 targets
[LRU_ACTIVE_ANON
] + 1;
2356 percentage
= nr_anon
* 100 / scan_target
;
2358 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2359 targets
[LRU_ACTIVE_FILE
] + 1;
2361 percentage
= nr_file
* 100 / scan_target
;
2364 /* Stop scanning the smaller of the LRU */
2366 nr
[lru
+ LRU_ACTIVE
] = 0;
2369 * Recalculate the other LRU scan count based on its original
2370 * scan target and the percentage scanning already complete
2372 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2373 nr_scanned
= targets
[lru
] - nr
[lru
];
2374 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2375 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2378 nr_scanned
= targets
[lru
] - nr
[lru
];
2379 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2380 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2382 scan_adjusted
= true;
2384 blk_finish_plug(&plug
);
2385 sc
->nr_reclaimed
+= nr_reclaimed
;
2388 * Even if we did not try to evict anon pages at all, we want to
2389 * rebalance the anon lru active/inactive ratio.
2391 if (inactive_list_is_low(lruvec
, false, sc
, true))
2392 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2393 sc
, LRU_ACTIVE_ANON
);
2396 /* Use reclaim/compaction for costly allocs or under memory pressure */
2397 static bool in_reclaim_compaction(struct scan_control
*sc
)
2399 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2400 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2401 sc
->priority
< DEF_PRIORITY
- 2))
2408 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2409 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2410 * true if more pages should be reclaimed such that when the page allocator
2411 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2412 * It will give up earlier than that if there is difficulty reclaiming pages.
2414 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2415 unsigned long nr_reclaimed
,
2416 unsigned long nr_scanned
,
2417 struct scan_control
*sc
)
2419 unsigned long pages_for_compaction
;
2420 unsigned long inactive_lru_pages
;
2423 /* If not in reclaim/compaction mode, stop */
2424 if (!in_reclaim_compaction(sc
))
2427 /* Consider stopping depending on scan and reclaim activity */
2428 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2430 * For __GFP_REPEAT allocations, stop reclaiming if the
2431 * full LRU list has been scanned and we are still failing
2432 * to reclaim pages. This full LRU scan is potentially
2433 * expensive but a __GFP_REPEAT caller really wants to succeed
2435 if (!nr_reclaimed
&& !nr_scanned
)
2439 * For non-__GFP_REPEAT allocations which can presumably
2440 * fail without consequence, stop if we failed to reclaim
2441 * any pages from the last SWAP_CLUSTER_MAX number of
2442 * pages that were scanned. This will return to the
2443 * caller faster at the risk reclaim/compaction and
2444 * the resulting allocation attempt fails
2451 * If we have not reclaimed enough pages for compaction and the
2452 * inactive lists are large enough, continue reclaiming
2454 pages_for_compaction
= compact_gap(sc
->order
);
2455 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2456 if (get_nr_swap_pages() > 0)
2457 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2458 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2459 inactive_lru_pages
> pages_for_compaction
)
2462 /* If compaction would go ahead or the allocation would succeed, stop */
2463 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2464 struct zone
*zone
= &pgdat
->node_zones
[z
];
2465 if (!managed_zone(zone
))
2468 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2469 case COMPACT_SUCCESS
:
2470 case COMPACT_CONTINUE
:
2473 /* check next zone */
2480 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2482 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2483 unsigned long nr_reclaimed
, nr_scanned
;
2484 bool reclaimable
= false;
2487 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2488 struct mem_cgroup_reclaim_cookie reclaim
= {
2490 .priority
= sc
->priority
,
2492 unsigned long node_lru_pages
= 0;
2493 struct mem_cgroup
*memcg
;
2495 nr_reclaimed
= sc
->nr_reclaimed
;
2496 nr_scanned
= sc
->nr_scanned
;
2498 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2500 unsigned long lru_pages
;
2501 unsigned long reclaimed
;
2502 unsigned long scanned
;
2504 if (mem_cgroup_low(root
, memcg
)) {
2505 if (!sc
->may_thrash
)
2507 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2510 reclaimed
= sc
->nr_reclaimed
;
2511 scanned
= sc
->nr_scanned
;
2513 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2514 node_lru_pages
+= lru_pages
;
2517 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2518 memcg
, sc
->nr_scanned
- scanned
,
2521 /* Record the group's reclaim efficiency */
2522 vmpressure(sc
->gfp_mask
, memcg
, false,
2523 sc
->nr_scanned
- scanned
,
2524 sc
->nr_reclaimed
- reclaimed
);
2527 * Direct reclaim and kswapd have to scan all memory
2528 * cgroups to fulfill the overall scan target for the
2531 * Limit reclaim, on the other hand, only cares about
2532 * nr_to_reclaim pages to be reclaimed and it will
2533 * retry with decreasing priority if one round over the
2534 * whole hierarchy is not sufficient.
2536 if (!global_reclaim(sc
) &&
2537 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2538 mem_cgroup_iter_break(root
, memcg
);
2541 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2544 * Shrink the slab caches in the same proportion that
2545 * the eligible LRU pages were scanned.
2547 if (global_reclaim(sc
))
2548 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2549 sc
->nr_scanned
- nr_scanned
,
2552 if (reclaim_state
) {
2553 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2554 reclaim_state
->reclaimed_slab
= 0;
2557 /* Record the subtree's reclaim efficiency */
2558 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2559 sc
->nr_scanned
- nr_scanned
,
2560 sc
->nr_reclaimed
- nr_reclaimed
);
2562 if (sc
->nr_reclaimed
- nr_reclaimed
)
2565 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2566 sc
->nr_scanned
- nr_scanned
, sc
));
2569 * Kswapd gives up on balancing particular nodes after too
2570 * many failures to reclaim anything from them and goes to
2571 * sleep. On reclaim progress, reset the failure counter. A
2572 * successful direct reclaim run will revive a dormant kswapd.
2575 pgdat
->kswapd_failures
= 0;
2581 * Returns true if compaction should go ahead for a costly-order request, or
2582 * the allocation would already succeed without compaction. Return false if we
2583 * should reclaim first.
2585 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2587 unsigned long watermark
;
2588 enum compact_result suitable
;
2590 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2591 if (suitable
== COMPACT_SUCCESS
)
2592 /* Allocation should succeed already. Don't reclaim. */
2594 if (suitable
== COMPACT_SKIPPED
)
2595 /* Compaction cannot yet proceed. Do reclaim. */
2599 * Compaction is already possible, but it takes time to run and there
2600 * are potentially other callers using the pages just freed. So proceed
2601 * with reclaim to make a buffer of free pages available to give
2602 * compaction a reasonable chance of completing and allocating the page.
2603 * Note that we won't actually reclaim the whole buffer in one attempt
2604 * as the target watermark in should_continue_reclaim() is lower. But if
2605 * we are already above the high+gap watermark, don't reclaim at all.
2607 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2609 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2613 * This is the direct reclaim path, for page-allocating processes. We only
2614 * try to reclaim pages from zones which will satisfy the caller's allocation
2617 * If a zone is deemed to be full of pinned pages then just give it a light
2618 * scan then give up on it.
2620 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2624 unsigned long nr_soft_reclaimed
;
2625 unsigned long nr_soft_scanned
;
2627 pg_data_t
*last_pgdat
= NULL
;
2630 * If the number of buffer_heads in the machine exceeds the maximum
2631 * allowed level, force direct reclaim to scan the highmem zone as
2632 * highmem pages could be pinning lowmem pages storing buffer_heads
2634 orig_mask
= sc
->gfp_mask
;
2635 if (buffer_heads_over_limit
) {
2636 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2637 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2640 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2641 sc
->reclaim_idx
, sc
->nodemask
) {
2643 * Take care memory controller reclaiming has small influence
2646 if (global_reclaim(sc
)) {
2647 if (!cpuset_zone_allowed(zone
,
2648 GFP_KERNEL
| __GFP_HARDWALL
))
2652 * If we already have plenty of memory free for
2653 * compaction in this zone, don't free any more.
2654 * Even though compaction is invoked for any
2655 * non-zero order, only frequent costly order
2656 * reclamation is disruptive enough to become a
2657 * noticeable problem, like transparent huge
2660 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2661 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2662 compaction_ready(zone
, sc
)) {
2663 sc
->compaction_ready
= true;
2668 * Shrink each node in the zonelist once. If the
2669 * zonelist is ordered by zone (not the default) then a
2670 * node may be shrunk multiple times but in that case
2671 * the user prefers lower zones being preserved.
2673 if (zone
->zone_pgdat
== last_pgdat
)
2677 * This steals pages from memory cgroups over softlimit
2678 * and returns the number of reclaimed pages and
2679 * scanned pages. This works for global memory pressure
2680 * and balancing, not for a memcg's limit.
2682 nr_soft_scanned
= 0;
2683 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2684 sc
->order
, sc
->gfp_mask
,
2686 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2687 sc
->nr_scanned
+= nr_soft_scanned
;
2688 /* need some check for avoid more shrink_zone() */
2691 /* See comment about same check for global reclaim above */
2692 if (zone
->zone_pgdat
== last_pgdat
)
2694 last_pgdat
= zone
->zone_pgdat
;
2695 shrink_node(zone
->zone_pgdat
, sc
);
2699 * Restore to original mask to avoid the impact on the caller if we
2700 * promoted it to __GFP_HIGHMEM.
2702 sc
->gfp_mask
= orig_mask
;
2706 * This is the main entry point to direct page reclaim.
2708 * If a full scan of the inactive list fails to free enough memory then we
2709 * are "out of memory" and something needs to be killed.
2711 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2712 * high - the zone may be full of dirty or under-writeback pages, which this
2713 * caller can't do much about. We kick the writeback threads and take explicit
2714 * naps in the hope that some of these pages can be written. But if the
2715 * allocating task holds filesystem locks which prevent writeout this might not
2716 * work, and the allocation attempt will fail.
2718 * returns: 0, if no pages reclaimed
2719 * else, the number of pages reclaimed
2721 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2722 struct scan_control
*sc
)
2724 int initial_priority
= sc
->priority
;
2726 delayacct_freepages_start();
2728 if (global_reclaim(sc
))
2729 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2732 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2735 shrink_zones(zonelist
, sc
);
2737 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2740 if (sc
->compaction_ready
)
2744 * If we're getting trouble reclaiming, start doing
2745 * writepage even in laptop mode.
2747 if (sc
->priority
< DEF_PRIORITY
- 2)
2748 sc
->may_writepage
= 1;
2749 } while (--sc
->priority
>= 0);
2751 delayacct_freepages_end();
2753 if (sc
->nr_reclaimed
)
2754 return sc
->nr_reclaimed
;
2756 /* Aborted reclaim to try compaction? don't OOM, then */
2757 if (sc
->compaction_ready
)
2760 /* Untapped cgroup reserves? Don't OOM, retry. */
2761 if (!sc
->may_thrash
) {
2762 sc
->priority
= initial_priority
;
2770 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2773 unsigned long pfmemalloc_reserve
= 0;
2774 unsigned long free_pages
= 0;
2778 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2781 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2782 zone
= &pgdat
->node_zones
[i
];
2783 if (!managed_zone(zone
))
2786 if (!zone_reclaimable_pages(zone
))
2789 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2790 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2793 /* If there are no reserves (unexpected config) then do not throttle */
2794 if (!pfmemalloc_reserve
)
2797 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2799 /* kswapd must be awake if processes are being throttled */
2800 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2801 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2802 (enum zone_type
)ZONE_NORMAL
);
2803 wake_up_interruptible(&pgdat
->kswapd_wait
);
2810 * Throttle direct reclaimers if backing storage is backed by the network
2811 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2812 * depleted. kswapd will continue to make progress and wake the processes
2813 * when the low watermark is reached.
2815 * Returns true if a fatal signal was delivered during throttling. If this
2816 * happens, the page allocator should not consider triggering the OOM killer.
2818 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2819 nodemask_t
*nodemask
)
2823 pg_data_t
*pgdat
= NULL
;
2826 * Kernel threads should not be throttled as they may be indirectly
2827 * responsible for cleaning pages necessary for reclaim to make forward
2828 * progress. kjournald for example may enter direct reclaim while
2829 * committing a transaction where throttling it could forcing other
2830 * processes to block on log_wait_commit().
2832 if (current
->flags
& PF_KTHREAD
)
2836 * If a fatal signal is pending, this process should not throttle.
2837 * It should return quickly so it can exit and free its memory
2839 if (fatal_signal_pending(current
))
2843 * Check if the pfmemalloc reserves are ok by finding the first node
2844 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2845 * GFP_KERNEL will be required for allocating network buffers when
2846 * swapping over the network so ZONE_HIGHMEM is unusable.
2848 * Throttling is based on the first usable node and throttled processes
2849 * wait on a queue until kswapd makes progress and wakes them. There
2850 * is an affinity then between processes waking up and where reclaim
2851 * progress has been made assuming the process wakes on the same node.
2852 * More importantly, processes running on remote nodes will not compete
2853 * for remote pfmemalloc reserves and processes on different nodes
2854 * should make reasonable progress.
2856 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2857 gfp_zone(gfp_mask
), nodemask
) {
2858 if (zone_idx(zone
) > ZONE_NORMAL
)
2861 /* Throttle based on the first usable node */
2862 pgdat
= zone
->zone_pgdat
;
2863 if (allow_direct_reclaim(pgdat
))
2868 /* If no zone was usable by the allocation flags then do not throttle */
2872 /* Account for the throttling */
2873 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2876 * If the caller cannot enter the filesystem, it's possible that it
2877 * is due to the caller holding an FS lock or performing a journal
2878 * transaction in the case of a filesystem like ext[3|4]. In this case,
2879 * it is not safe to block on pfmemalloc_wait as kswapd could be
2880 * blocked waiting on the same lock. Instead, throttle for up to a
2881 * second before continuing.
2883 if (!(gfp_mask
& __GFP_FS
)) {
2884 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2885 allow_direct_reclaim(pgdat
), HZ
);
2890 /* Throttle until kswapd wakes the process */
2891 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2892 allow_direct_reclaim(pgdat
));
2895 if (fatal_signal_pending(current
))
2902 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2903 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2905 unsigned long nr_reclaimed
;
2906 struct scan_control sc
= {
2907 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2908 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2909 .reclaim_idx
= gfp_zone(gfp_mask
),
2911 .nodemask
= nodemask
,
2912 .priority
= DEF_PRIORITY
,
2913 .may_writepage
= !laptop_mode
,
2919 * Do not enter reclaim if fatal signal was delivered while throttled.
2920 * 1 is returned so that the page allocator does not OOM kill at this
2923 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2926 trace_mm_vmscan_direct_reclaim_begin(order
,
2931 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2933 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2935 return nr_reclaimed
;
2940 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
2941 gfp_t gfp_mask
, bool noswap
,
2943 unsigned long *nr_scanned
)
2945 struct scan_control sc
= {
2946 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2947 .target_mem_cgroup
= memcg
,
2948 .may_writepage
= !laptop_mode
,
2950 .reclaim_idx
= MAX_NR_ZONES
- 1,
2951 .may_swap
= !noswap
,
2953 unsigned long lru_pages
;
2955 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2956 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2958 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2964 * NOTE: Although we can get the priority field, using it
2965 * here is not a good idea, since it limits the pages we can scan.
2966 * if we don't reclaim here, the shrink_node from balance_pgdat
2967 * will pick up pages from other mem cgroup's as well. We hack
2968 * the priority and make it zero.
2970 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
2972 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2974 *nr_scanned
= sc
.nr_scanned
;
2975 return sc
.nr_reclaimed
;
2978 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2979 unsigned long nr_pages
,
2983 struct zonelist
*zonelist
;
2984 unsigned long nr_reclaimed
;
2986 struct scan_control sc
= {
2987 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2988 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2989 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2990 .reclaim_idx
= MAX_NR_ZONES
- 1,
2991 .target_mem_cgroup
= memcg
,
2992 .priority
= DEF_PRIORITY
,
2993 .may_writepage
= !laptop_mode
,
2995 .may_swap
= may_swap
,
2999 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3000 * take care of from where we get pages. So the node where we start the
3001 * scan does not need to be the current node.
3003 nid
= mem_cgroup_select_victim_node(memcg
);
3005 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3007 trace_mm_vmscan_memcg_reclaim_begin(0,
3012 current
->flags
|= PF_MEMALLOC
;
3013 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3014 current
->flags
&= ~PF_MEMALLOC
;
3016 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3018 return nr_reclaimed
;
3022 static void age_active_anon(struct pglist_data
*pgdat
,
3023 struct scan_control
*sc
)
3025 struct mem_cgroup
*memcg
;
3027 if (!total_swap_pages
)
3030 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3032 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3034 if (inactive_list_is_low(lruvec
, false, sc
, true))
3035 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3036 sc
, LRU_ACTIVE_ANON
);
3038 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3042 static bool zone_balanced(struct zone
*zone
, int order
, int classzone_idx
)
3044 unsigned long mark
= high_wmark_pages(zone
);
3046 if (!zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3050 * If any eligible zone is balanced then the node is not considered
3051 * to be congested or dirty
3053 clear_bit(PGDAT_CONGESTED
, &zone
->zone_pgdat
->flags
);
3054 clear_bit(PGDAT_DIRTY
, &zone
->zone_pgdat
->flags
);
3055 clear_bit(PGDAT_WRITEBACK
, &zone
->zone_pgdat
->flags
);
3061 * Prepare kswapd for sleeping. This verifies that there are no processes
3062 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3064 * Returns true if kswapd is ready to sleep
3066 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3071 * The throttled processes are normally woken up in balance_pgdat() as
3072 * soon as allow_direct_reclaim() is true. But there is a potential
3073 * race between when kswapd checks the watermarks and a process gets
3074 * throttled. There is also a potential race if processes get
3075 * throttled, kswapd wakes, a large process exits thereby balancing the
3076 * zones, which causes kswapd to exit balance_pgdat() before reaching
3077 * the wake up checks. If kswapd is going to sleep, no process should
3078 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3079 * the wake up is premature, processes will wake kswapd and get
3080 * throttled again. The difference from wake ups in balance_pgdat() is
3081 * that here we are under prepare_to_wait().
3083 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3084 wake_up_all(&pgdat
->pfmemalloc_wait
);
3086 /* Hopeless node, leave it to direct reclaim */
3087 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3090 for (i
= 0; i
<= classzone_idx
; i
++) {
3091 struct zone
*zone
= pgdat
->node_zones
+ i
;
3093 if (!managed_zone(zone
))
3096 if (!zone_balanced(zone
, order
, classzone_idx
))
3104 * kswapd shrinks a node of pages that are at or below the highest usable
3105 * zone that is currently unbalanced.
3107 * Returns true if kswapd scanned at least the requested number of pages to
3108 * reclaim or if the lack of progress was due to pages under writeback.
3109 * This is used to determine if the scanning priority needs to be raised.
3111 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3112 struct scan_control
*sc
)
3117 /* Reclaim a number of pages proportional to the number of zones */
3118 sc
->nr_to_reclaim
= 0;
3119 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3120 zone
= pgdat
->node_zones
+ z
;
3121 if (!managed_zone(zone
))
3124 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3128 * Historically care was taken to put equal pressure on all zones but
3129 * now pressure is applied based on node LRU order.
3131 shrink_node(pgdat
, sc
);
3134 * Fragmentation may mean that the system cannot be rebalanced for
3135 * high-order allocations. If twice the allocation size has been
3136 * reclaimed then recheck watermarks only at order-0 to prevent
3137 * excessive reclaim. Assume that a process requested a high-order
3138 * can direct reclaim/compact.
3140 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3143 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3147 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3148 * that are eligible for use by the caller until at least one zone is
3151 * Returns the order kswapd finished reclaiming at.
3153 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3154 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3155 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3156 * or lower is eligible for reclaim until at least one usable zone is
3159 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3162 unsigned long nr_soft_reclaimed
;
3163 unsigned long nr_soft_scanned
;
3165 struct scan_control sc
= {
3166 .gfp_mask
= GFP_KERNEL
,
3168 .priority
= DEF_PRIORITY
,
3169 .may_writepage
= !laptop_mode
,
3173 count_vm_event(PAGEOUTRUN
);
3176 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3177 bool raise_priority
= true;
3179 sc
.reclaim_idx
= classzone_idx
;
3182 * If the number of buffer_heads exceeds the maximum allowed
3183 * then consider reclaiming from all zones. This has a dual
3184 * purpose -- on 64-bit systems it is expected that
3185 * buffer_heads are stripped during active rotation. On 32-bit
3186 * systems, highmem pages can pin lowmem memory and shrinking
3187 * buffers can relieve lowmem pressure. Reclaim may still not
3188 * go ahead if all eligible zones for the original allocation
3189 * request are balanced to avoid excessive reclaim from kswapd.
3191 if (buffer_heads_over_limit
) {
3192 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3193 zone
= pgdat
->node_zones
+ i
;
3194 if (!managed_zone(zone
))
3203 * Only reclaim if there are no eligible zones. Check from
3204 * high to low zone as allocations prefer higher zones.
3205 * Scanning from low to high zone would allow congestion to be
3206 * cleared during a very small window when a small low
3207 * zone was balanced even under extreme pressure when the
3208 * overall node may be congested. Note that sc.reclaim_idx
3209 * is not used as buffer_heads_over_limit may have adjusted
3212 for (i
= classzone_idx
; i
>= 0; i
--) {
3213 zone
= pgdat
->node_zones
+ i
;
3214 if (!managed_zone(zone
))
3217 if (zone_balanced(zone
, sc
.order
, classzone_idx
))
3222 * Do some background aging of the anon list, to give
3223 * pages a chance to be referenced before reclaiming. All
3224 * pages are rotated regardless of classzone as this is
3225 * about consistent aging.
3227 age_active_anon(pgdat
, &sc
);
3230 * If we're getting trouble reclaiming, start doing writepage
3231 * even in laptop mode.
3233 if (sc
.priority
< DEF_PRIORITY
- 2)
3234 sc
.may_writepage
= 1;
3236 /* Call soft limit reclaim before calling shrink_node. */
3238 nr_soft_scanned
= 0;
3239 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3240 sc
.gfp_mask
, &nr_soft_scanned
);
3241 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3244 * There should be no need to raise the scanning priority if
3245 * enough pages are already being scanned that that high
3246 * watermark would be met at 100% efficiency.
3248 if (kswapd_shrink_node(pgdat
, &sc
))
3249 raise_priority
= false;
3252 * If the low watermark is met there is no need for processes
3253 * to be throttled on pfmemalloc_wait as they should not be
3254 * able to safely make forward progress. Wake them
3256 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3257 allow_direct_reclaim(pgdat
))
3258 wake_up_all(&pgdat
->pfmemalloc_wait
);
3260 /* Check if kswapd should be suspending */
3261 if (try_to_freeze() || kthread_should_stop())
3265 * Raise priority if scanning rate is too low or there was no
3266 * progress in reclaiming pages
3268 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3269 if (raise_priority
|| !nr_reclaimed
)
3271 } while (sc
.priority
>= 1);
3273 if (!sc
.nr_reclaimed
)
3274 pgdat
->kswapd_failures
++;
3278 * Return the order kswapd stopped reclaiming at as
3279 * prepare_kswapd_sleep() takes it into account. If another caller
3280 * entered the allocator slow path while kswapd was awake, order will
3281 * remain at the higher level.
3286 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3287 unsigned int classzone_idx
)
3292 if (freezing(current
) || kthread_should_stop())
3295 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3297 /* Try to sleep for a short interval */
3298 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3300 * Compaction records what page blocks it recently failed to
3301 * isolate pages from and skips them in the future scanning.
3302 * When kswapd is going to sleep, it is reasonable to assume
3303 * that pages and compaction may succeed so reset the cache.
3305 reset_isolation_suitable(pgdat
);
3308 * We have freed the memory, now we should compact it to make
3309 * allocation of the requested order possible.
3311 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3313 remaining
= schedule_timeout(HZ
/10);
3316 * If woken prematurely then reset kswapd_classzone_idx and
3317 * order. The values will either be from a wakeup request or
3318 * the previous request that slept prematurely.
3321 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3322 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3325 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3326 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3330 * After a short sleep, check if it was a premature sleep. If not, then
3331 * go fully to sleep until explicitly woken up.
3334 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3335 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3338 * vmstat counters are not perfectly accurate and the estimated
3339 * value for counters such as NR_FREE_PAGES can deviate from the
3340 * true value by nr_online_cpus * threshold. To avoid the zone
3341 * watermarks being breached while under pressure, we reduce the
3342 * per-cpu vmstat threshold while kswapd is awake and restore
3343 * them before going back to sleep.
3345 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3347 if (!kthread_should_stop())
3350 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3353 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3355 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3357 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3361 * The background pageout daemon, started as a kernel thread
3362 * from the init process.
3364 * This basically trickles out pages so that we have _some_
3365 * free memory available even if there is no other activity
3366 * that frees anything up. This is needed for things like routing
3367 * etc, where we otherwise might have all activity going on in
3368 * asynchronous contexts that cannot page things out.
3370 * If there are applications that are active memory-allocators
3371 * (most normal use), this basically shouldn't matter.
3373 static int kswapd(void *p
)
3375 unsigned int alloc_order
, reclaim_order
, classzone_idx
;
3376 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3377 struct task_struct
*tsk
= current
;
3379 struct reclaim_state reclaim_state
= {
3380 .reclaimed_slab
= 0,
3382 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3384 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3386 if (!cpumask_empty(cpumask
))
3387 set_cpus_allowed_ptr(tsk
, cpumask
);
3388 current
->reclaim_state
= &reclaim_state
;
3391 * Tell the memory management that we're a "memory allocator",
3392 * and that if we need more memory we should get access to it
3393 * regardless (see "__alloc_pages()"). "kswapd" should
3394 * never get caught in the normal page freeing logic.
3396 * (Kswapd normally doesn't need memory anyway, but sometimes
3397 * you need a small amount of memory in order to be able to
3398 * page out something else, and this flag essentially protects
3399 * us from recursively trying to free more memory as we're
3400 * trying to free the first piece of memory in the first place).
3402 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3405 pgdat
->kswapd_order
= alloc_order
= reclaim_order
= 0;
3406 pgdat
->kswapd_classzone_idx
= classzone_idx
= 0;
3411 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3414 /* Read the new order and classzone_idx */
3415 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3416 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3417 pgdat
->kswapd_order
= 0;
3418 pgdat
->kswapd_classzone_idx
= 0;
3420 ret
= try_to_freeze();
3421 if (kthread_should_stop())
3425 * We can speed up thawing tasks if we don't call balance_pgdat
3426 * after returning from the refrigerator
3432 * Reclaim begins at the requested order but if a high-order
3433 * reclaim fails then kswapd falls back to reclaiming for
3434 * order-0. If that happens, kswapd will consider sleeping
3435 * for the order it finished reclaiming at (reclaim_order)
3436 * but kcompactd is woken to compact for the original
3437 * request (alloc_order).
3439 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3441 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3442 if (reclaim_order
< alloc_order
)
3443 goto kswapd_try_sleep
;
3445 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3446 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3449 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3450 current
->reclaim_state
= NULL
;
3451 lockdep_clear_current_reclaim_state();
3457 * A zone is low on free memory, so wake its kswapd task to service it.
3459 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3464 if (!managed_zone(zone
))
3467 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3469 pgdat
= zone
->zone_pgdat
;
3470 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3471 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3472 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3475 /* Hopeless node, leave it to direct reclaim */
3476 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3479 /* Only wake kswapd if all zones are unbalanced */
3480 for (z
= 0; z
<= classzone_idx
; z
++) {
3481 zone
= pgdat
->node_zones
+ z
;
3482 if (!managed_zone(zone
))
3485 if (zone_balanced(zone
, order
, classzone_idx
))
3489 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3490 wake_up_interruptible(&pgdat
->kswapd_wait
);
3493 #ifdef CONFIG_HIBERNATION
3495 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3498 * Rather than trying to age LRUs the aim is to preserve the overall
3499 * LRU order by reclaiming preferentially
3500 * inactive > active > active referenced > active mapped
3502 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3504 struct reclaim_state reclaim_state
;
3505 struct scan_control sc
= {
3506 .nr_to_reclaim
= nr_to_reclaim
,
3507 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3508 .reclaim_idx
= MAX_NR_ZONES
- 1,
3509 .priority
= DEF_PRIORITY
,
3513 .hibernation_mode
= 1,
3515 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3516 struct task_struct
*p
= current
;
3517 unsigned long nr_reclaimed
;
3519 p
->flags
|= PF_MEMALLOC
;
3520 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3521 reclaim_state
.reclaimed_slab
= 0;
3522 p
->reclaim_state
= &reclaim_state
;
3524 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3526 p
->reclaim_state
= NULL
;
3527 lockdep_clear_current_reclaim_state();
3528 p
->flags
&= ~PF_MEMALLOC
;
3530 return nr_reclaimed
;
3532 #endif /* CONFIG_HIBERNATION */
3534 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3535 not required for correctness. So if the last cpu in a node goes
3536 away, we get changed to run anywhere: as the first one comes back,
3537 restore their cpu bindings. */
3538 static int kswapd_cpu_online(unsigned int cpu
)
3542 for_each_node_state(nid
, N_MEMORY
) {
3543 pg_data_t
*pgdat
= NODE_DATA(nid
);
3544 const struct cpumask
*mask
;
3546 mask
= cpumask_of_node(pgdat
->node_id
);
3548 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3549 /* One of our CPUs online: restore mask */
3550 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3556 * This kswapd start function will be called by init and node-hot-add.
3557 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3559 int kswapd_run(int nid
)
3561 pg_data_t
*pgdat
= NODE_DATA(nid
);
3567 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3568 if (IS_ERR(pgdat
->kswapd
)) {
3569 /* failure at boot is fatal */
3570 BUG_ON(system_state
== SYSTEM_BOOTING
);
3571 pr_err("Failed to start kswapd on node %d\n", nid
);
3572 ret
= PTR_ERR(pgdat
->kswapd
);
3573 pgdat
->kswapd
= NULL
;
3579 * Called by memory hotplug when all memory in a node is offlined. Caller must
3580 * hold mem_hotplug_begin/end().
3582 void kswapd_stop(int nid
)
3584 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3587 kthread_stop(kswapd
);
3588 NODE_DATA(nid
)->kswapd
= NULL
;
3592 static int __init
kswapd_init(void)
3597 for_each_node_state(nid
, N_MEMORY
)
3599 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3600 "mm/vmscan:online", kswapd_cpu_online
,
3606 module_init(kswapd_init
)
3612 * If non-zero call node_reclaim when the number of free pages falls below
3615 int node_reclaim_mode __read_mostly
;
3617 #define RECLAIM_OFF 0
3618 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3619 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3620 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3623 * Priority for NODE_RECLAIM. This determines the fraction of pages
3624 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3627 #define NODE_RECLAIM_PRIORITY 4
3630 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3633 int sysctl_min_unmapped_ratio
= 1;
3636 * If the number of slab pages in a zone grows beyond this percentage then
3637 * slab reclaim needs to occur.
3639 int sysctl_min_slab_ratio
= 5;
3641 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3643 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3644 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3645 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3648 * It's possible for there to be more file mapped pages than
3649 * accounted for by the pages on the file LRU lists because
3650 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3652 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3655 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3656 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3658 unsigned long nr_pagecache_reclaimable
;
3659 unsigned long delta
= 0;
3662 * If RECLAIM_UNMAP is set, then all file pages are considered
3663 * potentially reclaimable. Otherwise, we have to worry about
3664 * pages like swapcache and node_unmapped_file_pages() provides
3667 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3668 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3670 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3672 /* If we can't clean pages, remove dirty pages from consideration */
3673 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3674 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3676 /* Watch for any possible underflows due to delta */
3677 if (unlikely(delta
> nr_pagecache_reclaimable
))
3678 delta
= nr_pagecache_reclaimable
;
3680 return nr_pagecache_reclaimable
- delta
;
3684 * Try to free up some pages from this node through reclaim.
3686 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3688 /* Minimum pages needed in order to stay on node */
3689 const unsigned long nr_pages
= 1 << order
;
3690 struct task_struct
*p
= current
;
3691 struct reclaim_state reclaim_state
;
3692 int classzone_idx
= gfp_zone(gfp_mask
);
3693 struct scan_control sc
= {
3694 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3695 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3697 .priority
= NODE_RECLAIM_PRIORITY
,
3698 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3699 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3701 .reclaim_idx
= classzone_idx
,
3706 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3707 * and we also need to be able to write out pages for RECLAIM_WRITE
3708 * and RECLAIM_UNMAP.
3710 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3711 lockdep_set_current_reclaim_state(gfp_mask
);
3712 reclaim_state
.reclaimed_slab
= 0;
3713 p
->reclaim_state
= &reclaim_state
;
3715 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3717 * Free memory by calling shrink zone with increasing
3718 * priorities until we have enough memory freed.
3721 shrink_node(pgdat
, &sc
);
3722 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3725 p
->reclaim_state
= NULL
;
3726 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3727 lockdep_clear_current_reclaim_state();
3728 return sc
.nr_reclaimed
>= nr_pages
;
3731 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3736 * Node reclaim reclaims unmapped file backed pages and
3737 * slab pages if we are over the defined limits.
3739 * A small portion of unmapped file backed pages is needed for
3740 * file I/O otherwise pages read by file I/O will be immediately
3741 * thrown out if the node is overallocated. So we do not reclaim
3742 * if less than a specified percentage of the node is used by
3743 * unmapped file backed pages.
3745 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3746 sum_zone_node_page_state(pgdat
->node_id
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3747 return NODE_RECLAIM_FULL
;
3750 * Do not scan if the allocation should not be delayed.
3752 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3753 return NODE_RECLAIM_NOSCAN
;
3756 * Only run node reclaim on the local node or on nodes that do not
3757 * have associated processors. This will favor the local processor
3758 * over remote processors and spread off node memory allocations
3759 * as wide as possible.
3761 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3762 return NODE_RECLAIM_NOSCAN
;
3764 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3765 return NODE_RECLAIM_NOSCAN
;
3767 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3768 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3771 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3778 * page_evictable - test whether a page is evictable
3779 * @page: the page to test
3781 * Test whether page is evictable--i.e., should be placed on active/inactive
3782 * lists vs unevictable list.
3784 * Reasons page might not be evictable:
3785 * (1) page's mapping marked unevictable
3786 * (2) page is part of an mlocked VMA
3789 int page_evictable(struct page
*page
)
3791 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3796 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3797 * @pages: array of pages to check
3798 * @nr_pages: number of pages to check
3800 * Checks pages for evictability and moves them to the appropriate lru list.
3802 * This function is only used for SysV IPC SHM_UNLOCK.
3804 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3806 struct lruvec
*lruvec
;
3807 struct pglist_data
*pgdat
= NULL
;
3812 for (i
= 0; i
< nr_pages
; i
++) {
3813 struct page
*page
= pages
[i
];
3814 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3817 if (pagepgdat
!= pgdat
) {
3819 spin_unlock_irq(&pgdat
->lru_lock
);
3821 spin_lock_irq(&pgdat
->lru_lock
);
3823 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
3825 if (!PageLRU(page
) || !PageUnevictable(page
))
3828 if (page_evictable(page
)) {
3829 enum lru_list lru
= page_lru_base_type(page
);
3831 VM_BUG_ON_PAGE(PageActive(page
), page
);
3832 ClearPageUnevictable(page
);
3833 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3834 add_page_to_lru_list(page
, lruvec
, lru
);
3840 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3841 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3842 spin_unlock_irq(&pgdat
->lru_lock
);
3845 #endif /* CONFIG_SHMEM */