1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim
;
68 /* This context's GFP mask */
71 /* Allocation order */
75 * Nodemask of nodes allowed by the caller. If NULL, all nodes
81 * The memory cgroup that hit its limit and as a result is the
82 * primary target of this reclaim invocation.
84 struct mem_cgroup
*target_mem_cgroup
;
86 /* Scan (total_size >> priority) pages at once */
89 /* The highest zone to isolate pages for reclaim from */
90 enum zone_type reclaim_idx
;
92 /* Writepage batching in laptop mode; RECLAIM_WRITE */
93 unsigned int may_writepage
:1;
95 /* Can mapped pages be reclaimed? */
96 unsigned int may_unmap
:1;
98 /* Can pages be swapped as part of reclaim? */
99 unsigned int may_swap
:1;
102 * Cgroups are not reclaimed below their configured memory.low,
103 * unless we threaten to OOM. If any cgroups are skipped due to
104 * memory.low and nothing was reclaimed, go back for memory.low.
106 unsigned int memcg_low_reclaim
:1;
107 unsigned int memcg_low_skipped
:1;
109 unsigned int hibernation_mode
:1;
111 /* One of the zones is ready for compaction */
112 unsigned int compaction_ready
:1;
114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned
;
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed
;
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field) \
124 if ((_page)->lru.prev != _base) { \
127 prev = lru_to_page(&(_page->lru)); \
128 prefetch(&prev->_field); \
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field) \
138 if ((_page)->lru.prev != _base) { \
141 prev = lru_to_page(&(_page->lru)); \
142 prefetchw(&prev->_field); \
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
150 * From 0 .. 100. Higher means more swappy.
152 int vm_swappiness
= 60;
154 * The total number of pages which are beyond the high watermark within all
157 unsigned long vm_total_pages
;
159 static LIST_HEAD(shrinker_list
);
160 static DECLARE_RWSEM(shrinker_rwsem
);
163 static bool global_reclaim(struct scan_control
*sc
)
165 return !sc
->target_mem_cgroup
;
169 * sane_reclaim - is the usual dirty throttling mechanism operational?
170 * @sc: scan_control in question
172 * The normal page dirty throttling mechanism in balance_dirty_pages() is
173 * completely broken with the legacy memcg and direct stalling in
174 * shrink_page_list() is used for throttling instead, which lacks all the
175 * niceties such as fairness, adaptive pausing, bandwidth proportional
176 * allocation and configurability.
178 * This function tests whether the vmscan currently in progress can assume
179 * that the normal dirty throttling mechanism is operational.
181 static bool sane_reclaim(struct scan_control
*sc
)
183 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
187 #ifdef CONFIG_CGROUP_WRITEBACK
188 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
194 static bool global_reclaim(struct scan_control
*sc
)
199 static bool sane_reclaim(struct scan_control
*sc
)
206 * This misses isolated pages which are not accounted for to save counters.
207 * As the data only determines if reclaim or compaction continues, it is
208 * not expected that isolated pages will be a dominating factor.
210 unsigned long zone_reclaimable_pages(struct zone
*zone
)
214 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
215 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
216 if (get_nr_swap_pages() > 0)
217 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
218 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
223 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
227 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
228 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
229 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
231 if (get_nr_swap_pages() > 0)
232 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
233 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
234 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
240 * lruvec_lru_size - Returns the number of pages on the given LRU list.
241 * @lruvec: lru vector
243 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
245 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
247 unsigned long lru_size
;
250 if (!mem_cgroup_disabled())
251 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
253 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
255 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
256 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
259 if (!managed_zone(zone
))
262 if (!mem_cgroup_disabled())
263 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
265 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
266 NR_ZONE_LRU_BASE
+ lru
);
267 lru_size
-= min(size
, lru_size
);
275 * Add a shrinker callback to be called from the vm.
277 int register_shrinker(struct shrinker
*shrinker
)
279 size_t size
= sizeof(*shrinker
->nr_deferred
);
281 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
284 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
285 if (!shrinker
->nr_deferred
)
288 down_write(&shrinker_rwsem
);
289 list_add_tail(&shrinker
->list
, &shrinker_list
);
290 up_write(&shrinker_rwsem
);
293 EXPORT_SYMBOL(register_shrinker
);
298 void unregister_shrinker(struct shrinker
*shrinker
)
300 if (!shrinker
->nr_deferred
)
302 down_write(&shrinker_rwsem
);
303 list_del(&shrinker
->list
);
304 up_write(&shrinker_rwsem
);
305 kfree(shrinker
->nr_deferred
);
306 shrinker
->nr_deferred
= NULL
;
308 EXPORT_SYMBOL(unregister_shrinker
);
310 #define SHRINK_BATCH 128
312 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
313 struct shrinker
*shrinker
,
314 unsigned long nr_scanned
,
315 unsigned long nr_eligible
)
317 unsigned long freed
= 0;
318 unsigned long long delta
;
323 int nid
= shrinkctl
->nid
;
324 long batch_size
= shrinker
->batch
? shrinker
->batch
326 long scanned
= 0, next_deferred
;
328 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
333 * copy the current shrinker scan count into a local variable
334 * and zero it so that other concurrent shrinker invocations
335 * don't also do this scanning work.
337 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
340 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
342 do_div(delta
, nr_eligible
+ 1);
344 if (total_scan
< 0) {
345 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
346 shrinker
->scan_objects
, total_scan
);
347 total_scan
= freeable
;
350 next_deferred
= total_scan
;
353 * We need to avoid excessive windup on filesystem shrinkers
354 * due to large numbers of GFP_NOFS allocations causing the
355 * shrinkers to return -1 all the time. This results in a large
356 * nr being built up so when a shrink that can do some work
357 * comes along it empties the entire cache due to nr >>>
358 * freeable. This is bad for sustaining a working set in
361 * Hence only allow the shrinker to scan the entire cache when
362 * a large delta change is calculated directly.
364 if (delta
< freeable
/ 4)
365 total_scan
= min(total_scan
, freeable
/ 2);
368 * Avoid risking looping forever due to too large nr value:
369 * never try to free more than twice the estimate number of
372 if (total_scan
> freeable
* 2)
373 total_scan
= freeable
* 2;
375 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
376 nr_scanned
, nr_eligible
,
377 freeable
, delta
, total_scan
);
380 * Normally, we should not scan less than batch_size objects in one
381 * pass to avoid too frequent shrinker calls, but if the slab has less
382 * than batch_size objects in total and we are really tight on memory,
383 * we will try to reclaim all available objects, otherwise we can end
384 * up failing allocations although there are plenty of reclaimable
385 * objects spread over several slabs with usage less than the
388 * We detect the "tight on memory" situations by looking at the total
389 * number of objects we want to scan (total_scan). If it is greater
390 * than the total number of objects on slab (freeable), we must be
391 * scanning at high prio and therefore should try to reclaim as much as
394 while (total_scan
>= batch_size
||
395 total_scan
>= freeable
) {
397 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
399 shrinkctl
->nr_to_scan
= nr_to_scan
;
400 shrinkctl
->nr_scanned
= nr_to_scan
;
401 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
402 if (ret
== SHRINK_STOP
)
406 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
407 total_scan
-= shrinkctl
->nr_scanned
;
408 scanned
+= shrinkctl
->nr_scanned
;
413 if (next_deferred
>= scanned
)
414 next_deferred
-= scanned
;
418 * move the unused scan count back into the shrinker in a
419 * manner that handles concurrent updates. If we exhausted the
420 * scan, there is no need to do an update.
422 if (next_deferred
> 0)
423 new_nr
= atomic_long_add_return(next_deferred
,
424 &shrinker
->nr_deferred
[nid
]);
426 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
428 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
433 * shrink_slab - shrink slab caches
434 * @gfp_mask: allocation context
435 * @nid: node whose slab caches to target
436 * @memcg: memory cgroup whose slab caches to target
437 * @nr_scanned: pressure numerator
438 * @nr_eligible: pressure denominator
440 * Call the shrink functions to age shrinkable caches.
442 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
443 * unaware shrinkers will receive a node id of 0 instead.
445 * @memcg specifies the memory cgroup to target. If it is not NULL,
446 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
447 * objects from the memory cgroup specified. Otherwise, only unaware
448 * shrinkers are called.
450 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
451 * the available objects should be scanned. Page reclaim for example
452 * passes the number of pages scanned and the number of pages on the
453 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
454 * when it encountered mapped pages. The ratio is further biased by
455 * the ->seeks setting of the shrink function, which indicates the
456 * cost to recreate an object relative to that of an LRU page.
458 * Returns the number of reclaimed slab objects.
460 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
461 struct mem_cgroup
*memcg
,
462 unsigned long nr_scanned
,
463 unsigned long nr_eligible
)
465 struct shrinker
*shrinker
;
466 unsigned long freed
= 0;
468 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
472 nr_scanned
= SWAP_CLUSTER_MAX
;
474 if (!down_read_trylock(&shrinker_rwsem
)) {
476 * If we would return 0, our callers would understand that we
477 * have nothing else to shrink and give up trying. By returning
478 * 1 we keep it going and assume we'll be able to shrink next
485 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
486 struct shrink_control sc
= {
487 .gfp_mask
= gfp_mask
,
493 * If kernel memory accounting is disabled, we ignore
494 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
495 * passing NULL for memcg.
497 if (memcg_kmem_enabled() &&
498 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
501 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
504 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
507 up_read(&shrinker_rwsem
);
513 void drop_slab_node(int nid
)
518 struct mem_cgroup
*memcg
= NULL
;
522 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
524 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
525 } while (freed
> 10);
532 for_each_online_node(nid
)
536 static inline int is_page_cache_freeable(struct page
*page
)
539 * A freeable page cache page is referenced only by the caller
540 * that isolated the page, the page cache radix tree and
541 * optional buffer heads at page->private.
543 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
545 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
548 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
550 if (current
->flags
& PF_SWAPWRITE
)
552 if (!inode_write_congested(inode
))
554 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
560 * We detected a synchronous write error writing a page out. Probably
561 * -ENOSPC. We need to propagate that into the address_space for a subsequent
562 * fsync(), msync() or close().
564 * The tricky part is that after writepage we cannot touch the mapping: nothing
565 * prevents it from being freed up. But we have a ref on the page and once
566 * that page is locked, the mapping is pinned.
568 * We're allowed to run sleeping lock_page() here because we know the caller has
571 static void handle_write_error(struct address_space
*mapping
,
572 struct page
*page
, int error
)
575 if (page_mapping(page
) == mapping
)
576 mapping_set_error(mapping
, error
);
580 /* possible outcome of pageout() */
582 /* failed to write page out, page is locked */
584 /* move page to the active list, page is locked */
586 /* page has been sent to the disk successfully, page is unlocked */
588 /* page is clean and locked */
593 * pageout is called by shrink_page_list() for each dirty page.
594 * Calls ->writepage().
596 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
597 struct scan_control
*sc
)
600 * If the page is dirty, only perform writeback if that write
601 * will be non-blocking. To prevent this allocation from being
602 * stalled by pagecache activity. But note that there may be
603 * stalls if we need to run get_block(). We could test
604 * PagePrivate for that.
606 * If this process is currently in __generic_file_write_iter() against
607 * this page's queue, we can perform writeback even if that
610 * If the page is swapcache, write it back even if that would
611 * block, for some throttling. This happens by accident, because
612 * swap_backing_dev_info is bust: it doesn't reflect the
613 * congestion state of the swapdevs. Easy to fix, if needed.
615 if (!is_page_cache_freeable(page
))
619 * Some data journaling orphaned pages can have
620 * page->mapping == NULL while being dirty with clean buffers.
622 if (page_has_private(page
)) {
623 if (try_to_free_buffers(page
)) {
624 ClearPageDirty(page
);
625 pr_info("%s: orphaned page\n", __func__
);
631 if (mapping
->a_ops
->writepage
== NULL
)
632 return PAGE_ACTIVATE
;
633 if (!may_write_to_inode(mapping
->host
, sc
))
636 if (clear_page_dirty_for_io(page
)) {
638 struct writeback_control wbc
= {
639 .sync_mode
= WB_SYNC_NONE
,
640 .nr_to_write
= SWAP_CLUSTER_MAX
,
642 .range_end
= LLONG_MAX
,
646 SetPageReclaim(page
);
647 res
= mapping
->a_ops
->writepage(page
, &wbc
);
649 handle_write_error(mapping
, page
, res
);
650 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
651 ClearPageReclaim(page
);
652 return PAGE_ACTIVATE
;
655 if (!PageWriteback(page
)) {
656 /* synchronous write or broken a_ops? */
657 ClearPageReclaim(page
);
659 trace_mm_vmscan_writepage(page
);
660 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
668 * Same as remove_mapping, but if the page is removed from the mapping, it
669 * gets returned with a refcount of 0.
671 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
677 BUG_ON(!PageLocked(page
));
678 BUG_ON(mapping
!= page_mapping(page
));
680 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
682 * The non racy check for a busy page.
684 * Must be careful with the order of the tests. When someone has
685 * a ref to the page, it may be possible that they dirty it then
686 * drop the reference. So if PageDirty is tested before page_count
687 * here, then the following race may occur:
689 * get_user_pages(&page);
690 * [user mapping goes away]
692 * !PageDirty(page) [good]
693 * SetPageDirty(page);
695 * !page_count(page) [good, discard it]
697 * [oops, our write_to data is lost]
699 * Reversing the order of the tests ensures such a situation cannot
700 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
701 * load is not satisfied before that of page->_refcount.
703 * Note that if SetPageDirty is always performed via set_page_dirty,
704 * and thus under tree_lock, then this ordering is not required.
706 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
707 refcount
= 1 + HPAGE_PMD_NR
;
710 if (!page_ref_freeze(page
, refcount
))
712 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
713 if (unlikely(PageDirty(page
))) {
714 page_ref_unfreeze(page
, refcount
);
718 if (PageSwapCache(page
)) {
719 swp_entry_t swap
= { .val
= page_private(page
) };
720 mem_cgroup_swapout(page
, swap
);
721 __delete_from_swap_cache(page
);
722 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
723 put_swap_page(page
, swap
);
725 void (*freepage
)(struct page
*);
728 freepage
= mapping
->a_ops
->freepage
;
730 * Remember a shadow entry for reclaimed file cache in
731 * order to detect refaults, thus thrashing, later on.
733 * But don't store shadows in an address space that is
734 * already exiting. This is not just an optizimation,
735 * inode reclaim needs to empty out the radix tree or
736 * the nodes are lost. Don't plant shadows behind its
739 * We also don't store shadows for DAX mappings because the
740 * only page cache pages found in these are zero pages
741 * covering holes, and because we don't want to mix DAX
742 * exceptional entries and shadow exceptional entries in the
745 if (reclaimed
&& page_is_file_cache(page
) &&
746 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
747 shadow
= workingset_eviction(mapping
, page
);
748 __delete_from_page_cache(page
, shadow
);
749 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
751 if (freepage
!= NULL
)
758 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
763 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
764 * someone else has a ref on the page, abort and return 0. If it was
765 * successfully detached, return 1. Assumes the caller has a single ref on
768 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
770 if (__remove_mapping(mapping
, page
, false)) {
772 * Unfreezing the refcount with 1 rather than 2 effectively
773 * drops the pagecache ref for us without requiring another
776 page_ref_unfreeze(page
, 1);
783 * putback_lru_page - put previously isolated page onto appropriate LRU list
784 * @page: page to be put back to appropriate lru list
786 * Add previously isolated @page to appropriate LRU list.
787 * Page may still be unevictable for other reasons.
789 * lru_lock must not be held, interrupts must be enabled.
791 void putback_lru_page(struct page
*page
)
794 int was_unevictable
= PageUnevictable(page
);
796 VM_BUG_ON_PAGE(PageLRU(page
), page
);
799 ClearPageUnevictable(page
);
801 if (page_evictable(page
)) {
803 * For evictable pages, we can use the cache.
804 * In event of a race, worst case is we end up with an
805 * unevictable page on [in]active list.
806 * We know how to handle that.
808 is_unevictable
= false;
812 * Put unevictable pages directly on zone's unevictable
815 is_unevictable
= true;
816 add_page_to_unevictable_list(page
);
818 * When racing with an mlock or AS_UNEVICTABLE clearing
819 * (page is unlocked) make sure that if the other thread
820 * does not observe our setting of PG_lru and fails
821 * isolation/check_move_unevictable_pages,
822 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
823 * the page back to the evictable list.
825 * The other side is TestClearPageMlocked() or shmem_lock().
831 * page's status can change while we move it among lru. If an evictable
832 * page is on unevictable list, it never be freed. To avoid that,
833 * check after we added it to the list, again.
835 if (is_unevictable
&& page_evictable(page
)) {
836 if (!isolate_lru_page(page
)) {
840 /* This means someone else dropped this page from LRU
841 * So, it will be freed or putback to LRU again. There is
842 * nothing to do here.
846 if (was_unevictable
&& !is_unevictable
)
847 count_vm_event(UNEVICTABLE_PGRESCUED
);
848 else if (!was_unevictable
&& is_unevictable
)
849 count_vm_event(UNEVICTABLE_PGCULLED
);
851 put_page(page
); /* drop ref from isolate */
854 enum page_references
{
856 PAGEREF_RECLAIM_CLEAN
,
861 static enum page_references
page_check_references(struct page
*page
,
862 struct scan_control
*sc
)
864 int referenced_ptes
, referenced_page
;
865 unsigned long vm_flags
;
867 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
869 referenced_page
= TestClearPageReferenced(page
);
872 * Mlock lost the isolation race with us. Let try_to_unmap()
873 * move the page to the unevictable list.
875 if (vm_flags
& VM_LOCKED
)
876 return PAGEREF_RECLAIM
;
878 if (referenced_ptes
) {
879 if (PageSwapBacked(page
))
880 return PAGEREF_ACTIVATE
;
882 * All mapped pages start out with page table
883 * references from the instantiating fault, so we need
884 * to look twice if a mapped file page is used more
887 * Mark it and spare it for another trip around the
888 * inactive list. Another page table reference will
889 * lead to its activation.
891 * Note: the mark is set for activated pages as well
892 * so that recently deactivated but used pages are
895 SetPageReferenced(page
);
897 if (referenced_page
|| referenced_ptes
> 1)
898 return PAGEREF_ACTIVATE
;
901 * Activate file-backed executable pages after first usage.
903 if (vm_flags
& VM_EXEC
)
904 return PAGEREF_ACTIVATE
;
909 /* Reclaim if clean, defer dirty pages to writeback */
910 if (referenced_page
&& !PageSwapBacked(page
))
911 return PAGEREF_RECLAIM_CLEAN
;
913 return PAGEREF_RECLAIM
;
916 /* Check if a page is dirty or under writeback */
917 static void page_check_dirty_writeback(struct page
*page
,
918 bool *dirty
, bool *writeback
)
920 struct address_space
*mapping
;
923 * Anonymous pages are not handled by flushers and must be written
924 * from reclaim context. Do not stall reclaim based on them
926 if (!page_is_file_cache(page
) ||
927 (PageAnon(page
) && !PageSwapBacked(page
))) {
933 /* By default assume that the page flags are accurate */
934 *dirty
= PageDirty(page
);
935 *writeback
= PageWriteback(page
);
937 /* Verify dirty/writeback state if the filesystem supports it */
938 if (!page_has_private(page
))
941 mapping
= page_mapping(page
);
942 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
943 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
946 struct reclaim_stat
{
948 unsigned nr_unqueued_dirty
;
949 unsigned nr_congested
;
950 unsigned nr_writeback
;
951 unsigned nr_immediate
;
952 unsigned nr_activate
;
953 unsigned nr_ref_keep
;
954 unsigned nr_unmap_fail
;
958 * shrink_page_list() returns the number of reclaimed pages
960 static unsigned long shrink_page_list(struct list_head
*page_list
,
961 struct pglist_data
*pgdat
,
962 struct scan_control
*sc
,
963 enum ttu_flags ttu_flags
,
964 struct reclaim_stat
*stat
,
967 LIST_HEAD(ret_pages
);
968 LIST_HEAD(free_pages
);
970 unsigned nr_unqueued_dirty
= 0;
971 unsigned nr_dirty
= 0;
972 unsigned nr_congested
= 0;
973 unsigned nr_reclaimed
= 0;
974 unsigned nr_writeback
= 0;
975 unsigned nr_immediate
= 0;
976 unsigned nr_ref_keep
= 0;
977 unsigned nr_unmap_fail
= 0;
981 while (!list_empty(page_list
)) {
982 struct address_space
*mapping
;
985 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
986 bool dirty
, writeback
;
990 page
= lru_to_page(page_list
);
991 list_del(&page
->lru
);
993 if (!trylock_page(page
))
996 VM_BUG_ON_PAGE(PageActive(page
), page
);
1000 if (unlikely(!page_evictable(page
)))
1001 goto activate_locked
;
1003 if (!sc
->may_unmap
&& page_mapped(page
))
1006 /* Double the slab pressure for mapped and swapcache pages */
1007 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1008 !(PageAnon(page
) && !PageSwapBacked(page
)))
1011 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1012 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1015 * The number of dirty pages determines if a zone is marked
1016 * reclaim_congested which affects wait_iff_congested. kswapd
1017 * will stall and start writing pages if the tail of the LRU
1018 * is all dirty unqueued pages.
1020 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1021 if (dirty
|| writeback
)
1024 if (dirty
&& !writeback
)
1025 nr_unqueued_dirty
++;
1028 * Treat this page as congested if the underlying BDI is or if
1029 * pages are cycling through the LRU so quickly that the
1030 * pages marked for immediate reclaim are making it to the
1031 * end of the LRU a second time.
1033 mapping
= page_mapping(page
);
1034 if (((dirty
|| writeback
) && mapping
&&
1035 inode_write_congested(mapping
->host
)) ||
1036 (writeback
&& PageReclaim(page
)))
1040 * If a page at the tail of the LRU is under writeback, there
1041 * are three cases to consider.
1043 * 1) If reclaim is encountering an excessive number of pages
1044 * under writeback and this page is both under writeback and
1045 * PageReclaim then it indicates that pages are being queued
1046 * for IO but are being recycled through the LRU before the
1047 * IO can complete. Waiting on the page itself risks an
1048 * indefinite stall if it is impossible to writeback the
1049 * page due to IO error or disconnected storage so instead
1050 * note that the LRU is being scanned too quickly and the
1051 * caller can stall after page list has been processed.
1053 * 2) Global or new memcg reclaim encounters a page that is
1054 * not marked for immediate reclaim, or the caller does not
1055 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1056 * not to fs). In this case mark the page for immediate
1057 * reclaim and continue scanning.
1059 * Require may_enter_fs because we would wait on fs, which
1060 * may not have submitted IO yet. And the loop driver might
1061 * enter reclaim, and deadlock if it waits on a page for
1062 * which it is needed to do the write (loop masks off
1063 * __GFP_IO|__GFP_FS for this reason); but more thought
1064 * would probably show more reasons.
1066 * 3) Legacy memcg encounters a page that is already marked
1067 * PageReclaim. memcg does not have any dirty pages
1068 * throttling so we could easily OOM just because too many
1069 * pages are in writeback and there is nothing else to
1070 * reclaim. Wait for the writeback to complete.
1072 * In cases 1) and 2) we activate the pages to get them out of
1073 * the way while we continue scanning for clean pages on the
1074 * inactive list and refilling from the active list. The
1075 * observation here is that waiting for disk writes is more
1076 * expensive than potentially causing reloads down the line.
1077 * Since they're marked for immediate reclaim, they won't put
1078 * memory pressure on the cache working set any longer than it
1079 * takes to write them to disk.
1081 if (PageWriteback(page
)) {
1083 if (current_is_kswapd() &&
1084 PageReclaim(page
) &&
1085 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1087 goto activate_locked
;
1090 } else if (sane_reclaim(sc
) ||
1091 !PageReclaim(page
) || !may_enter_fs
) {
1093 * This is slightly racy - end_page_writeback()
1094 * might have just cleared PageReclaim, then
1095 * setting PageReclaim here end up interpreted
1096 * as PageReadahead - but that does not matter
1097 * enough to care. What we do want is for this
1098 * page to have PageReclaim set next time memcg
1099 * reclaim reaches the tests above, so it will
1100 * then wait_on_page_writeback() to avoid OOM;
1101 * and it's also appropriate in global reclaim.
1103 SetPageReclaim(page
);
1105 goto activate_locked
;
1110 wait_on_page_writeback(page
);
1111 /* then go back and try same page again */
1112 list_add_tail(&page
->lru
, page_list
);
1118 references
= page_check_references(page
, sc
);
1120 switch (references
) {
1121 case PAGEREF_ACTIVATE
:
1122 goto activate_locked
;
1126 case PAGEREF_RECLAIM
:
1127 case PAGEREF_RECLAIM_CLEAN
:
1128 ; /* try to reclaim the page below */
1132 * Anonymous process memory has backing store?
1133 * Try to allocate it some swap space here.
1134 * Lazyfree page could be freed directly
1136 if (PageAnon(page
) && PageSwapBacked(page
)) {
1137 if (!PageSwapCache(page
)) {
1138 if (!(sc
->gfp_mask
& __GFP_IO
))
1140 if (PageTransHuge(page
)) {
1141 /* cannot split THP, skip it */
1142 if (!can_split_huge_page(page
, NULL
))
1143 goto activate_locked
;
1145 * Split pages without a PMD map right
1146 * away. Chances are some or all of the
1147 * tail pages can be freed without IO.
1149 if (!compound_mapcount(page
) &&
1150 split_huge_page_to_list(page
,
1152 goto activate_locked
;
1154 if (!add_to_swap(page
)) {
1155 if (!PageTransHuge(page
))
1156 goto activate_locked
;
1157 /* Fallback to swap normal pages */
1158 if (split_huge_page_to_list(page
,
1160 goto activate_locked
;
1161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1162 count_vm_event(THP_SWPOUT_FALLBACK
);
1164 if (!add_to_swap(page
))
1165 goto activate_locked
;
1170 /* Adding to swap updated mapping */
1171 mapping
= page_mapping(page
);
1173 } else if (unlikely(PageTransHuge(page
))) {
1174 /* Split file THP */
1175 if (split_huge_page_to_list(page
, page_list
))
1180 * The page is mapped into the page tables of one or more
1181 * processes. Try to unmap it here.
1183 if (page_mapped(page
)) {
1184 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1186 if (unlikely(PageTransHuge(page
)))
1187 flags
|= TTU_SPLIT_HUGE_PMD
;
1188 if (!try_to_unmap(page
, flags
)) {
1190 goto activate_locked
;
1194 if (PageDirty(page
)) {
1196 * Only kswapd can writeback filesystem pages
1197 * to avoid risk of stack overflow. But avoid
1198 * injecting inefficient single-page IO into
1199 * flusher writeback as much as possible: only
1200 * write pages when we've encountered many
1201 * dirty pages, and when we've already scanned
1202 * the rest of the LRU for clean pages and see
1203 * the same dirty pages again (PageReclaim).
1205 if (page_is_file_cache(page
) &&
1206 (!current_is_kswapd() || !PageReclaim(page
) ||
1207 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1209 * Immediately reclaim when written back.
1210 * Similar in principal to deactivate_page()
1211 * except we already have the page isolated
1212 * and know it's dirty
1214 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1215 SetPageReclaim(page
);
1217 goto activate_locked
;
1220 if (references
== PAGEREF_RECLAIM_CLEAN
)
1224 if (!sc
->may_writepage
)
1228 * Page is dirty. Flush the TLB if a writable entry
1229 * potentially exists to avoid CPU writes after IO
1230 * starts and then write it out here.
1232 try_to_unmap_flush_dirty();
1233 switch (pageout(page
, mapping
, sc
)) {
1237 goto activate_locked
;
1239 if (PageWriteback(page
))
1241 if (PageDirty(page
))
1245 * A synchronous write - probably a ramdisk. Go
1246 * ahead and try to reclaim the page.
1248 if (!trylock_page(page
))
1250 if (PageDirty(page
) || PageWriteback(page
))
1252 mapping
= page_mapping(page
);
1254 ; /* try to free the page below */
1259 * If the page has buffers, try to free the buffer mappings
1260 * associated with this page. If we succeed we try to free
1263 * We do this even if the page is PageDirty().
1264 * try_to_release_page() does not perform I/O, but it is
1265 * possible for a page to have PageDirty set, but it is actually
1266 * clean (all its buffers are clean). This happens if the
1267 * buffers were written out directly, with submit_bh(). ext3
1268 * will do this, as well as the blockdev mapping.
1269 * try_to_release_page() will discover that cleanness and will
1270 * drop the buffers and mark the page clean - it can be freed.
1272 * Rarely, pages can have buffers and no ->mapping. These are
1273 * the pages which were not successfully invalidated in
1274 * truncate_complete_page(). We try to drop those buffers here
1275 * and if that worked, and the page is no longer mapped into
1276 * process address space (page_count == 1) it can be freed.
1277 * Otherwise, leave the page on the LRU so it is swappable.
1279 if (page_has_private(page
)) {
1280 if (!try_to_release_page(page
, sc
->gfp_mask
))
1281 goto activate_locked
;
1282 if (!mapping
&& page_count(page
) == 1) {
1284 if (put_page_testzero(page
))
1288 * rare race with speculative reference.
1289 * the speculative reference will free
1290 * this page shortly, so we may
1291 * increment nr_reclaimed here (and
1292 * leave it off the LRU).
1300 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1301 /* follow __remove_mapping for reference */
1302 if (!page_ref_freeze(page
, 1))
1304 if (PageDirty(page
)) {
1305 page_ref_unfreeze(page
, 1);
1309 count_vm_event(PGLAZYFREED
);
1310 count_memcg_page_event(page
, PGLAZYFREED
);
1311 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1314 * At this point, we have no other references and there is
1315 * no way to pick any more up (removed from LRU, removed
1316 * from pagecache). Can use non-atomic bitops now (and
1317 * we obviously don't have to worry about waking up a process
1318 * waiting on the page lock, because there are no references.
1320 __ClearPageLocked(page
);
1325 * Is there need to periodically free_page_list? It would
1326 * appear not as the counts should be low
1328 if (unlikely(PageTransHuge(page
))) {
1329 mem_cgroup_uncharge(page
);
1330 (*get_compound_page_dtor(page
))(page
);
1332 list_add(&page
->lru
, &free_pages
);
1336 /* Not a candidate for swapping, so reclaim swap space. */
1337 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1339 try_to_free_swap(page
);
1340 VM_BUG_ON_PAGE(PageActive(page
), page
);
1341 if (!PageMlocked(page
)) {
1342 SetPageActive(page
);
1344 count_memcg_page_event(page
, PGACTIVATE
);
1349 list_add(&page
->lru
, &ret_pages
);
1350 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1353 mem_cgroup_uncharge_list(&free_pages
);
1354 try_to_unmap_flush();
1355 free_hot_cold_page_list(&free_pages
, true);
1357 list_splice(&ret_pages
, page_list
);
1358 count_vm_events(PGACTIVATE
, pgactivate
);
1361 stat
->nr_dirty
= nr_dirty
;
1362 stat
->nr_congested
= nr_congested
;
1363 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1364 stat
->nr_writeback
= nr_writeback
;
1365 stat
->nr_immediate
= nr_immediate
;
1366 stat
->nr_activate
= pgactivate
;
1367 stat
->nr_ref_keep
= nr_ref_keep
;
1368 stat
->nr_unmap_fail
= nr_unmap_fail
;
1370 return nr_reclaimed
;
1373 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1374 struct list_head
*page_list
)
1376 struct scan_control sc
= {
1377 .gfp_mask
= GFP_KERNEL
,
1378 .priority
= DEF_PRIORITY
,
1382 struct page
*page
, *next
;
1383 LIST_HEAD(clean_pages
);
1385 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1386 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1387 !__PageMovable(page
)) {
1388 ClearPageActive(page
);
1389 list_move(&page
->lru
, &clean_pages
);
1393 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1394 TTU_IGNORE_ACCESS
, NULL
, true);
1395 list_splice(&clean_pages
, page_list
);
1396 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1401 * Attempt to remove the specified page from its LRU. Only take this page
1402 * if it is of the appropriate PageActive status. Pages which are being
1403 * freed elsewhere are also ignored.
1405 * page: page to consider
1406 * mode: one of the LRU isolation modes defined above
1408 * returns 0 on success, -ve errno on failure.
1410 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1414 /* Only take pages on the LRU. */
1418 /* Compaction should not handle unevictable pages but CMA can do so */
1419 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1425 * To minimise LRU disruption, the caller can indicate that it only
1426 * wants to isolate pages it will be able to operate on without
1427 * blocking - clean pages for the most part.
1429 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1430 * that it is possible to migrate without blocking
1432 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1433 /* All the caller can do on PageWriteback is block */
1434 if (PageWriteback(page
))
1437 if (PageDirty(page
)) {
1438 struct address_space
*mapping
;
1442 * Only pages without mappings or that have a
1443 * ->migratepage callback are possible to migrate
1444 * without blocking. However, we can be racing with
1445 * truncation so it's necessary to lock the page
1446 * to stabilise the mapping as truncation holds
1447 * the page lock until after the page is removed
1448 * from the page cache.
1450 if (!trylock_page(page
))
1453 mapping
= page_mapping(page
);
1454 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1461 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1464 if (likely(get_page_unless_zero(page
))) {
1466 * Be careful not to clear PageLRU until after we're
1467 * sure the page is not being freed elsewhere -- the
1468 * page release code relies on it.
1479 * Update LRU sizes after isolating pages. The LRU size updates must
1480 * be complete before mem_cgroup_update_lru_size due to a santity check.
1482 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1483 enum lru_list lru
, unsigned long *nr_zone_taken
)
1487 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1488 if (!nr_zone_taken
[zid
])
1491 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1493 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1500 * zone_lru_lock is heavily contended. Some of the functions that
1501 * shrink the lists perform better by taking out a batch of pages
1502 * and working on them outside the LRU lock.
1504 * For pagecache intensive workloads, this function is the hottest
1505 * spot in the kernel (apart from copy_*_user functions).
1507 * Appropriate locks must be held before calling this function.
1509 * @nr_to_scan: The number of eligible pages to look through on the list.
1510 * @lruvec: The LRU vector to pull pages from.
1511 * @dst: The temp list to put pages on to.
1512 * @nr_scanned: The number of pages that were scanned.
1513 * @sc: The scan_control struct for this reclaim session
1514 * @mode: One of the LRU isolation modes
1515 * @lru: LRU list id for isolating
1517 * returns how many pages were moved onto *@dst.
1519 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1520 struct lruvec
*lruvec
, struct list_head
*dst
,
1521 unsigned long *nr_scanned
, struct scan_control
*sc
,
1522 isolate_mode_t mode
, enum lru_list lru
)
1524 struct list_head
*src
= &lruvec
->lists
[lru
];
1525 unsigned long nr_taken
= 0;
1526 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1527 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1528 unsigned long skipped
= 0;
1529 unsigned long scan
, total_scan
, nr_pages
;
1530 LIST_HEAD(pages_skipped
);
1533 for (total_scan
= 0;
1534 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1538 page
= lru_to_page(src
);
1539 prefetchw_prev_lru_page(page
, src
, flags
);
1541 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1543 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1544 list_move(&page
->lru
, &pages_skipped
);
1545 nr_skipped
[page_zonenum(page
)]++;
1550 * Do not count skipped pages because that makes the function
1551 * return with no isolated pages if the LRU mostly contains
1552 * ineligible pages. This causes the VM to not reclaim any
1553 * pages, triggering a premature OOM.
1556 switch (__isolate_lru_page(page
, mode
)) {
1558 nr_pages
= hpage_nr_pages(page
);
1559 nr_taken
+= nr_pages
;
1560 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1561 list_move(&page
->lru
, dst
);
1565 /* else it is being freed elsewhere */
1566 list_move(&page
->lru
, src
);
1575 * Splice any skipped pages to the start of the LRU list. Note that
1576 * this disrupts the LRU order when reclaiming for lower zones but
1577 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1578 * scanning would soon rescan the same pages to skip and put the
1579 * system at risk of premature OOM.
1581 if (!list_empty(&pages_skipped
)) {
1584 list_splice(&pages_skipped
, src
);
1585 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1586 if (!nr_skipped
[zid
])
1589 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1590 skipped
+= nr_skipped
[zid
];
1593 *nr_scanned
= total_scan
;
1594 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1595 total_scan
, skipped
, nr_taken
, mode
, lru
);
1596 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1601 * isolate_lru_page - tries to isolate a page from its LRU list
1602 * @page: page to isolate from its LRU list
1604 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1605 * vmstat statistic corresponding to whatever LRU list the page was on.
1607 * Returns 0 if the page was removed from an LRU list.
1608 * Returns -EBUSY if the page was not on an LRU list.
1610 * The returned page will have PageLRU() cleared. If it was found on
1611 * the active list, it will have PageActive set. If it was found on
1612 * the unevictable list, it will have the PageUnevictable bit set. That flag
1613 * may need to be cleared by the caller before letting the page go.
1615 * The vmstat statistic corresponding to the list on which the page was
1616 * found will be decremented.
1619 * (1) Must be called with an elevated refcount on the page. This is a
1620 * fundamentnal difference from isolate_lru_pages (which is called
1621 * without a stable reference).
1622 * (2) the lru_lock must not be held.
1623 * (3) interrupts must be enabled.
1625 int isolate_lru_page(struct page
*page
)
1629 VM_BUG_ON_PAGE(!page_count(page
), page
);
1630 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1632 if (PageLRU(page
)) {
1633 struct zone
*zone
= page_zone(page
);
1634 struct lruvec
*lruvec
;
1636 spin_lock_irq(zone_lru_lock(zone
));
1637 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1638 if (PageLRU(page
)) {
1639 int lru
= page_lru(page
);
1642 del_page_from_lru_list(page
, lruvec
, lru
);
1645 spin_unlock_irq(zone_lru_lock(zone
));
1651 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1652 * then get resheduled. When there are massive number of tasks doing page
1653 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1654 * the LRU list will go small and be scanned faster than necessary, leading to
1655 * unnecessary swapping, thrashing and OOM.
1657 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1658 struct scan_control
*sc
)
1660 unsigned long inactive
, isolated
;
1662 if (current_is_kswapd())
1665 if (!sane_reclaim(sc
))
1669 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1670 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1672 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1673 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1677 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1678 * won't get blocked by normal direct-reclaimers, forming a circular
1681 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1684 return isolated
> inactive
;
1687 static noinline_for_stack
void
1688 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1690 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1691 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1692 LIST_HEAD(pages_to_free
);
1695 * Put back any unfreeable pages.
1697 while (!list_empty(page_list
)) {
1698 struct page
*page
= lru_to_page(page_list
);
1701 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1702 list_del(&page
->lru
);
1703 if (unlikely(!page_evictable(page
))) {
1704 spin_unlock_irq(&pgdat
->lru_lock
);
1705 putback_lru_page(page
);
1706 spin_lock_irq(&pgdat
->lru_lock
);
1710 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1713 lru
= page_lru(page
);
1714 add_page_to_lru_list(page
, lruvec
, lru
);
1716 if (is_active_lru(lru
)) {
1717 int file
= is_file_lru(lru
);
1718 int numpages
= hpage_nr_pages(page
);
1719 reclaim_stat
->recent_rotated
[file
] += numpages
;
1721 if (put_page_testzero(page
)) {
1722 __ClearPageLRU(page
);
1723 __ClearPageActive(page
);
1724 del_page_from_lru_list(page
, lruvec
, lru
);
1726 if (unlikely(PageCompound(page
))) {
1727 spin_unlock_irq(&pgdat
->lru_lock
);
1728 mem_cgroup_uncharge(page
);
1729 (*get_compound_page_dtor(page
))(page
);
1730 spin_lock_irq(&pgdat
->lru_lock
);
1732 list_add(&page
->lru
, &pages_to_free
);
1737 * To save our caller's stack, now use input list for pages to free.
1739 list_splice(&pages_to_free
, page_list
);
1743 * If a kernel thread (such as nfsd for loop-back mounts) services
1744 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1745 * In that case we should only throttle if the backing device it is
1746 * writing to is congested. In other cases it is safe to throttle.
1748 static int current_may_throttle(void)
1750 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1751 current
->backing_dev_info
== NULL
||
1752 bdi_write_congested(current
->backing_dev_info
);
1756 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1757 * of reclaimed pages
1759 static noinline_for_stack
unsigned long
1760 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1761 struct scan_control
*sc
, enum lru_list lru
)
1763 LIST_HEAD(page_list
);
1764 unsigned long nr_scanned
;
1765 unsigned long nr_reclaimed
= 0;
1766 unsigned long nr_taken
;
1767 struct reclaim_stat stat
= {};
1768 isolate_mode_t isolate_mode
= 0;
1769 int file
= is_file_lru(lru
);
1770 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1771 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1772 bool stalled
= false;
1774 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1778 /* wait a bit for the reclaimer. */
1782 /* We are about to die and free our memory. Return now. */
1783 if (fatal_signal_pending(current
))
1784 return SWAP_CLUSTER_MAX
;
1790 isolate_mode
|= ISOLATE_UNMAPPED
;
1792 spin_lock_irq(&pgdat
->lru_lock
);
1794 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1795 &nr_scanned
, sc
, isolate_mode
, lru
);
1797 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1798 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1800 if (current_is_kswapd()) {
1801 if (global_reclaim(sc
))
1802 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1803 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1806 if (global_reclaim(sc
))
1807 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1808 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1811 spin_unlock_irq(&pgdat
->lru_lock
);
1816 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1819 spin_lock_irq(&pgdat
->lru_lock
);
1821 if (current_is_kswapd()) {
1822 if (global_reclaim(sc
))
1823 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1824 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1827 if (global_reclaim(sc
))
1828 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1829 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1833 putback_inactive_pages(lruvec
, &page_list
);
1835 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1837 spin_unlock_irq(&pgdat
->lru_lock
);
1839 mem_cgroup_uncharge_list(&page_list
);
1840 free_hot_cold_page_list(&page_list
, true);
1843 * If reclaim is isolating dirty pages under writeback, it implies
1844 * that the long-lived page allocation rate is exceeding the page
1845 * laundering rate. Either the global limits are not being effective
1846 * at throttling processes due to the page distribution throughout
1847 * zones or there is heavy usage of a slow backing device. The
1848 * only option is to throttle from reclaim context which is not ideal
1849 * as there is no guarantee the dirtying process is throttled in the
1850 * same way balance_dirty_pages() manages.
1852 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1853 * of pages under pages flagged for immediate reclaim and stall if any
1854 * are encountered in the nr_immediate check below.
1856 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1857 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1860 * If dirty pages are scanned that are not queued for IO, it
1861 * implies that flushers are not doing their job. This can
1862 * happen when memory pressure pushes dirty pages to the end of
1863 * the LRU before the dirty limits are breached and the dirty
1864 * data has expired. It can also happen when the proportion of
1865 * dirty pages grows not through writes but through memory
1866 * pressure reclaiming all the clean cache. And in some cases,
1867 * the flushers simply cannot keep up with the allocation
1868 * rate. Nudge the flusher threads in case they are asleep.
1870 if (stat
.nr_unqueued_dirty
== nr_taken
)
1871 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1874 * Legacy memcg will stall in page writeback so avoid forcibly
1877 if (sane_reclaim(sc
)) {
1879 * Tag a zone as congested if all the dirty pages scanned were
1880 * backed by a congested BDI and wait_iff_congested will stall.
1882 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1883 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1885 /* Allow kswapd to start writing pages during reclaim. */
1886 if (stat
.nr_unqueued_dirty
== nr_taken
)
1887 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1890 * If kswapd scans pages marked marked for immediate
1891 * reclaim and under writeback (nr_immediate), it implies
1892 * that pages are cycling through the LRU faster than
1893 * they are written so also forcibly stall.
1895 if (stat
.nr_immediate
&& current_may_throttle())
1896 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1900 * Stall direct reclaim for IO completions if underlying BDIs or zone
1901 * is congested. Allow kswapd to continue until it starts encountering
1902 * unqueued dirty pages or cycling through the LRU too quickly.
1904 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1905 current_may_throttle())
1906 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1908 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1909 nr_scanned
, nr_reclaimed
,
1910 stat
.nr_dirty
, stat
.nr_writeback
,
1911 stat
.nr_congested
, stat
.nr_immediate
,
1912 stat
.nr_activate
, stat
.nr_ref_keep
,
1914 sc
->priority
, file
);
1915 return nr_reclaimed
;
1919 * This moves pages from the active list to the inactive list.
1921 * We move them the other way if the page is referenced by one or more
1922 * processes, from rmap.
1924 * If the pages are mostly unmapped, the processing is fast and it is
1925 * appropriate to hold zone_lru_lock across the whole operation. But if
1926 * the pages are mapped, the processing is slow (page_referenced()) so we
1927 * should drop zone_lru_lock around each page. It's impossible to balance
1928 * this, so instead we remove the pages from the LRU while processing them.
1929 * It is safe to rely on PG_active against the non-LRU pages in here because
1930 * nobody will play with that bit on a non-LRU page.
1932 * The downside is that we have to touch page->_refcount against each page.
1933 * But we had to alter page->flags anyway.
1935 * Returns the number of pages moved to the given lru.
1938 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1939 struct list_head
*list
,
1940 struct list_head
*pages_to_free
,
1943 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1948 while (!list_empty(list
)) {
1949 page
= lru_to_page(list
);
1950 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1952 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1955 nr_pages
= hpage_nr_pages(page
);
1956 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1957 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1959 if (put_page_testzero(page
)) {
1960 __ClearPageLRU(page
);
1961 __ClearPageActive(page
);
1962 del_page_from_lru_list(page
, lruvec
, lru
);
1964 if (unlikely(PageCompound(page
))) {
1965 spin_unlock_irq(&pgdat
->lru_lock
);
1966 mem_cgroup_uncharge(page
);
1967 (*get_compound_page_dtor(page
))(page
);
1968 spin_lock_irq(&pgdat
->lru_lock
);
1970 list_add(&page
->lru
, pages_to_free
);
1972 nr_moved
+= nr_pages
;
1976 if (!is_active_lru(lru
)) {
1977 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1978 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
1985 static void shrink_active_list(unsigned long nr_to_scan
,
1986 struct lruvec
*lruvec
,
1987 struct scan_control
*sc
,
1990 unsigned long nr_taken
;
1991 unsigned long nr_scanned
;
1992 unsigned long vm_flags
;
1993 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1994 LIST_HEAD(l_active
);
1995 LIST_HEAD(l_inactive
);
1997 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1998 unsigned nr_deactivate
, nr_activate
;
1999 unsigned nr_rotated
= 0;
2000 isolate_mode_t isolate_mode
= 0;
2001 int file
= is_file_lru(lru
);
2002 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2007 isolate_mode
|= ISOLATE_UNMAPPED
;
2009 spin_lock_irq(&pgdat
->lru_lock
);
2011 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2012 &nr_scanned
, sc
, isolate_mode
, lru
);
2014 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2015 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2017 __count_vm_events(PGREFILL
, nr_scanned
);
2018 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2020 spin_unlock_irq(&pgdat
->lru_lock
);
2022 while (!list_empty(&l_hold
)) {
2024 page
= lru_to_page(&l_hold
);
2025 list_del(&page
->lru
);
2027 if (unlikely(!page_evictable(page
))) {
2028 putback_lru_page(page
);
2032 if (unlikely(buffer_heads_over_limit
)) {
2033 if (page_has_private(page
) && trylock_page(page
)) {
2034 if (page_has_private(page
))
2035 try_to_release_page(page
, 0);
2040 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2042 nr_rotated
+= hpage_nr_pages(page
);
2044 * Identify referenced, file-backed active pages and
2045 * give them one more trip around the active list. So
2046 * that executable code get better chances to stay in
2047 * memory under moderate memory pressure. Anon pages
2048 * are not likely to be evicted by use-once streaming
2049 * IO, plus JVM can create lots of anon VM_EXEC pages,
2050 * so we ignore them here.
2052 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2053 list_add(&page
->lru
, &l_active
);
2058 ClearPageActive(page
); /* we are de-activating */
2059 list_add(&page
->lru
, &l_inactive
);
2063 * Move pages back to the lru list.
2065 spin_lock_irq(&pgdat
->lru_lock
);
2067 * Count referenced pages from currently used mappings as rotated,
2068 * even though only some of them are actually re-activated. This
2069 * helps balance scan pressure between file and anonymous pages in
2072 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2074 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2075 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2076 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2077 spin_unlock_irq(&pgdat
->lru_lock
);
2079 mem_cgroup_uncharge_list(&l_hold
);
2080 free_hot_cold_page_list(&l_hold
, true);
2081 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2082 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2086 * The inactive anon list should be small enough that the VM never has
2087 * to do too much work.
2089 * The inactive file list should be small enough to leave most memory
2090 * to the established workingset on the scan-resistant active list,
2091 * but large enough to avoid thrashing the aggregate readahead window.
2093 * Both inactive lists should also be large enough that each inactive
2094 * page has a chance to be referenced again before it is reclaimed.
2096 * If that fails and refaulting is observed, the inactive list grows.
2098 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2099 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2100 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2103 * memory ratio inactive
2104 * -------------------------------------
2113 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2114 struct mem_cgroup
*memcg
,
2115 struct scan_control
*sc
, bool actual_reclaim
)
2117 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2118 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2119 enum lru_list inactive_lru
= file
* LRU_FILE
;
2120 unsigned long inactive
, active
;
2121 unsigned long inactive_ratio
;
2122 unsigned long refaults
;
2126 * If we don't have swap space, anonymous page deactivation
2129 if (!file
&& !total_swap_pages
)
2132 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2133 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2136 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2138 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2141 * When refaults are being observed, it means a new workingset
2142 * is being established. Disable active list protection to get
2143 * rid of the stale workingset quickly.
2145 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2148 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2150 inactive_ratio
= int_sqrt(10 * gb
);
2156 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2157 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2158 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2159 inactive_ratio
, file
);
2161 return inactive
* inactive_ratio
< active
;
2164 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2165 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2166 struct scan_control
*sc
)
2168 if (is_active_lru(lru
)) {
2169 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2171 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2175 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2186 * Determine how aggressively the anon and file LRU lists should be
2187 * scanned. The relative value of each set of LRU lists is determined
2188 * by looking at the fraction of the pages scanned we did rotate back
2189 * onto the active list instead of evict.
2191 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2192 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2194 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2195 struct scan_control
*sc
, unsigned long *nr
,
2196 unsigned long *lru_pages
)
2198 int swappiness
= mem_cgroup_swappiness(memcg
);
2199 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2201 u64 denominator
= 0; /* gcc */
2202 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2203 unsigned long anon_prio
, file_prio
;
2204 enum scan_balance scan_balance
;
2205 unsigned long anon
, file
;
2206 unsigned long ap
, fp
;
2209 /* If we have no swap space, do not bother scanning anon pages. */
2210 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2211 scan_balance
= SCAN_FILE
;
2216 * Global reclaim will swap to prevent OOM even with no
2217 * swappiness, but memcg users want to use this knob to
2218 * disable swapping for individual groups completely when
2219 * using the memory controller's swap limit feature would be
2222 if (!global_reclaim(sc
) && !swappiness
) {
2223 scan_balance
= SCAN_FILE
;
2228 * Do not apply any pressure balancing cleverness when the
2229 * system is close to OOM, scan both anon and file equally
2230 * (unless the swappiness setting disagrees with swapping).
2232 if (!sc
->priority
&& swappiness
) {
2233 scan_balance
= SCAN_EQUAL
;
2238 * Prevent the reclaimer from falling into the cache trap: as
2239 * cache pages start out inactive, every cache fault will tip
2240 * the scan balance towards the file LRU. And as the file LRU
2241 * shrinks, so does the window for rotation from references.
2242 * This means we have a runaway feedback loop where a tiny
2243 * thrashing file LRU becomes infinitely more attractive than
2244 * anon pages. Try to detect this based on file LRU size.
2246 if (global_reclaim(sc
)) {
2247 unsigned long pgdatfile
;
2248 unsigned long pgdatfree
;
2250 unsigned long total_high_wmark
= 0;
2252 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2253 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2254 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2256 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2257 struct zone
*zone
= &pgdat
->node_zones
[z
];
2258 if (!managed_zone(zone
))
2261 total_high_wmark
+= high_wmark_pages(zone
);
2264 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2266 * Force SCAN_ANON if there are enough inactive
2267 * anonymous pages on the LRU in eligible zones.
2268 * Otherwise, the small LRU gets thrashed.
2270 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2271 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2273 scan_balance
= SCAN_ANON
;
2280 * If there is enough inactive page cache, i.e. if the size of the
2281 * inactive list is greater than that of the active list *and* the
2282 * inactive list actually has some pages to scan on this priority, we
2283 * do not reclaim anything from the anonymous working set right now.
2284 * Without the second condition we could end up never scanning an
2285 * lruvec even if it has plenty of old anonymous pages unless the
2286 * system is under heavy pressure.
2288 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2289 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2290 scan_balance
= SCAN_FILE
;
2294 scan_balance
= SCAN_FRACT
;
2297 * With swappiness at 100, anonymous and file have the same priority.
2298 * This scanning priority is essentially the inverse of IO cost.
2300 anon_prio
= swappiness
;
2301 file_prio
= 200 - anon_prio
;
2304 * OK, so we have swap space and a fair amount of page cache
2305 * pages. We use the recently rotated / recently scanned
2306 * ratios to determine how valuable each cache is.
2308 * Because workloads change over time (and to avoid overflow)
2309 * we keep these statistics as a floating average, which ends
2310 * up weighing recent references more than old ones.
2312 * anon in [0], file in [1]
2315 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2316 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2317 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2318 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2320 spin_lock_irq(&pgdat
->lru_lock
);
2321 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2322 reclaim_stat
->recent_scanned
[0] /= 2;
2323 reclaim_stat
->recent_rotated
[0] /= 2;
2326 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2327 reclaim_stat
->recent_scanned
[1] /= 2;
2328 reclaim_stat
->recent_rotated
[1] /= 2;
2332 * The amount of pressure on anon vs file pages is inversely
2333 * proportional to the fraction of recently scanned pages on
2334 * each list that were recently referenced and in active use.
2336 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2337 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2339 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2340 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2341 spin_unlock_irq(&pgdat
->lru_lock
);
2345 denominator
= ap
+ fp
+ 1;
2348 for_each_evictable_lru(lru
) {
2349 int file
= is_file_lru(lru
);
2353 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2354 scan
= size
>> sc
->priority
;
2356 * If the cgroup's already been deleted, make sure to
2357 * scrape out the remaining cache.
2359 if (!scan
&& !mem_cgroup_online(memcg
))
2360 scan
= min(size
, SWAP_CLUSTER_MAX
);
2362 switch (scan_balance
) {
2364 /* Scan lists relative to size */
2368 * Scan types proportional to swappiness and
2369 * their relative recent reclaim efficiency.
2371 scan
= div64_u64(scan
* fraction
[file
],
2376 /* Scan one type exclusively */
2377 if ((scan_balance
== SCAN_FILE
) != file
) {
2383 /* Look ma, no brain */
2393 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2395 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2396 struct scan_control
*sc
, unsigned long *lru_pages
)
2398 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2399 unsigned long nr
[NR_LRU_LISTS
];
2400 unsigned long targets
[NR_LRU_LISTS
];
2401 unsigned long nr_to_scan
;
2403 unsigned long nr_reclaimed
= 0;
2404 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2405 struct blk_plug plug
;
2408 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2410 /* Record the original scan target for proportional adjustments later */
2411 memcpy(targets
, nr
, sizeof(nr
));
2414 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2415 * event that can occur when there is little memory pressure e.g.
2416 * multiple streaming readers/writers. Hence, we do not abort scanning
2417 * when the requested number of pages are reclaimed when scanning at
2418 * DEF_PRIORITY on the assumption that the fact we are direct
2419 * reclaiming implies that kswapd is not keeping up and it is best to
2420 * do a batch of work at once. For memcg reclaim one check is made to
2421 * abort proportional reclaim if either the file or anon lru has already
2422 * dropped to zero at the first pass.
2424 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2425 sc
->priority
== DEF_PRIORITY
);
2427 blk_start_plug(&plug
);
2428 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2429 nr
[LRU_INACTIVE_FILE
]) {
2430 unsigned long nr_anon
, nr_file
, percentage
;
2431 unsigned long nr_scanned
;
2433 for_each_evictable_lru(lru
) {
2435 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2436 nr
[lru
] -= nr_to_scan
;
2438 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2445 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2449 * For kswapd and memcg, reclaim at least the number of pages
2450 * requested. Ensure that the anon and file LRUs are scanned
2451 * proportionally what was requested by get_scan_count(). We
2452 * stop reclaiming one LRU and reduce the amount scanning
2453 * proportional to the original scan target.
2455 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2456 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2459 * It's just vindictive to attack the larger once the smaller
2460 * has gone to zero. And given the way we stop scanning the
2461 * smaller below, this makes sure that we only make one nudge
2462 * towards proportionality once we've got nr_to_reclaim.
2464 if (!nr_file
|| !nr_anon
)
2467 if (nr_file
> nr_anon
) {
2468 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2469 targets
[LRU_ACTIVE_ANON
] + 1;
2471 percentage
= nr_anon
* 100 / scan_target
;
2473 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2474 targets
[LRU_ACTIVE_FILE
] + 1;
2476 percentage
= nr_file
* 100 / scan_target
;
2479 /* Stop scanning the smaller of the LRU */
2481 nr
[lru
+ LRU_ACTIVE
] = 0;
2484 * Recalculate the other LRU scan count based on its original
2485 * scan target and the percentage scanning already complete
2487 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2488 nr_scanned
= targets
[lru
] - nr
[lru
];
2489 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2490 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2493 nr_scanned
= targets
[lru
] - nr
[lru
];
2494 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2495 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2497 scan_adjusted
= true;
2499 blk_finish_plug(&plug
);
2500 sc
->nr_reclaimed
+= nr_reclaimed
;
2503 * Even if we did not try to evict anon pages at all, we want to
2504 * rebalance the anon lru active/inactive ratio.
2506 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2507 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2508 sc
, LRU_ACTIVE_ANON
);
2511 /* Use reclaim/compaction for costly allocs or under memory pressure */
2512 static bool in_reclaim_compaction(struct scan_control
*sc
)
2514 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2515 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2516 sc
->priority
< DEF_PRIORITY
- 2))
2523 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2524 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2525 * true if more pages should be reclaimed such that when the page allocator
2526 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2527 * It will give up earlier than that if there is difficulty reclaiming pages.
2529 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2530 unsigned long nr_reclaimed
,
2531 unsigned long nr_scanned
,
2532 struct scan_control
*sc
)
2534 unsigned long pages_for_compaction
;
2535 unsigned long inactive_lru_pages
;
2538 /* If not in reclaim/compaction mode, stop */
2539 if (!in_reclaim_compaction(sc
))
2542 /* Consider stopping depending on scan and reclaim activity */
2543 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2545 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2546 * full LRU list has been scanned and we are still failing
2547 * to reclaim pages. This full LRU scan is potentially
2548 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2550 if (!nr_reclaimed
&& !nr_scanned
)
2554 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2555 * fail without consequence, stop if we failed to reclaim
2556 * any pages from the last SWAP_CLUSTER_MAX number of
2557 * pages that were scanned. This will return to the
2558 * caller faster at the risk reclaim/compaction and
2559 * the resulting allocation attempt fails
2566 * If we have not reclaimed enough pages for compaction and the
2567 * inactive lists are large enough, continue reclaiming
2569 pages_for_compaction
= compact_gap(sc
->order
);
2570 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2571 if (get_nr_swap_pages() > 0)
2572 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2573 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2574 inactive_lru_pages
> pages_for_compaction
)
2577 /* If compaction would go ahead or the allocation would succeed, stop */
2578 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2579 struct zone
*zone
= &pgdat
->node_zones
[z
];
2580 if (!managed_zone(zone
))
2583 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2584 case COMPACT_SUCCESS
:
2585 case COMPACT_CONTINUE
:
2588 /* check next zone */
2595 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2597 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2598 unsigned long nr_reclaimed
, nr_scanned
;
2599 bool reclaimable
= false;
2602 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2603 struct mem_cgroup_reclaim_cookie reclaim
= {
2605 .priority
= sc
->priority
,
2607 unsigned long node_lru_pages
= 0;
2608 struct mem_cgroup
*memcg
;
2610 nr_reclaimed
= sc
->nr_reclaimed
;
2611 nr_scanned
= sc
->nr_scanned
;
2613 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2615 unsigned long lru_pages
;
2616 unsigned long reclaimed
;
2617 unsigned long scanned
;
2619 if (mem_cgroup_low(root
, memcg
)) {
2620 if (!sc
->memcg_low_reclaim
) {
2621 sc
->memcg_low_skipped
= 1;
2624 mem_cgroup_event(memcg
, MEMCG_LOW
);
2627 reclaimed
= sc
->nr_reclaimed
;
2628 scanned
= sc
->nr_scanned
;
2630 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2631 node_lru_pages
+= lru_pages
;
2634 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2635 memcg
, sc
->nr_scanned
- scanned
,
2638 /* Record the group's reclaim efficiency */
2639 vmpressure(sc
->gfp_mask
, memcg
, false,
2640 sc
->nr_scanned
- scanned
,
2641 sc
->nr_reclaimed
- reclaimed
);
2644 * Direct reclaim and kswapd have to scan all memory
2645 * cgroups to fulfill the overall scan target for the
2648 * Limit reclaim, on the other hand, only cares about
2649 * nr_to_reclaim pages to be reclaimed and it will
2650 * retry with decreasing priority if one round over the
2651 * whole hierarchy is not sufficient.
2653 if (!global_reclaim(sc
) &&
2654 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2655 mem_cgroup_iter_break(root
, memcg
);
2658 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2661 * Shrink the slab caches in the same proportion that
2662 * the eligible LRU pages were scanned.
2664 if (global_reclaim(sc
))
2665 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2666 sc
->nr_scanned
- nr_scanned
,
2669 if (reclaim_state
) {
2670 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2671 reclaim_state
->reclaimed_slab
= 0;
2674 /* Record the subtree's reclaim efficiency */
2675 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2676 sc
->nr_scanned
- nr_scanned
,
2677 sc
->nr_reclaimed
- nr_reclaimed
);
2679 if (sc
->nr_reclaimed
- nr_reclaimed
)
2682 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2683 sc
->nr_scanned
- nr_scanned
, sc
));
2686 * Kswapd gives up on balancing particular nodes after too
2687 * many failures to reclaim anything from them and goes to
2688 * sleep. On reclaim progress, reset the failure counter. A
2689 * successful direct reclaim run will revive a dormant kswapd.
2692 pgdat
->kswapd_failures
= 0;
2698 * Returns true if compaction should go ahead for a costly-order request, or
2699 * the allocation would already succeed without compaction. Return false if we
2700 * should reclaim first.
2702 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2704 unsigned long watermark
;
2705 enum compact_result suitable
;
2707 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2708 if (suitable
== COMPACT_SUCCESS
)
2709 /* Allocation should succeed already. Don't reclaim. */
2711 if (suitable
== COMPACT_SKIPPED
)
2712 /* Compaction cannot yet proceed. Do reclaim. */
2716 * Compaction is already possible, but it takes time to run and there
2717 * are potentially other callers using the pages just freed. So proceed
2718 * with reclaim to make a buffer of free pages available to give
2719 * compaction a reasonable chance of completing and allocating the page.
2720 * Note that we won't actually reclaim the whole buffer in one attempt
2721 * as the target watermark in should_continue_reclaim() is lower. But if
2722 * we are already above the high+gap watermark, don't reclaim at all.
2724 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2726 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2730 * This is the direct reclaim path, for page-allocating processes. We only
2731 * try to reclaim pages from zones which will satisfy the caller's allocation
2734 * If a zone is deemed to be full of pinned pages then just give it a light
2735 * scan then give up on it.
2737 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2741 unsigned long nr_soft_reclaimed
;
2742 unsigned long nr_soft_scanned
;
2744 pg_data_t
*last_pgdat
= NULL
;
2747 * If the number of buffer_heads in the machine exceeds the maximum
2748 * allowed level, force direct reclaim to scan the highmem zone as
2749 * highmem pages could be pinning lowmem pages storing buffer_heads
2751 orig_mask
= sc
->gfp_mask
;
2752 if (buffer_heads_over_limit
) {
2753 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2754 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2757 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2758 sc
->reclaim_idx
, sc
->nodemask
) {
2760 * Take care memory controller reclaiming has small influence
2763 if (global_reclaim(sc
)) {
2764 if (!cpuset_zone_allowed(zone
,
2765 GFP_KERNEL
| __GFP_HARDWALL
))
2769 * If we already have plenty of memory free for
2770 * compaction in this zone, don't free any more.
2771 * Even though compaction is invoked for any
2772 * non-zero order, only frequent costly order
2773 * reclamation is disruptive enough to become a
2774 * noticeable problem, like transparent huge
2777 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2778 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2779 compaction_ready(zone
, sc
)) {
2780 sc
->compaction_ready
= true;
2785 * Shrink each node in the zonelist once. If the
2786 * zonelist is ordered by zone (not the default) then a
2787 * node may be shrunk multiple times but in that case
2788 * the user prefers lower zones being preserved.
2790 if (zone
->zone_pgdat
== last_pgdat
)
2794 * This steals pages from memory cgroups over softlimit
2795 * and returns the number of reclaimed pages and
2796 * scanned pages. This works for global memory pressure
2797 * and balancing, not for a memcg's limit.
2799 nr_soft_scanned
= 0;
2800 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2801 sc
->order
, sc
->gfp_mask
,
2803 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2804 sc
->nr_scanned
+= nr_soft_scanned
;
2805 /* need some check for avoid more shrink_zone() */
2808 /* See comment about same check for global reclaim above */
2809 if (zone
->zone_pgdat
== last_pgdat
)
2811 last_pgdat
= zone
->zone_pgdat
;
2812 shrink_node(zone
->zone_pgdat
, sc
);
2816 * Restore to original mask to avoid the impact on the caller if we
2817 * promoted it to __GFP_HIGHMEM.
2819 sc
->gfp_mask
= orig_mask
;
2822 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2824 struct mem_cgroup
*memcg
;
2826 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2828 unsigned long refaults
;
2829 struct lruvec
*lruvec
;
2832 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2834 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2836 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2837 lruvec
->refaults
= refaults
;
2838 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2842 * This is the main entry point to direct page reclaim.
2844 * If a full scan of the inactive list fails to free enough memory then we
2845 * are "out of memory" and something needs to be killed.
2847 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2848 * high - the zone may be full of dirty or under-writeback pages, which this
2849 * caller can't do much about. We kick the writeback threads and take explicit
2850 * naps in the hope that some of these pages can be written. But if the
2851 * allocating task holds filesystem locks which prevent writeout this might not
2852 * work, and the allocation attempt will fail.
2854 * returns: 0, if no pages reclaimed
2855 * else, the number of pages reclaimed
2857 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2858 struct scan_control
*sc
)
2860 int initial_priority
= sc
->priority
;
2861 pg_data_t
*last_pgdat
;
2865 delayacct_freepages_start();
2867 if (global_reclaim(sc
))
2868 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2871 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2874 shrink_zones(zonelist
, sc
);
2876 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2879 if (sc
->compaction_ready
)
2883 * If we're getting trouble reclaiming, start doing
2884 * writepage even in laptop mode.
2886 if (sc
->priority
< DEF_PRIORITY
- 2)
2887 sc
->may_writepage
= 1;
2888 } while (--sc
->priority
>= 0);
2891 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2893 if (zone
->zone_pgdat
== last_pgdat
)
2895 last_pgdat
= zone
->zone_pgdat
;
2896 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2899 delayacct_freepages_end();
2901 if (sc
->nr_reclaimed
)
2902 return sc
->nr_reclaimed
;
2904 /* Aborted reclaim to try compaction? don't OOM, then */
2905 if (sc
->compaction_ready
)
2908 /* Untapped cgroup reserves? Don't OOM, retry. */
2909 if (sc
->memcg_low_skipped
) {
2910 sc
->priority
= initial_priority
;
2911 sc
->memcg_low_reclaim
= 1;
2912 sc
->memcg_low_skipped
= 0;
2919 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2922 unsigned long pfmemalloc_reserve
= 0;
2923 unsigned long free_pages
= 0;
2927 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2930 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2931 zone
= &pgdat
->node_zones
[i
];
2932 if (!managed_zone(zone
))
2935 if (!zone_reclaimable_pages(zone
))
2938 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2939 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2942 /* If there are no reserves (unexpected config) then do not throttle */
2943 if (!pfmemalloc_reserve
)
2946 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2948 /* kswapd must be awake if processes are being throttled */
2949 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2950 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2951 (enum zone_type
)ZONE_NORMAL
);
2952 wake_up_interruptible(&pgdat
->kswapd_wait
);
2959 * Throttle direct reclaimers if backing storage is backed by the network
2960 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2961 * depleted. kswapd will continue to make progress and wake the processes
2962 * when the low watermark is reached.
2964 * Returns true if a fatal signal was delivered during throttling. If this
2965 * happens, the page allocator should not consider triggering the OOM killer.
2967 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2968 nodemask_t
*nodemask
)
2972 pg_data_t
*pgdat
= NULL
;
2975 * Kernel threads should not be throttled as they may be indirectly
2976 * responsible for cleaning pages necessary for reclaim to make forward
2977 * progress. kjournald for example may enter direct reclaim while
2978 * committing a transaction where throttling it could forcing other
2979 * processes to block on log_wait_commit().
2981 if (current
->flags
& PF_KTHREAD
)
2985 * If a fatal signal is pending, this process should not throttle.
2986 * It should return quickly so it can exit and free its memory
2988 if (fatal_signal_pending(current
))
2992 * Check if the pfmemalloc reserves are ok by finding the first node
2993 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2994 * GFP_KERNEL will be required for allocating network buffers when
2995 * swapping over the network so ZONE_HIGHMEM is unusable.
2997 * Throttling is based on the first usable node and throttled processes
2998 * wait on a queue until kswapd makes progress and wakes them. There
2999 * is an affinity then between processes waking up and where reclaim
3000 * progress has been made assuming the process wakes on the same node.
3001 * More importantly, processes running on remote nodes will not compete
3002 * for remote pfmemalloc reserves and processes on different nodes
3003 * should make reasonable progress.
3005 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3006 gfp_zone(gfp_mask
), nodemask
) {
3007 if (zone_idx(zone
) > ZONE_NORMAL
)
3010 /* Throttle based on the first usable node */
3011 pgdat
= zone
->zone_pgdat
;
3012 if (allow_direct_reclaim(pgdat
))
3017 /* If no zone was usable by the allocation flags then do not throttle */
3021 /* Account for the throttling */
3022 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3025 * If the caller cannot enter the filesystem, it's possible that it
3026 * is due to the caller holding an FS lock or performing a journal
3027 * transaction in the case of a filesystem like ext[3|4]. In this case,
3028 * it is not safe to block on pfmemalloc_wait as kswapd could be
3029 * blocked waiting on the same lock. Instead, throttle for up to a
3030 * second before continuing.
3032 if (!(gfp_mask
& __GFP_FS
)) {
3033 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3034 allow_direct_reclaim(pgdat
), HZ
);
3039 /* Throttle until kswapd wakes the process */
3040 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3041 allow_direct_reclaim(pgdat
));
3044 if (fatal_signal_pending(current
))
3051 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3052 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3054 unsigned long nr_reclaimed
;
3055 struct scan_control sc
= {
3056 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3057 .gfp_mask
= current_gfp_context(gfp_mask
),
3058 .reclaim_idx
= gfp_zone(gfp_mask
),
3060 .nodemask
= nodemask
,
3061 .priority
= DEF_PRIORITY
,
3062 .may_writepage
= !laptop_mode
,
3068 * Do not enter reclaim if fatal signal was delivered while throttled.
3069 * 1 is returned so that the page allocator does not OOM kill at this
3072 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3075 trace_mm_vmscan_direct_reclaim_begin(order
,
3080 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3082 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3084 return nr_reclaimed
;
3089 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3090 gfp_t gfp_mask
, bool noswap
,
3092 unsigned long *nr_scanned
)
3094 struct scan_control sc
= {
3095 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3096 .target_mem_cgroup
= memcg
,
3097 .may_writepage
= !laptop_mode
,
3099 .reclaim_idx
= MAX_NR_ZONES
- 1,
3100 .may_swap
= !noswap
,
3102 unsigned long lru_pages
;
3104 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3105 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3107 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3113 * NOTE: Although we can get the priority field, using it
3114 * here is not a good idea, since it limits the pages we can scan.
3115 * if we don't reclaim here, the shrink_node from balance_pgdat
3116 * will pick up pages from other mem cgroup's as well. We hack
3117 * the priority and make it zero.
3119 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3121 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3123 *nr_scanned
= sc
.nr_scanned
;
3124 return sc
.nr_reclaimed
;
3127 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3128 unsigned long nr_pages
,
3132 struct zonelist
*zonelist
;
3133 unsigned long nr_reclaimed
;
3135 unsigned int noreclaim_flag
;
3136 struct scan_control sc
= {
3137 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3138 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3139 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3140 .reclaim_idx
= MAX_NR_ZONES
- 1,
3141 .target_mem_cgroup
= memcg
,
3142 .priority
= DEF_PRIORITY
,
3143 .may_writepage
= !laptop_mode
,
3145 .may_swap
= may_swap
,
3149 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3150 * take care of from where we get pages. So the node where we start the
3151 * scan does not need to be the current node.
3153 nid
= mem_cgroup_select_victim_node(memcg
);
3155 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3157 trace_mm_vmscan_memcg_reclaim_begin(0,
3162 noreclaim_flag
= memalloc_noreclaim_save();
3163 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3164 memalloc_noreclaim_restore(noreclaim_flag
);
3166 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3168 return nr_reclaimed
;
3172 static void age_active_anon(struct pglist_data
*pgdat
,
3173 struct scan_control
*sc
)
3175 struct mem_cgroup
*memcg
;
3177 if (!total_swap_pages
)
3180 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3182 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3184 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3185 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3186 sc
, LRU_ACTIVE_ANON
);
3188 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3193 * Returns true if there is an eligible zone balanced for the request order
3196 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3199 unsigned long mark
= -1;
3202 for (i
= 0; i
<= classzone_idx
; i
++) {
3203 zone
= pgdat
->node_zones
+ i
;
3205 if (!managed_zone(zone
))
3208 mark
= high_wmark_pages(zone
);
3209 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3214 * If a node has no populated zone within classzone_idx, it does not
3215 * need balancing by definition. This can happen if a zone-restricted
3216 * allocation tries to wake a remote kswapd.
3224 /* Clear pgdat state for congested, dirty or under writeback. */
3225 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3227 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3228 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3229 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3233 * Prepare kswapd for sleeping. This verifies that there are no processes
3234 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3236 * Returns true if kswapd is ready to sleep
3238 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3241 * The throttled processes are normally woken up in balance_pgdat() as
3242 * soon as allow_direct_reclaim() is true. But there is a potential
3243 * race between when kswapd checks the watermarks and a process gets
3244 * throttled. There is also a potential race if processes get
3245 * throttled, kswapd wakes, a large process exits thereby balancing the
3246 * zones, which causes kswapd to exit balance_pgdat() before reaching
3247 * the wake up checks. If kswapd is going to sleep, no process should
3248 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3249 * the wake up is premature, processes will wake kswapd and get
3250 * throttled again. The difference from wake ups in balance_pgdat() is
3251 * that here we are under prepare_to_wait().
3253 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3254 wake_up_all(&pgdat
->pfmemalloc_wait
);
3256 /* Hopeless node, leave it to direct reclaim */
3257 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3260 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3261 clear_pgdat_congested(pgdat
);
3269 * kswapd shrinks a node of pages that are at or below the highest usable
3270 * zone that is currently unbalanced.
3272 * Returns true if kswapd scanned at least the requested number of pages to
3273 * reclaim or if the lack of progress was due to pages under writeback.
3274 * This is used to determine if the scanning priority needs to be raised.
3276 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3277 struct scan_control
*sc
)
3282 /* Reclaim a number of pages proportional to the number of zones */
3283 sc
->nr_to_reclaim
= 0;
3284 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3285 zone
= pgdat
->node_zones
+ z
;
3286 if (!managed_zone(zone
))
3289 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3293 * Historically care was taken to put equal pressure on all zones but
3294 * now pressure is applied based on node LRU order.
3296 shrink_node(pgdat
, sc
);
3299 * Fragmentation may mean that the system cannot be rebalanced for
3300 * high-order allocations. If twice the allocation size has been
3301 * reclaimed then recheck watermarks only at order-0 to prevent
3302 * excessive reclaim. Assume that a process requested a high-order
3303 * can direct reclaim/compact.
3305 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3308 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3312 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3313 * that are eligible for use by the caller until at least one zone is
3316 * Returns the order kswapd finished reclaiming at.
3318 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3319 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3320 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3321 * or lower is eligible for reclaim until at least one usable zone is
3324 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3327 unsigned long nr_soft_reclaimed
;
3328 unsigned long nr_soft_scanned
;
3330 struct scan_control sc
= {
3331 .gfp_mask
= GFP_KERNEL
,
3333 .priority
= DEF_PRIORITY
,
3334 .may_writepage
= !laptop_mode
,
3338 count_vm_event(PAGEOUTRUN
);
3341 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3342 bool raise_priority
= true;
3344 sc
.reclaim_idx
= classzone_idx
;
3347 * If the number of buffer_heads exceeds the maximum allowed
3348 * then consider reclaiming from all zones. This has a dual
3349 * purpose -- on 64-bit systems it is expected that
3350 * buffer_heads are stripped during active rotation. On 32-bit
3351 * systems, highmem pages can pin lowmem memory and shrinking
3352 * buffers can relieve lowmem pressure. Reclaim may still not
3353 * go ahead if all eligible zones for the original allocation
3354 * request are balanced to avoid excessive reclaim from kswapd.
3356 if (buffer_heads_over_limit
) {
3357 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3358 zone
= pgdat
->node_zones
+ i
;
3359 if (!managed_zone(zone
))
3368 * Only reclaim if there are no eligible zones. Note that
3369 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3372 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3376 * Do some background aging of the anon list, to give
3377 * pages a chance to be referenced before reclaiming. All
3378 * pages are rotated regardless of classzone as this is
3379 * about consistent aging.
3381 age_active_anon(pgdat
, &sc
);
3384 * If we're getting trouble reclaiming, start doing writepage
3385 * even in laptop mode.
3387 if (sc
.priority
< DEF_PRIORITY
- 2)
3388 sc
.may_writepage
= 1;
3390 /* Call soft limit reclaim before calling shrink_node. */
3392 nr_soft_scanned
= 0;
3393 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3394 sc
.gfp_mask
, &nr_soft_scanned
);
3395 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3398 * There should be no need to raise the scanning priority if
3399 * enough pages are already being scanned that that high
3400 * watermark would be met at 100% efficiency.
3402 if (kswapd_shrink_node(pgdat
, &sc
))
3403 raise_priority
= false;
3406 * If the low watermark is met there is no need for processes
3407 * to be throttled on pfmemalloc_wait as they should not be
3408 * able to safely make forward progress. Wake them
3410 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3411 allow_direct_reclaim(pgdat
))
3412 wake_up_all(&pgdat
->pfmemalloc_wait
);
3414 /* Check if kswapd should be suspending */
3415 if (try_to_freeze() || kthread_should_stop())
3419 * Raise priority if scanning rate is too low or there was no
3420 * progress in reclaiming pages
3422 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3423 if (raise_priority
|| !nr_reclaimed
)
3425 } while (sc
.priority
>= 1);
3427 if (!sc
.nr_reclaimed
)
3428 pgdat
->kswapd_failures
++;
3431 snapshot_refaults(NULL
, pgdat
);
3433 * Return the order kswapd stopped reclaiming at as
3434 * prepare_kswapd_sleep() takes it into account. If another caller
3435 * entered the allocator slow path while kswapd was awake, order will
3436 * remain at the higher level.
3442 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3443 * allocation request woke kswapd for. When kswapd has not woken recently,
3444 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3445 * given classzone and returns it or the highest classzone index kswapd
3446 * was recently woke for.
3448 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3449 enum zone_type classzone_idx
)
3451 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3452 return classzone_idx
;
3454 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3457 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3458 unsigned int classzone_idx
)
3463 if (freezing(current
) || kthread_should_stop())
3466 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3469 * Try to sleep for a short interval. Note that kcompactd will only be
3470 * woken if it is possible to sleep for a short interval. This is
3471 * deliberate on the assumption that if reclaim cannot keep an
3472 * eligible zone balanced that it's also unlikely that compaction will
3475 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3477 * Compaction records what page blocks it recently failed to
3478 * isolate pages from and skips them in the future scanning.
3479 * When kswapd is going to sleep, it is reasonable to assume
3480 * that pages and compaction may succeed so reset the cache.
3482 reset_isolation_suitable(pgdat
);
3485 * We have freed the memory, now we should compact it to make
3486 * allocation of the requested order possible.
3488 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3490 remaining
= schedule_timeout(HZ
/10);
3493 * If woken prematurely then reset kswapd_classzone_idx and
3494 * order. The values will either be from a wakeup request or
3495 * the previous request that slept prematurely.
3498 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3499 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3502 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3503 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3507 * After a short sleep, check if it was a premature sleep. If not, then
3508 * go fully to sleep until explicitly woken up.
3511 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3512 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3515 * vmstat counters are not perfectly accurate and the estimated
3516 * value for counters such as NR_FREE_PAGES can deviate from the
3517 * true value by nr_online_cpus * threshold. To avoid the zone
3518 * watermarks being breached while under pressure, we reduce the
3519 * per-cpu vmstat threshold while kswapd is awake and restore
3520 * them before going back to sleep.
3522 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3524 if (!kthread_should_stop())
3527 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3530 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3532 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3534 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3538 * The background pageout daemon, started as a kernel thread
3539 * from the init process.
3541 * This basically trickles out pages so that we have _some_
3542 * free memory available even if there is no other activity
3543 * that frees anything up. This is needed for things like routing
3544 * etc, where we otherwise might have all activity going on in
3545 * asynchronous contexts that cannot page things out.
3547 * If there are applications that are active memory-allocators
3548 * (most normal use), this basically shouldn't matter.
3550 static int kswapd(void *p
)
3552 unsigned int alloc_order
, reclaim_order
;
3553 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3554 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3555 struct task_struct
*tsk
= current
;
3557 struct reclaim_state reclaim_state
= {
3558 .reclaimed_slab
= 0,
3560 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3562 if (!cpumask_empty(cpumask
))
3563 set_cpus_allowed_ptr(tsk
, cpumask
);
3564 current
->reclaim_state
= &reclaim_state
;
3567 * Tell the memory management that we're a "memory allocator",
3568 * and that if we need more memory we should get access to it
3569 * regardless (see "__alloc_pages()"). "kswapd" should
3570 * never get caught in the normal page freeing logic.
3572 * (Kswapd normally doesn't need memory anyway, but sometimes
3573 * you need a small amount of memory in order to be able to
3574 * page out something else, and this flag essentially protects
3575 * us from recursively trying to free more memory as we're
3576 * trying to free the first piece of memory in the first place).
3578 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3581 pgdat
->kswapd_order
= 0;
3582 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3586 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3587 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3590 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3593 /* Read the new order and classzone_idx */
3594 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3595 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3596 pgdat
->kswapd_order
= 0;
3597 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3599 ret
= try_to_freeze();
3600 if (kthread_should_stop())
3604 * We can speed up thawing tasks if we don't call balance_pgdat
3605 * after returning from the refrigerator
3611 * Reclaim begins at the requested order but if a high-order
3612 * reclaim fails then kswapd falls back to reclaiming for
3613 * order-0. If that happens, kswapd will consider sleeping
3614 * for the order it finished reclaiming at (reclaim_order)
3615 * but kcompactd is woken to compact for the original
3616 * request (alloc_order).
3618 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3620 fs_reclaim_acquire(GFP_KERNEL
);
3621 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3622 fs_reclaim_release(GFP_KERNEL
);
3623 if (reclaim_order
< alloc_order
)
3624 goto kswapd_try_sleep
;
3627 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3628 current
->reclaim_state
= NULL
;
3634 * A zone is low on free memory, so wake its kswapd task to service it.
3636 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3640 if (!managed_zone(zone
))
3643 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3645 pgdat
= zone
->zone_pgdat
;
3646 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3648 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3649 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3652 /* Hopeless node, leave it to direct reclaim */
3653 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3656 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3659 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3660 wake_up_interruptible(&pgdat
->kswapd_wait
);
3663 #ifdef CONFIG_HIBERNATION
3665 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3668 * Rather than trying to age LRUs the aim is to preserve the overall
3669 * LRU order by reclaiming preferentially
3670 * inactive > active > active referenced > active mapped
3672 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3674 struct reclaim_state reclaim_state
;
3675 struct scan_control sc
= {
3676 .nr_to_reclaim
= nr_to_reclaim
,
3677 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3678 .reclaim_idx
= MAX_NR_ZONES
- 1,
3679 .priority
= DEF_PRIORITY
,
3683 .hibernation_mode
= 1,
3685 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3686 struct task_struct
*p
= current
;
3687 unsigned long nr_reclaimed
;
3688 unsigned int noreclaim_flag
;
3690 noreclaim_flag
= memalloc_noreclaim_save();
3691 fs_reclaim_acquire(sc
.gfp_mask
);
3692 reclaim_state
.reclaimed_slab
= 0;
3693 p
->reclaim_state
= &reclaim_state
;
3695 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3697 p
->reclaim_state
= NULL
;
3698 fs_reclaim_release(sc
.gfp_mask
);
3699 memalloc_noreclaim_restore(noreclaim_flag
);
3701 return nr_reclaimed
;
3703 #endif /* CONFIG_HIBERNATION */
3705 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3706 not required for correctness. So if the last cpu in a node goes
3707 away, we get changed to run anywhere: as the first one comes back,
3708 restore their cpu bindings. */
3709 static int kswapd_cpu_online(unsigned int cpu
)
3713 for_each_node_state(nid
, N_MEMORY
) {
3714 pg_data_t
*pgdat
= NODE_DATA(nid
);
3715 const struct cpumask
*mask
;
3717 mask
= cpumask_of_node(pgdat
->node_id
);
3719 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3720 /* One of our CPUs online: restore mask */
3721 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3727 * This kswapd start function will be called by init and node-hot-add.
3728 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3730 int kswapd_run(int nid
)
3732 pg_data_t
*pgdat
= NODE_DATA(nid
);
3738 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3739 if (IS_ERR(pgdat
->kswapd
)) {
3740 /* failure at boot is fatal */
3741 BUG_ON(system_state
< SYSTEM_RUNNING
);
3742 pr_err("Failed to start kswapd on node %d\n", nid
);
3743 ret
= PTR_ERR(pgdat
->kswapd
);
3744 pgdat
->kswapd
= NULL
;
3750 * Called by memory hotplug when all memory in a node is offlined. Caller must
3751 * hold mem_hotplug_begin/end().
3753 void kswapd_stop(int nid
)
3755 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3758 kthread_stop(kswapd
);
3759 NODE_DATA(nid
)->kswapd
= NULL
;
3763 static int __init
kswapd_init(void)
3768 for_each_node_state(nid
, N_MEMORY
)
3770 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3771 "mm/vmscan:online", kswapd_cpu_online
,
3777 module_init(kswapd_init
)
3783 * If non-zero call node_reclaim when the number of free pages falls below
3786 int node_reclaim_mode __read_mostly
;
3788 #define RECLAIM_OFF 0
3789 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3790 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3791 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3794 * Priority for NODE_RECLAIM. This determines the fraction of pages
3795 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3798 #define NODE_RECLAIM_PRIORITY 4
3801 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3804 int sysctl_min_unmapped_ratio
= 1;
3807 * If the number of slab pages in a zone grows beyond this percentage then
3808 * slab reclaim needs to occur.
3810 int sysctl_min_slab_ratio
= 5;
3812 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3814 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3815 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3816 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3819 * It's possible for there to be more file mapped pages than
3820 * accounted for by the pages on the file LRU lists because
3821 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3823 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3826 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3827 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3829 unsigned long nr_pagecache_reclaimable
;
3830 unsigned long delta
= 0;
3833 * If RECLAIM_UNMAP is set, then all file pages are considered
3834 * potentially reclaimable. Otherwise, we have to worry about
3835 * pages like swapcache and node_unmapped_file_pages() provides
3838 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3839 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3841 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3843 /* If we can't clean pages, remove dirty pages from consideration */
3844 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3845 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3847 /* Watch for any possible underflows due to delta */
3848 if (unlikely(delta
> nr_pagecache_reclaimable
))
3849 delta
= nr_pagecache_reclaimable
;
3851 return nr_pagecache_reclaimable
- delta
;
3855 * Try to free up some pages from this node through reclaim.
3857 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3859 /* Minimum pages needed in order to stay on node */
3860 const unsigned long nr_pages
= 1 << order
;
3861 struct task_struct
*p
= current
;
3862 struct reclaim_state reclaim_state
;
3863 unsigned int noreclaim_flag
;
3864 struct scan_control sc
= {
3865 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3866 .gfp_mask
= current_gfp_context(gfp_mask
),
3868 .priority
= NODE_RECLAIM_PRIORITY
,
3869 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3870 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3872 .reclaim_idx
= gfp_zone(gfp_mask
),
3877 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3878 * and we also need to be able to write out pages for RECLAIM_WRITE
3879 * and RECLAIM_UNMAP.
3881 noreclaim_flag
= memalloc_noreclaim_save();
3882 p
->flags
|= PF_SWAPWRITE
;
3883 fs_reclaim_acquire(sc
.gfp_mask
);
3884 reclaim_state
.reclaimed_slab
= 0;
3885 p
->reclaim_state
= &reclaim_state
;
3887 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3889 * Free memory by calling shrink zone with increasing
3890 * priorities until we have enough memory freed.
3893 shrink_node(pgdat
, &sc
);
3894 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3897 p
->reclaim_state
= NULL
;
3898 fs_reclaim_release(gfp_mask
);
3899 current
->flags
&= ~PF_SWAPWRITE
;
3900 memalloc_noreclaim_restore(noreclaim_flag
);
3901 return sc
.nr_reclaimed
>= nr_pages
;
3904 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3909 * Node reclaim reclaims unmapped file backed pages and
3910 * slab pages if we are over the defined limits.
3912 * A small portion of unmapped file backed pages is needed for
3913 * file I/O otherwise pages read by file I/O will be immediately
3914 * thrown out if the node is overallocated. So we do not reclaim
3915 * if less than a specified percentage of the node is used by
3916 * unmapped file backed pages.
3918 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3919 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3920 return NODE_RECLAIM_FULL
;
3923 * Do not scan if the allocation should not be delayed.
3925 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3926 return NODE_RECLAIM_NOSCAN
;
3929 * Only run node reclaim on the local node or on nodes that do not
3930 * have associated processors. This will favor the local processor
3931 * over remote processors and spread off node memory allocations
3932 * as wide as possible.
3934 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3935 return NODE_RECLAIM_NOSCAN
;
3937 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3938 return NODE_RECLAIM_NOSCAN
;
3940 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3941 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3944 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3951 * page_evictable - test whether a page is evictable
3952 * @page: the page to test
3954 * Test whether page is evictable--i.e., should be placed on active/inactive
3955 * lists vs unevictable list.
3957 * Reasons page might not be evictable:
3958 * (1) page's mapping marked unevictable
3959 * (2) page is part of an mlocked VMA
3962 int page_evictable(struct page
*page
)
3966 /* Prevent address_space of inode and swap cache from being freed */
3968 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3975 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3976 * @pages: array of pages to check
3977 * @nr_pages: number of pages to check
3979 * Checks pages for evictability and moves them to the appropriate lru list.
3981 * This function is only used for SysV IPC SHM_UNLOCK.
3983 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3985 struct lruvec
*lruvec
;
3986 struct pglist_data
*pgdat
= NULL
;
3991 for (i
= 0; i
< nr_pages
; i
++) {
3992 struct page
*page
= pages
[i
];
3993 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3996 if (pagepgdat
!= pgdat
) {
3998 spin_unlock_irq(&pgdat
->lru_lock
);
4000 spin_lock_irq(&pgdat
->lru_lock
);
4002 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4004 if (!PageLRU(page
) || !PageUnevictable(page
))
4007 if (page_evictable(page
)) {
4008 enum lru_list lru
= page_lru_base_type(page
);
4010 VM_BUG_ON_PAGE(PageActive(page
), page
);
4011 ClearPageUnevictable(page
);
4012 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4013 add_page_to_lru_list(page
, lruvec
, lru
);
4019 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4020 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4021 spin_unlock_irq(&pgdat
->lru_lock
);
4024 #endif /* CONFIG_SHMEM */