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;
101 * Cgroups are not reclaimed below their configured memory.low,
102 * unless we threaten to OOM. If any cgroups are skipped due to
103 * memory.low and nothing was reclaimed, go back for memory.low.
105 unsigned int memcg_low_reclaim
:1;
106 unsigned int memcg_low_skipped
:1;
108 unsigned int hibernation_mode
:1;
110 /* One of the zones is ready for compaction */
111 unsigned int compaction_ready
:1;
113 /* Incremented by the number of inactive pages that were scanned */
114 unsigned long nr_scanned
;
116 /* Number of pages freed so far during a call to shrink_zones() */
117 unsigned long nr_reclaimed
;
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness
= 60;
153 * The total number of pages which are beyond the high watermark within all
156 unsigned long vm_total_pages
;
158 static LIST_HEAD(shrinker_list
);
159 static DECLARE_RWSEM(shrinker_rwsem
);
162 static bool global_reclaim(struct scan_control
*sc
)
164 return !sc
->target_mem_cgroup
;
168 * sane_reclaim - is the usual dirty throttling mechanism operational?
169 * @sc: scan_control in question
171 * The normal page dirty throttling mechanism in balance_dirty_pages() is
172 * completely broken with the legacy memcg and direct stalling in
173 * shrink_page_list() is used for throttling instead, which lacks all the
174 * niceties such as fairness, adaptive pausing, bandwidth proportional
175 * allocation and configurability.
177 * This function tests whether the vmscan currently in progress can assume
178 * that the normal dirty throttling mechanism is operational.
180 static bool sane_reclaim(struct scan_control
*sc
)
182 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
186 #ifdef CONFIG_CGROUP_WRITEBACK
187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
193 static bool global_reclaim(struct scan_control
*sc
)
198 static bool sane_reclaim(struct scan_control
*sc
)
205 * This misses isolated pages which are not accounted for to save counters.
206 * As the data only determines if reclaim or compaction continues, it is
207 * not expected that isolated pages will be a dominating factor.
209 unsigned long zone_reclaimable_pages(struct zone
*zone
)
213 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
214 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
215 if (get_nr_swap_pages() > 0)
216 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
217 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
222 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
226 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
227 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
228 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
230 if (get_nr_swap_pages() > 0)
231 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
232 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
233 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
239 * lruvec_lru_size - Returns the number of pages on the given LRU list.
240 * @lruvec: lru vector
242 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
246 unsigned long lru_size
;
249 if (!mem_cgroup_disabled())
250 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
252 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
254 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
255 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
258 if (!managed_zone(zone
))
261 if (!mem_cgroup_disabled())
262 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
264 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
265 NR_ZONE_LRU_BASE
+ lru
);
266 lru_size
-= min(size
, lru_size
);
274 * Add a shrinker callback to be called from the vm.
276 int register_shrinker(struct shrinker
*shrinker
)
278 size_t size
= sizeof(*shrinker
->nr_deferred
);
280 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
283 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
284 if (!shrinker
->nr_deferred
)
287 down_write(&shrinker_rwsem
);
288 list_add_tail(&shrinker
->list
, &shrinker_list
);
289 up_write(&shrinker_rwsem
);
292 EXPORT_SYMBOL(register_shrinker
);
297 void unregister_shrinker(struct shrinker
*shrinker
)
299 down_write(&shrinker_rwsem
);
300 list_del(&shrinker
->list
);
301 up_write(&shrinker_rwsem
);
302 kfree(shrinker
->nr_deferred
);
304 EXPORT_SYMBOL(unregister_shrinker
);
306 #define SHRINK_BATCH 128
308 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
309 struct shrinker
*shrinker
,
310 unsigned long nr_scanned
,
311 unsigned long nr_eligible
)
313 unsigned long freed
= 0;
314 unsigned long long delta
;
319 int nid
= shrinkctl
->nid
;
320 long batch_size
= shrinker
->batch
? shrinker
->batch
322 long scanned
= 0, next_deferred
;
324 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
333 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
336 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
338 do_div(delta
, nr_eligible
+ 1);
340 if (total_scan
< 0) {
341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342 shrinker
->scan_objects
, total_scan
);
343 total_scan
= freeable
;
346 next_deferred
= total_scan
;
349 * We need to avoid excessive windup on filesystem shrinkers
350 * due to large numbers of GFP_NOFS allocations causing the
351 * shrinkers to return -1 all the time. This results in a large
352 * nr being built up so when a shrink that can do some work
353 * comes along it empties the entire cache due to nr >>>
354 * freeable. This is bad for sustaining a working set in
357 * Hence only allow the shrinker to scan the entire cache when
358 * a large delta change is calculated directly.
360 if (delta
< freeable
/ 4)
361 total_scan
= min(total_scan
, freeable
/ 2);
364 * Avoid risking looping forever due to too large nr value:
365 * never try to free more than twice the estimate number of
368 if (total_scan
> freeable
* 2)
369 total_scan
= freeable
* 2;
371 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
372 nr_scanned
, nr_eligible
,
373 freeable
, delta
, total_scan
);
376 * Normally, we should not scan less than batch_size objects in one
377 * pass to avoid too frequent shrinker calls, but if the slab has less
378 * than batch_size objects in total and we are really tight on memory,
379 * we will try to reclaim all available objects, otherwise we can end
380 * up failing allocations although there are plenty of reclaimable
381 * objects spread over several slabs with usage less than the
384 * We detect the "tight on memory" situations by looking at the total
385 * number of objects we want to scan (total_scan). If it is greater
386 * than the total number of objects on slab (freeable), we must be
387 * scanning at high prio and therefore should try to reclaim as much as
390 while (total_scan
>= batch_size
||
391 total_scan
>= freeable
) {
393 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
395 shrinkctl
->nr_to_scan
= nr_to_scan
;
396 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
397 if (ret
== SHRINK_STOP
)
401 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
402 total_scan
-= nr_to_scan
;
403 scanned
+= nr_to_scan
;
408 if (next_deferred
>= scanned
)
409 next_deferred
-= scanned
;
413 * move the unused scan count back into the shrinker in a
414 * manner that handles concurrent updates. If we exhausted the
415 * scan, there is no need to do an update.
417 if (next_deferred
> 0)
418 new_nr
= atomic_long_add_return(next_deferred
,
419 &shrinker
->nr_deferred
[nid
]);
421 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
423 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
428 * shrink_slab - shrink slab caches
429 * @gfp_mask: allocation context
430 * @nid: node whose slab caches to target
431 * @memcg: memory cgroup whose slab caches to target
432 * @nr_scanned: pressure numerator
433 * @nr_eligible: pressure denominator
435 * Call the shrink functions to age shrinkable caches.
437 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
438 * unaware shrinkers will receive a node id of 0 instead.
440 * @memcg specifies the memory cgroup to target. If it is not NULL,
441 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
442 * objects from the memory cgroup specified. Otherwise, only unaware
443 * shrinkers are called.
445 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
446 * the available objects should be scanned. Page reclaim for example
447 * passes the number of pages scanned and the number of pages on the
448 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
449 * when it encountered mapped pages. The ratio is further biased by
450 * the ->seeks setting of the shrink function, which indicates the
451 * cost to recreate an object relative to that of an LRU page.
453 * Returns the number of reclaimed slab objects.
455 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
456 struct mem_cgroup
*memcg
,
457 unsigned long nr_scanned
,
458 unsigned long nr_eligible
)
460 struct shrinker
*shrinker
;
461 unsigned long freed
= 0;
463 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
467 nr_scanned
= SWAP_CLUSTER_MAX
;
469 if (!down_read_trylock(&shrinker_rwsem
)) {
471 * If we would return 0, our callers would understand that we
472 * have nothing else to shrink and give up trying. By returning
473 * 1 we keep it going and assume we'll be able to shrink next
480 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
481 struct shrink_control sc
= {
482 .gfp_mask
= gfp_mask
,
488 * If kernel memory accounting is disabled, we ignore
489 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
490 * passing NULL for memcg.
492 if (memcg_kmem_enabled() &&
493 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
496 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
499 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
502 up_read(&shrinker_rwsem
);
508 void drop_slab_node(int nid
)
513 struct mem_cgroup
*memcg
= NULL
;
517 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
519 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
520 } while (freed
> 10);
527 for_each_online_node(nid
)
531 static inline int is_page_cache_freeable(struct page
*page
)
534 * A freeable page cache page is referenced only by the caller
535 * that isolated the page, the page cache radix tree and
536 * optional buffer heads at page->private.
538 return page_count(page
) - page_has_private(page
) == 2;
541 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
543 if (current
->flags
& PF_SWAPWRITE
)
545 if (!inode_write_congested(inode
))
547 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
553 * We detected a synchronous write error writing a page out. Probably
554 * -ENOSPC. We need to propagate that into the address_space for a subsequent
555 * fsync(), msync() or close().
557 * The tricky part is that after writepage we cannot touch the mapping: nothing
558 * prevents it from being freed up. But we have a ref on the page and once
559 * that page is locked, the mapping is pinned.
561 * We're allowed to run sleeping lock_page() here because we know the caller has
564 static void handle_write_error(struct address_space
*mapping
,
565 struct page
*page
, int error
)
568 if (page_mapping(page
) == mapping
)
569 mapping_set_error(mapping
, error
);
573 /* possible outcome of pageout() */
575 /* failed to write page out, page is locked */
577 /* move page to the active list, page is locked */
579 /* page has been sent to the disk successfully, page is unlocked */
581 /* page is clean and locked */
586 * pageout is called by shrink_page_list() for each dirty page.
587 * Calls ->writepage().
589 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
590 struct scan_control
*sc
)
593 * If the page is dirty, only perform writeback if that write
594 * will be non-blocking. To prevent this allocation from being
595 * stalled by pagecache activity. But note that there may be
596 * stalls if we need to run get_block(). We could test
597 * PagePrivate for that.
599 * If this process is currently in __generic_file_write_iter() against
600 * this page's queue, we can perform writeback even if that
603 * If the page is swapcache, write it back even if that would
604 * block, for some throttling. This happens by accident, because
605 * swap_backing_dev_info is bust: it doesn't reflect the
606 * congestion state of the swapdevs. Easy to fix, if needed.
608 if (!is_page_cache_freeable(page
))
612 * Some data journaling orphaned pages can have
613 * page->mapping == NULL while being dirty with clean buffers.
615 if (page_has_private(page
)) {
616 if (try_to_free_buffers(page
)) {
617 ClearPageDirty(page
);
618 pr_info("%s: orphaned page\n", __func__
);
624 if (mapping
->a_ops
->writepage
== NULL
)
625 return PAGE_ACTIVATE
;
626 if (!may_write_to_inode(mapping
->host
, sc
))
629 if (clear_page_dirty_for_io(page
)) {
631 struct writeback_control wbc
= {
632 .sync_mode
= WB_SYNC_NONE
,
633 .nr_to_write
= SWAP_CLUSTER_MAX
,
635 .range_end
= LLONG_MAX
,
639 SetPageReclaim(page
);
640 res
= mapping
->a_ops
->writepage(page
, &wbc
);
642 handle_write_error(mapping
, page
, res
);
643 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
644 ClearPageReclaim(page
);
645 return PAGE_ACTIVATE
;
648 if (!PageWriteback(page
)) {
649 /* synchronous write or broken a_ops? */
650 ClearPageReclaim(page
);
652 trace_mm_vmscan_writepage(page
);
653 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
661 * Same as remove_mapping, but if the page is removed from the mapping, it
662 * gets returned with a refcount of 0.
664 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
669 BUG_ON(!PageLocked(page
));
670 BUG_ON(mapping
!= page_mapping(page
));
672 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
674 * The non racy check for a busy page.
676 * Must be careful with the order of the tests. When someone has
677 * a ref to the page, it may be possible that they dirty it then
678 * drop the reference. So if PageDirty is tested before page_count
679 * here, then the following race may occur:
681 * get_user_pages(&page);
682 * [user mapping goes away]
684 * !PageDirty(page) [good]
685 * SetPageDirty(page);
687 * !page_count(page) [good, discard it]
689 * [oops, our write_to data is lost]
691 * Reversing the order of the tests ensures such a situation cannot
692 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
693 * load is not satisfied before that of page->_refcount.
695 * Note that if SetPageDirty is always performed via set_page_dirty,
696 * and thus under tree_lock, then this ordering is not required.
698 if (!page_ref_freeze(page
, 2))
700 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
701 if (unlikely(PageDirty(page
))) {
702 page_ref_unfreeze(page
, 2);
706 if (PageSwapCache(page
)) {
707 swp_entry_t swap
= { .val
= page_private(page
) };
708 mem_cgroup_swapout(page
, swap
);
709 __delete_from_swap_cache(page
);
710 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
711 put_swap_page(page
, swap
);
713 void (*freepage
)(struct page
*);
716 freepage
= mapping
->a_ops
->freepage
;
718 * Remember a shadow entry for reclaimed file cache in
719 * order to detect refaults, thus thrashing, later on.
721 * But don't store shadows in an address space that is
722 * already exiting. This is not just an optizimation,
723 * inode reclaim needs to empty out the radix tree or
724 * the nodes are lost. Don't plant shadows behind its
727 * We also don't store shadows for DAX mappings because the
728 * only page cache pages found in these are zero pages
729 * covering holes, and because we don't want to mix DAX
730 * exceptional entries and shadow exceptional entries in the
733 if (reclaimed
&& page_is_file_cache(page
) &&
734 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
735 shadow
= workingset_eviction(mapping
, page
);
736 __delete_from_page_cache(page
, shadow
);
737 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
739 if (freepage
!= NULL
)
746 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
751 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
752 * someone else has a ref on the page, abort and return 0. If it was
753 * successfully detached, return 1. Assumes the caller has a single ref on
756 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
758 if (__remove_mapping(mapping
, page
, false)) {
760 * Unfreezing the refcount with 1 rather than 2 effectively
761 * drops the pagecache ref for us without requiring another
764 page_ref_unfreeze(page
, 1);
771 * putback_lru_page - put previously isolated page onto appropriate LRU list
772 * @page: page to be put back to appropriate lru list
774 * Add previously isolated @page to appropriate LRU list.
775 * Page may still be unevictable for other reasons.
777 * lru_lock must not be held, interrupts must be enabled.
779 void putback_lru_page(struct page
*page
)
782 int was_unevictable
= PageUnevictable(page
);
784 VM_BUG_ON_PAGE(PageLRU(page
), page
);
787 ClearPageUnevictable(page
);
789 if (page_evictable(page
)) {
791 * For evictable pages, we can use the cache.
792 * In event of a race, worst case is we end up with an
793 * unevictable page on [in]active list.
794 * We know how to handle that.
796 is_unevictable
= false;
800 * Put unevictable pages directly on zone's unevictable
803 is_unevictable
= true;
804 add_page_to_unevictable_list(page
);
806 * When racing with an mlock or AS_UNEVICTABLE clearing
807 * (page is unlocked) make sure that if the other thread
808 * does not observe our setting of PG_lru and fails
809 * isolation/check_move_unevictable_pages,
810 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
811 * the page back to the evictable list.
813 * The other side is TestClearPageMlocked() or shmem_lock().
819 * page's status can change while we move it among lru. If an evictable
820 * page is on unevictable list, it never be freed. To avoid that,
821 * check after we added it to the list, again.
823 if (is_unevictable
&& page_evictable(page
)) {
824 if (!isolate_lru_page(page
)) {
828 /* This means someone else dropped this page from LRU
829 * So, it will be freed or putback to LRU again. There is
830 * nothing to do here.
834 if (was_unevictable
&& !is_unevictable
)
835 count_vm_event(UNEVICTABLE_PGRESCUED
);
836 else if (!was_unevictable
&& is_unevictable
)
837 count_vm_event(UNEVICTABLE_PGCULLED
);
839 put_page(page
); /* drop ref from isolate */
842 enum page_references
{
844 PAGEREF_RECLAIM_CLEAN
,
849 static enum page_references
page_check_references(struct page
*page
,
850 struct scan_control
*sc
)
852 int referenced_ptes
, referenced_page
;
853 unsigned long vm_flags
;
855 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
857 referenced_page
= TestClearPageReferenced(page
);
860 * Mlock lost the isolation race with us. Let try_to_unmap()
861 * move the page to the unevictable list.
863 if (vm_flags
& VM_LOCKED
)
864 return PAGEREF_RECLAIM
;
866 if (referenced_ptes
) {
867 if (PageSwapBacked(page
))
868 return PAGEREF_ACTIVATE
;
870 * All mapped pages start out with page table
871 * references from the instantiating fault, so we need
872 * to look twice if a mapped file page is used more
875 * Mark it and spare it for another trip around the
876 * inactive list. Another page table reference will
877 * lead to its activation.
879 * Note: the mark is set for activated pages as well
880 * so that recently deactivated but used pages are
883 SetPageReferenced(page
);
885 if (referenced_page
|| referenced_ptes
> 1)
886 return PAGEREF_ACTIVATE
;
889 * Activate file-backed executable pages after first usage.
891 if (vm_flags
& VM_EXEC
)
892 return PAGEREF_ACTIVATE
;
897 /* Reclaim if clean, defer dirty pages to writeback */
898 if (referenced_page
&& !PageSwapBacked(page
))
899 return PAGEREF_RECLAIM_CLEAN
;
901 return PAGEREF_RECLAIM
;
904 /* Check if a page is dirty or under writeback */
905 static void page_check_dirty_writeback(struct page
*page
,
906 bool *dirty
, bool *writeback
)
908 struct address_space
*mapping
;
911 * Anonymous pages are not handled by flushers and must be written
912 * from reclaim context. Do not stall reclaim based on them
914 if (!page_is_file_cache(page
) ||
915 (PageAnon(page
) && !PageSwapBacked(page
))) {
921 /* By default assume that the page flags are accurate */
922 *dirty
= PageDirty(page
);
923 *writeback
= PageWriteback(page
);
925 /* Verify dirty/writeback state if the filesystem supports it */
926 if (!page_has_private(page
))
929 mapping
= page_mapping(page
);
930 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
931 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
934 struct reclaim_stat
{
936 unsigned nr_unqueued_dirty
;
937 unsigned nr_congested
;
938 unsigned nr_writeback
;
939 unsigned nr_immediate
;
940 unsigned nr_activate
;
941 unsigned nr_ref_keep
;
942 unsigned nr_unmap_fail
;
946 * shrink_page_list() returns the number of reclaimed pages
948 static unsigned long shrink_page_list(struct list_head
*page_list
,
949 struct pglist_data
*pgdat
,
950 struct scan_control
*sc
,
951 enum ttu_flags ttu_flags
,
952 struct reclaim_stat
*stat
,
955 LIST_HEAD(ret_pages
);
956 LIST_HEAD(free_pages
);
958 unsigned nr_unqueued_dirty
= 0;
959 unsigned nr_dirty
= 0;
960 unsigned nr_congested
= 0;
961 unsigned nr_reclaimed
= 0;
962 unsigned nr_writeback
= 0;
963 unsigned nr_immediate
= 0;
964 unsigned nr_ref_keep
= 0;
965 unsigned nr_unmap_fail
= 0;
969 while (!list_empty(page_list
)) {
970 struct address_space
*mapping
;
973 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
974 bool dirty
, writeback
;
978 page
= lru_to_page(page_list
);
979 list_del(&page
->lru
);
981 if (!trylock_page(page
))
984 VM_BUG_ON_PAGE(PageActive(page
), page
);
988 if (unlikely(!page_evictable(page
)))
989 goto activate_locked
;
991 if (!sc
->may_unmap
&& page_mapped(page
))
994 /* Double the slab pressure for mapped and swapcache pages */
995 if ((page_mapped(page
) || PageSwapCache(page
)) &&
996 !(PageAnon(page
) && !PageSwapBacked(page
)))
999 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1000 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1003 * The number of dirty pages determines if a zone is marked
1004 * reclaim_congested which affects wait_iff_congested. kswapd
1005 * will stall and start writing pages if the tail of the LRU
1006 * is all dirty unqueued pages.
1008 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1009 if (dirty
|| writeback
)
1012 if (dirty
&& !writeback
)
1013 nr_unqueued_dirty
++;
1016 * Treat this page as congested if the underlying BDI is or if
1017 * pages are cycling through the LRU so quickly that the
1018 * pages marked for immediate reclaim are making it to the
1019 * end of the LRU a second time.
1021 mapping
= page_mapping(page
);
1022 if (((dirty
|| writeback
) && mapping
&&
1023 inode_write_congested(mapping
->host
)) ||
1024 (writeback
&& PageReclaim(page
)))
1028 * If a page at the tail of the LRU is under writeback, there
1029 * are three cases to consider.
1031 * 1) If reclaim is encountering an excessive number of pages
1032 * under writeback and this page is both under writeback and
1033 * PageReclaim then it indicates that pages are being queued
1034 * for IO but are being recycled through the LRU before the
1035 * IO can complete. Waiting on the page itself risks an
1036 * indefinite stall if it is impossible to writeback the
1037 * page due to IO error or disconnected storage so instead
1038 * note that the LRU is being scanned too quickly and the
1039 * caller can stall after page list has been processed.
1041 * 2) Global or new memcg reclaim encounters a page that is
1042 * not marked for immediate reclaim, or the caller does not
1043 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1044 * not to fs). In this case mark the page for immediate
1045 * reclaim and continue scanning.
1047 * Require may_enter_fs because we would wait on fs, which
1048 * may not have submitted IO yet. And the loop driver might
1049 * enter reclaim, and deadlock if it waits on a page for
1050 * which it is needed to do the write (loop masks off
1051 * __GFP_IO|__GFP_FS for this reason); but more thought
1052 * would probably show more reasons.
1054 * 3) Legacy memcg encounters a page that is already marked
1055 * PageReclaim. memcg does not have any dirty pages
1056 * throttling so we could easily OOM just because too many
1057 * pages are in writeback and there is nothing else to
1058 * reclaim. Wait for the writeback to complete.
1060 * In cases 1) and 2) we activate the pages to get them out of
1061 * the way while we continue scanning for clean pages on the
1062 * inactive list and refilling from the active list. The
1063 * observation here is that waiting for disk writes is more
1064 * expensive than potentially causing reloads down the line.
1065 * Since they're marked for immediate reclaim, they won't put
1066 * memory pressure on the cache working set any longer than it
1067 * takes to write them to disk.
1069 if (PageWriteback(page
)) {
1071 if (current_is_kswapd() &&
1072 PageReclaim(page
) &&
1073 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1075 goto activate_locked
;
1078 } else if (sane_reclaim(sc
) ||
1079 !PageReclaim(page
) || !may_enter_fs
) {
1081 * This is slightly racy - end_page_writeback()
1082 * might have just cleared PageReclaim, then
1083 * setting PageReclaim here end up interpreted
1084 * as PageReadahead - but that does not matter
1085 * enough to care. What we do want is for this
1086 * page to have PageReclaim set next time memcg
1087 * reclaim reaches the tests above, so it will
1088 * then wait_on_page_writeback() to avoid OOM;
1089 * and it's also appropriate in global reclaim.
1091 SetPageReclaim(page
);
1093 goto activate_locked
;
1098 wait_on_page_writeback(page
);
1099 /* then go back and try same page again */
1100 list_add_tail(&page
->lru
, page_list
);
1106 references
= page_check_references(page
, sc
);
1108 switch (references
) {
1109 case PAGEREF_ACTIVATE
:
1110 goto activate_locked
;
1114 case PAGEREF_RECLAIM
:
1115 case PAGEREF_RECLAIM_CLEAN
:
1116 ; /* try to reclaim the page below */
1120 * Anonymous process memory has backing store?
1121 * Try to allocate it some swap space here.
1122 * Lazyfree page could be freed directly
1124 if (PageAnon(page
) && PageSwapBacked(page
) &&
1125 !PageSwapCache(page
)) {
1126 if (!(sc
->gfp_mask
& __GFP_IO
))
1128 if (PageTransHuge(page
)) {
1129 /* cannot split THP, skip it */
1130 if (!can_split_huge_page(page
, NULL
))
1131 goto activate_locked
;
1133 * Split pages without a PMD map right
1134 * away. Chances are some or all of the
1135 * tail pages can be freed without IO.
1137 if (!compound_mapcount(page
) &&
1138 split_huge_page_to_list(page
, page_list
))
1139 goto activate_locked
;
1141 if (!add_to_swap(page
)) {
1142 if (!PageTransHuge(page
))
1143 goto activate_locked
;
1144 /* Split THP and swap individual base pages */
1145 if (split_huge_page_to_list(page
, page_list
))
1146 goto activate_locked
;
1147 if (!add_to_swap(page
))
1148 goto activate_locked
;
1151 /* XXX: We don't support THP writes */
1152 if (PageTransHuge(page
) &&
1153 split_huge_page_to_list(page
, page_list
)) {
1154 delete_from_swap_cache(page
);
1155 goto activate_locked
;
1160 /* Adding to swap updated mapping */
1161 mapping
= page_mapping(page
);
1162 } else if (unlikely(PageTransHuge(page
))) {
1163 /* Split file THP */
1164 if (split_huge_page_to_list(page
, page_list
))
1168 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1171 * The page is mapped into the page tables of one or more
1172 * processes. Try to unmap it here.
1174 if (page_mapped(page
)) {
1175 if (!try_to_unmap(page
, ttu_flags
| TTU_BATCH_FLUSH
)) {
1177 goto activate_locked
;
1181 if (PageDirty(page
)) {
1183 * Only kswapd can writeback filesystem pages
1184 * to avoid risk of stack overflow. But avoid
1185 * injecting inefficient single-page IO into
1186 * flusher writeback as much as possible: only
1187 * write pages when we've encountered many
1188 * dirty pages, and when we've already scanned
1189 * the rest of the LRU for clean pages and see
1190 * the same dirty pages again (PageReclaim).
1192 if (page_is_file_cache(page
) &&
1193 (!current_is_kswapd() || !PageReclaim(page
) ||
1194 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1196 * Immediately reclaim when written back.
1197 * Similar in principal to deactivate_page()
1198 * except we already have the page isolated
1199 * and know it's dirty
1201 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1202 SetPageReclaim(page
);
1204 goto activate_locked
;
1207 if (references
== PAGEREF_RECLAIM_CLEAN
)
1211 if (!sc
->may_writepage
)
1215 * Page is dirty. Flush the TLB if a writable entry
1216 * potentially exists to avoid CPU writes after IO
1217 * starts and then write it out here.
1219 try_to_unmap_flush_dirty();
1220 switch (pageout(page
, mapping
, sc
)) {
1224 goto activate_locked
;
1226 if (PageWriteback(page
))
1228 if (PageDirty(page
))
1232 * A synchronous write - probably a ramdisk. Go
1233 * ahead and try to reclaim the page.
1235 if (!trylock_page(page
))
1237 if (PageDirty(page
) || PageWriteback(page
))
1239 mapping
= page_mapping(page
);
1241 ; /* try to free the page below */
1246 * If the page has buffers, try to free the buffer mappings
1247 * associated with this page. If we succeed we try to free
1250 * We do this even if the page is PageDirty().
1251 * try_to_release_page() does not perform I/O, but it is
1252 * possible for a page to have PageDirty set, but it is actually
1253 * clean (all its buffers are clean). This happens if the
1254 * buffers were written out directly, with submit_bh(). ext3
1255 * will do this, as well as the blockdev mapping.
1256 * try_to_release_page() will discover that cleanness and will
1257 * drop the buffers and mark the page clean - it can be freed.
1259 * Rarely, pages can have buffers and no ->mapping. These are
1260 * the pages which were not successfully invalidated in
1261 * truncate_complete_page(). We try to drop those buffers here
1262 * and if that worked, and the page is no longer mapped into
1263 * process address space (page_count == 1) it can be freed.
1264 * Otherwise, leave the page on the LRU so it is swappable.
1266 if (page_has_private(page
)) {
1267 if (!try_to_release_page(page
, sc
->gfp_mask
))
1268 goto activate_locked
;
1269 if (!mapping
&& page_count(page
) == 1) {
1271 if (put_page_testzero(page
))
1275 * rare race with speculative reference.
1276 * the speculative reference will free
1277 * this page shortly, so we may
1278 * increment nr_reclaimed here (and
1279 * leave it off the LRU).
1287 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1288 /* follow __remove_mapping for reference */
1289 if (!page_ref_freeze(page
, 1))
1291 if (PageDirty(page
)) {
1292 page_ref_unfreeze(page
, 1);
1296 count_vm_event(PGLAZYFREED
);
1297 count_memcg_page_event(page
, PGLAZYFREED
);
1298 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1301 * At this point, we have no other references and there is
1302 * no way to pick any more up (removed from LRU, removed
1303 * from pagecache). Can use non-atomic bitops now (and
1304 * we obviously don't have to worry about waking up a process
1305 * waiting on the page lock, because there are no references.
1307 __ClearPageLocked(page
);
1312 * Is there need to periodically free_page_list? It would
1313 * appear not as the counts should be low
1315 list_add(&page
->lru
, &free_pages
);
1319 /* Not a candidate for swapping, so reclaim swap space. */
1320 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1322 try_to_free_swap(page
);
1323 VM_BUG_ON_PAGE(PageActive(page
), page
);
1324 if (!PageMlocked(page
)) {
1325 SetPageActive(page
);
1327 count_memcg_page_event(page
, PGACTIVATE
);
1332 list_add(&page
->lru
, &ret_pages
);
1333 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1336 mem_cgroup_uncharge_list(&free_pages
);
1337 try_to_unmap_flush();
1338 free_hot_cold_page_list(&free_pages
, true);
1340 list_splice(&ret_pages
, page_list
);
1341 count_vm_events(PGACTIVATE
, pgactivate
);
1344 stat
->nr_dirty
= nr_dirty
;
1345 stat
->nr_congested
= nr_congested
;
1346 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1347 stat
->nr_writeback
= nr_writeback
;
1348 stat
->nr_immediate
= nr_immediate
;
1349 stat
->nr_activate
= pgactivate
;
1350 stat
->nr_ref_keep
= nr_ref_keep
;
1351 stat
->nr_unmap_fail
= nr_unmap_fail
;
1353 return nr_reclaimed
;
1356 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1357 struct list_head
*page_list
)
1359 struct scan_control sc
= {
1360 .gfp_mask
= GFP_KERNEL
,
1361 .priority
= DEF_PRIORITY
,
1365 struct page
*page
, *next
;
1366 LIST_HEAD(clean_pages
);
1368 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1369 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1370 !__PageMovable(page
)) {
1371 ClearPageActive(page
);
1372 list_move(&page
->lru
, &clean_pages
);
1376 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1377 TTU_IGNORE_ACCESS
, NULL
, true);
1378 list_splice(&clean_pages
, page_list
);
1379 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1384 * Attempt to remove the specified page from its LRU. Only take this page
1385 * if it is of the appropriate PageActive status. Pages which are being
1386 * freed elsewhere are also ignored.
1388 * page: page to consider
1389 * mode: one of the LRU isolation modes defined above
1391 * returns 0 on success, -ve errno on failure.
1393 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1397 /* Only take pages on the LRU. */
1401 /* Compaction should not handle unevictable pages but CMA can do so */
1402 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1408 * To minimise LRU disruption, the caller can indicate that it only
1409 * wants to isolate pages it will be able to operate on without
1410 * blocking - clean pages for the most part.
1412 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1413 * that it is possible to migrate without blocking
1415 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1416 /* All the caller can do on PageWriteback is block */
1417 if (PageWriteback(page
))
1420 if (PageDirty(page
)) {
1421 struct address_space
*mapping
;
1424 * Only pages without mappings or that have a
1425 * ->migratepage callback are possible to migrate
1428 mapping
= page_mapping(page
);
1429 if (mapping
&& !mapping
->a_ops
->migratepage
)
1434 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1437 if (likely(get_page_unless_zero(page
))) {
1439 * Be careful not to clear PageLRU until after we're
1440 * sure the page is not being freed elsewhere -- the
1441 * page release code relies on it.
1452 * Update LRU sizes after isolating pages. The LRU size updates must
1453 * be complete before mem_cgroup_update_lru_size due to a santity check.
1455 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1456 enum lru_list lru
, unsigned long *nr_zone_taken
)
1460 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1461 if (!nr_zone_taken
[zid
])
1464 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1466 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1473 * zone_lru_lock is heavily contended. Some of the functions that
1474 * shrink the lists perform better by taking out a batch of pages
1475 * and working on them outside the LRU lock.
1477 * For pagecache intensive workloads, this function is the hottest
1478 * spot in the kernel (apart from copy_*_user functions).
1480 * Appropriate locks must be held before calling this function.
1482 * @nr_to_scan: The number of eligible pages to look through on the list.
1483 * @lruvec: The LRU vector to pull pages from.
1484 * @dst: The temp list to put pages on to.
1485 * @nr_scanned: The number of pages that were scanned.
1486 * @sc: The scan_control struct for this reclaim session
1487 * @mode: One of the LRU isolation modes
1488 * @lru: LRU list id for isolating
1490 * returns how many pages were moved onto *@dst.
1492 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1493 struct lruvec
*lruvec
, struct list_head
*dst
,
1494 unsigned long *nr_scanned
, struct scan_control
*sc
,
1495 isolate_mode_t mode
, enum lru_list lru
)
1497 struct list_head
*src
= &lruvec
->lists
[lru
];
1498 unsigned long nr_taken
= 0;
1499 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1500 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1501 unsigned long skipped
= 0;
1502 unsigned long scan
, total_scan
, nr_pages
;
1503 LIST_HEAD(pages_skipped
);
1506 for (total_scan
= 0;
1507 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1511 page
= lru_to_page(src
);
1512 prefetchw_prev_lru_page(page
, src
, flags
);
1514 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1516 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1517 list_move(&page
->lru
, &pages_skipped
);
1518 nr_skipped
[page_zonenum(page
)]++;
1523 * Do not count skipped pages because that makes the function
1524 * return with no isolated pages if the LRU mostly contains
1525 * ineligible pages. This causes the VM to not reclaim any
1526 * pages, triggering a premature OOM.
1529 switch (__isolate_lru_page(page
, mode
)) {
1531 nr_pages
= hpage_nr_pages(page
);
1532 nr_taken
+= nr_pages
;
1533 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1534 list_move(&page
->lru
, dst
);
1538 /* else it is being freed elsewhere */
1539 list_move(&page
->lru
, src
);
1548 * Splice any skipped pages to the start of the LRU list. Note that
1549 * this disrupts the LRU order when reclaiming for lower zones but
1550 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1551 * scanning would soon rescan the same pages to skip and put the
1552 * system at risk of premature OOM.
1554 if (!list_empty(&pages_skipped
)) {
1557 list_splice(&pages_skipped
, src
);
1558 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1559 if (!nr_skipped
[zid
])
1562 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1563 skipped
+= nr_skipped
[zid
];
1566 *nr_scanned
= total_scan
;
1567 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1568 total_scan
, skipped
, nr_taken
, mode
, lru
);
1569 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1574 * isolate_lru_page - tries to isolate a page from its LRU list
1575 * @page: page to isolate from its LRU list
1577 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1578 * vmstat statistic corresponding to whatever LRU list the page was on.
1580 * Returns 0 if the page was removed from an LRU list.
1581 * Returns -EBUSY if the page was not on an LRU list.
1583 * The returned page will have PageLRU() cleared. If it was found on
1584 * the active list, it will have PageActive set. If it was found on
1585 * the unevictable list, it will have the PageUnevictable bit set. That flag
1586 * may need to be cleared by the caller before letting the page go.
1588 * The vmstat statistic corresponding to the list on which the page was
1589 * found will be decremented.
1592 * (1) Must be called with an elevated refcount on the page. This is a
1593 * fundamentnal difference from isolate_lru_pages (which is called
1594 * without a stable reference).
1595 * (2) the lru_lock must not be held.
1596 * (3) interrupts must be enabled.
1598 int isolate_lru_page(struct page
*page
)
1602 VM_BUG_ON_PAGE(!page_count(page
), page
);
1603 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1605 if (PageLRU(page
)) {
1606 struct zone
*zone
= page_zone(page
);
1607 struct lruvec
*lruvec
;
1609 spin_lock_irq(zone_lru_lock(zone
));
1610 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1611 if (PageLRU(page
)) {
1612 int lru
= page_lru(page
);
1615 del_page_from_lru_list(page
, lruvec
, lru
);
1618 spin_unlock_irq(zone_lru_lock(zone
));
1624 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1625 * then get resheduled. When there are massive number of tasks doing page
1626 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1627 * the LRU list will go small and be scanned faster than necessary, leading to
1628 * unnecessary swapping, thrashing and OOM.
1630 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1631 struct scan_control
*sc
)
1633 unsigned long inactive
, isolated
;
1635 if (current_is_kswapd())
1638 if (!sane_reclaim(sc
))
1642 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1643 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1645 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1646 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1650 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1651 * won't get blocked by normal direct-reclaimers, forming a circular
1654 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1657 return isolated
> inactive
;
1660 static noinline_for_stack
void
1661 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1663 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1664 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1665 LIST_HEAD(pages_to_free
);
1668 * Put back any unfreeable pages.
1670 while (!list_empty(page_list
)) {
1671 struct page
*page
= lru_to_page(page_list
);
1674 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1675 list_del(&page
->lru
);
1676 if (unlikely(!page_evictable(page
))) {
1677 spin_unlock_irq(&pgdat
->lru_lock
);
1678 putback_lru_page(page
);
1679 spin_lock_irq(&pgdat
->lru_lock
);
1683 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1686 lru
= page_lru(page
);
1687 add_page_to_lru_list(page
, lruvec
, lru
);
1689 if (is_active_lru(lru
)) {
1690 int file
= is_file_lru(lru
);
1691 int numpages
= hpage_nr_pages(page
);
1692 reclaim_stat
->recent_rotated
[file
] += numpages
;
1694 if (put_page_testzero(page
)) {
1695 __ClearPageLRU(page
);
1696 __ClearPageActive(page
);
1697 del_page_from_lru_list(page
, lruvec
, lru
);
1699 if (unlikely(PageCompound(page
))) {
1700 spin_unlock_irq(&pgdat
->lru_lock
);
1701 mem_cgroup_uncharge(page
);
1702 (*get_compound_page_dtor(page
))(page
);
1703 spin_lock_irq(&pgdat
->lru_lock
);
1705 list_add(&page
->lru
, &pages_to_free
);
1710 * To save our caller's stack, now use input list for pages to free.
1712 list_splice(&pages_to_free
, page_list
);
1716 * If a kernel thread (such as nfsd for loop-back mounts) services
1717 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1718 * In that case we should only throttle if the backing device it is
1719 * writing to is congested. In other cases it is safe to throttle.
1721 static int current_may_throttle(void)
1723 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1724 current
->backing_dev_info
== NULL
||
1725 bdi_write_congested(current
->backing_dev_info
);
1729 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1730 * of reclaimed pages
1732 static noinline_for_stack
unsigned long
1733 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1734 struct scan_control
*sc
, enum lru_list lru
)
1736 LIST_HEAD(page_list
);
1737 unsigned long nr_scanned
;
1738 unsigned long nr_reclaimed
= 0;
1739 unsigned long nr_taken
;
1740 struct reclaim_stat stat
= {};
1741 isolate_mode_t isolate_mode
= 0;
1742 int file
= is_file_lru(lru
);
1743 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1744 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1746 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1747 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1749 /* We are about to die and free our memory. Return now. */
1750 if (fatal_signal_pending(current
))
1751 return SWAP_CLUSTER_MAX
;
1757 isolate_mode
|= ISOLATE_UNMAPPED
;
1759 spin_lock_irq(&pgdat
->lru_lock
);
1761 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1762 &nr_scanned
, sc
, isolate_mode
, lru
);
1764 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1765 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1767 if (current_is_kswapd()) {
1768 if (global_reclaim(sc
))
1769 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1770 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1773 if (global_reclaim(sc
))
1774 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1775 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1778 spin_unlock_irq(&pgdat
->lru_lock
);
1783 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1786 spin_lock_irq(&pgdat
->lru_lock
);
1788 if (current_is_kswapd()) {
1789 if (global_reclaim(sc
))
1790 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1791 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1794 if (global_reclaim(sc
))
1795 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1796 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1800 putback_inactive_pages(lruvec
, &page_list
);
1802 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1804 spin_unlock_irq(&pgdat
->lru_lock
);
1806 mem_cgroup_uncharge_list(&page_list
);
1807 free_hot_cold_page_list(&page_list
, true);
1810 * If reclaim is isolating dirty pages under writeback, it implies
1811 * that the long-lived page allocation rate is exceeding the page
1812 * laundering rate. Either the global limits are not being effective
1813 * at throttling processes due to the page distribution throughout
1814 * zones or there is heavy usage of a slow backing device. The
1815 * only option is to throttle from reclaim context which is not ideal
1816 * as there is no guarantee the dirtying process is throttled in the
1817 * same way balance_dirty_pages() manages.
1819 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1820 * of pages under pages flagged for immediate reclaim and stall if any
1821 * are encountered in the nr_immediate check below.
1823 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1824 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1827 * Legacy memcg will stall in page writeback so avoid forcibly
1830 if (sane_reclaim(sc
)) {
1832 * Tag a zone as congested if all the dirty pages scanned were
1833 * backed by a congested BDI and wait_iff_congested will stall.
1835 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1836 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1839 * If dirty pages are scanned that are not queued for IO, it
1840 * implies that flushers are not doing their job. This can
1841 * happen when memory pressure pushes dirty pages to the end of
1842 * the LRU before the dirty limits are breached and the dirty
1843 * data has expired. It can also happen when the proportion of
1844 * dirty pages grows not through writes but through memory
1845 * pressure reclaiming all the clean cache. And in some cases,
1846 * the flushers simply cannot keep up with the allocation
1847 * rate. Nudge the flusher threads in case they are asleep, but
1848 * also allow kswapd to start writing pages during reclaim.
1850 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1851 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1852 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1856 * If kswapd scans pages marked marked for immediate
1857 * reclaim and under writeback (nr_immediate), it implies
1858 * that pages are cycling through the LRU faster than
1859 * they are written so also forcibly stall.
1861 if (stat
.nr_immediate
&& current_may_throttle())
1862 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1866 * Stall direct reclaim for IO completions if underlying BDIs or zone
1867 * is congested. Allow kswapd to continue until it starts encountering
1868 * unqueued dirty pages or cycling through the LRU too quickly.
1870 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1871 current_may_throttle())
1872 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1874 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1875 nr_scanned
, nr_reclaimed
,
1876 stat
.nr_dirty
, stat
.nr_writeback
,
1877 stat
.nr_congested
, stat
.nr_immediate
,
1878 stat
.nr_activate
, stat
.nr_ref_keep
,
1880 sc
->priority
, file
);
1881 return nr_reclaimed
;
1885 * This moves pages from the active list to the inactive list.
1887 * We move them the other way if the page is referenced by one or more
1888 * processes, from rmap.
1890 * If the pages are mostly unmapped, the processing is fast and it is
1891 * appropriate to hold zone_lru_lock across the whole operation. But if
1892 * the pages are mapped, the processing is slow (page_referenced()) so we
1893 * should drop zone_lru_lock around each page. It's impossible to balance
1894 * this, so instead we remove the pages from the LRU while processing them.
1895 * It is safe to rely on PG_active against the non-LRU pages in here because
1896 * nobody will play with that bit on a non-LRU page.
1898 * The downside is that we have to touch page->_refcount against each page.
1899 * But we had to alter page->flags anyway.
1901 * Returns the number of pages moved to the given lru.
1904 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1905 struct list_head
*list
,
1906 struct list_head
*pages_to_free
,
1909 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1914 while (!list_empty(list
)) {
1915 page
= lru_to_page(list
);
1916 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1918 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1921 nr_pages
= hpage_nr_pages(page
);
1922 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1923 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1925 if (put_page_testzero(page
)) {
1926 __ClearPageLRU(page
);
1927 __ClearPageActive(page
);
1928 del_page_from_lru_list(page
, lruvec
, lru
);
1930 if (unlikely(PageCompound(page
))) {
1931 spin_unlock_irq(&pgdat
->lru_lock
);
1932 mem_cgroup_uncharge(page
);
1933 (*get_compound_page_dtor(page
))(page
);
1934 spin_lock_irq(&pgdat
->lru_lock
);
1936 list_add(&page
->lru
, pages_to_free
);
1938 nr_moved
+= nr_pages
;
1942 if (!is_active_lru(lru
)) {
1943 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1944 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
1951 static void shrink_active_list(unsigned long nr_to_scan
,
1952 struct lruvec
*lruvec
,
1953 struct scan_control
*sc
,
1956 unsigned long nr_taken
;
1957 unsigned long nr_scanned
;
1958 unsigned long vm_flags
;
1959 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1960 LIST_HEAD(l_active
);
1961 LIST_HEAD(l_inactive
);
1963 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1964 unsigned nr_deactivate
, nr_activate
;
1965 unsigned nr_rotated
= 0;
1966 isolate_mode_t isolate_mode
= 0;
1967 int file
= is_file_lru(lru
);
1968 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1973 isolate_mode
|= ISOLATE_UNMAPPED
;
1975 spin_lock_irq(&pgdat
->lru_lock
);
1977 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1978 &nr_scanned
, sc
, isolate_mode
, lru
);
1980 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1981 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1983 __count_vm_events(PGREFILL
, nr_scanned
);
1984 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
1986 spin_unlock_irq(&pgdat
->lru_lock
);
1988 while (!list_empty(&l_hold
)) {
1990 page
= lru_to_page(&l_hold
);
1991 list_del(&page
->lru
);
1993 if (unlikely(!page_evictable(page
))) {
1994 putback_lru_page(page
);
1998 if (unlikely(buffer_heads_over_limit
)) {
1999 if (page_has_private(page
) && trylock_page(page
)) {
2000 if (page_has_private(page
))
2001 try_to_release_page(page
, 0);
2006 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2008 nr_rotated
+= hpage_nr_pages(page
);
2010 * Identify referenced, file-backed active pages and
2011 * give them one more trip around the active list. So
2012 * that executable code get better chances to stay in
2013 * memory under moderate memory pressure. Anon pages
2014 * are not likely to be evicted by use-once streaming
2015 * IO, plus JVM can create lots of anon VM_EXEC pages,
2016 * so we ignore them here.
2018 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2019 list_add(&page
->lru
, &l_active
);
2024 ClearPageActive(page
); /* we are de-activating */
2025 list_add(&page
->lru
, &l_inactive
);
2029 * Move pages back to the lru list.
2031 spin_lock_irq(&pgdat
->lru_lock
);
2033 * Count referenced pages from currently used mappings as rotated,
2034 * even though only some of them are actually re-activated. This
2035 * helps balance scan pressure between file and anonymous pages in
2038 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2040 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2041 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2042 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2043 spin_unlock_irq(&pgdat
->lru_lock
);
2045 mem_cgroup_uncharge_list(&l_hold
);
2046 free_hot_cold_page_list(&l_hold
, true);
2047 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2048 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2052 * The inactive anon list should be small enough that the VM never has
2053 * to do too much work.
2055 * The inactive file list should be small enough to leave most memory
2056 * to the established workingset on the scan-resistant active list,
2057 * but large enough to avoid thrashing the aggregate readahead window.
2059 * Both inactive lists should also be large enough that each inactive
2060 * page has a chance to be referenced again before it is reclaimed.
2062 * If that fails and refaulting is observed, the inactive list grows.
2064 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2065 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2066 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2069 * memory ratio inactive
2070 * -------------------------------------
2079 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2080 struct mem_cgroup
*memcg
,
2081 struct scan_control
*sc
, bool actual_reclaim
)
2083 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2084 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2085 enum lru_list inactive_lru
= file
* LRU_FILE
;
2086 unsigned long inactive
, active
;
2087 unsigned long inactive_ratio
;
2088 unsigned long refaults
;
2092 * If we don't have swap space, anonymous page deactivation
2095 if (!file
&& !total_swap_pages
)
2098 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2099 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2102 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2104 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2107 * When refaults are being observed, it means a new workingset
2108 * is being established. Disable active list protection to get
2109 * rid of the stale workingset quickly.
2111 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2114 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2116 inactive_ratio
= int_sqrt(10 * gb
);
2122 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2123 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2124 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2125 inactive_ratio
, file
);
2127 return inactive
* inactive_ratio
< active
;
2130 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2131 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2132 struct scan_control
*sc
)
2134 if (is_active_lru(lru
)) {
2135 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2137 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2141 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2152 * Determine how aggressively the anon and file LRU lists should be
2153 * scanned. The relative value of each set of LRU lists is determined
2154 * by looking at the fraction of the pages scanned we did rotate back
2155 * onto the active list instead of evict.
2157 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2158 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2160 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2161 struct scan_control
*sc
, unsigned long *nr
,
2162 unsigned long *lru_pages
)
2164 int swappiness
= mem_cgroup_swappiness(memcg
);
2165 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2167 u64 denominator
= 0; /* gcc */
2168 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2169 unsigned long anon_prio
, file_prio
;
2170 enum scan_balance scan_balance
;
2171 unsigned long anon
, file
;
2172 unsigned long ap
, fp
;
2175 /* If we have no swap space, do not bother scanning anon pages. */
2176 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2177 scan_balance
= SCAN_FILE
;
2182 * Global reclaim will swap to prevent OOM even with no
2183 * swappiness, but memcg users want to use this knob to
2184 * disable swapping for individual groups completely when
2185 * using the memory controller's swap limit feature would be
2188 if (!global_reclaim(sc
) && !swappiness
) {
2189 scan_balance
= SCAN_FILE
;
2194 * Do not apply any pressure balancing cleverness when the
2195 * system is close to OOM, scan both anon and file equally
2196 * (unless the swappiness setting disagrees with swapping).
2198 if (!sc
->priority
&& swappiness
) {
2199 scan_balance
= SCAN_EQUAL
;
2204 * Prevent the reclaimer from falling into the cache trap: as
2205 * cache pages start out inactive, every cache fault will tip
2206 * the scan balance towards the file LRU. And as the file LRU
2207 * shrinks, so does the window for rotation from references.
2208 * This means we have a runaway feedback loop where a tiny
2209 * thrashing file LRU becomes infinitely more attractive than
2210 * anon pages. Try to detect this based on file LRU size.
2212 if (global_reclaim(sc
)) {
2213 unsigned long pgdatfile
;
2214 unsigned long pgdatfree
;
2216 unsigned long total_high_wmark
= 0;
2218 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2219 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2220 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2222 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2223 struct zone
*zone
= &pgdat
->node_zones
[z
];
2224 if (!managed_zone(zone
))
2227 total_high_wmark
+= high_wmark_pages(zone
);
2230 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2231 scan_balance
= SCAN_ANON
;
2237 * If there is enough inactive page cache, i.e. if the size of the
2238 * inactive list is greater than that of the active list *and* the
2239 * inactive list actually has some pages to scan on this priority, we
2240 * do not reclaim anything from the anonymous working set right now.
2241 * Without the second condition we could end up never scanning an
2242 * lruvec even if it has plenty of old anonymous pages unless the
2243 * system is under heavy pressure.
2245 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2246 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2247 scan_balance
= SCAN_FILE
;
2251 scan_balance
= SCAN_FRACT
;
2254 * With swappiness at 100, anonymous and file have the same priority.
2255 * This scanning priority is essentially the inverse of IO cost.
2257 anon_prio
= swappiness
;
2258 file_prio
= 200 - anon_prio
;
2261 * OK, so we have swap space and a fair amount of page cache
2262 * pages. We use the recently rotated / recently scanned
2263 * ratios to determine how valuable each cache is.
2265 * Because workloads change over time (and to avoid overflow)
2266 * we keep these statistics as a floating average, which ends
2267 * up weighing recent references more than old ones.
2269 * anon in [0], file in [1]
2272 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2273 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2274 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2275 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2277 spin_lock_irq(&pgdat
->lru_lock
);
2278 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2279 reclaim_stat
->recent_scanned
[0] /= 2;
2280 reclaim_stat
->recent_rotated
[0] /= 2;
2283 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2284 reclaim_stat
->recent_scanned
[1] /= 2;
2285 reclaim_stat
->recent_rotated
[1] /= 2;
2289 * The amount of pressure on anon vs file pages is inversely
2290 * proportional to the fraction of recently scanned pages on
2291 * each list that were recently referenced and in active use.
2293 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2294 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2296 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2297 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2298 spin_unlock_irq(&pgdat
->lru_lock
);
2302 denominator
= ap
+ fp
+ 1;
2305 for_each_evictable_lru(lru
) {
2306 int file
= is_file_lru(lru
);
2310 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2311 scan
= size
>> sc
->priority
;
2313 * If the cgroup's already been deleted, make sure to
2314 * scrape out the remaining cache.
2316 if (!scan
&& !mem_cgroup_online(memcg
))
2317 scan
= min(size
, SWAP_CLUSTER_MAX
);
2319 switch (scan_balance
) {
2321 /* Scan lists relative to size */
2325 * Scan types proportional to swappiness and
2326 * their relative recent reclaim efficiency.
2328 scan
= div64_u64(scan
* fraction
[file
],
2333 /* Scan one type exclusively */
2334 if ((scan_balance
== SCAN_FILE
) != file
) {
2340 /* Look ma, no brain */
2350 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2352 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2353 struct scan_control
*sc
, unsigned long *lru_pages
)
2355 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2356 unsigned long nr
[NR_LRU_LISTS
];
2357 unsigned long targets
[NR_LRU_LISTS
];
2358 unsigned long nr_to_scan
;
2360 unsigned long nr_reclaimed
= 0;
2361 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2362 struct blk_plug plug
;
2365 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2367 /* Record the original scan target for proportional adjustments later */
2368 memcpy(targets
, nr
, sizeof(nr
));
2371 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2372 * event that can occur when there is little memory pressure e.g.
2373 * multiple streaming readers/writers. Hence, we do not abort scanning
2374 * when the requested number of pages are reclaimed when scanning at
2375 * DEF_PRIORITY on the assumption that the fact we are direct
2376 * reclaiming implies that kswapd is not keeping up and it is best to
2377 * do a batch of work at once. For memcg reclaim one check is made to
2378 * abort proportional reclaim if either the file or anon lru has already
2379 * dropped to zero at the first pass.
2381 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2382 sc
->priority
== DEF_PRIORITY
);
2384 blk_start_plug(&plug
);
2385 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2386 nr
[LRU_INACTIVE_FILE
]) {
2387 unsigned long nr_anon
, nr_file
, percentage
;
2388 unsigned long nr_scanned
;
2390 for_each_evictable_lru(lru
) {
2392 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2393 nr
[lru
] -= nr_to_scan
;
2395 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2402 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2406 * For kswapd and memcg, reclaim at least the number of pages
2407 * requested. Ensure that the anon and file LRUs are scanned
2408 * proportionally what was requested by get_scan_count(). We
2409 * stop reclaiming one LRU and reduce the amount scanning
2410 * proportional to the original scan target.
2412 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2413 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2416 * It's just vindictive to attack the larger once the smaller
2417 * has gone to zero. And given the way we stop scanning the
2418 * smaller below, this makes sure that we only make one nudge
2419 * towards proportionality once we've got nr_to_reclaim.
2421 if (!nr_file
|| !nr_anon
)
2424 if (nr_file
> nr_anon
) {
2425 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2426 targets
[LRU_ACTIVE_ANON
] + 1;
2428 percentage
= nr_anon
* 100 / scan_target
;
2430 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2431 targets
[LRU_ACTIVE_FILE
] + 1;
2433 percentage
= nr_file
* 100 / scan_target
;
2436 /* Stop scanning the smaller of the LRU */
2438 nr
[lru
+ LRU_ACTIVE
] = 0;
2441 * Recalculate the other LRU scan count based on its original
2442 * scan target and the percentage scanning already complete
2444 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2445 nr_scanned
= targets
[lru
] - nr
[lru
];
2446 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2447 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2450 nr_scanned
= targets
[lru
] - nr
[lru
];
2451 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2452 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2454 scan_adjusted
= true;
2456 blk_finish_plug(&plug
);
2457 sc
->nr_reclaimed
+= nr_reclaimed
;
2460 * Even if we did not try to evict anon pages at all, we want to
2461 * rebalance the anon lru active/inactive ratio.
2463 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2464 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2465 sc
, LRU_ACTIVE_ANON
);
2468 /* Use reclaim/compaction for costly allocs or under memory pressure */
2469 static bool in_reclaim_compaction(struct scan_control
*sc
)
2471 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2472 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2473 sc
->priority
< DEF_PRIORITY
- 2))
2480 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2481 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2482 * true if more pages should be reclaimed such that when the page allocator
2483 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2484 * It will give up earlier than that if there is difficulty reclaiming pages.
2486 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2487 unsigned long nr_reclaimed
,
2488 unsigned long nr_scanned
,
2489 struct scan_control
*sc
)
2491 unsigned long pages_for_compaction
;
2492 unsigned long inactive_lru_pages
;
2495 /* If not in reclaim/compaction mode, stop */
2496 if (!in_reclaim_compaction(sc
))
2499 /* Consider stopping depending on scan and reclaim activity */
2500 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2502 * For __GFP_REPEAT allocations, stop reclaiming if the
2503 * full LRU list has been scanned and we are still failing
2504 * to reclaim pages. This full LRU scan is potentially
2505 * expensive but a __GFP_REPEAT caller really wants to succeed
2507 if (!nr_reclaimed
&& !nr_scanned
)
2511 * For non-__GFP_REPEAT allocations which can presumably
2512 * fail without consequence, stop if we failed to reclaim
2513 * any pages from the last SWAP_CLUSTER_MAX number of
2514 * pages that were scanned. This will return to the
2515 * caller faster at the risk reclaim/compaction and
2516 * the resulting allocation attempt fails
2523 * If we have not reclaimed enough pages for compaction and the
2524 * inactive lists are large enough, continue reclaiming
2526 pages_for_compaction
= compact_gap(sc
->order
);
2527 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2528 if (get_nr_swap_pages() > 0)
2529 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2530 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2531 inactive_lru_pages
> pages_for_compaction
)
2534 /* If compaction would go ahead or the allocation would succeed, stop */
2535 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2536 struct zone
*zone
= &pgdat
->node_zones
[z
];
2537 if (!managed_zone(zone
))
2540 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2541 case COMPACT_SUCCESS
:
2542 case COMPACT_CONTINUE
:
2545 /* check next zone */
2552 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2554 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2555 unsigned long nr_reclaimed
, nr_scanned
;
2556 bool reclaimable
= false;
2559 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2560 struct mem_cgroup_reclaim_cookie reclaim
= {
2562 .priority
= sc
->priority
,
2564 unsigned long node_lru_pages
= 0;
2565 struct mem_cgroup
*memcg
;
2567 nr_reclaimed
= sc
->nr_reclaimed
;
2568 nr_scanned
= sc
->nr_scanned
;
2570 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2572 unsigned long lru_pages
;
2573 unsigned long reclaimed
;
2574 unsigned long scanned
;
2576 if (mem_cgroup_low(root
, memcg
)) {
2577 if (!sc
->memcg_low_reclaim
) {
2578 sc
->memcg_low_skipped
= 1;
2581 mem_cgroup_event(memcg
, MEMCG_LOW
);
2584 reclaimed
= sc
->nr_reclaimed
;
2585 scanned
= sc
->nr_scanned
;
2587 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2588 node_lru_pages
+= lru_pages
;
2591 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2592 memcg
, sc
->nr_scanned
- scanned
,
2595 /* Record the group's reclaim efficiency */
2596 vmpressure(sc
->gfp_mask
, memcg
, false,
2597 sc
->nr_scanned
- scanned
,
2598 sc
->nr_reclaimed
- reclaimed
);
2601 * Direct reclaim and kswapd have to scan all memory
2602 * cgroups to fulfill the overall scan target for the
2605 * Limit reclaim, on the other hand, only cares about
2606 * nr_to_reclaim pages to be reclaimed and it will
2607 * retry with decreasing priority if one round over the
2608 * whole hierarchy is not sufficient.
2610 if (!global_reclaim(sc
) &&
2611 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2612 mem_cgroup_iter_break(root
, memcg
);
2615 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2618 * Shrink the slab caches in the same proportion that
2619 * the eligible LRU pages were scanned.
2621 if (global_reclaim(sc
))
2622 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2623 sc
->nr_scanned
- nr_scanned
,
2626 if (reclaim_state
) {
2627 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2628 reclaim_state
->reclaimed_slab
= 0;
2631 /* Record the subtree's reclaim efficiency */
2632 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2633 sc
->nr_scanned
- nr_scanned
,
2634 sc
->nr_reclaimed
- nr_reclaimed
);
2636 if (sc
->nr_reclaimed
- nr_reclaimed
)
2639 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2640 sc
->nr_scanned
- nr_scanned
, sc
));
2643 * Kswapd gives up on balancing particular nodes after too
2644 * many failures to reclaim anything from them and goes to
2645 * sleep. On reclaim progress, reset the failure counter. A
2646 * successful direct reclaim run will revive a dormant kswapd.
2649 pgdat
->kswapd_failures
= 0;
2655 * Returns true if compaction should go ahead for a costly-order request, or
2656 * the allocation would already succeed without compaction. Return false if we
2657 * should reclaim first.
2659 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2661 unsigned long watermark
;
2662 enum compact_result suitable
;
2664 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2665 if (suitable
== COMPACT_SUCCESS
)
2666 /* Allocation should succeed already. Don't reclaim. */
2668 if (suitable
== COMPACT_SKIPPED
)
2669 /* Compaction cannot yet proceed. Do reclaim. */
2673 * Compaction is already possible, but it takes time to run and there
2674 * are potentially other callers using the pages just freed. So proceed
2675 * with reclaim to make a buffer of free pages available to give
2676 * compaction a reasonable chance of completing and allocating the page.
2677 * Note that we won't actually reclaim the whole buffer in one attempt
2678 * as the target watermark in should_continue_reclaim() is lower. But if
2679 * we are already above the high+gap watermark, don't reclaim at all.
2681 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2683 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2687 * This is the direct reclaim path, for page-allocating processes. We only
2688 * try to reclaim pages from zones which will satisfy the caller's allocation
2691 * If a zone is deemed to be full of pinned pages then just give it a light
2692 * scan then give up on it.
2694 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2698 unsigned long nr_soft_reclaimed
;
2699 unsigned long nr_soft_scanned
;
2701 pg_data_t
*last_pgdat
= NULL
;
2704 * If the number of buffer_heads in the machine exceeds the maximum
2705 * allowed level, force direct reclaim to scan the highmem zone as
2706 * highmem pages could be pinning lowmem pages storing buffer_heads
2708 orig_mask
= sc
->gfp_mask
;
2709 if (buffer_heads_over_limit
) {
2710 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2711 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2714 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2715 sc
->reclaim_idx
, sc
->nodemask
) {
2717 * Take care memory controller reclaiming has small influence
2720 if (global_reclaim(sc
)) {
2721 if (!cpuset_zone_allowed(zone
,
2722 GFP_KERNEL
| __GFP_HARDWALL
))
2726 * If we already have plenty of memory free for
2727 * compaction in this zone, don't free any more.
2728 * Even though compaction is invoked for any
2729 * non-zero order, only frequent costly order
2730 * reclamation is disruptive enough to become a
2731 * noticeable problem, like transparent huge
2734 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2735 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2736 compaction_ready(zone
, sc
)) {
2737 sc
->compaction_ready
= true;
2742 * Shrink each node in the zonelist once. If the
2743 * zonelist is ordered by zone (not the default) then a
2744 * node may be shrunk multiple times but in that case
2745 * the user prefers lower zones being preserved.
2747 if (zone
->zone_pgdat
== last_pgdat
)
2751 * This steals pages from memory cgroups over softlimit
2752 * and returns the number of reclaimed pages and
2753 * scanned pages. This works for global memory pressure
2754 * and balancing, not for a memcg's limit.
2756 nr_soft_scanned
= 0;
2757 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2758 sc
->order
, sc
->gfp_mask
,
2760 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2761 sc
->nr_scanned
+= nr_soft_scanned
;
2762 /* need some check for avoid more shrink_zone() */
2765 /* See comment about same check for global reclaim above */
2766 if (zone
->zone_pgdat
== last_pgdat
)
2768 last_pgdat
= zone
->zone_pgdat
;
2769 shrink_node(zone
->zone_pgdat
, sc
);
2773 * Restore to original mask to avoid the impact on the caller if we
2774 * promoted it to __GFP_HIGHMEM.
2776 sc
->gfp_mask
= orig_mask
;
2779 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2781 struct mem_cgroup
*memcg
;
2783 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2785 unsigned long refaults
;
2786 struct lruvec
*lruvec
;
2789 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2791 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2793 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2794 lruvec
->refaults
= refaults
;
2795 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2799 * This is the main entry point to direct page reclaim.
2801 * If a full scan of the inactive list fails to free enough memory then we
2802 * are "out of memory" and something needs to be killed.
2804 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2805 * high - the zone may be full of dirty or under-writeback pages, which this
2806 * caller can't do much about. We kick the writeback threads and take explicit
2807 * naps in the hope that some of these pages can be written. But if the
2808 * allocating task holds filesystem locks which prevent writeout this might not
2809 * work, and the allocation attempt will fail.
2811 * returns: 0, if no pages reclaimed
2812 * else, the number of pages reclaimed
2814 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2815 struct scan_control
*sc
)
2817 int initial_priority
= sc
->priority
;
2818 pg_data_t
*last_pgdat
;
2822 delayacct_freepages_start();
2824 if (global_reclaim(sc
))
2825 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2828 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2831 shrink_zones(zonelist
, sc
);
2833 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2836 if (sc
->compaction_ready
)
2840 * If we're getting trouble reclaiming, start doing
2841 * writepage even in laptop mode.
2843 if (sc
->priority
< DEF_PRIORITY
- 2)
2844 sc
->may_writepage
= 1;
2845 } while (--sc
->priority
>= 0);
2848 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2850 if (zone
->zone_pgdat
== last_pgdat
)
2852 last_pgdat
= zone
->zone_pgdat
;
2853 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2856 delayacct_freepages_end();
2858 if (sc
->nr_reclaimed
)
2859 return sc
->nr_reclaimed
;
2861 /* Aborted reclaim to try compaction? don't OOM, then */
2862 if (sc
->compaction_ready
)
2865 /* Untapped cgroup reserves? Don't OOM, retry. */
2866 if (sc
->memcg_low_skipped
) {
2867 sc
->priority
= initial_priority
;
2868 sc
->memcg_low_reclaim
= 1;
2869 sc
->memcg_low_skipped
= 0;
2876 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2879 unsigned long pfmemalloc_reserve
= 0;
2880 unsigned long free_pages
= 0;
2884 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2887 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2888 zone
= &pgdat
->node_zones
[i
];
2889 if (!managed_zone(zone
))
2892 if (!zone_reclaimable_pages(zone
))
2895 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2896 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2899 /* If there are no reserves (unexpected config) then do not throttle */
2900 if (!pfmemalloc_reserve
)
2903 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2905 /* kswapd must be awake if processes are being throttled */
2906 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2907 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2908 (enum zone_type
)ZONE_NORMAL
);
2909 wake_up_interruptible(&pgdat
->kswapd_wait
);
2916 * Throttle direct reclaimers if backing storage is backed by the network
2917 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2918 * depleted. kswapd will continue to make progress and wake the processes
2919 * when the low watermark is reached.
2921 * Returns true if a fatal signal was delivered during throttling. If this
2922 * happens, the page allocator should not consider triggering the OOM killer.
2924 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2925 nodemask_t
*nodemask
)
2929 pg_data_t
*pgdat
= NULL
;
2932 * Kernel threads should not be throttled as they may be indirectly
2933 * responsible for cleaning pages necessary for reclaim to make forward
2934 * progress. kjournald for example may enter direct reclaim while
2935 * committing a transaction where throttling it could forcing other
2936 * processes to block on log_wait_commit().
2938 if (current
->flags
& PF_KTHREAD
)
2942 * If a fatal signal is pending, this process should not throttle.
2943 * It should return quickly so it can exit and free its memory
2945 if (fatal_signal_pending(current
))
2949 * Check if the pfmemalloc reserves are ok by finding the first node
2950 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2951 * GFP_KERNEL will be required for allocating network buffers when
2952 * swapping over the network so ZONE_HIGHMEM is unusable.
2954 * Throttling is based on the first usable node and throttled processes
2955 * wait on a queue until kswapd makes progress and wakes them. There
2956 * is an affinity then between processes waking up and where reclaim
2957 * progress has been made assuming the process wakes on the same node.
2958 * More importantly, processes running on remote nodes will not compete
2959 * for remote pfmemalloc reserves and processes on different nodes
2960 * should make reasonable progress.
2962 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2963 gfp_zone(gfp_mask
), nodemask
) {
2964 if (zone_idx(zone
) > ZONE_NORMAL
)
2967 /* Throttle based on the first usable node */
2968 pgdat
= zone
->zone_pgdat
;
2969 if (allow_direct_reclaim(pgdat
))
2974 /* If no zone was usable by the allocation flags then do not throttle */
2978 /* Account for the throttling */
2979 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2982 * If the caller cannot enter the filesystem, it's possible that it
2983 * is due to the caller holding an FS lock or performing a journal
2984 * transaction in the case of a filesystem like ext[3|4]. In this case,
2985 * it is not safe to block on pfmemalloc_wait as kswapd could be
2986 * blocked waiting on the same lock. Instead, throttle for up to a
2987 * second before continuing.
2989 if (!(gfp_mask
& __GFP_FS
)) {
2990 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2991 allow_direct_reclaim(pgdat
), HZ
);
2996 /* Throttle until kswapd wakes the process */
2997 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2998 allow_direct_reclaim(pgdat
));
3001 if (fatal_signal_pending(current
))
3008 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3009 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3011 unsigned long nr_reclaimed
;
3012 struct scan_control sc
= {
3013 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3014 .gfp_mask
= current_gfp_context(gfp_mask
),
3015 .reclaim_idx
= gfp_zone(gfp_mask
),
3017 .nodemask
= nodemask
,
3018 .priority
= DEF_PRIORITY
,
3019 .may_writepage
= !laptop_mode
,
3025 * Do not enter reclaim if fatal signal was delivered while throttled.
3026 * 1 is returned so that the page allocator does not OOM kill at this
3029 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3032 trace_mm_vmscan_direct_reclaim_begin(order
,
3037 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3039 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3041 return nr_reclaimed
;
3046 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3047 gfp_t gfp_mask
, bool noswap
,
3049 unsigned long *nr_scanned
)
3051 struct scan_control sc
= {
3052 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3053 .target_mem_cgroup
= memcg
,
3054 .may_writepage
= !laptop_mode
,
3056 .reclaim_idx
= MAX_NR_ZONES
- 1,
3057 .may_swap
= !noswap
,
3059 unsigned long lru_pages
;
3061 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3062 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3064 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3070 * NOTE: Although we can get the priority field, using it
3071 * here is not a good idea, since it limits the pages we can scan.
3072 * if we don't reclaim here, the shrink_node from balance_pgdat
3073 * will pick up pages from other mem cgroup's as well. We hack
3074 * the priority and make it zero.
3076 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3078 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3080 *nr_scanned
= sc
.nr_scanned
;
3081 return sc
.nr_reclaimed
;
3084 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3085 unsigned long nr_pages
,
3089 struct zonelist
*zonelist
;
3090 unsigned long nr_reclaimed
;
3092 unsigned int noreclaim_flag
;
3093 struct scan_control sc
= {
3094 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3095 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3096 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3097 .reclaim_idx
= MAX_NR_ZONES
- 1,
3098 .target_mem_cgroup
= memcg
,
3099 .priority
= DEF_PRIORITY
,
3100 .may_writepage
= !laptop_mode
,
3102 .may_swap
= may_swap
,
3106 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3107 * take care of from where we get pages. So the node where we start the
3108 * scan does not need to be the current node.
3110 nid
= mem_cgroup_select_victim_node(memcg
);
3112 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3114 trace_mm_vmscan_memcg_reclaim_begin(0,
3119 noreclaim_flag
= memalloc_noreclaim_save();
3120 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3121 memalloc_noreclaim_restore(noreclaim_flag
);
3123 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3125 return nr_reclaimed
;
3129 static void age_active_anon(struct pglist_data
*pgdat
,
3130 struct scan_control
*sc
)
3132 struct mem_cgroup
*memcg
;
3134 if (!total_swap_pages
)
3137 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3139 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3141 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3142 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3143 sc
, LRU_ACTIVE_ANON
);
3145 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3150 * Returns true if there is an eligible zone balanced for the request order
3153 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3156 unsigned long mark
= -1;
3159 for (i
= 0; i
<= classzone_idx
; i
++) {
3160 zone
= pgdat
->node_zones
+ i
;
3162 if (!managed_zone(zone
))
3165 mark
= high_wmark_pages(zone
);
3166 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3171 * If a node has no populated zone within classzone_idx, it does not
3172 * need balancing by definition. This can happen if a zone-restricted
3173 * allocation tries to wake a remote kswapd.
3181 /* Clear pgdat state for congested, dirty or under writeback. */
3182 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3184 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3185 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3186 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3190 * Prepare kswapd for sleeping. This verifies that there are no processes
3191 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3193 * Returns true if kswapd is ready to sleep
3195 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3198 * The throttled processes are normally woken up in balance_pgdat() as
3199 * soon as allow_direct_reclaim() is true. But there is a potential
3200 * race between when kswapd checks the watermarks and a process gets
3201 * throttled. There is also a potential race if processes get
3202 * throttled, kswapd wakes, a large process exits thereby balancing the
3203 * zones, which causes kswapd to exit balance_pgdat() before reaching
3204 * the wake up checks. If kswapd is going to sleep, no process should
3205 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3206 * the wake up is premature, processes will wake kswapd and get
3207 * throttled again. The difference from wake ups in balance_pgdat() is
3208 * that here we are under prepare_to_wait().
3210 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3211 wake_up_all(&pgdat
->pfmemalloc_wait
);
3213 /* Hopeless node, leave it to direct reclaim */
3214 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3217 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3218 clear_pgdat_congested(pgdat
);
3226 * kswapd shrinks a node of pages that are at or below the highest usable
3227 * zone that is currently unbalanced.
3229 * Returns true if kswapd scanned at least the requested number of pages to
3230 * reclaim or if the lack of progress was due to pages under writeback.
3231 * This is used to determine if the scanning priority needs to be raised.
3233 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3234 struct scan_control
*sc
)
3239 /* Reclaim a number of pages proportional to the number of zones */
3240 sc
->nr_to_reclaim
= 0;
3241 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3242 zone
= pgdat
->node_zones
+ z
;
3243 if (!managed_zone(zone
))
3246 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3250 * Historically care was taken to put equal pressure on all zones but
3251 * now pressure is applied based on node LRU order.
3253 shrink_node(pgdat
, sc
);
3256 * Fragmentation may mean that the system cannot be rebalanced for
3257 * high-order allocations. If twice the allocation size has been
3258 * reclaimed then recheck watermarks only at order-0 to prevent
3259 * excessive reclaim. Assume that a process requested a high-order
3260 * can direct reclaim/compact.
3262 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3265 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3269 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3270 * that are eligible for use by the caller until at least one zone is
3273 * Returns the order kswapd finished reclaiming at.
3275 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3276 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3277 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3278 * or lower is eligible for reclaim until at least one usable zone is
3281 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3284 unsigned long nr_soft_reclaimed
;
3285 unsigned long nr_soft_scanned
;
3287 struct scan_control sc
= {
3288 .gfp_mask
= GFP_KERNEL
,
3290 .priority
= DEF_PRIORITY
,
3291 .may_writepage
= !laptop_mode
,
3295 count_vm_event(PAGEOUTRUN
);
3298 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3299 bool raise_priority
= true;
3301 sc
.reclaim_idx
= classzone_idx
;
3304 * If the number of buffer_heads exceeds the maximum allowed
3305 * then consider reclaiming from all zones. This has a dual
3306 * purpose -- on 64-bit systems it is expected that
3307 * buffer_heads are stripped during active rotation. On 32-bit
3308 * systems, highmem pages can pin lowmem memory and shrinking
3309 * buffers can relieve lowmem pressure. Reclaim may still not
3310 * go ahead if all eligible zones for the original allocation
3311 * request are balanced to avoid excessive reclaim from kswapd.
3313 if (buffer_heads_over_limit
) {
3314 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3315 zone
= pgdat
->node_zones
+ i
;
3316 if (!managed_zone(zone
))
3325 * Only reclaim if there are no eligible zones. Note that
3326 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3329 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3333 * Do some background aging of the anon list, to give
3334 * pages a chance to be referenced before reclaiming. All
3335 * pages are rotated regardless of classzone as this is
3336 * about consistent aging.
3338 age_active_anon(pgdat
, &sc
);
3341 * If we're getting trouble reclaiming, start doing writepage
3342 * even in laptop mode.
3344 if (sc
.priority
< DEF_PRIORITY
- 2)
3345 sc
.may_writepage
= 1;
3347 /* Call soft limit reclaim before calling shrink_node. */
3349 nr_soft_scanned
= 0;
3350 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3351 sc
.gfp_mask
, &nr_soft_scanned
);
3352 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3355 * There should be no need to raise the scanning priority if
3356 * enough pages are already being scanned that that high
3357 * watermark would be met at 100% efficiency.
3359 if (kswapd_shrink_node(pgdat
, &sc
))
3360 raise_priority
= false;
3363 * If the low watermark is met there is no need for processes
3364 * to be throttled on pfmemalloc_wait as they should not be
3365 * able to safely make forward progress. Wake them
3367 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3368 allow_direct_reclaim(pgdat
))
3369 wake_up_all(&pgdat
->pfmemalloc_wait
);
3371 /* Check if kswapd should be suspending */
3372 if (try_to_freeze() || kthread_should_stop())
3376 * Raise priority if scanning rate is too low or there was no
3377 * progress in reclaiming pages
3379 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3380 if (raise_priority
|| !nr_reclaimed
)
3382 } while (sc
.priority
>= 1);
3384 if (!sc
.nr_reclaimed
)
3385 pgdat
->kswapd_failures
++;
3388 snapshot_refaults(NULL
, pgdat
);
3390 * Return the order kswapd stopped reclaiming at as
3391 * prepare_kswapd_sleep() takes it into account. If another caller
3392 * entered the allocator slow path while kswapd was awake, order will
3393 * remain at the higher level.
3399 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3400 * allocation request woke kswapd for. When kswapd has not woken recently,
3401 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3402 * given classzone and returns it or the highest classzone index kswapd
3403 * was recently woke for.
3405 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3406 enum zone_type classzone_idx
)
3408 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3409 return classzone_idx
;
3411 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3414 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3415 unsigned int classzone_idx
)
3420 if (freezing(current
) || kthread_should_stop())
3423 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3426 * Try to sleep for a short interval. Note that kcompactd will only be
3427 * woken if it is possible to sleep for a short interval. This is
3428 * deliberate on the assumption that if reclaim cannot keep an
3429 * eligible zone balanced that it's also unlikely that compaction will
3432 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3434 * Compaction records what page blocks it recently failed to
3435 * isolate pages from and skips them in the future scanning.
3436 * When kswapd is going to sleep, it is reasonable to assume
3437 * that pages and compaction may succeed so reset the cache.
3439 reset_isolation_suitable(pgdat
);
3442 * We have freed the memory, now we should compact it to make
3443 * allocation of the requested order possible.
3445 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3447 remaining
= schedule_timeout(HZ
/10);
3450 * If woken prematurely then reset kswapd_classzone_idx and
3451 * order. The values will either be from a wakeup request or
3452 * the previous request that slept prematurely.
3455 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3456 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3459 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3460 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3464 * After a short sleep, check if it was a premature sleep. If not, then
3465 * go fully to sleep until explicitly woken up.
3468 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3469 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3472 * vmstat counters are not perfectly accurate and the estimated
3473 * value for counters such as NR_FREE_PAGES can deviate from the
3474 * true value by nr_online_cpus * threshold. To avoid the zone
3475 * watermarks being breached while under pressure, we reduce the
3476 * per-cpu vmstat threshold while kswapd is awake and restore
3477 * them before going back to sleep.
3479 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3481 if (!kthread_should_stop())
3484 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3487 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3489 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3491 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3495 * The background pageout daemon, started as a kernel thread
3496 * from the init process.
3498 * This basically trickles out pages so that we have _some_
3499 * free memory available even if there is no other activity
3500 * that frees anything up. This is needed for things like routing
3501 * etc, where we otherwise might have all activity going on in
3502 * asynchronous contexts that cannot page things out.
3504 * If there are applications that are active memory-allocators
3505 * (most normal use), this basically shouldn't matter.
3507 static int kswapd(void *p
)
3509 unsigned int alloc_order
, reclaim_order
;
3510 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3511 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3512 struct task_struct
*tsk
= current
;
3514 struct reclaim_state reclaim_state
= {
3515 .reclaimed_slab
= 0,
3517 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3519 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3521 if (!cpumask_empty(cpumask
))
3522 set_cpus_allowed_ptr(tsk
, cpumask
);
3523 current
->reclaim_state
= &reclaim_state
;
3526 * Tell the memory management that we're a "memory allocator",
3527 * and that if we need more memory we should get access to it
3528 * regardless (see "__alloc_pages()"). "kswapd" should
3529 * never get caught in the normal page freeing logic.
3531 * (Kswapd normally doesn't need memory anyway, but sometimes
3532 * you need a small amount of memory in order to be able to
3533 * page out something else, and this flag essentially protects
3534 * us from recursively trying to free more memory as we're
3535 * trying to free the first piece of memory in the first place).
3537 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3540 pgdat
->kswapd_order
= 0;
3541 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3545 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3546 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3549 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3552 /* Read the new order and classzone_idx */
3553 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3554 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3555 pgdat
->kswapd_order
= 0;
3556 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3558 ret
= try_to_freeze();
3559 if (kthread_should_stop())
3563 * We can speed up thawing tasks if we don't call balance_pgdat
3564 * after returning from the refrigerator
3570 * Reclaim begins at the requested order but if a high-order
3571 * reclaim fails then kswapd falls back to reclaiming for
3572 * order-0. If that happens, kswapd will consider sleeping
3573 * for the order it finished reclaiming at (reclaim_order)
3574 * but kcompactd is woken to compact for the original
3575 * request (alloc_order).
3577 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3579 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3580 if (reclaim_order
< alloc_order
)
3581 goto kswapd_try_sleep
;
3584 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3585 current
->reclaim_state
= NULL
;
3586 lockdep_clear_current_reclaim_state();
3592 * A zone is low on free memory, so wake its kswapd task to service it.
3594 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3598 if (!managed_zone(zone
))
3601 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3603 pgdat
= zone
->zone_pgdat
;
3604 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3606 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3607 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3610 /* Hopeless node, leave it to direct reclaim */
3611 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3614 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3617 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3618 wake_up_interruptible(&pgdat
->kswapd_wait
);
3621 #ifdef CONFIG_HIBERNATION
3623 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3626 * Rather than trying to age LRUs the aim is to preserve the overall
3627 * LRU order by reclaiming preferentially
3628 * inactive > active > active referenced > active mapped
3630 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3632 struct reclaim_state reclaim_state
;
3633 struct scan_control sc
= {
3634 .nr_to_reclaim
= nr_to_reclaim
,
3635 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3636 .reclaim_idx
= MAX_NR_ZONES
- 1,
3637 .priority
= DEF_PRIORITY
,
3641 .hibernation_mode
= 1,
3643 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3644 struct task_struct
*p
= current
;
3645 unsigned long nr_reclaimed
;
3646 unsigned int noreclaim_flag
;
3648 noreclaim_flag
= memalloc_noreclaim_save();
3649 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3650 reclaim_state
.reclaimed_slab
= 0;
3651 p
->reclaim_state
= &reclaim_state
;
3653 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3655 p
->reclaim_state
= NULL
;
3656 lockdep_clear_current_reclaim_state();
3657 memalloc_noreclaim_restore(noreclaim_flag
);
3659 return nr_reclaimed
;
3661 #endif /* CONFIG_HIBERNATION */
3663 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3664 not required for correctness. So if the last cpu in a node goes
3665 away, we get changed to run anywhere: as the first one comes back,
3666 restore their cpu bindings. */
3667 static int kswapd_cpu_online(unsigned int cpu
)
3671 for_each_node_state(nid
, N_MEMORY
) {
3672 pg_data_t
*pgdat
= NODE_DATA(nid
);
3673 const struct cpumask
*mask
;
3675 mask
= cpumask_of_node(pgdat
->node_id
);
3677 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3678 /* One of our CPUs online: restore mask */
3679 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3685 * This kswapd start function will be called by init and node-hot-add.
3686 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3688 int kswapd_run(int nid
)
3690 pg_data_t
*pgdat
= NODE_DATA(nid
);
3696 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3697 if (IS_ERR(pgdat
->kswapd
)) {
3698 /* failure at boot is fatal */
3699 BUG_ON(system_state
< SYSTEM_RUNNING
);
3700 pr_err("Failed to start kswapd on node %d\n", nid
);
3701 ret
= PTR_ERR(pgdat
->kswapd
);
3702 pgdat
->kswapd
= NULL
;
3708 * Called by memory hotplug when all memory in a node is offlined. Caller must
3709 * hold mem_hotplug_begin/end().
3711 void kswapd_stop(int nid
)
3713 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3716 kthread_stop(kswapd
);
3717 NODE_DATA(nid
)->kswapd
= NULL
;
3721 static int __init
kswapd_init(void)
3726 for_each_node_state(nid
, N_MEMORY
)
3728 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3729 "mm/vmscan:online", kswapd_cpu_online
,
3735 module_init(kswapd_init
)
3741 * If non-zero call node_reclaim when the number of free pages falls below
3744 int node_reclaim_mode __read_mostly
;
3746 #define RECLAIM_OFF 0
3747 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3748 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3749 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3752 * Priority for NODE_RECLAIM. This determines the fraction of pages
3753 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3756 #define NODE_RECLAIM_PRIORITY 4
3759 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3762 int sysctl_min_unmapped_ratio
= 1;
3765 * If the number of slab pages in a zone grows beyond this percentage then
3766 * slab reclaim needs to occur.
3768 int sysctl_min_slab_ratio
= 5;
3770 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3772 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3773 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3774 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3777 * It's possible for there to be more file mapped pages than
3778 * accounted for by the pages on the file LRU lists because
3779 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3781 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3784 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3785 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3787 unsigned long nr_pagecache_reclaimable
;
3788 unsigned long delta
= 0;
3791 * If RECLAIM_UNMAP is set, then all file pages are considered
3792 * potentially reclaimable. Otherwise, we have to worry about
3793 * pages like swapcache and node_unmapped_file_pages() provides
3796 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3797 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3799 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3801 /* If we can't clean pages, remove dirty pages from consideration */
3802 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3803 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3805 /* Watch for any possible underflows due to delta */
3806 if (unlikely(delta
> nr_pagecache_reclaimable
))
3807 delta
= nr_pagecache_reclaimable
;
3809 return nr_pagecache_reclaimable
- delta
;
3813 * Try to free up some pages from this node through reclaim.
3815 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3817 /* Minimum pages needed in order to stay on node */
3818 const unsigned long nr_pages
= 1 << order
;
3819 struct task_struct
*p
= current
;
3820 struct reclaim_state reclaim_state
;
3821 unsigned int noreclaim_flag
;
3822 struct scan_control sc
= {
3823 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3824 .gfp_mask
= current_gfp_context(gfp_mask
),
3826 .priority
= NODE_RECLAIM_PRIORITY
,
3827 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3828 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3830 .reclaim_idx
= gfp_zone(gfp_mask
),
3835 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3836 * and we also need to be able to write out pages for RECLAIM_WRITE
3837 * and RECLAIM_UNMAP.
3839 noreclaim_flag
= memalloc_noreclaim_save();
3840 p
->flags
|= PF_SWAPWRITE
;
3841 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3842 reclaim_state
.reclaimed_slab
= 0;
3843 p
->reclaim_state
= &reclaim_state
;
3845 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3847 * Free memory by calling shrink zone with increasing
3848 * priorities until we have enough memory freed.
3851 shrink_node(pgdat
, &sc
);
3852 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3855 p
->reclaim_state
= NULL
;
3856 current
->flags
&= ~PF_SWAPWRITE
;
3857 memalloc_noreclaim_restore(noreclaim_flag
);
3858 lockdep_clear_current_reclaim_state();
3859 return sc
.nr_reclaimed
>= nr_pages
;
3862 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3867 * Node reclaim reclaims unmapped file backed pages and
3868 * slab pages if we are over the defined limits.
3870 * A small portion of unmapped file backed pages is needed for
3871 * file I/O otherwise pages read by file I/O will be immediately
3872 * thrown out if the node is overallocated. So we do not reclaim
3873 * if less than a specified percentage of the node is used by
3874 * unmapped file backed pages.
3876 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3877 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3878 return NODE_RECLAIM_FULL
;
3881 * Do not scan if the allocation should not be delayed.
3883 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3884 return NODE_RECLAIM_NOSCAN
;
3887 * Only run node reclaim on the local node or on nodes that do not
3888 * have associated processors. This will favor the local processor
3889 * over remote processors and spread off node memory allocations
3890 * as wide as possible.
3892 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3893 return NODE_RECLAIM_NOSCAN
;
3895 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3896 return NODE_RECLAIM_NOSCAN
;
3898 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3899 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3902 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3909 * page_evictable - test whether a page is evictable
3910 * @page: the page to test
3912 * Test whether page is evictable--i.e., should be placed on active/inactive
3913 * lists vs unevictable list.
3915 * Reasons page might not be evictable:
3916 * (1) page's mapping marked unevictable
3917 * (2) page is part of an mlocked VMA
3920 int page_evictable(struct page
*page
)
3922 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3927 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3928 * @pages: array of pages to check
3929 * @nr_pages: number of pages to check
3931 * Checks pages for evictability and moves them to the appropriate lru list.
3933 * This function is only used for SysV IPC SHM_UNLOCK.
3935 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3937 struct lruvec
*lruvec
;
3938 struct pglist_data
*pgdat
= NULL
;
3943 for (i
= 0; i
< nr_pages
; i
++) {
3944 struct page
*page
= pages
[i
];
3945 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3948 if (pagepgdat
!= pgdat
) {
3950 spin_unlock_irq(&pgdat
->lru_lock
);
3952 spin_lock_irq(&pgdat
->lru_lock
);
3954 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
3956 if (!PageLRU(page
) || !PageUnevictable(page
))
3959 if (page_evictable(page
)) {
3960 enum lru_list lru
= page_lru_base_type(page
);
3962 VM_BUG_ON_PAGE(PageActive(page
), page
);
3963 ClearPageUnevictable(page
);
3964 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3965 add_page_to_lru_list(page
, lruvec
, lru
);
3971 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3972 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3973 spin_unlock_irq(&pgdat
->lru_lock
);
3976 #endif /* CONFIG_SHMEM */