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 unsigned int may_writepage
:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap
:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap
:1;
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash
:1;
101 unsigned int hibernation_mode
:1;
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready
:1;
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned
;
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed
;
113 #ifdef ARCH_HAS_PREFETCH
114 #define prefetch_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
130 if ((_page)->lru.prev != _base) { \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
142 * From 0 .. 100. Higher means more swappy.
144 int vm_swappiness
= 60;
146 * The total number of pages which are beyond the high watermark within all
149 unsigned long vm_total_pages
;
151 static LIST_HEAD(shrinker_list
);
152 static DECLARE_RWSEM(shrinker_rwsem
);
155 static bool global_reclaim(struct scan_control
*sc
)
157 return !sc
->target_mem_cgroup
;
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
173 static bool sane_reclaim(struct scan_control
*sc
)
175 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
186 static bool global_reclaim(struct scan_control
*sc
)
191 static bool sane_reclaim(struct scan_control
*sc
)
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
202 unsigned long zone_reclaimable_pages(struct zone
*zone
)
206 nr
= zone_page_state_snapshot(zone
, NR_ZONE_LRU_FILE
);
207 if (get_nr_swap_pages() > 0)
208 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_LRU_ANON
);
213 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
217 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
218 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
219 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
221 if (get_nr_swap_pages() > 0)
222 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
223 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
224 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
229 bool pgdat_reclaimable(struct pglist_data
*pgdat
)
231 return node_page_state_snapshot(pgdat
, NR_PAGES_SCANNED
) <
232 pgdat_reclaimable_pages(pgdat
) * 6;
235 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
237 if (!mem_cgroup_disabled())
238 return mem_cgroup_get_lru_size(lruvec
, lru
);
240 return node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
244 * Add a shrinker callback to be called from the vm.
246 int register_shrinker(struct shrinker
*shrinker
)
248 size_t size
= sizeof(*shrinker
->nr_deferred
);
250 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
253 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
254 if (!shrinker
->nr_deferred
)
257 down_write(&shrinker_rwsem
);
258 list_add_tail(&shrinker
->list
, &shrinker_list
);
259 up_write(&shrinker_rwsem
);
262 EXPORT_SYMBOL(register_shrinker
);
267 void unregister_shrinker(struct shrinker
*shrinker
)
269 down_write(&shrinker_rwsem
);
270 list_del(&shrinker
->list
);
271 up_write(&shrinker_rwsem
);
272 kfree(shrinker
->nr_deferred
);
274 EXPORT_SYMBOL(unregister_shrinker
);
276 #define SHRINK_BATCH 128
278 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
279 struct shrinker
*shrinker
,
280 unsigned long nr_scanned
,
281 unsigned long nr_eligible
)
283 unsigned long freed
= 0;
284 unsigned long long delta
;
289 int nid
= shrinkctl
->nid
;
290 long batch_size
= shrinker
->batch
? shrinker
->batch
293 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
298 * copy the current shrinker scan count into a local variable
299 * and zero it so that other concurrent shrinker invocations
300 * don't also do this scanning work.
302 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
305 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
307 do_div(delta
, nr_eligible
+ 1);
309 if (total_scan
< 0) {
310 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
311 shrinker
->scan_objects
, total_scan
);
312 total_scan
= freeable
;
316 * We need to avoid excessive windup on filesystem shrinkers
317 * due to large numbers of GFP_NOFS allocations causing the
318 * shrinkers to return -1 all the time. This results in a large
319 * nr being built up so when a shrink that can do some work
320 * comes along it empties the entire cache due to nr >>>
321 * freeable. This is bad for sustaining a working set in
324 * Hence only allow the shrinker to scan the entire cache when
325 * a large delta change is calculated directly.
327 if (delta
< freeable
/ 4)
328 total_scan
= min(total_scan
, freeable
/ 2);
331 * Avoid risking looping forever due to too large nr value:
332 * never try to free more than twice the estimate number of
335 if (total_scan
> freeable
* 2)
336 total_scan
= freeable
* 2;
338 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
339 nr_scanned
, nr_eligible
,
340 freeable
, delta
, total_scan
);
343 * Normally, we should not scan less than batch_size objects in one
344 * pass to avoid too frequent shrinker calls, but if the slab has less
345 * than batch_size objects in total and we are really tight on memory,
346 * we will try to reclaim all available objects, otherwise we can end
347 * up failing allocations although there are plenty of reclaimable
348 * objects spread over several slabs with usage less than the
351 * We detect the "tight on memory" situations by looking at the total
352 * number of objects we want to scan (total_scan). If it is greater
353 * than the total number of objects on slab (freeable), we must be
354 * scanning at high prio and therefore should try to reclaim as much as
357 while (total_scan
>= batch_size
||
358 total_scan
>= freeable
) {
360 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
362 shrinkctl
->nr_to_scan
= nr_to_scan
;
363 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
364 if (ret
== SHRINK_STOP
)
368 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
369 total_scan
-= nr_to_scan
;
375 * move the unused scan count back into the shrinker in a
376 * manner that handles concurrent updates. If we exhausted the
377 * scan, there is no need to do an update.
380 new_nr
= atomic_long_add_return(total_scan
,
381 &shrinker
->nr_deferred
[nid
]);
383 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
385 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
390 * shrink_slab - shrink slab caches
391 * @gfp_mask: allocation context
392 * @nid: node whose slab caches to target
393 * @memcg: memory cgroup whose slab caches to target
394 * @nr_scanned: pressure numerator
395 * @nr_eligible: pressure denominator
397 * Call the shrink functions to age shrinkable caches.
399 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
400 * unaware shrinkers will receive a node id of 0 instead.
402 * @memcg specifies the memory cgroup to target. If it is not NULL,
403 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
404 * objects from the memory cgroup specified. Otherwise, only unaware
405 * shrinkers are called.
407 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
408 * the available objects should be scanned. Page reclaim for example
409 * passes the number of pages scanned and the number of pages on the
410 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
411 * when it encountered mapped pages. The ratio is further biased by
412 * the ->seeks setting of the shrink function, which indicates the
413 * cost to recreate an object relative to that of an LRU page.
415 * Returns the number of reclaimed slab objects.
417 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
418 struct mem_cgroup
*memcg
,
419 unsigned long nr_scanned
,
420 unsigned long nr_eligible
)
422 struct shrinker
*shrinker
;
423 unsigned long freed
= 0;
425 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
429 nr_scanned
= SWAP_CLUSTER_MAX
;
431 if (!down_read_trylock(&shrinker_rwsem
)) {
433 * If we would return 0, our callers would understand that we
434 * have nothing else to shrink and give up trying. By returning
435 * 1 we keep it going and assume we'll be able to shrink next
442 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
443 struct shrink_control sc
= {
444 .gfp_mask
= gfp_mask
,
450 * If kernel memory accounting is disabled, we ignore
451 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
452 * passing NULL for memcg.
454 if (memcg_kmem_enabled() &&
455 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
458 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
461 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
464 up_read(&shrinker_rwsem
);
470 void drop_slab_node(int nid
)
475 struct mem_cgroup
*memcg
= NULL
;
479 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
481 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
482 } while (freed
> 10);
489 for_each_online_node(nid
)
493 static inline int is_page_cache_freeable(struct page
*page
)
496 * A freeable page cache page is referenced only by the caller
497 * that isolated the page, the page cache radix tree and
498 * optional buffer heads at page->private.
500 return page_count(page
) - page_has_private(page
) == 2;
503 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
505 if (current
->flags
& PF_SWAPWRITE
)
507 if (!inode_write_congested(inode
))
509 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
515 * We detected a synchronous write error writing a page out. Probably
516 * -ENOSPC. We need to propagate that into the address_space for a subsequent
517 * fsync(), msync() or close().
519 * The tricky part is that after writepage we cannot touch the mapping: nothing
520 * prevents it from being freed up. But we have a ref on the page and once
521 * that page is locked, the mapping is pinned.
523 * We're allowed to run sleeping lock_page() here because we know the caller has
526 static void handle_write_error(struct address_space
*mapping
,
527 struct page
*page
, int error
)
530 if (page_mapping(page
) == mapping
)
531 mapping_set_error(mapping
, error
);
535 /* possible outcome of pageout() */
537 /* failed to write page out, page is locked */
539 /* move page to the active list, page is locked */
541 /* page has been sent to the disk successfully, page is unlocked */
543 /* page is clean and locked */
548 * pageout is called by shrink_page_list() for each dirty page.
549 * Calls ->writepage().
551 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
552 struct scan_control
*sc
)
555 * If the page is dirty, only perform writeback if that write
556 * will be non-blocking. To prevent this allocation from being
557 * stalled by pagecache activity. But note that there may be
558 * stalls if we need to run get_block(). We could test
559 * PagePrivate for that.
561 * If this process is currently in __generic_file_write_iter() against
562 * this page's queue, we can perform writeback even if that
565 * If the page is swapcache, write it back even if that would
566 * block, for some throttling. This happens by accident, because
567 * swap_backing_dev_info is bust: it doesn't reflect the
568 * congestion state of the swapdevs. Easy to fix, if needed.
570 if (!is_page_cache_freeable(page
))
574 * Some data journaling orphaned pages can have
575 * page->mapping == NULL while being dirty with clean buffers.
577 if (page_has_private(page
)) {
578 if (try_to_free_buffers(page
)) {
579 ClearPageDirty(page
);
580 pr_info("%s: orphaned page\n", __func__
);
586 if (mapping
->a_ops
->writepage
== NULL
)
587 return PAGE_ACTIVATE
;
588 if (!may_write_to_inode(mapping
->host
, sc
))
591 if (clear_page_dirty_for_io(page
)) {
593 struct writeback_control wbc
= {
594 .sync_mode
= WB_SYNC_NONE
,
595 .nr_to_write
= SWAP_CLUSTER_MAX
,
597 .range_end
= LLONG_MAX
,
601 SetPageReclaim(page
);
602 res
= mapping
->a_ops
->writepage(page
, &wbc
);
604 handle_write_error(mapping
, page
, res
);
605 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
606 ClearPageReclaim(page
);
607 return PAGE_ACTIVATE
;
610 if (!PageWriteback(page
)) {
611 /* synchronous write or broken a_ops? */
612 ClearPageReclaim(page
);
614 trace_mm_vmscan_writepage(page
);
615 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
623 * Same as remove_mapping, but if the page is removed from the mapping, it
624 * gets returned with a refcount of 0.
626 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
631 BUG_ON(!PageLocked(page
));
632 BUG_ON(mapping
!= page_mapping(page
));
634 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
636 * The non racy check for a busy page.
638 * Must be careful with the order of the tests. When someone has
639 * a ref to the page, it may be possible that they dirty it then
640 * drop the reference. So if PageDirty is tested before page_count
641 * here, then the following race may occur:
643 * get_user_pages(&page);
644 * [user mapping goes away]
646 * !PageDirty(page) [good]
647 * SetPageDirty(page);
649 * !page_count(page) [good, discard it]
651 * [oops, our write_to data is lost]
653 * Reversing the order of the tests ensures such a situation cannot
654 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
655 * load is not satisfied before that of page->_refcount.
657 * Note that if SetPageDirty is always performed via set_page_dirty,
658 * and thus under tree_lock, then this ordering is not required.
660 if (!page_ref_freeze(page
, 2))
662 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
663 if (unlikely(PageDirty(page
))) {
664 page_ref_unfreeze(page
, 2);
668 if (PageSwapCache(page
)) {
669 swp_entry_t swap
= { .val
= page_private(page
) };
670 mem_cgroup_swapout(page
, swap
);
671 __delete_from_swap_cache(page
);
672 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
673 swapcache_free(swap
);
675 void (*freepage
)(struct page
*);
678 freepage
= mapping
->a_ops
->freepage
;
680 * Remember a shadow entry for reclaimed file cache in
681 * order to detect refaults, thus thrashing, later on.
683 * But don't store shadows in an address space that is
684 * already exiting. This is not just an optizimation,
685 * inode reclaim needs to empty out the radix tree or
686 * the nodes are lost. Don't plant shadows behind its
689 * We also don't store shadows for DAX mappings because the
690 * only page cache pages found in these are zero pages
691 * covering holes, and because we don't want to mix DAX
692 * exceptional entries and shadow exceptional entries in the
695 if (reclaimed
&& page_is_file_cache(page
) &&
696 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
697 shadow
= workingset_eviction(mapping
, page
);
698 __delete_from_page_cache(page
, shadow
);
699 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
701 if (freepage
!= NULL
)
708 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
713 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
714 * someone else has a ref on the page, abort and return 0. If it was
715 * successfully detached, return 1. Assumes the caller has a single ref on
718 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
720 if (__remove_mapping(mapping
, page
, false)) {
722 * Unfreezing the refcount with 1 rather than 2 effectively
723 * drops the pagecache ref for us without requiring another
726 page_ref_unfreeze(page
, 1);
733 * putback_lru_page - put previously isolated page onto appropriate LRU list
734 * @page: page to be put back to appropriate lru list
736 * Add previously isolated @page to appropriate LRU list.
737 * Page may still be unevictable for other reasons.
739 * lru_lock must not be held, interrupts must be enabled.
741 void putback_lru_page(struct page
*page
)
744 int was_unevictable
= PageUnevictable(page
);
746 VM_BUG_ON_PAGE(PageLRU(page
), page
);
749 ClearPageUnevictable(page
);
751 if (page_evictable(page
)) {
753 * For evictable pages, we can use the cache.
754 * In event of a race, worst case is we end up with an
755 * unevictable page on [in]active list.
756 * We know how to handle that.
758 is_unevictable
= false;
762 * Put unevictable pages directly on zone's unevictable
765 is_unevictable
= true;
766 add_page_to_unevictable_list(page
);
768 * When racing with an mlock or AS_UNEVICTABLE clearing
769 * (page is unlocked) make sure that if the other thread
770 * does not observe our setting of PG_lru and fails
771 * isolation/check_move_unevictable_pages,
772 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
773 * the page back to the evictable list.
775 * The other side is TestClearPageMlocked() or shmem_lock().
781 * page's status can change while we move it among lru. If an evictable
782 * page is on unevictable list, it never be freed. To avoid that,
783 * check after we added it to the list, again.
785 if (is_unevictable
&& page_evictable(page
)) {
786 if (!isolate_lru_page(page
)) {
790 /* This means someone else dropped this page from LRU
791 * So, it will be freed or putback to LRU again. There is
792 * nothing to do here.
796 if (was_unevictable
&& !is_unevictable
)
797 count_vm_event(UNEVICTABLE_PGRESCUED
);
798 else if (!was_unevictable
&& is_unevictable
)
799 count_vm_event(UNEVICTABLE_PGCULLED
);
801 put_page(page
); /* drop ref from isolate */
804 enum page_references
{
806 PAGEREF_RECLAIM_CLEAN
,
811 static enum page_references
page_check_references(struct page
*page
,
812 struct scan_control
*sc
)
814 int referenced_ptes
, referenced_page
;
815 unsigned long vm_flags
;
817 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
819 referenced_page
= TestClearPageReferenced(page
);
822 * Mlock lost the isolation race with us. Let try_to_unmap()
823 * move the page to the unevictable list.
825 if (vm_flags
& VM_LOCKED
)
826 return PAGEREF_RECLAIM
;
828 if (referenced_ptes
) {
829 if (PageSwapBacked(page
))
830 return PAGEREF_ACTIVATE
;
832 * All mapped pages start out with page table
833 * references from the instantiating fault, so we need
834 * to look twice if a mapped file page is used more
837 * Mark it and spare it for another trip around the
838 * inactive list. Another page table reference will
839 * lead to its activation.
841 * Note: the mark is set for activated pages as well
842 * so that recently deactivated but used pages are
845 SetPageReferenced(page
);
847 if (referenced_page
|| referenced_ptes
> 1)
848 return PAGEREF_ACTIVATE
;
851 * Activate file-backed executable pages after first usage.
853 if (vm_flags
& VM_EXEC
)
854 return PAGEREF_ACTIVATE
;
859 /* Reclaim if clean, defer dirty pages to writeback */
860 if (referenced_page
&& !PageSwapBacked(page
))
861 return PAGEREF_RECLAIM_CLEAN
;
863 return PAGEREF_RECLAIM
;
866 /* Check if a page is dirty or under writeback */
867 static void page_check_dirty_writeback(struct page
*page
,
868 bool *dirty
, bool *writeback
)
870 struct address_space
*mapping
;
873 * Anonymous pages are not handled by flushers and must be written
874 * from reclaim context. Do not stall reclaim based on them
876 if (!page_is_file_cache(page
)) {
882 /* By default assume that the page flags are accurate */
883 *dirty
= PageDirty(page
);
884 *writeback
= PageWriteback(page
);
886 /* Verify dirty/writeback state if the filesystem supports it */
887 if (!page_has_private(page
))
890 mapping
= page_mapping(page
);
891 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
892 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
896 * shrink_page_list() returns the number of reclaimed pages
898 static unsigned long shrink_page_list(struct list_head
*page_list
,
899 struct pglist_data
*pgdat
,
900 struct scan_control
*sc
,
901 enum ttu_flags ttu_flags
,
902 unsigned long *ret_nr_dirty
,
903 unsigned long *ret_nr_unqueued_dirty
,
904 unsigned long *ret_nr_congested
,
905 unsigned long *ret_nr_writeback
,
906 unsigned long *ret_nr_immediate
,
909 LIST_HEAD(ret_pages
);
910 LIST_HEAD(free_pages
);
912 unsigned long nr_unqueued_dirty
= 0;
913 unsigned long nr_dirty
= 0;
914 unsigned long nr_congested
= 0;
915 unsigned long nr_reclaimed
= 0;
916 unsigned long nr_writeback
= 0;
917 unsigned long nr_immediate
= 0;
921 while (!list_empty(page_list
)) {
922 struct address_space
*mapping
;
925 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
926 bool dirty
, writeback
;
927 bool lazyfree
= false;
928 int ret
= SWAP_SUCCESS
;
932 page
= lru_to_page(page_list
);
933 list_del(&page
->lru
);
935 if (!trylock_page(page
))
938 VM_BUG_ON_PAGE(PageActive(page
), page
);
942 if (unlikely(!page_evictable(page
)))
945 if (!sc
->may_unmap
&& page_mapped(page
))
948 /* Double the slab pressure for mapped and swapcache pages */
949 if (page_mapped(page
) || PageSwapCache(page
))
952 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
953 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
956 * The number of dirty pages determines if a zone is marked
957 * reclaim_congested which affects wait_iff_congested. kswapd
958 * will stall and start writing pages if the tail of the LRU
959 * is all dirty unqueued pages.
961 page_check_dirty_writeback(page
, &dirty
, &writeback
);
962 if (dirty
|| writeback
)
965 if (dirty
&& !writeback
)
969 * Treat this page as congested if the underlying BDI is or if
970 * pages are cycling through the LRU so quickly that the
971 * pages marked for immediate reclaim are making it to the
972 * end of the LRU a second time.
974 mapping
= page_mapping(page
);
975 if (((dirty
|| writeback
) && mapping
&&
976 inode_write_congested(mapping
->host
)) ||
977 (writeback
&& PageReclaim(page
)))
981 * If a page at the tail of the LRU is under writeback, there
982 * are three cases to consider.
984 * 1) If reclaim is encountering an excessive number of pages
985 * under writeback and this page is both under writeback and
986 * PageReclaim then it indicates that pages are being queued
987 * for IO but are being recycled through the LRU before the
988 * IO can complete. Waiting on the page itself risks an
989 * indefinite stall if it is impossible to writeback the
990 * page due to IO error or disconnected storage so instead
991 * note that the LRU is being scanned too quickly and the
992 * caller can stall after page list has been processed.
994 * 2) Global or new memcg reclaim encounters a page that is
995 * not marked for immediate reclaim, or the caller does not
996 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
997 * not to fs). In this case mark the page for immediate
998 * reclaim and continue scanning.
1000 * Require may_enter_fs because we would wait on fs, which
1001 * may not have submitted IO yet. And the loop driver might
1002 * enter reclaim, and deadlock if it waits on a page for
1003 * which it is needed to do the write (loop masks off
1004 * __GFP_IO|__GFP_FS for this reason); but more thought
1005 * would probably show more reasons.
1007 * 3) Legacy memcg encounters a page that is already marked
1008 * PageReclaim. memcg does not have any dirty pages
1009 * throttling so we could easily OOM just because too many
1010 * pages are in writeback and there is nothing else to
1011 * reclaim. Wait for the writeback to complete.
1013 if (PageWriteback(page
)) {
1015 if (current_is_kswapd() &&
1016 PageReclaim(page
) &&
1017 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1022 } else if (sane_reclaim(sc
) ||
1023 !PageReclaim(page
) || !may_enter_fs
) {
1025 * This is slightly racy - end_page_writeback()
1026 * might have just cleared PageReclaim, then
1027 * setting PageReclaim here end up interpreted
1028 * as PageReadahead - but that does not matter
1029 * enough to care. What we do want is for this
1030 * page to have PageReclaim set next time memcg
1031 * reclaim reaches the tests above, so it will
1032 * then wait_on_page_writeback() to avoid OOM;
1033 * and it's also appropriate in global reclaim.
1035 SetPageReclaim(page
);
1042 wait_on_page_writeback(page
);
1043 /* then go back and try same page again */
1044 list_add_tail(&page
->lru
, page_list
);
1050 references
= page_check_references(page
, sc
);
1052 switch (references
) {
1053 case PAGEREF_ACTIVATE
:
1054 goto activate_locked
;
1057 case PAGEREF_RECLAIM
:
1058 case PAGEREF_RECLAIM_CLEAN
:
1059 ; /* try to reclaim the page below */
1063 * Anonymous process memory has backing store?
1064 * Try to allocate it some swap space here.
1066 if (PageAnon(page
) && !PageSwapCache(page
)) {
1067 if (!(sc
->gfp_mask
& __GFP_IO
))
1069 if (!add_to_swap(page
, page_list
))
1070 goto activate_locked
;
1074 /* Adding to swap updated mapping */
1075 mapping
= page_mapping(page
);
1076 } else if (unlikely(PageTransHuge(page
))) {
1077 /* Split file THP */
1078 if (split_huge_page_to_list(page
, page_list
))
1082 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1085 * The page is mapped into the page tables of one or more
1086 * processes. Try to unmap it here.
1088 if (page_mapped(page
) && mapping
) {
1089 switch (ret
= try_to_unmap(page
, lazyfree
?
1090 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1091 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1093 goto activate_locked
;
1101 ; /* try to free the page below */
1105 if (PageDirty(page
)) {
1107 * Only kswapd can writeback filesystem pages to
1108 * avoid risk of stack overflow but only writeback
1109 * if many dirty pages have been encountered.
1111 if (page_is_file_cache(page
) &&
1112 (!current_is_kswapd() ||
1113 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1115 * Immediately reclaim when written back.
1116 * Similar in principal to deactivate_page()
1117 * except we already have the page isolated
1118 * and know it's dirty
1120 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1121 SetPageReclaim(page
);
1126 if (references
== PAGEREF_RECLAIM_CLEAN
)
1130 if (!sc
->may_writepage
)
1134 * Page is dirty. Flush the TLB if a writable entry
1135 * potentially exists to avoid CPU writes after IO
1136 * starts and then write it out here.
1138 try_to_unmap_flush_dirty();
1139 switch (pageout(page
, mapping
, sc
)) {
1143 goto activate_locked
;
1145 if (PageWriteback(page
))
1147 if (PageDirty(page
))
1151 * A synchronous write - probably a ramdisk. Go
1152 * ahead and try to reclaim the page.
1154 if (!trylock_page(page
))
1156 if (PageDirty(page
) || PageWriteback(page
))
1158 mapping
= page_mapping(page
);
1160 ; /* try to free the page below */
1165 * If the page has buffers, try to free the buffer mappings
1166 * associated with this page. If we succeed we try to free
1169 * We do this even if the page is PageDirty().
1170 * try_to_release_page() does not perform I/O, but it is
1171 * possible for a page to have PageDirty set, but it is actually
1172 * clean (all its buffers are clean). This happens if the
1173 * buffers were written out directly, with submit_bh(). ext3
1174 * will do this, as well as the blockdev mapping.
1175 * try_to_release_page() will discover that cleanness and will
1176 * drop the buffers and mark the page clean - it can be freed.
1178 * Rarely, pages can have buffers and no ->mapping. These are
1179 * the pages which were not successfully invalidated in
1180 * truncate_complete_page(). We try to drop those buffers here
1181 * and if that worked, and the page is no longer mapped into
1182 * process address space (page_count == 1) it can be freed.
1183 * Otherwise, leave the page on the LRU so it is swappable.
1185 if (page_has_private(page
)) {
1186 if (!try_to_release_page(page
, sc
->gfp_mask
))
1187 goto activate_locked
;
1188 if (!mapping
&& page_count(page
) == 1) {
1190 if (put_page_testzero(page
))
1194 * rare race with speculative reference.
1195 * the speculative reference will free
1196 * this page shortly, so we may
1197 * increment nr_reclaimed here (and
1198 * leave it off the LRU).
1207 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1211 * At this point, we have no other references and there is
1212 * no way to pick any more up (removed from LRU, removed
1213 * from pagecache). Can use non-atomic bitops now (and
1214 * we obviously don't have to worry about waking up a process
1215 * waiting on the page lock, because there are no references.
1217 __ClearPageLocked(page
);
1219 if (ret
== SWAP_LZFREE
)
1220 count_vm_event(PGLAZYFREED
);
1225 * Is there need to periodically free_page_list? It would
1226 * appear not as the counts should be low
1228 list_add(&page
->lru
, &free_pages
);
1232 if (PageSwapCache(page
))
1233 try_to_free_swap(page
);
1235 list_add(&page
->lru
, &ret_pages
);
1239 /* Not a candidate for swapping, so reclaim swap space. */
1240 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1241 try_to_free_swap(page
);
1242 VM_BUG_ON_PAGE(PageActive(page
), page
);
1243 SetPageActive(page
);
1248 list_add(&page
->lru
, &ret_pages
);
1249 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1252 mem_cgroup_uncharge_list(&free_pages
);
1253 try_to_unmap_flush();
1254 free_hot_cold_page_list(&free_pages
, true);
1256 list_splice(&ret_pages
, page_list
);
1257 count_vm_events(PGACTIVATE
, pgactivate
);
1259 *ret_nr_dirty
+= nr_dirty
;
1260 *ret_nr_congested
+= nr_congested
;
1261 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1262 *ret_nr_writeback
+= nr_writeback
;
1263 *ret_nr_immediate
+= nr_immediate
;
1264 return nr_reclaimed
;
1267 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1268 struct list_head
*page_list
)
1270 struct scan_control sc
= {
1271 .gfp_mask
= GFP_KERNEL
,
1272 .priority
= DEF_PRIORITY
,
1275 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1276 struct page
*page
, *next
;
1277 LIST_HEAD(clean_pages
);
1279 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1280 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1281 !__PageMovable(page
)) {
1282 ClearPageActive(page
);
1283 list_move(&page
->lru
, &clean_pages
);
1287 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1288 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1289 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1290 list_splice(&clean_pages
, page_list
);
1291 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1296 * Attempt to remove the specified page from its LRU. Only take this page
1297 * if it is of the appropriate PageActive status. Pages which are being
1298 * freed elsewhere are also ignored.
1300 * page: page to consider
1301 * mode: one of the LRU isolation modes defined above
1303 * returns 0 on success, -ve errno on failure.
1305 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1309 /* Only take pages on the LRU. */
1313 /* Compaction should not handle unevictable pages but CMA can do so */
1314 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1320 * To minimise LRU disruption, the caller can indicate that it only
1321 * wants to isolate pages it will be able to operate on without
1322 * blocking - clean pages for the most part.
1324 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1325 * is used by reclaim when it is cannot write to backing storage
1327 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1328 * that it is possible to migrate without blocking
1330 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1331 /* All the caller can do on PageWriteback is block */
1332 if (PageWriteback(page
))
1335 if (PageDirty(page
)) {
1336 struct address_space
*mapping
;
1338 /* ISOLATE_CLEAN means only clean pages */
1339 if (mode
& ISOLATE_CLEAN
)
1343 * Only pages without mappings or that have a
1344 * ->migratepage callback are possible to migrate
1347 mapping
= page_mapping(page
);
1348 if (mapping
&& !mapping
->a_ops
->migratepage
)
1353 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1356 if (likely(get_page_unless_zero(page
))) {
1358 * Be careful not to clear PageLRU until after we're
1359 * sure the page is not being freed elsewhere -- the
1360 * page release code relies on it.
1370 * zone_lru_lock is heavily contended. Some of the functions that
1371 * shrink the lists perform better by taking out a batch of pages
1372 * and working on them outside the LRU lock.
1374 * For pagecache intensive workloads, this function is the hottest
1375 * spot in the kernel (apart from copy_*_user functions).
1377 * Appropriate locks must be held before calling this function.
1379 * @nr_to_scan: The number of pages to look through on the list.
1380 * @lruvec: The LRU vector to pull pages from.
1381 * @dst: The temp list to put pages on to.
1382 * @nr_scanned: The number of pages that were scanned.
1383 * @sc: The scan_control struct for this reclaim session
1384 * @mode: One of the LRU isolation modes
1385 * @lru: LRU list id for isolating
1387 * returns how many pages were moved onto *@dst.
1389 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1390 struct lruvec
*lruvec
, struct list_head
*dst
,
1391 unsigned long *nr_scanned
, struct scan_control
*sc
,
1392 isolate_mode_t mode
, enum lru_list lru
)
1394 struct list_head
*src
= &lruvec
->lists
[lru
];
1395 unsigned long nr_taken
= 0;
1396 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1397 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1398 unsigned long scan
, nr_pages
;
1399 LIST_HEAD(pages_skipped
);
1401 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1402 !list_empty(src
); scan
++) {
1405 page
= lru_to_page(src
);
1406 prefetchw_prev_lru_page(page
, src
, flags
);
1408 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1410 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1411 list_move(&page
->lru
, &pages_skipped
);
1412 nr_skipped
[page_zonenum(page
)]++;
1416 switch (__isolate_lru_page(page
, mode
)) {
1418 nr_pages
= hpage_nr_pages(page
);
1419 nr_taken
+= nr_pages
;
1420 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1421 list_move(&page
->lru
, dst
);
1425 /* else it is being freed elsewhere */
1426 list_move(&page
->lru
, src
);
1435 * Splice any skipped pages to the start of the LRU list. Note that
1436 * this disrupts the LRU order when reclaiming for lower zones but
1437 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1438 * scanning would soon rescan the same pages to skip and put the
1439 * system at risk of premature OOM.
1441 if (!list_empty(&pages_skipped
)) {
1444 list_splice(&pages_skipped
, src
);
1445 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1446 if (!nr_skipped
[zid
])
1449 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1453 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
, scan
,
1454 nr_taken
, mode
, is_file_lru(lru
));
1455 for (scan
= 0; scan
< MAX_NR_ZONES
; scan
++) {
1456 nr_pages
= nr_zone_taken
[scan
];
1460 update_lru_size(lruvec
, lru
, scan
, -nr_pages
);
1466 * isolate_lru_page - tries to isolate a page from its LRU list
1467 * @page: page to isolate from its LRU list
1469 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1470 * vmstat statistic corresponding to whatever LRU list the page was on.
1472 * Returns 0 if the page was removed from an LRU list.
1473 * Returns -EBUSY if the page was not on an LRU list.
1475 * The returned page will have PageLRU() cleared. If it was found on
1476 * the active list, it will have PageActive set. If it was found on
1477 * the unevictable list, it will have the PageUnevictable bit set. That flag
1478 * may need to be cleared by the caller before letting the page go.
1480 * The vmstat statistic corresponding to the list on which the page was
1481 * found will be decremented.
1484 * (1) Must be called with an elevated refcount on the page. This is a
1485 * fundamentnal difference from isolate_lru_pages (which is called
1486 * without a stable reference).
1487 * (2) the lru_lock must not be held.
1488 * (3) interrupts must be enabled.
1490 int isolate_lru_page(struct page
*page
)
1494 VM_BUG_ON_PAGE(!page_count(page
), page
);
1495 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1497 if (PageLRU(page
)) {
1498 struct zone
*zone
= page_zone(page
);
1499 struct lruvec
*lruvec
;
1501 spin_lock_irq(zone_lru_lock(zone
));
1502 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1503 if (PageLRU(page
)) {
1504 int lru
= page_lru(page
);
1507 del_page_from_lru_list(page
, lruvec
, lru
);
1510 spin_unlock_irq(zone_lru_lock(zone
));
1516 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1517 * then get resheduled. When there are massive number of tasks doing page
1518 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1519 * the LRU list will go small and be scanned faster than necessary, leading to
1520 * unnecessary swapping, thrashing and OOM.
1522 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1523 struct scan_control
*sc
)
1525 unsigned long inactive
, isolated
;
1527 if (current_is_kswapd())
1530 if (!sane_reclaim(sc
))
1534 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1535 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1537 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1538 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1542 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1543 * won't get blocked by normal direct-reclaimers, forming a circular
1546 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1549 return isolated
> inactive
;
1552 static noinline_for_stack
void
1553 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1555 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1556 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1557 LIST_HEAD(pages_to_free
);
1560 * Put back any unfreeable pages.
1562 while (!list_empty(page_list
)) {
1563 struct page
*page
= lru_to_page(page_list
);
1566 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1567 list_del(&page
->lru
);
1568 if (unlikely(!page_evictable(page
))) {
1569 spin_unlock_irq(&pgdat
->lru_lock
);
1570 putback_lru_page(page
);
1571 spin_lock_irq(&pgdat
->lru_lock
);
1575 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1578 lru
= page_lru(page
);
1579 add_page_to_lru_list(page
, lruvec
, lru
);
1581 if (is_active_lru(lru
)) {
1582 int file
= is_file_lru(lru
);
1583 int numpages
= hpage_nr_pages(page
);
1584 reclaim_stat
->recent_rotated
[file
] += numpages
;
1586 if (put_page_testzero(page
)) {
1587 __ClearPageLRU(page
);
1588 __ClearPageActive(page
);
1589 del_page_from_lru_list(page
, lruvec
, lru
);
1591 if (unlikely(PageCompound(page
))) {
1592 spin_unlock_irq(&pgdat
->lru_lock
);
1593 mem_cgroup_uncharge(page
);
1594 (*get_compound_page_dtor(page
))(page
);
1595 spin_lock_irq(&pgdat
->lru_lock
);
1597 list_add(&page
->lru
, &pages_to_free
);
1602 * To save our caller's stack, now use input list for pages to free.
1604 list_splice(&pages_to_free
, page_list
);
1608 * If a kernel thread (such as nfsd for loop-back mounts) services
1609 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1610 * In that case we should only throttle if the backing device it is
1611 * writing to is congested. In other cases it is safe to throttle.
1613 static int current_may_throttle(void)
1615 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1616 current
->backing_dev_info
== NULL
||
1617 bdi_write_congested(current
->backing_dev_info
);
1621 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1622 * of reclaimed pages
1624 static noinline_for_stack
unsigned long
1625 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1626 struct scan_control
*sc
, enum lru_list lru
)
1628 LIST_HEAD(page_list
);
1629 unsigned long nr_scanned
;
1630 unsigned long nr_reclaimed
= 0;
1631 unsigned long nr_taken
;
1632 unsigned long nr_dirty
= 0;
1633 unsigned long nr_congested
= 0;
1634 unsigned long nr_unqueued_dirty
= 0;
1635 unsigned long nr_writeback
= 0;
1636 unsigned long nr_immediate
= 0;
1637 isolate_mode_t isolate_mode
= 0;
1638 int file
= is_file_lru(lru
);
1639 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1640 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1642 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1643 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1645 /* We are about to die and free our memory. Return now. */
1646 if (fatal_signal_pending(current
))
1647 return SWAP_CLUSTER_MAX
;
1653 isolate_mode
|= ISOLATE_UNMAPPED
;
1654 if (!sc
->may_writepage
)
1655 isolate_mode
|= ISOLATE_CLEAN
;
1657 spin_lock_irq(&pgdat
->lru_lock
);
1659 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1660 &nr_scanned
, sc
, isolate_mode
, lru
);
1662 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1663 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1665 if (global_reclaim(sc
)) {
1666 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1667 if (current_is_kswapd())
1668 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1670 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1672 spin_unlock_irq(&pgdat
->lru_lock
);
1677 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, TTU_UNMAP
,
1678 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1679 &nr_writeback
, &nr_immediate
,
1682 spin_lock_irq(&pgdat
->lru_lock
);
1684 if (global_reclaim(sc
)) {
1685 if (current_is_kswapd())
1686 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1688 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1691 putback_inactive_pages(lruvec
, &page_list
);
1693 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1695 spin_unlock_irq(&pgdat
->lru_lock
);
1697 mem_cgroup_uncharge_list(&page_list
);
1698 free_hot_cold_page_list(&page_list
, true);
1701 * If reclaim is isolating dirty pages under writeback, it implies
1702 * that the long-lived page allocation rate is exceeding the page
1703 * laundering rate. Either the global limits are not being effective
1704 * at throttling processes due to the page distribution throughout
1705 * zones or there is heavy usage of a slow backing device. The
1706 * only option is to throttle from reclaim context which is not ideal
1707 * as there is no guarantee the dirtying process is throttled in the
1708 * same way balance_dirty_pages() manages.
1710 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1711 * of pages under pages flagged for immediate reclaim and stall if any
1712 * are encountered in the nr_immediate check below.
1714 if (nr_writeback
&& nr_writeback
== nr_taken
)
1715 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1718 * Legacy memcg will stall in page writeback so avoid forcibly
1721 if (sane_reclaim(sc
)) {
1723 * Tag a zone as congested if all the dirty pages scanned were
1724 * backed by a congested BDI and wait_iff_congested will stall.
1726 if (nr_dirty
&& nr_dirty
== nr_congested
)
1727 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1730 * If dirty pages are scanned that are not queued for IO, it
1731 * implies that flushers are not keeping up. In this case, flag
1732 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1735 if (nr_unqueued_dirty
== nr_taken
)
1736 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1739 * If kswapd scans pages marked marked for immediate
1740 * reclaim and under writeback (nr_immediate), it implies
1741 * that pages are cycling through the LRU faster than
1742 * they are written so also forcibly stall.
1744 if (nr_immediate
&& current_may_throttle())
1745 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1749 * Stall direct reclaim for IO completions if underlying BDIs or zone
1750 * is congested. Allow kswapd to continue until it starts encountering
1751 * unqueued dirty pages or cycling through the LRU too quickly.
1753 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1754 current_may_throttle())
1755 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1757 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1758 nr_scanned
, nr_reclaimed
,
1759 sc
->priority
, file
);
1760 return nr_reclaimed
;
1764 * This moves pages from the active list to the inactive list.
1766 * We move them the other way if the page is referenced by one or more
1767 * processes, from rmap.
1769 * If the pages are mostly unmapped, the processing is fast and it is
1770 * appropriate to hold zone_lru_lock across the whole operation. But if
1771 * the pages are mapped, the processing is slow (page_referenced()) so we
1772 * should drop zone_lru_lock around each page. It's impossible to balance
1773 * this, so instead we remove the pages from the LRU while processing them.
1774 * It is safe to rely on PG_active against the non-LRU pages in here because
1775 * nobody will play with that bit on a non-LRU page.
1777 * The downside is that we have to touch page->_refcount against each page.
1778 * But we had to alter page->flags anyway.
1781 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1782 struct list_head
*list
,
1783 struct list_head
*pages_to_free
,
1786 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1787 unsigned long pgmoved
= 0;
1791 while (!list_empty(list
)) {
1792 page
= lru_to_page(list
);
1793 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1795 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1798 nr_pages
= hpage_nr_pages(page
);
1799 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1800 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1801 pgmoved
+= nr_pages
;
1803 if (put_page_testzero(page
)) {
1804 __ClearPageLRU(page
);
1805 __ClearPageActive(page
);
1806 del_page_from_lru_list(page
, lruvec
, lru
);
1808 if (unlikely(PageCompound(page
))) {
1809 spin_unlock_irq(&pgdat
->lru_lock
);
1810 mem_cgroup_uncharge(page
);
1811 (*get_compound_page_dtor(page
))(page
);
1812 spin_lock_irq(&pgdat
->lru_lock
);
1814 list_add(&page
->lru
, pages_to_free
);
1818 if (!is_active_lru(lru
))
1819 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1822 static void shrink_active_list(unsigned long nr_to_scan
,
1823 struct lruvec
*lruvec
,
1824 struct scan_control
*sc
,
1827 unsigned long nr_taken
;
1828 unsigned long nr_scanned
;
1829 unsigned long vm_flags
;
1830 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1831 LIST_HEAD(l_active
);
1832 LIST_HEAD(l_inactive
);
1834 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1835 unsigned long nr_rotated
= 0;
1836 isolate_mode_t isolate_mode
= 0;
1837 int file
= is_file_lru(lru
);
1838 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1843 isolate_mode
|= ISOLATE_UNMAPPED
;
1844 if (!sc
->may_writepage
)
1845 isolate_mode
|= ISOLATE_CLEAN
;
1847 spin_lock_irq(&pgdat
->lru_lock
);
1849 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1850 &nr_scanned
, sc
, isolate_mode
, lru
);
1852 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1853 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1855 if (global_reclaim(sc
))
1856 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1857 __count_vm_events(PGREFILL
, nr_scanned
);
1859 spin_unlock_irq(&pgdat
->lru_lock
);
1861 while (!list_empty(&l_hold
)) {
1863 page
= lru_to_page(&l_hold
);
1864 list_del(&page
->lru
);
1866 if (unlikely(!page_evictable(page
))) {
1867 putback_lru_page(page
);
1871 if (unlikely(buffer_heads_over_limit
)) {
1872 if (page_has_private(page
) && trylock_page(page
)) {
1873 if (page_has_private(page
))
1874 try_to_release_page(page
, 0);
1879 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1881 nr_rotated
+= hpage_nr_pages(page
);
1883 * Identify referenced, file-backed active pages and
1884 * give them one more trip around the active list. So
1885 * that executable code get better chances to stay in
1886 * memory under moderate memory pressure. Anon pages
1887 * are not likely to be evicted by use-once streaming
1888 * IO, plus JVM can create lots of anon VM_EXEC pages,
1889 * so we ignore them here.
1891 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1892 list_add(&page
->lru
, &l_active
);
1897 ClearPageActive(page
); /* we are de-activating */
1898 list_add(&page
->lru
, &l_inactive
);
1902 * Move pages back to the lru list.
1904 spin_lock_irq(&pgdat
->lru_lock
);
1906 * Count referenced pages from currently used mappings as rotated,
1907 * even though only some of them are actually re-activated. This
1908 * helps balance scan pressure between file and anonymous pages in
1911 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1913 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1914 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1915 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1916 spin_unlock_irq(&pgdat
->lru_lock
);
1918 mem_cgroup_uncharge_list(&l_hold
);
1919 free_hot_cold_page_list(&l_hold
, true);
1923 * The inactive anon list should be small enough that the VM never has
1924 * to do too much work.
1926 * The inactive file list should be small enough to leave most memory
1927 * to the established workingset on the scan-resistant active list,
1928 * but large enough to avoid thrashing the aggregate readahead window.
1930 * Both inactive lists should also be large enough that each inactive
1931 * page has a chance to be referenced again before it is reclaimed.
1933 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1934 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1935 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1938 * memory ratio inactive
1939 * -------------------------------------
1948 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
)
1950 unsigned long inactive_ratio
;
1951 unsigned long inactive
;
1952 unsigned long active
;
1956 * If we don't have swap space, anonymous page deactivation
1959 if (!file
&& !total_swap_pages
)
1962 inactive
= lruvec_lru_size(lruvec
, file
* LRU_FILE
);
1963 active
= lruvec_lru_size(lruvec
, file
* LRU_FILE
+ LRU_ACTIVE
);
1965 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1967 inactive_ratio
= int_sqrt(10 * gb
);
1971 return inactive
* inactive_ratio
< active
;
1974 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1975 struct lruvec
*lruvec
, struct scan_control
*sc
)
1977 if (is_active_lru(lru
)) {
1978 if (inactive_list_is_low(lruvec
, is_file_lru(lru
)))
1979 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1983 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1994 * Determine how aggressively the anon and file LRU lists should be
1995 * scanned. The relative value of each set of LRU lists is determined
1996 * by looking at the fraction of the pages scanned we did rotate back
1997 * onto the active list instead of evict.
1999 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2000 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2002 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2003 struct scan_control
*sc
, unsigned long *nr
,
2004 unsigned long *lru_pages
)
2006 int swappiness
= mem_cgroup_swappiness(memcg
);
2007 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2009 u64 denominator
= 0; /* gcc */
2010 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2011 unsigned long anon_prio
, file_prio
;
2012 enum scan_balance scan_balance
;
2013 unsigned long anon
, file
;
2014 bool force_scan
= false;
2015 unsigned long ap
, fp
;
2021 * If the zone or memcg is small, nr[l] can be 0. This
2022 * results in no scanning on this priority and a potential
2023 * priority drop. Global direct reclaim can go to the next
2024 * zone and tends to have no problems. Global kswapd is for
2025 * zone balancing and it needs to scan a minimum amount. When
2026 * reclaiming for a memcg, a priority drop can cause high
2027 * latencies, so it's better to scan a minimum amount there as
2030 if (current_is_kswapd()) {
2031 if (!pgdat_reclaimable(pgdat
))
2033 if (!mem_cgroup_online(memcg
))
2036 if (!global_reclaim(sc
))
2039 /* If we have no swap space, do not bother scanning anon pages. */
2040 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2041 scan_balance
= SCAN_FILE
;
2046 * Global reclaim will swap to prevent OOM even with no
2047 * swappiness, but memcg users want to use this knob to
2048 * disable swapping for individual groups completely when
2049 * using the memory controller's swap limit feature would be
2052 if (!global_reclaim(sc
) && !swappiness
) {
2053 scan_balance
= SCAN_FILE
;
2058 * Do not apply any pressure balancing cleverness when the
2059 * system is close to OOM, scan both anon and file equally
2060 * (unless the swappiness setting disagrees with swapping).
2062 if (!sc
->priority
&& swappiness
) {
2063 scan_balance
= SCAN_EQUAL
;
2068 * Prevent the reclaimer from falling into the cache trap: as
2069 * cache pages start out inactive, every cache fault will tip
2070 * the scan balance towards the file LRU. And as the file LRU
2071 * shrinks, so does the window for rotation from references.
2072 * This means we have a runaway feedback loop where a tiny
2073 * thrashing file LRU becomes infinitely more attractive than
2074 * anon pages. Try to detect this based on file LRU size.
2076 if (global_reclaim(sc
)) {
2077 unsigned long pgdatfile
;
2078 unsigned long pgdatfree
;
2080 unsigned long total_high_wmark
= 0;
2082 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2083 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2084 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2086 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2087 struct zone
*zone
= &pgdat
->node_zones
[z
];
2088 if (!populated_zone(zone
))
2091 total_high_wmark
+= high_wmark_pages(zone
);
2094 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2095 scan_balance
= SCAN_ANON
;
2101 * If there is enough inactive page cache, i.e. if the size of the
2102 * inactive list is greater than that of the active list *and* the
2103 * inactive list actually has some pages to scan on this priority, we
2104 * do not reclaim anything from the anonymous working set right now.
2105 * Without the second condition we could end up never scanning an
2106 * lruvec even if it has plenty of old anonymous pages unless the
2107 * system is under heavy pressure.
2109 if (!inactive_list_is_low(lruvec
, true) &&
2110 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2111 scan_balance
= SCAN_FILE
;
2115 scan_balance
= SCAN_FRACT
;
2118 * With swappiness at 100, anonymous and file have the same priority.
2119 * This scanning priority is essentially the inverse of IO cost.
2121 anon_prio
= swappiness
;
2122 file_prio
= 200 - anon_prio
;
2125 * OK, so we have swap space and a fair amount of page cache
2126 * pages. We use the recently rotated / recently scanned
2127 * ratios to determine how valuable each cache is.
2129 * Because workloads change over time (and to avoid overflow)
2130 * we keep these statistics as a floating average, which ends
2131 * up weighing recent references more than old ones.
2133 * anon in [0], file in [1]
2136 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2137 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2138 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2139 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2141 spin_lock_irq(&pgdat
->lru_lock
);
2142 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2143 reclaim_stat
->recent_scanned
[0] /= 2;
2144 reclaim_stat
->recent_rotated
[0] /= 2;
2147 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2148 reclaim_stat
->recent_scanned
[1] /= 2;
2149 reclaim_stat
->recent_rotated
[1] /= 2;
2153 * The amount of pressure on anon vs file pages is inversely
2154 * proportional to the fraction of recently scanned pages on
2155 * each list that were recently referenced and in active use.
2157 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2158 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2160 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2161 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2162 spin_unlock_irq(&pgdat
->lru_lock
);
2166 denominator
= ap
+ fp
+ 1;
2168 some_scanned
= false;
2169 /* Only use force_scan on second pass. */
2170 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2172 for_each_evictable_lru(lru
) {
2173 int file
= is_file_lru(lru
);
2177 size
= lruvec_lru_size(lruvec
, lru
);
2178 scan
= size
>> sc
->priority
;
2180 if (!scan
&& pass
&& force_scan
)
2181 scan
= min(size
, SWAP_CLUSTER_MAX
);
2183 switch (scan_balance
) {
2185 /* Scan lists relative to size */
2189 * Scan types proportional to swappiness and
2190 * their relative recent reclaim efficiency.
2192 scan
= div64_u64(scan
* fraction
[file
],
2197 /* Scan one type exclusively */
2198 if ((scan_balance
== SCAN_FILE
) != file
) {
2204 /* Look ma, no brain */
2212 * Skip the second pass and don't force_scan,
2213 * if we found something to scan.
2215 some_scanned
|= !!scan
;
2220 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2221 static void init_tlb_ubc(void)
2224 * This deliberately does not clear the cpumask as it's expensive
2225 * and unnecessary. If there happens to be data in there then the
2226 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2227 * then will be cleared.
2229 current
->tlb_ubc
.flush_required
= false;
2232 static inline void init_tlb_ubc(void)
2235 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2238 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2240 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2241 struct scan_control
*sc
, unsigned long *lru_pages
)
2243 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2244 unsigned long nr
[NR_LRU_LISTS
];
2245 unsigned long targets
[NR_LRU_LISTS
];
2246 unsigned long nr_to_scan
;
2248 unsigned long nr_reclaimed
= 0;
2249 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2250 struct blk_plug plug
;
2253 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2255 /* Record the original scan target for proportional adjustments later */
2256 memcpy(targets
, nr
, sizeof(nr
));
2259 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2260 * event that can occur when there is little memory pressure e.g.
2261 * multiple streaming readers/writers. Hence, we do not abort scanning
2262 * when the requested number of pages are reclaimed when scanning at
2263 * DEF_PRIORITY on the assumption that the fact we are direct
2264 * reclaiming implies that kswapd is not keeping up and it is best to
2265 * do a batch of work at once. For memcg reclaim one check is made to
2266 * abort proportional reclaim if either the file or anon lru has already
2267 * dropped to zero at the first pass.
2269 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2270 sc
->priority
== DEF_PRIORITY
);
2274 blk_start_plug(&plug
);
2275 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2276 nr
[LRU_INACTIVE_FILE
]) {
2277 unsigned long nr_anon
, nr_file
, percentage
;
2278 unsigned long nr_scanned
;
2280 for_each_evictable_lru(lru
) {
2282 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2283 nr
[lru
] -= nr_to_scan
;
2285 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2290 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2294 * For kswapd and memcg, reclaim at least the number of pages
2295 * requested. Ensure that the anon and file LRUs are scanned
2296 * proportionally what was requested by get_scan_count(). We
2297 * stop reclaiming one LRU and reduce the amount scanning
2298 * proportional to the original scan target.
2300 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2301 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2304 * It's just vindictive to attack the larger once the smaller
2305 * has gone to zero. And given the way we stop scanning the
2306 * smaller below, this makes sure that we only make one nudge
2307 * towards proportionality once we've got nr_to_reclaim.
2309 if (!nr_file
|| !nr_anon
)
2312 if (nr_file
> nr_anon
) {
2313 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2314 targets
[LRU_ACTIVE_ANON
] + 1;
2316 percentage
= nr_anon
* 100 / scan_target
;
2318 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2319 targets
[LRU_ACTIVE_FILE
] + 1;
2321 percentage
= nr_file
* 100 / scan_target
;
2324 /* Stop scanning the smaller of the LRU */
2326 nr
[lru
+ LRU_ACTIVE
] = 0;
2329 * Recalculate the other LRU scan count based on its original
2330 * scan target and the percentage scanning already complete
2332 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2333 nr_scanned
= targets
[lru
] - nr
[lru
];
2334 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2335 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2338 nr_scanned
= targets
[lru
] - nr
[lru
];
2339 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2340 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2342 scan_adjusted
= true;
2344 blk_finish_plug(&plug
);
2345 sc
->nr_reclaimed
+= nr_reclaimed
;
2348 * Even if we did not try to evict anon pages at all, we want to
2349 * rebalance the anon lru active/inactive ratio.
2351 if (inactive_list_is_low(lruvec
, false))
2352 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2353 sc
, LRU_ACTIVE_ANON
);
2355 throttle_vm_writeout(sc
->gfp_mask
);
2358 /* Use reclaim/compaction for costly allocs or under memory pressure */
2359 static bool in_reclaim_compaction(struct scan_control
*sc
)
2361 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2362 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2363 sc
->priority
< DEF_PRIORITY
- 2))
2370 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2371 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2372 * true if more pages should be reclaimed such that when the page allocator
2373 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2374 * It will give up earlier than that if there is difficulty reclaiming pages.
2376 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2377 unsigned long nr_reclaimed
,
2378 unsigned long nr_scanned
,
2379 struct scan_control
*sc
)
2381 unsigned long pages_for_compaction
;
2382 unsigned long inactive_lru_pages
;
2385 /* If not in reclaim/compaction mode, stop */
2386 if (!in_reclaim_compaction(sc
))
2389 /* Consider stopping depending on scan and reclaim activity */
2390 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2392 * For __GFP_REPEAT allocations, stop reclaiming if the
2393 * full LRU list has been scanned and we are still failing
2394 * to reclaim pages. This full LRU scan is potentially
2395 * expensive but a __GFP_REPEAT caller really wants to succeed
2397 if (!nr_reclaimed
&& !nr_scanned
)
2401 * For non-__GFP_REPEAT allocations which can presumably
2402 * fail without consequence, stop if we failed to reclaim
2403 * any pages from the last SWAP_CLUSTER_MAX number of
2404 * pages that were scanned. This will return to the
2405 * caller faster at the risk reclaim/compaction and
2406 * the resulting allocation attempt fails
2413 * If we have not reclaimed enough pages for compaction and the
2414 * inactive lists are large enough, continue reclaiming
2416 pages_for_compaction
= (2UL << sc
->order
);
2417 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2418 if (get_nr_swap_pages() > 0)
2419 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2420 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2421 inactive_lru_pages
> pages_for_compaction
)
2424 /* If compaction would go ahead or the allocation would succeed, stop */
2425 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2426 struct zone
*zone
= &pgdat
->node_zones
[z
];
2427 if (!populated_zone(zone
))
2430 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2431 case COMPACT_PARTIAL
:
2432 case COMPACT_CONTINUE
:
2435 /* check next zone */
2442 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2444 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2445 unsigned long nr_reclaimed
, nr_scanned
;
2446 bool reclaimable
= false;
2449 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2450 struct mem_cgroup_reclaim_cookie reclaim
= {
2452 .priority
= sc
->priority
,
2454 unsigned long node_lru_pages
= 0;
2455 struct mem_cgroup
*memcg
;
2457 nr_reclaimed
= sc
->nr_reclaimed
;
2458 nr_scanned
= sc
->nr_scanned
;
2460 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2462 unsigned long lru_pages
;
2463 unsigned long reclaimed
;
2464 unsigned long scanned
;
2466 if (mem_cgroup_low(root
, memcg
)) {
2467 if (!sc
->may_thrash
)
2469 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2472 reclaimed
= sc
->nr_reclaimed
;
2473 scanned
= sc
->nr_scanned
;
2475 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2476 node_lru_pages
+= lru_pages
;
2478 if (!global_reclaim(sc
))
2479 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2480 memcg
, sc
->nr_scanned
- scanned
,
2483 /* Record the group's reclaim efficiency */
2484 vmpressure(sc
->gfp_mask
, memcg
, false,
2485 sc
->nr_scanned
- scanned
,
2486 sc
->nr_reclaimed
- reclaimed
);
2489 * Direct reclaim and kswapd have to scan all memory
2490 * cgroups to fulfill the overall scan target for the
2493 * Limit reclaim, on the other hand, only cares about
2494 * nr_to_reclaim pages to be reclaimed and it will
2495 * retry with decreasing priority if one round over the
2496 * whole hierarchy is not sufficient.
2498 if (!global_reclaim(sc
) &&
2499 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2500 mem_cgroup_iter_break(root
, memcg
);
2503 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2506 * Shrink the slab caches in the same proportion that
2507 * the eligible LRU pages were scanned.
2509 if (global_reclaim(sc
))
2510 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2511 sc
->nr_scanned
- nr_scanned
,
2514 if (reclaim_state
) {
2515 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2516 reclaim_state
->reclaimed_slab
= 0;
2519 /* Record the subtree's reclaim efficiency */
2520 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2521 sc
->nr_scanned
- nr_scanned
,
2522 sc
->nr_reclaimed
- nr_reclaimed
);
2524 if (sc
->nr_reclaimed
- nr_reclaimed
)
2527 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2528 sc
->nr_scanned
- nr_scanned
, sc
));
2534 * Returns true if compaction should go ahead for a high-order request, or
2535 * the high-order allocation would succeed without compaction.
2537 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2539 unsigned long watermark
;
2543 * Compaction takes time to run and there are potentially other
2544 * callers using the pages just freed. Continue reclaiming until
2545 * there is a buffer of free pages available to give compaction
2546 * a reasonable chance of completing and allocating the page
2548 watermark
= high_wmark_pages(zone
) + (2UL << sc
->order
);
2549 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2552 * If compaction is deferred, reclaim up to a point where
2553 * compaction will have a chance of success when re-enabled
2555 if (compaction_deferred(zone
, sc
->order
))
2556 return watermark_ok
;
2559 * If compaction is not ready to start and allocation is not likely
2560 * to succeed without it, then keep reclaiming.
2562 if (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
) == COMPACT_SKIPPED
)
2565 return watermark_ok
;
2569 * This is the direct reclaim path, for page-allocating processes. We only
2570 * try to reclaim pages from zones which will satisfy the caller's allocation
2573 * If a zone is deemed to be full of pinned pages then just give it a light
2574 * scan then give up on it.
2576 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2580 unsigned long nr_soft_reclaimed
;
2581 unsigned long nr_soft_scanned
;
2583 pg_data_t
*last_pgdat
= NULL
;
2586 * If the number of buffer_heads in the machine exceeds the maximum
2587 * allowed level, force direct reclaim to scan the highmem zone as
2588 * highmem pages could be pinning lowmem pages storing buffer_heads
2590 orig_mask
= sc
->gfp_mask
;
2591 if (buffer_heads_over_limit
) {
2592 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2593 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2596 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2597 sc
->reclaim_idx
, sc
->nodemask
) {
2598 if (!populated_zone(zone
))
2602 * Take care memory controller reclaiming has small influence
2605 if (global_reclaim(sc
)) {
2606 if (!cpuset_zone_allowed(zone
,
2607 GFP_KERNEL
| __GFP_HARDWALL
))
2610 if (sc
->priority
!= DEF_PRIORITY
&&
2611 !pgdat_reclaimable(zone
->zone_pgdat
))
2612 continue; /* Let kswapd poll it */
2615 * If we already have plenty of memory free for
2616 * compaction in this zone, don't free any more.
2617 * Even though compaction is invoked for any
2618 * non-zero order, only frequent costly order
2619 * reclamation is disruptive enough to become a
2620 * noticeable problem, like transparent huge
2623 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2624 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2625 compaction_ready(zone
, sc
)) {
2626 sc
->compaction_ready
= true;
2631 * Shrink each node in the zonelist once. If the
2632 * zonelist is ordered by zone (not the default) then a
2633 * node may be shrunk multiple times but in that case
2634 * the user prefers lower zones being preserved.
2636 if (zone
->zone_pgdat
== last_pgdat
)
2640 * This steals pages from memory cgroups over softlimit
2641 * and returns the number of reclaimed pages and
2642 * scanned pages. This works for global memory pressure
2643 * and balancing, not for a memcg's limit.
2645 nr_soft_scanned
= 0;
2646 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2647 sc
->order
, sc
->gfp_mask
,
2649 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2650 sc
->nr_scanned
+= nr_soft_scanned
;
2651 /* need some check for avoid more shrink_zone() */
2654 /* See comment about same check for global reclaim above */
2655 if (zone
->zone_pgdat
== last_pgdat
)
2657 last_pgdat
= zone
->zone_pgdat
;
2658 shrink_node(zone
->zone_pgdat
, sc
);
2662 * Restore to original mask to avoid the impact on the caller if we
2663 * promoted it to __GFP_HIGHMEM.
2665 sc
->gfp_mask
= orig_mask
;
2669 * This is the main entry point to direct page reclaim.
2671 * If a full scan of the inactive list fails to free enough memory then we
2672 * are "out of memory" and something needs to be killed.
2674 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2675 * high - the zone may be full of dirty or under-writeback pages, which this
2676 * caller can't do much about. We kick the writeback threads and take explicit
2677 * naps in the hope that some of these pages can be written. But if the
2678 * allocating task holds filesystem locks which prevent writeout this might not
2679 * work, and the allocation attempt will fail.
2681 * returns: 0, if no pages reclaimed
2682 * else, the number of pages reclaimed
2684 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2685 struct scan_control
*sc
)
2687 int initial_priority
= sc
->priority
;
2688 unsigned long total_scanned
= 0;
2689 unsigned long writeback_threshold
;
2691 delayacct_freepages_start();
2693 if (global_reclaim(sc
))
2694 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2697 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2700 shrink_zones(zonelist
, sc
);
2702 total_scanned
+= sc
->nr_scanned
;
2703 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2706 if (sc
->compaction_ready
)
2710 * If we're getting trouble reclaiming, start doing
2711 * writepage even in laptop mode.
2713 if (sc
->priority
< DEF_PRIORITY
- 2)
2714 sc
->may_writepage
= 1;
2717 * Try to write back as many pages as we just scanned. This
2718 * tends to cause slow streaming writers to write data to the
2719 * disk smoothly, at the dirtying rate, which is nice. But
2720 * that's undesirable in laptop mode, where we *want* lumpy
2721 * writeout. So in laptop mode, write out the whole world.
2723 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2724 if (total_scanned
> writeback_threshold
) {
2725 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2726 WB_REASON_TRY_TO_FREE_PAGES
);
2727 sc
->may_writepage
= 1;
2729 } while (--sc
->priority
>= 0);
2731 delayacct_freepages_end();
2733 if (sc
->nr_reclaimed
)
2734 return sc
->nr_reclaimed
;
2736 /* Aborted reclaim to try compaction? don't OOM, then */
2737 if (sc
->compaction_ready
)
2740 /* Untapped cgroup reserves? Don't OOM, retry. */
2741 if (!sc
->may_thrash
) {
2742 sc
->priority
= initial_priority
;
2750 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2753 unsigned long pfmemalloc_reserve
= 0;
2754 unsigned long free_pages
= 0;
2758 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2759 zone
= &pgdat
->node_zones
[i
];
2760 if (!populated_zone(zone
) ||
2761 pgdat_reclaimable_pages(pgdat
) == 0)
2764 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2765 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2768 /* If there are no reserves (unexpected config) then do not throttle */
2769 if (!pfmemalloc_reserve
)
2772 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2774 /* kswapd must be awake if processes are being throttled */
2775 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2776 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2777 (enum zone_type
)ZONE_NORMAL
);
2778 wake_up_interruptible(&pgdat
->kswapd_wait
);
2785 * Throttle direct reclaimers if backing storage is backed by the network
2786 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2787 * depleted. kswapd will continue to make progress and wake the processes
2788 * when the low watermark is reached.
2790 * Returns true if a fatal signal was delivered during throttling. If this
2791 * happens, the page allocator should not consider triggering the OOM killer.
2793 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2794 nodemask_t
*nodemask
)
2798 pg_data_t
*pgdat
= NULL
;
2801 * Kernel threads should not be throttled as they may be indirectly
2802 * responsible for cleaning pages necessary for reclaim to make forward
2803 * progress. kjournald for example may enter direct reclaim while
2804 * committing a transaction where throttling it could forcing other
2805 * processes to block on log_wait_commit().
2807 if (current
->flags
& PF_KTHREAD
)
2811 * If a fatal signal is pending, this process should not throttle.
2812 * It should return quickly so it can exit and free its memory
2814 if (fatal_signal_pending(current
))
2818 * Check if the pfmemalloc reserves are ok by finding the first node
2819 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2820 * GFP_KERNEL will be required for allocating network buffers when
2821 * swapping over the network so ZONE_HIGHMEM is unusable.
2823 * Throttling is based on the first usable node and throttled processes
2824 * wait on a queue until kswapd makes progress and wakes them. There
2825 * is an affinity then between processes waking up and where reclaim
2826 * progress has been made assuming the process wakes on the same node.
2827 * More importantly, processes running on remote nodes will not compete
2828 * for remote pfmemalloc reserves and processes on different nodes
2829 * should make reasonable progress.
2831 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2832 gfp_zone(gfp_mask
), nodemask
) {
2833 if (zone_idx(zone
) > ZONE_NORMAL
)
2836 /* Throttle based on the first usable node */
2837 pgdat
= zone
->zone_pgdat
;
2838 if (pfmemalloc_watermark_ok(pgdat
))
2843 /* If no zone was usable by the allocation flags then do not throttle */
2847 /* Account for the throttling */
2848 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2851 * If the caller cannot enter the filesystem, it's possible that it
2852 * is due to the caller holding an FS lock or performing a journal
2853 * transaction in the case of a filesystem like ext[3|4]. In this case,
2854 * it is not safe to block on pfmemalloc_wait as kswapd could be
2855 * blocked waiting on the same lock. Instead, throttle for up to a
2856 * second before continuing.
2858 if (!(gfp_mask
& __GFP_FS
)) {
2859 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2860 pfmemalloc_watermark_ok(pgdat
), HZ
);
2865 /* Throttle until kswapd wakes the process */
2866 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2867 pfmemalloc_watermark_ok(pgdat
));
2870 if (fatal_signal_pending(current
))
2877 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2878 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2880 unsigned long nr_reclaimed
;
2881 struct scan_control sc
= {
2882 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2883 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2884 .reclaim_idx
= gfp_zone(gfp_mask
),
2886 .nodemask
= nodemask
,
2887 .priority
= DEF_PRIORITY
,
2888 .may_writepage
= !laptop_mode
,
2894 * Do not enter reclaim if fatal signal was delivered while throttled.
2895 * 1 is returned so that the page allocator does not OOM kill at this
2898 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2901 trace_mm_vmscan_direct_reclaim_begin(order
,
2906 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2908 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2910 return nr_reclaimed
;
2915 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
2916 gfp_t gfp_mask
, bool noswap
,
2918 unsigned long *nr_scanned
)
2920 struct scan_control sc
= {
2921 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2922 .target_mem_cgroup
= memcg
,
2923 .may_writepage
= !laptop_mode
,
2925 .reclaim_idx
= MAX_NR_ZONES
- 1,
2926 .may_swap
= !noswap
,
2928 unsigned long lru_pages
;
2930 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2931 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2933 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2939 * NOTE: Although we can get the priority field, using it
2940 * here is not a good idea, since it limits the pages we can scan.
2941 * if we don't reclaim here, the shrink_node from balance_pgdat
2942 * will pick up pages from other mem cgroup's as well. We hack
2943 * the priority and make it zero.
2945 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
2947 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2949 *nr_scanned
= sc
.nr_scanned
;
2950 return sc
.nr_reclaimed
;
2953 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2954 unsigned long nr_pages
,
2958 struct zonelist
*zonelist
;
2959 unsigned long nr_reclaimed
;
2961 struct scan_control sc
= {
2962 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2963 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2964 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2965 .reclaim_idx
= MAX_NR_ZONES
- 1,
2966 .target_mem_cgroup
= memcg
,
2967 .priority
= DEF_PRIORITY
,
2968 .may_writepage
= !laptop_mode
,
2970 .may_swap
= may_swap
,
2974 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2975 * take care of from where we get pages. So the node where we start the
2976 * scan does not need to be the current node.
2978 nid
= mem_cgroup_select_victim_node(memcg
);
2980 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2982 trace_mm_vmscan_memcg_reclaim_begin(0,
2987 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2989 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2991 return nr_reclaimed
;
2995 static void age_active_anon(struct pglist_data
*pgdat
,
2996 struct scan_control
*sc
)
2998 struct mem_cgroup
*memcg
;
3000 if (!total_swap_pages
)
3003 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3005 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3007 if (inactive_list_is_low(lruvec
, false))
3008 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3009 sc
, LRU_ACTIVE_ANON
);
3011 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3015 static bool zone_balanced(struct zone
*zone
, int order
, int classzone_idx
)
3017 unsigned long mark
= high_wmark_pages(zone
);
3019 if (!zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3023 * If any eligible zone is balanced then the node is not considered
3024 * to be congested or dirty
3026 clear_bit(PGDAT_CONGESTED
, &zone
->zone_pgdat
->flags
);
3027 clear_bit(PGDAT_DIRTY
, &zone
->zone_pgdat
->flags
);
3033 * Prepare kswapd for sleeping. This verifies that there are no processes
3034 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3036 * Returns true if kswapd is ready to sleep
3038 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3043 * The throttled processes are normally woken up in balance_pgdat() as
3044 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3045 * race between when kswapd checks the watermarks and a process gets
3046 * throttled. There is also a potential race if processes get
3047 * throttled, kswapd wakes, a large process exits thereby balancing the
3048 * zones, which causes kswapd to exit balance_pgdat() before reaching
3049 * the wake up checks. If kswapd is going to sleep, no process should
3050 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3051 * the wake up is premature, processes will wake kswapd and get
3052 * throttled again. The difference from wake ups in balance_pgdat() is
3053 * that here we are under prepare_to_wait().
3055 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3056 wake_up_all(&pgdat
->pfmemalloc_wait
);
3058 for (i
= 0; i
<= classzone_idx
; i
++) {
3059 struct zone
*zone
= pgdat
->node_zones
+ i
;
3061 if (!populated_zone(zone
))
3064 if (!zone_balanced(zone
, order
, classzone_idx
))
3072 * kswapd shrinks a node of pages that are at or below the highest usable
3073 * zone that is currently unbalanced.
3075 * Returns true if kswapd scanned at least the requested number of pages to
3076 * reclaim or if the lack of progress was due to pages under writeback.
3077 * This is used to determine if the scanning priority needs to be raised.
3079 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3080 struct scan_control
*sc
)
3085 /* Reclaim a number of pages proportional to the number of zones */
3086 sc
->nr_to_reclaim
= 0;
3087 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3088 zone
= pgdat
->node_zones
+ z
;
3089 if (!populated_zone(zone
))
3092 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3096 * Historically care was taken to put equal pressure on all zones but
3097 * now pressure is applied based on node LRU order.
3099 shrink_node(pgdat
, sc
);
3102 * Fragmentation may mean that the system cannot be rebalanced for
3103 * high-order allocations. If twice the allocation size has been
3104 * reclaimed then recheck watermarks only at order-0 to prevent
3105 * excessive reclaim. Assume that a process requested a high-order
3106 * can direct reclaim/compact.
3108 if (sc
->order
&& sc
->nr_reclaimed
>= 2UL << sc
->order
)
3111 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3115 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3116 * that are eligible for use by the caller until at least one zone is
3119 * Returns the order kswapd finished reclaiming at.
3121 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3122 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3123 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3124 * or lower is eligible for reclaim until at least one usable zone is
3127 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3130 unsigned long nr_soft_reclaimed
;
3131 unsigned long nr_soft_scanned
;
3133 struct scan_control sc
= {
3134 .gfp_mask
= GFP_KERNEL
,
3136 .priority
= DEF_PRIORITY
,
3137 .may_writepage
= !laptop_mode
,
3141 count_vm_event(PAGEOUTRUN
);
3144 bool raise_priority
= true;
3146 sc
.nr_reclaimed
= 0;
3147 sc
.reclaim_idx
= classzone_idx
;
3150 * If the number of buffer_heads exceeds the maximum allowed
3151 * then consider reclaiming from all zones. This has a dual
3152 * purpose -- on 64-bit systems it is expected that
3153 * buffer_heads are stripped during active rotation. On 32-bit
3154 * systems, highmem pages can pin lowmem memory and shrinking
3155 * buffers can relieve lowmem pressure. Reclaim may still not
3156 * go ahead if all eligible zones for the original allocation
3157 * request are balanced to avoid excessive reclaim from kswapd.
3159 if (buffer_heads_over_limit
) {
3160 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3161 zone
= pgdat
->node_zones
+ i
;
3162 if (!populated_zone(zone
))
3171 * Only reclaim if there are no eligible zones. Check from
3172 * high to low zone as allocations prefer higher zones.
3173 * Scanning from low to high zone would allow congestion to be
3174 * cleared during a very small window when a small low
3175 * zone was balanced even under extreme pressure when the
3176 * overall node may be congested. Note that sc.reclaim_idx
3177 * is not used as buffer_heads_over_limit may have adjusted
3180 for (i
= classzone_idx
; i
>= 0; i
--) {
3181 zone
= pgdat
->node_zones
+ i
;
3182 if (!populated_zone(zone
))
3185 if (zone_balanced(zone
, sc
.order
, classzone_idx
))
3190 * Do some background aging of the anon list, to give
3191 * pages a chance to be referenced before reclaiming. All
3192 * pages are rotated regardless of classzone as this is
3193 * about consistent aging.
3195 age_active_anon(pgdat
, &sc
);
3198 * If we're getting trouble reclaiming, start doing writepage
3199 * even in laptop mode.
3201 if (sc
.priority
< DEF_PRIORITY
- 2 || !pgdat_reclaimable(pgdat
))
3202 sc
.may_writepage
= 1;
3204 /* Call soft limit reclaim before calling shrink_node. */
3206 nr_soft_scanned
= 0;
3207 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3208 sc
.gfp_mask
, &nr_soft_scanned
);
3209 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3212 * There should be no need to raise the scanning priority if
3213 * enough pages are already being scanned that that high
3214 * watermark would be met at 100% efficiency.
3216 if (kswapd_shrink_node(pgdat
, &sc
))
3217 raise_priority
= false;
3220 * If the low watermark is met there is no need for processes
3221 * to be throttled on pfmemalloc_wait as they should not be
3222 * able to safely make forward progress. Wake them
3224 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3225 pfmemalloc_watermark_ok(pgdat
))
3226 wake_up_all(&pgdat
->pfmemalloc_wait
);
3228 /* Check if kswapd should be suspending */
3229 if (try_to_freeze() || kthread_should_stop())
3233 * Raise priority if scanning rate is too low or there was no
3234 * progress in reclaiming pages
3236 if (raise_priority
|| !sc
.nr_reclaimed
)
3238 } while (sc
.priority
>= 1);
3242 * Return the order kswapd stopped reclaiming at as
3243 * prepare_kswapd_sleep() takes it into account. If another caller
3244 * entered the allocator slow path while kswapd was awake, order will
3245 * remain at the higher level.
3250 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3251 unsigned int classzone_idx
)
3256 if (freezing(current
) || kthread_should_stop())
3259 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3261 /* Try to sleep for a short interval */
3262 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3264 * Compaction records what page blocks it recently failed to
3265 * isolate pages from and skips them in the future scanning.
3266 * When kswapd is going to sleep, it is reasonable to assume
3267 * that pages and compaction may succeed so reset the cache.
3269 reset_isolation_suitable(pgdat
);
3272 * We have freed the memory, now we should compact it to make
3273 * allocation of the requested order possible.
3275 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3277 remaining
= schedule_timeout(HZ
/10);
3280 * If woken prematurely then reset kswapd_classzone_idx and
3281 * order. The values will either be from a wakeup request or
3282 * the previous request that slept prematurely.
3285 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3286 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3289 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3290 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3294 * After a short sleep, check if it was a premature sleep. If not, then
3295 * go fully to sleep until explicitly woken up.
3298 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3299 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3302 * vmstat counters are not perfectly accurate and the estimated
3303 * value for counters such as NR_FREE_PAGES can deviate from the
3304 * true value by nr_online_cpus * threshold. To avoid the zone
3305 * watermarks being breached while under pressure, we reduce the
3306 * per-cpu vmstat threshold while kswapd is awake and restore
3307 * them before going back to sleep.
3309 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3311 if (!kthread_should_stop())
3314 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3317 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3319 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3321 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3325 * The background pageout daemon, started as a kernel thread
3326 * from the init process.
3328 * This basically trickles out pages so that we have _some_
3329 * free memory available even if there is no other activity
3330 * that frees anything up. This is needed for things like routing
3331 * etc, where we otherwise might have all activity going on in
3332 * asynchronous contexts that cannot page things out.
3334 * If there are applications that are active memory-allocators
3335 * (most normal use), this basically shouldn't matter.
3337 static int kswapd(void *p
)
3339 unsigned int alloc_order
, reclaim_order
, classzone_idx
;
3340 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3341 struct task_struct
*tsk
= current
;
3343 struct reclaim_state reclaim_state
= {
3344 .reclaimed_slab
= 0,
3346 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3348 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3350 if (!cpumask_empty(cpumask
))
3351 set_cpus_allowed_ptr(tsk
, cpumask
);
3352 current
->reclaim_state
= &reclaim_state
;
3355 * Tell the memory management that we're a "memory allocator",
3356 * and that if we need more memory we should get access to it
3357 * regardless (see "__alloc_pages()"). "kswapd" should
3358 * never get caught in the normal page freeing logic.
3360 * (Kswapd normally doesn't need memory anyway, but sometimes
3361 * you need a small amount of memory in order to be able to
3362 * page out something else, and this flag essentially protects
3363 * us from recursively trying to free more memory as we're
3364 * trying to free the first piece of memory in the first place).
3366 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3369 pgdat
->kswapd_order
= alloc_order
= reclaim_order
= 0;
3370 pgdat
->kswapd_classzone_idx
= classzone_idx
= 0;
3375 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3378 /* Read the new order and classzone_idx */
3379 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3380 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3381 pgdat
->kswapd_order
= 0;
3382 pgdat
->kswapd_classzone_idx
= 0;
3384 ret
= try_to_freeze();
3385 if (kthread_should_stop())
3389 * We can speed up thawing tasks if we don't call balance_pgdat
3390 * after returning from the refrigerator
3396 * Reclaim begins at the requested order but if a high-order
3397 * reclaim fails then kswapd falls back to reclaiming for
3398 * order-0. If that happens, kswapd will consider sleeping
3399 * for the order it finished reclaiming at (reclaim_order)
3400 * but kcompactd is woken to compact for the original
3401 * request (alloc_order).
3403 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3405 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3406 if (reclaim_order
< alloc_order
)
3407 goto kswapd_try_sleep
;
3409 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3410 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3413 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3414 current
->reclaim_state
= NULL
;
3415 lockdep_clear_current_reclaim_state();
3421 * A zone is low on free memory, so wake its kswapd task to service it.
3423 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3428 if (!populated_zone(zone
))
3431 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3433 pgdat
= zone
->zone_pgdat
;
3434 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3435 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3436 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3439 /* Only wake kswapd if all zones are unbalanced */
3440 for (z
= 0; z
<= classzone_idx
; z
++) {
3441 zone
= pgdat
->node_zones
+ z
;
3442 if (!populated_zone(zone
))
3445 if (zone_balanced(zone
, order
, classzone_idx
))
3449 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3450 wake_up_interruptible(&pgdat
->kswapd_wait
);
3453 #ifdef CONFIG_HIBERNATION
3455 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3458 * Rather than trying to age LRUs the aim is to preserve the overall
3459 * LRU order by reclaiming preferentially
3460 * inactive > active > active referenced > active mapped
3462 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3464 struct reclaim_state reclaim_state
;
3465 struct scan_control sc
= {
3466 .nr_to_reclaim
= nr_to_reclaim
,
3467 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3468 .reclaim_idx
= MAX_NR_ZONES
- 1,
3469 .priority
= DEF_PRIORITY
,
3473 .hibernation_mode
= 1,
3475 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3476 struct task_struct
*p
= current
;
3477 unsigned long nr_reclaimed
;
3479 p
->flags
|= PF_MEMALLOC
;
3480 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3481 reclaim_state
.reclaimed_slab
= 0;
3482 p
->reclaim_state
= &reclaim_state
;
3484 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3486 p
->reclaim_state
= NULL
;
3487 lockdep_clear_current_reclaim_state();
3488 p
->flags
&= ~PF_MEMALLOC
;
3490 return nr_reclaimed
;
3492 #endif /* CONFIG_HIBERNATION */
3494 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3495 not required for correctness. So if the last cpu in a node goes
3496 away, we get changed to run anywhere: as the first one comes back,
3497 restore their cpu bindings. */
3498 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3503 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3504 for_each_node_state(nid
, N_MEMORY
) {
3505 pg_data_t
*pgdat
= NODE_DATA(nid
);
3506 const struct cpumask
*mask
;
3508 mask
= cpumask_of_node(pgdat
->node_id
);
3510 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3511 /* One of our CPUs online: restore mask */
3512 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3519 * This kswapd start function will be called by init and node-hot-add.
3520 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3522 int kswapd_run(int nid
)
3524 pg_data_t
*pgdat
= NODE_DATA(nid
);
3530 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3531 if (IS_ERR(pgdat
->kswapd
)) {
3532 /* failure at boot is fatal */
3533 BUG_ON(system_state
== SYSTEM_BOOTING
);
3534 pr_err("Failed to start kswapd on node %d\n", nid
);
3535 ret
= PTR_ERR(pgdat
->kswapd
);
3536 pgdat
->kswapd
= NULL
;
3542 * Called by memory hotplug when all memory in a node is offlined. Caller must
3543 * hold mem_hotplug_begin/end().
3545 void kswapd_stop(int nid
)
3547 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3550 kthread_stop(kswapd
);
3551 NODE_DATA(nid
)->kswapd
= NULL
;
3555 static int __init
kswapd_init(void)
3560 for_each_node_state(nid
, N_MEMORY
)
3562 hotcpu_notifier(cpu_callback
, 0);
3566 module_init(kswapd_init
)
3572 * If non-zero call node_reclaim when the number of free pages falls below
3575 int node_reclaim_mode __read_mostly
;
3577 #define RECLAIM_OFF 0
3578 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3579 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3580 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3583 * Priority for NODE_RECLAIM. This determines the fraction of pages
3584 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3587 #define NODE_RECLAIM_PRIORITY 4
3590 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3593 int sysctl_min_unmapped_ratio
= 1;
3596 * If the number of slab pages in a zone grows beyond this percentage then
3597 * slab reclaim needs to occur.
3599 int sysctl_min_slab_ratio
= 5;
3601 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3603 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3604 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3605 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3608 * It's possible for there to be more file mapped pages than
3609 * accounted for by the pages on the file LRU lists because
3610 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3612 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3615 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3616 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3618 unsigned long nr_pagecache_reclaimable
;
3619 unsigned long delta
= 0;
3622 * If RECLAIM_UNMAP is set, then all file pages are considered
3623 * potentially reclaimable. Otherwise, we have to worry about
3624 * pages like swapcache and node_unmapped_file_pages() provides
3627 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3628 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3630 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3632 /* If we can't clean pages, remove dirty pages from consideration */
3633 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3634 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3636 /* Watch for any possible underflows due to delta */
3637 if (unlikely(delta
> nr_pagecache_reclaimable
))
3638 delta
= nr_pagecache_reclaimable
;
3640 return nr_pagecache_reclaimable
- delta
;
3644 * Try to free up some pages from this node through reclaim.
3646 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3648 /* Minimum pages needed in order to stay on node */
3649 const unsigned long nr_pages
= 1 << order
;
3650 struct task_struct
*p
= current
;
3651 struct reclaim_state reclaim_state
;
3652 int classzone_idx
= gfp_zone(gfp_mask
);
3653 struct scan_control sc
= {
3654 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3655 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3657 .priority
= NODE_RECLAIM_PRIORITY
,
3658 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3659 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3661 .reclaim_idx
= classzone_idx
,
3666 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3667 * and we also need to be able to write out pages for RECLAIM_WRITE
3668 * and RECLAIM_UNMAP.
3670 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3671 lockdep_set_current_reclaim_state(gfp_mask
);
3672 reclaim_state
.reclaimed_slab
= 0;
3673 p
->reclaim_state
= &reclaim_state
;
3675 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3677 * Free memory by calling shrink zone with increasing
3678 * priorities until we have enough memory freed.
3681 shrink_node(pgdat
, &sc
);
3682 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3685 p
->reclaim_state
= NULL
;
3686 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3687 lockdep_clear_current_reclaim_state();
3688 return sc
.nr_reclaimed
>= nr_pages
;
3691 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3696 * Node reclaim reclaims unmapped file backed pages and
3697 * slab pages if we are over the defined limits.
3699 * A small portion of unmapped file backed pages is needed for
3700 * file I/O otherwise pages read by file I/O will be immediately
3701 * thrown out if the node is overallocated. So we do not reclaim
3702 * if less than a specified percentage of the node is used by
3703 * unmapped file backed pages.
3705 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3706 sum_zone_node_page_state(pgdat
->node_id
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3707 return NODE_RECLAIM_FULL
;
3709 if (!pgdat_reclaimable(pgdat
))
3710 return NODE_RECLAIM_FULL
;
3713 * Do not scan if the allocation should not be delayed.
3715 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3716 return NODE_RECLAIM_NOSCAN
;
3719 * Only run node reclaim on the local node or on nodes that do not
3720 * have associated processors. This will favor the local processor
3721 * over remote processors and spread off node memory allocations
3722 * as wide as possible.
3724 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3725 return NODE_RECLAIM_NOSCAN
;
3727 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3728 return NODE_RECLAIM_NOSCAN
;
3730 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3731 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3734 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3741 * page_evictable - test whether a page is evictable
3742 * @page: the page to test
3744 * Test whether page is evictable--i.e., should be placed on active/inactive
3745 * lists vs unevictable list.
3747 * Reasons page might not be evictable:
3748 * (1) page's mapping marked unevictable
3749 * (2) page is part of an mlocked VMA
3752 int page_evictable(struct page
*page
)
3754 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3759 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3760 * @pages: array of pages to check
3761 * @nr_pages: number of pages to check
3763 * Checks pages for evictability and moves them to the appropriate lru list.
3765 * This function is only used for SysV IPC SHM_UNLOCK.
3767 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3769 struct lruvec
*lruvec
;
3770 struct zone
*zone
= NULL
;
3775 for (i
= 0; i
< nr_pages
; i
++) {
3776 struct page
*page
= pages
[i
];
3777 struct zone
*pagezone
;
3780 pagezone
= page_zone(page
);
3781 if (pagezone
!= zone
) {
3783 spin_unlock_irq(zone_lru_lock(zone
));
3785 spin_lock_irq(zone_lru_lock(zone
));
3787 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
3789 if (!PageLRU(page
) || !PageUnevictable(page
))
3792 if (page_evictable(page
)) {
3793 enum lru_list lru
= page_lru_base_type(page
);
3795 VM_BUG_ON_PAGE(PageActive(page
), page
);
3796 ClearPageUnevictable(page
);
3797 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3798 add_page_to_lru_list(page
, lruvec
, lru
);
3804 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3805 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3806 spin_unlock_irq(zone_lru_lock(zone
));
3809 #endif /* CONFIG_SHMEM */