4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup
*target_mem_cgroup
;
84 /* Scan (total_size >> priority) pages at once */
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx
;
90 /* Writepage batching in laptop mode; RECLAIM_WRITE */
91 unsigned int may_writepage
:1;
93 /* Can mapped pages be reclaimed? */
94 unsigned int may_unmap
:1;
96 /* Can pages be swapped as part of reclaim? */
97 unsigned int may_swap
:1;
99 /* Can cgroups be reclaimed below their normal consumption range? */
100 unsigned int may_thrash
:1;
102 unsigned int hibernation_mode
:1;
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready
:1;
107 /* Incremented by the number of inactive pages that were scanned */
108 unsigned long nr_scanned
;
110 /* Number of pages freed so far during a call to shrink_zones() */
111 unsigned long nr_reclaimed
;
114 #ifdef ARCH_HAS_PREFETCH
115 #define prefetch_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetch(&prev->_field); \
125 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
128 #ifdef ARCH_HAS_PREFETCHW
129 #define prefetchw_prev_lru_page(_page, _base, _field) \
131 if ((_page)->lru.prev != _base) { \
134 prev = lru_to_page(&(_page->lru)); \
135 prefetchw(&prev->_field); \
139 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
143 * From 0 .. 100. Higher means more swappy.
145 int vm_swappiness
= 60;
147 * The total number of pages which are beyond the high watermark within all
150 unsigned long vm_total_pages
;
152 static LIST_HEAD(shrinker_list
);
153 static DECLARE_RWSEM(shrinker_rwsem
);
156 static bool global_reclaim(struct scan_control
*sc
)
158 return !sc
->target_mem_cgroup
;
162 * sane_reclaim - is the usual dirty throttling mechanism operational?
163 * @sc: scan_control in question
165 * The normal page dirty throttling mechanism in balance_dirty_pages() is
166 * completely broken with the legacy memcg and direct stalling in
167 * shrink_page_list() is used for throttling instead, which lacks all the
168 * niceties such as fairness, adaptive pausing, bandwidth proportional
169 * allocation and configurability.
171 * This function tests whether the vmscan currently in progress can assume
172 * that the normal dirty throttling mechanism is operational.
174 static bool sane_reclaim(struct scan_control
*sc
)
176 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
180 #ifdef CONFIG_CGROUP_WRITEBACK
181 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
187 static bool global_reclaim(struct scan_control
*sc
)
192 static bool sane_reclaim(struct scan_control
*sc
)
199 * This misses isolated pages which are not accounted for to save counters.
200 * As the data only determines if reclaim or compaction continues, it is
201 * not expected that isolated pages will be a dominating factor.
203 unsigned long zone_reclaimable_pages(struct zone
*zone
)
207 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
208 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
209 if (get_nr_swap_pages() > 0)
210 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
211 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
216 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
220 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
221 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
222 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
224 if (get_nr_swap_pages() > 0)
225 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
226 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
227 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
232 bool pgdat_reclaimable(struct pglist_data
*pgdat
)
234 return node_page_state_snapshot(pgdat
, NR_PAGES_SCANNED
) <
235 pgdat_reclaimable_pages(pgdat
) * 6;
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 swapcache_free(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
)) {
920 /* By default assume that the page flags are accurate */
921 *dirty
= PageDirty(page
);
922 *writeback
= PageWriteback(page
);
924 /* Verify dirty/writeback state if the filesystem supports it */
925 if (!page_has_private(page
))
928 mapping
= page_mapping(page
);
929 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
930 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
933 struct reclaim_stat
{
935 unsigned nr_unqueued_dirty
;
936 unsigned nr_congested
;
937 unsigned nr_writeback
;
938 unsigned nr_immediate
;
939 unsigned nr_activate
;
940 unsigned nr_ref_keep
;
941 unsigned nr_unmap_fail
;
945 * shrink_page_list() returns the number of reclaimed pages
947 static unsigned long shrink_page_list(struct list_head
*page_list
,
948 struct pglist_data
*pgdat
,
949 struct scan_control
*sc
,
950 enum ttu_flags ttu_flags
,
951 struct reclaim_stat
*stat
,
954 LIST_HEAD(ret_pages
);
955 LIST_HEAD(free_pages
);
957 unsigned nr_unqueued_dirty
= 0;
958 unsigned nr_dirty
= 0;
959 unsigned nr_congested
= 0;
960 unsigned nr_reclaimed
= 0;
961 unsigned nr_writeback
= 0;
962 unsigned nr_immediate
= 0;
963 unsigned nr_ref_keep
= 0;
964 unsigned nr_unmap_fail
= 0;
968 while (!list_empty(page_list
)) {
969 struct address_space
*mapping
;
972 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
973 bool dirty
, writeback
;
974 bool lazyfree
= false;
975 int ret
= SWAP_SUCCESS
;
979 page
= lru_to_page(page_list
);
980 list_del(&page
->lru
);
982 if (!trylock_page(page
))
985 VM_BUG_ON_PAGE(PageActive(page
), page
);
989 if (unlikely(!page_evictable(page
)))
992 if (!sc
->may_unmap
&& page_mapped(page
))
995 /* Double the slab pressure for mapped and swapcache pages */
996 if (page_mapped(page
) || PageSwapCache(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.
1123 if (PageAnon(page
) && !PageSwapCache(page
)) {
1124 if (!(sc
->gfp_mask
& __GFP_IO
))
1126 if (!add_to_swap(page
, page_list
))
1127 goto activate_locked
;
1131 /* Adding to swap updated mapping */
1132 mapping
= page_mapping(page
);
1133 } else if (unlikely(PageTransHuge(page
))) {
1134 /* Split file THP */
1135 if (split_huge_page_to_list(page
, page_list
))
1139 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1142 * The page is mapped into the page tables of one or more
1143 * processes. Try to unmap it here.
1145 if (page_mapped(page
) && mapping
) {
1146 switch (ret
= try_to_unmap(page
, lazyfree
?
1147 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1148 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1151 goto activate_locked
;
1159 ; /* try to free the page below */
1163 if (PageDirty(page
)) {
1165 * Only kswapd can writeback filesystem pages
1166 * to avoid risk of stack overflow. But avoid
1167 * injecting inefficient single-page IO into
1168 * flusher writeback as much as possible: only
1169 * write pages when we've encountered many
1170 * dirty pages, and when we've already scanned
1171 * the rest of the LRU for clean pages and see
1172 * the same dirty pages again (PageReclaim).
1174 if (page_is_file_cache(page
) &&
1175 (!current_is_kswapd() || !PageReclaim(page
) ||
1176 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1178 * Immediately reclaim when written back.
1179 * Similar in principal to deactivate_page()
1180 * except we already have the page isolated
1181 * and know it's dirty
1183 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1184 SetPageReclaim(page
);
1186 goto activate_locked
;
1189 if (references
== PAGEREF_RECLAIM_CLEAN
)
1193 if (!sc
->may_writepage
)
1197 * Page is dirty. Flush the TLB if a writable entry
1198 * potentially exists to avoid CPU writes after IO
1199 * starts and then write it out here.
1201 try_to_unmap_flush_dirty();
1202 switch (pageout(page
, mapping
, sc
)) {
1206 goto activate_locked
;
1208 if (PageWriteback(page
))
1210 if (PageDirty(page
))
1214 * A synchronous write - probably a ramdisk. Go
1215 * ahead and try to reclaim the page.
1217 if (!trylock_page(page
))
1219 if (PageDirty(page
) || PageWriteback(page
))
1221 mapping
= page_mapping(page
);
1223 ; /* try to free the page below */
1228 * If the page has buffers, try to free the buffer mappings
1229 * associated with this page. If we succeed we try to free
1232 * We do this even if the page is PageDirty().
1233 * try_to_release_page() does not perform I/O, but it is
1234 * possible for a page to have PageDirty set, but it is actually
1235 * clean (all its buffers are clean). This happens if the
1236 * buffers were written out directly, with submit_bh(). ext3
1237 * will do this, as well as the blockdev mapping.
1238 * try_to_release_page() will discover that cleanness and will
1239 * drop the buffers and mark the page clean - it can be freed.
1241 * Rarely, pages can have buffers and no ->mapping. These are
1242 * the pages which were not successfully invalidated in
1243 * truncate_complete_page(). We try to drop those buffers here
1244 * and if that worked, and the page is no longer mapped into
1245 * process address space (page_count == 1) it can be freed.
1246 * Otherwise, leave the page on the LRU so it is swappable.
1248 if (page_has_private(page
)) {
1249 if (!try_to_release_page(page
, sc
->gfp_mask
))
1250 goto activate_locked
;
1251 if (!mapping
&& page_count(page
) == 1) {
1253 if (put_page_testzero(page
))
1257 * rare race with speculative reference.
1258 * the speculative reference will free
1259 * this page shortly, so we may
1260 * increment nr_reclaimed here (and
1261 * leave it off the LRU).
1270 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1274 * At this point, we have no other references and there is
1275 * no way to pick any more up (removed from LRU, removed
1276 * from pagecache). Can use non-atomic bitops now (and
1277 * we obviously don't have to worry about waking up a process
1278 * waiting on the page lock, because there are no references.
1280 __ClearPageLocked(page
);
1282 if (ret
== SWAP_LZFREE
)
1283 count_vm_event(PGLAZYFREED
);
1288 * Is there need to periodically free_page_list? It would
1289 * appear not as the counts should be low
1291 list_add(&page
->lru
, &free_pages
);
1295 if (PageSwapCache(page
))
1296 try_to_free_swap(page
);
1298 list_add(&page
->lru
, &ret_pages
);
1302 /* Not a candidate for swapping, so reclaim swap space. */
1303 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1304 try_to_free_swap(page
);
1305 VM_BUG_ON_PAGE(PageActive(page
), page
);
1306 SetPageActive(page
);
1311 list_add(&page
->lru
, &ret_pages
);
1312 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1315 mem_cgroup_uncharge_list(&free_pages
);
1316 try_to_unmap_flush();
1317 free_hot_cold_page_list(&free_pages
, true);
1319 list_splice(&ret_pages
, page_list
);
1320 count_vm_events(PGACTIVATE
, pgactivate
);
1323 stat
->nr_dirty
= nr_dirty
;
1324 stat
->nr_congested
= nr_congested
;
1325 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1326 stat
->nr_writeback
= nr_writeback
;
1327 stat
->nr_immediate
= nr_immediate
;
1328 stat
->nr_activate
= pgactivate
;
1329 stat
->nr_ref_keep
= nr_ref_keep
;
1330 stat
->nr_unmap_fail
= nr_unmap_fail
;
1332 return nr_reclaimed
;
1335 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1336 struct list_head
*page_list
)
1338 struct scan_control sc
= {
1339 .gfp_mask
= GFP_KERNEL
,
1340 .priority
= DEF_PRIORITY
,
1344 struct page
*page
, *next
;
1345 LIST_HEAD(clean_pages
);
1347 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1348 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1349 !__PageMovable(page
)) {
1350 ClearPageActive(page
);
1351 list_move(&page
->lru
, &clean_pages
);
1355 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1356 TTU_UNMAP
|TTU_IGNORE_ACCESS
, NULL
, true);
1357 list_splice(&clean_pages
, page_list
);
1358 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1363 * Attempt to remove the specified page from its LRU. Only take this page
1364 * if it is of the appropriate PageActive status. Pages which are being
1365 * freed elsewhere are also ignored.
1367 * page: page to consider
1368 * mode: one of the LRU isolation modes defined above
1370 * returns 0 on success, -ve errno on failure.
1372 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1376 /* Only take pages on the LRU. */
1380 /* Compaction should not handle unevictable pages but CMA can do so */
1381 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1387 * To minimise LRU disruption, the caller can indicate that it only
1388 * wants to isolate pages it will be able to operate on without
1389 * blocking - clean pages for the most part.
1391 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1392 * that it is possible to migrate without blocking
1394 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1395 /* All the caller can do on PageWriteback is block */
1396 if (PageWriteback(page
))
1399 if (PageDirty(page
)) {
1400 struct address_space
*mapping
;
1403 * Only pages without mappings or that have a
1404 * ->migratepage callback are possible to migrate
1407 mapping
= page_mapping(page
);
1408 if (mapping
&& !mapping
->a_ops
->migratepage
)
1413 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1416 if (likely(get_page_unless_zero(page
))) {
1418 * Be careful not to clear PageLRU until after we're
1419 * sure the page is not being freed elsewhere -- the
1420 * page release code relies on it.
1431 * Update LRU sizes after isolating pages. The LRU size updates must
1432 * be complete before mem_cgroup_update_lru_size due to a santity check.
1434 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1435 enum lru_list lru
, unsigned long *nr_zone_taken
)
1439 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1440 if (!nr_zone_taken
[zid
])
1443 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1445 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1452 * zone_lru_lock is heavily contended. Some of the functions that
1453 * shrink the lists perform better by taking out a batch of pages
1454 * and working on them outside the LRU lock.
1456 * For pagecache intensive workloads, this function is the hottest
1457 * spot in the kernel (apart from copy_*_user functions).
1459 * Appropriate locks must be held before calling this function.
1461 * @nr_to_scan: The number of pages to look through on the list.
1462 * @lruvec: The LRU vector to pull pages from.
1463 * @dst: The temp list to put pages on to.
1464 * @nr_scanned: The number of pages that were scanned.
1465 * @sc: The scan_control struct for this reclaim session
1466 * @mode: One of the LRU isolation modes
1467 * @lru: LRU list id for isolating
1469 * returns how many pages were moved onto *@dst.
1471 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1472 struct lruvec
*lruvec
, struct list_head
*dst
,
1473 unsigned long *nr_scanned
, struct scan_control
*sc
,
1474 isolate_mode_t mode
, enum lru_list lru
)
1476 struct list_head
*src
= &lruvec
->lists
[lru
];
1477 unsigned long nr_taken
= 0;
1478 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1479 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1480 unsigned long skipped
= 0, total_skipped
= 0;
1481 unsigned long scan
, nr_pages
;
1482 LIST_HEAD(pages_skipped
);
1484 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1485 !list_empty(src
);) {
1488 page
= lru_to_page(src
);
1489 prefetchw_prev_lru_page(page
, src
, flags
);
1491 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1493 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1494 list_move(&page
->lru
, &pages_skipped
);
1495 nr_skipped
[page_zonenum(page
)]++;
1500 * Account for scanned and skipped separetly to avoid the pgdat
1501 * being prematurely marked unreclaimable by pgdat_reclaimable.
1505 switch (__isolate_lru_page(page
, mode
)) {
1507 nr_pages
= hpage_nr_pages(page
);
1508 nr_taken
+= nr_pages
;
1509 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1510 list_move(&page
->lru
, dst
);
1514 /* else it is being freed elsewhere */
1515 list_move(&page
->lru
, src
);
1524 * Splice any skipped pages to the start of the LRU list. Note that
1525 * this disrupts the LRU order when reclaiming for lower zones but
1526 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1527 * scanning would soon rescan the same pages to skip and put the
1528 * system at risk of premature OOM.
1530 if (!list_empty(&pages_skipped
)) {
1533 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1534 if (!nr_skipped
[zid
])
1537 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1538 skipped
+= nr_skipped
[zid
];
1542 * Account skipped pages as a partial scan as the pgdat may be
1543 * close to unreclaimable. If the LRU list is empty, account
1544 * skipped pages as a full scan.
1546 total_skipped
= list_empty(src
) ? skipped
: skipped
>> 2;
1548 list_splice(&pages_skipped
, src
);
1550 *nr_scanned
= scan
+ total_skipped
;
1551 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1552 scan
, skipped
, nr_taken
, mode
, lru
);
1553 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1558 * isolate_lru_page - tries to isolate a page from its LRU list
1559 * @page: page to isolate from its LRU list
1561 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1562 * vmstat statistic corresponding to whatever LRU list the page was on.
1564 * Returns 0 if the page was removed from an LRU list.
1565 * Returns -EBUSY if the page was not on an LRU list.
1567 * The returned page will have PageLRU() cleared. If it was found on
1568 * the active list, it will have PageActive set. If it was found on
1569 * the unevictable list, it will have the PageUnevictable bit set. That flag
1570 * may need to be cleared by the caller before letting the page go.
1572 * The vmstat statistic corresponding to the list on which the page was
1573 * found will be decremented.
1576 * (1) Must be called with an elevated refcount on the page. This is a
1577 * fundamentnal difference from isolate_lru_pages (which is called
1578 * without a stable reference).
1579 * (2) the lru_lock must not be held.
1580 * (3) interrupts must be enabled.
1582 int isolate_lru_page(struct page
*page
)
1586 VM_BUG_ON_PAGE(!page_count(page
), page
);
1587 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1589 if (PageLRU(page
)) {
1590 struct zone
*zone
= page_zone(page
);
1591 struct lruvec
*lruvec
;
1593 spin_lock_irq(zone_lru_lock(zone
));
1594 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1595 if (PageLRU(page
)) {
1596 int lru
= page_lru(page
);
1599 del_page_from_lru_list(page
, lruvec
, lru
);
1602 spin_unlock_irq(zone_lru_lock(zone
));
1608 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1609 * then get resheduled. When there are massive number of tasks doing page
1610 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1611 * the LRU list will go small and be scanned faster than necessary, leading to
1612 * unnecessary swapping, thrashing and OOM.
1614 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1615 struct scan_control
*sc
)
1617 unsigned long inactive
, isolated
;
1619 if (current_is_kswapd())
1622 if (!sane_reclaim(sc
))
1626 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1627 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1629 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1630 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1634 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1635 * won't get blocked by normal direct-reclaimers, forming a circular
1638 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1641 return isolated
> inactive
;
1644 static noinline_for_stack
void
1645 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1647 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1648 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1649 LIST_HEAD(pages_to_free
);
1652 * Put back any unfreeable pages.
1654 while (!list_empty(page_list
)) {
1655 struct page
*page
= lru_to_page(page_list
);
1658 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1659 list_del(&page
->lru
);
1660 if (unlikely(!page_evictable(page
))) {
1661 spin_unlock_irq(&pgdat
->lru_lock
);
1662 putback_lru_page(page
);
1663 spin_lock_irq(&pgdat
->lru_lock
);
1667 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1670 lru
= page_lru(page
);
1671 add_page_to_lru_list(page
, lruvec
, lru
);
1673 if (is_active_lru(lru
)) {
1674 int file
= is_file_lru(lru
);
1675 int numpages
= hpage_nr_pages(page
);
1676 reclaim_stat
->recent_rotated
[file
] += numpages
;
1678 if (put_page_testzero(page
)) {
1679 __ClearPageLRU(page
);
1680 __ClearPageActive(page
);
1681 del_page_from_lru_list(page
, lruvec
, lru
);
1683 if (unlikely(PageCompound(page
))) {
1684 spin_unlock_irq(&pgdat
->lru_lock
);
1685 mem_cgroup_uncharge(page
);
1686 (*get_compound_page_dtor(page
))(page
);
1687 spin_lock_irq(&pgdat
->lru_lock
);
1689 list_add(&page
->lru
, &pages_to_free
);
1694 * To save our caller's stack, now use input list for pages to free.
1696 list_splice(&pages_to_free
, page_list
);
1700 * If a kernel thread (such as nfsd for loop-back mounts) services
1701 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1702 * In that case we should only throttle if the backing device it is
1703 * writing to is congested. In other cases it is safe to throttle.
1705 static int current_may_throttle(void)
1707 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1708 current
->backing_dev_info
== NULL
||
1709 bdi_write_congested(current
->backing_dev_info
);
1713 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1714 * of reclaimed pages
1716 static noinline_for_stack
unsigned long
1717 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1718 struct scan_control
*sc
, enum lru_list lru
)
1720 LIST_HEAD(page_list
);
1721 unsigned long nr_scanned
;
1722 unsigned long nr_reclaimed
= 0;
1723 unsigned long nr_taken
;
1724 struct reclaim_stat stat
= {};
1725 isolate_mode_t isolate_mode
= 0;
1726 int file
= is_file_lru(lru
);
1727 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1728 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1730 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1731 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1733 /* We are about to die and free our memory. Return now. */
1734 if (fatal_signal_pending(current
))
1735 return SWAP_CLUSTER_MAX
;
1741 isolate_mode
|= ISOLATE_UNMAPPED
;
1743 spin_lock_irq(&pgdat
->lru_lock
);
1745 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1746 &nr_scanned
, sc
, isolate_mode
, lru
);
1748 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1749 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1751 if (global_reclaim(sc
)) {
1752 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1753 if (current_is_kswapd())
1754 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1756 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1758 spin_unlock_irq(&pgdat
->lru_lock
);
1763 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, TTU_UNMAP
,
1766 spin_lock_irq(&pgdat
->lru_lock
);
1768 if (global_reclaim(sc
)) {
1769 if (current_is_kswapd())
1770 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1772 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1775 putback_inactive_pages(lruvec
, &page_list
);
1777 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1779 spin_unlock_irq(&pgdat
->lru_lock
);
1781 mem_cgroup_uncharge_list(&page_list
);
1782 free_hot_cold_page_list(&page_list
, true);
1785 * If reclaim is isolating dirty pages under writeback, it implies
1786 * that the long-lived page allocation rate is exceeding the page
1787 * laundering rate. Either the global limits are not being effective
1788 * at throttling processes due to the page distribution throughout
1789 * zones or there is heavy usage of a slow backing device. The
1790 * only option is to throttle from reclaim context which is not ideal
1791 * as there is no guarantee the dirtying process is throttled in the
1792 * same way balance_dirty_pages() manages.
1794 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1795 * of pages under pages flagged for immediate reclaim and stall if any
1796 * are encountered in the nr_immediate check below.
1798 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1799 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1802 * Legacy memcg will stall in page writeback so avoid forcibly
1805 if (sane_reclaim(sc
)) {
1807 * Tag a zone as congested if all the dirty pages scanned were
1808 * backed by a congested BDI and wait_iff_congested will stall.
1810 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1811 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1814 * If dirty pages are scanned that are not queued for IO, it
1815 * implies that flushers are not doing their job. This can
1816 * happen when memory pressure pushes dirty pages to the end of
1817 * the LRU before the dirty limits are breached and the dirty
1818 * data has expired. It can also happen when the proportion of
1819 * dirty pages grows not through writes but through memory
1820 * pressure reclaiming all the clean cache. And in some cases,
1821 * the flushers simply cannot keep up with the allocation
1822 * rate. Nudge the flusher threads in case they are asleep, but
1823 * also allow kswapd to start writing pages during reclaim.
1825 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1826 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1827 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1831 * If kswapd scans pages marked marked for immediate
1832 * reclaim and under writeback (nr_immediate), it implies
1833 * that pages are cycling through the LRU faster than
1834 * they are written so also forcibly stall.
1836 if (stat
.nr_immediate
&& current_may_throttle())
1837 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1841 * Stall direct reclaim for IO completions if underlying BDIs or zone
1842 * is congested. Allow kswapd to continue until it starts encountering
1843 * unqueued dirty pages or cycling through the LRU too quickly.
1845 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1846 current_may_throttle())
1847 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1849 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1850 nr_scanned
, nr_reclaimed
,
1851 stat
.nr_dirty
, stat
.nr_writeback
,
1852 stat
.nr_congested
, stat
.nr_immediate
,
1853 stat
.nr_activate
, stat
.nr_ref_keep
,
1855 sc
->priority
, file
);
1856 return nr_reclaimed
;
1860 * This moves pages from the active list to the inactive list.
1862 * We move them the other way if the page is referenced by one or more
1863 * processes, from rmap.
1865 * If the pages are mostly unmapped, the processing is fast and it is
1866 * appropriate to hold zone_lru_lock across the whole operation. But if
1867 * the pages are mapped, the processing is slow (page_referenced()) so we
1868 * should drop zone_lru_lock around each page. It's impossible to balance
1869 * this, so instead we remove the pages from the LRU while processing them.
1870 * It is safe to rely on PG_active against the non-LRU pages in here because
1871 * nobody will play with that bit on a non-LRU page.
1873 * The downside is that we have to touch page->_refcount against each page.
1874 * But we had to alter page->flags anyway.
1876 * Returns the number of pages moved to the given lru.
1879 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1880 struct list_head
*list
,
1881 struct list_head
*pages_to_free
,
1884 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1889 while (!list_empty(list
)) {
1890 page
= lru_to_page(list
);
1891 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1893 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1896 nr_pages
= hpage_nr_pages(page
);
1897 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1898 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1900 if (put_page_testzero(page
)) {
1901 __ClearPageLRU(page
);
1902 __ClearPageActive(page
);
1903 del_page_from_lru_list(page
, lruvec
, lru
);
1905 if (unlikely(PageCompound(page
))) {
1906 spin_unlock_irq(&pgdat
->lru_lock
);
1907 mem_cgroup_uncharge(page
);
1908 (*get_compound_page_dtor(page
))(page
);
1909 spin_lock_irq(&pgdat
->lru_lock
);
1911 list_add(&page
->lru
, pages_to_free
);
1913 nr_moved
+= nr_pages
;
1917 if (!is_active_lru(lru
))
1918 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1923 static void shrink_active_list(unsigned long nr_to_scan
,
1924 struct lruvec
*lruvec
,
1925 struct scan_control
*sc
,
1928 unsigned long nr_taken
;
1929 unsigned long nr_scanned
;
1930 unsigned long vm_flags
;
1931 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1932 LIST_HEAD(l_active
);
1933 LIST_HEAD(l_inactive
);
1935 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1936 unsigned nr_deactivate
, nr_activate
;
1937 unsigned nr_rotated
= 0;
1938 isolate_mode_t isolate_mode
= 0;
1939 int file
= is_file_lru(lru
);
1940 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1945 isolate_mode
|= ISOLATE_UNMAPPED
;
1947 spin_lock_irq(&pgdat
->lru_lock
);
1949 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1950 &nr_scanned
, sc
, isolate_mode
, lru
);
1952 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1953 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1955 if (global_reclaim(sc
))
1956 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1957 __count_vm_events(PGREFILL
, nr_scanned
);
1959 spin_unlock_irq(&pgdat
->lru_lock
);
1961 while (!list_empty(&l_hold
)) {
1963 page
= lru_to_page(&l_hold
);
1964 list_del(&page
->lru
);
1966 if (unlikely(!page_evictable(page
))) {
1967 putback_lru_page(page
);
1971 if (unlikely(buffer_heads_over_limit
)) {
1972 if (page_has_private(page
) && trylock_page(page
)) {
1973 if (page_has_private(page
))
1974 try_to_release_page(page
, 0);
1979 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1981 nr_rotated
+= hpage_nr_pages(page
);
1983 * Identify referenced, file-backed active pages and
1984 * give them one more trip around the active list. So
1985 * that executable code get better chances to stay in
1986 * memory under moderate memory pressure. Anon pages
1987 * are not likely to be evicted by use-once streaming
1988 * IO, plus JVM can create lots of anon VM_EXEC pages,
1989 * so we ignore them here.
1991 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1992 list_add(&page
->lru
, &l_active
);
1997 ClearPageActive(page
); /* we are de-activating */
1998 list_add(&page
->lru
, &l_inactive
);
2002 * Move pages back to the lru list.
2004 spin_lock_irq(&pgdat
->lru_lock
);
2006 * Count referenced pages from currently used mappings as rotated,
2007 * even though only some of them are actually re-activated. This
2008 * helps balance scan pressure between file and anonymous pages in
2011 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2013 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2014 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2015 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2016 spin_unlock_irq(&pgdat
->lru_lock
);
2018 mem_cgroup_uncharge_list(&l_hold
);
2019 free_hot_cold_page_list(&l_hold
, true);
2020 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2021 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2025 * The inactive anon list should be small enough that the VM never has
2026 * to do too much work.
2028 * The inactive file list should be small enough to leave most memory
2029 * to the established workingset on the scan-resistant active list,
2030 * but large enough to avoid thrashing the aggregate readahead window.
2032 * Both inactive lists should also be large enough that each inactive
2033 * page has a chance to be referenced again before it is reclaimed.
2035 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2036 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2037 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2040 * memory ratio inactive
2041 * -------------------------------------
2050 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2051 struct scan_control
*sc
, bool trace
)
2053 unsigned long inactive_ratio
;
2054 unsigned long inactive
, active
;
2055 enum lru_list inactive_lru
= file
* LRU_FILE
;
2056 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2060 * If we don't have swap space, anonymous page deactivation
2063 if (!file
&& !total_swap_pages
)
2066 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2067 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2069 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2071 inactive_ratio
= int_sqrt(10 * gb
);
2076 trace_mm_vmscan_inactive_list_is_low(lruvec_pgdat(lruvec
)->node_id
,
2078 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2079 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2080 inactive_ratio
, file
);
2082 return inactive
* inactive_ratio
< active
;
2085 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2086 struct lruvec
*lruvec
, struct scan_control
*sc
)
2088 if (is_active_lru(lru
)) {
2089 if (inactive_list_is_low(lruvec
, is_file_lru(lru
), sc
, true))
2090 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2094 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2105 * Determine how aggressively the anon and file LRU lists should be
2106 * scanned. The relative value of each set of LRU lists is determined
2107 * by looking at the fraction of the pages scanned we did rotate back
2108 * onto the active list instead of evict.
2110 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2111 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2113 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2114 struct scan_control
*sc
, unsigned long *nr
,
2115 unsigned long *lru_pages
)
2117 int swappiness
= mem_cgroup_swappiness(memcg
);
2118 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2120 u64 denominator
= 0; /* gcc */
2121 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2122 unsigned long anon_prio
, file_prio
;
2123 enum scan_balance scan_balance
;
2124 unsigned long anon
, file
;
2125 bool force_scan
= false;
2126 unsigned long ap
, fp
;
2132 * If the zone or memcg is small, nr[l] can be 0. This
2133 * results in no scanning on this priority and a potential
2134 * priority drop. Global direct reclaim can go to the next
2135 * zone and tends to have no problems. Global kswapd is for
2136 * zone balancing and it needs to scan a minimum amount. When
2137 * reclaiming for a memcg, a priority drop can cause high
2138 * latencies, so it's better to scan a minimum amount there as
2141 if (current_is_kswapd()) {
2142 if (!pgdat_reclaimable(pgdat
))
2144 if (!mem_cgroup_online(memcg
))
2147 if (!global_reclaim(sc
))
2150 /* If we have no swap space, do not bother scanning anon pages. */
2151 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2152 scan_balance
= SCAN_FILE
;
2157 * Global reclaim will swap to prevent OOM even with no
2158 * swappiness, but memcg users want to use this knob to
2159 * disable swapping for individual groups completely when
2160 * using the memory controller's swap limit feature would be
2163 if (!global_reclaim(sc
) && !swappiness
) {
2164 scan_balance
= SCAN_FILE
;
2169 * Do not apply any pressure balancing cleverness when the
2170 * system is close to OOM, scan both anon and file equally
2171 * (unless the swappiness setting disagrees with swapping).
2173 if (!sc
->priority
&& swappiness
) {
2174 scan_balance
= SCAN_EQUAL
;
2179 * Prevent the reclaimer from falling into the cache trap: as
2180 * cache pages start out inactive, every cache fault will tip
2181 * the scan balance towards the file LRU. And as the file LRU
2182 * shrinks, so does the window for rotation from references.
2183 * This means we have a runaway feedback loop where a tiny
2184 * thrashing file LRU becomes infinitely more attractive than
2185 * anon pages. Try to detect this based on file LRU size.
2187 if (global_reclaim(sc
)) {
2188 unsigned long pgdatfile
;
2189 unsigned long pgdatfree
;
2191 unsigned long total_high_wmark
= 0;
2193 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2194 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2195 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2197 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2198 struct zone
*zone
= &pgdat
->node_zones
[z
];
2199 if (!managed_zone(zone
))
2202 total_high_wmark
+= high_wmark_pages(zone
);
2205 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2206 scan_balance
= SCAN_ANON
;
2212 * If there is enough inactive page cache, i.e. if the size of the
2213 * inactive list is greater than that of the active list *and* the
2214 * inactive list actually has some pages to scan on this priority, we
2215 * do not reclaim anything from the anonymous working set right now.
2216 * Without the second condition we could end up never scanning an
2217 * lruvec even if it has plenty of old anonymous pages unless the
2218 * system is under heavy pressure.
2220 if (!inactive_list_is_low(lruvec
, true, sc
, false) &&
2221 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2222 scan_balance
= SCAN_FILE
;
2226 scan_balance
= SCAN_FRACT
;
2229 * With swappiness at 100, anonymous and file have the same priority.
2230 * This scanning priority is essentially the inverse of IO cost.
2232 anon_prio
= swappiness
;
2233 file_prio
= 200 - anon_prio
;
2236 * OK, so we have swap space and a fair amount of page cache
2237 * pages. We use the recently rotated / recently scanned
2238 * ratios to determine how valuable each cache is.
2240 * Because workloads change over time (and to avoid overflow)
2241 * we keep these statistics as a floating average, which ends
2242 * up weighing recent references more than old ones.
2244 * anon in [0], file in [1]
2247 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2248 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2249 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2250 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2252 spin_lock_irq(&pgdat
->lru_lock
);
2253 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2254 reclaim_stat
->recent_scanned
[0] /= 2;
2255 reclaim_stat
->recent_rotated
[0] /= 2;
2258 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2259 reclaim_stat
->recent_scanned
[1] /= 2;
2260 reclaim_stat
->recent_rotated
[1] /= 2;
2264 * The amount of pressure on anon vs file pages is inversely
2265 * proportional to the fraction of recently scanned pages on
2266 * each list that were recently referenced and in active use.
2268 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2269 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2271 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2272 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2273 spin_unlock_irq(&pgdat
->lru_lock
);
2277 denominator
= ap
+ fp
+ 1;
2279 some_scanned
= false;
2280 /* Only use force_scan on second pass. */
2281 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2283 for_each_evictable_lru(lru
) {
2284 int file
= is_file_lru(lru
);
2288 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2289 scan
= size
>> sc
->priority
;
2291 if (!scan
&& pass
&& force_scan
)
2292 scan
= min(size
, SWAP_CLUSTER_MAX
);
2294 switch (scan_balance
) {
2296 /* Scan lists relative to size */
2300 * Scan types proportional to swappiness and
2301 * their relative recent reclaim efficiency.
2303 scan
= div64_u64(scan
* fraction
[file
],
2308 /* Scan one type exclusively */
2309 if ((scan_balance
== SCAN_FILE
) != file
) {
2315 /* Look ma, no brain */
2323 * Skip the second pass and don't force_scan,
2324 * if we found something to scan.
2326 some_scanned
|= !!scan
;
2332 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2334 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2335 struct scan_control
*sc
, unsigned long *lru_pages
)
2337 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2338 unsigned long nr
[NR_LRU_LISTS
];
2339 unsigned long targets
[NR_LRU_LISTS
];
2340 unsigned long nr_to_scan
;
2342 unsigned long nr_reclaimed
= 0;
2343 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2344 struct blk_plug plug
;
2347 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2349 /* Record the original scan target for proportional adjustments later */
2350 memcpy(targets
, nr
, sizeof(nr
));
2353 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2354 * event that can occur when there is little memory pressure e.g.
2355 * multiple streaming readers/writers. Hence, we do not abort scanning
2356 * when the requested number of pages are reclaimed when scanning at
2357 * DEF_PRIORITY on the assumption that the fact we are direct
2358 * reclaiming implies that kswapd is not keeping up and it is best to
2359 * do a batch of work at once. For memcg reclaim one check is made to
2360 * abort proportional reclaim if either the file or anon lru has already
2361 * dropped to zero at the first pass.
2363 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2364 sc
->priority
== DEF_PRIORITY
);
2366 blk_start_plug(&plug
);
2367 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2368 nr
[LRU_INACTIVE_FILE
]) {
2369 unsigned long nr_anon
, nr_file
, percentage
;
2370 unsigned long nr_scanned
;
2372 for_each_evictable_lru(lru
) {
2374 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2375 nr
[lru
] -= nr_to_scan
;
2377 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2384 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2388 * For kswapd and memcg, reclaim at least the number of pages
2389 * requested. Ensure that the anon and file LRUs are scanned
2390 * proportionally what was requested by get_scan_count(). We
2391 * stop reclaiming one LRU and reduce the amount scanning
2392 * proportional to the original scan target.
2394 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2395 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2398 * It's just vindictive to attack the larger once the smaller
2399 * has gone to zero. And given the way we stop scanning the
2400 * smaller below, this makes sure that we only make one nudge
2401 * towards proportionality once we've got nr_to_reclaim.
2403 if (!nr_file
|| !nr_anon
)
2406 if (nr_file
> nr_anon
) {
2407 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2408 targets
[LRU_ACTIVE_ANON
] + 1;
2410 percentage
= nr_anon
* 100 / scan_target
;
2412 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2413 targets
[LRU_ACTIVE_FILE
] + 1;
2415 percentage
= nr_file
* 100 / scan_target
;
2418 /* Stop scanning the smaller of the LRU */
2420 nr
[lru
+ LRU_ACTIVE
] = 0;
2423 * Recalculate the other LRU scan count based on its original
2424 * scan target and the percentage scanning already complete
2426 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2427 nr_scanned
= targets
[lru
] - nr
[lru
];
2428 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2429 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2432 nr_scanned
= targets
[lru
] - nr
[lru
];
2433 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2434 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2436 scan_adjusted
= true;
2438 blk_finish_plug(&plug
);
2439 sc
->nr_reclaimed
+= nr_reclaimed
;
2442 * Even if we did not try to evict anon pages at all, we want to
2443 * rebalance the anon lru active/inactive ratio.
2445 if (inactive_list_is_low(lruvec
, false, sc
, true))
2446 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2447 sc
, LRU_ACTIVE_ANON
);
2450 /* Use reclaim/compaction for costly allocs or under memory pressure */
2451 static bool in_reclaim_compaction(struct scan_control
*sc
)
2453 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2454 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2455 sc
->priority
< DEF_PRIORITY
- 2))
2462 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2463 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2464 * true if more pages should be reclaimed such that when the page allocator
2465 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2466 * It will give up earlier than that if there is difficulty reclaiming pages.
2468 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2469 unsigned long nr_reclaimed
,
2470 unsigned long nr_scanned
,
2471 struct scan_control
*sc
)
2473 unsigned long pages_for_compaction
;
2474 unsigned long inactive_lru_pages
;
2477 /* If not in reclaim/compaction mode, stop */
2478 if (!in_reclaim_compaction(sc
))
2481 /* Consider stopping depending on scan and reclaim activity */
2482 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2484 * For __GFP_REPEAT allocations, stop reclaiming if the
2485 * full LRU list has been scanned and we are still failing
2486 * to reclaim pages. This full LRU scan is potentially
2487 * expensive but a __GFP_REPEAT caller really wants to succeed
2489 if (!nr_reclaimed
&& !nr_scanned
)
2493 * For non-__GFP_REPEAT allocations which can presumably
2494 * fail without consequence, stop if we failed to reclaim
2495 * any pages from the last SWAP_CLUSTER_MAX number of
2496 * pages that were scanned. This will return to the
2497 * caller faster at the risk reclaim/compaction and
2498 * the resulting allocation attempt fails
2505 * If we have not reclaimed enough pages for compaction and the
2506 * inactive lists are large enough, continue reclaiming
2508 pages_for_compaction
= compact_gap(sc
->order
);
2509 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2510 if (get_nr_swap_pages() > 0)
2511 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2512 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2513 inactive_lru_pages
> pages_for_compaction
)
2516 /* If compaction would go ahead or the allocation would succeed, stop */
2517 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2518 struct zone
*zone
= &pgdat
->node_zones
[z
];
2519 if (!managed_zone(zone
))
2522 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2523 case COMPACT_SUCCESS
:
2524 case COMPACT_CONTINUE
:
2527 /* check next zone */
2534 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2536 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2537 unsigned long nr_reclaimed
, nr_scanned
;
2538 bool reclaimable
= false;
2541 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2542 struct mem_cgroup_reclaim_cookie reclaim
= {
2544 .priority
= sc
->priority
,
2546 unsigned long node_lru_pages
= 0;
2547 struct mem_cgroup
*memcg
;
2549 nr_reclaimed
= sc
->nr_reclaimed
;
2550 nr_scanned
= sc
->nr_scanned
;
2552 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2554 unsigned long lru_pages
;
2555 unsigned long reclaimed
;
2556 unsigned long scanned
;
2558 if (mem_cgroup_low(root
, memcg
)) {
2559 if (!sc
->may_thrash
)
2561 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2564 reclaimed
= sc
->nr_reclaimed
;
2565 scanned
= sc
->nr_scanned
;
2567 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2568 node_lru_pages
+= lru_pages
;
2571 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2572 memcg
, sc
->nr_scanned
- scanned
,
2575 /* Record the group's reclaim efficiency */
2576 vmpressure(sc
->gfp_mask
, memcg
, false,
2577 sc
->nr_scanned
- scanned
,
2578 sc
->nr_reclaimed
- reclaimed
);
2581 * Direct reclaim and kswapd have to scan all memory
2582 * cgroups to fulfill the overall scan target for the
2585 * Limit reclaim, on the other hand, only cares about
2586 * nr_to_reclaim pages to be reclaimed and it will
2587 * retry with decreasing priority if one round over the
2588 * whole hierarchy is not sufficient.
2590 if (!global_reclaim(sc
) &&
2591 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2592 mem_cgroup_iter_break(root
, memcg
);
2595 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2598 * Shrink the slab caches in the same proportion that
2599 * the eligible LRU pages were scanned.
2601 if (global_reclaim(sc
))
2602 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2603 sc
->nr_scanned
- nr_scanned
,
2606 if (reclaim_state
) {
2607 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2608 reclaim_state
->reclaimed_slab
= 0;
2611 /* Record the subtree's reclaim efficiency */
2612 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2613 sc
->nr_scanned
- nr_scanned
,
2614 sc
->nr_reclaimed
- nr_reclaimed
);
2616 if (sc
->nr_reclaimed
- nr_reclaimed
)
2619 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2620 sc
->nr_scanned
- nr_scanned
, sc
));
2626 * Returns true if compaction should go ahead for a costly-order request, or
2627 * the allocation would already succeed without compaction. Return false if we
2628 * should reclaim first.
2630 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2632 unsigned long watermark
;
2633 enum compact_result suitable
;
2635 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2636 if (suitable
== COMPACT_SUCCESS
)
2637 /* Allocation should succeed already. Don't reclaim. */
2639 if (suitable
== COMPACT_SKIPPED
)
2640 /* Compaction cannot yet proceed. Do reclaim. */
2644 * Compaction is already possible, but it takes time to run and there
2645 * are potentially other callers using the pages just freed. So proceed
2646 * with reclaim to make a buffer of free pages available to give
2647 * compaction a reasonable chance of completing and allocating the page.
2648 * Note that we won't actually reclaim the whole buffer in one attempt
2649 * as the target watermark in should_continue_reclaim() is lower. But if
2650 * we are already above the high+gap watermark, don't reclaim at all.
2652 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2654 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2658 * This is the direct reclaim path, for page-allocating processes. We only
2659 * try to reclaim pages from zones which will satisfy the caller's allocation
2662 * If a zone is deemed to be full of pinned pages then just give it a light
2663 * scan then give up on it.
2665 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2669 unsigned long nr_soft_reclaimed
;
2670 unsigned long nr_soft_scanned
;
2672 pg_data_t
*last_pgdat
= NULL
;
2675 * If the number of buffer_heads in the machine exceeds the maximum
2676 * allowed level, force direct reclaim to scan the highmem zone as
2677 * highmem pages could be pinning lowmem pages storing buffer_heads
2679 orig_mask
= sc
->gfp_mask
;
2680 if (buffer_heads_over_limit
) {
2681 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2682 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2685 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2686 sc
->reclaim_idx
, sc
->nodemask
) {
2688 * Take care memory controller reclaiming has small influence
2691 if (global_reclaim(sc
)) {
2692 if (!cpuset_zone_allowed(zone
,
2693 GFP_KERNEL
| __GFP_HARDWALL
))
2696 if (sc
->priority
!= DEF_PRIORITY
&&
2697 !pgdat_reclaimable(zone
->zone_pgdat
))
2698 continue; /* Let kswapd poll it */
2701 * If we already have plenty of memory free for
2702 * compaction in this zone, don't free any more.
2703 * Even though compaction is invoked for any
2704 * non-zero order, only frequent costly order
2705 * reclamation is disruptive enough to become a
2706 * noticeable problem, like transparent huge
2709 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2710 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2711 compaction_ready(zone
, sc
)) {
2712 sc
->compaction_ready
= true;
2717 * Shrink each node in the zonelist once. If the
2718 * zonelist is ordered by zone (not the default) then a
2719 * node may be shrunk multiple times but in that case
2720 * the user prefers lower zones being preserved.
2722 if (zone
->zone_pgdat
== last_pgdat
)
2726 * This steals pages from memory cgroups over softlimit
2727 * and returns the number of reclaimed pages and
2728 * scanned pages. This works for global memory pressure
2729 * and balancing, not for a memcg's limit.
2731 nr_soft_scanned
= 0;
2732 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2733 sc
->order
, sc
->gfp_mask
,
2735 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2736 sc
->nr_scanned
+= nr_soft_scanned
;
2737 /* need some check for avoid more shrink_zone() */
2740 /* See comment about same check for global reclaim above */
2741 if (zone
->zone_pgdat
== last_pgdat
)
2743 last_pgdat
= zone
->zone_pgdat
;
2744 shrink_node(zone
->zone_pgdat
, sc
);
2748 * Restore to original mask to avoid the impact on the caller if we
2749 * promoted it to __GFP_HIGHMEM.
2751 sc
->gfp_mask
= orig_mask
;
2755 * This is the main entry point to direct page reclaim.
2757 * If a full scan of the inactive list fails to free enough memory then we
2758 * are "out of memory" and something needs to be killed.
2760 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2761 * high - the zone may be full of dirty or under-writeback pages, which this
2762 * caller can't do much about. We kick the writeback threads and take explicit
2763 * naps in the hope that some of these pages can be written. But if the
2764 * allocating task holds filesystem locks which prevent writeout this might not
2765 * work, and the allocation attempt will fail.
2767 * returns: 0, if no pages reclaimed
2768 * else, the number of pages reclaimed
2770 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2771 struct scan_control
*sc
)
2773 int initial_priority
= sc
->priority
;
2775 delayacct_freepages_start();
2777 if (global_reclaim(sc
))
2778 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2781 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2784 shrink_zones(zonelist
, sc
);
2786 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2789 if (sc
->compaction_ready
)
2793 * If we're getting trouble reclaiming, start doing
2794 * writepage even in laptop mode.
2796 if (sc
->priority
< DEF_PRIORITY
- 2)
2797 sc
->may_writepage
= 1;
2798 } while (--sc
->priority
>= 0);
2800 delayacct_freepages_end();
2802 if (sc
->nr_reclaimed
)
2803 return sc
->nr_reclaimed
;
2805 /* Aborted reclaim to try compaction? don't OOM, then */
2806 if (sc
->compaction_ready
)
2809 /* Untapped cgroup reserves? Don't OOM, retry. */
2810 if (!sc
->may_thrash
) {
2811 sc
->priority
= initial_priority
;
2819 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2822 unsigned long pfmemalloc_reserve
= 0;
2823 unsigned long free_pages
= 0;
2827 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2828 zone
= &pgdat
->node_zones
[i
];
2829 if (!managed_zone(zone
) ||
2830 pgdat_reclaimable_pages(pgdat
) == 0)
2833 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2834 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2837 /* If there are no reserves (unexpected config) then do not throttle */
2838 if (!pfmemalloc_reserve
)
2841 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2843 /* kswapd must be awake if processes are being throttled */
2844 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2845 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2846 (enum zone_type
)ZONE_NORMAL
);
2847 wake_up_interruptible(&pgdat
->kswapd_wait
);
2854 * Throttle direct reclaimers if backing storage is backed by the network
2855 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2856 * depleted. kswapd will continue to make progress and wake the processes
2857 * when the low watermark is reached.
2859 * Returns true if a fatal signal was delivered during throttling. If this
2860 * happens, the page allocator should not consider triggering the OOM killer.
2862 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2863 nodemask_t
*nodemask
)
2867 pg_data_t
*pgdat
= NULL
;
2870 * Kernel threads should not be throttled as they may be indirectly
2871 * responsible for cleaning pages necessary for reclaim to make forward
2872 * progress. kjournald for example may enter direct reclaim while
2873 * committing a transaction where throttling it could forcing other
2874 * processes to block on log_wait_commit().
2876 if (current
->flags
& PF_KTHREAD
)
2880 * If a fatal signal is pending, this process should not throttle.
2881 * It should return quickly so it can exit and free its memory
2883 if (fatal_signal_pending(current
))
2887 * Check if the pfmemalloc reserves are ok by finding the first node
2888 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2889 * GFP_KERNEL will be required for allocating network buffers when
2890 * swapping over the network so ZONE_HIGHMEM is unusable.
2892 * Throttling is based on the first usable node and throttled processes
2893 * wait on a queue until kswapd makes progress and wakes them. There
2894 * is an affinity then between processes waking up and where reclaim
2895 * progress has been made assuming the process wakes on the same node.
2896 * More importantly, processes running on remote nodes will not compete
2897 * for remote pfmemalloc reserves and processes on different nodes
2898 * should make reasonable progress.
2900 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2901 gfp_zone(gfp_mask
), nodemask
) {
2902 if (zone_idx(zone
) > ZONE_NORMAL
)
2905 /* Throttle based on the first usable node */
2906 pgdat
= zone
->zone_pgdat
;
2907 if (pfmemalloc_watermark_ok(pgdat
))
2912 /* If no zone was usable by the allocation flags then do not throttle */
2916 /* Account for the throttling */
2917 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2920 * If the caller cannot enter the filesystem, it's possible that it
2921 * is due to the caller holding an FS lock or performing a journal
2922 * transaction in the case of a filesystem like ext[3|4]. In this case,
2923 * it is not safe to block on pfmemalloc_wait as kswapd could be
2924 * blocked waiting on the same lock. Instead, throttle for up to a
2925 * second before continuing.
2927 if (!(gfp_mask
& __GFP_FS
)) {
2928 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2929 pfmemalloc_watermark_ok(pgdat
), HZ
);
2934 /* Throttle until kswapd wakes the process */
2935 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2936 pfmemalloc_watermark_ok(pgdat
));
2939 if (fatal_signal_pending(current
))
2946 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2947 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2949 unsigned long nr_reclaimed
;
2950 struct scan_control sc
= {
2951 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2952 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2953 .reclaim_idx
= gfp_zone(gfp_mask
),
2955 .nodemask
= nodemask
,
2956 .priority
= DEF_PRIORITY
,
2957 .may_writepage
= !laptop_mode
,
2963 * Do not enter reclaim if fatal signal was delivered while throttled.
2964 * 1 is returned so that the page allocator does not OOM kill at this
2967 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2970 trace_mm_vmscan_direct_reclaim_begin(order
,
2975 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2977 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2979 return nr_reclaimed
;
2984 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
2985 gfp_t gfp_mask
, bool noswap
,
2987 unsigned long *nr_scanned
)
2989 struct scan_control sc
= {
2990 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2991 .target_mem_cgroup
= memcg
,
2992 .may_writepage
= !laptop_mode
,
2994 .reclaim_idx
= MAX_NR_ZONES
- 1,
2995 .may_swap
= !noswap
,
2997 unsigned long lru_pages
;
2999 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3000 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3002 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3008 * NOTE: Although we can get the priority field, using it
3009 * here is not a good idea, since it limits the pages we can scan.
3010 * if we don't reclaim here, the shrink_node from balance_pgdat
3011 * will pick up pages from other mem cgroup's as well. We hack
3012 * the priority and make it zero.
3014 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3016 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3018 *nr_scanned
= sc
.nr_scanned
;
3019 return sc
.nr_reclaimed
;
3022 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3023 unsigned long nr_pages
,
3027 struct zonelist
*zonelist
;
3028 unsigned long nr_reclaimed
;
3030 struct scan_control sc
= {
3031 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3032 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3033 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3034 .reclaim_idx
= MAX_NR_ZONES
- 1,
3035 .target_mem_cgroup
= memcg
,
3036 .priority
= DEF_PRIORITY
,
3037 .may_writepage
= !laptop_mode
,
3039 .may_swap
= may_swap
,
3043 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3044 * take care of from where we get pages. So the node where we start the
3045 * scan does not need to be the current node.
3047 nid
= mem_cgroup_select_victim_node(memcg
);
3049 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3051 trace_mm_vmscan_memcg_reclaim_begin(0,
3056 current
->flags
|= PF_MEMALLOC
;
3057 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3058 current
->flags
&= ~PF_MEMALLOC
;
3060 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3062 return nr_reclaimed
;
3066 static void age_active_anon(struct pglist_data
*pgdat
,
3067 struct scan_control
*sc
)
3069 struct mem_cgroup
*memcg
;
3071 if (!total_swap_pages
)
3074 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3076 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3078 if (inactive_list_is_low(lruvec
, false, sc
, true))
3079 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3080 sc
, LRU_ACTIVE_ANON
);
3082 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3086 static bool zone_balanced(struct zone
*zone
, int order
, int classzone_idx
)
3088 unsigned long mark
= high_wmark_pages(zone
);
3090 if (!zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3094 * If any eligible zone is balanced then the node is not considered
3095 * to be congested or dirty
3097 clear_bit(PGDAT_CONGESTED
, &zone
->zone_pgdat
->flags
);
3098 clear_bit(PGDAT_DIRTY
, &zone
->zone_pgdat
->flags
);
3099 clear_bit(PGDAT_WRITEBACK
, &zone
->zone_pgdat
->flags
);
3105 * Prepare kswapd for sleeping. This verifies that there are no processes
3106 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3108 * Returns true if kswapd is ready to sleep
3110 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3115 * The throttled processes are normally woken up in balance_pgdat() as
3116 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3117 * race between when kswapd checks the watermarks and a process gets
3118 * throttled. There is also a potential race if processes get
3119 * throttled, kswapd wakes, a large process exits thereby balancing the
3120 * zones, which causes kswapd to exit balance_pgdat() before reaching
3121 * the wake up checks. If kswapd is going to sleep, no process should
3122 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3123 * the wake up is premature, processes will wake kswapd and get
3124 * throttled again. The difference from wake ups in balance_pgdat() is
3125 * that here we are under prepare_to_wait().
3127 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3128 wake_up_all(&pgdat
->pfmemalloc_wait
);
3130 for (i
= 0; i
<= classzone_idx
; i
++) {
3131 struct zone
*zone
= pgdat
->node_zones
+ i
;
3133 if (!managed_zone(zone
))
3136 if (!zone_balanced(zone
, order
, classzone_idx
))
3144 * kswapd shrinks a node of pages that are at or below the highest usable
3145 * zone that is currently unbalanced.
3147 * Returns true if kswapd scanned at least the requested number of pages to
3148 * reclaim or if the lack of progress was due to pages under writeback.
3149 * This is used to determine if the scanning priority needs to be raised.
3151 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3152 struct scan_control
*sc
)
3157 /* Reclaim a number of pages proportional to the number of zones */
3158 sc
->nr_to_reclaim
= 0;
3159 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3160 zone
= pgdat
->node_zones
+ z
;
3161 if (!managed_zone(zone
))
3164 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3168 * Historically care was taken to put equal pressure on all zones but
3169 * now pressure is applied based on node LRU order.
3171 shrink_node(pgdat
, sc
);
3174 * Fragmentation may mean that the system cannot be rebalanced for
3175 * high-order allocations. If twice the allocation size has been
3176 * reclaimed then recheck watermarks only at order-0 to prevent
3177 * excessive reclaim. Assume that a process requested a high-order
3178 * can direct reclaim/compact.
3180 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3183 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3187 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3188 * that are eligible for use by the caller until at least one zone is
3191 * Returns the order kswapd finished reclaiming at.
3193 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3194 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3195 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3196 * or lower is eligible for reclaim until at least one usable zone is
3199 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3202 unsigned long nr_soft_reclaimed
;
3203 unsigned long nr_soft_scanned
;
3205 struct scan_control sc
= {
3206 .gfp_mask
= GFP_KERNEL
,
3208 .priority
= DEF_PRIORITY
,
3209 .may_writepage
= !laptop_mode
,
3213 count_vm_event(PAGEOUTRUN
);
3216 bool raise_priority
= true;
3218 sc
.nr_reclaimed
= 0;
3219 sc
.reclaim_idx
= classzone_idx
;
3222 * If the number of buffer_heads exceeds the maximum allowed
3223 * then consider reclaiming from all zones. This has a dual
3224 * purpose -- on 64-bit systems it is expected that
3225 * buffer_heads are stripped during active rotation. On 32-bit
3226 * systems, highmem pages can pin lowmem memory and shrinking
3227 * buffers can relieve lowmem pressure. Reclaim may still not
3228 * go ahead if all eligible zones for the original allocation
3229 * request are balanced to avoid excessive reclaim from kswapd.
3231 if (buffer_heads_over_limit
) {
3232 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3233 zone
= pgdat
->node_zones
+ i
;
3234 if (!managed_zone(zone
))
3243 * Only reclaim if there are no eligible zones. Check from
3244 * high to low zone as allocations prefer higher zones.
3245 * Scanning from low to high zone would allow congestion to be
3246 * cleared during a very small window when a small low
3247 * zone was balanced even under extreme pressure when the
3248 * overall node may be congested. Note that sc.reclaim_idx
3249 * is not used as buffer_heads_over_limit may have adjusted
3252 for (i
= classzone_idx
; i
>= 0; i
--) {
3253 zone
= pgdat
->node_zones
+ i
;
3254 if (!managed_zone(zone
))
3257 if (zone_balanced(zone
, sc
.order
, classzone_idx
))
3262 * Do some background aging of the anon list, to give
3263 * pages a chance to be referenced before reclaiming. All
3264 * pages are rotated regardless of classzone as this is
3265 * about consistent aging.
3267 age_active_anon(pgdat
, &sc
);
3270 * If we're getting trouble reclaiming, start doing writepage
3271 * even in laptop mode.
3273 if (sc
.priority
< DEF_PRIORITY
- 2 || !pgdat_reclaimable(pgdat
))
3274 sc
.may_writepage
= 1;
3276 /* Call soft limit reclaim before calling shrink_node. */
3278 nr_soft_scanned
= 0;
3279 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3280 sc
.gfp_mask
, &nr_soft_scanned
);
3281 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3284 * There should be no need to raise the scanning priority if
3285 * enough pages are already being scanned that that high
3286 * watermark would be met at 100% efficiency.
3288 if (kswapd_shrink_node(pgdat
, &sc
))
3289 raise_priority
= false;
3292 * If the low watermark is met there is no need for processes
3293 * to be throttled on pfmemalloc_wait as they should not be
3294 * able to safely make forward progress. Wake them
3296 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3297 pfmemalloc_watermark_ok(pgdat
))
3298 wake_up_all(&pgdat
->pfmemalloc_wait
);
3300 /* Check if kswapd should be suspending */
3301 if (try_to_freeze() || kthread_should_stop())
3305 * Raise priority if scanning rate is too low or there was no
3306 * progress in reclaiming pages
3308 if (raise_priority
|| !sc
.nr_reclaimed
)
3310 } while (sc
.priority
>= 1);
3314 * Return the order kswapd stopped reclaiming at as
3315 * prepare_kswapd_sleep() takes it into account. If another caller
3316 * entered the allocator slow path while kswapd was awake, order will
3317 * remain at the higher level.
3322 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3323 unsigned int classzone_idx
)
3328 if (freezing(current
) || kthread_should_stop())
3331 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3333 /* Try to sleep for a short interval */
3334 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3336 * Compaction records what page blocks it recently failed to
3337 * isolate pages from and skips them in the future scanning.
3338 * When kswapd is going to sleep, it is reasonable to assume
3339 * that pages and compaction may succeed so reset the cache.
3341 reset_isolation_suitable(pgdat
);
3344 * We have freed the memory, now we should compact it to make
3345 * allocation of the requested order possible.
3347 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3349 remaining
= schedule_timeout(HZ
/10);
3352 * If woken prematurely then reset kswapd_classzone_idx and
3353 * order. The values will either be from a wakeup request or
3354 * the previous request that slept prematurely.
3357 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3358 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3361 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3362 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3366 * After a short sleep, check if it was a premature sleep. If not, then
3367 * go fully to sleep until explicitly woken up.
3370 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3371 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3374 * vmstat counters are not perfectly accurate and the estimated
3375 * value for counters such as NR_FREE_PAGES can deviate from the
3376 * true value by nr_online_cpus * threshold. To avoid the zone
3377 * watermarks being breached while under pressure, we reduce the
3378 * per-cpu vmstat threshold while kswapd is awake and restore
3379 * them before going back to sleep.
3381 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3383 if (!kthread_should_stop())
3386 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3389 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3391 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3393 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3397 * The background pageout daemon, started as a kernel thread
3398 * from the init process.
3400 * This basically trickles out pages so that we have _some_
3401 * free memory available even if there is no other activity
3402 * that frees anything up. This is needed for things like routing
3403 * etc, where we otherwise might have all activity going on in
3404 * asynchronous contexts that cannot page things out.
3406 * If there are applications that are active memory-allocators
3407 * (most normal use), this basically shouldn't matter.
3409 static int kswapd(void *p
)
3411 unsigned int alloc_order
, reclaim_order
, classzone_idx
;
3412 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3413 struct task_struct
*tsk
= current
;
3415 struct reclaim_state reclaim_state
= {
3416 .reclaimed_slab
= 0,
3418 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3420 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3422 if (!cpumask_empty(cpumask
))
3423 set_cpus_allowed_ptr(tsk
, cpumask
);
3424 current
->reclaim_state
= &reclaim_state
;
3427 * Tell the memory management that we're a "memory allocator",
3428 * and that if we need more memory we should get access to it
3429 * regardless (see "__alloc_pages()"). "kswapd" should
3430 * never get caught in the normal page freeing logic.
3432 * (Kswapd normally doesn't need memory anyway, but sometimes
3433 * you need a small amount of memory in order to be able to
3434 * page out something else, and this flag essentially protects
3435 * us from recursively trying to free more memory as we're
3436 * trying to free the first piece of memory in the first place).
3438 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3441 pgdat
->kswapd_order
= alloc_order
= reclaim_order
= 0;
3442 pgdat
->kswapd_classzone_idx
= classzone_idx
= 0;
3447 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3450 /* Read the new order and classzone_idx */
3451 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3452 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3453 pgdat
->kswapd_order
= 0;
3454 pgdat
->kswapd_classzone_idx
= 0;
3456 ret
= try_to_freeze();
3457 if (kthread_should_stop())
3461 * We can speed up thawing tasks if we don't call balance_pgdat
3462 * after returning from the refrigerator
3468 * Reclaim begins at the requested order but if a high-order
3469 * reclaim fails then kswapd falls back to reclaiming for
3470 * order-0. If that happens, kswapd will consider sleeping
3471 * for the order it finished reclaiming at (reclaim_order)
3472 * but kcompactd is woken to compact for the original
3473 * request (alloc_order).
3475 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3477 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3478 if (reclaim_order
< alloc_order
)
3479 goto kswapd_try_sleep
;
3481 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3482 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3485 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3486 current
->reclaim_state
= NULL
;
3487 lockdep_clear_current_reclaim_state();
3493 * A zone is low on free memory, so wake its kswapd task to service it.
3495 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3500 if (!managed_zone(zone
))
3503 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3505 pgdat
= zone
->zone_pgdat
;
3506 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3507 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3508 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3511 /* Only wake kswapd if all zones are unbalanced */
3512 for (z
= 0; z
<= classzone_idx
; z
++) {
3513 zone
= pgdat
->node_zones
+ z
;
3514 if (!managed_zone(zone
))
3517 if (zone_balanced(zone
, order
, classzone_idx
))
3521 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3522 wake_up_interruptible(&pgdat
->kswapd_wait
);
3525 #ifdef CONFIG_HIBERNATION
3527 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3530 * Rather than trying to age LRUs the aim is to preserve the overall
3531 * LRU order by reclaiming preferentially
3532 * inactive > active > active referenced > active mapped
3534 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3536 struct reclaim_state reclaim_state
;
3537 struct scan_control sc
= {
3538 .nr_to_reclaim
= nr_to_reclaim
,
3539 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3540 .reclaim_idx
= MAX_NR_ZONES
- 1,
3541 .priority
= DEF_PRIORITY
,
3545 .hibernation_mode
= 1,
3547 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3548 struct task_struct
*p
= current
;
3549 unsigned long nr_reclaimed
;
3551 p
->flags
|= PF_MEMALLOC
;
3552 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3553 reclaim_state
.reclaimed_slab
= 0;
3554 p
->reclaim_state
= &reclaim_state
;
3556 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3558 p
->reclaim_state
= NULL
;
3559 lockdep_clear_current_reclaim_state();
3560 p
->flags
&= ~PF_MEMALLOC
;
3562 return nr_reclaimed
;
3564 #endif /* CONFIG_HIBERNATION */
3566 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3567 not required for correctness. So if the last cpu in a node goes
3568 away, we get changed to run anywhere: as the first one comes back,
3569 restore their cpu bindings. */
3570 static int kswapd_cpu_online(unsigned int cpu
)
3574 for_each_node_state(nid
, N_MEMORY
) {
3575 pg_data_t
*pgdat
= NODE_DATA(nid
);
3576 const struct cpumask
*mask
;
3578 mask
= cpumask_of_node(pgdat
->node_id
);
3580 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3581 /* One of our CPUs online: restore mask */
3582 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3588 * This kswapd start function will be called by init and node-hot-add.
3589 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3591 int kswapd_run(int nid
)
3593 pg_data_t
*pgdat
= NODE_DATA(nid
);
3599 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3600 if (IS_ERR(pgdat
->kswapd
)) {
3601 /* failure at boot is fatal */
3602 BUG_ON(system_state
== SYSTEM_BOOTING
);
3603 pr_err("Failed to start kswapd on node %d\n", nid
);
3604 ret
= PTR_ERR(pgdat
->kswapd
);
3605 pgdat
->kswapd
= NULL
;
3611 * Called by memory hotplug when all memory in a node is offlined. Caller must
3612 * hold mem_hotplug_begin/end().
3614 void kswapd_stop(int nid
)
3616 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3619 kthread_stop(kswapd
);
3620 NODE_DATA(nid
)->kswapd
= NULL
;
3624 static int __init
kswapd_init(void)
3629 for_each_node_state(nid
, N_MEMORY
)
3631 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3632 "mm/vmscan:online", kswapd_cpu_online
,
3638 module_init(kswapd_init
)
3644 * If non-zero call node_reclaim when the number of free pages falls below
3647 int node_reclaim_mode __read_mostly
;
3649 #define RECLAIM_OFF 0
3650 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3651 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3652 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3655 * Priority for NODE_RECLAIM. This determines the fraction of pages
3656 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3659 #define NODE_RECLAIM_PRIORITY 4
3662 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3665 int sysctl_min_unmapped_ratio
= 1;
3668 * If the number of slab pages in a zone grows beyond this percentage then
3669 * slab reclaim needs to occur.
3671 int sysctl_min_slab_ratio
= 5;
3673 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3675 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3676 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3677 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3680 * It's possible for there to be more file mapped pages than
3681 * accounted for by the pages on the file LRU lists because
3682 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3684 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3687 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3688 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3690 unsigned long nr_pagecache_reclaimable
;
3691 unsigned long delta
= 0;
3694 * If RECLAIM_UNMAP is set, then all file pages are considered
3695 * potentially reclaimable. Otherwise, we have to worry about
3696 * pages like swapcache and node_unmapped_file_pages() provides
3699 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3700 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3702 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3704 /* If we can't clean pages, remove dirty pages from consideration */
3705 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3706 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3708 /* Watch for any possible underflows due to delta */
3709 if (unlikely(delta
> nr_pagecache_reclaimable
))
3710 delta
= nr_pagecache_reclaimable
;
3712 return nr_pagecache_reclaimable
- delta
;
3716 * Try to free up some pages from this node through reclaim.
3718 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3720 /* Minimum pages needed in order to stay on node */
3721 const unsigned long nr_pages
= 1 << order
;
3722 struct task_struct
*p
= current
;
3723 struct reclaim_state reclaim_state
;
3724 int classzone_idx
= gfp_zone(gfp_mask
);
3725 struct scan_control sc
= {
3726 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3727 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3729 .priority
= NODE_RECLAIM_PRIORITY
,
3730 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3731 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3733 .reclaim_idx
= classzone_idx
,
3738 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3739 * and we also need to be able to write out pages for RECLAIM_WRITE
3740 * and RECLAIM_UNMAP.
3742 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3743 lockdep_set_current_reclaim_state(gfp_mask
);
3744 reclaim_state
.reclaimed_slab
= 0;
3745 p
->reclaim_state
= &reclaim_state
;
3747 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3749 * Free memory by calling shrink zone with increasing
3750 * priorities until we have enough memory freed.
3753 shrink_node(pgdat
, &sc
);
3754 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3757 p
->reclaim_state
= NULL
;
3758 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3759 lockdep_clear_current_reclaim_state();
3760 return sc
.nr_reclaimed
>= nr_pages
;
3763 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3768 * Node reclaim reclaims unmapped file backed pages and
3769 * slab pages if we are over the defined limits.
3771 * A small portion of unmapped file backed pages is needed for
3772 * file I/O otherwise pages read by file I/O will be immediately
3773 * thrown out if the node is overallocated. So we do not reclaim
3774 * if less than a specified percentage of the node is used by
3775 * unmapped file backed pages.
3777 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3778 sum_zone_node_page_state(pgdat
->node_id
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3779 return NODE_RECLAIM_FULL
;
3781 if (!pgdat_reclaimable(pgdat
))
3782 return NODE_RECLAIM_FULL
;
3785 * Do not scan if the allocation should not be delayed.
3787 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3788 return NODE_RECLAIM_NOSCAN
;
3791 * Only run node reclaim on the local node or on nodes that do not
3792 * have associated processors. This will favor the local processor
3793 * over remote processors and spread off node memory allocations
3794 * as wide as possible.
3796 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3797 return NODE_RECLAIM_NOSCAN
;
3799 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3800 return NODE_RECLAIM_NOSCAN
;
3802 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3803 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3806 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3813 * page_evictable - test whether a page is evictable
3814 * @page: the page to test
3816 * Test whether page is evictable--i.e., should be placed on active/inactive
3817 * lists vs unevictable list.
3819 * Reasons page might not be evictable:
3820 * (1) page's mapping marked unevictable
3821 * (2) page is part of an mlocked VMA
3824 int page_evictable(struct page
*page
)
3826 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3831 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3832 * @pages: array of pages to check
3833 * @nr_pages: number of pages to check
3835 * Checks pages for evictability and moves them to the appropriate lru list.
3837 * This function is only used for SysV IPC SHM_UNLOCK.
3839 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3841 struct lruvec
*lruvec
;
3842 struct pglist_data
*pgdat
= NULL
;
3847 for (i
= 0; i
< nr_pages
; i
++) {
3848 struct page
*page
= pages
[i
];
3849 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3852 if (pagepgdat
!= pgdat
) {
3854 spin_unlock_irq(&pgdat
->lru_lock
);
3856 spin_lock_irq(&pgdat
->lru_lock
);
3858 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
3860 if (!PageLRU(page
) || !PageUnevictable(page
))
3863 if (page_evictable(page
)) {
3864 enum lru_list lru
= page_lru_base_type(page
);
3866 VM_BUG_ON_PAGE(PageActive(page
), page
);
3867 ClearPageUnevictable(page
);
3868 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3869 add_page_to_lru_list(page
, lruvec
, lru
);
3875 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3876 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3877 spin_unlock_irq(&pgdat
->lru_lock
);
3880 #endif /* CONFIG_SHMEM */