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>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
53 #include <linux/swapops.h>
54 #include <linux/balloon_compaction.h>
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/vmscan.h>
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim
;
65 /* This context's GFP mask */
68 /* Allocation order */
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup
*target_mem_cgroup
;
83 /* Scan (total_size >> priority) pages at once */
86 unsigned int may_writepage
:1;
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap
:1;
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap
:1;
94 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash
:1;
97 unsigned int hibernation_mode
:1;
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready
:1;
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned
;
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed
;
109 #ifdef ARCH_HAS_PREFETCH
110 #define prefetch_prev_lru_page(_page, _base, _field) \
112 if ((_page)->lru.prev != _base) { \
115 prev = lru_to_page(&(_page->lru)); \
116 prefetch(&prev->_field); \
120 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
123 #ifdef ARCH_HAS_PREFETCHW
124 #define prefetchw_prev_lru_page(_page, _base, _field) \
126 if ((_page)->lru.prev != _base) { \
129 prev = lru_to_page(&(_page->lru)); \
130 prefetchw(&prev->_field); \
134 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
138 * From 0 .. 100. Higher means more swappy.
140 int vm_swappiness
= 60;
142 * The total number of pages which are beyond the high watermark within all
145 unsigned long vm_total_pages
;
147 static LIST_HEAD(shrinker_list
);
148 static DECLARE_RWSEM(shrinker_rwsem
);
151 static bool global_reclaim(struct scan_control
*sc
)
153 return !sc
->target_mem_cgroup
;
157 * sane_reclaim - is the usual dirty throttling mechanism operational?
158 * @sc: scan_control in question
160 * The normal page dirty throttling mechanism in balance_dirty_pages() is
161 * completely broken with the legacy memcg and direct stalling in
162 * shrink_page_list() is used for throttling instead, which lacks all the
163 * niceties such as fairness, adaptive pausing, bandwidth proportional
164 * allocation and configurability.
166 * This function tests whether the vmscan currently in progress can assume
167 * that the normal dirty throttling mechanism is operational.
169 static bool sane_reclaim(struct scan_control
*sc
)
171 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
175 #ifdef CONFIG_CGROUP_WRITEBACK
176 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
182 static bool global_reclaim(struct scan_control
*sc
)
187 static bool sane_reclaim(struct scan_control
*sc
)
193 static unsigned long zone_reclaimable_pages(struct zone
*zone
)
197 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
198 zone_page_state(zone
, NR_INACTIVE_FILE
) +
199 zone_page_state(zone
, NR_ISOLATED_FILE
);
201 if (get_nr_swap_pages() > 0)
202 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
203 zone_page_state(zone
, NR_INACTIVE_ANON
) +
204 zone_page_state(zone
, NR_ISOLATED_ANON
);
209 bool zone_reclaimable(struct zone
*zone
)
211 return zone_page_state(zone
, NR_PAGES_SCANNED
) <
212 zone_reclaimable_pages(zone
) * 6;
215 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
217 if (!mem_cgroup_disabled())
218 return mem_cgroup_get_lru_size(lruvec
, lru
);
220 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
224 * Add a shrinker callback to be called from the vm.
226 int register_shrinker(struct shrinker
*shrinker
)
228 size_t size
= sizeof(*shrinker
->nr_deferred
);
231 * If we only have one possible node in the system anyway, save
232 * ourselves the trouble and disable NUMA aware behavior. This way we
233 * will save memory and some small loop time later.
235 if (nr_node_ids
== 1)
236 shrinker
->flags
&= ~SHRINKER_NUMA_AWARE
;
238 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
241 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
242 if (!shrinker
->nr_deferred
)
245 down_write(&shrinker_rwsem
);
246 list_add_tail(&shrinker
->list
, &shrinker_list
);
247 up_write(&shrinker_rwsem
);
250 EXPORT_SYMBOL(register_shrinker
);
255 void unregister_shrinker(struct shrinker
*shrinker
)
257 down_write(&shrinker_rwsem
);
258 list_del(&shrinker
->list
);
259 up_write(&shrinker_rwsem
);
260 kfree(shrinker
->nr_deferred
);
262 EXPORT_SYMBOL(unregister_shrinker
);
264 #define SHRINK_BATCH 128
266 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
267 struct shrinker
*shrinker
,
268 unsigned long nr_scanned
,
269 unsigned long nr_eligible
)
271 unsigned long freed
= 0;
272 unsigned long long delta
;
277 int nid
= shrinkctl
->nid
;
278 long batch_size
= shrinker
->batch
? shrinker
->batch
281 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
286 * copy the current shrinker scan count into a local variable
287 * and zero it so that other concurrent shrinker invocations
288 * don't also do this scanning work.
290 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
293 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
295 do_div(delta
, nr_eligible
+ 1);
297 if (total_scan
< 0) {
298 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
299 shrinker
->scan_objects
, total_scan
);
300 total_scan
= freeable
;
304 * We need to avoid excessive windup on filesystem shrinkers
305 * due to large numbers of GFP_NOFS allocations causing the
306 * shrinkers to return -1 all the time. This results in a large
307 * nr being built up so when a shrink that can do some work
308 * comes along it empties the entire cache due to nr >>>
309 * freeable. This is bad for sustaining a working set in
312 * Hence only allow the shrinker to scan the entire cache when
313 * a large delta change is calculated directly.
315 if (delta
< freeable
/ 4)
316 total_scan
= min(total_scan
, freeable
/ 2);
319 * Avoid risking looping forever due to too large nr value:
320 * never try to free more than twice the estimate number of
323 if (total_scan
> freeable
* 2)
324 total_scan
= freeable
* 2;
326 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
327 nr_scanned
, nr_eligible
,
328 freeable
, delta
, total_scan
);
331 * Normally, we should not scan less than batch_size objects in one
332 * pass to avoid too frequent shrinker calls, but if the slab has less
333 * than batch_size objects in total and we are really tight on memory,
334 * we will try to reclaim all available objects, otherwise we can end
335 * up failing allocations although there are plenty of reclaimable
336 * objects spread over several slabs with usage less than the
339 * We detect the "tight on memory" situations by looking at the total
340 * number of objects we want to scan (total_scan). If it is greater
341 * than the total number of objects on slab (freeable), we must be
342 * scanning at high prio and therefore should try to reclaim as much as
345 while (total_scan
>= batch_size
||
346 total_scan
>= freeable
) {
348 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
350 shrinkctl
->nr_to_scan
= nr_to_scan
;
351 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
352 if (ret
== SHRINK_STOP
)
356 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
357 total_scan
-= nr_to_scan
;
363 * move the unused scan count back into the shrinker in a
364 * manner that handles concurrent updates. If we exhausted the
365 * scan, there is no need to do an update.
368 new_nr
= atomic_long_add_return(total_scan
,
369 &shrinker
->nr_deferred
[nid
]);
371 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
373 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
378 * shrink_slab - shrink slab caches
379 * @gfp_mask: allocation context
380 * @nid: node whose slab caches to target
381 * @memcg: memory cgroup whose slab caches to target
382 * @nr_scanned: pressure numerator
383 * @nr_eligible: pressure denominator
385 * Call the shrink functions to age shrinkable caches.
387 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
388 * unaware shrinkers will receive a node id of 0 instead.
390 * @memcg specifies the memory cgroup to target. If it is not NULL,
391 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
392 * objects from the memory cgroup specified. Otherwise all shrinkers
393 * are called, and memcg aware shrinkers are supposed to scan the
396 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
397 * the available objects should be scanned. Page reclaim for example
398 * passes the number of pages scanned and the number of pages on the
399 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
400 * when it encountered mapped pages. The ratio is further biased by
401 * the ->seeks setting of the shrink function, which indicates the
402 * cost to recreate an object relative to that of an LRU page.
404 * Returns the number of reclaimed slab objects.
406 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
407 struct mem_cgroup
*memcg
,
408 unsigned long nr_scanned
,
409 unsigned long nr_eligible
)
411 struct shrinker
*shrinker
;
412 unsigned long freed
= 0;
414 if (memcg
&& !memcg_kmem_is_active(memcg
))
418 nr_scanned
= SWAP_CLUSTER_MAX
;
420 if (!down_read_trylock(&shrinker_rwsem
)) {
422 * If we would return 0, our callers would understand that we
423 * have nothing else to shrink and give up trying. By returning
424 * 1 we keep it going and assume we'll be able to shrink next
431 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
432 struct shrink_control sc
= {
433 .gfp_mask
= gfp_mask
,
438 if (memcg
&& !(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
441 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
444 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
447 up_read(&shrinker_rwsem
);
453 void drop_slab_node(int nid
)
458 struct mem_cgroup
*memcg
= NULL
;
462 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
464 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
465 } while (freed
> 10);
472 for_each_online_node(nid
)
476 static inline int is_page_cache_freeable(struct page
*page
)
479 * A freeable page cache page is referenced only by the caller
480 * that isolated the page, the page cache radix tree and
481 * optional buffer heads at page->private.
483 return page_count(page
) - page_has_private(page
) == 2;
486 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
488 if (current
->flags
& PF_SWAPWRITE
)
490 if (!inode_write_congested(inode
))
492 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
498 * We detected a synchronous write error writing a page out. Probably
499 * -ENOSPC. We need to propagate that into the address_space for a subsequent
500 * fsync(), msync() or close().
502 * The tricky part is that after writepage we cannot touch the mapping: nothing
503 * prevents it from being freed up. But we have a ref on the page and once
504 * that page is locked, the mapping is pinned.
506 * We're allowed to run sleeping lock_page() here because we know the caller has
509 static void handle_write_error(struct address_space
*mapping
,
510 struct page
*page
, int error
)
513 if (page_mapping(page
) == mapping
)
514 mapping_set_error(mapping
, error
);
518 /* possible outcome of pageout() */
520 /* failed to write page out, page is locked */
522 /* move page to the active list, page is locked */
524 /* page has been sent to the disk successfully, page is unlocked */
526 /* page is clean and locked */
531 * pageout is called by shrink_page_list() for each dirty page.
532 * Calls ->writepage().
534 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
535 struct scan_control
*sc
)
538 * If the page is dirty, only perform writeback if that write
539 * will be non-blocking. To prevent this allocation from being
540 * stalled by pagecache activity. But note that there may be
541 * stalls if we need to run get_block(). We could test
542 * PagePrivate for that.
544 * If this process is currently in __generic_file_write_iter() against
545 * this page's queue, we can perform writeback even if that
548 * If the page is swapcache, write it back even if that would
549 * block, for some throttling. This happens by accident, because
550 * swap_backing_dev_info is bust: it doesn't reflect the
551 * congestion state of the swapdevs. Easy to fix, if needed.
553 if (!is_page_cache_freeable(page
))
557 * Some data journaling orphaned pages can have
558 * page->mapping == NULL while being dirty with clean buffers.
560 if (page_has_private(page
)) {
561 if (try_to_free_buffers(page
)) {
562 ClearPageDirty(page
);
563 pr_info("%s: orphaned page\n", __func__
);
569 if (mapping
->a_ops
->writepage
== NULL
)
570 return PAGE_ACTIVATE
;
571 if (!may_write_to_inode(mapping
->host
, sc
))
574 if (clear_page_dirty_for_io(page
)) {
576 struct writeback_control wbc
= {
577 .sync_mode
= WB_SYNC_NONE
,
578 .nr_to_write
= SWAP_CLUSTER_MAX
,
580 .range_end
= LLONG_MAX
,
584 SetPageReclaim(page
);
585 res
= mapping
->a_ops
->writepage(page
, &wbc
);
587 handle_write_error(mapping
, page
, res
);
588 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
589 ClearPageReclaim(page
);
590 return PAGE_ACTIVATE
;
593 if (!PageWriteback(page
)) {
594 /* synchronous write or broken a_ops? */
595 ClearPageReclaim(page
);
597 trace_mm_vmscan_writepage(page
);
598 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
606 * Same as remove_mapping, but if the page is removed from the mapping, it
607 * gets returned with a refcount of 0.
609 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
613 struct mem_cgroup
*memcg
;
615 BUG_ON(!PageLocked(page
));
616 BUG_ON(mapping
!= page_mapping(page
));
618 memcg
= mem_cgroup_begin_page_stat(page
);
619 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
621 * The non racy check for a busy page.
623 * Must be careful with the order of the tests. When someone has
624 * a ref to the page, it may be possible that they dirty it then
625 * drop the reference. So if PageDirty is tested before page_count
626 * here, then the following race may occur:
628 * get_user_pages(&page);
629 * [user mapping goes away]
631 * !PageDirty(page) [good]
632 * SetPageDirty(page);
634 * !page_count(page) [good, discard it]
636 * [oops, our write_to data is lost]
638 * Reversing the order of the tests ensures such a situation cannot
639 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
640 * load is not satisfied before that of page->_count.
642 * Note that if SetPageDirty is always performed via set_page_dirty,
643 * and thus under tree_lock, then this ordering is not required.
645 if (!page_freeze_refs(page
, 2))
647 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
648 if (unlikely(PageDirty(page
))) {
649 page_unfreeze_refs(page
, 2);
653 if (PageSwapCache(page
)) {
654 swp_entry_t swap
= { .val
= page_private(page
) };
655 mem_cgroup_swapout(page
, swap
);
656 __delete_from_swap_cache(page
);
657 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
658 mem_cgroup_end_page_stat(memcg
);
659 swapcache_free(swap
);
661 void (*freepage
)(struct page
*);
664 freepage
= mapping
->a_ops
->freepage
;
666 * Remember a shadow entry for reclaimed file cache in
667 * order to detect refaults, thus thrashing, later on.
669 * But don't store shadows in an address space that is
670 * already exiting. This is not just an optizimation,
671 * inode reclaim needs to empty out the radix tree or
672 * the nodes are lost. Don't plant shadows behind its
675 if (reclaimed
&& page_is_file_cache(page
) &&
676 !mapping_exiting(mapping
))
677 shadow
= workingset_eviction(mapping
, page
);
678 __delete_from_page_cache(page
, shadow
, memcg
);
679 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
680 mem_cgroup_end_page_stat(memcg
);
682 if (freepage
!= NULL
)
689 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
690 mem_cgroup_end_page_stat(memcg
);
695 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
696 * someone else has a ref on the page, abort and return 0. If it was
697 * successfully detached, return 1. Assumes the caller has a single ref on
700 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
702 if (__remove_mapping(mapping
, page
, false)) {
704 * Unfreezing the refcount with 1 rather than 2 effectively
705 * drops the pagecache ref for us without requiring another
708 page_unfreeze_refs(page
, 1);
715 * putback_lru_page - put previously isolated page onto appropriate LRU list
716 * @page: page to be put back to appropriate lru list
718 * Add previously isolated @page to appropriate LRU list.
719 * Page may still be unevictable for other reasons.
721 * lru_lock must not be held, interrupts must be enabled.
723 void putback_lru_page(struct page
*page
)
726 int was_unevictable
= PageUnevictable(page
);
728 VM_BUG_ON_PAGE(PageLRU(page
), page
);
731 ClearPageUnevictable(page
);
733 if (page_evictable(page
)) {
735 * For evictable pages, we can use the cache.
736 * In event of a race, worst case is we end up with an
737 * unevictable page on [in]active list.
738 * We know how to handle that.
740 is_unevictable
= false;
744 * Put unevictable pages directly on zone's unevictable
747 is_unevictable
= true;
748 add_page_to_unevictable_list(page
);
750 * When racing with an mlock or AS_UNEVICTABLE clearing
751 * (page is unlocked) make sure that if the other thread
752 * does not observe our setting of PG_lru and fails
753 * isolation/check_move_unevictable_pages,
754 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
755 * the page back to the evictable list.
757 * The other side is TestClearPageMlocked() or shmem_lock().
763 * page's status can change while we move it among lru. If an evictable
764 * page is on unevictable list, it never be freed. To avoid that,
765 * check after we added it to the list, again.
767 if (is_unevictable
&& page_evictable(page
)) {
768 if (!isolate_lru_page(page
)) {
772 /* This means someone else dropped this page from LRU
773 * So, it will be freed or putback to LRU again. There is
774 * nothing to do here.
778 if (was_unevictable
&& !is_unevictable
)
779 count_vm_event(UNEVICTABLE_PGRESCUED
);
780 else if (!was_unevictable
&& is_unevictable
)
781 count_vm_event(UNEVICTABLE_PGCULLED
);
783 put_page(page
); /* drop ref from isolate */
786 enum page_references
{
788 PAGEREF_RECLAIM_CLEAN
,
793 static enum page_references
page_check_references(struct page
*page
,
794 struct scan_control
*sc
)
796 int referenced_ptes
, referenced_page
;
797 unsigned long vm_flags
;
799 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
801 referenced_page
= TestClearPageReferenced(page
);
804 * Mlock lost the isolation race with us. Let try_to_unmap()
805 * move the page to the unevictable list.
807 if (vm_flags
& VM_LOCKED
)
808 return PAGEREF_RECLAIM
;
810 if (referenced_ptes
) {
811 if (PageSwapBacked(page
))
812 return PAGEREF_ACTIVATE
;
814 * All mapped pages start out with page table
815 * references from the instantiating fault, so we need
816 * to look twice if a mapped file page is used more
819 * Mark it and spare it for another trip around the
820 * inactive list. Another page table reference will
821 * lead to its activation.
823 * Note: the mark is set for activated pages as well
824 * so that recently deactivated but used pages are
827 SetPageReferenced(page
);
829 if (referenced_page
|| referenced_ptes
> 1)
830 return PAGEREF_ACTIVATE
;
833 * Activate file-backed executable pages after first usage.
835 if (vm_flags
& VM_EXEC
)
836 return PAGEREF_ACTIVATE
;
841 /* Reclaim if clean, defer dirty pages to writeback */
842 if (referenced_page
&& !PageSwapBacked(page
))
843 return PAGEREF_RECLAIM_CLEAN
;
845 return PAGEREF_RECLAIM
;
848 /* Check if a page is dirty or under writeback */
849 static void page_check_dirty_writeback(struct page
*page
,
850 bool *dirty
, bool *writeback
)
852 struct address_space
*mapping
;
855 * Anonymous pages are not handled by flushers and must be written
856 * from reclaim context. Do not stall reclaim based on them
858 if (!page_is_file_cache(page
)) {
864 /* By default assume that the page flags are accurate */
865 *dirty
= PageDirty(page
);
866 *writeback
= PageWriteback(page
);
868 /* Verify dirty/writeback state if the filesystem supports it */
869 if (!page_has_private(page
))
872 mapping
= page_mapping(page
);
873 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
874 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
878 * shrink_page_list() returns the number of reclaimed pages
880 static unsigned long shrink_page_list(struct list_head
*page_list
,
882 struct scan_control
*sc
,
883 enum ttu_flags ttu_flags
,
884 unsigned long *ret_nr_dirty
,
885 unsigned long *ret_nr_unqueued_dirty
,
886 unsigned long *ret_nr_congested
,
887 unsigned long *ret_nr_writeback
,
888 unsigned long *ret_nr_immediate
,
891 LIST_HEAD(ret_pages
);
892 LIST_HEAD(free_pages
);
894 unsigned long nr_unqueued_dirty
= 0;
895 unsigned long nr_dirty
= 0;
896 unsigned long nr_congested
= 0;
897 unsigned long nr_reclaimed
= 0;
898 unsigned long nr_writeback
= 0;
899 unsigned long nr_immediate
= 0;
903 while (!list_empty(page_list
)) {
904 struct address_space
*mapping
;
907 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
908 bool dirty
, writeback
;
909 bool lazyfree
= false;
910 int ret
= SWAP_SUCCESS
;
914 page
= lru_to_page(page_list
);
915 list_del(&page
->lru
);
917 if (!trylock_page(page
))
920 VM_BUG_ON_PAGE(PageActive(page
), page
);
921 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
925 if (unlikely(!page_evictable(page
)))
928 if (!sc
->may_unmap
&& page_mapped(page
))
931 /* Double the slab pressure for mapped and swapcache pages */
932 if (page_mapped(page
) || PageSwapCache(page
))
935 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
936 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
939 * The number of dirty pages determines if a zone is marked
940 * reclaim_congested which affects wait_iff_congested. kswapd
941 * will stall and start writing pages if the tail of the LRU
942 * is all dirty unqueued pages.
944 page_check_dirty_writeback(page
, &dirty
, &writeback
);
945 if (dirty
|| writeback
)
948 if (dirty
&& !writeback
)
952 * Treat this page as congested if the underlying BDI is or if
953 * pages are cycling through the LRU so quickly that the
954 * pages marked for immediate reclaim are making it to the
955 * end of the LRU a second time.
957 mapping
= page_mapping(page
);
958 if (((dirty
|| writeback
) && mapping
&&
959 inode_write_congested(mapping
->host
)) ||
960 (writeback
&& PageReclaim(page
)))
964 * If a page at the tail of the LRU is under writeback, there
965 * are three cases to consider.
967 * 1) If reclaim is encountering an excessive number of pages
968 * under writeback and this page is both under writeback and
969 * PageReclaim then it indicates that pages are being queued
970 * for IO but are being recycled through the LRU before the
971 * IO can complete. Waiting on the page itself risks an
972 * indefinite stall if it is impossible to writeback the
973 * page due to IO error or disconnected storage so instead
974 * note that the LRU is being scanned too quickly and the
975 * caller can stall after page list has been processed.
977 * 2) Global or new memcg reclaim encounters a page that is
978 * not marked for immediate reclaim, or the caller does not
979 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
980 * not to fs). In this case mark the page for immediate
981 * reclaim and continue scanning.
983 * Require may_enter_fs because we would wait on fs, which
984 * may not have submitted IO yet. And the loop driver might
985 * enter reclaim, and deadlock if it waits on a page for
986 * which it is needed to do the write (loop masks off
987 * __GFP_IO|__GFP_FS for this reason); but more thought
988 * would probably show more reasons.
990 * 3) Legacy memcg encounters a page that is already marked
991 * PageReclaim. memcg does not have any dirty pages
992 * throttling so we could easily OOM just because too many
993 * pages are in writeback and there is nothing else to
994 * reclaim. Wait for the writeback to complete.
996 if (PageWriteback(page
)) {
998 if (current_is_kswapd() &&
1000 test_bit(ZONE_WRITEBACK
, &zone
->flags
)) {
1005 } else if (sane_reclaim(sc
) ||
1006 !PageReclaim(page
) || !may_enter_fs
) {
1008 * This is slightly racy - end_page_writeback()
1009 * might have just cleared PageReclaim, then
1010 * setting PageReclaim here end up interpreted
1011 * as PageReadahead - but that does not matter
1012 * enough to care. What we do want is for this
1013 * page to have PageReclaim set next time memcg
1014 * reclaim reaches the tests above, so it will
1015 * then wait_on_page_writeback() to avoid OOM;
1016 * and it's also appropriate in global reclaim.
1018 SetPageReclaim(page
);
1025 wait_on_page_writeback(page
);
1026 /* then go back and try same page again */
1027 list_add_tail(&page
->lru
, page_list
);
1033 references
= page_check_references(page
, sc
);
1035 switch (references
) {
1036 case PAGEREF_ACTIVATE
:
1037 goto activate_locked
;
1040 case PAGEREF_RECLAIM
:
1041 case PAGEREF_RECLAIM_CLEAN
:
1042 ; /* try to reclaim the page below */
1046 * Anonymous process memory has backing store?
1047 * Try to allocate it some swap space here.
1049 if (PageAnon(page
) && !PageSwapCache(page
)) {
1050 if (!(sc
->gfp_mask
& __GFP_IO
))
1052 if (!add_to_swap(page
, page_list
))
1053 goto activate_locked
;
1057 /* Adding to swap updated mapping */
1058 mapping
= page_mapping(page
);
1062 * The page is mapped into the page tables of one or more
1063 * processes. Try to unmap it here.
1065 if (page_mapped(page
) && mapping
) {
1066 switch (ret
= try_to_unmap(page
, lazyfree
?
1067 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1068 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1070 goto activate_locked
;
1078 ; /* try to free the page below */
1082 if (PageDirty(page
)) {
1084 * Only kswapd can writeback filesystem pages to
1085 * avoid risk of stack overflow but only writeback
1086 * if many dirty pages have been encountered.
1088 if (page_is_file_cache(page
) &&
1089 (!current_is_kswapd() ||
1090 !test_bit(ZONE_DIRTY
, &zone
->flags
))) {
1092 * Immediately reclaim when written back.
1093 * Similar in principal to deactivate_page()
1094 * except we already have the page isolated
1095 * and know it's dirty
1097 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1098 SetPageReclaim(page
);
1103 if (references
== PAGEREF_RECLAIM_CLEAN
)
1107 if (!sc
->may_writepage
)
1111 * Page is dirty. Flush the TLB if a writable entry
1112 * potentially exists to avoid CPU writes after IO
1113 * starts and then write it out here.
1115 try_to_unmap_flush_dirty();
1116 switch (pageout(page
, mapping
, sc
)) {
1120 goto activate_locked
;
1122 if (PageWriteback(page
))
1124 if (PageDirty(page
))
1128 * A synchronous write - probably a ramdisk. Go
1129 * ahead and try to reclaim the page.
1131 if (!trylock_page(page
))
1133 if (PageDirty(page
) || PageWriteback(page
))
1135 mapping
= page_mapping(page
);
1137 ; /* try to free the page below */
1142 * If the page has buffers, try to free the buffer mappings
1143 * associated with this page. If we succeed we try to free
1146 * We do this even if the page is PageDirty().
1147 * try_to_release_page() does not perform I/O, but it is
1148 * possible for a page to have PageDirty set, but it is actually
1149 * clean (all its buffers are clean). This happens if the
1150 * buffers were written out directly, with submit_bh(). ext3
1151 * will do this, as well as the blockdev mapping.
1152 * try_to_release_page() will discover that cleanness and will
1153 * drop the buffers and mark the page clean - it can be freed.
1155 * Rarely, pages can have buffers and no ->mapping. These are
1156 * the pages which were not successfully invalidated in
1157 * truncate_complete_page(). We try to drop those buffers here
1158 * and if that worked, and the page is no longer mapped into
1159 * process address space (page_count == 1) it can be freed.
1160 * Otherwise, leave the page on the LRU so it is swappable.
1162 if (page_has_private(page
)) {
1163 if (!try_to_release_page(page
, sc
->gfp_mask
))
1164 goto activate_locked
;
1165 if (!mapping
&& page_count(page
) == 1) {
1167 if (put_page_testzero(page
))
1171 * rare race with speculative reference.
1172 * the speculative reference will free
1173 * this page shortly, so we may
1174 * increment nr_reclaimed here (and
1175 * leave it off the LRU).
1184 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1188 * At this point, we have no other references and there is
1189 * no way to pick any more up (removed from LRU, removed
1190 * from pagecache). Can use non-atomic bitops now (and
1191 * we obviously don't have to worry about waking up a process
1192 * waiting on the page lock, because there are no references.
1194 __ClearPageLocked(page
);
1196 if (ret
== SWAP_LZFREE
)
1197 count_vm_event(PGLAZYFREED
);
1202 * Is there need to periodically free_page_list? It would
1203 * appear not as the counts should be low
1205 list_add(&page
->lru
, &free_pages
);
1209 if (PageSwapCache(page
))
1210 try_to_free_swap(page
);
1212 list_add(&page
->lru
, &ret_pages
);
1216 /* Not a candidate for swapping, so reclaim swap space. */
1217 if (PageSwapCache(page
) && vm_swap_full())
1218 try_to_free_swap(page
);
1219 VM_BUG_ON_PAGE(PageActive(page
), page
);
1220 SetPageActive(page
);
1225 list_add(&page
->lru
, &ret_pages
);
1226 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1229 mem_cgroup_uncharge_list(&free_pages
);
1230 try_to_unmap_flush();
1231 free_hot_cold_page_list(&free_pages
, true);
1233 list_splice(&ret_pages
, page_list
);
1234 count_vm_events(PGACTIVATE
, pgactivate
);
1236 *ret_nr_dirty
+= nr_dirty
;
1237 *ret_nr_congested
+= nr_congested
;
1238 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1239 *ret_nr_writeback
+= nr_writeback
;
1240 *ret_nr_immediate
+= nr_immediate
;
1241 return nr_reclaimed
;
1244 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1245 struct list_head
*page_list
)
1247 struct scan_control sc
= {
1248 .gfp_mask
= GFP_KERNEL
,
1249 .priority
= DEF_PRIORITY
,
1252 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1253 struct page
*page
, *next
;
1254 LIST_HEAD(clean_pages
);
1256 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1257 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1258 !isolated_balloon_page(page
)) {
1259 ClearPageActive(page
);
1260 list_move(&page
->lru
, &clean_pages
);
1264 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1265 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1266 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1267 list_splice(&clean_pages
, page_list
);
1268 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1273 * Attempt to remove the specified page from its LRU. Only take this page
1274 * if it is of the appropriate PageActive status. Pages which are being
1275 * freed elsewhere are also ignored.
1277 * page: page to consider
1278 * mode: one of the LRU isolation modes defined above
1280 * returns 0 on success, -ve errno on failure.
1282 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1286 /* Only take pages on the LRU. */
1290 /* Compaction should not handle unevictable pages but CMA can do so */
1291 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1297 * To minimise LRU disruption, the caller can indicate that it only
1298 * wants to isolate pages it will be able to operate on without
1299 * blocking - clean pages for the most part.
1301 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1302 * is used by reclaim when it is cannot write to backing storage
1304 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1305 * that it is possible to migrate without blocking
1307 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1308 /* All the caller can do on PageWriteback is block */
1309 if (PageWriteback(page
))
1312 if (PageDirty(page
)) {
1313 struct address_space
*mapping
;
1315 /* ISOLATE_CLEAN means only clean pages */
1316 if (mode
& ISOLATE_CLEAN
)
1320 * Only pages without mappings or that have a
1321 * ->migratepage callback are possible to migrate
1324 mapping
= page_mapping(page
);
1325 if (mapping
&& !mapping
->a_ops
->migratepage
)
1330 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1333 if (likely(get_page_unless_zero(page
))) {
1335 * Be careful not to clear PageLRU until after we're
1336 * sure the page is not being freed elsewhere -- the
1337 * page release code relies on it.
1347 * zone->lru_lock is heavily contended. Some of the functions that
1348 * shrink the lists perform better by taking out a batch of pages
1349 * and working on them outside the LRU lock.
1351 * For pagecache intensive workloads, this function is the hottest
1352 * spot in the kernel (apart from copy_*_user functions).
1354 * Appropriate locks must be held before calling this function.
1356 * @nr_to_scan: The number of pages to look through on the list.
1357 * @lruvec: The LRU vector to pull pages from.
1358 * @dst: The temp list to put pages on to.
1359 * @nr_scanned: The number of pages that were scanned.
1360 * @sc: The scan_control struct for this reclaim session
1361 * @mode: One of the LRU isolation modes
1362 * @lru: LRU list id for isolating
1364 * returns how many pages were moved onto *@dst.
1366 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1367 struct lruvec
*lruvec
, struct list_head
*dst
,
1368 unsigned long *nr_scanned
, struct scan_control
*sc
,
1369 isolate_mode_t mode
, enum lru_list lru
)
1371 struct list_head
*src
= &lruvec
->lists
[lru
];
1372 unsigned long nr_taken
= 0;
1375 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1376 !list_empty(src
); scan
++) {
1380 page
= lru_to_page(src
);
1381 prefetchw_prev_lru_page(page
, src
, flags
);
1383 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1385 switch (__isolate_lru_page(page
, mode
)) {
1387 nr_pages
= hpage_nr_pages(page
);
1388 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1389 list_move(&page
->lru
, dst
);
1390 nr_taken
+= nr_pages
;
1394 /* else it is being freed elsewhere */
1395 list_move(&page
->lru
, src
);
1404 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1405 nr_taken
, mode
, is_file_lru(lru
));
1410 * isolate_lru_page - tries to isolate a page from its LRU list
1411 * @page: page to isolate from its LRU list
1413 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1414 * vmstat statistic corresponding to whatever LRU list the page was on.
1416 * Returns 0 if the page was removed from an LRU list.
1417 * Returns -EBUSY if the page was not on an LRU list.
1419 * The returned page will have PageLRU() cleared. If it was found on
1420 * the active list, it will have PageActive set. If it was found on
1421 * the unevictable list, it will have the PageUnevictable bit set. That flag
1422 * may need to be cleared by the caller before letting the page go.
1424 * The vmstat statistic corresponding to the list on which the page was
1425 * found will be decremented.
1428 * (1) Must be called with an elevated refcount on the page. This is a
1429 * fundamentnal difference from isolate_lru_pages (which is called
1430 * without a stable reference).
1431 * (2) the lru_lock must not be held.
1432 * (3) interrupts must be enabled.
1434 int isolate_lru_page(struct page
*page
)
1438 VM_BUG_ON_PAGE(!page_count(page
), page
);
1439 VM_BUG_ON_PAGE(PageTail(page
), page
);
1441 if (PageLRU(page
)) {
1442 struct zone
*zone
= page_zone(page
);
1443 struct lruvec
*lruvec
;
1445 spin_lock_irq(&zone
->lru_lock
);
1446 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1447 if (PageLRU(page
)) {
1448 int lru
= page_lru(page
);
1451 del_page_from_lru_list(page
, lruvec
, lru
);
1454 spin_unlock_irq(&zone
->lru_lock
);
1460 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1461 * then get resheduled. When there are massive number of tasks doing page
1462 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1463 * the LRU list will go small and be scanned faster than necessary, leading to
1464 * unnecessary swapping, thrashing and OOM.
1466 static int too_many_isolated(struct zone
*zone
, int file
,
1467 struct scan_control
*sc
)
1469 unsigned long inactive
, isolated
;
1471 if (current_is_kswapd())
1474 if (!sane_reclaim(sc
))
1478 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1479 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1481 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1482 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1486 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1487 * won't get blocked by normal direct-reclaimers, forming a circular
1490 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1493 return isolated
> inactive
;
1496 static noinline_for_stack
void
1497 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1499 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1500 struct zone
*zone
= lruvec_zone(lruvec
);
1501 LIST_HEAD(pages_to_free
);
1504 * Put back any unfreeable pages.
1506 while (!list_empty(page_list
)) {
1507 struct page
*page
= lru_to_page(page_list
);
1510 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1511 list_del(&page
->lru
);
1512 if (unlikely(!page_evictable(page
))) {
1513 spin_unlock_irq(&zone
->lru_lock
);
1514 putback_lru_page(page
);
1515 spin_lock_irq(&zone
->lru_lock
);
1519 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1522 lru
= page_lru(page
);
1523 add_page_to_lru_list(page
, lruvec
, lru
);
1525 if (is_active_lru(lru
)) {
1526 int file
= is_file_lru(lru
);
1527 int numpages
= hpage_nr_pages(page
);
1528 reclaim_stat
->recent_rotated
[file
] += numpages
;
1530 if (put_page_testzero(page
)) {
1531 __ClearPageLRU(page
);
1532 __ClearPageActive(page
);
1533 del_page_from_lru_list(page
, lruvec
, lru
);
1535 if (unlikely(PageCompound(page
))) {
1536 spin_unlock_irq(&zone
->lru_lock
);
1537 mem_cgroup_uncharge(page
);
1538 (*get_compound_page_dtor(page
))(page
);
1539 spin_lock_irq(&zone
->lru_lock
);
1541 list_add(&page
->lru
, &pages_to_free
);
1546 * To save our caller's stack, now use input list for pages to free.
1548 list_splice(&pages_to_free
, page_list
);
1552 * If a kernel thread (such as nfsd for loop-back mounts) services
1553 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1554 * In that case we should only throttle if the backing device it is
1555 * writing to is congested. In other cases it is safe to throttle.
1557 static int current_may_throttle(void)
1559 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1560 current
->backing_dev_info
== NULL
||
1561 bdi_write_congested(current
->backing_dev_info
);
1565 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1566 * of reclaimed pages
1568 static noinline_for_stack
unsigned long
1569 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1570 struct scan_control
*sc
, enum lru_list lru
)
1572 LIST_HEAD(page_list
);
1573 unsigned long nr_scanned
;
1574 unsigned long nr_reclaimed
= 0;
1575 unsigned long nr_taken
;
1576 unsigned long nr_dirty
= 0;
1577 unsigned long nr_congested
= 0;
1578 unsigned long nr_unqueued_dirty
= 0;
1579 unsigned long nr_writeback
= 0;
1580 unsigned long nr_immediate
= 0;
1581 isolate_mode_t isolate_mode
= 0;
1582 int file
= is_file_lru(lru
);
1583 struct zone
*zone
= lruvec_zone(lruvec
);
1584 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1586 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1587 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1589 /* We are about to die and free our memory. Return now. */
1590 if (fatal_signal_pending(current
))
1591 return SWAP_CLUSTER_MAX
;
1597 isolate_mode
|= ISOLATE_UNMAPPED
;
1598 if (!sc
->may_writepage
)
1599 isolate_mode
|= ISOLATE_CLEAN
;
1601 spin_lock_irq(&zone
->lru_lock
);
1603 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1604 &nr_scanned
, sc
, isolate_mode
, lru
);
1606 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1607 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1609 if (global_reclaim(sc
)) {
1610 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1611 if (current_is_kswapd())
1612 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1614 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1616 spin_unlock_irq(&zone
->lru_lock
);
1621 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1622 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1623 &nr_writeback
, &nr_immediate
,
1626 spin_lock_irq(&zone
->lru_lock
);
1628 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1630 if (global_reclaim(sc
)) {
1631 if (current_is_kswapd())
1632 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1635 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1639 putback_inactive_pages(lruvec
, &page_list
);
1641 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1643 spin_unlock_irq(&zone
->lru_lock
);
1645 mem_cgroup_uncharge_list(&page_list
);
1646 free_hot_cold_page_list(&page_list
, true);
1649 * If reclaim is isolating dirty pages under writeback, it implies
1650 * that the long-lived page allocation rate is exceeding the page
1651 * laundering rate. Either the global limits are not being effective
1652 * at throttling processes due to the page distribution throughout
1653 * zones or there is heavy usage of a slow backing device. The
1654 * only option is to throttle from reclaim context which is not ideal
1655 * as there is no guarantee the dirtying process is throttled in the
1656 * same way balance_dirty_pages() manages.
1658 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1659 * of pages under pages flagged for immediate reclaim and stall if any
1660 * are encountered in the nr_immediate check below.
1662 if (nr_writeback
&& nr_writeback
== nr_taken
)
1663 set_bit(ZONE_WRITEBACK
, &zone
->flags
);
1666 * Legacy memcg will stall in page writeback so avoid forcibly
1669 if (sane_reclaim(sc
)) {
1671 * Tag a zone as congested if all the dirty pages scanned were
1672 * backed by a congested BDI and wait_iff_congested will stall.
1674 if (nr_dirty
&& nr_dirty
== nr_congested
)
1675 set_bit(ZONE_CONGESTED
, &zone
->flags
);
1678 * If dirty pages are scanned that are not queued for IO, it
1679 * implies that flushers are not keeping up. In this case, flag
1680 * the zone ZONE_DIRTY and kswapd will start writing pages from
1683 if (nr_unqueued_dirty
== nr_taken
)
1684 set_bit(ZONE_DIRTY
, &zone
->flags
);
1687 * If kswapd scans pages marked marked for immediate
1688 * reclaim and under writeback (nr_immediate), it implies
1689 * that pages are cycling through the LRU faster than
1690 * they are written so also forcibly stall.
1692 if (nr_immediate
&& current_may_throttle())
1693 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1697 * Stall direct reclaim for IO completions if underlying BDIs or zone
1698 * is congested. Allow kswapd to continue until it starts encountering
1699 * unqueued dirty pages or cycling through the LRU too quickly.
1701 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1702 current_may_throttle())
1703 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1705 trace_mm_vmscan_lru_shrink_inactive(zone
, nr_scanned
, nr_reclaimed
,
1706 sc
->priority
, file
);
1707 return nr_reclaimed
;
1711 * This moves pages from the active list to the inactive list.
1713 * We move them the other way if the page is referenced by one or more
1714 * processes, from rmap.
1716 * If the pages are mostly unmapped, the processing is fast and it is
1717 * appropriate to hold zone->lru_lock across the whole operation. But if
1718 * the pages are mapped, the processing is slow (page_referenced()) so we
1719 * should drop zone->lru_lock around each page. It's impossible to balance
1720 * this, so instead we remove the pages from the LRU while processing them.
1721 * It is safe to rely on PG_active against the non-LRU pages in here because
1722 * nobody will play with that bit on a non-LRU page.
1724 * The downside is that we have to touch page->_count against each page.
1725 * But we had to alter page->flags anyway.
1728 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1729 struct list_head
*list
,
1730 struct list_head
*pages_to_free
,
1733 struct zone
*zone
= lruvec_zone(lruvec
);
1734 unsigned long pgmoved
= 0;
1738 while (!list_empty(list
)) {
1739 page
= lru_to_page(list
);
1740 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1742 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1745 nr_pages
= hpage_nr_pages(page
);
1746 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1747 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1748 pgmoved
+= nr_pages
;
1750 if (put_page_testzero(page
)) {
1751 __ClearPageLRU(page
);
1752 __ClearPageActive(page
);
1753 del_page_from_lru_list(page
, lruvec
, lru
);
1755 if (unlikely(PageCompound(page
))) {
1756 spin_unlock_irq(&zone
->lru_lock
);
1757 mem_cgroup_uncharge(page
);
1758 (*get_compound_page_dtor(page
))(page
);
1759 spin_lock_irq(&zone
->lru_lock
);
1761 list_add(&page
->lru
, pages_to_free
);
1764 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1765 if (!is_active_lru(lru
))
1766 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1769 static void shrink_active_list(unsigned long nr_to_scan
,
1770 struct lruvec
*lruvec
,
1771 struct scan_control
*sc
,
1774 unsigned long nr_taken
;
1775 unsigned long nr_scanned
;
1776 unsigned long vm_flags
;
1777 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1778 LIST_HEAD(l_active
);
1779 LIST_HEAD(l_inactive
);
1781 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1782 unsigned long nr_rotated
= 0;
1783 isolate_mode_t isolate_mode
= 0;
1784 int file
= is_file_lru(lru
);
1785 struct zone
*zone
= lruvec_zone(lruvec
);
1790 isolate_mode
|= ISOLATE_UNMAPPED
;
1791 if (!sc
->may_writepage
)
1792 isolate_mode
|= ISOLATE_CLEAN
;
1794 spin_lock_irq(&zone
->lru_lock
);
1796 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1797 &nr_scanned
, sc
, isolate_mode
, lru
);
1798 if (global_reclaim(sc
))
1799 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1801 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1803 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1804 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1805 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1806 spin_unlock_irq(&zone
->lru_lock
);
1808 while (!list_empty(&l_hold
)) {
1810 page
= lru_to_page(&l_hold
);
1811 list_del(&page
->lru
);
1813 if (unlikely(!page_evictable(page
))) {
1814 putback_lru_page(page
);
1818 if (unlikely(buffer_heads_over_limit
)) {
1819 if (page_has_private(page
) && trylock_page(page
)) {
1820 if (page_has_private(page
))
1821 try_to_release_page(page
, 0);
1826 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1828 nr_rotated
+= hpage_nr_pages(page
);
1830 * Identify referenced, file-backed active pages and
1831 * give them one more trip around the active list. So
1832 * that executable code get better chances to stay in
1833 * memory under moderate memory pressure. Anon pages
1834 * are not likely to be evicted by use-once streaming
1835 * IO, plus JVM can create lots of anon VM_EXEC pages,
1836 * so we ignore them here.
1838 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1839 list_add(&page
->lru
, &l_active
);
1844 ClearPageActive(page
); /* we are de-activating */
1845 list_add(&page
->lru
, &l_inactive
);
1849 * Move pages back to the lru list.
1851 spin_lock_irq(&zone
->lru_lock
);
1853 * Count referenced pages from currently used mappings as rotated,
1854 * even though only some of them are actually re-activated. This
1855 * helps balance scan pressure between file and anonymous pages in
1858 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1860 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1861 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1862 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1863 spin_unlock_irq(&zone
->lru_lock
);
1865 mem_cgroup_uncharge_list(&l_hold
);
1866 free_hot_cold_page_list(&l_hold
, true);
1870 static bool inactive_anon_is_low_global(struct zone
*zone
)
1872 unsigned long active
, inactive
;
1874 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1875 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1877 return inactive
* zone
->inactive_ratio
< active
;
1881 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1882 * @lruvec: LRU vector to check
1884 * Returns true if the zone does not have enough inactive anon pages,
1885 * meaning some active anon pages need to be deactivated.
1887 static bool inactive_anon_is_low(struct lruvec
*lruvec
)
1890 * If we don't have swap space, anonymous page deactivation
1893 if (!total_swap_pages
)
1896 if (!mem_cgroup_disabled())
1897 return mem_cgroup_inactive_anon_is_low(lruvec
);
1899 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1902 static inline bool inactive_anon_is_low(struct lruvec
*lruvec
)
1909 * inactive_file_is_low - check if file pages need to be deactivated
1910 * @lruvec: LRU vector to check
1912 * When the system is doing streaming IO, memory pressure here
1913 * ensures that active file pages get deactivated, until more
1914 * than half of the file pages are on the inactive list.
1916 * Once we get to that situation, protect the system's working
1917 * set from being evicted by disabling active file page aging.
1919 * This uses a different ratio than the anonymous pages, because
1920 * the page cache uses a use-once replacement algorithm.
1922 static bool inactive_file_is_low(struct lruvec
*lruvec
)
1924 unsigned long inactive
;
1925 unsigned long active
;
1927 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1928 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1930 return active
> inactive
;
1933 static bool inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1935 if (is_file_lru(lru
))
1936 return inactive_file_is_low(lruvec
);
1938 return inactive_anon_is_low(lruvec
);
1941 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1942 struct lruvec
*lruvec
, struct scan_control
*sc
)
1944 if (is_active_lru(lru
)) {
1945 if (inactive_list_is_low(lruvec
, lru
))
1946 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1950 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1961 * Determine how aggressively the anon and file LRU lists should be
1962 * scanned. The relative value of each set of LRU lists is determined
1963 * by looking at the fraction of the pages scanned we did rotate back
1964 * onto the active list instead of evict.
1966 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1967 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1969 static void get_scan_count(struct lruvec
*lruvec
, int swappiness
,
1970 struct scan_control
*sc
, unsigned long *nr
,
1971 unsigned long *lru_pages
)
1973 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1975 u64 denominator
= 0; /* gcc */
1976 struct zone
*zone
= lruvec_zone(lruvec
);
1977 unsigned long anon_prio
, file_prio
;
1978 enum scan_balance scan_balance
;
1979 unsigned long anon
, file
;
1980 bool force_scan
= false;
1981 unsigned long ap
, fp
;
1987 * If the zone or memcg is small, nr[l] can be 0. This
1988 * results in no scanning on this priority and a potential
1989 * priority drop. Global direct reclaim can go to the next
1990 * zone and tends to have no problems. Global kswapd is for
1991 * zone balancing and it needs to scan a minimum amount. When
1992 * reclaiming for a memcg, a priority drop can cause high
1993 * latencies, so it's better to scan a minimum amount there as
1996 if (current_is_kswapd()) {
1997 if (!zone_reclaimable(zone
))
1999 if (!mem_cgroup_lruvec_online(lruvec
))
2002 if (!global_reclaim(sc
))
2005 /* If we have no swap space, do not bother scanning anon pages. */
2006 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
2007 scan_balance
= SCAN_FILE
;
2012 * Global reclaim will swap to prevent OOM even with no
2013 * swappiness, but memcg users want to use this knob to
2014 * disable swapping for individual groups completely when
2015 * using the memory controller's swap limit feature would be
2018 if (!global_reclaim(sc
) && !swappiness
) {
2019 scan_balance
= SCAN_FILE
;
2024 * Do not apply any pressure balancing cleverness when the
2025 * system is close to OOM, scan both anon and file equally
2026 * (unless the swappiness setting disagrees with swapping).
2028 if (!sc
->priority
&& swappiness
) {
2029 scan_balance
= SCAN_EQUAL
;
2034 * Prevent the reclaimer from falling into the cache trap: as
2035 * cache pages start out inactive, every cache fault will tip
2036 * the scan balance towards the file LRU. And as the file LRU
2037 * shrinks, so does the window for rotation from references.
2038 * This means we have a runaway feedback loop where a tiny
2039 * thrashing file LRU becomes infinitely more attractive than
2040 * anon pages. Try to detect this based on file LRU size.
2042 if (global_reclaim(sc
)) {
2043 unsigned long zonefile
;
2044 unsigned long zonefree
;
2046 zonefree
= zone_page_state(zone
, NR_FREE_PAGES
);
2047 zonefile
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2048 zone_page_state(zone
, NR_INACTIVE_FILE
);
2050 if (unlikely(zonefile
+ zonefree
<= high_wmark_pages(zone
))) {
2051 scan_balance
= SCAN_ANON
;
2057 * If there is enough inactive page cache, i.e. if the size of the
2058 * inactive list is greater than that of the active list *and* the
2059 * inactive list actually has some pages to scan on this priority, we
2060 * do not reclaim anything from the anonymous working set right now.
2061 * Without the second condition we could end up never scanning an
2062 * lruvec even if it has plenty of old anonymous pages unless the
2063 * system is under heavy pressure.
2065 if (!inactive_file_is_low(lruvec
) &&
2066 get_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2067 scan_balance
= SCAN_FILE
;
2071 scan_balance
= SCAN_FRACT
;
2074 * With swappiness at 100, anonymous and file have the same priority.
2075 * This scanning priority is essentially the inverse of IO cost.
2077 anon_prio
= swappiness
;
2078 file_prio
= 200 - anon_prio
;
2081 * OK, so we have swap space and a fair amount of page cache
2082 * pages. We use the recently rotated / recently scanned
2083 * ratios to determine how valuable each cache is.
2085 * Because workloads change over time (and to avoid overflow)
2086 * we keep these statistics as a floating average, which ends
2087 * up weighing recent references more than old ones.
2089 * anon in [0], file in [1]
2092 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2093 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2094 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2095 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2097 spin_lock_irq(&zone
->lru_lock
);
2098 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2099 reclaim_stat
->recent_scanned
[0] /= 2;
2100 reclaim_stat
->recent_rotated
[0] /= 2;
2103 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2104 reclaim_stat
->recent_scanned
[1] /= 2;
2105 reclaim_stat
->recent_rotated
[1] /= 2;
2109 * The amount of pressure on anon vs file pages is inversely
2110 * proportional to the fraction of recently scanned pages on
2111 * each list that were recently referenced and in active use.
2113 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2114 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2116 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2117 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2118 spin_unlock_irq(&zone
->lru_lock
);
2122 denominator
= ap
+ fp
+ 1;
2124 some_scanned
= false;
2125 /* Only use force_scan on second pass. */
2126 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2128 for_each_evictable_lru(lru
) {
2129 int file
= is_file_lru(lru
);
2133 size
= get_lru_size(lruvec
, lru
);
2134 scan
= size
>> sc
->priority
;
2136 if (!scan
&& pass
&& force_scan
)
2137 scan
= min(size
, SWAP_CLUSTER_MAX
);
2139 switch (scan_balance
) {
2141 /* Scan lists relative to size */
2145 * Scan types proportional to swappiness and
2146 * their relative recent reclaim efficiency.
2148 scan
= div64_u64(scan
* fraction
[file
],
2153 /* Scan one type exclusively */
2154 if ((scan_balance
== SCAN_FILE
) != file
) {
2160 /* Look ma, no brain */
2168 * Skip the second pass and don't force_scan,
2169 * if we found something to scan.
2171 some_scanned
|= !!scan
;
2176 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2177 static void init_tlb_ubc(void)
2180 * This deliberately does not clear the cpumask as it's expensive
2181 * and unnecessary. If there happens to be data in there then the
2182 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2183 * then will be cleared.
2185 current
->tlb_ubc
.flush_required
= false;
2188 static inline void init_tlb_ubc(void)
2191 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2194 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2196 static void shrink_lruvec(struct lruvec
*lruvec
, int swappiness
,
2197 struct scan_control
*sc
, unsigned long *lru_pages
)
2199 unsigned long nr
[NR_LRU_LISTS
];
2200 unsigned long targets
[NR_LRU_LISTS
];
2201 unsigned long nr_to_scan
;
2203 unsigned long nr_reclaimed
= 0;
2204 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2205 struct blk_plug plug
;
2208 get_scan_count(lruvec
, swappiness
, sc
, nr
, lru_pages
);
2210 /* Record the original scan target for proportional adjustments later */
2211 memcpy(targets
, nr
, sizeof(nr
));
2214 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2215 * event that can occur when there is little memory pressure e.g.
2216 * multiple streaming readers/writers. Hence, we do not abort scanning
2217 * when the requested number of pages are reclaimed when scanning at
2218 * DEF_PRIORITY on the assumption that the fact we are direct
2219 * reclaiming implies that kswapd is not keeping up and it is best to
2220 * do a batch of work at once. For memcg reclaim one check is made to
2221 * abort proportional reclaim if either the file or anon lru has already
2222 * dropped to zero at the first pass.
2224 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2225 sc
->priority
== DEF_PRIORITY
);
2229 blk_start_plug(&plug
);
2230 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2231 nr
[LRU_INACTIVE_FILE
]) {
2232 unsigned long nr_anon
, nr_file
, percentage
;
2233 unsigned long nr_scanned
;
2235 for_each_evictable_lru(lru
) {
2237 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2238 nr
[lru
] -= nr_to_scan
;
2240 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2245 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2249 * For kswapd and memcg, reclaim at least the number of pages
2250 * requested. Ensure that the anon and file LRUs are scanned
2251 * proportionally what was requested by get_scan_count(). We
2252 * stop reclaiming one LRU and reduce the amount scanning
2253 * proportional to the original scan target.
2255 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2256 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2259 * It's just vindictive to attack the larger once the smaller
2260 * has gone to zero. And given the way we stop scanning the
2261 * smaller below, this makes sure that we only make one nudge
2262 * towards proportionality once we've got nr_to_reclaim.
2264 if (!nr_file
|| !nr_anon
)
2267 if (nr_file
> nr_anon
) {
2268 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2269 targets
[LRU_ACTIVE_ANON
] + 1;
2271 percentage
= nr_anon
* 100 / scan_target
;
2273 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2274 targets
[LRU_ACTIVE_FILE
] + 1;
2276 percentage
= nr_file
* 100 / scan_target
;
2279 /* Stop scanning the smaller of the LRU */
2281 nr
[lru
+ LRU_ACTIVE
] = 0;
2284 * Recalculate the other LRU scan count based on its original
2285 * scan target and the percentage scanning already complete
2287 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2288 nr_scanned
= targets
[lru
] - nr
[lru
];
2289 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2290 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2293 nr_scanned
= targets
[lru
] - nr
[lru
];
2294 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2295 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2297 scan_adjusted
= true;
2299 blk_finish_plug(&plug
);
2300 sc
->nr_reclaimed
+= nr_reclaimed
;
2303 * Even if we did not try to evict anon pages at all, we want to
2304 * rebalance the anon lru active/inactive ratio.
2306 if (inactive_anon_is_low(lruvec
))
2307 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2308 sc
, LRU_ACTIVE_ANON
);
2310 throttle_vm_writeout(sc
->gfp_mask
);
2313 /* Use reclaim/compaction for costly allocs or under memory pressure */
2314 static bool in_reclaim_compaction(struct scan_control
*sc
)
2316 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2317 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2318 sc
->priority
< DEF_PRIORITY
- 2))
2325 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2326 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2327 * true if more pages should be reclaimed such that when the page allocator
2328 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2329 * It will give up earlier than that if there is difficulty reclaiming pages.
2331 static inline bool should_continue_reclaim(struct zone
*zone
,
2332 unsigned long nr_reclaimed
,
2333 unsigned long nr_scanned
,
2334 struct scan_control
*sc
)
2336 unsigned long pages_for_compaction
;
2337 unsigned long inactive_lru_pages
;
2339 /* If not in reclaim/compaction mode, stop */
2340 if (!in_reclaim_compaction(sc
))
2343 /* Consider stopping depending on scan and reclaim activity */
2344 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2346 * For __GFP_REPEAT allocations, stop reclaiming if the
2347 * full LRU list has been scanned and we are still failing
2348 * to reclaim pages. This full LRU scan is potentially
2349 * expensive but a __GFP_REPEAT caller really wants to succeed
2351 if (!nr_reclaimed
&& !nr_scanned
)
2355 * For non-__GFP_REPEAT allocations which can presumably
2356 * fail without consequence, stop if we failed to reclaim
2357 * any pages from the last SWAP_CLUSTER_MAX number of
2358 * pages that were scanned. This will return to the
2359 * caller faster at the risk reclaim/compaction and
2360 * the resulting allocation attempt fails
2367 * If we have not reclaimed enough pages for compaction and the
2368 * inactive lists are large enough, continue reclaiming
2370 pages_for_compaction
= (2UL << sc
->order
);
2371 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2372 if (get_nr_swap_pages() > 0)
2373 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2374 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2375 inactive_lru_pages
> pages_for_compaction
)
2378 /* If compaction would go ahead or the allocation would succeed, stop */
2379 switch (compaction_suitable(zone
, sc
->order
, 0, 0)) {
2380 case COMPACT_PARTIAL
:
2381 case COMPACT_CONTINUE
:
2388 static bool shrink_zone(struct zone
*zone
, struct scan_control
*sc
,
2391 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2392 unsigned long nr_reclaimed
, nr_scanned
;
2393 bool reclaimable
= false;
2396 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2397 struct mem_cgroup_reclaim_cookie reclaim
= {
2399 .priority
= sc
->priority
,
2401 unsigned long zone_lru_pages
= 0;
2402 struct mem_cgroup
*memcg
;
2404 nr_reclaimed
= sc
->nr_reclaimed
;
2405 nr_scanned
= sc
->nr_scanned
;
2407 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2409 unsigned long lru_pages
;
2410 unsigned long reclaimed
;
2411 unsigned long scanned
;
2412 struct lruvec
*lruvec
;
2415 if (mem_cgroup_low(root
, memcg
)) {
2416 if (!sc
->may_thrash
)
2418 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2421 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2422 swappiness
= mem_cgroup_swappiness(memcg
);
2423 reclaimed
= sc
->nr_reclaimed
;
2424 scanned
= sc
->nr_scanned
;
2426 shrink_lruvec(lruvec
, swappiness
, sc
, &lru_pages
);
2427 zone_lru_pages
+= lru_pages
;
2429 if (memcg
&& is_classzone
)
2430 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
),
2431 memcg
, sc
->nr_scanned
- scanned
,
2434 /* Record the group's reclaim efficiency */
2435 vmpressure(sc
->gfp_mask
, memcg
, false,
2436 sc
->nr_scanned
- scanned
,
2437 sc
->nr_reclaimed
- reclaimed
);
2440 * Direct reclaim and kswapd have to scan all memory
2441 * cgroups to fulfill the overall scan target for the
2444 * Limit reclaim, on the other hand, only cares about
2445 * nr_to_reclaim pages to be reclaimed and it will
2446 * retry with decreasing priority if one round over the
2447 * whole hierarchy is not sufficient.
2449 if (!global_reclaim(sc
) &&
2450 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2451 mem_cgroup_iter_break(root
, memcg
);
2454 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2457 * Shrink the slab caches in the same proportion that
2458 * the eligible LRU pages were scanned.
2460 if (global_reclaim(sc
) && is_classzone
)
2461 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
), NULL
,
2462 sc
->nr_scanned
- nr_scanned
,
2465 if (reclaim_state
) {
2466 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2467 reclaim_state
->reclaimed_slab
= 0;
2470 /* Record the subtree's reclaim efficiency */
2471 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2472 sc
->nr_scanned
- nr_scanned
,
2473 sc
->nr_reclaimed
- nr_reclaimed
);
2475 if (sc
->nr_reclaimed
- nr_reclaimed
)
2478 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2479 sc
->nr_scanned
- nr_scanned
, sc
));
2485 * Returns true if compaction should go ahead for a high-order request, or
2486 * the high-order allocation would succeed without compaction.
2488 static inline bool compaction_ready(struct zone
*zone
, int order
)
2490 unsigned long balance_gap
, watermark
;
2494 * Compaction takes time to run and there are potentially other
2495 * callers using the pages just freed. Continue reclaiming until
2496 * there is a buffer of free pages available to give compaction
2497 * a reasonable chance of completing and allocating the page
2499 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2500 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2501 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << order
);
2502 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0);
2505 * If compaction is deferred, reclaim up to a point where
2506 * compaction will have a chance of success when re-enabled
2508 if (compaction_deferred(zone
, order
))
2509 return watermark_ok
;
2512 * If compaction is not ready to start and allocation is not likely
2513 * to succeed without it, then keep reclaiming.
2515 if (compaction_suitable(zone
, order
, 0, 0) == COMPACT_SKIPPED
)
2518 return watermark_ok
;
2522 * This is the direct reclaim path, for page-allocating processes. We only
2523 * try to reclaim pages from zones which will satisfy the caller's allocation
2526 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2528 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2530 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2531 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2532 * zone defense algorithm.
2534 * If a zone is deemed to be full of pinned pages then just give it a light
2535 * scan then give up on it.
2537 * Returns true if a zone was reclaimable.
2539 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2543 unsigned long nr_soft_reclaimed
;
2544 unsigned long nr_soft_scanned
;
2546 enum zone_type requested_highidx
= gfp_zone(sc
->gfp_mask
);
2547 bool reclaimable
= false;
2550 * If the number of buffer_heads in the machine exceeds the maximum
2551 * allowed level, force direct reclaim to scan the highmem zone as
2552 * highmem pages could be pinning lowmem pages storing buffer_heads
2554 orig_mask
= sc
->gfp_mask
;
2555 if (buffer_heads_over_limit
)
2556 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2558 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2559 requested_highidx
, sc
->nodemask
) {
2560 enum zone_type classzone_idx
;
2562 if (!populated_zone(zone
))
2565 classzone_idx
= requested_highidx
;
2566 while (!populated_zone(zone
->zone_pgdat
->node_zones
+
2571 * Take care memory controller reclaiming has small influence
2574 if (global_reclaim(sc
)) {
2575 if (!cpuset_zone_allowed(zone
,
2576 GFP_KERNEL
| __GFP_HARDWALL
))
2579 if (sc
->priority
!= DEF_PRIORITY
&&
2580 !zone_reclaimable(zone
))
2581 continue; /* Let kswapd poll it */
2584 * If we already have plenty of memory free for
2585 * compaction in this zone, don't free any more.
2586 * Even though compaction is invoked for any
2587 * non-zero order, only frequent costly order
2588 * reclamation is disruptive enough to become a
2589 * noticeable problem, like transparent huge
2592 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2593 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2594 zonelist_zone_idx(z
) <= requested_highidx
&&
2595 compaction_ready(zone
, sc
->order
)) {
2596 sc
->compaction_ready
= true;
2601 * This steals pages from memory cgroups over softlimit
2602 * and returns the number of reclaimed pages and
2603 * scanned pages. This works for global memory pressure
2604 * and balancing, not for a memcg's limit.
2606 nr_soft_scanned
= 0;
2607 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2608 sc
->order
, sc
->gfp_mask
,
2610 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2611 sc
->nr_scanned
+= nr_soft_scanned
;
2612 if (nr_soft_reclaimed
)
2614 /* need some check for avoid more shrink_zone() */
2617 if (shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
))
2620 if (global_reclaim(sc
) &&
2621 !reclaimable
&& zone_reclaimable(zone
))
2626 * Restore to original mask to avoid the impact on the caller if we
2627 * promoted it to __GFP_HIGHMEM.
2629 sc
->gfp_mask
= orig_mask
;
2635 * This is the main entry point to direct page reclaim.
2637 * If a full scan of the inactive list fails to free enough memory then we
2638 * are "out of memory" and something needs to be killed.
2640 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2641 * high - the zone may be full of dirty or under-writeback pages, which this
2642 * caller can't do much about. We kick the writeback threads and take explicit
2643 * naps in the hope that some of these pages can be written. But if the
2644 * allocating task holds filesystem locks which prevent writeout this might not
2645 * work, and the allocation attempt will fail.
2647 * returns: 0, if no pages reclaimed
2648 * else, the number of pages reclaimed
2650 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2651 struct scan_control
*sc
)
2653 int initial_priority
= sc
->priority
;
2654 unsigned long total_scanned
= 0;
2655 unsigned long writeback_threshold
;
2656 bool zones_reclaimable
;
2658 delayacct_freepages_start();
2660 if (global_reclaim(sc
))
2661 count_vm_event(ALLOCSTALL
);
2664 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2667 zones_reclaimable
= shrink_zones(zonelist
, sc
);
2669 total_scanned
+= sc
->nr_scanned
;
2670 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2673 if (sc
->compaction_ready
)
2677 * If we're getting trouble reclaiming, start doing
2678 * writepage even in laptop mode.
2680 if (sc
->priority
< DEF_PRIORITY
- 2)
2681 sc
->may_writepage
= 1;
2684 * Try to write back as many pages as we just scanned. This
2685 * tends to cause slow streaming writers to write data to the
2686 * disk smoothly, at the dirtying rate, which is nice. But
2687 * that's undesirable in laptop mode, where we *want* lumpy
2688 * writeout. So in laptop mode, write out the whole world.
2690 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2691 if (total_scanned
> writeback_threshold
) {
2692 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2693 WB_REASON_TRY_TO_FREE_PAGES
);
2694 sc
->may_writepage
= 1;
2696 } while (--sc
->priority
>= 0);
2698 delayacct_freepages_end();
2700 if (sc
->nr_reclaimed
)
2701 return sc
->nr_reclaimed
;
2703 /* Aborted reclaim to try compaction? don't OOM, then */
2704 if (sc
->compaction_ready
)
2707 /* Untapped cgroup reserves? Don't OOM, retry. */
2708 if (!sc
->may_thrash
) {
2709 sc
->priority
= initial_priority
;
2714 /* Any of the zones still reclaimable? Don't OOM. */
2715 if (zones_reclaimable
)
2721 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2724 unsigned long pfmemalloc_reserve
= 0;
2725 unsigned long free_pages
= 0;
2729 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2730 zone
= &pgdat
->node_zones
[i
];
2731 if (!populated_zone(zone
) ||
2732 zone_reclaimable_pages(zone
) == 0)
2735 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2736 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2739 /* If there are no reserves (unexpected config) then do not throttle */
2740 if (!pfmemalloc_reserve
)
2743 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2745 /* kswapd must be awake if processes are being throttled */
2746 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2747 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2748 (enum zone_type
)ZONE_NORMAL
);
2749 wake_up_interruptible(&pgdat
->kswapd_wait
);
2756 * Throttle direct reclaimers if backing storage is backed by the network
2757 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2758 * depleted. kswapd will continue to make progress and wake the processes
2759 * when the low watermark is reached.
2761 * Returns true if a fatal signal was delivered during throttling. If this
2762 * happens, the page allocator should not consider triggering the OOM killer.
2764 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2765 nodemask_t
*nodemask
)
2769 pg_data_t
*pgdat
= NULL
;
2772 * Kernel threads should not be throttled as they may be indirectly
2773 * responsible for cleaning pages necessary for reclaim to make forward
2774 * progress. kjournald for example may enter direct reclaim while
2775 * committing a transaction where throttling it could forcing other
2776 * processes to block on log_wait_commit().
2778 if (current
->flags
& PF_KTHREAD
)
2782 * If a fatal signal is pending, this process should not throttle.
2783 * It should return quickly so it can exit and free its memory
2785 if (fatal_signal_pending(current
))
2789 * Check if the pfmemalloc reserves are ok by finding the first node
2790 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2791 * GFP_KERNEL will be required for allocating network buffers when
2792 * swapping over the network so ZONE_HIGHMEM is unusable.
2794 * Throttling is based on the first usable node and throttled processes
2795 * wait on a queue until kswapd makes progress and wakes them. There
2796 * is an affinity then between processes waking up and where reclaim
2797 * progress has been made assuming the process wakes on the same node.
2798 * More importantly, processes running on remote nodes will not compete
2799 * for remote pfmemalloc reserves and processes on different nodes
2800 * should make reasonable progress.
2802 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2803 gfp_zone(gfp_mask
), nodemask
) {
2804 if (zone_idx(zone
) > ZONE_NORMAL
)
2807 /* Throttle based on the first usable node */
2808 pgdat
= zone
->zone_pgdat
;
2809 if (pfmemalloc_watermark_ok(pgdat
))
2814 /* If no zone was usable by the allocation flags then do not throttle */
2818 /* Account for the throttling */
2819 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2822 * If the caller cannot enter the filesystem, it's possible that it
2823 * is due to the caller holding an FS lock or performing a journal
2824 * transaction in the case of a filesystem like ext[3|4]. In this case,
2825 * it is not safe to block on pfmemalloc_wait as kswapd could be
2826 * blocked waiting on the same lock. Instead, throttle for up to a
2827 * second before continuing.
2829 if (!(gfp_mask
& __GFP_FS
)) {
2830 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2831 pfmemalloc_watermark_ok(pgdat
), HZ
);
2836 /* Throttle until kswapd wakes the process */
2837 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2838 pfmemalloc_watermark_ok(pgdat
));
2841 if (fatal_signal_pending(current
))
2848 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2849 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2851 unsigned long nr_reclaimed
;
2852 struct scan_control sc
= {
2853 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2854 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2856 .nodemask
= nodemask
,
2857 .priority
= DEF_PRIORITY
,
2858 .may_writepage
= !laptop_mode
,
2864 * Do not enter reclaim if fatal signal was delivered while throttled.
2865 * 1 is returned so that the page allocator does not OOM kill at this
2868 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2871 trace_mm_vmscan_direct_reclaim_begin(order
,
2875 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2877 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2879 return nr_reclaimed
;
2884 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2885 gfp_t gfp_mask
, bool noswap
,
2887 unsigned long *nr_scanned
)
2889 struct scan_control sc
= {
2890 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2891 .target_mem_cgroup
= memcg
,
2892 .may_writepage
= !laptop_mode
,
2894 .may_swap
= !noswap
,
2896 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2897 int swappiness
= mem_cgroup_swappiness(memcg
);
2898 unsigned long lru_pages
;
2900 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2901 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2903 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2908 * NOTE: Although we can get the priority field, using it
2909 * here is not a good idea, since it limits the pages we can scan.
2910 * if we don't reclaim here, the shrink_zone from balance_pgdat
2911 * will pick up pages from other mem cgroup's as well. We hack
2912 * the priority and make it zero.
2914 shrink_lruvec(lruvec
, swappiness
, &sc
, &lru_pages
);
2916 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2918 *nr_scanned
= sc
.nr_scanned
;
2919 return sc
.nr_reclaimed
;
2922 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2923 unsigned long nr_pages
,
2927 struct zonelist
*zonelist
;
2928 unsigned long nr_reclaimed
;
2930 struct scan_control sc
= {
2931 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2932 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2933 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2934 .target_mem_cgroup
= memcg
,
2935 .priority
= DEF_PRIORITY
,
2936 .may_writepage
= !laptop_mode
,
2938 .may_swap
= may_swap
,
2942 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2943 * take care of from where we get pages. So the node where we start the
2944 * scan does not need to be the current node.
2946 nid
= mem_cgroup_select_victim_node(memcg
);
2948 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2950 trace_mm_vmscan_memcg_reclaim_begin(0,
2954 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2956 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2958 return nr_reclaimed
;
2962 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2964 struct mem_cgroup
*memcg
;
2966 if (!total_swap_pages
)
2969 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2971 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2973 if (inactive_anon_is_low(lruvec
))
2974 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2975 sc
, LRU_ACTIVE_ANON
);
2977 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2981 static bool zone_balanced(struct zone
*zone
, int order
,
2982 unsigned long balance_gap
, int classzone_idx
)
2984 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2985 balance_gap
, classzone_idx
))
2988 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&& compaction_suitable(zone
,
2989 order
, 0, classzone_idx
) == COMPACT_SKIPPED
)
2996 * pgdat_balanced() is used when checking if a node is balanced.
2998 * For order-0, all zones must be balanced!
3000 * For high-order allocations only zones that meet watermarks and are in a
3001 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3002 * total of balanced pages must be at least 25% of the zones allowed by
3003 * classzone_idx for the node to be considered balanced. Forcing all zones to
3004 * be balanced for high orders can cause excessive reclaim when there are
3006 * The choice of 25% is due to
3007 * o a 16M DMA zone that is balanced will not balance a zone on any
3008 * reasonable sized machine
3009 * o On all other machines, the top zone must be at least a reasonable
3010 * percentage of the middle zones. For example, on 32-bit x86, highmem
3011 * would need to be at least 256M for it to be balance a whole node.
3012 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3013 * to balance a node on its own. These seemed like reasonable ratios.
3015 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3017 unsigned long managed_pages
= 0;
3018 unsigned long balanced_pages
= 0;
3021 /* Check the watermark levels */
3022 for (i
= 0; i
<= classzone_idx
; i
++) {
3023 struct zone
*zone
= pgdat
->node_zones
+ i
;
3025 if (!populated_zone(zone
))
3028 managed_pages
+= zone
->managed_pages
;
3031 * A special case here:
3033 * balance_pgdat() skips over all_unreclaimable after
3034 * DEF_PRIORITY. Effectively, it considers them balanced so
3035 * they must be considered balanced here as well!
3037 if (!zone_reclaimable(zone
)) {
3038 balanced_pages
+= zone
->managed_pages
;
3042 if (zone_balanced(zone
, order
, 0, i
))
3043 balanced_pages
+= zone
->managed_pages
;
3049 return balanced_pages
>= (managed_pages
>> 2);
3055 * Prepare kswapd for sleeping. This verifies that there are no processes
3056 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3058 * Returns true if kswapd is ready to sleep
3060 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
3063 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3068 * The throttled processes are normally woken up in balance_pgdat() as
3069 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3070 * race between when kswapd checks the watermarks and a process gets
3071 * throttled. There is also a potential race if processes get
3072 * throttled, kswapd wakes, a large process exits thereby balancing the
3073 * zones, which causes kswapd to exit balance_pgdat() before reaching
3074 * the wake up checks. If kswapd is going to sleep, no process should
3075 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3076 * the wake up is premature, processes will wake kswapd and get
3077 * throttled again. The difference from wake ups in balance_pgdat() is
3078 * that here we are under prepare_to_wait().
3080 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3081 wake_up_all(&pgdat
->pfmemalloc_wait
);
3083 return pgdat_balanced(pgdat
, order
, classzone_idx
);
3087 * kswapd shrinks the zone by the number of pages required to reach
3088 * the high watermark.
3090 * Returns true if kswapd scanned at least the requested number of pages to
3091 * reclaim or if the lack of progress was due to pages under writeback.
3092 * This is used to determine if the scanning priority needs to be raised.
3094 static bool kswapd_shrink_zone(struct zone
*zone
,
3096 struct scan_control
*sc
,
3097 unsigned long *nr_attempted
)
3099 int testorder
= sc
->order
;
3100 unsigned long balance_gap
;
3101 bool lowmem_pressure
;
3103 /* Reclaim above the high watermark. */
3104 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
3107 * Kswapd reclaims only single pages with compaction enabled. Trying
3108 * too hard to reclaim until contiguous free pages have become
3109 * available can hurt performance by evicting too much useful data
3110 * from memory. Do not reclaim more than needed for compaction.
3112 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
3113 compaction_suitable(zone
, sc
->order
, 0, classzone_idx
)
3118 * We put equal pressure on every zone, unless one zone has way too
3119 * many pages free already. The "too many pages" is defined as the
3120 * high wmark plus a "gap" where the gap is either the low
3121 * watermark or 1% of the zone, whichever is smaller.
3123 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
3124 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
3127 * If there is no low memory pressure or the zone is balanced then no
3128 * reclaim is necessary
3130 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
3131 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
3132 balance_gap
, classzone_idx
))
3135 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
3137 /* Account for the number of pages attempted to reclaim */
3138 *nr_attempted
+= sc
->nr_to_reclaim
;
3140 clear_bit(ZONE_WRITEBACK
, &zone
->flags
);
3143 * If a zone reaches its high watermark, consider it to be no longer
3144 * congested. It's possible there are dirty pages backed by congested
3145 * BDIs but as pressure is relieved, speculatively avoid congestion
3148 if (zone_reclaimable(zone
) &&
3149 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
3150 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3151 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3154 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3158 * For kswapd, balance_pgdat() will work across all this node's zones until
3159 * they are all at high_wmark_pages(zone).
3161 * Returns the final order kswapd was reclaiming at
3163 * There is special handling here for zones which are full of pinned pages.
3164 * This can happen if the pages are all mlocked, or if they are all used by
3165 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3166 * What we do is to detect the case where all pages in the zone have been
3167 * scanned twice and there has been zero successful reclaim. Mark the zone as
3168 * dead and from now on, only perform a short scan. Basically we're polling
3169 * the zone for when the problem goes away.
3171 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3172 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3173 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3174 * lower zones regardless of the number of free pages in the lower zones. This
3175 * interoperates with the page allocator fallback scheme to ensure that aging
3176 * of pages is balanced across the zones.
3178 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
3182 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
3183 unsigned long nr_soft_reclaimed
;
3184 unsigned long nr_soft_scanned
;
3185 struct scan_control sc
= {
3186 .gfp_mask
= GFP_KERNEL
,
3188 .priority
= DEF_PRIORITY
,
3189 .may_writepage
= !laptop_mode
,
3193 count_vm_event(PAGEOUTRUN
);
3196 unsigned long nr_attempted
= 0;
3197 bool raise_priority
= true;
3198 bool pgdat_needs_compaction
= (order
> 0);
3200 sc
.nr_reclaimed
= 0;
3203 * Scan in the highmem->dma direction for the highest
3204 * zone which needs scanning
3206 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
3207 struct zone
*zone
= pgdat
->node_zones
+ i
;
3209 if (!populated_zone(zone
))
3212 if (sc
.priority
!= DEF_PRIORITY
&&
3213 !zone_reclaimable(zone
))
3217 * Do some background aging of the anon list, to give
3218 * pages a chance to be referenced before reclaiming.
3220 age_active_anon(zone
, &sc
);
3223 * If the number of buffer_heads in the machine
3224 * exceeds the maximum allowed level and this node
3225 * has a highmem zone, force kswapd to reclaim from
3226 * it to relieve lowmem pressure.
3228 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
3233 if (!zone_balanced(zone
, order
, 0, 0)) {
3238 * If balanced, clear the dirty and congested
3241 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3242 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3249 for (i
= 0; i
<= end_zone
; i
++) {
3250 struct zone
*zone
= pgdat
->node_zones
+ i
;
3252 if (!populated_zone(zone
))
3256 * If any zone is currently balanced then kswapd will
3257 * not call compaction as it is expected that the
3258 * necessary pages are already available.
3260 if (pgdat_needs_compaction
&&
3261 zone_watermark_ok(zone
, order
,
3262 low_wmark_pages(zone
),
3264 pgdat_needs_compaction
= false;
3268 * If we're getting trouble reclaiming, start doing writepage
3269 * even in laptop mode.
3271 if (sc
.priority
< DEF_PRIORITY
- 2)
3272 sc
.may_writepage
= 1;
3275 * Now scan the zone in the dma->highmem direction, stopping
3276 * at the last zone which needs scanning.
3278 * We do this because the page allocator works in the opposite
3279 * direction. This prevents the page allocator from allocating
3280 * pages behind kswapd's direction of progress, which would
3281 * cause too much scanning of the lower zones.
3283 for (i
= 0; i
<= end_zone
; i
++) {
3284 struct zone
*zone
= pgdat
->node_zones
+ i
;
3286 if (!populated_zone(zone
))
3289 if (sc
.priority
!= DEF_PRIORITY
&&
3290 !zone_reclaimable(zone
))
3295 nr_soft_scanned
= 0;
3297 * Call soft limit reclaim before calling shrink_zone.
3299 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3302 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3305 * There should be no need to raise the scanning
3306 * priority if enough pages are already being scanned
3307 * that that high watermark would be met at 100%
3310 if (kswapd_shrink_zone(zone
, end_zone
,
3311 &sc
, &nr_attempted
))
3312 raise_priority
= false;
3316 * If the low watermark is met there is no need for processes
3317 * to be throttled on pfmemalloc_wait as they should not be
3318 * able to safely make forward progress. Wake them
3320 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3321 pfmemalloc_watermark_ok(pgdat
))
3322 wake_up_all(&pgdat
->pfmemalloc_wait
);
3325 * Fragmentation may mean that the system cannot be rebalanced
3326 * for high-order allocations in all zones. If twice the
3327 * allocation size has been reclaimed and the zones are still
3328 * not balanced then recheck the watermarks at order-0 to
3329 * prevent kswapd reclaiming excessively. Assume that a
3330 * process requested a high-order can direct reclaim/compact.
3332 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3333 order
= sc
.order
= 0;
3335 /* Check if kswapd should be suspending */
3336 if (try_to_freeze() || kthread_should_stop())
3340 * Compact if necessary and kswapd is reclaiming at least the
3341 * high watermark number of pages as requsted
3343 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3344 compact_pgdat(pgdat
, order
);
3347 * Raise priority if scanning rate is too low or there was no
3348 * progress in reclaiming pages
3350 if (raise_priority
|| !sc
.nr_reclaimed
)
3352 } while (sc
.priority
>= 1 &&
3353 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3357 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3358 * makes a decision on the order we were last reclaiming at. However,
3359 * if another caller entered the allocator slow path while kswapd
3360 * was awake, order will remain at the higher level
3362 *classzone_idx
= end_zone
;
3366 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3371 if (freezing(current
) || kthread_should_stop())
3374 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3376 /* Try to sleep for a short interval */
3377 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3378 remaining
= schedule_timeout(HZ
/10);
3379 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3380 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3384 * After a short sleep, check if it was a premature sleep. If not, then
3385 * go fully to sleep until explicitly woken up.
3387 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3388 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3391 * vmstat counters are not perfectly accurate and the estimated
3392 * value for counters such as NR_FREE_PAGES can deviate from the
3393 * true value by nr_online_cpus * threshold. To avoid the zone
3394 * watermarks being breached while under pressure, we reduce the
3395 * per-cpu vmstat threshold while kswapd is awake and restore
3396 * them before going back to sleep.
3398 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3401 * Compaction records what page blocks it recently failed to
3402 * isolate pages from and skips them in the future scanning.
3403 * When kswapd is going to sleep, it is reasonable to assume
3404 * that pages and compaction may succeed so reset the cache.
3406 reset_isolation_suitable(pgdat
);
3408 if (!kthread_should_stop())
3411 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3414 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3416 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3418 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3422 * The background pageout daemon, started as a kernel thread
3423 * from the init process.
3425 * This basically trickles out pages so that we have _some_
3426 * free memory available even if there is no other activity
3427 * that frees anything up. This is needed for things like routing
3428 * etc, where we otherwise might have all activity going on in
3429 * asynchronous contexts that cannot page things out.
3431 * If there are applications that are active memory-allocators
3432 * (most normal use), this basically shouldn't matter.
3434 static int kswapd(void *p
)
3436 unsigned long order
, new_order
;
3437 unsigned balanced_order
;
3438 int classzone_idx
, new_classzone_idx
;
3439 int balanced_classzone_idx
;
3440 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3441 struct task_struct
*tsk
= current
;
3443 struct reclaim_state reclaim_state
= {
3444 .reclaimed_slab
= 0,
3446 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3448 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3450 if (!cpumask_empty(cpumask
))
3451 set_cpus_allowed_ptr(tsk
, cpumask
);
3452 current
->reclaim_state
= &reclaim_state
;
3455 * Tell the memory management that we're a "memory allocator",
3456 * and that if we need more memory we should get access to it
3457 * regardless (see "__alloc_pages()"). "kswapd" should
3458 * never get caught in the normal page freeing logic.
3460 * (Kswapd normally doesn't need memory anyway, but sometimes
3461 * you need a small amount of memory in order to be able to
3462 * page out something else, and this flag essentially protects
3463 * us from recursively trying to free more memory as we're
3464 * trying to free the first piece of memory in the first place).
3466 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3469 order
= new_order
= 0;
3471 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3472 balanced_classzone_idx
= classzone_idx
;
3477 * If the last balance_pgdat was unsuccessful it's unlikely a
3478 * new request of a similar or harder type will succeed soon
3479 * so consider going to sleep on the basis we reclaimed at
3481 if (balanced_classzone_idx
>= new_classzone_idx
&&
3482 balanced_order
== new_order
) {
3483 new_order
= pgdat
->kswapd_max_order
;
3484 new_classzone_idx
= pgdat
->classzone_idx
;
3485 pgdat
->kswapd_max_order
= 0;
3486 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3489 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3491 * Don't sleep if someone wants a larger 'order'
3492 * allocation or has tigher zone constraints
3495 classzone_idx
= new_classzone_idx
;
3497 kswapd_try_to_sleep(pgdat
, balanced_order
,
3498 balanced_classzone_idx
);
3499 order
= pgdat
->kswapd_max_order
;
3500 classzone_idx
= pgdat
->classzone_idx
;
3502 new_classzone_idx
= classzone_idx
;
3503 pgdat
->kswapd_max_order
= 0;
3504 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3507 ret
= try_to_freeze();
3508 if (kthread_should_stop())
3512 * We can speed up thawing tasks if we don't call balance_pgdat
3513 * after returning from the refrigerator
3516 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3517 balanced_classzone_idx
= classzone_idx
;
3518 balanced_order
= balance_pgdat(pgdat
, order
,
3519 &balanced_classzone_idx
);
3523 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3524 current
->reclaim_state
= NULL
;
3525 lockdep_clear_current_reclaim_state();
3531 * A zone is low on free memory, so wake its kswapd task to service it.
3533 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3537 if (!populated_zone(zone
))
3540 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3542 pgdat
= zone
->zone_pgdat
;
3543 if (pgdat
->kswapd_max_order
< order
) {
3544 pgdat
->kswapd_max_order
= order
;
3545 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3547 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3549 if (zone_balanced(zone
, order
, 0, 0))
3552 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3553 wake_up_interruptible(&pgdat
->kswapd_wait
);
3556 #ifdef CONFIG_HIBERNATION
3558 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3561 * Rather than trying to age LRUs the aim is to preserve the overall
3562 * LRU order by reclaiming preferentially
3563 * inactive > active > active referenced > active mapped
3565 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3567 struct reclaim_state reclaim_state
;
3568 struct scan_control sc
= {
3569 .nr_to_reclaim
= nr_to_reclaim
,
3570 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3571 .priority
= DEF_PRIORITY
,
3575 .hibernation_mode
= 1,
3577 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3578 struct task_struct
*p
= current
;
3579 unsigned long nr_reclaimed
;
3581 p
->flags
|= PF_MEMALLOC
;
3582 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3583 reclaim_state
.reclaimed_slab
= 0;
3584 p
->reclaim_state
= &reclaim_state
;
3586 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3588 p
->reclaim_state
= NULL
;
3589 lockdep_clear_current_reclaim_state();
3590 p
->flags
&= ~PF_MEMALLOC
;
3592 return nr_reclaimed
;
3594 #endif /* CONFIG_HIBERNATION */
3596 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3597 not required for correctness. So if the last cpu in a node goes
3598 away, we get changed to run anywhere: as the first one comes back,
3599 restore their cpu bindings. */
3600 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3605 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3606 for_each_node_state(nid
, N_MEMORY
) {
3607 pg_data_t
*pgdat
= NODE_DATA(nid
);
3608 const struct cpumask
*mask
;
3610 mask
= cpumask_of_node(pgdat
->node_id
);
3612 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3613 /* One of our CPUs online: restore mask */
3614 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3621 * This kswapd start function will be called by init and node-hot-add.
3622 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3624 int kswapd_run(int nid
)
3626 pg_data_t
*pgdat
= NODE_DATA(nid
);
3632 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3633 if (IS_ERR(pgdat
->kswapd
)) {
3634 /* failure at boot is fatal */
3635 BUG_ON(system_state
== SYSTEM_BOOTING
);
3636 pr_err("Failed to start kswapd on node %d\n", nid
);
3637 ret
= PTR_ERR(pgdat
->kswapd
);
3638 pgdat
->kswapd
= NULL
;
3644 * Called by memory hotplug when all memory in a node is offlined. Caller must
3645 * hold mem_hotplug_begin/end().
3647 void kswapd_stop(int nid
)
3649 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3652 kthread_stop(kswapd
);
3653 NODE_DATA(nid
)->kswapd
= NULL
;
3657 static int __init
kswapd_init(void)
3662 for_each_node_state(nid
, N_MEMORY
)
3664 hotcpu_notifier(cpu_callback
, 0);
3668 module_init(kswapd_init
)
3674 * If non-zero call zone_reclaim when the number of free pages falls below
3677 int zone_reclaim_mode __read_mostly
;
3679 #define RECLAIM_OFF 0
3680 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3681 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3682 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3685 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3686 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3689 #define ZONE_RECLAIM_PRIORITY 4
3692 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3695 int sysctl_min_unmapped_ratio
= 1;
3698 * If the number of slab pages in a zone grows beyond this percentage then
3699 * slab reclaim needs to occur.
3701 int sysctl_min_slab_ratio
= 5;
3703 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3705 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3706 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3707 zone_page_state(zone
, NR_ACTIVE_FILE
);
3710 * It's possible for there to be more file mapped pages than
3711 * accounted for by the pages on the file LRU lists because
3712 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3714 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3717 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3718 static unsigned long zone_pagecache_reclaimable(struct zone
*zone
)
3720 unsigned long nr_pagecache_reclaimable
;
3721 unsigned long delta
= 0;
3724 * If RECLAIM_UNMAP is set, then all file pages are considered
3725 * potentially reclaimable. Otherwise, we have to worry about
3726 * pages like swapcache and zone_unmapped_file_pages() provides
3729 if (zone_reclaim_mode
& RECLAIM_UNMAP
)
3730 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3732 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3734 /* If we can't clean pages, remove dirty pages from consideration */
3735 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3736 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3738 /* Watch for any possible underflows due to delta */
3739 if (unlikely(delta
> nr_pagecache_reclaimable
))
3740 delta
= nr_pagecache_reclaimable
;
3742 return nr_pagecache_reclaimable
- delta
;
3746 * Try to free up some pages from this zone through reclaim.
3748 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3750 /* Minimum pages needed in order to stay on node */
3751 const unsigned long nr_pages
= 1 << order
;
3752 struct task_struct
*p
= current
;
3753 struct reclaim_state reclaim_state
;
3754 struct scan_control sc
= {
3755 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3756 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3758 .priority
= ZONE_RECLAIM_PRIORITY
,
3759 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3760 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_UNMAP
),
3766 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3767 * and we also need to be able to write out pages for RECLAIM_WRITE
3768 * and RECLAIM_UNMAP.
3770 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3771 lockdep_set_current_reclaim_state(gfp_mask
);
3772 reclaim_state
.reclaimed_slab
= 0;
3773 p
->reclaim_state
= &reclaim_state
;
3775 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3777 * Free memory by calling shrink zone with increasing
3778 * priorities until we have enough memory freed.
3781 shrink_zone(zone
, &sc
, true);
3782 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3785 p
->reclaim_state
= NULL
;
3786 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3787 lockdep_clear_current_reclaim_state();
3788 return sc
.nr_reclaimed
>= nr_pages
;
3791 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3797 * Zone reclaim reclaims unmapped file backed pages and
3798 * slab pages if we are over the defined limits.
3800 * A small portion of unmapped file backed pages is needed for
3801 * file I/O otherwise pages read by file I/O will be immediately
3802 * thrown out if the zone is overallocated. So we do not reclaim
3803 * if less than a specified percentage of the zone is used by
3804 * unmapped file backed pages.
3806 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3807 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3808 return ZONE_RECLAIM_FULL
;
3810 if (!zone_reclaimable(zone
))
3811 return ZONE_RECLAIM_FULL
;
3814 * Do not scan if the allocation should not be delayed.
3816 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3817 return ZONE_RECLAIM_NOSCAN
;
3820 * Only run zone reclaim on the local zone or on zones that do not
3821 * have associated processors. This will favor the local processor
3822 * over remote processors and spread off node memory allocations
3823 * as wide as possible.
3825 node_id
= zone_to_nid(zone
);
3826 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3827 return ZONE_RECLAIM_NOSCAN
;
3829 if (test_and_set_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
))
3830 return ZONE_RECLAIM_NOSCAN
;
3832 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3833 clear_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
);
3836 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3843 * page_evictable - test whether a page is evictable
3844 * @page: the page to test
3846 * Test whether page is evictable--i.e., should be placed on active/inactive
3847 * lists vs unevictable list.
3849 * Reasons page might not be evictable:
3850 * (1) page's mapping marked unevictable
3851 * (2) page is part of an mlocked VMA
3854 int page_evictable(struct page
*page
)
3856 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3861 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3862 * @pages: array of pages to check
3863 * @nr_pages: number of pages to check
3865 * Checks pages for evictability and moves them to the appropriate lru list.
3867 * This function is only used for SysV IPC SHM_UNLOCK.
3869 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3871 struct lruvec
*lruvec
;
3872 struct zone
*zone
= NULL
;
3877 for (i
= 0; i
< nr_pages
; i
++) {
3878 struct page
*page
= pages
[i
];
3879 struct zone
*pagezone
;
3882 pagezone
= page_zone(page
);
3883 if (pagezone
!= zone
) {
3885 spin_unlock_irq(&zone
->lru_lock
);
3887 spin_lock_irq(&zone
->lru_lock
);
3889 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3891 if (!PageLRU(page
) || !PageUnevictable(page
))
3894 if (page_evictable(page
)) {
3895 enum lru_list lru
= page_lru_base_type(page
);
3897 VM_BUG_ON_PAGE(PageActive(page
), page
);
3898 ClearPageUnevictable(page
);
3899 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3900 add_page_to_lru_list(page
, lruvec
, lru
);
3906 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3907 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3908 spin_unlock_irq(&zone
->lru_lock
);
3911 #endif /* CONFIG_SHMEM */