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 unsigned int may_writepage
:1;
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap
:1;
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap
:1;
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash
:1;
98 unsigned int hibernation_mode
:1;
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready
:1;
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned
;
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed
;
110 #ifdef ARCH_HAS_PREFETCH
111 #define prefetch_prev_lru_page(_page, _base, _field) \
113 if ((_page)->lru.prev != _base) { \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
121 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
124 #ifdef ARCH_HAS_PREFETCHW
125 #define prefetchw_prev_lru_page(_page, _base, _field) \
127 if ((_page)->lru.prev != _base) { \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
135 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 * From 0 .. 100. Higher means more swappy.
141 int vm_swappiness
= 60;
143 * The total number of pages which are beyond the high watermark within all
146 unsigned long vm_total_pages
;
148 static LIST_HEAD(shrinker_list
);
149 static DECLARE_RWSEM(shrinker_rwsem
);
152 static bool global_reclaim(struct scan_control
*sc
)
154 return !sc
->target_mem_cgroup
;
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
170 static bool sane_reclaim(struct scan_control
*sc
)
172 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
176 #ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
183 static bool global_reclaim(struct scan_control
*sc
)
188 static bool sane_reclaim(struct scan_control
*sc
)
194 unsigned long zone_reclaimable_pages(struct zone
*zone
)
198 nr
= zone_page_state_snapshot(zone
, NR_ACTIVE_FILE
) +
199 zone_page_state_snapshot(zone
, NR_INACTIVE_FILE
) +
200 zone_page_state_snapshot(zone
, NR_ISOLATED_FILE
);
202 if (get_nr_swap_pages() > 0)
203 nr
+= zone_page_state_snapshot(zone
, NR_ACTIVE_ANON
) +
204 zone_page_state_snapshot(zone
, NR_INACTIVE_ANON
) +
205 zone_page_state_snapshot(zone
, NR_ISOLATED_ANON
);
210 bool zone_reclaimable(struct zone
*zone
)
212 return zone_page_state_snapshot(zone
, NR_PAGES_SCANNED
) <
213 zone_reclaimable_pages(zone
) * 6;
216 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec
, lru
);
221 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
225 * Add a shrinker callback to be called from the vm.
227 int register_shrinker(struct shrinker
*shrinker
)
229 size_t size
= sizeof(*shrinker
->nr_deferred
);
231 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
234 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
235 if (!shrinker
->nr_deferred
)
238 down_write(&shrinker_rwsem
);
239 list_add_tail(&shrinker
->list
, &shrinker_list
);
240 up_write(&shrinker_rwsem
);
243 EXPORT_SYMBOL(register_shrinker
);
248 void unregister_shrinker(struct shrinker
*shrinker
)
250 down_write(&shrinker_rwsem
);
251 list_del(&shrinker
->list
);
252 up_write(&shrinker_rwsem
);
253 kfree(shrinker
->nr_deferred
);
255 EXPORT_SYMBOL(unregister_shrinker
);
257 #define SHRINK_BATCH 128
259 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
260 struct shrinker
*shrinker
,
261 unsigned long nr_scanned
,
262 unsigned long nr_eligible
)
264 unsigned long freed
= 0;
265 unsigned long long delta
;
270 int nid
= shrinkctl
->nid
;
271 long batch_size
= shrinker
->batch
? shrinker
->batch
274 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
283 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
286 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
288 do_div(delta
, nr_eligible
+ 1);
290 if (total_scan
< 0) {
291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
292 shrinker
->scan_objects
, total_scan
);
293 total_scan
= freeable
;
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
302 * freeable. This is bad for sustaining a working set in
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
308 if (delta
< freeable
/ 4)
309 total_scan
= min(total_scan
, freeable
/ 2);
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
316 if (total_scan
> freeable
* 2)
317 total_scan
= freeable
* 2;
319 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
320 nr_scanned
, nr_eligible
,
321 freeable
, delta
, total_scan
);
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
334 * than the total number of objects on slab (freeable), we must be
335 * scanning at high prio and therefore should try to reclaim as much as
338 while (total_scan
>= batch_size
||
339 total_scan
>= freeable
) {
341 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
343 shrinkctl
->nr_to_scan
= nr_to_scan
;
344 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
345 if (ret
== SHRINK_STOP
)
349 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
350 total_scan
-= nr_to_scan
;
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
361 new_nr
= atomic_long_add_return(total_scan
,
362 &shrinker
->nr_deferred
[nid
]);
364 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
366 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
371 * shrink_slab - shrink slab caches
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
374 * @memcg: memory cgroup whose slab caches to target
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
378 * Call the shrink functions to age shrinkable caches.
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
396 * Returns the number of reclaimed slab objects.
398 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
399 struct mem_cgroup
*memcg
,
400 unsigned long nr_scanned
,
401 unsigned long nr_eligible
)
403 struct shrinker
*shrinker
;
404 unsigned long freed
= 0;
406 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
410 nr_scanned
= SWAP_CLUSTER_MAX
;
412 if (!down_read_trylock(&shrinker_rwsem
)) {
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
423 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
424 struct shrink_control sc
= {
425 .gfp_mask
= gfp_mask
,
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
435 if (memcg_kmem_enabled() &&
436 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
439 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
442 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
445 up_read(&shrinker_rwsem
);
451 void drop_slab_node(int nid
)
456 struct mem_cgroup
*memcg
= NULL
;
460 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
462 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
463 } while (freed
> 10);
470 for_each_online_node(nid
)
474 static inline int is_page_cache_freeable(struct page
*page
)
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
481 return page_count(page
) - page_has_private(page
) == 2;
484 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
486 if (current
->flags
& PF_SWAPWRITE
)
488 if (!inode_write_congested(inode
))
490 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
504 * We're allowed to run sleeping lock_page() here because we know the caller has
507 static void handle_write_error(struct address_space
*mapping
,
508 struct page
*page
, int error
)
511 if (page_mapping(page
) == mapping
)
512 mapping_set_error(mapping
, error
);
516 /* possible outcome of pageout() */
518 /* failed to write page out, page is locked */
520 /* move page to the active list, page is locked */
522 /* page has been sent to the disk successfully, page is unlocked */
524 /* page is clean and locked */
529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
532 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
533 struct scan_control
*sc
)
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
542 * If this process is currently in __generic_file_write_iter() against
543 * this page's queue, we can perform writeback even if that
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
551 if (!is_page_cache_freeable(page
))
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
558 if (page_has_private(page
)) {
559 if (try_to_free_buffers(page
)) {
560 ClearPageDirty(page
);
561 pr_info("%s: orphaned page\n", __func__
);
567 if (mapping
->a_ops
->writepage
== NULL
)
568 return PAGE_ACTIVATE
;
569 if (!may_write_to_inode(mapping
->host
, sc
))
572 if (clear_page_dirty_for_io(page
)) {
574 struct writeback_control wbc
= {
575 .sync_mode
= WB_SYNC_NONE
,
576 .nr_to_write
= SWAP_CLUSTER_MAX
,
578 .range_end
= LLONG_MAX
,
582 SetPageReclaim(page
);
583 res
= mapping
->a_ops
->writepage(page
, &wbc
);
585 handle_write_error(mapping
, page
, res
);
586 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
587 ClearPageReclaim(page
);
588 return PAGE_ACTIVATE
;
591 if (!PageWriteback(page
)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page
);
595 trace_mm_vmscan_writepage(page
);
596 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
607 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
612 BUG_ON(!PageLocked(page
));
613 BUG_ON(mapping
!= page_mapping(page
));
615 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
617 * The non racy check for a busy page.
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
624 * get_user_pages(&page);
625 * [user mapping goes away]
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
630 * !page_count(page) [good, discard it]
632 * [oops, our write_to data is lost]
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
636 * load is not satisfied before that of page->_refcount.
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
641 if (!page_ref_freeze(page
, 2))
643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page
))) {
645 page_ref_unfreeze(page
, 2);
649 if (PageSwapCache(page
)) {
650 swp_entry_t swap
= { .val
= page_private(page
) };
651 mem_cgroup_swapout(page
, swap
);
652 __delete_from_swap_cache(page
);
653 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
654 swapcache_free(swap
);
656 void (*freepage
)(struct page
*);
659 freepage
= mapping
->a_ops
->freepage
;
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
676 if (reclaimed
&& page_is_file_cache(page
) &&
677 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
678 shadow
= workingset_eviction(mapping
, page
);
679 __delete_from_page_cache(page
, shadow
);
680 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
682 if (freepage
!= NULL
)
689 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
699 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
701 if (__remove_mapping(mapping
, page
, false)) {
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
707 page_ref_unfreeze(page
, 1);
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
720 * lru_lock must not be held, interrupts must be enabled.
722 void putback_lru_page(struct page
*page
)
725 int was_unevictable
= PageUnevictable(page
);
727 VM_BUG_ON_PAGE(PageLRU(page
), page
);
730 ClearPageUnevictable(page
);
732 if (page_evictable(page
)) {
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
739 is_unevictable
= false;
743 * Put unevictable pages directly on zone's unevictable
746 is_unevictable
= true;
747 add_page_to_unevictable_list(page
);
749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
752 * isolation/check_move_unevictable_pages,
753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
754 * the page back to the evictable list.
756 * The other side is TestClearPageMlocked() or shmem_lock().
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
766 if (is_unevictable
&& page_evictable(page
)) {
767 if (!isolate_lru_page(page
)) {
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
777 if (was_unevictable
&& !is_unevictable
)
778 count_vm_event(UNEVICTABLE_PGRESCUED
);
779 else if (!was_unevictable
&& is_unevictable
)
780 count_vm_event(UNEVICTABLE_PGCULLED
);
782 put_page(page
); /* drop ref from isolate */
785 enum page_references
{
787 PAGEREF_RECLAIM_CLEAN
,
792 static enum page_references
page_check_references(struct page
*page
,
793 struct scan_control
*sc
)
795 int referenced_ptes
, referenced_page
;
796 unsigned long vm_flags
;
798 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
800 referenced_page
= TestClearPageReferenced(page
);
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
806 if (vm_flags
& VM_LOCKED
)
807 return PAGEREF_RECLAIM
;
809 if (referenced_ptes
) {
810 if (PageSwapBacked(page
))
811 return PAGEREF_ACTIVATE
;
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
826 SetPageReferenced(page
);
828 if (referenced_page
|| referenced_ptes
> 1)
829 return PAGEREF_ACTIVATE
;
832 * Activate file-backed executable pages after first usage.
834 if (vm_flags
& VM_EXEC
)
835 return PAGEREF_ACTIVATE
;
840 /* Reclaim if clean, defer dirty pages to writeback */
841 if (referenced_page
&& !PageSwapBacked(page
))
842 return PAGEREF_RECLAIM_CLEAN
;
844 return PAGEREF_RECLAIM
;
847 /* Check if a page is dirty or under writeback */
848 static void page_check_dirty_writeback(struct page
*page
,
849 bool *dirty
, bool *writeback
)
851 struct address_space
*mapping
;
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
857 if (!page_is_file_cache(page
)) {
863 /* By default assume that the page flags are accurate */
864 *dirty
= PageDirty(page
);
865 *writeback
= PageWriteback(page
);
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page
))
871 mapping
= page_mapping(page
);
872 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
873 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
877 * shrink_page_list() returns the number of reclaimed pages
879 static unsigned long shrink_page_list(struct list_head
*page_list
,
881 struct scan_control
*sc
,
882 enum ttu_flags ttu_flags
,
883 unsigned long *ret_nr_dirty
,
884 unsigned long *ret_nr_unqueued_dirty
,
885 unsigned long *ret_nr_congested
,
886 unsigned long *ret_nr_writeback
,
887 unsigned long *ret_nr_immediate
,
890 LIST_HEAD(ret_pages
);
891 LIST_HEAD(free_pages
);
893 unsigned long nr_unqueued_dirty
= 0;
894 unsigned long nr_dirty
= 0;
895 unsigned long nr_congested
= 0;
896 unsigned long nr_reclaimed
= 0;
897 unsigned long nr_writeback
= 0;
898 unsigned long nr_immediate
= 0;
902 while (!list_empty(page_list
)) {
903 struct address_space
*mapping
;
906 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
907 bool dirty
, writeback
;
908 bool lazyfree
= false;
909 int ret
= SWAP_SUCCESS
;
913 page
= lru_to_page(page_list
);
914 list_del(&page
->lru
);
916 if (!trylock_page(page
))
919 VM_BUG_ON_PAGE(PageActive(page
), page
);
920 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
924 if (unlikely(!page_evictable(page
)))
927 if (!sc
->may_unmap
&& page_mapped(page
))
930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page
) || PageSwapCache(page
))
934 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
935 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
943 page_check_dirty_writeback(page
, &dirty
, &writeback
);
944 if (dirty
|| writeback
)
947 if (dirty
&& !writeback
)
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
956 mapping
= page_mapping(page
);
957 if (((dirty
|| writeback
) && mapping
&&
958 inode_write_congested(mapping
->host
)) ||
959 (writeback
&& PageReclaim(page
)))
963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
976 * 2) Global or new memcg reclaim encounters a page that is
977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
980 * reclaim and continue scanning.
982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
989 * 3) Legacy memcg encounters a page that is already marked
990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
995 if (PageWriteback(page
)) {
997 if (current_is_kswapd() &&
999 test_bit(ZONE_WRITEBACK
, &zone
->flags
)) {
1004 } else if (sane_reclaim(sc
) ||
1005 !PageReclaim(page
) || !may_enter_fs
) {
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1017 SetPageReclaim(page
);
1024 wait_on_page_writeback(page
);
1025 /* then go back and try same page again */
1026 list_add_tail(&page
->lru
, page_list
);
1032 references
= page_check_references(page
, sc
);
1034 switch (references
) {
1035 case PAGEREF_ACTIVATE
:
1036 goto activate_locked
;
1039 case PAGEREF_RECLAIM
:
1040 case PAGEREF_RECLAIM_CLEAN
:
1041 ; /* try to reclaim the page below */
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1048 if (PageAnon(page
) && !PageSwapCache(page
)) {
1049 if (!(sc
->gfp_mask
& __GFP_IO
))
1051 if (!add_to_swap(page
, page_list
))
1052 goto activate_locked
;
1056 /* Adding to swap updated mapping */
1057 mapping
= page_mapping(page
);
1058 } else if (unlikely(PageTransHuge(page
))) {
1059 /* Split file THP */
1060 if (split_huge_page_to_list(page
, page_list
))
1064 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1067 * The page is mapped into the page tables of one or more
1068 * processes. Try to unmap it here.
1070 if (page_mapped(page
) && mapping
) {
1071 switch (ret
= try_to_unmap(page
, lazyfree
?
1072 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1073 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1075 goto activate_locked
;
1083 ; /* try to free the page below */
1087 if (PageDirty(page
)) {
1089 * Only kswapd can writeback filesystem pages to
1090 * avoid risk of stack overflow but only writeback
1091 * if many dirty pages have been encountered.
1093 if (page_is_file_cache(page
) &&
1094 (!current_is_kswapd() ||
1095 !test_bit(ZONE_DIRTY
, &zone
->flags
))) {
1097 * Immediately reclaim when written back.
1098 * Similar in principal to deactivate_page()
1099 * except we already have the page isolated
1100 * and know it's dirty
1102 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1103 SetPageReclaim(page
);
1108 if (references
== PAGEREF_RECLAIM_CLEAN
)
1112 if (!sc
->may_writepage
)
1116 * Page is dirty. Flush the TLB if a writable entry
1117 * potentially exists to avoid CPU writes after IO
1118 * starts and then write it out here.
1120 try_to_unmap_flush_dirty();
1121 switch (pageout(page
, mapping
, sc
)) {
1125 goto activate_locked
;
1127 if (PageWriteback(page
))
1129 if (PageDirty(page
))
1133 * A synchronous write - probably a ramdisk. Go
1134 * ahead and try to reclaim the page.
1136 if (!trylock_page(page
))
1138 if (PageDirty(page
) || PageWriteback(page
))
1140 mapping
= page_mapping(page
);
1142 ; /* try to free the page below */
1147 * If the page has buffers, try to free the buffer mappings
1148 * associated with this page. If we succeed we try to free
1151 * We do this even if the page is PageDirty().
1152 * try_to_release_page() does not perform I/O, but it is
1153 * possible for a page to have PageDirty set, but it is actually
1154 * clean (all its buffers are clean). This happens if the
1155 * buffers were written out directly, with submit_bh(). ext3
1156 * will do this, as well as the blockdev mapping.
1157 * try_to_release_page() will discover that cleanness and will
1158 * drop the buffers and mark the page clean - it can be freed.
1160 * Rarely, pages can have buffers and no ->mapping. These are
1161 * the pages which were not successfully invalidated in
1162 * truncate_complete_page(). We try to drop those buffers here
1163 * and if that worked, and the page is no longer mapped into
1164 * process address space (page_count == 1) it can be freed.
1165 * Otherwise, leave the page on the LRU so it is swappable.
1167 if (page_has_private(page
)) {
1168 if (!try_to_release_page(page
, sc
->gfp_mask
))
1169 goto activate_locked
;
1170 if (!mapping
&& page_count(page
) == 1) {
1172 if (put_page_testzero(page
))
1176 * rare race with speculative reference.
1177 * the speculative reference will free
1178 * this page shortly, so we may
1179 * increment nr_reclaimed here (and
1180 * leave it off the LRU).
1189 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1193 * At this point, we have no other references and there is
1194 * no way to pick any more up (removed from LRU, removed
1195 * from pagecache). Can use non-atomic bitops now (and
1196 * we obviously don't have to worry about waking up a process
1197 * waiting on the page lock, because there are no references.
1199 __ClearPageLocked(page
);
1201 if (ret
== SWAP_LZFREE
)
1202 count_vm_event(PGLAZYFREED
);
1207 * Is there need to periodically free_page_list? It would
1208 * appear not as the counts should be low
1210 list_add(&page
->lru
, &free_pages
);
1214 if (PageSwapCache(page
))
1215 try_to_free_swap(page
);
1217 list_add(&page
->lru
, &ret_pages
);
1221 /* Not a candidate for swapping, so reclaim swap space. */
1222 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1223 try_to_free_swap(page
);
1224 VM_BUG_ON_PAGE(PageActive(page
), page
);
1225 SetPageActive(page
);
1230 list_add(&page
->lru
, &ret_pages
);
1231 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1234 mem_cgroup_uncharge_list(&free_pages
);
1235 try_to_unmap_flush();
1236 free_hot_cold_page_list(&free_pages
, true);
1238 list_splice(&ret_pages
, page_list
);
1239 count_vm_events(PGACTIVATE
, pgactivate
);
1241 *ret_nr_dirty
+= nr_dirty
;
1242 *ret_nr_congested
+= nr_congested
;
1243 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1244 *ret_nr_writeback
+= nr_writeback
;
1245 *ret_nr_immediate
+= nr_immediate
;
1246 return nr_reclaimed
;
1249 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1250 struct list_head
*page_list
)
1252 struct scan_control sc
= {
1253 .gfp_mask
= GFP_KERNEL
,
1254 .priority
= DEF_PRIORITY
,
1257 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1258 struct page
*page
, *next
;
1259 LIST_HEAD(clean_pages
);
1261 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1262 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1263 !__PageMovable(page
)) {
1264 ClearPageActive(page
);
1265 list_move(&page
->lru
, &clean_pages
);
1269 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1270 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1271 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1272 list_splice(&clean_pages
, page_list
);
1273 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1278 * Attempt to remove the specified page from its LRU. Only take this page
1279 * if it is of the appropriate PageActive status. Pages which are being
1280 * freed elsewhere are also ignored.
1282 * page: page to consider
1283 * mode: one of the LRU isolation modes defined above
1285 * returns 0 on success, -ve errno on failure.
1287 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1291 /* Only take pages on the LRU. */
1295 /* Compaction should not handle unevictable pages but CMA can do so */
1296 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1302 * To minimise LRU disruption, the caller can indicate that it only
1303 * wants to isolate pages it will be able to operate on without
1304 * blocking - clean pages for the most part.
1306 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1307 * is used by reclaim when it is cannot write to backing storage
1309 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1310 * that it is possible to migrate without blocking
1312 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1313 /* All the caller can do on PageWriteback is block */
1314 if (PageWriteback(page
))
1317 if (PageDirty(page
)) {
1318 struct address_space
*mapping
;
1320 /* ISOLATE_CLEAN means only clean pages */
1321 if (mode
& ISOLATE_CLEAN
)
1325 * Only pages without mappings or that have a
1326 * ->migratepage callback are possible to migrate
1329 mapping
= page_mapping(page
);
1330 if (mapping
&& !mapping
->a_ops
->migratepage
)
1335 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1338 if (likely(get_page_unless_zero(page
))) {
1340 * Be careful not to clear PageLRU until after we're
1341 * sure the page is not being freed elsewhere -- the
1342 * page release code relies on it.
1352 * zone->lru_lock is heavily contended. Some of the functions that
1353 * shrink the lists perform better by taking out a batch of pages
1354 * and working on them outside the LRU lock.
1356 * For pagecache intensive workloads, this function is the hottest
1357 * spot in the kernel (apart from copy_*_user functions).
1359 * Appropriate locks must be held before calling this function.
1361 * @nr_to_scan: The number of pages to look through on the list.
1362 * @lruvec: The LRU vector to pull pages from.
1363 * @dst: The temp list to put pages on to.
1364 * @nr_scanned: The number of pages that were scanned.
1365 * @sc: The scan_control struct for this reclaim session
1366 * @mode: One of the LRU isolation modes
1367 * @lru: LRU list id for isolating
1369 * returns how many pages were moved onto *@dst.
1371 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1372 struct lruvec
*lruvec
, struct list_head
*dst
,
1373 unsigned long *nr_scanned
, struct scan_control
*sc
,
1374 isolate_mode_t mode
, enum lru_list lru
)
1376 struct list_head
*src
= &lruvec
->lists
[lru
];
1377 unsigned long nr_taken
= 0;
1380 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1381 !list_empty(src
); scan
++) {
1384 page
= lru_to_page(src
);
1385 prefetchw_prev_lru_page(page
, src
, flags
);
1387 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1389 switch (__isolate_lru_page(page
, mode
)) {
1391 nr_taken
+= hpage_nr_pages(page
);
1392 list_move(&page
->lru
, dst
);
1396 /* else it is being freed elsewhere */
1397 list_move(&page
->lru
, src
);
1406 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1407 nr_taken
, mode
, is_file_lru(lru
));
1412 * isolate_lru_page - tries to isolate a page from its LRU list
1413 * @page: page to isolate from its LRU list
1415 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1416 * vmstat statistic corresponding to whatever LRU list the page was on.
1418 * Returns 0 if the page was removed from an LRU list.
1419 * Returns -EBUSY if the page was not on an LRU list.
1421 * The returned page will have PageLRU() cleared. If it was found on
1422 * the active list, it will have PageActive set. If it was found on
1423 * the unevictable list, it will have the PageUnevictable bit set. That flag
1424 * may need to be cleared by the caller before letting the page go.
1426 * The vmstat statistic corresponding to the list on which the page was
1427 * found will be decremented.
1430 * (1) Must be called with an elevated refcount on the page. This is a
1431 * fundamentnal difference from isolate_lru_pages (which is called
1432 * without a stable reference).
1433 * (2) the lru_lock must not be held.
1434 * (3) interrupts must be enabled.
1436 int isolate_lru_page(struct page
*page
)
1440 VM_BUG_ON_PAGE(!page_count(page
), page
);
1441 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1443 if (PageLRU(page
)) {
1444 struct zone
*zone
= page_zone(page
);
1445 struct lruvec
*lruvec
;
1447 spin_lock_irq(&zone
->lru_lock
);
1448 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1449 if (PageLRU(page
)) {
1450 int lru
= page_lru(page
);
1453 del_page_from_lru_list(page
, lruvec
, lru
);
1456 spin_unlock_irq(&zone
->lru_lock
);
1462 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1463 * then get resheduled. When there are massive number of tasks doing page
1464 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1465 * the LRU list will go small and be scanned faster than necessary, leading to
1466 * unnecessary swapping, thrashing and OOM.
1468 static int too_many_isolated(struct zone
*zone
, int file
,
1469 struct scan_control
*sc
)
1471 unsigned long inactive
, isolated
;
1473 if (current_is_kswapd())
1476 if (!sane_reclaim(sc
))
1480 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1481 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1483 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1484 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1488 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1489 * won't get blocked by normal direct-reclaimers, forming a circular
1492 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1495 return isolated
> inactive
;
1498 static noinline_for_stack
void
1499 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1501 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1502 struct zone
*zone
= lruvec_zone(lruvec
);
1503 LIST_HEAD(pages_to_free
);
1506 * Put back any unfreeable pages.
1508 while (!list_empty(page_list
)) {
1509 struct page
*page
= lru_to_page(page_list
);
1512 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1513 list_del(&page
->lru
);
1514 if (unlikely(!page_evictable(page
))) {
1515 spin_unlock_irq(&zone
->lru_lock
);
1516 putback_lru_page(page
);
1517 spin_lock_irq(&zone
->lru_lock
);
1521 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1524 lru
= page_lru(page
);
1525 add_page_to_lru_list(page
, lruvec
, lru
);
1527 if (is_active_lru(lru
)) {
1528 int file
= is_file_lru(lru
);
1529 int numpages
= hpage_nr_pages(page
);
1530 reclaim_stat
->recent_rotated
[file
] += numpages
;
1532 if (put_page_testzero(page
)) {
1533 __ClearPageLRU(page
);
1534 __ClearPageActive(page
);
1535 del_page_from_lru_list(page
, lruvec
, lru
);
1537 if (unlikely(PageCompound(page
))) {
1538 spin_unlock_irq(&zone
->lru_lock
);
1539 mem_cgroup_uncharge(page
);
1540 (*get_compound_page_dtor(page
))(page
);
1541 spin_lock_irq(&zone
->lru_lock
);
1543 list_add(&page
->lru
, &pages_to_free
);
1548 * To save our caller's stack, now use input list for pages to free.
1550 list_splice(&pages_to_free
, page_list
);
1554 * If a kernel thread (such as nfsd for loop-back mounts) services
1555 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1556 * In that case we should only throttle if the backing device it is
1557 * writing to is congested. In other cases it is safe to throttle.
1559 static int current_may_throttle(void)
1561 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1562 current
->backing_dev_info
== NULL
||
1563 bdi_write_congested(current
->backing_dev_info
);
1567 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1568 * of reclaimed pages
1570 static noinline_for_stack
unsigned long
1571 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1572 struct scan_control
*sc
, enum lru_list lru
)
1574 LIST_HEAD(page_list
);
1575 unsigned long nr_scanned
;
1576 unsigned long nr_reclaimed
= 0;
1577 unsigned long nr_taken
;
1578 unsigned long nr_dirty
= 0;
1579 unsigned long nr_congested
= 0;
1580 unsigned long nr_unqueued_dirty
= 0;
1581 unsigned long nr_writeback
= 0;
1582 unsigned long nr_immediate
= 0;
1583 isolate_mode_t isolate_mode
= 0;
1584 int file
= is_file_lru(lru
);
1585 struct zone
*zone
= lruvec_zone(lruvec
);
1586 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1588 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1589 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1591 /* We are about to die and free our memory. Return now. */
1592 if (fatal_signal_pending(current
))
1593 return SWAP_CLUSTER_MAX
;
1599 isolate_mode
|= ISOLATE_UNMAPPED
;
1600 if (!sc
->may_writepage
)
1601 isolate_mode
|= ISOLATE_CLEAN
;
1603 spin_lock_irq(&zone
->lru_lock
);
1605 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1606 &nr_scanned
, sc
, isolate_mode
, lru
);
1608 update_lru_size(lruvec
, lru
, -nr_taken
);
1609 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1610 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1612 if (global_reclaim(sc
)) {
1613 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1614 if (current_is_kswapd())
1615 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1617 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1619 spin_unlock_irq(&zone
->lru_lock
);
1624 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1625 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1626 &nr_writeback
, &nr_immediate
,
1629 spin_lock_irq(&zone
->lru_lock
);
1631 if (global_reclaim(sc
)) {
1632 if (current_is_kswapd())
1633 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1636 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1640 putback_inactive_pages(lruvec
, &page_list
);
1642 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1644 spin_unlock_irq(&zone
->lru_lock
);
1646 mem_cgroup_uncharge_list(&page_list
);
1647 free_hot_cold_page_list(&page_list
, true);
1650 * If reclaim is isolating dirty pages under writeback, it implies
1651 * that the long-lived page allocation rate is exceeding the page
1652 * laundering rate. Either the global limits are not being effective
1653 * at throttling processes due to the page distribution throughout
1654 * zones or there is heavy usage of a slow backing device. The
1655 * only option is to throttle from reclaim context which is not ideal
1656 * as there is no guarantee the dirtying process is throttled in the
1657 * same way balance_dirty_pages() manages.
1659 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1660 * of pages under pages flagged for immediate reclaim and stall if any
1661 * are encountered in the nr_immediate check below.
1663 if (nr_writeback
&& nr_writeback
== nr_taken
)
1664 set_bit(ZONE_WRITEBACK
, &zone
->flags
);
1667 * Legacy memcg will stall in page writeback so avoid forcibly
1670 if (sane_reclaim(sc
)) {
1672 * Tag a zone as congested if all the dirty pages scanned were
1673 * backed by a congested BDI and wait_iff_congested will stall.
1675 if (nr_dirty
&& nr_dirty
== nr_congested
)
1676 set_bit(ZONE_CONGESTED
, &zone
->flags
);
1679 * If dirty pages are scanned that are not queued for IO, it
1680 * implies that flushers are not keeping up. In this case, flag
1681 * the zone ZONE_DIRTY and kswapd will start writing pages from
1684 if (nr_unqueued_dirty
== nr_taken
)
1685 set_bit(ZONE_DIRTY
, &zone
->flags
);
1688 * If kswapd scans pages marked marked for immediate
1689 * reclaim and under writeback (nr_immediate), it implies
1690 * that pages are cycling through the LRU faster than
1691 * they are written so also forcibly stall.
1693 if (nr_immediate
&& current_may_throttle())
1694 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1698 * Stall direct reclaim for IO completions if underlying BDIs or zone
1699 * is congested. Allow kswapd to continue until it starts encountering
1700 * unqueued dirty pages or cycling through the LRU too quickly.
1702 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1703 current_may_throttle())
1704 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1706 trace_mm_vmscan_lru_shrink_inactive(zone
, nr_scanned
, nr_reclaimed
,
1707 sc
->priority
, file
);
1708 return nr_reclaimed
;
1712 * This moves pages from the active list to the inactive list.
1714 * We move them the other way if the page is referenced by one or more
1715 * processes, from rmap.
1717 * If the pages are mostly unmapped, the processing is fast and it is
1718 * appropriate to hold zone->lru_lock across the whole operation. But if
1719 * the pages are mapped, the processing is slow (page_referenced()) so we
1720 * should drop zone->lru_lock around each page. It's impossible to balance
1721 * this, so instead we remove the pages from the LRU while processing them.
1722 * It is safe to rely on PG_active against the non-LRU pages in here because
1723 * nobody will play with that bit on a non-LRU page.
1725 * The downside is that we have to touch page->_refcount against each page.
1726 * But we had to alter page->flags anyway.
1729 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1730 struct list_head
*list
,
1731 struct list_head
*pages_to_free
,
1734 struct zone
*zone
= lruvec_zone(lruvec
);
1735 unsigned long pgmoved
= 0;
1739 while (!list_empty(list
)) {
1740 page
= lru_to_page(list
);
1741 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1743 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1746 nr_pages
= hpage_nr_pages(page
);
1747 update_lru_size(lruvec
, lru
, nr_pages
);
1748 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1749 pgmoved
+= nr_pages
;
1751 if (put_page_testzero(page
)) {
1752 __ClearPageLRU(page
);
1753 __ClearPageActive(page
);
1754 del_page_from_lru_list(page
, lruvec
, lru
);
1756 if (unlikely(PageCompound(page
))) {
1757 spin_unlock_irq(&zone
->lru_lock
);
1758 mem_cgroup_uncharge(page
);
1759 (*get_compound_page_dtor(page
))(page
);
1760 spin_lock_irq(&zone
->lru_lock
);
1762 list_add(&page
->lru
, pages_to_free
);
1766 if (!is_active_lru(lru
))
1767 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1770 static void shrink_active_list(unsigned long nr_to_scan
,
1771 struct lruvec
*lruvec
,
1772 struct scan_control
*sc
,
1775 unsigned long nr_taken
;
1776 unsigned long nr_scanned
;
1777 unsigned long vm_flags
;
1778 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1779 LIST_HEAD(l_active
);
1780 LIST_HEAD(l_inactive
);
1782 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1783 unsigned long nr_rotated
= 0;
1784 isolate_mode_t isolate_mode
= 0;
1785 int file
= is_file_lru(lru
);
1786 struct zone
*zone
= lruvec_zone(lruvec
);
1791 isolate_mode
|= ISOLATE_UNMAPPED
;
1792 if (!sc
->may_writepage
)
1793 isolate_mode
|= ISOLATE_CLEAN
;
1795 spin_lock_irq(&zone
->lru_lock
);
1797 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1798 &nr_scanned
, sc
, isolate_mode
, lru
);
1800 update_lru_size(lruvec
, lru
, -nr_taken
);
1801 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1802 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1804 if (global_reclaim(sc
))
1805 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1806 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1808 spin_unlock_irq(&zone
->lru_lock
);
1810 while (!list_empty(&l_hold
)) {
1812 page
= lru_to_page(&l_hold
);
1813 list_del(&page
->lru
);
1815 if (unlikely(!page_evictable(page
))) {
1816 putback_lru_page(page
);
1820 if (unlikely(buffer_heads_over_limit
)) {
1821 if (page_has_private(page
) && trylock_page(page
)) {
1822 if (page_has_private(page
))
1823 try_to_release_page(page
, 0);
1828 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1830 nr_rotated
+= hpage_nr_pages(page
);
1832 * Identify referenced, file-backed active pages and
1833 * give them one more trip around the active list. So
1834 * that executable code get better chances to stay in
1835 * memory under moderate memory pressure. Anon pages
1836 * are not likely to be evicted by use-once streaming
1837 * IO, plus JVM can create lots of anon VM_EXEC pages,
1838 * so we ignore them here.
1840 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1841 list_add(&page
->lru
, &l_active
);
1846 ClearPageActive(page
); /* we are de-activating */
1847 list_add(&page
->lru
, &l_inactive
);
1851 * Move pages back to the lru list.
1853 spin_lock_irq(&zone
->lru_lock
);
1855 * Count referenced pages from currently used mappings as rotated,
1856 * even though only some of them are actually re-activated. This
1857 * helps balance scan pressure between file and anonymous pages in
1860 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1862 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1863 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1864 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1865 spin_unlock_irq(&zone
->lru_lock
);
1867 mem_cgroup_uncharge_list(&l_hold
);
1868 free_hot_cold_page_list(&l_hold
, true);
1872 * The inactive anon list should be small enough that the VM never has
1873 * to do too much work.
1875 * The inactive file list should be small enough to leave most memory
1876 * to the established workingset on the scan-resistant active list,
1877 * but large enough to avoid thrashing the aggregate readahead window.
1879 * Both inactive lists should also be large enough that each inactive
1880 * page has a chance to be referenced again before it is reclaimed.
1882 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1883 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1884 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1887 * memory ratio inactive
1888 * -------------------------------------
1897 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
)
1899 unsigned long inactive_ratio
;
1900 unsigned long inactive
;
1901 unsigned long active
;
1905 * If we don't have swap space, anonymous page deactivation
1908 if (!file
&& !total_swap_pages
)
1911 inactive
= lruvec_lru_size(lruvec
, file
* LRU_FILE
);
1912 active
= lruvec_lru_size(lruvec
, file
* LRU_FILE
+ LRU_ACTIVE
);
1914 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1916 inactive_ratio
= int_sqrt(10 * gb
);
1920 return inactive
* inactive_ratio
< active
;
1923 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1924 struct lruvec
*lruvec
, struct scan_control
*sc
)
1926 if (is_active_lru(lru
)) {
1927 if (inactive_list_is_low(lruvec
, is_file_lru(lru
)))
1928 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1932 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1943 * Determine how aggressively the anon and file LRU lists should be
1944 * scanned. The relative value of each set of LRU lists is determined
1945 * by looking at the fraction of the pages scanned we did rotate back
1946 * onto the active list instead of evict.
1948 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1949 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1951 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
1952 struct scan_control
*sc
, unsigned long *nr
,
1953 unsigned long *lru_pages
)
1955 int swappiness
= mem_cgroup_swappiness(memcg
);
1956 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1958 u64 denominator
= 0; /* gcc */
1959 struct zone
*zone
= lruvec_zone(lruvec
);
1960 unsigned long anon_prio
, file_prio
;
1961 enum scan_balance scan_balance
;
1962 unsigned long anon
, file
;
1963 bool force_scan
= false;
1964 unsigned long ap
, fp
;
1970 * If the zone or memcg is small, nr[l] can be 0. This
1971 * results in no scanning on this priority and a potential
1972 * priority drop. Global direct reclaim can go to the next
1973 * zone and tends to have no problems. Global kswapd is for
1974 * zone balancing and it needs to scan a minimum amount. When
1975 * reclaiming for a memcg, a priority drop can cause high
1976 * latencies, so it's better to scan a minimum amount there as
1979 if (current_is_kswapd()) {
1980 if (!zone_reclaimable(zone
))
1982 if (!mem_cgroup_online(memcg
))
1985 if (!global_reclaim(sc
))
1988 /* If we have no swap space, do not bother scanning anon pages. */
1989 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
1990 scan_balance
= SCAN_FILE
;
1995 * Global reclaim will swap to prevent OOM even with no
1996 * swappiness, but memcg users want to use this knob to
1997 * disable swapping for individual groups completely when
1998 * using the memory controller's swap limit feature would be
2001 if (!global_reclaim(sc
) && !swappiness
) {
2002 scan_balance
= SCAN_FILE
;
2007 * Do not apply any pressure balancing cleverness when the
2008 * system is close to OOM, scan both anon and file equally
2009 * (unless the swappiness setting disagrees with swapping).
2011 if (!sc
->priority
&& swappiness
) {
2012 scan_balance
= SCAN_EQUAL
;
2017 * Prevent the reclaimer from falling into the cache trap: as
2018 * cache pages start out inactive, every cache fault will tip
2019 * the scan balance towards the file LRU. And as the file LRU
2020 * shrinks, so does the window for rotation from references.
2021 * This means we have a runaway feedback loop where a tiny
2022 * thrashing file LRU becomes infinitely more attractive than
2023 * anon pages. Try to detect this based on file LRU size.
2025 if (global_reclaim(sc
)) {
2026 unsigned long zonefile
;
2027 unsigned long zonefree
;
2029 zonefree
= zone_page_state(zone
, NR_FREE_PAGES
);
2030 zonefile
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2031 zone_page_state(zone
, NR_INACTIVE_FILE
);
2033 if (unlikely(zonefile
+ zonefree
<= high_wmark_pages(zone
))) {
2034 scan_balance
= SCAN_ANON
;
2040 * If there is enough inactive page cache, i.e. if the size of the
2041 * inactive list is greater than that of the active list *and* the
2042 * inactive list actually has some pages to scan on this priority, we
2043 * do not reclaim anything from the anonymous working set right now.
2044 * Without the second condition we could end up never scanning an
2045 * lruvec even if it has plenty of old anonymous pages unless the
2046 * system is under heavy pressure.
2048 if (!inactive_list_is_low(lruvec
, true) &&
2049 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2050 scan_balance
= SCAN_FILE
;
2054 scan_balance
= SCAN_FRACT
;
2057 * With swappiness at 100, anonymous and file have the same priority.
2058 * This scanning priority is essentially the inverse of IO cost.
2060 anon_prio
= swappiness
;
2061 file_prio
= 200 - anon_prio
;
2064 * OK, so we have swap space and a fair amount of page cache
2065 * pages. We use the recently rotated / recently scanned
2066 * ratios to determine how valuable each cache is.
2068 * Because workloads change over time (and to avoid overflow)
2069 * we keep these statistics as a floating average, which ends
2070 * up weighing recent references more than old ones.
2072 * anon in [0], file in [1]
2075 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2076 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2077 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2078 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2080 spin_lock_irq(&zone
->lru_lock
);
2081 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2082 reclaim_stat
->recent_scanned
[0] /= 2;
2083 reclaim_stat
->recent_rotated
[0] /= 2;
2086 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2087 reclaim_stat
->recent_scanned
[1] /= 2;
2088 reclaim_stat
->recent_rotated
[1] /= 2;
2092 * The amount of pressure on anon vs file pages is inversely
2093 * proportional to the fraction of recently scanned pages on
2094 * each list that were recently referenced and in active use.
2096 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2097 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2099 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2100 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2101 spin_unlock_irq(&zone
->lru_lock
);
2105 denominator
= ap
+ fp
+ 1;
2107 some_scanned
= false;
2108 /* Only use force_scan on second pass. */
2109 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2111 for_each_evictable_lru(lru
) {
2112 int file
= is_file_lru(lru
);
2116 size
= lruvec_lru_size(lruvec
, lru
);
2117 scan
= size
>> sc
->priority
;
2119 if (!scan
&& pass
&& force_scan
)
2120 scan
= min(size
, SWAP_CLUSTER_MAX
);
2122 switch (scan_balance
) {
2124 /* Scan lists relative to size */
2128 * Scan types proportional to swappiness and
2129 * their relative recent reclaim efficiency.
2131 scan
= div64_u64(scan
* fraction
[file
],
2136 /* Scan one type exclusively */
2137 if ((scan_balance
== SCAN_FILE
) != file
) {
2143 /* Look ma, no brain */
2151 * Skip the second pass and don't force_scan,
2152 * if we found something to scan.
2154 some_scanned
|= !!scan
;
2159 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2160 static void init_tlb_ubc(void)
2163 * This deliberately does not clear the cpumask as it's expensive
2164 * and unnecessary. If there happens to be data in there then the
2165 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2166 * then will be cleared.
2168 current
->tlb_ubc
.flush_required
= false;
2171 static inline void init_tlb_ubc(void)
2174 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2177 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2179 static void shrink_zone_memcg(struct zone
*zone
, struct mem_cgroup
*memcg
,
2180 struct scan_control
*sc
, unsigned long *lru_pages
)
2182 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2183 unsigned long nr
[NR_LRU_LISTS
];
2184 unsigned long targets
[NR_LRU_LISTS
];
2185 unsigned long nr_to_scan
;
2187 unsigned long nr_reclaimed
= 0;
2188 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2189 struct blk_plug plug
;
2192 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2194 /* Record the original scan target for proportional adjustments later */
2195 memcpy(targets
, nr
, sizeof(nr
));
2198 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2199 * event that can occur when there is little memory pressure e.g.
2200 * multiple streaming readers/writers. Hence, we do not abort scanning
2201 * when the requested number of pages are reclaimed when scanning at
2202 * DEF_PRIORITY on the assumption that the fact we are direct
2203 * reclaiming implies that kswapd is not keeping up and it is best to
2204 * do a batch of work at once. For memcg reclaim one check is made to
2205 * abort proportional reclaim if either the file or anon lru has already
2206 * dropped to zero at the first pass.
2208 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2209 sc
->priority
== DEF_PRIORITY
);
2213 blk_start_plug(&plug
);
2214 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2215 nr
[LRU_INACTIVE_FILE
]) {
2216 unsigned long nr_anon
, nr_file
, percentage
;
2217 unsigned long nr_scanned
;
2219 for_each_evictable_lru(lru
) {
2221 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2222 nr
[lru
] -= nr_to_scan
;
2224 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2229 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2233 * For kswapd and memcg, reclaim at least the number of pages
2234 * requested. Ensure that the anon and file LRUs are scanned
2235 * proportionally what was requested by get_scan_count(). We
2236 * stop reclaiming one LRU and reduce the amount scanning
2237 * proportional to the original scan target.
2239 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2240 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2243 * It's just vindictive to attack the larger once the smaller
2244 * has gone to zero. And given the way we stop scanning the
2245 * smaller below, this makes sure that we only make one nudge
2246 * towards proportionality once we've got nr_to_reclaim.
2248 if (!nr_file
|| !nr_anon
)
2251 if (nr_file
> nr_anon
) {
2252 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2253 targets
[LRU_ACTIVE_ANON
] + 1;
2255 percentage
= nr_anon
* 100 / scan_target
;
2257 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2258 targets
[LRU_ACTIVE_FILE
] + 1;
2260 percentage
= nr_file
* 100 / scan_target
;
2263 /* Stop scanning the smaller of the LRU */
2265 nr
[lru
+ LRU_ACTIVE
] = 0;
2268 * Recalculate the other LRU scan count based on its original
2269 * scan target and the percentage scanning already complete
2271 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2272 nr_scanned
= targets
[lru
] - nr
[lru
];
2273 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2274 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2277 nr_scanned
= targets
[lru
] - nr
[lru
];
2278 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2279 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2281 scan_adjusted
= true;
2283 blk_finish_plug(&plug
);
2284 sc
->nr_reclaimed
+= nr_reclaimed
;
2287 * Even if we did not try to evict anon pages at all, we want to
2288 * rebalance the anon lru active/inactive ratio.
2290 if (inactive_list_is_low(lruvec
, false))
2291 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2292 sc
, LRU_ACTIVE_ANON
);
2294 throttle_vm_writeout(sc
->gfp_mask
);
2297 /* Use reclaim/compaction for costly allocs or under memory pressure */
2298 static bool in_reclaim_compaction(struct scan_control
*sc
)
2300 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2301 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2302 sc
->priority
< DEF_PRIORITY
- 2))
2309 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2310 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2311 * true if more pages should be reclaimed such that when the page allocator
2312 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2313 * It will give up earlier than that if there is difficulty reclaiming pages.
2315 static inline bool should_continue_reclaim(struct zone
*zone
,
2316 unsigned long nr_reclaimed
,
2317 unsigned long nr_scanned
,
2318 struct scan_control
*sc
)
2320 unsigned long pages_for_compaction
;
2321 unsigned long inactive_lru_pages
;
2323 /* If not in reclaim/compaction mode, stop */
2324 if (!in_reclaim_compaction(sc
))
2327 /* Consider stopping depending on scan and reclaim activity */
2328 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2330 * For __GFP_REPEAT allocations, stop reclaiming if the
2331 * full LRU list has been scanned and we are still failing
2332 * to reclaim pages. This full LRU scan is potentially
2333 * expensive but a __GFP_REPEAT caller really wants to succeed
2335 if (!nr_reclaimed
&& !nr_scanned
)
2339 * For non-__GFP_REPEAT allocations which can presumably
2340 * fail without consequence, stop if we failed to reclaim
2341 * any pages from the last SWAP_CLUSTER_MAX number of
2342 * pages that were scanned. This will return to the
2343 * caller faster at the risk reclaim/compaction and
2344 * the resulting allocation attempt fails
2351 * If we have not reclaimed enough pages for compaction and the
2352 * inactive lists are large enough, continue reclaiming
2354 pages_for_compaction
= (2UL << sc
->order
);
2355 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2356 if (get_nr_swap_pages() > 0)
2357 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2358 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2359 inactive_lru_pages
> pages_for_compaction
)
2362 /* If compaction would go ahead or the allocation would succeed, stop */
2363 switch (compaction_suitable(zone
, sc
->order
, 0, 0)) {
2364 case COMPACT_PARTIAL
:
2365 case COMPACT_CONTINUE
:
2372 static bool shrink_zone(struct zone
*zone
, struct scan_control
*sc
,
2375 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2376 unsigned long nr_reclaimed
, nr_scanned
;
2377 bool reclaimable
= false;
2380 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2381 struct mem_cgroup_reclaim_cookie reclaim
= {
2383 .priority
= sc
->priority
,
2385 unsigned long zone_lru_pages
= 0;
2386 struct mem_cgroup
*memcg
;
2388 nr_reclaimed
= sc
->nr_reclaimed
;
2389 nr_scanned
= sc
->nr_scanned
;
2391 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2393 unsigned long lru_pages
;
2394 unsigned long reclaimed
;
2395 unsigned long scanned
;
2397 if (mem_cgroup_low(root
, memcg
)) {
2398 if (!sc
->may_thrash
)
2400 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2403 reclaimed
= sc
->nr_reclaimed
;
2404 scanned
= sc
->nr_scanned
;
2406 shrink_zone_memcg(zone
, memcg
, sc
, &lru_pages
);
2407 zone_lru_pages
+= lru_pages
;
2409 if (memcg
&& is_classzone
)
2410 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
),
2411 memcg
, sc
->nr_scanned
- scanned
,
2414 /* Record the group's reclaim efficiency */
2415 vmpressure(sc
->gfp_mask
, memcg
, false,
2416 sc
->nr_scanned
- scanned
,
2417 sc
->nr_reclaimed
- reclaimed
);
2420 * Direct reclaim and kswapd have to scan all memory
2421 * cgroups to fulfill the overall scan target for the
2424 * Limit reclaim, on the other hand, only cares about
2425 * nr_to_reclaim pages to be reclaimed and it will
2426 * retry with decreasing priority if one round over the
2427 * whole hierarchy is not sufficient.
2429 if (!global_reclaim(sc
) &&
2430 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2431 mem_cgroup_iter_break(root
, memcg
);
2434 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2437 * Shrink the slab caches in the same proportion that
2438 * the eligible LRU pages were scanned.
2440 if (global_reclaim(sc
) && is_classzone
)
2441 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
), NULL
,
2442 sc
->nr_scanned
- nr_scanned
,
2445 if (reclaim_state
) {
2446 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2447 reclaim_state
->reclaimed_slab
= 0;
2450 /* Record the subtree's reclaim efficiency */
2451 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2452 sc
->nr_scanned
- nr_scanned
,
2453 sc
->nr_reclaimed
- nr_reclaimed
);
2455 if (sc
->nr_reclaimed
- nr_reclaimed
)
2458 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2459 sc
->nr_scanned
- nr_scanned
, sc
));
2465 * Returns true if compaction should go ahead for a high-order request, or
2466 * the high-order allocation would succeed without compaction.
2468 static inline bool compaction_ready(struct zone
*zone
, int order
, int classzone_idx
)
2470 unsigned long balance_gap
, watermark
;
2474 * Compaction takes time to run and there are potentially other
2475 * callers using the pages just freed. Continue reclaiming until
2476 * there is a buffer of free pages available to give compaction
2477 * a reasonable chance of completing and allocating the page
2479 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2480 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2481 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << order
);
2482 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, classzone_idx
);
2485 * If compaction is deferred, reclaim up to a point where
2486 * compaction will have a chance of success when re-enabled
2488 if (compaction_deferred(zone
, order
))
2489 return watermark_ok
;
2492 * If compaction is not ready to start and allocation is not likely
2493 * to succeed without it, then keep reclaiming.
2495 if (compaction_suitable(zone
, order
, 0, classzone_idx
) == COMPACT_SKIPPED
)
2498 return watermark_ok
;
2502 * This is the direct reclaim path, for page-allocating processes. We only
2503 * try to reclaim pages from zones which will satisfy the caller's allocation
2506 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2508 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2510 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2511 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2512 * zone defense algorithm.
2514 * If a zone is deemed to be full of pinned pages then just give it a light
2515 * scan then give up on it.
2517 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2521 unsigned long nr_soft_reclaimed
;
2522 unsigned long nr_soft_scanned
;
2524 enum zone_type requested_highidx
= gfp_zone(sc
->gfp_mask
);
2527 * If the number of buffer_heads in the machine exceeds the maximum
2528 * allowed level, force direct reclaim to scan the highmem zone as
2529 * highmem pages could be pinning lowmem pages storing buffer_heads
2531 orig_mask
= sc
->gfp_mask
;
2532 if (buffer_heads_over_limit
)
2533 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2535 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2536 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2537 enum zone_type classzone_idx
;
2539 if (!populated_zone(zone
))
2542 classzone_idx
= requested_highidx
;
2543 while (!populated_zone(zone
->zone_pgdat
->node_zones
+
2548 * Take care memory controller reclaiming has small influence
2551 if (global_reclaim(sc
)) {
2552 if (!cpuset_zone_allowed(zone
,
2553 GFP_KERNEL
| __GFP_HARDWALL
))
2556 if (sc
->priority
!= DEF_PRIORITY
&&
2557 !zone_reclaimable(zone
))
2558 continue; /* Let kswapd poll it */
2561 * If we already have plenty of memory free for
2562 * compaction in this zone, don't free any more.
2563 * Even though compaction is invoked for any
2564 * non-zero order, only frequent costly order
2565 * reclamation is disruptive enough to become a
2566 * noticeable problem, like transparent huge
2569 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2570 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2571 zonelist_zone_idx(z
) <= requested_highidx
&&
2572 compaction_ready(zone
, sc
->order
, requested_highidx
)) {
2573 sc
->compaction_ready
= true;
2578 * This steals pages from memory cgroups over softlimit
2579 * and returns the number of reclaimed pages and
2580 * scanned pages. This works for global memory pressure
2581 * and balancing, not for a memcg's limit.
2583 nr_soft_scanned
= 0;
2584 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2585 sc
->order
, sc
->gfp_mask
,
2587 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2588 sc
->nr_scanned
+= nr_soft_scanned
;
2589 /* need some check for avoid more shrink_zone() */
2592 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
2596 * Restore to original mask to avoid the impact on the caller if we
2597 * promoted it to __GFP_HIGHMEM.
2599 sc
->gfp_mask
= orig_mask
;
2603 * This is the main entry point to direct page reclaim.
2605 * If a full scan of the inactive list fails to free enough memory then we
2606 * are "out of memory" and something needs to be killed.
2608 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2609 * high - the zone may be full of dirty or under-writeback pages, which this
2610 * caller can't do much about. We kick the writeback threads and take explicit
2611 * naps in the hope that some of these pages can be written. But if the
2612 * allocating task holds filesystem locks which prevent writeout this might not
2613 * work, and the allocation attempt will fail.
2615 * returns: 0, if no pages reclaimed
2616 * else, the number of pages reclaimed
2618 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2619 struct scan_control
*sc
)
2621 int initial_priority
= sc
->priority
;
2622 unsigned long total_scanned
= 0;
2623 unsigned long writeback_threshold
;
2625 delayacct_freepages_start();
2627 if (global_reclaim(sc
))
2628 count_vm_event(ALLOCSTALL
);
2631 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2634 shrink_zones(zonelist
, sc
);
2636 total_scanned
+= sc
->nr_scanned
;
2637 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2640 if (sc
->compaction_ready
)
2644 * If we're getting trouble reclaiming, start doing
2645 * writepage even in laptop mode.
2647 if (sc
->priority
< DEF_PRIORITY
- 2)
2648 sc
->may_writepage
= 1;
2651 * Try to write back as many pages as we just scanned. This
2652 * tends to cause slow streaming writers to write data to the
2653 * disk smoothly, at the dirtying rate, which is nice. But
2654 * that's undesirable in laptop mode, where we *want* lumpy
2655 * writeout. So in laptop mode, write out the whole world.
2657 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2658 if (total_scanned
> writeback_threshold
) {
2659 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2660 WB_REASON_TRY_TO_FREE_PAGES
);
2661 sc
->may_writepage
= 1;
2663 } while (--sc
->priority
>= 0);
2665 delayacct_freepages_end();
2667 if (sc
->nr_reclaimed
)
2668 return sc
->nr_reclaimed
;
2670 /* Aborted reclaim to try compaction? don't OOM, then */
2671 if (sc
->compaction_ready
)
2674 /* Untapped cgroup reserves? Don't OOM, retry. */
2675 if (!sc
->may_thrash
) {
2676 sc
->priority
= initial_priority
;
2684 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2687 unsigned long pfmemalloc_reserve
= 0;
2688 unsigned long free_pages
= 0;
2692 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2693 zone
= &pgdat
->node_zones
[i
];
2694 if (!populated_zone(zone
) ||
2695 zone_reclaimable_pages(zone
) == 0)
2698 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2699 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2702 /* If there are no reserves (unexpected config) then do not throttle */
2703 if (!pfmemalloc_reserve
)
2706 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2708 /* kswapd must be awake if processes are being throttled */
2709 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2710 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2711 (enum zone_type
)ZONE_NORMAL
);
2712 wake_up_interruptible(&pgdat
->kswapd_wait
);
2719 * Throttle direct reclaimers if backing storage is backed by the network
2720 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2721 * depleted. kswapd will continue to make progress and wake the processes
2722 * when the low watermark is reached.
2724 * Returns true if a fatal signal was delivered during throttling. If this
2725 * happens, the page allocator should not consider triggering the OOM killer.
2727 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2728 nodemask_t
*nodemask
)
2732 pg_data_t
*pgdat
= NULL
;
2735 * Kernel threads should not be throttled as they may be indirectly
2736 * responsible for cleaning pages necessary for reclaim to make forward
2737 * progress. kjournald for example may enter direct reclaim while
2738 * committing a transaction where throttling it could forcing other
2739 * processes to block on log_wait_commit().
2741 if (current
->flags
& PF_KTHREAD
)
2745 * If a fatal signal is pending, this process should not throttle.
2746 * It should return quickly so it can exit and free its memory
2748 if (fatal_signal_pending(current
))
2752 * Check if the pfmemalloc reserves are ok by finding the first node
2753 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2754 * GFP_KERNEL will be required for allocating network buffers when
2755 * swapping over the network so ZONE_HIGHMEM is unusable.
2757 * Throttling is based on the first usable node and throttled processes
2758 * wait on a queue until kswapd makes progress and wakes them. There
2759 * is an affinity then between processes waking up and where reclaim
2760 * progress has been made assuming the process wakes on the same node.
2761 * More importantly, processes running on remote nodes will not compete
2762 * for remote pfmemalloc reserves and processes on different nodes
2763 * should make reasonable progress.
2765 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2766 gfp_zone(gfp_mask
), nodemask
) {
2767 if (zone_idx(zone
) > ZONE_NORMAL
)
2770 /* Throttle based on the first usable node */
2771 pgdat
= zone
->zone_pgdat
;
2772 if (pfmemalloc_watermark_ok(pgdat
))
2777 /* If no zone was usable by the allocation flags then do not throttle */
2781 /* Account for the throttling */
2782 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2785 * If the caller cannot enter the filesystem, it's possible that it
2786 * is due to the caller holding an FS lock or performing a journal
2787 * transaction in the case of a filesystem like ext[3|4]. In this case,
2788 * it is not safe to block on pfmemalloc_wait as kswapd could be
2789 * blocked waiting on the same lock. Instead, throttle for up to a
2790 * second before continuing.
2792 if (!(gfp_mask
& __GFP_FS
)) {
2793 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2794 pfmemalloc_watermark_ok(pgdat
), HZ
);
2799 /* Throttle until kswapd wakes the process */
2800 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2801 pfmemalloc_watermark_ok(pgdat
));
2804 if (fatal_signal_pending(current
))
2811 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2812 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2814 unsigned long nr_reclaimed
;
2815 struct scan_control sc
= {
2816 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2817 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2819 .nodemask
= nodemask
,
2820 .priority
= DEF_PRIORITY
,
2821 .may_writepage
= !laptop_mode
,
2827 * Do not enter reclaim if fatal signal was delivered while throttled.
2828 * 1 is returned so that the page allocator does not OOM kill at this
2831 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2834 trace_mm_vmscan_direct_reclaim_begin(order
,
2838 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2840 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2842 return nr_reclaimed
;
2847 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2848 gfp_t gfp_mask
, bool noswap
,
2850 unsigned long *nr_scanned
)
2852 struct scan_control sc
= {
2853 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2854 .target_mem_cgroup
= memcg
,
2855 .may_writepage
= !laptop_mode
,
2857 .may_swap
= !noswap
,
2859 unsigned long lru_pages
;
2861 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2862 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2864 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2869 * NOTE: Although we can get the priority field, using it
2870 * here is not a good idea, since it limits the pages we can scan.
2871 * if we don't reclaim here, the shrink_zone from balance_pgdat
2872 * will pick up pages from other mem cgroup's as well. We hack
2873 * the priority and make it zero.
2875 shrink_zone_memcg(zone
, memcg
, &sc
, &lru_pages
);
2877 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2879 *nr_scanned
= sc
.nr_scanned
;
2880 return sc
.nr_reclaimed
;
2883 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2884 unsigned long nr_pages
,
2888 struct zonelist
*zonelist
;
2889 unsigned long nr_reclaimed
;
2891 struct scan_control sc
= {
2892 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2893 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2894 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2895 .target_mem_cgroup
= memcg
,
2896 .priority
= DEF_PRIORITY
,
2897 .may_writepage
= !laptop_mode
,
2899 .may_swap
= may_swap
,
2903 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2904 * take care of from where we get pages. So the node where we start the
2905 * scan does not need to be the current node.
2907 nid
= mem_cgroup_select_victim_node(memcg
);
2909 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2911 trace_mm_vmscan_memcg_reclaim_begin(0,
2915 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2917 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2919 return nr_reclaimed
;
2923 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2925 struct mem_cgroup
*memcg
;
2927 if (!total_swap_pages
)
2930 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2932 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2934 if (inactive_list_is_low(lruvec
, false))
2935 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2936 sc
, LRU_ACTIVE_ANON
);
2938 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2942 static bool zone_balanced(struct zone
*zone
, int order
, bool highorder
,
2943 unsigned long balance_gap
, int classzone_idx
)
2945 unsigned long mark
= high_wmark_pages(zone
) + balance_gap
;
2948 * When checking from pgdat_balanced(), kswapd should stop and sleep
2949 * when it reaches the high order-0 watermark and let kcompactd take
2950 * over. Other callers such as wakeup_kswapd() want to determine the
2951 * true high-order watermark.
2953 if (IS_ENABLED(CONFIG_COMPACTION
) && !highorder
) {
2954 mark
+= (1UL << order
);
2958 return zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
);
2962 * pgdat_balanced() is used when checking if a node is balanced.
2964 * For order-0, all zones must be balanced!
2966 * For high-order allocations only zones that meet watermarks and are in a
2967 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2968 * total of balanced pages must be at least 25% of the zones allowed by
2969 * classzone_idx for the node to be considered balanced. Forcing all zones to
2970 * be balanced for high orders can cause excessive reclaim when there are
2972 * The choice of 25% is due to
2973 * o a 16M DMA zone that is balanced will not balance a zone on any
2974 * reasonable sized machine
2975 * o On all other machines, the top zone must be at least a reasonable
2976 * percentage of the middle zones. For example, on 32-bit x86, highmem
2977 * would need to be at least 256M for it to be balance a whole node.
2978 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2979 * to balance a node on its own. These seemed like reasonable ratios.
2981 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2983 unsigned long managed_pages
= 0;
2984 unsigned long balanced_pages
= 0;
2987 /* Check the watermark levels */
2988 for (i
= 0; i
<= classzone_idx
; i
++) {
2989 struct zone
*zone
= pgdat
->node_zones
+ i
;
2991 if (!populated_zone(zone
))
2994 managed_pages
+= zone
->managed_pages
;
2997 * A special case here:
2999 * balance_pgdat() skips over all_unreclaimable after
3000 * DEF_PRIORITY. Effectively, it considers them balanced so
3001 * they must be considered balanced here as well!
3003 if (!zone_reclaimable(zone
)) {
3004 balanced_pages
+= zone
->managed_pages
;
3008 if (zone_balanced(zone
, order
, false, 0, i
))
3009 balanced_pages
+= zone
->managed_pages
;
3015 return balanced_pages
>= (managed_pages
>> 2);
3021 * Prepare kswapd for sleeping. This verifies that there are no processes
3022 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3024 * Returns true if kswapd is ready to sleep
3026 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
3029 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3034 * The throttled processes are normally woken up in balance_pgdat() as
3035 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3036 * race between when kswapd checks the watermarks and a process gets
3037 * throttled. There is also a potential race if processes get
3038 * throttled, kswapd wakes, a large process exits thereby balancing the
3039 * zones, which causes kswapd to exit balance_pgdat() before reaching
3040 * the wake up checks. If kswapd is going to sleep, no process should
3041 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3042 * the wake up is premature, processes will wake kswapd and get
3043 * throttled again. The difference from wake ups in balance_pgdat() is
3044 * that here we are under prepare_to_wait().
3046 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3047 wake_up_all(&pgdat
->pfmemalloc_wait
);
3049 return pgdat_balanced(pgdat
, order
, classzone_idx
);
3053 * kswapd shrinks the zone by the number of pages required to reach
3054 * the high watermark.
3056 * Returns true if kswapd scanned at least the requested number of pages to
3057 * reclaim or if the lack of progress was due to pages under writeback.
3058 * This is used to determine if the scanning priority needs to be raised.
3060 static bool kswapd_shrink_zone(struct zone
*zone
,
3062 struct scan_control
*sc
)
3064 unsigned long balance_gap
;
3065 bool lowmem_pressure
;
3067 /* Reclaim above the high watermark. */
3068 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
3071 * We put equal pressure on every zone, unless one zone has way too
3072 * many pages free already. The "too many pages" is defined as the
3073 * high wmark plus a "gap" where the gap is either the low
3074 * watermark or 1% of the zone, whichever is smaller.
3076 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
3077 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
3080 * If there is no low memory pressure or the zone is balanced then no
3081 * reclaim is necessary
3083 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
3084 if (!lowmem_pressure
&& zone_balanced(zone
, sc
->order
, false,
3085 balance_gap
, classzone_idx
))
3088 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
3090 clear_bit(ZONE_WRITEBACK
, &zone
->flags
);
3093 * If a zone reaches its high watermark, consider it to be no longer
3094 * congested. It's possible there are dirty pages backed by congested
3095 * BDIs but as pressure is relieved, speculatively avoid congestion
3098 if (zone_reclaimable(zone
) &&
3099 zone_balanced(zone
, sc
->order
, false, 0, classzone_idx
)) {
3100 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3101 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3104 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3108 * For kswapd, balance_pgdat() will work across all this node's zones until
3109 * they are all at high_wmark_pages(zone).
3111 * Returns the highest zone idx kswapd was reclaiming at
3113 * There is special handling here for zones which are full of pinned pages.
3114 * This can happen if the pages are all mlocked, or if they are all used by
3115 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3116 * What we do is to detect the case where all pages in the zone have been
3117 * scanned twice and there has been zero successful reclaim. Mark the zone as
3118 * dead and from now on, only perform a short scan. Basically we're polling
3119 * the zone for when the problem goes away.
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), we scan that zone and the
3124 * lower zones regardless of the number of free pages in the lower zones. This
3125 * interoperates with the page allocator fallback scheme to ensure that aging
3126 * of pages is balanced across the zones.
3128 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3131 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
3132 unsigned long nr_soft_reclaimed
;
3133 unsigned long nr_soft_scanned
;
3134 struct scan_control sc
= {
3135 .gfp_mask
= GFP_KERNEL
,
3137 .priority
= DEF_PRIORITY
,
3138 .may_writepage
= !laptop_mode
,
3142 count_vm_event(PAGEOUTRUN
);
3145 bool raise_priority
= true;
3147 sc
.nr_reclaimed
= 0;
3150 * Scan in the highmem->dma direction for the highest
3151 * zone which needs scanning
3153 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
3154 struct zone
*zone
= pgdat
->node_zones
+ i
;
3156 if (!populated_zone(zone
))
3159 if (sc
.priority
!= DEF_PRIORITY
&&
3160 !zone_reclaimable(zone
))
3164 * Do some background aging of the anon list, to give
3165 * pages a chance to be referenced before reclaiming.
3167 age_active_anon(zone
, &sc
);
3170 * If the number of buffer_heads in the machine
3171 * exceeds the maximum allowed level and this node
3172 * has a highmem zone, force kswapd to reclaim from
3173 * it to relieve lowmem pressure.
3175 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
3180 if (!zone_balanced(zone
, order
, false, 0, 0)) {
3185 * If balanced, clear the dirty and congested
3188 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3189 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3197 * If we're getting trouble reclaiming, start doing writepage
3198 * even in laptop mode.
3200 if (sc
.priority
< DEF_PRIORITY
- 2)
3201 sc
.may_writepage
= 1;
3204 * Now scan the zone in the dma->highmem direction, stopping
3205 * at the last zone which needs scanning.
3207 * We do this because the page allocator works in the opposite
3208 * direction. This prevents the page allocator from allocating
3209 * pages behind kswapd's direction of progress, which would
3210 * cause too much scanning of the lower zones.
3212 for (i
= 0; i
<= end_zone
; i
++) {
3213 struct zone
*zone
= pgdat
->node_zones
+ i
;
3215 if (!populated_zone(zone
))
3218 if (sc
.priority
!= DEF_PRIORITY
&&
3219 !zone_reclaimable(zone
))
3224 nr_soft_scanned
= 0;
3226 * Call soft limit reclaim before calling shrink_zone.
3228 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3231 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3234 * There should be no need to raise the scanning
3235 * priority if enough pages are already being scanned
3236 * that that high watermark would be met at 100%
3239 if (kswapd_shrink_zone(zone
, end_zone
, &sc
))
3240 raise_priority
= false;
3244 * If the low watermark is met there is no need for processes
3245 * to be throttled on pfmemalloc_wait as they should not be
3246 * able to safely make forward progress. Wake them
3248 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3249 pfmemalloc_watermark_ok(pgdat
))
3250 wake_up_all(&pgdat
->pfmemalloc_wait
);
3252 /* Check if kswapd should be suspending */
3253 if (try_to_freeze() || kthread_should_stop())
3257 * Raise priority if scanning rate is too low or there was no
3258 * progress in reclaiming pages
3260 if (raise_priority
|| !sc
.nr_reclaimed
)
3262 } while (sc
.priority
>= 1 &&
3263 !pgdat_balanced(pgdat
, order
, classzone_idx
));
3267 * Return the highest zone idx we were reclaiming at so
3268 * prepare_kswapd_sleep() makes the same decisions as here.
3273 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
,
3274 int classzone_idx
, int balanced_classzone_idx
)
3279 if (freezing(current
) || kthread_should_stop())
3282 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3284 /* Try to sleep for a short interval */
3285 if (prepare_kswapd_sleep(pgdat
, order
, remaining
,
3286 balanced_classzone_idx
)) {
3288 * Compaction records what page blocks it recently failed to
3289 * isolate pages from and skips them in the future scanning.
3290 * When kswapd is going to sleep, it is reasonable to assume
3291 * that pages and compaction may succeed so reset the cache.
3293 reset_isolation_suitable(pgdat
);
3296 * We have freed the memory, now we should compact it to make
3297 * allocation of the requested order possible.
3299 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3301 remaining
= schedule_timeout(HZ
/10);
3302 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3303 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3307 * After a short sleep, check if it was a premature sleep. If not, then
3308 * go fully to sleep until explicitly woken up.
3310 if (prepare_kswapd_sleep(pgdat
, order
, remaining
,
3311 balanced_classzone_idx
)) {
3312 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3315 * vmstat counters are not perfectly accurate and the estimated
3316 * value for counters such as NR_FREE_PAGES can deviate from the
3317 * true value by nr_online_cpus * threshold. To avoid the zone
3318 * watermarks being breached while under pressure, we reduce the
3319 * per-cpu vmstat threshold while kswapd is awake and restore
3320 * them before going back to sleep.
3322 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3324 if (!kthread_should_stop())
3327 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3330 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3332 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3334 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3338 * The background pageout daemon, started as a kernel thread
3339 * from the init process.
3341 * This basically trickles out pages so that we have _some_
3342 * free memory available even if there is no other activity
3343 * that frees anything up. This is needed for things like routing
3344 * etc, where we otherwise might have all activity going on in
3345 * asynchronous contexts that cannot page things out.
3347 * If there are applications that are active memory-allocators
3348 * (most normal use), this basically shouldn't matter.
3350 static int kswapd(void *p
)
3352 unsigned long order
, new_order
;
3353 int classzone_idx
, new_classzone_idx
;
3354 int balanced_classzone_idx
;
3355 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3356 struct task_struct
*tsk
= current
;
3358 struct reclaim_state reclaim_state
= {
3359 .reclaimed_slab
= 0,
3361 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3363 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3365 if (!cpumask_empty(cpumask
))
3366 set_cpus_allowed_ptr(tsk
, cpumask
);
3367 current
->reclaim_state
= &reclaim_state
;
3370 * Tell the memory management that we're a "memory allocator",
3371 * and that if we need more memory we should get access to it
3372 * regardless (see "__alloc_pages()"). "kswapd" should
3373 * never get caught in the normal page freeing logic.
3375 * (Kswapd normally doesn't need memory anyway, but sometimes
3376 * you need a small amount of memory in order to be able to
3377 * page out something else, and this flag essentially protects
3378 * us from recursively trying to free more memory as we're
3379 * trying to free the first piece of memory in the first place).
3381 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3384 order
= new_order
= 0;
3385 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3386 balanced_classzone_idx
= classzone_idx
;
3391 * While we were reclaiming, there might have been another
3392 * wakeup, so check the values.
3394 new_order
= pgdat
->kswapd_max_order
;
3395 new_classzone_idx
= pgdat
->classzone_idx
;
3396 pgdat
->kswapd_max_order
= 0;
3397 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3399 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3401 * Don't sleep if someone wants a larger 'order'
3402 * allocation or has tigher zone constraints
3405 classzone_idx
= new_classzone_idx
;
3407 kswapd_try_to_sleep(pgdat
, order
, classzone_idx
,
3408 balanced_classzone_idx
);
3409 order
= pgdat
->kswapd_max_order
;
3410 classzone_idx
= pgdat
->classzone_idx
;
3412 new_classzone_idx
= classzone_idx
;
3413 pgdat
->kswapd_max_order
= 0;
3414 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3417 ret
= try_to_freeze();
3418 if (kthread_should_stop())
3422 * We can speed up thawing tasks if we don't call balance_pgdat
3423 * after returning from the refrigerator
3426 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3427 balanced_classzone_idx
= balance_pgdat(pgdat
, order
,
3432 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3433 current
->reclaim_state
= NULL
;
3434 lockdep_clear_current_reclaim_state();
3440 * A zone is low on free memory, so wake its kswapd task to service it.
3442 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3446 if (!populated_zone(zone
))
3449 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3451 pgdat
= zone
->zone_pgdat
;
3452 if (pgdat
->kswapd_max_order
< order
) {
3453 pgdat
->kswapd_max_order
= order
;
3454 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3456 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3458 if (zone_balanced(zone
, order
, true, 0, 0))
3461 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3462 wake_up_interruptible(&pgdat
->kswapd_wait
);
3465 #ifdef CONFIG_HIBERNATION
3467 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3470 * Rather than trying to age LRUs the aim is to preserve the overall
3471 * LRU order by reclaiming preferentially
3472 * inactive > active > active referenced > active mapped
3474 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3476 struct reclaim_state reclaim_state
;
3477 struct scan_control sc
= {
3478 .nr_to_reclaim
= nr_to_reclaim
,
3479 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3480 .priority
= DEF_PRIORITY
,
3484 .hibernation_mode
= 1,
3486 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3487 struct task_struct
*p
= current
;
3488 unsigned long nr_reclaimed
;
3490 p
->flags
|= PF_MEMALLOC
;
3491 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3492 reclaim_state
.reclaimed_slab
= 0;
3493 p
->reclaim_state
= &reclaim_state
;
3495 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3497 p
->reclaim_state
= NULL
;
3498 lockdep_clear_current_reclaim_state();
3499 p
->flags
&= ~PF_MEMALLOC
;
3501 return nr_reclaimed
;
3503 #endif /* CONFIG_HIBERNATION */
3505 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3506 not required for correctness. So if the last cpu in a node goes
3507 away, we get changed to run anywhere: as the first one comes back,
3508 restore their cpu bindings. */
3509 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3514 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3515 for_each_node_state(nid
, N_MEMORY
) {
3516 pg_data_t
*pgdat
= NODE_DATA(nid
);
3517 const struct cpumask
*mask
;
3519 mask
= cpumask_of_node(pgdat
->node_id
);
3521 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3522 /* One of our CPUs online: restore mask */
3523 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3530 * This kswapd start function will be called by init and node-hot-add.
3531 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3533 int kswapd_run(int nid
)
3535 pg_data_t
*pgdat
= NODE_DATA(nid
);
3541 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3542 if (IS_ERR(pgdat
->kswapd
)) {
3543 /* failure at boot is fatal */
3544 BUG_ON(system_state
== SYSTEM_BOOTING
);
3545 pr_err("Failed to start kswapd on node %d\n", nid
);
3546 ret
= PTR_ERR(pgdat
->kswapd
);
3547 pgdat
->kswapd
= NULL
;
3553 * Called by memory hotplug when all memory in a node is offlined. Caller must
3554 * hold mem_hotplug_begin/end().
3556 void kswapd_stop(int nid
)
3558 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3561 kthread_stop(kswapd
);
3562 NODE_DATA(nid
)->kswapd
= NULL
;
3566 static int __init
kswapd_init(void)
3571 for_each_node_state(nid
, N_MEMORY
)
3573 hotcpu_notifier(cpu_callback
, 0);
3577 module_init(kswapd_init
)
3583 * If non-zero call zone_reclaim when the number of free pages falls below
3586 int zone_reclaim_mode __read_mostly
;
3588 #define RECLAIM_OFF 0
3589 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3590 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3591 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3594 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3595 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3598 #define ZONE_RECLAIM_PRIORITY 4
3601 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3604 int sysctl_min_unmapped_ratio
= 1;
3607 * If the number of slab pages in a zone grows beyond this percentage then
3608 * slab reclaim needs to occur.
3610 int sysctl_min_slab_ratio
= 5;
3612 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3614 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3615 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3616 zone_page_state(zone
, NR_ACTIVE_FILE
);
3619 * It's possible for there to be more file mapped pages than
3620 * accounted for by the pages on the file LRU lists because
3621 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3623 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3626 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3627 static unsigned long zone_pagecache_reclaimable(struct zone
*zone
)
3629 unsigned long nr_pagecache_reclaimable
;
3630 unsigned long delta
= 0;
3633 * If RECLAIM_UNMAP is set, then all file pages are considered
3634 * potentially reclaimable. Otherwise, we have to worry about
3635 * pages like swapcache and zone_unmapped_file_pages() provides
3638 if (zone_reclaim_mode
& RECLAIM_UNMAP
)
3639 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3641 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3643 /* If we can't clean pages, remove dirty pages from consideration */
3644 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3645 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3647 /* Watch for any possible underflows due to delta */
3648 if (unlikely(delta
> nr_pagecache_reclaimable
))
3649 delta
= nr_pagecache_reclaimable
;
3651 return nr_pagecache_reclaimable
- delta
;
3655 * Try to free up some pages from this zone through reclaim.
3657 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3659 /* Minimum pages needed in order to stay on node */
3660 const unsigned long nr_pages
= 1 << order
;
3661 struct task_struct
*p
= current
;
3662 struct reclaim_state reclaim_state
;
3663 struct scan_control sc
= {
3664 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3665 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3667 .priority
= ZONE_RECLAIM_PRIORITY
,
3668 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3669 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_UNMAP
),
3675 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3676 * and we also need to be able to write out pages for RECLAIM_WRITE
3677 * and RECLAIM_UNMAP.
3679 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3680 lockdep_set_current_reclaim_state(gfp_mask
);
3681 reclaim_state
.reclaimed_slab
= 0;
3682 p
->reclaim_state
= &reclaim_state
;
3684 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3686 * Free memory by calling shrink zone with increasing
3687 * priorities until we have enough memory freed.
3690 shrink_zone(zone
, &sc
, true);
3691 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3694 p
->reclaim_state
= NULL
;
3695 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3696 lockdep_clear_current_reclaim_state();
3697 return sc
.nr_reclaimed
>= nr_pages
;
3700 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3706 * Zone reclaim reclaims unmapped file backed pages and
3707 * slab pages if we are over the defined limits.
3709 * A small portion of unmapped file backed pages is needed for
3710 * file I/O otherwise pages read by file I/O will be immediately
3711 * thrown out if the zone is overallocated. So we do not reclaim
3712 * if less than a specified percentage of the zone is used by
3713 * unmapped file backed pages.
3715 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3716 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3717 return ZONE_RECLAIM_FULL
;
3719 if (!zone_reclaimable(zone
))
3720 return ZONE_RECLAIM_FULL
;
3723 * Do not scan if the allocation should not be delayed.
3725 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3726 return ZONE_RECLAIM_NOSCAN
;
3729 * Only run zone reclaim on the local zone or on zones that do not
3730 * have associated processors. This will favor the local processor
3731 * over remote processors and spread off node memory allocations
3732 * as wide as possible.
3734 node_id
= zone_to_nid(zone
);
3735 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3736 return ZONE_RECLAIM_NOSCAN
;
3738 if (test_and_set_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
))
3739 return ZONE_RECLAIM_NOSCAN
;
3741 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3742 clear_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
);
3745 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3752 * page_evictable - test whether a page is evictable
3753 * @page: the page to test
3755 * Test whether page is evictable--i.e., should be placed on active/inactive
3756 * lists vs unevictable list.
3758 * Reasons page might not be evictable:
3759 * (1) page's mapping marked unevictable
3760 * (2) page is part of an mlocked VMA
3763 int page_evictable(struct page
*page
)
3765 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3770 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3771 * @pages: array of pages to check
3772 * @nr_pages: number of pages to check
3774 * Checks pages for evictability and moves them to the appropriate lru list.
3776 * This function is only used for SysV IPC SHM_UNLOCK.
3778 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3780 struct lruvec
*lruvec
;
3781 struct zone
*zone
= NULL
;
3786 for (i
= 0; i
< nr_pages
; i
++) {
3787 struct page
*page
= pages
[i
];
3788 struct zone
*pagezone
;
3791 pagezone
= page_zone(page
);
3792 if (pagezone
!= zone
) {
3794 spin_unlock_irq(&zone
->lru_lock
);
3796 spin_lock_irq(&zone
->lru_lock
);
3798 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3800 if (!PageLRU(page
) || !PageUnevictable(page
))
3803 if (page_evictable(page
)) {
3804 enum lru_list lru
= page_lru_base_type(page
);
3806 VM_BUG_ON_PAGE(PageActive(page
), page
);
3807 ClearPageUnevictable(page
);
3808 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3809 add_page_to_lru_list(page
, lruvec
, lru
);
3815 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3816 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3817 spin_unlock_irq(&zone
->lru_lock
);
3820 #endif /* CONFIG_SHMEM */