4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup
*target_mem_cgroup
;
84 /* Scan (total_size >> priority) pages at once */
87 unsigned int may_writepage
:1;
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap
:1;
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap
:1;
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash
:1;
98 unsigned int hibernation_mode
:1;
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready
:1;
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned
;
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed
;
110 #ifdef ARCH_HAS_PREFETCH
111 #define prefetch_prev_lru_page(_page, _base, _field) \
113 if ((_page)->lru.prev != _base) { \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
121 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
124 #ifdef ARCH_HAS_PREFETCHW
125 #define prefetchw_prev_lru_page(_page, _base, _field) \
127 if ((_page)->lru.prev != _base) { \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
135 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 * From 0 .. 100. Higher means more swappy.
141 int vm_swappiness
= 60;
143 * The total number of pages which are beyond the high watermark within all
146 unsigned long vm_total_pages
;
148 static LIST_HEAD(shrinker_list
);
149 static DECLARE_RWSEM(shrinker_rwsem
);
152 static bool global_reclaim(struct scan_control
*sc
)
154 return !sc
->target_mem_cgroup
;
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
170 static bool sane_reclaim(struct scan_control
*sc
)
172 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
176 #ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
183 static bool global_reclaim(struct scan_control
*sc
)
188 static bool sane_reclaim(struct scan_control
*sc
)
194 static unsigned long zone_reclaimable_pages(struct zone
*zone
)
198 nr
= zone_page_state_snapshot(zone
, NR_ACTIVE_FILE
) +
199 zone_page_state_snapshot(zone
, NR_INACTIVE_FILE
) +
200 zone_page_state_snapshot(zone
, NR_ISOLATED_FILE
);
202 if (get_nr_swap_pages() > 0)
203 nr
+= zone_page_state_snapshot(zone
, NR_ACTIVE_ANON
) +
204 zone_page_state_snapshot(zone
, NR_INACTIVE_ANON
) +
205 zone_page_state_snapshot(zone
, NR_ISOLATED_ANON
);
210 bool zone_reclaimable(struct zone
*zone
)
212 return zone_page_state_snapshot(zone
, NR_PAGES_SCANNED
) <
213 zone_reclaimable_pages(zone
) * 6;
216 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec
, lru
);
221 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
225 * Add a shrinker callback to be called from the vm.
227 int register_shrinker(struct shrinker
*shrinker
)
229 size_t size
= sizeof(*shrinker
->nr_deferred
);
231 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
234 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
235 if (!shrinker
->nr_deferred
)
238 down_write(&shrinker_rwsem
);
239 list_add_tail(&shrinker
->list
, &shrinker_list
);
240 up_write(&shrinker_rwsem
);
243 EXPORT_SYMBOL(register_shrinker
);
248 void unregister_shrinker(struct shrinker
*shrinker
)
250 down_write(&shrinker_rwsem
);
251 list_del(&shrinker
->list
);
252 up_write(&shrinker_rwsem
);
253 kfree(shrinker
->nr_deferred
);
255 EXPORT_SYMBOL(unregister_shrinker
);
257 #define SHRINK_BATCH 128
259 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
260 struct shrinker
*shrinker
,
261 unsigned long nr_scanned
,
262 unsigned long nr_eligible
)
264 unsigned long freed
= 0;
265 unsigned long long delta
;
270 int nid
= shrinkctl
->nid
;
271 long batch_size
= shrinker
->batch
? shrinker
->batch
274 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
283 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
286 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
288 do_div(delta
, nr_eligible
+ 1);
290 if (total_scan
< 0) {
291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
292 shrinker
->scan_objects
, total_scan
);
293 total_scan
= freeable
;
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
302 * freeable. This is bad for sustaining a working set in
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
308 if (delta
< freeable
/ 4)
309 total_scan
= min(total_scan
, freeable
/ 2);
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
316 if (total_scan
> freeable
* 2)
317 total_scan
= freeable
* 2;
319 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
320 nr_scanned
, nr_eligible
,
321 freeable
, delta
, total_scan
);
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
334 * than the total number of objects on slab (freeable), we must be
335 * scanning at high prio and therefore should try to reclaim as much as
338 while (total_scan
>= batch_size
||
339 total_scan
>= freeable
) {
341 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
343 shrinkctl
->nr_to_scan
= nr_to_scan
;
344 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
345 if (ret
== SHRINK_STOP
)
349 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
350 total_scan
-= nr_to_scan
;
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
361 new_nr
= atomic_long_add_return(total_scan
,
362 &shrinker
->nr_deferred
[nid
]);
364 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
366 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
371 * shrink_slab - shrink slab caches
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
374 * @memcg: memory cgroup whose slab caches to target
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
378 * Call the shrink functions to age shrinkable caches.
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
385 * objects from the memory cgroup specified. Otherwise all shrinkers
386 * are called, and memcg aware shrinkers are supposed to scan the
389 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
390 * the available objects should be scanned. Page reclaim for example
391 * passes the number of pages scanned and the number of pages on the
392 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
393 * when it encountered mapped pages. The ratio is further biased by
394 * the ->seeks setting of the shrink function, which indicates the
395 * cost to recreate an object relative to that of an LRU page.
397 * Returns the number of reclaimed slab objects.
399 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
400 struct mem_cgroup
*memcg
,
401 unsigned long nr_scanned
,
402 unsigned long nr_eligible
)
404 struct shrinker
*shrinker
;
405 unsigned long freed
= 0;
407 if (memcg
&& !memcg_kmem_online(memcg
))
411 nr_scanned
= SWAP_CLUSTER_MAX
;
413 if (!down_read_trylock(&shrinker_rwsem
)) {
415 * If we would return 0, our callers would understand that we
416 * have nothing else to shrink and give up trying. By returning
417 * 1 we keep it going and assume we'll be able to shrink next
424 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
425 struct shrink_control sc
= {
426 .gfp_mask
= gfp_mask
,
431 if (memcg
&& !(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
434 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
437 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
440 up_read(&shrinker_rwsem
);
446 void drop_slab_node(int nid
)
451 struct mem_cgroup
*memcg
= NULL
;
455 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
457 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
458 } while (freed
> 10);
465 for_each_online_node(nid
)
469 static inline int is_page_cache_freeable(struct page
*page
)
472 * A freeable page cache page is referenced only by the caller
473 * that isolated the page, the page cache radix tree and
474 * optional buffer heads at page->private.
476 return page_count(page
) - page_has_private(page
) == 2;
479 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
481 if (current
->flags
& PF_SWAPWRITE
)
483 if (!inode_write_congested(inode
))
485 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
491 * We detected a synchronous write error writing a page out. Probably
492 * -ENOSPC. We need to propagate that into the address_space for a subsequent
493 * fsync(), msync() or close().
495 * The tricky part is that after writepage we cannot touch the mapping: nothing
496 * prevents it from being freed up. But we have a ref on the page and once
497 * that page is locked, the mapping is pinned.
499 * We're allowed to run sleeping lock_page() here because we know the caller has
502 static void handle_write_error(struct address_space
*mapping
,
503 struct page
*page
, int error
)
506 if (page_mapping(page
) == mapping
)
507 mapping_set_error(mapping
, error
);
511 /* possible outcome of pageout() */
513 /* failed to write page out, page is locked */
515 /* move page to the active list, page is locked */
517 /* page has been sent to the disk successfully, page is unlocked */
519 /* page is clean and locked */
524 * pageout is called by shrink_page_list() for each dirty page.
525 * Calls ->writepage().
527 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
528 struct scan_control
*sc
)
531 * If the page is dirty, only perform writeback if that write
532 * will be non-blocking. To prevent this allocation from being
533 * stalled by pagecache activity. But note that there may be
534 * stalls if we need to run get_block(). We could test
535 * PagePrivate for that.
537 * If this process is currently in __generic_file_write_iter() against
538 * this page's queue, we can perform writeback even if that
541 * If the page is swapcache, write it back even if that would
542 * block, for some throttling. This happens by accident, because
543 * swap_backing_dev_info is bust: it doesn't reflect the
544 * congestion state of the swapdevs. Easy to fix, if needed.
546 if (!is_page_cache_freeable(page
))
550 * Some data journaling orphaned pages can have
551 * page->mapping == NULL while being dirty with clean buffers.
553 if (page_has_private(page
)) {
554 if (try_to_free_buffers(page
)) {
555 ClearPageDirty(page
);
556 pr_info("%s: orphaned page\n", __func__
);
562 if (mapping
->a_ops
->writepage
== NULL
)
563 return PAGE_ACTIVATE
;
564 if (!may_write_to_inode(mapping
->host
, sc
))
567 if (clear_page_dirty_for_io(page
)) {
569 struct writeback_control wbc
= {
570 .sync_mode
= WB_SYNC_NONE
,
571 .nr_to_write
= SWAP_CLUSTER_MAX
,
573 .range_end
= LLONG_MAX
,
577 SetPageReclaim(page
);
578 res
= mapping
->a_ops
->writepage(page
, &wbc
);
580 handle_write_error(mapping
, page
, res
);
581 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
582 ClearPageReclaim(page
);
583 return PAGE_ACTIVATE
;
586 if (!PageWriteback(page
)) {
587 /* synchronous write or broken a_ops? */
588 ClearPageReclaim(page
);
590 trace_mm_vmscan_writepage(page
);
591 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
599 * Same as remove_mapping, but if the page is removed from the mapping, it
600 * gets returned with a refcount of 0.
602 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
607 BUG_ON(!PageLocked(page
));
608 BUG_ON(mapping
!= page_mapping(page
));
610 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
612 * The non racy check for a busy page.
614 * Must be careful with the order of the tests. When someone has
615 * a ref to the page, it may be possible that they dirty it then
616 * drop the reference. So if PageDirty is tested before page_count
617 * here, then the following race may occur:
619 * get_user_pages(&page);
620 * [user mapping goes away]
622 * !PageDirty(page) [good]
623 * SetPageDirty(page);
625 * !page_count(page) [good, discard it]
627 * [oops, our write_to data is lost]
629 * Reversing the order of the tests ensures such a situation cannot
630 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
631 * load is not satisfied before that of page->_count.
633 * Note that if SetPageDirty is always performed via set_page_dirty,
634 * and thus under tree_lock, then this ordering is not required.
636 if (!page_freeze_refs(page
, 2))
638 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
639 if (unlikely(PageDirty(page
))) {
640 page_unfreeze_refs(page
, 2);
644 if (PageSwapCache(page
)) {
645 swp_entry_t swap
= { .val
= page_private(page
) };
646 mem_cgroup_swapout(page
, swap
);
647 __delete_from_swap_cache(page
);
648 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
649 swapcache_free(swap
);
651 void (*freepage
)(struct page
*);
654 freepage
= mapping
->a_ops
->freepage
;
656 * Remember a shadow entry for reclaimed file cache in
657 * order to detect refaults, thus thrashing, later on.
659 * But don't store shadows in an address space that is
660 * already exiting. This is not just an optizimation,
661 * inode reclaim needs to empty out the radix tree or
662 * the nodes are lost. Don't plant shadows behind its
665 * We also don't store shadows for DAX mappings because the
666 * only page cache pages found in these are zero pages
667 * covering holes, and because we don't want to mix DAX
668 * exceptional entries and shadow exceptional entries in the
671 if (reclaimed
&& page_is_file_cache(page
) &&
672 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
673 shadow
= workingset_eviction(mapping
, page
);
674 __delete_from_page_cache(page
, shadow
);
675 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
677 if (freepage
!= NULL
)
684 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
689 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
690 * someone else has a ref on the page, abort and return 0. If it was
691 * successfully detached, return 1. Assumes the caller has a single ref on
694 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
696 if (__remove_mapping(mapping
, page
, false)) {
698 * Unfreezing the refcount with 1 rather than 2 effectively
699 * drops the pagecache ref for us without requiring another
702 page_unfreeze_refs(page
, 1);
709 * putback_lru_page - put previously isolated page onto appropriate LRU list
710 * @page: page to be put back to appropriate lru list
712 * Add previously isolated @page to appropriate LRU list.
713 * Page may still be unevictable for other reasons.
715 * lru_lock must not be held, interrupts must be enabled.
717 void putback_lru_page(struct page
*page
)
720 int was_unevictable
= PageUnevictable(page
);
722 VM_BUG_ON_PAGE(PageLRU(page
), page
);
725 ClearPageUnevictable(page
);
727 if (page_evictable(page
)) {
729 * For evictable pages, we can use the cache.
730 * In event of a race, worst case is we end up with an
731 * unevictable page on [in]active list.
732 * We know how to handle that.
734 is_unevictable
= false;
738 * Put unevictable pages directly on zone's unevictable
741 is_unevictable
= true;
742 add_page_to_unevictable_list(page
);
744 * When racing with an mlock or AS_UNEVICTABLE clearing
745 * (page is unlocked) make sure that if the other thread
746 * does not observe our setting of PG_lru and fails
747 * isolation/check_move_unevictable_pages,
748 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
749 * the page back to the evictable list.
751 * The other side is TestClearPageMlocked() or shmem_lock().
757 * page's status can change while we move it among lru. If an evictable
758 * page is on unevictable list, it never be freed. To avoid that,
759 * check after we added it to the list, again.
761 if (is_unevictable
&& page_evictable(page
)) {
762 if (!isolate_lru_page(page
)) {
766 /* This means someone else dropped this page from LRU
767 * So, it will be freed or putback to LRU again. There is
768 * nothing to do here.
772 if (was_unevictable
&& !is_unevictable
)
773 count_vm_event(UNEVICTABLE_PGRESCUED
);
774 else if (!was_unevictable
&& is_unevictable
)
775 count_vm_event(UNEVICTABLE_PGCULLED
);
777 put_page(page
); /* drop ref from isolate */
780 enum page_references
{
782 PAGEREF_RECLAIM_CLEAN
,
787 static enum page_references
page_check_references(struct page
*page
,
788 struct scan_control
*sc
)
790 int referenced_ptes
, referenced_page
;
791 unsigned long vm_flags
;
793 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
795 referenced_page
= TestClearPageReferenced(page
);
798 * Mlock lost the isolation race with us. Let try_to_unmap()
799 * move the page to the unevictable list.
801 if (vm_flags
& VM_LOCKED
)
802 return PAGEREF_RECLAIM
;
804 if (referenced_ptes
) {
805 if (PageSwapBacked(page
))
806 return PAGEREF_ACTIVATE
;
808 * All mapped pages start out with page table
809 * references from the instantiating fault, so we need
810 * to look twice if a mapped file page is used more
813 * Mark it and spare it for another trip around the
814 * inactive list. Another page table reference will
815 * lead to its activation.
817 * Note: the mark is set for activated pages as well
818 * so that recently deactivated but used pages are
821 SetPageReferenced(page
);
823 if (referenced_page
|| referenced_ptes
> 1)
824 return PAGEREF_ACTIVATE
;
827 * Activate file-backed executable pages after first usage.
829 if (vm_flags
& VM_EXEC
)
830 return PAGEREF_ACTIVATE
;
835 /* Reclaim if clean, defer dirty pages to writeback */
836 if (referenced_page
&& !PageSwapBacked(page
))
837 return PAGEREF_RECLAIM_CLEAN
;
839 return PAGEREF_RECLAIM
;
842 /* Check if a page is dirty or under writeback */
843 static void page_check_dirty_writeback(struct page
*page
,
844 bool *dirty
, bool *writeback
)
846 struct address_space
*mapping
;
849 * Anonymous pages are not handled by flushers and must be written
850 * from reclaim context. Do not stall reclaim based on them
852 if (!page_is_file_cache(page
)) {
858 /* By default assume that the page flags are accurate */
859 *dirty
= PageDirty(page
);
860 *writeback
= PageWriteback(page
);
862 /* Verify dirty/writeback state if the filesystem supports it */
863 if (!page_has_private(page
))
866 mapping
= page_mapping(page
);
867 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
868 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
872 * shrink_page_list() returns the number of reclaimed pages
874 static unsigned long shrink_page_list(struct list_head
*page_list
,
876 struct scan_control
*sc
,
877 enum ttu_flags ttu_flags
,
878 unsigned long *ret_nr_dirty
,
879 unsigned long *ret_nr_unqueued_dirty
,
880 unsigned long *ret_nr_congested
,
881 unsigned long *ret_nr_writeback
,
882 unsigned long *ret_nr_immediate
,
885 LIST_HEAD(ret_pages
);
886 LIST_HEAD(free_pages
);
888 unsigned long nr_unqueued_dirty
= 0;
889 unsigned long nr_dirty
= 0;
890 unsigned long nr_congested
= 0;
891 unsigned long nr_reclaimed
= 0;
892 unsigned long nr_writeback
= 0;
893 unsigned long nr_immediate
= 0;
897 while (!list_empty(page_list
)) {
898 struct address_space
*mapping
;
901 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
902 bool dirty
, writeback
;
903 bool lazyfree
= false;
904 int ret
= SWAP_SUCCESS
;
908 page
= lru_to_page(page_list
);
909 list_del(&page
->lru
);
911 if (!trylock_page(page
))
914 VM_BUG_ON_PAGE(PageActive(page
), page
);
915 VM_BUG_ON_PAGE(page_zone(page
) != zone
, page
);
919 if (unlikely(!page_evictable(page
)))
922 if (!sc
->may_unmap
&& page_mapped(page
))
925 /* Double the slab pressure for mapped and swapcache pages */
926 if (page_mapped(page
) || PageSwapCache(page
))
929 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
930 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
933 * The number of dirty pages determines if a zone is marked
934 * reclaim_congested which affects wait_iff_congested. kswapd
935 * will stall and start writing pages if the tail of the LRU
936 * is all dirty unqueued pages.
938 page_check_dirty_writeback(page
, &dirty
, &writeback
);
939 if (dirty
|| writeback
)
942 if (dirty
&& !writeback
)
946 * Treat this page as congested if the underlying BDI is or if
947 * pages are cycling through the LRU so quickly that the
948 * pages marked for immediate reclaim are making it to the
949 * end of the LRU a second time.
951 mapping
= page_mapping(page
);
952 if (((dirty
|| writeback
) && mapping
&&
953 inode_write_congested(mapping
->host
)) ||
954 (writeback
&& PageReclaim(page
)))
958 * If a page at the tail of the LRU is under writeback, there
959 * are three cases to consider.
961 * 1) If reclaim is encountering an excessive number of pages
962 * under writeback and this page is both under writeback and
963 * PageReclaim then it indicates that pages are being queued
964 * for IO but are being recycled through the LRU before the
965 * IO can complete. Waiting on the page itself risks an
966 * indefinite stall if it is impossible to writeback the
967 * page due to IO error or disconnected storage so instead
968 * note that the LRU is being scanned too quickly and the
969 * caller can stall after page list has been processed.
971 * 2) Global or new memcg reclaim encounters a page that is
972 * not marked for immediate reclaim, or the caller does not
973 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
974 * not to fs). In this case mark the page for immediate
975 * reclaim and continue scanning.
977 * Require may_enter_fs because we would wait on fs, which
978 * may not have submitted IO yet. And the loop driver might
979 * enter reclaim, and deadlock if it waits on a page for
980 * which it is needed to do the write (loop masks off
981 * __GFP_IO|__GFP_FS for this reason); but more thought
982 * would probably show more reasons.
984 * 3) Legacy memcg encounters a page that is already marked
985 * PageReclaim. memcg does not have any dirty pages
986 * throttling so we could easily OOM just because too many
987 * pages are in writeback and there is nothing else to
988 * reclaim. Wait for the writeback to complete.
990 if (PageWriteback(page
)) {
992 if (current_is_kswapd() &&
994 test_bit(ZONE_WRITEBACK
, &zone
->flags
)) {
999 } else if (sane_reclaim(sc
) ||
1000 !PageReclaim(page
) || !may_enter_fs
) {
1002 * This is slightly racy - end_page_writeback()
1003 * might have just cleared PageReclaim, then
1004 * setting PageReclaim here end up interpreted
1005 * as PageReadahead - but that does not matter
1006 * enough to care. What we do want is for this
1007 * page to have PageReclaim set next time memcg
1008 * reclaim reaches the tests above, so it will
1009 * then wait_on_page_writeback() to avoid OOM;
1010 * and it's also appropriate in global reclaim.
1012 SetPageReclaim(page
);
1019 wait_on_page_writeback(page
);
1020 /* then go back and try same page again */
1021 list_add_tail(&page
->lru
, page_list
);
1027 references
= page_check_references(page
, sc
);
1029 switch (references
) {
1030 case PAGEREF_ACTIVATE
:
1031 goto activate_locked
;
1034 case PAGEREF_RECLAIM
:
1035 case PAGEREF_RECLAIM_CLEAN
:
1036 ; /* try to reclaim the page below */
1040 * Anonymous process memory has backing store?
1041 * Try to allocate it some swap space here.
1043 if (PageAnon(page
) && !PageSwapCache(page
)) {
1044 if (!(sc
->gfp_mask
& __GFP_IO
))
1046 if (!add_to_swap(page
, page_list
))
1047 goto activate_locked
;
1051 /* Adding to swap updated mapping */
1052 mapping
= page_mapping(page
);
1056 * The page is mapped into the page tables of one or more
1057 * processes. Try to unmap it here.
1059 if (page_mapped(page
) && mapping
) {
1060 switch (ret
= try_to_unmap(page
, lazyfree
?
1061 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1062 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1064 goto activate_locked
;
1072 ; /* try to free the page below */
1076 if (PageDirty(page
)) {
1078 * Only kswapd can writeback filesystem pages to
1079 * avoid risk of stack overflow but only writeback
1080 * if many dirty pages have been encountered.
1082 if (page_is_file_cache(page
) &&
1083 (!current_is_kswapd() ||
1084 !test_bit(ZONE_DIRTY
, &zone
->flags
))) {
1086 * Immediately reclaim when written back.
1087 * Similar in principal to deactivate_page()
1088 * except we already have the page isolated
1089 * and know it's dirty
1091 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1092 SetPageReclaim(page
);
1097 if (references
== PAGEREF_RECLAIM_CLEAN
)
1101 if (!sc
->may_writepage
)
1105 * Page is dirty. Flush the TLB if a writable entry
1106 * potentially exists to avoid CPU writes after IO
1107 * starts and then write it out here.
1109 try_to_unmap_flush_dirty();
1110 switch (pageout(page
, mapping
, sc
)) {
1114 goto activate_locked
;
1116 if (PageWriteback(page
))
1118 if (PageDirty(page
))
1122 * A synchronous write - probably a ramdisk. Go
1123 * ahead and try to reclaim the page.
1125 if (!trylock_page(page
))
1127 if (PageDirty(page
) || PageWriteback(page
))
1129 mapping
= page_mapping(page
);
1131 ; /* try to free the page below */
1136 * If the page has buffers, try to free the buffer mappings
1137 * associated with this page. If we succeed we try to free
1140 * We do this even if the page is PageDirty().
1141 * try_to_release_page() does not perform I/O, but it is
1142 * possible for a page to have PageDirty set, but it is actually
1143 * clean (all its buffers are clean). This happens if the
1144 * buffers were written out directly, with submit_bh(). ext3
1145 * will do this, as well as the blockdev mapping.
1146 * try_to_release_page() will discover that cleanness and will
1147 * drop the buffers and mark the page clean - it can be freed.
1149 * Rarely, pages can have buffers and no ->mapping. These are
1150 * the pages which were not successfully invalidated in
1151 * truncate_complete_page(). We try to drop those buffers here
1152 * and if that worked, and the page is no longer mapped into
1153 * process address space (page_count == 1) it can be freed.
1154 * Otherwise, leave the page on the LRU so it is swappable.
1156 if (page_has_private(page
)) {
1157 if (!try_to_release_page(page
, sc
->gfp_mask
))
1158 goto activate_locked
;
1159 if (!mapping
&& page_count(page
) == 1) {
1161 if (put_page_testzero(page
))
1165 * rare race with speculative reference.
1166 * the speculative reference will free
1167 * this page shortly, so we may
1168 * increment nr_reclaimed here (and
1169 * leave it off the LRU).
1178 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1182 * At this point, we have no other references and there is
1183 * no way to pick any more up (removed from LRU, removed
1184 * from pagecache). Can use non-atomic bitops now (and
1185 * we obviously don't have to worry about waking up a process
1186 * waiting on the page lock, because there are no references.
1188 __ClearPageLocked(page
);
1190 if (ret
== SWAP_LZFREE
)
1191 count_vm_event(PGLAZYFREED
);
1196 * Is there need to periodically free_page_list? It would
1197 * appear not as the counts should be low
1199 list_add(&page
->lru
, &free_pages
);
1203 if (PageSwapCache(page
))
1204 try_to_free_swap(page
);
1206 list_add(&page
->lru
, &ret_pages
);
1210 /* Not a candidate for swapping, so reclaim swap space. */
1211 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1212 try_to_free_swap(page
);
1213 VM_BUG_ON_PAGE(PageActive(page
), page
);
1214 SetPageActive(page
);
1219 list_add(&page
->lru
, &ret_pages
);
1220 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1223 mem_cgroup_uncharge_list(&free_pages
);
1224 try_to_unmap_flush();
1225 free_hot_cold_page_list(&free_pages
, true);
1227 list_splice(&ret_pages
, page_list
);
1228 count_vm_events(PGACTIVATE
, pgactivate
);
1230 *ret_nr_dirty
+= nr_dirty
;
1231 *ret_nr_congested
+= nr_congested
;
1232 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1233 *ret_nr_writeback
+= nr_writeback
;
1234 *ret_nr_immediate
+= nr_immediate
;
1235 return nr_reclaimed
;
1238 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1239 struct list_head
*page_list
)
1241 struct scan_control sc
= {
1242 .gfp_mask
= GFP_KERNEL
,
1243 .priority
= DEF_PRIORITY
,
1246 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1247 struct page
*page
, *next
;
1248 LIST_HEAD(clean_pages
);
1250 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1251 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1252 !isolated_balloon_page(page
)) {
1253 ClearPageActive(page
);
1254 list_move(&page
->lru
, &clean_pages
);
1258 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1259 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1260 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1261 list_splice(&clean_pages
, page_list
);
1262 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1267 * Attempt to remove the specified page from its LRU. Only take this page
1268 * if it is of the appropriate PageActive status. Pages which are being
1269 * freed elsewhere are also ignored.
1271 * page: page to consider
1272 * mode: one of the LRU isolation modes defined above
1274 * returns 0 on success, -ve errno on failure.
1276 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1280 /* Only take pages on the LRU. */
1284 /* Compaction should not handle unevictable pages but CMA can do so */
1285 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1291 * To minimise LRU disruption, the caller can indicate that it only
1292 * wants to isolate pages it will be able to operate on without
1293 * blocking - clean pages for the most part.
1295 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1296 * is used by reclaim when it is cannot write to backing storage
1298 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1299 * that it is possible to migrate without blocking
1301 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1302 /* All the caller can do on PageWriteback is block */
1303 if (PageWriteback(page
))
1306 if (PageDirty(page
)) {
1307 struct address_space
*mapping
;
1309 /* ISOLATE_CLEAN means only clean pages */
1310 if (mode
& ISOLATE_CLEAN
)
1314 * Only pages without mappings or that have a
1315 * ->migratepage callback are possible to migrate
1318 mapping
= page_mapping(page
);
1319 if (mapping
&& !mapping
->a_ops
->migratepage
)
1324 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1327 if (likely(get_page_unless_zero(page
))) {
1329 * Be careful not to clear PageLRU until after we're
1330 * sure the page is not being freed elsewhere -- the
1331 * page release code relies on it.
1341 * zone->lru_lock is heavily contended. Some of the functions that
1342 * shrink the lists perform better by taking out a batch of pages
1343 * and working on them outside the LRU lock.
1345 * For pagecache intensive workloads, this function is the hottest
1346 * spot in the kernel (apart from copy_*_user functions).
1348 * Appropriate locks must be held before calling this function.
1350 * @nr_to_scan: The number of pages to look through on the list.
1351 * @lruvec: The LRU vector to pull pages from.
1352 * @dst: The temp list to put pages on to.
1353 * @nr_scanned: The number of pages that were scanned.
1354 * @sc: The scan_control struct for this reclaim session
1355 * @mode: One of the LRU isolation modes
1356 * @lru: LRU list id for isolating
1358 * returns how many pages were moved onto *@dst.
1360 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1361 struct lruvec
*lruvec
, struct list_head
*dst
,
1362 unsigned long *nr_scanned
, struct scan_control
*sc
,
1363 isolate_mode_t mode
, enum lru_list lru
)
1365 struct list_head
*src
= &lruvec
->lists
[lru
];
1366 unsigned long nr_taken
= 0;
1369 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1370 !list_empty(src
); scan
++) {
1374 page
= lru_to_page(src
);
1375 prefetchw_prev_lru_page(page
, src
, flags
);
1377 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1379 switch (__isolate_lru_page(page
, mode
)) {
1381 nr_pages
= hpage_nr_pages(page
);
1382 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1383 list_move(&page
->lru
, dst
);
1384 nr_taken
+= nr_pages
;
1388 /* else it is being freed elsewhere */
1389 list_move(&page
->lru
, src
);
1398 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1399 nr_taken
, mode
, is_file_lru(lru
));
1404 * isolate_lru_page - tries to isolate a page from its LRU list
1405 * @page: page to isolate from its LRU list
1407 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1408 * vmstat statistic corresponding to whatever LRU list the page was on.
1410 * Returns 0 if the page was removed from an LRU list.
1411 * Returns -EBUSY if the page was not on an LRU list.
1413 * The returned page will have PageLRU() cleared. If it was found on
1414 * the active list, it will have PageActive set. If it was found on
1415 * the unevictable list, it will have the PageUnevictable bit set. That flag
1416 * may need to be cleared by the caller before letting the page go.
1418 * The vmstat statistic corresponding to the list on which the page was
1419 * found will be decremented.
1422 * (1) Must be called with an elevated refcount on the page. This is a
1423 * fundamentnal difference from isolate_lru_pages (which is called
1424 * without a stable reference).
1425 * (2) the lru_lock must not be held.
1426 * (3) interrupts must be enabled.
1428 int isolate_lru_page(struct page
*page
)
1432 VM_BUG_ON_PAGE(!page_count(page
), page
);
1433 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1435 if (PageLRU(page
)) {
1436 struct zone
*zone
= page_zone(page
);
1437 struct lruvec
*lruvec
;
1439 spin_lock_irq(&zone
->lru_lock
);
1440 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1441 if (PageLRU(page
)) {
1442 int lru
= page_lru(page
);
1445 del_page_from_lru_list(page
, lruvec
, lru
);
1448 spin_unlock_irq(&zone
->lru_lock
);
1454 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1455 * then get resheduled. When there are massive number of tasks doing page
1456 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1457 * the LRU list will go small and be scanned faster than necessary, leading to
1458 * unnecessary swapping, thrashing and OOM.
1460 static int too_many_isolated(struct zone
*zone
, int file
,
1461 struct scan_control
*sc
)
1463 unsigned long inactive
, isolated
;
1465 if (current_is_kswapd())
1468 if (!sane_reclaim(sc
))
1472 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1473 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1475 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1476 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1480 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1481 * won't get blocked by normal direct-reclaimers, forming a circular
1484 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1487 return isolated
> inactive
;
1490 static noinline_for_stack
void
1491 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1493 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1494 struct zone
*zone
= lruvec_zone(lruvec
);
1495 LIST_HEAD(pages_to_free
);
1498 * Put back any unfreeable pages.
1500 while (!list_empty(page_list
)) {
1501 struct page
*page
= lru_to_page(page_list
);
1504 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1505 list_del(&page
->lru
);
1506 if (unlikely(!page_evictable(page
))) {
1507 spin_unlock_irq(&zone
->lru_lock
);
1508 putback_lru_page(page
);
1509 spin_lock_irq(&zone
->lru_lock
);
1513 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1516 lru
= page_lru(page
);
1517 add_page_to_lru_list(page
, lruvec
, lru
);
1519 if (is_active_lru(lru
)) {
1520 int file
= is_file_lru(lru
);
1521 int numpages
= hpage_nr_pages(page
);
1522 reclaim_stat
->recent_rotated
[file
] += numpages
;
1524 if (put_page_testzero(page
)) {
1525 __ClearPageLRU(page
);
1526 __ClearPageActive(page
);
1527 del_page_from_lru_list(page
, lruvec
, lru
);
1529 if (unlikely(PageCompound(page
))) {
1530 spin_unlock_irq(&zone
->lru_lock
);
1531 mem_cgroup_uncharge(page
);
1532 (*get_compound_page_dtor(page
))(page
);
1533 spin_lock_irq(&zone
->lru_lock
);
1535 list_add(&page
->lru
, &pages_to_free
);
1540 * To save our caller's stack, now use input list for pages to free.
1542 list_splice(&pages_to_free
, page_list
);
1546 * If a kernel thread (such as nfsd for loop-back mounts) services
1547 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1548 * In that case we should only throttle if the backing device it is
1549 * writing to is congested. In other cases it is safe to throttle.
1551 static int current_may_throttle(void)
1553 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1554 current
->backing_dev_info
== NULL
||
1555 bdi_write_congested(current
->backing_dev_info
);
1559 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1560 * of reclaimed pages
1562 static noinline_for_stack
unsigned long
1563 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1564 struct scan_control
*sc
, enum lru_list lru
)
1566 LIST_HEAD(page_list
);
1567 unsigned long nr_scanned
;
1568 unsigned long nr_reclaimed
= 0;
1569 unsigned long nr_taken
;
1570 unsigned long nr_dirty
= 0;
1571 unsigned long nr_congested
= 0;
1572 unsigned long nr_unqueued_dirty
= 0;
1573 unsigned long nr_writeback
= 0;
1574 unsigned long nr_immediate
= 0;
1575 isolate_mode_t isolate_mode
= 0;
1576 int file
= is_file_lru(lru
);
1577 struct zone
*zone
= lruvec_zone(lruvec
);
1578 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1580 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1581 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1583 /* We are about to die and free our memory. Return now. */
1584 if (fatal_signal_pending(current
))
1585 return SWAP_CLUSTER_MAX
;
1591 isolate_mode
|= ISOLATE_UNMAPPED
;
1592 if (!sc
->may_writepage
)
1593 isolate_mode
|= ISOLATE_CLEAN
;
1595 spin_lock_irq(&zone
->lru_lock
);
1597 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1598 &nr_scanned
, sc
, isolate_mode
, lru
);
1600 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1601 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1603 if (global_reclaim(sc
)) {
1604 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1605 if (current_is_kswapd())
1606 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1608 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1610 spin_unlock_irq(&zone
->lru_lock
);
1615 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1616 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1617 &nr_writeback
, &nr_immediate
,
1620 spin_lock_irq(&zone
->lru_lock
);
1622 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1624 if (global_reclaim(sc
)) {
1625 if (current_is_kswapd())
1626 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1629 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1633 putback_inactive_pages(lruvec
, &page_list
);
1635 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1637 spin_unlock_irq(&zone
->lru_lock
);
1639 mem_cgroup_uncharge_list(&page_list
);
1640 free_hot_cold_page_list(&page_list
, true);
1643 * If reclaim is isolating dirty pages under writeback, it implies
1644 * that the long-lived page allocation rate is exceeding the page
1645 * laundering rate. Either the global limits are not being effective
1646 * at throttling processes due to the page distribution throughout
1647 * zones or there is heavy usage of a slow backing device. The
1648 * only option is to throttle from reclaim context which is not ideal
1649 * as there is no guarantee the dirtying process is throttled in the
1650 * same way balance_dirty_pages() manages.
1652 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1653 * of pages under pages flagged for immediate reclaim and stall if any
1654 * are encountered in the nr_immediate check below.
1656 if (nr_writeback
&& nr_writeback
== nr_taken
)
1657 set_bit(ZONE_WRITEBACK
, &zone
->flags
);
1660 * Legacy memcg will stall in page writeback so avoid forcibly
1663 if (sane_reclaim(sc
)) {
1665 * Tag a zone as congested if all the dirty pages scanned were
1666 * backed by a congested BDI and wait_iff_congested will stall.
1668 if (nr_dirty
&& nr_dirty
== nr_congested
)
1669 set_bit(ZONE_CONGESTED
, &zone
->flags
);
1672 * If dirty pages are scanned that are not queued for IO, it
1673 * implies that flushers are not keeping up. In this case, flag
1674 * the zone ZONE_DIRTY and kswapd will start writing pages from
1677 if (nr_unqueued_dirty
== nr_taken
)
1678 set_bit(ZONE_DIRTY
, &zone
->flags
);
1681 * If kswapd scans pages marked marked for immediate
1682 * reclaim and under writeback (nr_immediate), it implies
1683 * that pages are cycling through the LRU faster than
1684 * they are written so also forcibly stall.
1686 if (nr_immediate
&& current_may_throttle())
1687 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1691 * Stall direct reclaim for IO completions if underlying BDIs or zone
1692 * is congested. Allow kswapd to continue until it starts encountering
1693 * unqueued dirty pages or cycling through the LRU too quickly.
1695 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1696 current_may_throttle())
1697 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1699 trace_mm_vmscan_lru_shrink_inactive(zone
, nr_scanned
, nr_reclaimed
,
1700 sc
->priority
, file
);
1701 return nr_reclaimed
;
1705 * This moves pages from the active list to the inactive list.
1707 * We move them the other way if the page is referenced by one or more
1708 * processes, from rmap.
1710 * If the pages are mostly unmapped, the processing is fast and it is
1711 * appropriate to hold zone->lru_lock across the whole operation. But if
1712 * the pages are mapped, the processing is slow (page_referenced()) so we
1713 * should drop zone->lru_lock around each page. It's impossible to balance
1714 * this, so instead we remove the pages from the LRU while processing them.
1715 * It is safe to rely on PG_active against the non-LRU pages in here because
1716 * nobody will play with that bit on a non-LRU page.
1718 * The downside is that we have to touch page->_count against each page.
1719 * But we had to alter page->flags anyway.
1722 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1723 struct list_head
*list
,
1724 struct list_head
*pages_to_free
,
1727 struct zone
*zone
= lruvec_zone(lruvec
);
1728 unsigned long pgmoved
= 0;
1732 while (!list_empty(list
)) {
1733 page
= lru_to_page(list
);
1734 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1736 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1739 nr_pages
= hpage_nr_pages(page
);
1740 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1741 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1742 pgmoved
+= nr_pages
;
1744 if (put_page_testzero(page
)) {
1745 __ClearPageLRU(page
);
1746 __ClearPageActive(page
);
1747 del_page_from_lru_list(page
, lruvec
, lru
);
1749 if (unlikely(PageCompound(page
))) {
1750 spin_unlock_irq(&zone
->lru_lock
);
1751 mem_cgroup_uncharge(page
);
1752 (*get_compound_page_dtor(page
))(page
);
1753 spin_lock_irq(&zone
->lru_lock
);
1755 list_add(&page
->lru
, pages_to_free
);
1758 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1759 if (!is_active_lru(lru
))
1760 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1763 static void shrink_active_list(unsigned long nr_to_scan
,
1764 struct lruvec
*lruvec
,
1765 struct scan_control
*sc
,
1768 unsigned long nr_taken
;
1769 unsigned long nr_scanned
;
1770 unsigned long vm_flags
;
1771 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1772 LIST_HEAD(l_active
);
1773 LIST_HEAD(l_inactive
);
1775 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1776 unsigned long nr_rotated
= 0;
1777 isolate_mode_t isolate_mode
= 0;
1778 int file
= is_file_lru(lru
);
1779 struct zone
*zone
= lruvec_zone(lruvec
);
1784 isolate_mode
|= ISOLATE_UNMAPPED
;
1785 if (!sc
->may_writepage
)
1786 isolate_mode
|= ISOLATE_CLEAN
;
1788 spin_lock_irq(&zone
->lru_lock
);
1790 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1791 &nr_scanned
, sc
, isolate_mode
, lru
);
1792 if (global_reclaim(sc
))
1793 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, nr_scanned
);
1795 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1797 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1798 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1799 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1800 spin_unlock_irq(&zone
->lru_lock
);
1802 while (!list_empty(&l_hold
)) {
1804 page
= lru_to_page(&l_hold
);
1805 list_del(&page
->lru
);
1807 if (unlikely(!page_evictable(page
))) {
1808 putback_lru_page(page
);
1812 if (unlikely(buffer_heads_over_limit
)) {
1813 if (page_has_private(page
) && trylock_page(page
)) {
1814 if (page_has_private(page
))
1815 try_to_release_page(page
, 0);
1820 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1822 nr_rotated
+= hpage_nr_pages(page
);
1824 * Identify referenced, file-backed active pages and
1825 * give them one more trip around the active list. So
1826 * that executable code get better chances to stay in
1827 * memory under moderate memory pressure. Anon pages
1828 * are not likely to be evicted by use-once streaming
1829 * IO, plus JVM can create lots of anon VM_EXEC pages,
1830 * so we ignore them here.
1832 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1833 list_add(&page
->lru
, &l_active
);
1838 ClearPageActive(page
); /* we are de-activating */
1839 list_add(&page
->lru
, &l_inactive
);
1843 * Move pages back to the lru list.
1845 spin_lock_irq(&zone
->lru_lock
);
1847 * Count referenced pages from currently used mappings as rotated,
1848 * even though only some of them are actually re-activated. This
1849 * helps balance scan pressure between file and anonymous pages in
1852 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1854 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1855 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1856 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1857 spin_unlock_irq(&zone
->lru_lock
);
1859 mem_cgroup_uncharge_list(&l_hold
);
1860 free_hot_cold_page_list(&l_hold
, true);
1864 static bool inactive_anon_is_low_global(struct zone
*zone
)
1866 unsigned long active
, inactive
;
1868 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1869 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1871 return inactive
* zone
->inactive_ratio
< active
;
1875 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1876 * @lruvec: LRU vector to check
1878 * Returns true if the zone does not have enough inactive anon pages,
1879 * meaning some active anon pages need to be deactivated.
1881 static bool inactive_anon_is_low(struct lruvec
*lruvec
)
1884 * If we don't have swap space, anonymous page deactivation
1887 if (!total_swap_pages
)
1890 if (!mem_cgroup_disabled())
1891 return mem_cgroup_inactive_anon_is_low(lruvec
);
1893 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1896 static inline bool inactive_anon_is_low(struct lruvec
*lruvec
)
1903 * inactive_file_is_low - check if file pages need to be deactivated
1904 * @lruvec: LRU vector to check
1906 * When the system is doing streaming IO, memory pressure here
1907 * ensures that active file pages get deactivated, until more
1908 * than half of the file pages are on the inactive list.
1910 * Once we get to that situation, protect the system's working
1911 * set from being evicted by disabling active file page aging.
1913 * This uses a different ratio than the anonymous pages, because
1914 * the page cache uses a use-once replacement algorithm.
1916 static bool inactive_file_is_low(struct lruvec
*lruvec
)
1918 unsigned long inactive
;
1919 unsigned long active
;
1921 inactive
= lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1922 active
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1924 return active
> inactive
;
1927 static bool inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1929 if (is_file_lru(lru
))
1930 return inactive_file_is_low(lruvec
);
1932 return inactive_anon_is_low(lruvec
);
1935 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1936 struct lruvec
*lruvec
, struct scan_control
*sc
)
1938 if (is_active_lru(lru
)) {
1939 if (inactive_list_is_low(lruvec
, lru
))
1940 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1944 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1955 * Determine how aggressively the anon and file LRU lists should be
1956 * scanned. The relative value of each set of LRU lists is determined
1957 * by looking at the fraction of the pages scanned we did rotate back
1958 * onto the active list instead of evict.
1960 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1961 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1963 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
1964 struct scan_control
*sc
, unsigned long *nr
,
1965 unsigned long *lru_pages
)
1967 int swappiness
= mem_cgroup_swappiness(memcg
);
1968 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1970 u64 denominator
= 0; /* gcc */
1971 struct zone
*zone
= lruvec_zone(lruvec
);
1972 unsigned long anon_prio
, file_prio
;
1973 enum scan_balance scan_balance
;
1974 unsigned long anon
, file
;
1975 bool force_scan
= false;
1976 unsigned long ap
, fp
;
1982 * If the zone or memcg is small, nr[l] can be 0. This
1983 * results in no scanning on this priority and a potential
1984 * priority drop. Global direct reclaim can go to the next
1985 * zone and tends to have no problems. Global kswapd is for
1986 * zone balancing and it needs to scan a minimum amount. When
1987 * reclaiming for a memcg, a priority drop can cause high
1988 * latencies, so it's better to scan a minimum amount there as
1991 if (current_is_kswapd()) {
1992 if (!zone_reclaimable(zone
))
1994 if (!mem_cgroup_online(memcg
))
1997 if (!global_reclaim(sc
))
2000 /* If we have no swap space, do not bother scanning anon pages. */
2001 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2002 scan_balance
= SCAN_FILE
;
2007 * Global reclaim will swap to prevent OOM even with no
2008 * swappiness, but memcg users want to use this knob to
2009 * disable swapping for individual groups completely when
2010 * using the memory controller's swap limit feature would be
2013 if (!global_reclaim(sc
) && !swappiness
) {
2014 scan_balance
= SCAN_FILE
;
2019 * Do not apply any pressure balancing cleverness when the
2020 * system is close to OOM, scan both anon and file equally
2021 * (unless the swappiness setting disagrees with swapping).
2023 if (!sc
->priority
&& swappiness
) {
2024 scan_balance
= SCAN_EQUAL
;
2029 * Prevent the reclaimer from falling into the cache trap: as
2030 * cache pages start out inactive, every cache fault will tip
2031 * the scan balance towards the file LRU. And as the file LRU
2032 * shrinks, so does the window for rotation from references.
2033 * This means we have a runaway feedback loop where a tiny
2034 * thrashing file LRU becomes infinitely more attractive than
2035 * anon pages. Try to detect this based on file LRU size.
2037 if (global_reclaim(sc
)) {
2038 unsigned long zonefile
;
2039 unsigned long zonefree
;
2041 zonefree
= zone_page_state(zone
, NR_FREE_PAGES
);
2042 zonefile
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2043 zone_page_state(zone
, NR_INACTIVE_FILE
);
2045 if (unlikely(zonefile
+ zonefree
<= high_wmark_pages(zone
))) {
2046 scan_balance
= SCAN_ANON
;
2052 * If there is enough inactive page cache, i.e. if the size of the
2053 * inactive list is greater than that of the active list *and* the
2054 * inactive list actually has some pages to scan on this priority, we
2055 * do not reclaim anything from the anonymous working set right now.
2056 * Without the second condition we could end up never scanning an
2057 * lruvec even if it has plenty of old anonymous pages unless the
2058 * system is under heavy pressure.
2060 if (!inactive_file_is_low(lruvec
) &&
2061 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2062 scan_balance
= SCAN_FILE
;
2066 scan_balance
= SCAN_FRACT
;
2069 * With swappiness at 100, anonymous and file have the same priority.
2070 * This scanning priority is essentially the inverse of IO cost.
2072 anon_prio
= swappiness
;
2073 file_prio
= 200 - anon_prio
;
2076 * OK, so we have swap space and a fair amount of page cache
2077 * pages. We use the recently rotated / recently scanned
2078 * ratios to determine how valuable each cache is.
2080 * Because workloads change over time (and to avoid overflow)
2081 * we keep these statistics as a floating average, which ends
2082 * up weighing recent references more than old ones.
2084 * anon in [0], file in [1]
2087 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2088 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2089 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2090 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2092 spin_lock_irq(&zone
->lru_lock
);
2093 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2094 reclaim_stat
->recent_scanned
[0] /= 2;
2095 reclaim_stat
->recent_rotated
[0] /= 2;
2098 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2099 reclaim_stat
->recent_scanned
[1] /= 2;
2100 reclaim_stat
->recent_rotated
[1] /= 2;
2104 * The amount of pressure on anon vs file pages is inversely
2105 * proportional to the fraction of recently scanned pages on
2106 * each list that were recently referenced and in active use.
2108 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2109 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2111 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2112 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2113 spin_unlock_irq(&zone
->lru_lock
);
2117 denominator
= ap
+ fp
+ 1;
2119 some_scanned
= false;
2120 /* Only use force_scan on second pass. */
2121 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2123 for_each_evictable_lru(lru
) {
2124 int file
= is_file_lru(lru
);
2128 size
= lruvec_lru_size(lruvec
, lru
);
2129 scan
= size
>> sc
->priority
;
2131 if (!scan
&& pass
&& force_scan
)
2132 scan
= min(size
, SWAP_CLUSTER_MAX
);
2134 switch (scan_balance
) {
2136 /* Scan lists relative to size */
2140 * Scan types proportional to swappiness and
2141 * their relative recent reclaim efficiency.
2143 scan
= div64_u64(scan
* fraction
[file
],
2148 /* Scan one type exclusively */
2149 if ((scan_balance
== SCAN_FILE
) != file
) {
2155 /* Look ma, no brain */
2163 * Skip the second pass and don't force_scan,
2164 * if we found something to scan.
2166 some_scanned
|= !!scan
;
2171 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2172 static void init_tlb_ubc(void)
2175 * This deliberately does not clear the cpumask as it's expensive
2176 * and unnecessary. If there happens to be data in there then the
2177 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2178 * then will be cleared.
2180 current
->tlb_ubc
.flush_required
= false;
2183 static inline void init_tlb_ubc(void)
2186 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2189 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2191 static void shrink_zone_memcg(struct zone
*zone
, struct mem_cgroup
*memcg
,
2192 struct scan_control
*sc
, unsigned long *lru_pages
)
2194 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2195 unsigned long nr
[NR_LRU_LISTS
];
2196 unsigned long targets
[NR_LRU_LISTS
];
2197 unsigned long nr_to_scan
;
2199 unsigned long nr_reclaimed
= 0;
2200 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2201 struct blk_plug plug
;
2204 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2206 /* Record the original scan target for proportional adjustments later */
2207 memcpy(targets
, nr
, sizeof(nr
));
2210 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2211 * event that can occur when there is little memory pressure e.g.
2212 * multiple streaming readers/writers. Hence, we do not abort scanning
2213 * when the requested number of pages are reclaimed when scanning at
2214 * DEF_PRIORITY on the assumption that the fact we are direct
2215 * reclaiming implies that kswapd is not keeping up and it is best to
2216 * do a batch of work at once. For memcg reclaim one check is made to
2217 * abort proportional reclaim if either the file or anon lru has already
2218 * dropped to zero at the first pass.
2220 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2221 sc
->priority
== DEF_PRIORITY
);
2225 blk_start_plug(&plug
);
2226 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2227 nr
[LRU_INACTIVE_FILE
]) {
2228 unsigned long nr_anon
, nr_file
, percentage
;
2229 unsigned long nr_scanned
;
2231 for_each_evictable_lru(lru
) {
2233 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2234 nr
[lru
] -= nr_to_scan
;
2236 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2241 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2245 * For kswapd and memcg, reclaim at least the number of pages
2246 * requested. Ensure that the anon and file LRUs are scanned
2247 * proportionally what was requested by get_scan_count(). We
2248 * stop reclaiming one LRU and reduce the amount scanning
2249 * proportional to the original scan target.
2251 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2252 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2255 * It's just vindictive to attack the larger once the smaller
2256 * has gone to zero. And given the way we stop scanning the
2257 * smaller below, this makes sure that we only make one nudge
2258 * towards proportionality once we've got nr_to_reclaim.
2260 if (!nr_file
|| !nr_anon
)
2263 if (nr_file
> nr_anon
) {
2264 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2265 targets
[LRU_ACTIVE_ANON
] + 1;
2267 percentage
= nr_anon
* 100 / scan_target
;
2269 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2270 targets
[LRU_ACTIVE_FILE
] + 1;
2272 percentage
= nr_file
* 100 / scan_target
;
2275 /* Stop scanning the smaller of the LRU */
2277 nr
[lru
+ LRU_ACTIVE
] = 0;
2280 * Recalculate the other LRU scan count based on its original
2281 * scan target and the percentage scanning already complete
2283 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2284 nr_scanned
= targets
[lru
] - nr
[lru
];
2285 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2286 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2289 nr_scanned
= targets
[lru
] - nr
[lru
];
2290 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2291 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2293 scan_adjusted
= true;
2295 blk_finish_plug(&plug
);
2296 sc
->nr_reclaimed
+= nr_reclaimed
;
2299 * Even if we did not try to evict anon pages at all, we want to
2300 * rebalance the anon lru active/inactive ratio.
2302 if (inactive_anon_is_low(lruvec
))
2303 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2304 sc
, LRU_ACTIVE_ANON
);
2306 throttle_vm_writeout(sc
->gfp_mask
);
2309 /* Use reclaim/compaction for costly allocs or under memory pressure */
2310 static bool in_reclaim_compaction(struct scan_control
*sc
)
2312 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2313 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2314 sc
->priority
< DEF_PRIORITY
- 2))
2321 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2322 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2323 * true if more pages should be reclaimed such that when the page allocator
2324 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2325 * It will give up earlier than that if there is difficulty reclaiming pages.
2327 static inline bool should_continue_reclaim(struct zone
*zone
,
2328 unsigned long nr_reclaimed
,
2329 unsigned long nr_scanned
,
2330 struct scan_control
*sc
)
2332 unsigned long pages_for_compaction
;
2333 unsigned long inactive_lru_pages
;
2335 /* If not in reclaim/compaction mode, stop */
2336 if (!in_reclaim_compaction(sc
))
2339 /* Consider stopping depending on scan and reclaim activity */
2340 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2342 * For __GFP_REPEAT allocations, stop reclaiming if the
2343 * full LRU list has been scanned and we are still failing
2344 * to reclaim pages. This full LRU scan is potentially
2345 * expensive but a __GFP_REPEAT caller really wants to succeed
2347 if (!nr_reclaimed
&& !nr_scanned
)
2351 * For non-__GFP_REPEAT allocations which can presumably
2352 * fail without consequence, stop if we failed to reclaim
2353 * any pages from the last SWAP_CLUSTER_MAX number of
2354 * pages that were scanned. This will return to the
2355 * caller faster at the risk reclaim/compaction and
2356 * the resulting allocation attempt fails
2363 * If we have not reclaimed enough pages for compaction and the
2364 * inactive lists are large enough, continue reclaiming
2366 pages_for_compaction
= (2UL << sc
->order
);
2367 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2368 if (get_nr_swap_pages() > 0)
2369 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2370 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2371 inactive_lru_pages
> pages_for_compaction
)
2374 /* If compaction would go ahead or the allocation would succeed, stop */
2375 switch (compaction_suitable(zone
, sc
->order
, 0, 0)) {
2376 case COMPACT_PARTIAL
:
2377 case COMPACT_CONTINUE
:
2384 static bool shrink_zone(struct zone
*zone
, struct scan_control
*sc
,
2387 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2388 unsigned long nr_reclaimed
, nr_scanned
;
2389 bool reclaimable
= false;
2392 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2393 struct mem_cgroup_reclaim_cookie reclaim
= {
2395 .priority
= sc
->priority
,
2397 unsigned long zone_lru_pages
= 0;
2398 struct mem_cgroup
*memcg
;
2400 nr_reclaimed
= sc
->nr_reclaimed
;
2401 nr_scanned
= sc
->nr_scanned
;
2403 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2405 unsigned long lru_pages
;
2406 unsigned long reclaimed
;
2407 unsigned long scanned
;
2409 if (mem_cgroup_low(root
, memcg
)) {
2410 if (!sc
->may_thrash
)
2412 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2415 reclaimed
= sc
->nr_reclaimed
;
2416 scanned
= sc
->nr_scanned
;
2418 shrink_zone_memcg(zone
, memcg
, sc
, &lru_pages
);
2419 zone_lru_pages
+= lru_pages
;
2421 if (memcg
&& is_classzone
)
2422 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
),
2423 memcg
, sc
->nr_scanned
- scanned
,
2426 /* Record the group's reclaim efficiency */
2427 vmpressure(sc
->gfp_mask
, memcg
, false,
2428 sc
->nr_scanned
- scanned
,
2429 sc
->nr_reclaimed
- reclaimed
);
2432 * Direct reclaim and kswapd have to scan all memory
2433 * cgroups to fulfill the overall scan target for the
2436 * Limit reclaim, on the other hand, only cares about
2437 * nr_to_reclaim pages to be reclaimed and it will
2438 * retry with decreasing priority if one round over the
2439 * whole hierarchy is not sufficient.
2441 if (!global_reclaim(sc
) &&
2442 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2443 mem_cgroup_iter_break(root
, memcg
);
2446 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2449 * Shrink the slab caches in the same proportion that
2450 * the eligible LRU pages were scanned.
2452 if (global_reclaim(sc
) && is_classzone
)
2453 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
), NULL
,
2454 sc
->nr_scanned
- nr_scanned
,
2457 if (reclaim_state
) {
2458 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2459 reclaim_state
->reclaimed_slab
= 0;
2462 /* Record the subtree's reclaim efficiency */
2463 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2464 sc
->nr_scanned
- nr_scanned
,
2465 sc
->nr_reclaimed
- nr_reclaimed
);
2467 if (sc
->nr_reclaimed
- nr_reclaimed
)
2470 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2471 sc
->nr_scanned
- nr_scanned
, sc
));
2477 * Returns true if compaction should go ahead for a high-order request, or
2478 * the high-order allocation would succeed without compaction.
2480 static inline bool compaction_ready(struct zone
*zone
, int order
)
2482 unsigned long balance_gap
, watermark
;
2486 * Compaction takes time to run and there are potentially other
2487 * callers using the pages just freed. Continue reclaiming until
2488 * there is a buffer of free pages available to give compaction
2489 * a reasonable chance of completing and allocating the page
2491 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2492 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2493 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << order
);
2494 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0);
2497 * If compaction is deferred, reclaim up to a point where
2498 * compaction will have a chance of success when re-enabled
2500 if (compaction_deferred(zone
, order
))
2501 return watermark_ok
;
2504 * If compaction is not ready to start and allocation is not likely
2505 * to succeed without it, then keep reclaiming.
2507 if (compaction_suitable(zone
, order
, 0, 0) == COMPACT_SKIPPED
)
2510 return watermark_ok
;
2514 * This is the direct reclaim path, for page-allocating processes. We only
2515 * try to reclaim pages from zones which will satisfy the caller's allocation
2518 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2520 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2522 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2523 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2524 * zone defense algorithm.
2526 * If a zone is deemed to be full of pinned pages then just give it a light
2527 * scan then give up on it.
2529 * Returns true if a zone was reclaimable.
2531 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2535 unsigned long nr_soft_reclaimed
;
2536 unsigned long nr_soft_scanned
;
2538 enum zone_type requested_highidx
= gfp_zone(sc
->gfp_mask
);
2539 bool reclaimable
= false;
2542 * If the number of buffer_heads in the machine exceeds the maximum
2543 * allowed level, force direct reclaim to scan the highmem zone as
2544 * highmem pages could be pinning lowmem pages storing buffer_heads
2546 orig_mask
= sc
->gfp_mask
;
2547 if (buffer_heads_over_limit
)
2548 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2550 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2551 requested_highidx
, sc
->nodemask
) {
2552 enum zone_type classzone_idx
;
2554 if (!populated_zone(zone
))
2557 classzone_idx
= requested_highidx
;
2558 while (!populated_zone(zone
->zone_pgdat
->node_zones
+
2563 * Take care memory controller reclaiming has small influence
2566 if (global_reclaim(sc
)) {
2567 if (!cpuset_zone_allowed(zone
,
2568 GFP_KERNEL
| __GFP_HARDWALL
))
2571 if (sc
->priority
!= DEF_PRIORITY
&&
2572 !zone_reclaimable(zone
))
2573 continue; /* Let kswapd poll it */
2576 * If we already have plenty of memory free for
2577 * compaction in this zone, don't free any more.
2578 * Even though compaction is invoked for any
2579 * non-zero order, only frequent costly order
2580 * reclamation is disruptive enough to become a
2581 * noticeable problem, like transparent huge
2584 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2585 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2586 zonelist_zone_idx(z
) <= requested_highidx
&&
2587 compaction_ready(zone
, sc
->order
)) {
2588 sc
->compaction_ready
= true;
2593 * This steals pages from memory cgroups over softlimit
2594 * and returns the number of reclaimed pages and
2595 * scanned pages. This works for global memory pressure
2596 * and balancing, not for a memcg's limit.
2598 nr_soft_scanned
= 0;
2599 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2600 sc
->order
, sc
->gfp_mask
,
2602 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2603 sc
->nr_scanned
+= nr_soft_scanned
;
2604 if (nr_soft_reclaimed
)
2606 /* need some check for avoid more shrink_zone() */
2609 if (shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
))
2612 if (global_reclaim(sc
) &&
2613 !reclaimable
&& zone_reclaimable(zone
))
2618 * Restore to original mask to avoid the impact on the caller if we
2619 * promoted it to __GFP_HIGHMEM.
2621 sc
->gfp_mask
= orig_mask
;
2627 * This is the main entry point to direct page reclaim.
2629 * If a full scan of the inactive list fails to free enough memory then we
2630 * are "out of memory" and something needs to be killed.
2632 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2633 * high - the zone may be full of dirty or under-writeback pages, which this
2634 * caller can't do much about. We kick the writeback threads and take explicit
2635 * naps in the hope that some of these pages can be written. But if the
2636 * allocating task holds filesystem locks which prevent writeout this might not
2637 * work, and the allocation attempt will fail.
2639 * returns: 0, if no pages reclaimed
2640 * else, the number of pages reclaimed
2642 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2643 struct scan_control
*sc
)
2645 int initial_priority
= sc
->priority
;
2646 unsigned long total_scanned
= 0;
2647 unsigned long writeback_threshold
;
2648 bool zones_reclaimable
;
2650 delayacct_freepages_start();
2652 if (global_reclaim(sc
))
2653 count_vm_event(ALLOCSTALL
);
2656 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2659 zones_reclaimable
= shrink_zones(zonelist
, sc
);
2661 total_scanned
+= sc
->nr_scanned
;
2662 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2665 if (sc
->compaction_ready
)
2669 * If we're getting trouble reclaiming, start doing
2670 * writepage even in laptop mode.
2672 if (sc
->priority
< DEF_PRIORITY
- 2)
2673 sc
->may_writepage
= 1;
2676 * Try to write back as many pages as we just scanned. This
2677 * tends to cause slow streaming writers to write data to the
2678 * disk smoothly, at the dirtying rate, which is nice. But
2679 * that's undesirable in laptop mode, where we *want* lumpy
2680 * writeout. So in laptop mode, write out the whole world.
2682 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2683 if (total_scanned
> writeback_threshold
) {
2684 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2685 WB_REASON_TRY_TO_FREE_PAGES
);
2686 sc
->may_writepage
= 1;
2688 } while (--sc
->priority
>= 0);
2690 delayacct_freepages_end();
2692 if (sc
->nr_reclaimed
)
2693 return sc
->nr_reclaimed
;
2695 /* Aborted reclaim to try compaction? don't OOM, then */
2696 if (sc
->compaction_ready
)
2699 /* Untapped cgroup reserves? Don't OOM, retry. */
2700 if (!sc
->may_thrash
) {
2701 sc
->priority
= initial_priority
;
2706 /* Any of the zones still reclaimable? Don't OOM. */
2707 if (zones_reclaimable
)
2713 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2716 unsigned long pfmemalloc_reserve
= 0;
2717 unsigned long free_pages
= 0;
2721 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2722 zone
= &pgdat
->node_zones
[i
];
2723 if (!populated_zone(zone
) ||
2724 zone_reclaimable_pages(zone
) == 0)
2727 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2728 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2731 /* If there are no reserves (unexpected config) then do not throttle */
2732 if (!pfmemalloc_reserve
)
2735 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2737 /* kswapd must be awake if processes are being throttled */
2738 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2739 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2740 (enum zone_type
)ZONE_NORMAL
);
2741 wake_up_interruptible(&pgdat
->kswapd_wait
);
2748 * Throttle direct reclaimers if backing storage is backed by the network
2749 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2750 * depleted. kswapd will continue to make progress and wake the processes
2751 * when the low watermark is reached.
2753 * Returns true if a fatal signal was delivered during throttling. If this
2754 * happens, the page allocator should not consider triggering the OOM killer.
2756 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2757 nodemask_t
*nodemask
)
2761 pg_data_t
*pgdat
= NULL
;
2764 * Kernel threads should not be throttled as they may be indirectly
2765 * responsible for cleaning pages necessary for reclaim to make forward
2766 * progress. kjournald for example may enter direct reclaim while
2767 * committing a transaction where throttling it could forcing other
2768 * processes to block on log_wait_commit().
2770 if (current
->flags
& PF_KTHREAD
)
2774 * If a fatal signal is pending, this process should not throttle.
2775 * It should return quickly so it can exit and free its memory
2777 if (fatal_signal_pending(current
))
2781 * Check if the pfmemalloc reserves are ok by finding the first node
2782 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2783 * GFP_KERNEL will be required for allocating network buffers when
2784 * swapping over the network so ZONE_HIGHMEM is unusable.
2786 * Throttling is based on the first usable node and throttled processes
2787 * wait on a queue until kswapd makes progress and wakes them. There
2788 * is an affinity then between processes waking up and where reclaim
2789 * progress has been made assuming the process wakes on the same node.
2790 * More importantly, processes running on remote nodes will not compete
2791 * for remote pfmemalloc reserves and processes on different nodes
2792 * should make reasonable progress.
2794 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2795 gfp_zone(gfp_mask
), nodemask
) {
2796 if (zone_idx(zone
) > ZONE_NORMAL
)
2799 /* Throttle based on the first usable node */
2800 pgdat
= zone
->zone_pgdat
;
2801 if (pfmemalloc_watermark_ok(pgdat
))
2806 /* If no zone was usable by the allocation flags then do not throttle */
2810 /* Account for the throttling */
2811 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2814 * If the caller cannot enter the filesystem, it's possible that it
2815 * is due to the caller holding an FS lock or performing a journal
2816 * transaction in the case of a filesystem like ext[3|4]. In this case,
2817 * it is not safe to block on pfmemalloc_wait as kswapd could be
2818 * blocked waiting on the same lock. Instead, throttle for up to a
2819 * second before continuing.
2821 if (!(gfp_mask
& __GFP_FS
)) {
2822 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2823 pfmemalloc_watermark_ok(pgdat
), HZ
);
2828 /* Throttle until kswapd wakes the process */
2829 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2830 pfmemalloc_watermark_ok(pgdat
));
2833 if (fatal_signal_pending(current
))
2840 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2841 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2843 unsigned long nr_reclaimed
;
2844 struct scan_control sc
= {
2845 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2846 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2848 .nodemask
= nodemask
,
2849 .priority
= DEF_PRIORITY
,
2850 .may_writepage
= !laptop_mode
,
2856 * Do not enter reclaim if fatal signal was delivered while throttled.
2857 * 1 is returned so that the page allocator does not OOM kill at this
2860 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2863 trace_mm_vmscan_direct_reclaim_begin(order
,
2867 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2869 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2871 return nr_reclaimed
;
2876 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2877 gfp_t gfp_mask
, bool noswap
,
2879 unsigned long *nr_scanned
)
2881 struct scan_control sc
= {
2882 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2883 .target_mem_cgroup
= memcg
,
2884 .may_writepage
= !laptop_mode
,
2886 .may_swap
= !noswap
,
2888 unsigned long lru_pages
;
2890 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2891 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2893 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2898 * NOTE: Although we can get the priority field, using it
2899 * here is not a good idea, since it limits the pages we can scan.
2900 * if we don't reclaim here, the shrink_zone from balance_pgdat
2901 * will pick up pages from other mem cgroup's as well. We hack
2902 * the priority and make it zero.
2904 shrink_zone_memcg(zone
, memcg
, &sc
, &lru_pages
);
2906 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2908 *nr_scanned
= sc
.nr_scanned
;
2909 return sc
.nr_reclaimed
;
2912 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2913 unsigned long nr_pages
,
2917 struct zonelist
*zonelist
;
2918 unsigned long nr_reclaimed
;
2920 struct scan_control sc
= {
2921 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2922 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2923 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2924 .target_mem_cgroup
= memcg
,
2925 .priority
= DEF_PRIORITY
,
2926 .may_writepage
= !laptop_mode
,
2928 .may_swap
= may_swap
,
2932 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2933 * take care of from where we get pages. So the node where we start the
2934 * scan does not need to be the current node.
2936 nid
= mem_cgroup_select_victim_node(memcg
);
2938 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2940 trace_mm_vmscan_memcg_reclaim_begin(0,
2944 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2946 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2948 return nr_reclaimed
;
2952 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2954 struct mem_cgroup
*memcg
;
2956 if (!total_swap_pages
)
2959 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2961 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2963 if (inactive_anon_is_low(lruvec
))
2964 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2965 sc
, LRU_ACTIVE_ANON
);
2967 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2971 static bool zone_balanced(struct zone
*zone
, int order
,
2972 unsigned long balance_gap
, int classzone_idx
)
2974 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2975 balance_gap
, classzone_idx
))
2978 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&& compaction_suitable(zone
,
2979 order
, 0, classzone_idx
) == COMPACT_SKIPPED
)
2986 * pgdat_balanced() is used when checking if a node is balanced.
2988 * For order-0, all zones must be balanced!
2990 * For high-order allocations only zones that meet watermarks and are in a
2991 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2992 * total of balanced pages must be at least 25% of the zones allowed by
2993 * classzone_idx for the node to be considered balanced. Forcing all zones to
2994 * be balanced for high orders can cause excessive reclaim when there are
2996 * The choice of 25% is due to
2997 * o a 16M DMA zone that is balanced will not balance a zone on any
2998 * reasonable sized machine
2999 * o On all other machines, the top zone must be at least a reasonable
3000 * percentage of the middle zones. For example, on 32-bit x86, highmem
3001 * would need to be at least 256M for it to be balance a whole node.
3002 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3003 * to balance a node on its own. These seemed like reasonable ratios.
3005 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3007 unsigned long managed_pages
= 0;
3008 unsigned long balanced_pages
= 0;
3011 /* Check the watermark levels */
3012 for (i
= 0; i
<= classzone_idx
; i
++) {
3013 struct zone
*zone
= pgdat
->node_zones
+ i
;
3015 if (!populated_zone(zone
))
3018 managed_pages
+= zone
->managed_pages
;
3021 * A special case here:
3023 * balance_pgdat() skips over all_unreclaimable after
3024 * DEF_PRIORITY. Effectively, it considers them balanced so
3025 * they must be considered balanced here as well!
3027 if (!zone_reclaimable(zone
)) {
3028 balanced_pages
+= zone
->managed_pages
;
3032 if (zone_balanced(zone
, order
, 0, i
))
3033 balanced_pages
+= zone
->managed_pages
;
3039 return balanced_pages
>= (managed_pages
>> 2);
3045 * Prepare kswapd for sleeping. This verifies that there are no processes
3046 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3048 * Returns true if kswapd is ready to sleep
3050 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
3053 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3058 * The throttled processes are normally woken up in balance_pgdat() as
3059 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3060 * race between when kswapd checks the watermarks and a process gets
3061 * throttled. There is also a potential race if processes get
3062 * throttled, kswapd wakes, a large process exits thereby balancing the
3063 * zones, which causes kswapd to exit balance_pgdat() before reaching
3064 * the wake up checks. If kswapd is going to sleep, no process should
3065 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3066 * the wake up is premature, processes will wake kswapd and get
3067 * throttled again. The difference from wake ups in balance_pgdat() is
3068 * that here we are under prepare_to_wait().
3070 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3071 wake_up_all(&pgdat
->pfmemalloc_wait
);
3073 return pgdat_balanced(pgdat
, order
, classzone_idx
);
3077 * kswapd shrinks the zone by the number of pages required to reach
3078 * the high watermark.
3080 * Returns true if kswapd scanned at least the requested number of pages to
3081 * reclaim or if the lack of progress was due to pages under writeback.
3082 * This is used to determine if the scanning priority needs to be raised.
3084 static bool kswapd_shrink_zone(struct zone
*zone
,
3086 struct scan_control
*sc
,
3087 unsigned long *nr_attempted
)
3089 int testorder
= sc
->order
;
3090 unsigned long balance_gap
;
3091 bool lowmem_pressure
;
3093 /* Reclaim above the high watermark. */
3094 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
3097 * Kswapd reclaims only single pages with compaction enabled. Trying
3098 * too hard to reclaim until contiguous free pages have become
3099 * available can hurt performance by evicting too much useful data
3100 * from memory. Do not reclaim more than needed for compaction.
3102 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
3103 compaction_suitable(zone
, sc
->order
, 0, classzone_idx
)
3108 * We put equal pressure on every zone, unless one zone has way too
3109 * many pages free already. The "too many pages" is defined as the
3110 * high wmark plus a "gap" where the gap is either the low
3111 * watermark or 1% of the zone, whichever is smaller.
3113 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
3114 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
3117 * If there is no low memory pressure or the zone is balanced then no
3118 * reclaim is necessary
3120 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
3121 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
3122 balance_gap
, classzone_idx
))
3125 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
3127 /* Account for the number of pages attempted to reclaim */
3128 *nr_attempted
+= sc
->nr_to_reclaim
;
3130 clear_bit(ZONE_WRITEBACK
, &zone
->flags
);
3133 * If a zone reaches its high watermark, consider it to be no longer
3134 * congested. It's possible there are dirty pages backed by congested
3135 * BDIs but as pressure is relieved, speculatively avoid congestion
3138 if (zone_reclaimable(zone
) &&
3139 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
3140 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3141 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3144 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3148 * For kswapd, balance_pgdat() will work across all this node's zones until
3149 * they are all at high_wmark_pages(zone).
3151 * Returns the final order kswapd was reclaiming at
3153 * There is special handling here for zones which are full of pinned pages.
3154 * This can happen if the pages are all mlocked, or if they are all used by
3155 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3156 * What we do is to detect the case where all pages in the zone have been
3157 * scanned twice and there has been zero successful reclaim. Mark the zone as
3158 * dead and from now on, only perform a short scan. Basically we're polling
3159 * the zone for when the problem goes away.
3161 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3162 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3163 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3164 * lower zones regardless of the number of free pages in the lower zones. This
3165 * interoperates with the page allocator fallback scheme to ensure that aging
3166 * of pages is balanced across the zones.
3168 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
3172 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
3173 unsigned long nr_soft_reclaimed
;
3174 unsigned long nr_soft_scanned
;
3175 struct scan_control sc
= {
3176 .gfp_mask
= GFP_KERNEL
,
3178 .priority
= DEF_PRIORITY
,
3179 .may_writepage
= !laptop_mode
,
3183 count_vm_event(PAGEOUTRUN
);
3186 unsigned long nr_attempted
= 0;
3187 bool raise_priority
= true;
3188 bool pgdat_needs_compaction
= (order
> 0);
3190 sc
.nr_reclaimed
= 0;
3193 * Scan in the highmem->dma direction for the highest
3194 * zone which needs scanning
3196 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
3197 struct zone
*zone
= pgdat
->node_zones
+ i
;
3199 if (!populated_zone(zone
))
3202 if (sc
.priority
!= DEF_PRIORITY
&&
3203 !zone_reclaimable(zone
))
3207 * Do some background aging of the anon list, to give
3208 * pages a chance to be referenced before reclaiming.
3210 age_active_anon(zone
, &sc
);
3213 * If the number of buffer_heads in the machine
3214 * exceeds the maximum allowed level and this node
3215 * has a highmem zone, force kswapd to reclaim from
3216 * it to relieve lowmem pressure.
3218 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
3223 if (!zone_balanced(zone
, order
, 0, 0)) {
3228 * If balanced, clear the dirty and congested
3231 clear_bit(ZONE_CONGESTED
, &zone
->flags
);
3232 clear_bit(ZONE_DIRTY
, &zone
->flags
);
3239 for (i
= 0; i
<= end_zone
; i
++) {
3240 struct zone
*zone
= pgdat
->node_zones
+ i
;
3242 if (!populated_zone(zone
))
3246 * If any zone is currently balanced then kswapd will
3247 * not call compaction as it is expected that the
3248 * necessary pages are already available.
3250 if (pgdat_needs_compaction
&&
3251 zone_watermark_ok(zone
, order
,
3252 low_wmark_pages(zone
),
3254 pgdat_needs_compaction
= false;
3258 * If we're getting trouble reclaiming, start doing writepage
3259 * even in laptop mode.
3261 if (sc
.priority
< DEF_PRIORITY
- 2)
3262 sc
.may_writepage
= 1;
3265 * Now scan the zone in the dma->highmem direction, stopping
3266 * at the last zone which needs scanning.
3268 * We do this because the page allocator works in the opposite
3269 * direction. This prevents the page allocator from allocating
3270 * pages behind kswapd's direction of progress, which would
3271 * cause too much scanning of the lower zones.
3273 for (i
= 0; i
<= end_zone
; i
++) {
3274 struct zone
*zone
= pgdat
->node_zones
+ i
;
3276 if (!populated_zone(zone
))
3279 if (sc
.priority
!= DEF_PRIORITY
&&
3280 !zone_reclaimable(zone
))
3285 nr_soft_scanned
= 0;
3287 * Call soft limit reclaim before calling shrink_zone.
3289 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3292 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3295 * There should be no need to raise the scanning
3296 * priority if enough pages are already being scanned
3297 * that that high watermark would be met at 100%
3300 if (kswapd_shrink_zone(zone
, end_zone
,
3301 &sc
, &nr_attempted
))
3302 raise_priority
= false;
3306 * If the low watermark is met there is no need for processes
3307 * to be throttled on pfmemalloc_wait as they should not be
3308 * able to safely make forward progress. Wake them
3310 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3311 pfmemalloc_watermark_ok(pgdat
))
3312 wake_up_all(&pgdat
->pfmemalloc_wait
);
3315 * Fragmentation may mean that the system cannot be rebalanced
3316 * for high-order allocations in all zones. If twice the
3317 * allocation size has been reclaimed and the zones are still
3318 * not balanced then recheck the watermarks at order-0 to
3319 * prevent kswapd reclaiming excessively. Assume that a
3320 * process requested a high-order can direct reclaim/compact.
3322 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3323 order
= sc
.order
= 0;
3325 /* Check if kswapd should be suspending */
3326 if (try_to_freeze() || kthread_should_stop())
3330 * Compact if necessary and kswapd is reclaiming at least the
3331 * high watermark number of pages as requsted
3333 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3334 compact_pgdat(pgdat
, order
);
3337 * Raise priority if scanning rate is too low or there was no
3338 * progress in reclaiming pages
3340 if (raise_priority
|| !sc
.nr_reclaimed
)
3342 } while (sc
.priority
>= 1 &&
3343 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3347 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3348 * makes a decision on the order we were last reclaiming at. However,
3349 * if another caller entered the allocator slow path while kswapd
3350 * was awake, order will remain at the higher level
3352 *classzone_idx
= end_zone
;
3356 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3361 if (freezing(current
) || kthread_should_stop())
3364 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3366 /* Try to sleep for a short interval */
3367 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3368 remaining
= schedule_timeout(HZ
/10);
3369 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3370 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3374 * After a short sleep, check if it was a premature sleep. If not, then
3375 * go fully to sleep until explicitly woken up.
3377 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3378 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3381 * vmstat counters are not perfectly accurate and the estimated
3382 * value for counters such as NR_FREE_PAGES can deviate from the
3383 * true value by nr_online_cpus * threshold. To avoid the zone
3384 * watermarks being breached while under pressure, we reduce the
3385 * per-cpu vmstat threshold while kswapd is awake and restore
3386 * them before going back to sleep.
3388 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3391 * Compaction records what page blocks it recently failed to
3392 * isolate pages from and skips them in the future scanning.
3393 * When kswapd is going to sleep, it is reasonable to assume
3394 * that pages and compaction may succeed so reset the cache.
3396 reset_isolation_suitable(pgdat
);
3398 if (!kthread_should_stop())
3401 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3404 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3406 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3408 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3412 * The background pageout daemon, started as a kernel thread
3413 * from the init process.
3415 * This basically trickles out pages so that we have _some_
3416 * free memory available even if there is no other activity
3417 * that frees anything up. This is needed for things like routing
3418 * etc, where we otherwise might have all activity going on in
3419 * asynchronous contexts that cannot page things out.
3421 * If there are applications that are active memory-allocators
3422 * (most normal use), this basically shouldn't matter.
3424 static int kswapd(void *p
)
3426 unsigned long order
, new_order
;
3427 unsigned balanced_order
;
3428 int classzone_idx
, new_classzone_idx
;
3429 int balanced_classzone_idx
;
3430 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3431 struct task_struct
*tsk
= current
;
3433 struct reclaim_state reclaim_state
= {
3434 .reclaimed_slab
= 0,
3436 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3438 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3440 if (!cpumask_empty(cpumask
))
3441 set_cpus_allowed_ptr(tsk
, cpumask
);
3442 current
->reclaim_state
= &reclaim_state
;
3445 * Tell the memory management that we're a "memory allocator",
3446 * and that if we need more memory we should get access to it
3447 * regardless (see "__alloc_pages()"). "kswapd" should
3448 * never get caught in the normal page freeing logic.
3450 * (Kswapd normally doesn't need memory anyway, but sometimes
3451 * you need a small amount of memory in order to be able to
3452 * page out something else, and this flag essentially protects
3453 * us from recursively trying to free more memory as we're
3454 * trying to free the first piece of memory in the first place).
3456 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3459 order
= new_order
= 0;
3461 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3462 balanced_classzone_idx
= classzone_idx
;
3467 * If the last balance_pgdat was unsuccessful it's unlikely a
3468 * new request of a similar or harder type will succeed soon
3469 * so consider going to sleep on the basis we reclaimed at
3471 if (balanced_classzone_idx
>= new_classzone_idx
&&
3472 balanced_order
== new_order
) {
3473 new_order
= pgdat
->kswapd_max_order
;
3474 new_classzone_idx
= pgdat
->classzone_idx
;
3475 pgdat
->kswapd_max_order
= 0;
3476 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3479 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3481 * Don't sleep if someone wants a larger 'order'
3482 * allocation or has tigher zone constraints
3485 classzone_idx
= new_classzone_idx
;
3487 kswapd_try_to_sleep(pgdat
, balanced_order
,
3488 balanced_classzone_idx
);
3489 order
= pgdat
->kswapd_max_order
;
3490 classzone_idx
= pgdat
->classzone_idx
;
3492 new_classzone_idx
= classzone_idx
;
3493 pgdat
->kswapd_max_order
= 0;
3494 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3497 ret
= try_to_freeze();
3498 if (kthread_should_stop())
3502 * We can speed up thawing tasks if we don't call balance_pgdat
3503 * after returning from the refrigerator
3506 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3507 balanced_classzone_idx
= classzone_idx
;
3508 balanced_order
= balance_pgdat(pgdat
, order
,
3509 &balanced_classzone_idx
);
3513 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3514 current
->reclaim_state
= NULL
;
3515 lockdep_clear_current_reclaim_state();
3521 * A zone is low on free memory, so wake its kswapd task to service it.
3523 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3527 if (!populated_zone(zone
))
3530 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3532 pgdat
= zone
->zone_pgdat
;
3533 if (pgdat
->kswapd_max_order
< order
) {
3534 pgdat
->kswapd_max_order
= order
;
3535 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3537 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3539 if (zone_balanced(zone
, order
, 0, 0))
3542 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3543 wake_up_interruptible(&pgdat
->kswapd_wait
);
3546 #ifdef CONFIG_HIBERNATION
3548 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3551 * Rather than trying to age LRUs the aim is to preserve the overall
3552 * LRU order by reclaiming preferentially
3553 * inactive > active > active referenced > active mapped
3555 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3557 struct reclaim_state reclaim_state
;
3558 struct scan_control sc
= {
3559 .nr_to_reclaim
= nr_to_reclaim
,
3560 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3561 .priority
= DEF_PRIORITY
,
3565 .hibernation_mode
= 1,
3567 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3568 struct task_struct
*p
= current
;
3569 unsigned long nr_reclaimed
;
3571 p
->flags
|= PF_MEMALLOC
;
3572 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3573 reclaim_state
.reclaimed_slab
= 0;
3574 p
->reclaim_state
= &reclaim_state
;
3576 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3578 p
->reclaim_state
= NULL
;
3579 lockdep_clear_current_reclaim_state();
3580 p
->flags
&= ~PF_MEMALLOC
;
3582 return nr_reclaimed
;
3584 #endif /* CONFIG_HIBERNATION */
3586 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3587 not required for correctness. So if the last cpu in a node goes
3588 away, we get changed to run anywhere: as the first one comes back,
3589 restore their cpu bindings. */
3590 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3595 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3596 for_each_node_state(nid
, N_MEMORY
) {
3597 pg_data_t
*pgdat
= NODE_DATA(nid
);
3598 const struct cpumask
*mask
;
3600 mask
= cpumask_of_node(pgdat
->node_id
);
3602 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3603 /* One of our CPUs online: restore mask */
3604 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3611 * This kswapd start function will be called by init and node-hot-add.
3612 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3614 int kswapd_run(int nid
)
3616 pg_data_t
*pgdat
= NODE_DATA(nid
);
3622 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3623 if (IS_ERR(pgdat
->kswapd
)) {
3624 /* failure at boot is fatal */
3625 BUG_ON(system_state
== SYSTEM_BOOTING
);
3626 pr_err("Failed to start kswapd on node %d\n", nid
);
3627 ret
= PTR_ERR(pgdat
->kswapd
);
3628 pgdat
->kswapd
= NULL
;
3634 * Called by memory hotplug when all memory in a node is offlined. Caller must
3635 * hold mem_hotplug_begin/end().
3637 void kswapd_stop(int nid
)
3639 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3642 kthread_stop(kswapd
);
3643 NODE_DATA(nid
)->kswapd
= NULL
;
3647 static int __init
kswapd_init(void)
3652 for_each_node_state(nid
, N_MEMORY
)
3654 hotcpu_notifier(cpu_callback
, 0);
3658 module_init(kswapd_init
)
3664 * If non-zero call zone_reclaim when the number of free pages falls below
3667 int zone_reclaim_mode __read_mostly
;
3669 #define RECLAIM_OFF 0
3670 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3671 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3672 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3675 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3676 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3679 #define ZONE_RECLAIM_PRIORITY 4
3682 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3685 int sysctl_min_unmapped_ratio
= 1;
3688 * If the number of slab pages in a zone grows beyond this percentage then
3689 * slab reclaim needs to occur.
3691 int sysctl_min_slab_ratio
= 5;
3693 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3695 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3696 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3697 zone_page_state(zone
, NR_ACTIVE_FILE
);
3700 * It's possible for there to be more file mapped pages than
3701 * accounted for by the pages on the file LRU lists because
3702 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3704 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3707 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3708 static unsigned long zone_pagecache_reclaimable(struct zone
*zone
)
3710 unsigned long nr_pagecache_reclaimable
;
3711 unsigned long delta
= 0;
3714 * If RECLAIM_UNMAP is set, then all file pages are considered
3715 * potentially reclaimable. Otherwise, we have to worry about
3716 * pages like swapcache and zone_unmapped_file_pages() provides
3719 if (zone_reclaim_mode
& RECLAIM_UNMAP
)
3720 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3722 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3724 /* If we can't clean pages, remove dirty pages from consideration */
3725 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3726 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3728 /* Watch for any possible underflows due to delta */
3729 if (unlikely(delta
> nr_pagecache_reclaimable
))
3730 delta
= nr_pagecache_reclaimable
;
3732 return nr_pagecache_reclaimable
- delta
;
3736 * Try to free up some pages from this zone through reclaim.
3738 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3740 /* Minimum pages needed in order to stay on node */
3741 const unsigned long nr_pages
= 1 << order
;
3742 struct task_struct
*p
= current
;
3743 struct reclaim_state reclaim_state
;
3744 struct scan_control sc
= {
3745 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3746 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3748 .priority
= ZONE_RECLAIM_PRIORITY
,
3749 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3750 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_UNMAP
),
3756 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3757 * and we also need to be able to write out pages for RECLAIM_WRITE
3758 * and RECLAIM_UNMAP.
3760 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3761 lockdep_set_current_reclaim_state(gfp_mask
);
3762 reclaim_state
.reclaimed_slab
= 0;
3763 p
->reclaim_state
= &reclaim_state
;
3765 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3767 * Free memory by calling shrink zone with increasing
3768 * priorities until we have enough memory freed.
3771 shrink_zone(zone
, &sc
, true);
3772 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3775 p
->reclaim_state
= NULL
;
3776 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3777 lockdep_clear_current_reclaim_state();
3778 return sc
.nr_reclaimed
>= nr_pages
;
3781 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3787 * Zone reclaim reclaims unmapped file backed pages and
3788 * slab pages if we are over the defined limits.
3790 * A small portion of unmapped file backed pages is needed for
3791 * file I/O otherwise pages read by file I/O will be immediately
3792 * thrown out if the zone is overallocated. So we do not reclaim
3793 * if less than a specified percentage of the zone is used by
3794 * unmapped file backed pages.
3796 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3797 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3798 return ZONE_RECLAIM_FULL
;
3800 if (!zone_reclaimable(zone
))
3801 return ZONE_RECLAIM_FULL
;
3804 * Do not scan if the allocation should not be delayed.
3806 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3807 return ZONE_RECLAIM_NOSCAN
;
3810 * Only run zone reclaim on the local zone or on zones that do not
3811 * have associated processors. This will favor the local processor
3812 * over remote processors and spread off node memory allocations
3813 * as wide as possible.
3815 node_id
= zone_to_nid(zone
);
3816 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3817 return ZONE_RECLAIM_NOSCAN
;
3819 if (test_and_set_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
))
3820 return ZONE_RECLAIM_NOSCAN
;
3822 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3823 clear_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
);
3826 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3833 * page_evictable - test whether a page is evictable
3834 * @page: the page to test
3836 * Test whether page is evictable--i.e., should be placed on active/inactive
3837 * lists vs unevictable list.
3839 * Reasons page might not be evictable:
3840 * (1) page's mapping marked unevictable
3841 * (2) page is part of an mlocked VMA
3844 int page_evictable(struct page
*page
)
3846 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3851 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3852 * @pages: array of pages to check
3853 * @nr_pages: number of pages to check
3855 * Checks pages for evictability and moves them to the appropriate lru list.
3857 * This function is only used for SysV IPC SHM_UNLOCK.
3859 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3861 struct lruvec
*lruvec
;
3862 struct zone
*zone
= NULL
;
3867 for (i
= 0; i
< nr_pages
; i
++) {
3868 struct page
*page
= pages
[i
];
3869 struct zone
*pagezone
;
3872 pagezone
= page_zone(page
);
3873 if (pagezone
!= zone
) {
3875 spin_unlock_irq(&zone
->lru_lock
);
3877 spin_lock_irq(&zone
->lru_lock
);
3879 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3881 if (!PageLRU(page
) || !PageUnevictable(page
))
3884 if (page_evictable(page
)) {
3885 enum lru_list lru
= page_lru_base_type(page
);
3887 VM_BUG_ON_PAGE(PageActive(page
), page
);
3888 ClearPageUnevictable(page
);
3889 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3890 add_page_to_lru_list(page
, lruvec
, lru
);
3896 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3897 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3898 spin_unlock_irq(&zone
->lru_lock
);
3901 #endif /* CONFIG_SHMEM */