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/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim
;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup
*target_mem_cgroup
;
83 * Scan pressure balancing between anon and file LRUs
85 unsigned long anon_cost
;
86 unsigned long file_cost
;
88 /* Can active pages be deactivated as part of reclaim? */
89 #define DEACTIVATE_ANON 1
90 #define DEACTIVATE_FILE 2
91 unsigned int may_deactivate
:2;
92 unsigned int force_deactivate
:1;
93 unsigned int skipped_deactivate
:1;
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage
:1;
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap
:1;
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap
:1;
105 * Cgroups are not reclaimed below their configured memory.low,
106 * unless we threaten to OOM. If any cgroups are skipped due to
107 * memory.low and nothing was reclaimed, go back for memory.low.
109 unsigned int memcg_low_reclaim
:1;
110 unsigned int memcg_low_skipped
:1;
112 unsigned int hibernation_mode
:1;
114 /* One of the zones is ready for compaction */
115 unsigned int compaction_ready
:1;
117 /* There is easily reclaimable cold cache in the current node */
118 unsigned int cache_trim_mode
:1;
120 /* The file pages on the current node are dangerously low */
121 unsigned int file_is_tiny
:1;
123 /* Allocation order */
126 /* Scan (total_size >> priority) pages at once */
129 /* The highest zone to isolate pages for reclaim from */
132 /* This context's GFP mask */
135 /* Incremented by the number of inactive pages that were scanned */
136 unsigned long nr_scanned
;
138 /* Number of pages freed so far during a call to shrink_zones() */
139 unsigned long nr_reclaimed
;
143 unsigned int unqueued_dirty
;
144 unsigned int congested
;
145 unsigned int writeback
;
146 unsigned int immediate
;
147 unsigned int file_taken
;
151 /* for recording the reclaimed slab by now */
152 struct reclaim_state reclaim_state
;
155 #ifdef ARCH_HAS_PREFETCHW
156 #define prefetchw_prev_lru_page(_page, _base, _field) \
158 if ((_page)->lru.prev != _base) { \
161 prev = lru_to_page(&(_page->lru)); \
162 prefetchw(&prev->_field); \
166 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
170 * From 0 .. 200. Higher means more swappy.
172 int vm_swappiness
= 60;
174 static void set_task_reclaim_state(struct task_struct
*task
,
175 struct reclaim_state
*rs
)
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
183 task
->reclaim_state
= rs
;
186 static LIST_HEAD(shrinker_list
);
187 static DECLARE_RWSEM(shrinker_rwsem
);
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
201 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
203 static DEFINE_IDR(shrinker_idr
);
204 static int shrinker_nr_max
;
206 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
208 int id
, ret
= -ENOMEM
;
210 down_write(&shrinker_rwsem
);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
216 if (id
>= shrinker_nr_max
) {
217 if (memcg_expand_shrinker_maps(id
)) {
218 idr_remove(&shrinker_idr
, id
);
222 shrinker_nr_max
= id
+ 1;
227 up_write(&shrinker_rwsem
);
231 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
233 int id
= shrinker
->id
;
237 down_write(&shrinker_rwsem
);
238 idr_remove(&shrinker_idr
, id
);
239 up_write(&shrinker_rwsem
);
242 static bool cgroup_reclaim(struct scan_control
*sc
)
244 return sc
->target_mem_cgroup
;
248 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool writeback_throttling_sane(struct scan_control
*sc
)
262 if (!cgroup_reclaim(sc
))
264 #ifdef CONFIG_CGROUP_WRITEBACK
265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
271 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
276 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
280 static bool cgroup_reclaim(struct scan_control
*sc
)
285 static bool writeback_throttling_sane(struct scan_control
*sc
)
292 * This misses isolated pages which are not accounted for to save counters.
293 * As the data only determines if reclaim or compaction continues, it is
294 * not expected that isolated pages will be a dominating factor.
296 unsigned long zone_reclaimable_pages(struct zone
*zone
)
300 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
301 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
302 if (get_nr_swap_pages() > 0)
303 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
304 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
310 * lruvec_lru_size - Returns the number of pages on the given LRU list.
311 * @lruvec: lru vector
313 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
315 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
317 unsigned long size
= 0;
320 for (zid
= 0; zid
<= zone_idx
&& zid
< MAX_NR_ZONES
; zid
++) {
321 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
323 if (!managed_zone(zone
))
326 if (!mem_cgroup_disabled())
327 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
329 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
335 * Add a shrinker callback to be called from the vm.
337 int prealloc_shrinker(struct shrinker
*shrinker
)
339 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
341 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
344 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
345 if (!shrinker
->nr_deferred
)
348 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
349 if (prealloc_memcg_shrinker(shrinker
))
356 kfree(shrinker
->nr_deferred
);
357 shrinker
->nr_deferred
= NULL
;
361 void free_prealloced_shrinker(struct shrinker
*shrinker
)
363 if (!shrinker
->nr_deferred
)
366 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
367 unregister_memcg_shrinker(shrinker
);
369 kfree(shrinker
->nr_deferred
);
370 shrinker
->nr_deferred
= NULL
;
373 void register_shrinker_prepared(struct shrinker
*shrinker
)
375 down_write(&shrinker_rwsem
);
376 list_add_tail(&shrinker
->list
, &shrinker_list
);
378 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
379 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
381 up_write(&shrinker_rwsem
);
384 int register_shrinker(struct shrinker
*shrinker
)
386 int err
= prealloc_shrinker(shrinker
);
390 register_shrinker_prepared(shrinker
);
393 EXPORT_SYMBOL(register_shrinker
);
398 void unregister_shrinker(struct shrinker
*shrinker
)
400 if (!shrinker
->nr_deferred
)
402 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
403 unregister_memcg_shrinker(shrinker
);
404 down_write(&shrinker_rwsem
);
405 list_del(&shrinker
->list
);
406 up_write(&shrinker_rwsem
);
407 kfree(shrinker
->nr_deferred
);
408 shrinker
->nr_deferred
= NULL
;
410 EXPORT_SYMBOL(unregister_shrinker
);
412 #define SHRINK_BATCH 128
414 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
415 struct shrinker
*shrinker
, int priority
)
417 unsigned long freed
= 0;
418 unsigned long long delta
;
423 int nid
= shrinkctl
->nid
;
424 long batch_size
= shrinker
->batch
? shrinker
->batch
426 long scanned
= 0, next_deferred
;
428 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
431 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
432 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
436 * copy the current shrinker scan count into a local variable
437 * and zero it so that other concurrent shrinker invocations
438 * don't also do this scanning work.
440 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
443 if (shrinker
->seeks
) {
444 delta
= freeable
>> priority
;
446 do_div(delta
, shrinker
->seeks
);
449 * These objects don't require any IO to create. Trim
450 * them aggressively under memory pressure to keep
451 * them from causing refetches in the IO caches.
453 delta
= freeable
/ 2;
457 if (total_scan
< 0) {
458 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
459 shrinker
->scan_objects
, total_scan
);
460 total_scan
= freeable
;
463 next_deferred
= total_scan
;
466 * We need to avoid excessive windup on filesystem shrinkers
467 * due to large numbers of GFP_NOFS allocations causing the
468 * shrinkers to return -1 all the time. This results in a large
469 * nr being built up so when a shrink that can do some work
470 * comes along it empties the entire cache due to nr >>>
471 * freeable. This is bad for sustaining a working set in
474 * Hence only allow the shrinker to scan the entire cache when
475 * a large delta change is calculated directly.
477 if (delta
< freeable
/ 4)
478 total_scan
= min(total_scan
, freeable
/ 2);
481 * Avoid risking looping forever due to too large nr value:
482 * never try to free more than twice the estimate number of
485 if (total_scan
> freeable
* 2)
486 total_scan
= freeable
* 2;
488 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
489 freeable
, delta
, total_scan
, priority
);
492 * Normally, we should not scan less than batch_size objects in one
493 * pass to avoid too frequent shrinker calls, but if the slab has less
494 * than batch_size objects in total and we are really tight on memory,
495 * we will try to reclaim all available objects, otherwise we can end
496 * up failing allocations although there are plenty of reclaimable
497 * objects spread over several slabs with usage less than the
500 * We detect the "tight on memory" situations by looking at the total
501 * number of objects we want to scan (total_scan). If it is greater
502 * than the total number of objects on slab (freeable), we must be
503 * scanning at high prio and therefore should try to reclaim as much as
506 while (total_scan
>= batch_size
||
507 total_scan
>= freeable
) {
509 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
511 shrinkctl
->nr_to_scan
= nr_to_scan
;
512 shrinkctl
->nr_scanned
= nr_to_scan
;
513 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
514 if (ret
== SHRINK_STOP
)
518 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
519 total_scan
-= shrinkctl
->nr_scanned
;
520 scanned
+= shrinkctl
->nr_scanned
;
525 if (next_deferred
>= scanned
)
526 next_deferred
-= scanned
;
530 * move the unused scan count back into the shrinker in a
531 * manner that handles concurrent updates. If we exhausted the
532 * scan, there is no need to do an update.
534 if (next_deferred
> 0)
535 new_nr
= atomic_long_add_return(next_deferred
,
536 &shrinker
->nr_deferred
[nid
]);
538 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
540 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
545 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
546 struct mem_cgroup
*memcg
, int priority
)
548 struct memcg_shrinker_map
*map
;
549 unsigned long ret
, freed
= 0;
552 if (!mem_cgroup_online(memcg
))
555 if (!down_read_trylock(&shrinker_rwsem
))
558 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
563 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
564 struct shrink_control sc
= {
565 .gfp_mask
= gfp_mask
,
569 struct shrinker
*shrinker
;
571 shrinker
= idr_find(&shrinker_idr
, i
);
572 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
574 clear_bit(i
, map
->map
);
578 /* Call non-slab shrinkers even though kmem is disabled */
579 if (!memcg_kmem_enabled() &&
580 !(shrinker
->flags
& SHRINKER_NONSLAB
))
583 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
584 if (ret
== SHRINK_EMPTY
) {
585 clear_bit(i
, map
->map
);
587 * After the shrinker reported that it had no objects to
588 * free, but before we cleared the corresponding bit in
589 * the memcg shrinker map, a new object might have been
590 * added. To make sure, we have the bit set in this
591 * case, we invoke the shrinker one more time and reset
592 * the bit if it reports that it is not empty anymore.
593 * The memory barrier here pairs with the barrier in
594 * memcg_set_shrinker_bit():
596 * list_lru_add() shrink_slab_memcg()
597 * list_add_tail() clear_bit()
599 * set_bit() do_shrink_slab()
601 smp_mb__after_atomic();
602 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
603 if (ret
== SHRINK_EMPTY
)
606 memcg_set_shrinker_bit(memcg
, nid
, i
);
610 if (rwsem_is_contended(&shrinker_rwsem
)) {
616 up_read(&shrinker_rwsem
);
619 #else /* CONFIG_MEMCG */
620 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
621 struct mem_cgroup
*memcg
, int priority
)
625 #endif /* CONFIG_MEMCG */
628 * shrink_slab - shrink slab caches
629 * @gfp_mask: allocation context
630 * @nid: node whose slab caches to target
631 * @memcg: memory cgroup whose slab caches to target
632 * @priority: the reclaim priority
634 * Call the shrink functions to age shrinkable caches.
636 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
637 * unaware shrinkers will receive a node id of 0 instead.
639 * @memcg specifies the memory cgroup to target. Unaware shrinkers
640 * are called only if it is the root cgroup.
642 * @priority is sc->priority, we take the number of objects and >> by priority
643 * in order to get the scan target.
645 * Returns the number of reclaimed slab objects.
647 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
648 struct mem_cgroup
*memcg
,
651 unsigned long ret
, freed
= 0;
652 struct shrinker
*shrinker
;
655 * The root memcg might be allocated even though memcg is disabled
656 * via "cgroup_disable=memory" boot parameter. This could make
657 * mem_cgroup_is_root() return false, then just run memcg slab
658 * shrink, but skip global shrink. This may result in premature
661 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
662 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
664 if (!down_read_trylock(&shrinker_rwsem
))
667 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
668 struct shrink_control sc
= {
669 .gfp_mask
= gfp_mask
,
674 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
675 if (ret
== SHRINK_EMPTY
)
679 * Bail out if someone want to register a new shrinker to
680 * prevent the registration from being stalled for long periods
681 * by parallel ongoing shrinking.
683 if (rwsem_is_contended(&shrinker_rwsem
)) {
689 up_read(&shrinker_rwsem
);
695 void drop_slab_node(int nid
)
700 struct mem_cgroup
*memcg
= NULL
;
703 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
705 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
706 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
707 } while (freed
> 10);
714 for_each_online_node(nid
)
718 static inline int is_page_cache_freeable(struct page
*page
)
721 * A freeable page cache page is referenced only by the caller
722 * that isolated the page, the page cache and optional buffer
723 * heads at page->private.
725 int page_cache_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
727 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
730 static int may_write_to_inode(struct inode
*inode
)
732 if (current
->flags
& PF_SWAPWRITE
)
734 if (!inode_write_congested(inode
))
736 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
742 * We detected a synchronous write error writing a page out. Probably
743 * -ENOSPC. We need to propagate that into the address_space for a subsequent
744 * fsync(), msync() or close().
746 * The tricky part is that after writepage we cannot touch the mapping: nothing
747 * prevents it from being freed up. But we have a ref on the page and once
748 * that page is locked, the mapping is pinned.
750 * We're allowed to run sleeping lock_page() here because we know the caller has
753 static void handle_write_error(struct address_space
*mapping
,
754 struct page
*page
, int error
)
757 if (page_mapping(page
) == mapping
)
758 mapping_set_error(mapping
, error
);
762 /* possible outcome of pageout() */
764 /* failed to write page out, page is locked */
766 /* move page to the active list, page is locked */
768 /* page has been sent to the disk successfully, page is unlocked */
770 /* page is clean and locked */
775 * pageout is called by shrink_page_list() for each dirty page.
776 * Calls ->writepage().
778 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
781 * If the page is dirty, only perform writeback if that write
782 * will be non-blocking. To prevent this allocation from being
783 * stalled by pagecache activity. But note that there may be
784 * stalls if we need to run get_block(). We could test
785 * PagePrivate for that.
787 * If this process is currently in __generic_file_write_iter() against
788 * this page's queue, we can perform writeback even if that
791 * If the page is swapcache, write it back even if that would
792 * block, for some throttling. This happens by accident, because
793 * swap_backing_dev_info is bust: it doesn't reflect the
794 * congestion state of the swapdevs. Easy to fix, if needed.
796 if (!is_page_cache_freeable(page
))
800 * Some data journaling orphaned pages can have
801 * page->mapping == NULL while being dirty with clean buffers.
803 if (page_has_private(page
)) {
804 if (try_to_free_buffers(page
)) {
805 ClearPageDirty(page
);
806 pr_info("%s: orphaned page\n", __func__
);
812 if (mapping
->a_ops
->writepage
== NULL
)
813 return PAGE_ACTIVATE
;
814 if (!may_write_to_inode(mapping
->host
))
817 if (clear_page_dirty_for_io(page
)) {
819 struct writeback_control wbc
= {
820 .sync_mode
= WB_SYNC_NONE
,
821 .nr_to_write
= SWAP_CLUSTER_MAX
,
823 .range_end
= LLONG_MAX
,
827 SetPageReclaim(page
);
828 res
= mapping
->a_ops
->writepage(page
, &wbc
);
830 handle_write_error(mapping
, page
, res
);
831 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
832 ClearPageReclaim(page
);
833 return PAGE_ACTIVATE
;
836 if (!PageWriteback(page
)) {
837 /* synchronous write or broken a_ops? */
838 ClearPageReclaim(page
);
840 trace_mm_vmscan_writepage(page
);
841 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
849 * Same as remove_mapping, but if the page is removed from the mapping, it
850 * gets returned with a refcount of 0.
852 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
853 bool reclaimed
, struct mem_cgroup
*target_memcg
)
859 BUG_ON(!PageLocked(page
));
860 BUG_ON(mapping
!= page_mapping(page
));
862 xa_lock_irqsave(&mapping
->i_pages
, flags
);
864 * The non racy check for a busy page.
866 * Must be careful with the order of the tests. When someone has
867 * a ref to the page, it may be possible that they dirty it then
868 * drop the reference. So if PageDirty is tested before page_count
869 * here, then the following race may occur:
871 * get_user_pages(&page);
872 * [user mapping goes away]
874 * !PageDirty(page) [good]
875 * SetPageDirty(page);
877 * !page_count(page) [good, discard it]
879 * [oops, our write_to data is lost]
881 * Reversing the order of the tests ensures such a situation cannot
882 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
883 * load is not satisfied before that of page->_refcount.
885 * Note that if SetPageDirty is always performed via set_page_dirty,
886 * and thus under the i_pages lock, then this ordering is not required.
888 refcount
= 1 + compound_nr(page
);
889 if (!page_ref_freeze(page
, refcount
))
891 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
892 if (unlikely(PageDirty(page
))) {
893 page_ref_unfreeze(page
, refcount
);
897 if (PageSwapCache(page
)) {
898 swp_entry_t swap
= { .val
= page_private(page
) };
899 mem_cgroup_swapout(page
, swap
);
900 if (reclaimed
&& !mapping_exiting(mapping
))
901 shadow
= workingset_eviction(page
, target_memcg
);
902 __delete_from_swap_cache(page
, swap
, shadow
);
903 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
904 put_swap_page(page
, swap
);
906 void (*freepage
)(struct page
*);
908 freepage
= mapping
->a_ops
->freepage
;
910 * Remember a shadow entry for reclaimed file cache in
911 * order to detect refaults, thus thrashing, later on.
913 * But don't store shadows in an address space that is
914 * already exiting. This is not just an optimization,
915 * inode reclaim needs to empty out the radix tree or
916 * the nodes are lost. Don't plant shadows behind its
919 * We also don't store shadows for DAX mappings because the
920 * only page cache pages found in these are zero pages
921 * covering holes, and because we don't want to mix DAX
922 * exceptional entries and shadow exceptional entries in the
923 * same address_space.
925 if (reclaimed
&& page_is_file_lru(page
) &&
926 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
927 shadow
= workingset_eviction(page
, target_memcg
);
928 __delete_from_page_cache(page
, shadow
);
929 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
931 if (freepage
!= NULL
)
938 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
943 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
944 * someone else has a ref on the page, abort and return 0. If it was
945 * successfully detached, return 1. Assumes the caller has a single ref on
948 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
950 if (__remove_mapping(mapping
, page
, false, NULL
)) {
952 * Unfreezing the refcount with 1 rather than 2 effectively
953 * drops the pagecache ref for us without requiring another
956 page_ref_unfreeze(page
, 1);
963 * putback_lru_page - put previously isolated page onto appropriate LRU list
964 * @page: page to be put back to appropriate lru list
966 * Add previously isolated @page to appropriate LRU list.
967 * Page may still be unevictable for other reasons.
969 * lru_lock must not be held, interrupts must be enabled.
971 void putback_lru_page(struct page
*page
)
974 put_page(page
); /* drop ref from isolate */
977 enum page_references
{
979 PAGEREF_RECLAIM_CLEAN
,
984 static enum page_references
page_check_references(struct page
*page
,
985 struct scan_control
*sc
)
987 int referenced_ptes
, referenced_page
;
988 unsigned long vm_flags
;
990 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
992 referenced_page
= TestClearPageReferenced(page
);
995 * Mlock lost the isolation race with us. Let try_to_unmap()
996 * move the page to the unevictable list.
998 if (vm_flags
& VM_LOCKED
)
999 return PAGEREF_RECLAIM
;
1001 if (referenced_ptes
) {
1003 * All mapped pages start out with page table
1004 * references from the instantiating fault, so we need
1005 * to look twice if a mapped file page is used more
1008 * Mark it and spare it for another trip around the
1009 * inactive list. Another page table reference will
1010 * lead to its activation.
1012 * Note: the mark is set for activated pages as well
1013 * so that recently deactivated but used pages are
1014 * quickly recovered.
1016 SetPageReferenced(page
);
1018 if (referenced_page
|| referenced_ptes
> 1)
1019 return PAGEREF_ACTIVATE
;
1022 * Activate file-backed executable pages after first usage.
1024 if ((vm_flags
& VM_EXEC
) && !PageSwapBacked(page
))
1025 return PAGEREF_ACTIVATE
;
1027 return PAGEREF_KEEP
;
1030 /* Reclaim if clean, defer dirty pages to writeback */
1031 if (referenced_page
&& !PageSwapBacked(page
))
1032 return PAGEREF_RECLAIM_CLEAN
;
1034 return PAGEREF_RECLAIM
;
1037 /* Check if a page is dirty or under writeback */
1038 static void page_check_dirty_writeback(struct page
*page
,
1039 bool *dirty
, bool *writeback
)
1041 struct address_space
*mapping
;
1044 * Anonymous pages are not handled by flushers and must be written
1045 * from reclaim context. Do not stall reclaim based on them
1047 if (!page_is_file_lru(page
) ||
1048 (PageAnon(page
) && !PageSwapBacked(page
))) {
1054 /* By default assume that the page flags are accurate */
1055 *dirty
= PageDirty(page
);
1056 *writeback
= PageWriteback(page
);
1058 /* Verify dirty/writeback state if the filesystem supports it */
1059 if (!page_has_private(page
))
1062 mapping
= page_mapping(page
);
1063 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1064 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1068 * shrink_page_list() returns the number of reclaimed pages
1070 static unsigned int shrink_page_list(struct list_head
*page_list
,
1071 struct pglist_data
*pgdat
,
1072 struct scan_control
*sc
,
1073 enum ttu_flags ttu_flags
,
1074 struct reclaim_stat
*stat
,
1075 bool ignore_references
)
1077 LIST_HEAD(ret_pages
);
1078 LIST_HEAD(free_pages
);
1079 unsigned int nr_reclaimed
= 0;
1080 unsigned int pgactivate
= 0;
1082 memset(stat
, 0, sizeof(*stat
));
1085 while (!list_empty(page_list
)) {
1086 struct address_space
*mapping
;
1088 enum page_references references
= PAGEREF_RECLAIM
;
1089 bool dirty
, writeback
, may_enter_fs
;
1090 unsigned int nr_pages
;
1094 page
= lru_to_page(page_list
);
1095 list_del(&page
->lru
);
1097 if (!trylock_page(page
))
1100 VM_BUG_ON_PAGE(PageActive(page
), page
);
1102 nr_pages
= compound_nr(page
);
1104 /* Account the number of base pages even though THP */
1105 sc
->nr_scanned
+= nr_pages
;
1107 if (unlikely(!page_evictable(page
)))
1108 goto activate_locked
;
1110 if (!sc
->may_unmap
&& page_mapped(page
))
1113 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1114 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1117 * The number of dirty pages determines if a node is marked
1118 * reclaim_congested which affects wait_iff_congested. kswapd
1119 * will stall and start writing pages if the tail of the LRU
1120 * is all dirty unqueued pages.
1122 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1123 if (dirty
|| writeback
)
1126 if (dirty
&& !writeback
)
1127 stat
->nr_unqueued_dirty
++;
1130 * Treat this page as congested if the underlying BDI is or if
1131 * pages are cycling through the LRU so quickly that the
1132 * pages marked for immediate reclaim are making it to the
1133 * end of the LRU a second time.
1135 mapping
= page_mapping(page
);
1136 if (((dirty
|| writeback
) && mapping
&&
1137 inode_write_congested(mapping
->host
)) ||
1138 (writeback
&& PageReclaim(page
)))
1139 stat
->nr_congested
++;
1142 * If a page at the tail of the LRU is under writeback, there
1143 * are three cases to consider.
1145 * 1) If reclaim is encountering an excessive number of pages
1146 * under writeback and this page is both under writeback and
1147 * PageReclaim then it indicates that pages are being queued
1148 * for IO but are being recycled through the LRU before the
1149 * IO can complete. Waiting on the page itself risks an
1150 * indefinite stall if it is impossible to writeback the
1151 * page due to IO error or disconnected storage so instead
1152 * note that the LRU is being scanned too quickly and the
1153 * caller can stall after page list has been processed.
1155 * 2) Global or new memcg reclaim encounters a page that is
1156 * not marked for immediate reclaim, or the caller does not
1157 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1158 * not to fs). In this case mark the page for immediate
1159 * reclaim and continue scanning.
1161 * Require may_enter_fs because we would wait on fs, which
1162 * may not have submitted IO yet. And the loop driver might
1163 * enter reclaim, and deadlock if it waits on a page for
1164 * which it is needed to do the write (loop masks off
1165 * __GFP_IO|__GFP_FS for this reason); but more thought
1166 * would probably show more reasons.
1168 * 3) Legacy memcg encounters a page that is already marked
1169 * PageReclaim. memcg does not have any dirty pages
1170 * throttling so we could easily OOM just because too many
1171 * pages are in writeback and there is nothing else to
1172 * reclaim. Wait for the writeback to complete.
1174 * In cases 1) and 2) we activate the pages to get them out of
1175 * the way while we continue scanning for clean pages on the
1176 * inactive list and refilling from the active list. The
1177 * observation here is that waiting for disk writes is more
1178 * expensive than potentially causing reloads down the line.
1179 * Since they're marked for immediate reclaim, they won't put
1180 * memory pressure on the cache working set any longer than it
1181 * takes to write them to disk.
1183 if (PageWriteback(page
)) {
1185 if (current_is_kswapd() &&
1186 PageReclaim(page
) &&
1187 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1188 stat
->nr_immediate
++;
1189 goto activate_locked
;
1192 } else if (writeback_throttling_sane(sc
) ||
1193 !PageReclaim(page
) || !may_enter_fs
) {
1195 * This is slightly racy - end_page_writeback()
1196 * might have just cleared PageReclaim, then
1197 * setting PageReclaim here end up interpreted
1198 * as PageReadahead - but that does not matter
1199 * enough to care. What we do want is for this
1200 * page to have PageReclaim set next time memcg
1201 * reclaim reaches the tests above, so it will
1202 * then wait_on_page_writeback() to avoid OOM;
1203 * and it's also appropriate in global reclaim.
1205 SetPageReclaim(page
);
1206 stat
->nr_writeback
++;
1207 goto activate_locked
;
1212 wait_on_page_writeback(page
);
1213 /* then go back and try same page again */
1214 list_add_tail(&page
->lru
, page_list
);
1219 if (!ignore_references
)
1220 references
= page_check_references(page
, sc
);
1222 switch (references
) {
1223 case PAGEREF_ACTIVATE
:
1224 goto activate_locked
;
1226 stat
->nr_ref_keep
+= nr_pages
;
1228 case PAGEREF_RECLAIM
:
1229 case PAGEREF_RECLAIM_CLEAN
:
1230 ; /* try to reclaim the page below */
1234 * Anonymous process memory has backing store?
1235 * Try to allocate it some swap space here.
1236 * Lazyfree page could be freed directly
1238 if (PageAnon(page
) && PageSwapBacked(page
)) {
1239 if (!PageSwapCache(page
)) {
1240 if (!(sc
->gfp_mask
& __GFP_IO
))
1242 if (PageTransHuge(page
)) {
1243 /* cannot split THP, skip it */
1244 if (!can_split_huge_page(page
, NULL
))
1245 goto activate_locked
;
1247 * Split pages without a PMD map right
1248 * away. Chances are some or all of the
1249 * tail pages can be freed without IO.
1251 if (!compound_mapcount(page
) &&
1252 split_huge_page_to_list(page
,
1254 goto activate_locked
;
1256 if (!add_to_swap(page
)) {
1257 if (!PageTransHuge(page
))
1258 goto activate_locked_split
;
1259 /* Fallback to swap normal pages */
1260 if (split_huge_page_to_list(page
,
1262 goto activate_locked
;
1263 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1264 count_vm_event(THP_SWPOUT_FALLBACK
);
1266 if (!add_to_swap(page
))
1267 goto activate_locked_split
;
1270 may_enter_fs
= true;
1272 /* Adding to swap updated mapping */
1273 mapping
= page_mapping(page
);
1275 } else if (unlikely(PageTransHuge(page
))) {
1276 /* Split file THP */
1277 if (split_huge_page_to_list(page
, page_list
))
1282 * THP may get split above, need minus tail pages and update
1283 * nr_pages to avoid accounting tail pages twice.
1285 * The tail pages that are added into swap cache successfully
1288 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1289 sc
->nr_scanned
-= (nr_pages
- 1);
1294 * The page is mapped into the page tables of one or more
1295 * processes. Try to unmap it here.
1297 if (page_mapped(page
)) {
1298 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1299 bool was_swapbacked
= PageSwapBacked(page
);
1301 if (unlikely(PageTransHuge(page
)))
1302 flags
|= TTU_SPLIT_HUGE_PMD
;
1304 if (!try_to_unmap(page
, flags
)) {
1305 stat
->nr_unmap_fail
+= nr_pages
;
1306 if (!was_swapbacked
&& PageSwapBacked(page
))
1307 stat
->nr_lazyfree_fail
+= nr_pages
;
1308 goto activate_locked
;
1312 if (PageDirty(page
)) {
1314 * Only kswapd can writeback filesystem pages
1315 * to avoid risk of stack overflow. But avoid
1316 * injecting inefficient single-page IO into
1317 * flusher writeback as much as possible: only
1318 * write pages when we've encountered many
1319 * dirty pages, and when we've already scanned
1320 * the rest of the LRU for clean pages and see
1321 * the same dirty pages again (PageReclaim).
1323 if (page_is_file_lru(page
) &&
1324 (!current_is_kswapd() || !PageReclaim(page
) ||
1325 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1327 * Immediately reclaim when written back.
1328 * Similar in principal to deactivate_page()
1329 * except we already have the page isolated
1330 * and know it's dirty
1332 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1333 SetPageReclaim(page
);
1335 goto activate_locked
;
1338 if (references
== PAGEREF_RECLAIM_CLEAN
)
1342 if (!sc
->may_writepage
)
1346 * Page is dirty. Flush the TLB if a writable entry
1347 * potentially exists to avoid CPU writes after IO
1348 * starts and then write it out here.
1350 try_to_unmap_flush_dirty();
1351 switch (pageout(page
, mapping
)) {
1355 goto activate_locked
;
1357 stat
->nr_pageout
+= hpage_nr_pages(page
);
1359 if (PageWriteback(page
))
1361 if (PageDirty(page
))
1365 * A synchronous write - probably a ramdisk. Go
1366 * ahead and try to reclaim the page.
1368 if (!trylock_page(page
))
1370 if (PageDirty(page
) || PageWriteback(page
))
1372 mapping
= page_mapping(page
);
1374 ; /* try to free the page below */
1379 * If the page has buffers, try to free the buffer mappings
1380 * associated with this page. If we succeed we try to free
1383 * We do this even if the page is PageDirty().
1384 * try_to_release_page() does not perform I/O, but it is
1385 * possible for a page to have PageDirty set, but it is actually
1386 * clean (all its buffers are clean). This happens if the
1387 * buffers were written out directly, with submit_bh(). ext3
1388 * will do this, as well as the blockdev mapping.
1389 * try_to_release_page() will discover that cleanness and will
1390 * drop the buffers and mark the page clean - it can be freed.
1392 * Rarely, pages can have buffers and no ->mapping. These are
1393 * the pages which were not successfully invalidated in
1394 * truncate_complete_page(). We try to drop those buffers here
1395 * and if that worked, and the page is no longer mapped into
1396 * process address space (page_count == 1) it can be freed.
1397 * Otherwise, leave the page on the LRU so it is swappable.
1399 if (page_has_private(page
)) {
1400 if (!try_to_release_page(page
, sc
->gfp_mask
))
1401 goto activate_locked
;
1402 if (!mapping
&& page_count(page
) == 1) {
1404 if (put_page_testzero(page
))
1408 * rare race with speculative reference.
1409 * the speculative reference will free
1410 * this page shortly, so we may
1411 * increment nr_reclaimed here (and
1412 * leave it off the LRU).
1420 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1421 /* follow __remove_mapping for reference */
1422 if (!page_ref_freeze(page
, 1))
1424 if (PageDirty(page
)) {
1425 page_ref_unfreeze(page
, 1);
1429 count_vm_event(PGLAZYFREED
);
1430 count_memcg_page_event(page
, PGLAZYFREED
);
1431 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true,
1432 sc
->target_mem_cgroup
))
1438 * THP may get swapped out in a whole, need account
1441 nr_reclaimed
+= nr_pages
;
1444 * Is there need to periodically free_page_list? It would
1445 * appear not as the counts should be low
1447 if (unlikely(PageTransHuge(page
)))
1448 destroy_compound_page(page
);
1450 list_add(&page
->lru
, &free_pages
);
1453 activate_locked_split
:
1455 * The tail pages that are failed to add into swap cache
1456 * reach here. Fixup nr_scanned and nr_pages.
1459 sc
->nr_scanned
-= (nr_pages
- 1);
1463 /* Not a candidate for swapping, so reclaim swap space. */
1464 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1466 try_to_free_swap(page
);
1467 VM_BUG_ON_PAGE(PageActive(page
), page
);
1468 if (!PageMlocked(page
)) {
1469 int type
= page_is_file_lru(page
);
1470 SetPageActive(page
);
1471 stat
->nr_activate
[type
] += nr_pages
;
1472 count_memcg_page_event(page
, PGACTIVATE
);
1477 list_add(&page
->lru
, &ret_pages
);
1478 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1481 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1483 mem_cgroup_uncharge_list(&free_pages
);
1484 try_to_unmap_flush();
1485 free_unref_page_list(&free_pages
);
1487 list_splice(&ret_pages
, page_list
);
1488 count_vm_events(PGACTIVATE
, pgactivate
);
1490 return nr_reclaimed
;
1493 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1494 struct list_head
*page_list
)
1496 struct scan_control sc
= {
1497 .gfp_mask
= GFP_KERNEL
,
1498 .priority
= DEF_PRIORITY
,
1501 struct reclaim_stat stat
;
1502 unsigned int nr_reclaimed
;
1503 struct page
*page
, *next
;
1504 LIST_HEAD(clean_pages
);
1506 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1507 if (page_is_file_lru(page
) && !PageDirty(page
) &&
1508 !__PageMovable(page
) && !PageUnevictable(page
)) {
1509 ClearPageActive(page
);
1510 list_move(&page
->lru
, &clean_pages
);
1514 nr_reclaimed
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1515 TTU_IGNORE_ACCESS
, &stat
, true);
1516 list_splice(&clean_pages
, page_list
);
1517 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -nr_reclaimed
);
1519 * Since lazyfree pages are isolated from file LRU from the beginning,
1520 * they will rotate back to anonymous LRU in the end if it failed to
1521 * discard so isolated count will be mismatched.
1522 * Compensate the isolated count for both LRU lists.
1524 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
1525 stat
.nr_lazyfree_fail
);
1526 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1527 -stat
.nr_lazyfree_fail
);
1528 return nr_reclaimed
;
1532 * Attempt to remove the specified page from its LRU. Only take this page
1533 * if it is of the appropriate PageActive status. Pages which are being
1534 * freed elsewhere are also ignored.
1536 * page: page to consider
1537 * mode: one of the LRU isolation modes defined above
1539 * returns 0 on success, -ve errno on failure.
1541 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1545 /* Only take pages on the LRU. */
1549 /* Compaction should not handle unevictable pages but CMA can do so */
1550 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1556 * To minimise LRU disruption, the caller can indicate that it only
1557 * wants to isolate pages it will be able to operate on without
1558 * blocking - clean pages for the most part.
1560 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1561 * that it is possible to migrate without blocking
1563 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1564 /* All the caller can do on PageWriteback is block */
1565 if (PageWriteback(page
))
1568 if (PageDirty(page
)) {
1569 struct address_space
*mapping
;
1573 * Only pages without mappings or that have a
1574 * ->migratepage callback are possible to migrate
1575 * without blocking. However, we can be racing with
1576 * truncation so it's necessary to lock the page
1577 * to stabilise the mapping as truncation holds
1578 * the page lock until after the page is removed
1579 * from the page cache.
1581 if (!trylock_page(page
))
1584 mapping
= page_mapping(page
);
1585 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1592 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1595 if (likely(get_page_unless_zero(page
))) {
1597 * Be careful not to clear PageLRU until after we're
1598 * sure the page is not being freed elsewhere -- the
1599 * page release code relies on it.
1610 * Update LRU sizes after isolating pages. The LRU size updates must
1611 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1613 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1614 enum lru_list lru
, unsigned long *nr_zone_taken
)
1618 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1619 if (!nr_zone_taken
[zid
])
1622 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1628 * pgdat->lru_lock is heavily contended. Some of the functions that
1629 * shrink the lists perform better by taking out a batch of pages
1630 * and working on them outside the LRU lock.
1632 * For pagecache intensive workloads, this function is the hottest
1633 * spot in the kernel (apart from copy_*_user functions).
1635 * Appropriate locks must be held before calling this function.
1637 * @nr_to_scan: The number of eligible pages to look through on the list.
1638 * @lruvec: The LRU vector to pull pages from.
1639 * @dst: The temp list to put pages on to.
1640 * @nr_scanned: The number of pages that were scanned.
1641 * @sc: The scan_control struct for this reclaim session
1642 * @lru: LRU list id for isolating
1644 * returns how many pages were moved onto *@dst.
1646 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1647 struct lruvec
*lruvec
, struct list_head
*dst
,
1648 unsigned long *nr_scanned
, struct scan_control
*sc
,
1651 struct list_head
*src
= &lruvec
->lists
[lru
];
1652 unsigned long nr_taken
= 0;
1653 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1654 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1655 unsigned long skipped
= 0;
1656 unsigned long scan
, total_scan
, nr_pages
;
1657 LIST_HEAD(pages_skipped
);
1658 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1662 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1665 page
= lru_to_page(src
);
1666 prefetchw_prev_lru_page(page
, src
, flags
);
1668 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1670 nr_pages
= compound_nr(page
);
1671 total_scan
+= nr_pages
;
1673 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1674 list_move(&page
->lru
, &pages_skipped
);
1675 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1680 * Do not count skipped pages because that makes the function
1681 * return with no isolated pages if the LRU mostly contains
1682 * ineligible pages. This causes the VM to not reclaim any
1683 * pages, triggering a premature OOM.
1685 * Account all tail pages of THP. This would not cause
1686 * premature OOM since __isolate_lru_page() returns -EBUSY
1687 * only when the page is being freed somewhere else.
1690 switch (__isolate_lru_page(page
, mode
)) {
1692 nr_taken
+= nr_pages
;
1693 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1694 list_move(&page
->lru
, dst
);
1698 /* else it is being freed elsewhere */
1699 list_move(&page
->lru
, src
);
1708 * Splice any skipped pages to the start of the LRU list. Note that
1709 * this disrupts the LRU order when reclaiming for lower zones but
1710 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1711 * scanning would soon rescan the same pages to skip and put the
1712 * system at risk of premature OOM.
1714 if (!list_empty(&pages_skipped
)) {
1717 list_splice(&pages_skipped
, src
);
1718 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1719 if (!nr_skipped
[zid
])
1722 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1723 skipped
+= nr_skipped
[zid
];
1726 *nr_scanned
= total_scan
;
1727 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1728 total_scan
, skipped
, nr_taken
, mode
, lru
);
1729 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1734 * isolate_lru_page - tries to isolate a page from its LRU list
1735 * @page: page to isolate from its LRU list
1737 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1738 * vmstat statistic corresponding to whatever LRU list the page was on.
1740 * Returns 0 if the page was removed from an LRU list.
1741 * Returns -EBUSY if the page was not on an LRU list.
1743 * The returned page will have PageLRU() cleared. If it was found on
1744 * the active list, it will have PageActive set. If it was found on
1745 * the unevictable list, it will have the PageUnevictable bit set. That flag
1746 * may need to be cleared by the caller before letting the page go.
1748 * The vmstat statistic corresponding to the list on which the page was
1749 * found will be decremented.
1753 * (1) Must be called with an elevated refcount on the page. This is a
1754 * fundamentnal difference from isolate_lru_pages (which is called
1755 * without a stable reference).
1756 * (2) the lru_lock must not be held.
1757 * (3) interrupts must be enabled.
1759 int isolate_lru_page(struct page
*page
)
1763 VM_BUG_ON_PAGE(!page_count(page
), page
);
1764 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1766 if (PageLRU(page
)) {
1767 pg_data_t
*pgdat
= page_pgdat(page
);
1768 struct lruvec
*lruvec
;
1770 spin_lock_irq(&pgdat
->lru_lock
);
1771 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1772 if (PageLRU(page
)) {
1773 int lru
= page_lru(page
);
1776 del_page_from_lru_list(page
, lruvec
, lru
);
1779 spin_unlock_irq(&pgdat
->lru_lock
);
1785 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1786 * then get rescheduled. When there are massive number of tasks doing page
1787 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1788 * the LRU list will go small and be scanned faster than necessary, leading to
1789 * unnecessary swapping, thrashing and OOM.
1791 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1792 struct scan_control
*sc
)
1794 unsigned long inactive
, isolated
;
1796 if (current_is_kswapd())
1799 if (!writeback_throttling_sane(sc
))
1803 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1804 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1806 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1807 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1811 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1812 * won't get blocked by normal direct-reclaimers, forming a circular
1815 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1818 return isolated
> inactive
;
1822 * This moves pages from @list to corresponding LRU list.
1824 * We move them the other way if the page is referenced by one or more
1825 * processes, from rmap.
1827 * If the pages are mostly unmapped, the processing is fast and it is
1828 * appropriate to hold zone_lru_lock across the whole operation. But if
1829 * the pages are mapped, the processing is slow (page_referenced()) so we
1830 * should drop zone_lru_lock around each page. It's impossible to balance
1831 * this, so instead we remove the pages from the LRU while processing them.
1832 * It is safe to rely on PG_active against the non-LRU pages in here because
1833 * nobody will play with that bit on a non-LRU page.
1835 * The downside is that we have to touch page->_refcount against each page.
1836 * But we had to alter page->flags anyway.
1838 * Returns the number of pages moved to the given lruvec.
1841 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1842 struct list_head
*list
)
1844 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1845 int nr_pages
, nr_moved
= 0;
1846 LIST_HEAD(pages_to_free
);
1850 while (!list_empty(list
)) {
1851 page
= lru_to_page(list
);
1852 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1853 if (unlikely(!page_evictable(page
))) {
1854 list_del(&page
->lru
);
1855 spin_unlock_irq(&pgdat
->lru_lock
);
1856 putback_lru_page(page
);
1857 spin_lock_irq(&pgdat
->lru_lock
);
1860 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1863 lru
= page_lru(page
);
1865 nr_pages
= hpage_nr_pages(page
);
1866 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1867 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1869 if (put_page_testzero(page
)) {
1870 __ClearPageLRU(page
);
1871 __ClearPageActive(page
);
1872 del_page_from_lru_list(page
, lruvec
, lru
);
1874 if (unlikely(PageCompound(page
))) {
1875 spin_unlock_irq(&pgdat
->lru_lock
);
1876 destroy_compound_page(page
);
1877 spin_lock_irq(&pgdat
->lru_lock
);
1879 list_add(&page
->lru
, &pages_to_free
);
1881 nr_moved
+= nr_pages
;
1882 if (PageActive(page
))
1883 workingset_age_nonresident(lruvec
, nr_pages
);
1888 * To save our caller's stack, now use input list for pages to free.
1890 list_splice(&pages_to_free
, list
);
1896 * If a kernel thread (such as nfsd for loop-back mounts) services
1897 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1898 * In that case we should only throttle if the backing device it is
1899 * writing to is congested. In other cases it is safe to throttle.
1901 static int current_may_throttle(void)
1903 return !(current
->flags
& PF_LOCAL_THROTTLE
) ||
1904 current
->backing_dev_info
== NULL
||
1905 bdi_write_congested(current
->backing_dev_info
);
1909 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1910 * of reclaimed pages
1912 static noinline_for_stack
unsigned long
1913 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1914 struct scan_control
*sc
, enum lru_list lru
)
1916 LIST_HEAD(page_list
);
1917 unsigned long nr_scanned
;
1918 unsigned int nr_reclaimed
= 0;
1919 unsigned long nr_taken
;
1920 struct reclaim_stat stat
;
1921 bool file
= is_file_lru(lru
);
1922 enum vm_event_item item
;
1923 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1924 bool stalled
= false;
1926 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1930 /* wait a bit for the reclaimer. */
1934 /* We are about to die and free our memory. Return now. */
1935 if (fatal_signal_pending(current
))
1936 return SWAP_CLUSTER_MAX
;
1941 spin_lock_irq(&pgdat
->lru_lock
);
1943 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1944 &nr_scanned
, sc
, lru
);
1946 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1947 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1948 if (!cgroup_reclaim(sc
))
1949 __count_vm_events(item
, nr_scanned
);
1950 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1951 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
1953 spin_unlock_irq(&pgdat
->lru_lock
);
1958 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1961 spin_lock_irq(&pgdat
->lru_lock
);
1963 move_pages_to_lru(lruvec
, &page_list
);
1965 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1966 lru_note_cost(lruvec
, file
, stat
.nr_pageout
);
1967 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
1968 if (!cgroup_reclaim(sc
))
1969 __count_vm_events(item
, nr_reclaimed
);
1970 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
1971 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
1973 spin_unlock_irq(&pgdat
->lru_lock
);
1975 mem_cgroup_uncharge_list(&page_list
);
1976 free_unref_page_list(&page_list
);
1979 * If dirty pages are scanned that are not queued for IO, it
1980 * implies that flushers are not doing their job. This can
1981 * happen when memory pressure pushes dirty pages to the end of
1982 * the LRU before the dirty limits are breached and the dirty
1983 * data has expired. It can also happen when the proportion of
1984 * dirty pages grows not through writes but through memory
1985 * pressure reclaiming all the clean cache. And in some cases,
1986 * the flushers simply cannot keep up with the allocation
1987 * rate. Nudge the flusher threads in case they are asleep.
1989 if (stat
.nr_unqueued_dirty
== nr_taken
)
1990 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1992 sc
->nr
.dirty
+= stat
.nr_dirty
;
1993 sc
->nr
.congested
+= stat
.nr_congested
;
1994 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1995 sc
->nr
.writeback
+= stat
.nr_writeback
;
1996 sc
->nr
.immediate
+= stat
.nr_immediate
;
1997 sc
->nr
.taken
+= nr_taken
;
1999 sc
->nr
.file_taken
+= nr_taken
;
2001 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2002 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2003 return nr_reclaimed
;
2006 static void shrink_active_list(unsigned long nr_to_scan
,
2007 struct lruvec
*lruvec
,
2008 struct scan_control
*sc
,
2011 unsigned long nr_taken
;
2012 unsigned long nr_scanned
;
2013 unsigned long vm_flags
;
2014 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2015 LIST_HEAD(l_active
);
2016 LIST_HEAD(l_inactive
);
2018 unsigned nr_deactivate
, nr_activate
;
2019 unsigned nr_rotated
= 0;
2020 int file
= is_file_lru(lru
);
2021 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2025 spin_lock_irq(&pgdat
->lru_lock
);
2027 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2028 &nr_scanned
, sc
, lru
);
2030 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2032 if (!cgroup_reclaim(sc
))
2033 __count_vm_events(PGREFILL
, nr_scanned
);
2034 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2036 spin_unlock_irq(&pgdat
->lru_lock
);
2038 while (!list_empty(&l_hold
)) {
2040 page
= lru_to_page(&l_hold
);
2041 list_del(&page
->lru
);
2043 if (unlikely(!page_evictable(page
))) {
2044 putback_lru_page(page
);
2048 if (unlikely(buffer_heads_over_limit
)) {
2049 if (page_has_private(page
) && trylock_page(page
)) {
2050 if (page_has_private(page
))
2051 try_to_release_page(page
, 0);
2056 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2059 * Identify referenced, file-backed active pages and
2060 * give them one more trip around the active list. So
2061 * that executable code get better chances to stay in
2062 * memory under moderate memory pressure. Anon pages
2063 * are not likely to be evicted by use-once streaming
2064 * IO, plus JVM can create lots of anon VM_EXEC pages,
2065 * so we ignore them here.
2067 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2068 nr_rotated
+= hpage_nr_pages(page
);
2069 list_add(&page
->lru
, &l_active
);
2074 ClearPageActive(page
); /* we are de-activating */
2075 SetPageWorkingset(page
);
2076 list_add(&page
->lru
, &l_inactive
);
2080 * Move pages back to the lru list.
2082 spin_lock_irq(&pgdat
->lru_lock
);
2084 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2085 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2086 /* Keep all free pages in l_active list */
2087 list_splice(&l_inactive
, &l_active
);
2089 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2090 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2092 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2093 spin_unlock_irq(&pgdat
->lru_lock
);
2095 mem_cgroup_uncharge_list(&l_active
);
2096 free_unref_page_list(&l_active
);
2097 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2098 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2101 unsigned long reclaim_pages(struct list_head
*page_list
)
2103 int nid
= NUMA_NO_NODE
;
2104 unsigned int nr_reclaimed
= 0;
2105 LIST_HEAD(node_page_list
);
2106 struct reclaim_stat dummy_stat
;
2108 struct scan_control sc
= {
2109 .gfp_mask
= GFP_KERNEL
,
2110 .priority
= DEF_PRIORITY
,
2116 while (!list_empty(page_list
)) {
2117 page
= lru_to_page(page_list
);
2118 if (nid
== NUMA_NO_NODE
) {
2119 nid
= page_to_nid(page
);
2120 INIT_LIST_HEAD(&node_page_list
);
2123 if (nid
== page_to_nid(page
)) {
2124 ClearPageActive(page
);
2125 list_move(&page
->lru
, &node_page_list
);
2129 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2132 &dummy_stat
, false);
2133 while (!list_empty(&node_page_list
)) {
2134 page
= lru_to_page(&node_page_list
);
2135 list_del(&page
->lru
);
2136 putback_lru_page(page
);
2142 if (!list_empty(&node_page_list
)) {
2143 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2146 &dummy_stat
, false);
2147 while (!list_empty(&node_page_list
)) {
2148 page
= lru_to_page(&node_page_list
);
2149 list_del(&page
->lru
);
2150 putback_lru_page(page
);
2154 return nr_reclaimed
;
2157 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2158 struct lruvec
*lruvec
, struct scan_control
*sc
)
2160 if (is_active_lru(lru
)) {
2161 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2162 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2164 sc
->skipped_deactivate
= 1;
2168 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2172 * The inactive anon list should be small enough that the VM never has
2173 * to do too much work.
2175 * The inactive file list should be small enough to leave most memory
2176 * to the established workingset on the scan-resistant active list,
2177 * but large enough to avoid thrashing the aggregate readahead window.
2179 * Both inactive lists should also be large enough that each inactive
2180 * page has a chance to be referenced again before it is reclaimed.
2182 * If that fails and refaulting is observed, the inactive list grows.
2184 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2185 * on this LRU, maintained by the pageout code. An inactive_ratio
2186 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2189 * memory ratio inactive
2190 * -------------------------------------
2199 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2201 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2202 unsigned long inactive
, active
;
2203 unsigned long inactive_ratio
;
2206 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2207 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2209 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2211 inactive_ratio
= int_sqrt(10 * gb
);
2215 return inactive
* inactive_ratio
< active
;
2226 * Determine how aggressively the anon and file LRU lists should be
2227 * scanned. The relative value of each set of LRU lists is determined
2228 * by looking at the fraction of the pages scanned we did rotate back
2229 * onto the active list instead of evict.
2231 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2232 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2234 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2237 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2238 unsigned long anon_cost
, file_cost
, total_cost
;
2239 int swappiness
= mem_cgroup_swappiness(memcg
);
2241 u64 denominator
= 0; /* gcc */
2242 enum scan_balance scan_balance
;
2243 unsigned long ap
, fp
;
2246 /* If we have no swap space, do not bother scanning anon pages. */
2247 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2248 scan_balance
= SCAN_FILE
;
2253 * Global reclaim will swap to prevent OOM even with no
2254 * swappiness, but memcg users want to use this knob to
2255 * disable swapping for individual groups completely when
2256 * using the memory controller's swap limit feature would be
2259 if (cgroup_reclaim(sc
) && !swappiness
) {
2260 scan_balance
= SCAN_FILE
;
2265 * Do not apply any pressure balancing cleverness when the
2266 * system is close to OOM, scan both anon and file equally
2267 * (unless the swappiness setting disagrees with swapping).
2269 if (!sc
->priority
&& swappiness
) {
2270 scan_balance
= SCAN_EQUAL
;
2275 * If the system is almost out of file pages, force-scan anon.
2277 if (sc
->file_is_tiny
) {
2278 scan_balance
= SCAN_ANON
;
2283 * If there is enough inactive page cache, we do not reclaim
2284 * anything from the anonymous working right now.
2286 if (sc
->cache_trim_mode
) {
2287 scan_balance
= SCAN_FILE
;
2291 scan_balance
= SCAN_FRACT
;
2293 * Calculate the pressure balance between anon and file pages.
2295 * The amount of pressure we put on each LRU is inversely
2296 * proportional to the cost of reclaiming each list, as
2297 * determined by the share of pages that are refaulting, times
2298 * the relative IO cost of bringing back a swapped out
2299 * anonymous page vs reloading a filesystem page (swappiness).
2301 * Although we limit that influence to ensure no list gets
2302 * left behind completely: at least a third of the pressure is
2303 * applied, before swappiness.
2305 * With swappiness at 100, anon and file have equal IO cost.
2307 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2308 anon_cost
= total_cost
+ sc
->anon_cost
;
2309 file_cost
= total_cost
+ sc
->file_cost
;
2310 total_cost
= anon_cost
+ file_cost
;
2312 ap
= swappiness
* (total_cost
+ 1);
2313 ap
/= anon_cost
+ 1;
2315 fp
= (200 - swappiness
) * (total_cost
+ 1);
2316 fp
/= file_cost
+ 1;
2320 denominator
= ap
+ fp
;
2322 for_each_evictable_lru(lru
) {
2323 int file
= is_file_lru(lru
);
2324 unsigned long lruvec_size
;
2326 unsigned long protection
;
2328 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2329 protection
= mem_cgroup_protection(sc
->target_mem_cgroup
,
2331 sc
->memcg_low_reclaim
);
2335 * Scale a cgroup's reclaim pressure by proportioning
2336 * its current usage to its memory.low or memory.min
2339 * This is important, as otherwise scanning aggression
2340 * becomes extremely binary -- from nothing as we
2341 * approach the memory protection threshold, to totally
2342 * nominal as we exceed it. This results in requiring
2343 * setting extremely liberal protection thresholds. It
2344 * also means we simply get no protection at all if we
2345 * set it too low, which is not ideal.
2347 * If there is any protection in place, we reduce scan
2348 * pressure by how much of the total memory used is
2349 * within protection thresholds.
2351 * There is one special case: in the first reclaim pass,
2352 * we skip over all groups that are within their low
2353 * protection. If that fails to reclaim enough pages to
2354 * satisfy the reclaim goal, we come back and override
2355 * the best-effort low protection. However, we still
2356 * ideally want to honor how well-behaved groups are in
2357 * that case instead of simply punishing them all
2358 * equally. As such, we reclaim them based on how much
2359 * memory they are using, reducing the scan pressure
2360 * again by how much of the total memory used is under
2363 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2365 /* Avoid TOCTOU with earlier protection check */
2366 cgroup_size
= max(cgroup_size
, protection
);
2368 scan
= lruvec_size
- lruvec_size
* protection
/
2372 * Minimally target SWAP_CLUSTER_MAX pages to keep
2373 * reclaim moving forwards, avoiding decrementing
2374 * sc->priority further than desirable.
2376 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2381 scan
>>= sc
->priority
;
2384 * If the cgroup's already been deleted, make sure to
2385 * scrape out the remaining cache.
2387 if (!scan
&& !mem_cgroup_online(memcg
))
2388 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2390 switch (scan_balance
) {
2392 /* Scan lists relative to size */
2396 * Scan types proportional to swappiness and
2397 * their relative recent reclaim efficiency.
2398 * Make sure we don't miss the last page on
2399 * the offlined memory cgroups because of a
2402 scan
= mem_cgroup_online(memcg
) ?
2403 div64_u64(scan
* fraction
[file
], denominator
) :
2404 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2409 /* Scan one type exclusively */
2410 if ((scan_balance
== SCAN_FILE
) != file
)
2414 /* Look ma, no brain */
2422 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2424 unsigned long nr
[NR_LRU_LISTS
];
2425 unsigned long targets
[NR_LRU_LISTS
];
2426 unsigned long nr_to_scan
;
2428 unsigned long nr_reclaimed
= 0;
2429 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2430 struct blk_plug plug
;
2433 get_scan_count(lruvec
, sc
, nr
);
2435 /* Record the original scan target for proportional adjustments later */
2436 memcpy(targets
, nr
, sizeof(nr
));
2439 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2440 * event that can occur when there is little memory pressure e.g.
2441 * multiple streaming readers/writers. Hence, we do not abort scanning
2442 * when the requested number of pages are reclaimed when scanning at
2443 * DEF_PRIORITY on the assumption that the fact we are direct
2444 * reclaiming implies that kswapd is not keeping up and it is best to
2445 * do a batch of work at once. For memcg reclaim one check is made to
2446 * abort proportional reclaim if either the file or anon lru has already
2447 * dropped to zero at the first pass.
2449 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2450 sc
->priority
== DEF_PRIORITY
);
2452 blk_start_plug(&plug
);
2453 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2454 nr
[LRU_INACTIVE_FILE
]) {
2455 unsigned long nr_anon
, nr_file
, percentage
;
2456 unsigned long nr_scanned
;
2458 for_each_evictable_lru(lru
) {
2460 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2461 nr
[lru
] -= nr_to_scan
;
2463 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2470 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2474 * For kswapd and memcg, reclaim at least the number of pages
2475 * requested. Ensure that the anon and file LRUs are scanned
2476 * proportionally what was requested by get_scan_count(). We
2477 * stop reclaiming one LRU and reduce the amount scanning
2478 * proportional to the original scan target.
2480 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2481 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2484 * It's just vindictive to attack the larger once the smaller
2485 * has gone to zero. And given the way we stop scanning the
2486 * smaller below, this makes sure that we only make one nudge
2487 * towards proportionality once we've got nr_to_reclaim.
2489 if (!nr_file
|| !nr_anon
)
2492 if (nr_file
> nr_anon
) {
2493 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2494 targets
[LRU_ACTIVE_ANON
] + 1;
2496 percentage
= nr_anon
* 100 / scan_target
;
2498 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2499 targets
[LRU_ACTIVE_FILE
] + 1;
2501 percentage
= nr_file
* 100 / scan_target
;
2504 /* Stop scanning the smaller of the LRU */
2506 nr
[lru
+ LRU_ACTIVE
] = 0;
2509 * Recalculate the other LRU scan count based on its original
2510 * scan target and the percentage scanning already complete
2512 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2513 nr_scanned
= targets
[lru
] - nr
[lru
];
2514 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2515 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2518 nr_scanned
= targets
[lru
] - nr
[lru
];
2519 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2520 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2522 scan_adjusted
= true;
2524 blk_finish_plug(&plug
);
2525 sc
->nr_reclaimed
+= nr_reclaimed
;
2528 * Even if we did not try to evict anon pages at all, we want to
2529 * rebalance the anon lru active/inactive ratio.
2531 if (total_swap_pages
&& inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
2532 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2533 sc
, LRU_ACTIVE_ANON
);
2536 /* Use reclaim/compaction for costly allocs or under memory pressure */
2537 static bool in_reclaim_compaction(struct scan_control
*sc
)
2539 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2540 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2541 sc
->priority
< DEF_PRIORITY
- 2))
2548 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2549 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2550 * true if more pages should be reclaimed such that when the page allocator
2551 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2552 * It will give up earlier than that if there is difficulty reclaiming pages.
2554 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2555 unsigned long nr_reclaimed
,
2556 struct scan_control
*sc
)
2558 unsigned long pages_for_compaction
;
2559 unsigned long inactive_lru_pages
;
2562 /* If not in reclaim/compaction mode, stop */
2563 if (!in_reclaim_compaction(sc
))
2567 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2568 * number of pages that were scanned. This will return to the caller
2569 * with the risk reclaim/compaction and the resulting allocation attempt
2570 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2571 * allocations through requiring that the full LRU list has been scanned
2572 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2573 * scan, but that approximation was wrong, and there were corner cases
2574 * where always a non-zero amount of pages were scanned.
2579 /* If compaction would go ahead or the allocation would succeed, stop */
2580 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2581 struct zone
*zone
= &pgdat
->node_zones
[z
];
2582 if (!managed_zone(zone
))
2585 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2586 case COMPACT_SUCCESS
:
2587 case COMPACT_CONTINUE
:
2590 /* check next zone */
2596 * If we have not reclaimed enough pages for compaction and the
2597 * inactive lists are large enough, continue reclaiming
2599 pages_for_compaction
= compact_gap(sc
->order
);
2600 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2601 if (get_nr_swap_pages() > 0)
2602 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2604 return inactive_lru_pages
> pages_for_compaction
;
2607 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
2609 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
2610 struct mem_cgroup
*memcg
;
2612 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
2614 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
2615 unsigned long reclaimed
;
2616 unsigned long scanned
;
2618 mem_cgroup_calculate_protection(target_memcg
, memcg
);
2620 if (mem_cgroup_below_min(memcg
)) {
2623 * If there is no reclaimable memory, OOM.
2626 } else if (mem_cgroup_below_low(memcg
)) {
2629 * Respect the protection only as long as
2630 * there is an unprotected supply
2631 * of reclaimable memory from other cgroups.
2633 if (!sc
->memcg_low_reclaim
) {
2634 sc
->memcg_low_skipped
= 1;
2637 memcg_memory_event(memcg
, MEMCG_LOW
);
2640 reclaimed
= sc
->nr_reclaimed
;
2641 scanned
= sc
->nr_scanned
;
2643 shrink_lruvec(lruvec
, sc
);
2645 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2648 /* Record the group's reclaim efficiency */
2649 vmpressure(sc
->gfp_mask
, memcg
, false,
2650 sc
->nr_scanned
- scanned
,
2651 sc
->nr_reclaimed
- reclaimed
);
2653 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
2656 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2658 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2659 unsigned long nr_reclaimed
, nr_scanned
;
2660 struct lruvec
*target_lruvec
;
2661 bool reclaimable
= false;
2664 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
2667 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2669 nr_reclaimed
= sc
->nr_reclaimed
;
2670 nr_scanned
= sc
->nr_scanned
;
2673 * Determine the scan balance between anon and file LRUs.
2675 spin_lock_irq(&pgdat
->lru_lock
);
2676 sc
->anon_cost
= target_lruvec
->anon_cost
;
2677 sc
->file_cost
= target_lruvec
->file_cost
;
2678 spin_unlock_irq(&pgdat
->lru_lock
);
2681 * Target desirable inactive:active list ratios for the anon
2682 * and file LRU lists.
2684 if (!sc
->force_deactivate
) {
2685 unsigned long refaults
;
2687 refaults
= lruvec_page_state(target_lruvec
,
2688 WORKINGSET_ACTIVATE_ANON
);
2689 if (refaults
!= target_lruvec
->refaults
[0] ||
2690 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
2691 sc
->may_deactivate
|= DEACTIVATE_ANON
;
2693 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
2696 * When refaults are being observed, it means a new
2697 * workingset is being established. Deactivate to get
2698 * rid of any stale active pages quickly.
2700 refaults
= lruvec_page_state(target_lruvec
,
2701 WORKINGSET_ACTIVATE_FILE
);
2702 if (refaults
!= target_lruvec
->refaults
[1] ||
2703 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
2704 sc
->may_deactivate
|= DEACTIVATE_FILE
;
2706 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
2708 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
2711 * If we have plenty of inactive file pages that aren't
2712 * thrashing, try to reclaim those first before touching
2715 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
2716 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
2717 sc
->cache_trim_mode
= 1;
2719 sc
->cache_trim_mode
= 0;
2722 * Prevent the reclaimer from falling into the cache trap: as
2723 * cache pages start out inactive, every cache fault will tip
2724 * the scan balance towards the file LRU. And as the file LRU
2725 * shrinks, so does the window for rotation from references.
2726 * This means we have a runaway feedback loop where a tiny
2727 * thrashing file LRU becomes infinitely more attractive than
2728 * anon pages. Try to detect this based on file LRU size.
2730 if (!cgroup_reclaim(sc
)) {
2731 unsigned long total_high_wmark
= 0;
2732 unsigned long free
, anon
;
2735 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2736 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2737 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2739 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2740 struct zone
*zone
= &pgdat
->node_zones
[z
];
2741 if (!managed_zone(zone
))
2744 total_high_wmark
+= high_wmark_pages(zone
);
2748 * Consider anon: if that's low too, this isn't a
2749 * runaway file reclaim problem, but rather just
2750 * extreme pressure. Reclaim as per usual then.
2752 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2755 file
+ free
<= total_high_wmark
&&
2756 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
2757 anon
>> sc
->priority
;
2760 shrink_node_memcgs(pgdat
, sc
);
2762 if (reclaim_state
) {
2763 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2764 reclaim_state
->reclaimed_slab
= 0;
2767 /* Record the subtree's reclaim efficiency */
2768 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2769 sc
->nr_scanned
- nr_scanned
,
2770 sc
->nr_reclaimed
- nr_reclaimed
);
2772 if (sc
->nr_reclaimed
- nr_reclaimed
)
2775 if (current_is_kswapd()) {
2777 * If reclaim is isolating dirty pages under writeback,
2778 * it implies that the long-lived page allocation rate
2779 * is exceeding the page laundering rate. Either the
2780 * global limits are not being effective at throttling
2781 * processes due to the page distribution throughout
2782 * zones or there is heavy usage of a slow backing
2783 * device. The only option is to throttle from reclaim
2784 * context which is not ideal as there is no guarantee
2785 * the dirtying process is throttled in the same way
2786 * balance_dirty_pages() manages.
2788 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2789 * count the number of pages under pages flagged for
2790 * immediate reclaim and stall if any are encountered
2791 * in the nr_immediate check below.
2793 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2794 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2796 /* Allow kswapd to start writing pages during reclaim.*/
2797 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2798 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2801 * If kswapd scans pages marked for immediate
2802 * reclaim and under writeback (nr_immediate), it
2803 * implies that pages are cycling through the LRU
2804 * faster than they are written so also forcibly stall.
2806 if (sc
->nr
.immediate
)
2807 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2811 * Tag a node/memcg as congested if all the dirty pages
2812 * scanned were backed by a congested BDI and
2813 * wait_iff_congested will stall.
2815 * Legacy memcg will stall in page writeback so avoid forcibly
2816 * stalling in wait_iff_congested().
2818 if ((current_is_kswapd() ||
2819 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
2820 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2821 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
2824 * Stall direct reclaim for IO completions if underlying BDIs
2825 * and node is congested. Allow kswapd to continue until it
2826 * starts encountering unqueued dirty pages or cycling through
2827 * the LRU too quickly.
2829 if (!current_is_kswapd() && current_may_throttle() &&
2830 !sc
->hibernation_mode
&&
2831 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
2832 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2834 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2839 * Kswapd gives up on balancing particular nodes after too
2840 * many failures to reclaim anything from them and goes to
2841 * sleep. On reclaim progress, reset the failure counter. A
2842 * successful direct reclaim run will revive a dormant kswapd.
2845 pgdat
->kswapd_failures
= 0;
2849 * Returns true if compaction should go ahead for a costly-order request, or
2850 * the allocation would already succeed without compaction. Return false if we
2851 * should reclaim first.
2853 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2855 unsigned long watermark
;
2856 enum compact_result suitable
;
2858 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2859 if (suitable
== COMPACT_SUCCESS
)
2860 /* Allocation should succeed already. Don't reclaim. */
2862 if (suitable
== COMPACT_SKIPPED
)
2863 /* Compaction cannot yet proceed. Do reclaim. */
2867 * Compaction is already possible, but it takes time to run and there
2868 * are potentially other callers using the pages just freed. So proceed
2869 * with reclaim to make a buffer of free pages available to give
2870 * compaction a reasonable chance of completing and allocating the page.
2871 * Note that we won't actually reclaim the whole buffer in one attempt
2872 * as the target watermark in should_continue_reclaim() is lower. But if
2873 * we are already above the high+gap watermark, don't reclaim at all.
2875 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2877 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2881 * This is the direct reclaim path, for page-allocating processes. We only
2882 * try to reclaim pages from zones which will satisfy the caller's allocation
2885 * If a zone is deemed to be full of pinned pages then just give it a light
2886 * scan then give up on it.
2888 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2892 unsigned long nr_soft_reclaimed
;
2893 unsigned long nr_soft_scanned
;
2895 pg_data_t
*last_pgdat
= NULL
;
2898 * If the number of buffer_heads in the machine exceeds the maximum
2899 * allowed level, force direct reclaim to scan the highmem zone as
2900 * highmem pages could be pinning lowmem pages storing buffer_heads
2902 orig_mask
= sc
->gfp_mask
;
2903 if (buffer_heads_over_limit
) {
2904 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2905 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2908 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2909 sc
->reclaim_idx
, sc
->nodemask
) {
2911 * Take care memory controller reclaiming has small influence
2914 if (!cgroup_reclaim(sc
)) {
2915 if (!cpuset_zone_allowed(zone
,
2916 GFP_KERNEL
| __GFP_HARDWALL
))
2920 * If we already have plenty of memory free for
2921 * compaction in this zone, don't free any more.
2922 * Even though compaction is invoked for any
2923 * non-zero order, only frequent costly order
2924 * reclamation is disruptive enough to become a
2925 * noticeable problem, like transparent huge
2928 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2929 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2930 compaction_ready(zone
, sc
)) {
2931 sc
->compaction_ready
= true;
2936 * Shrink each node in the zonelist once. If the
2937 * zonelist is ordered by zone (not the default) then a
2938 * node may be shrunk multiple times but in that case
2939 * the user prefers lower zones being preserved.
2941 if (zone
->zone_pgdat
== last_pgdat
)
2945 * This steals pages from memory cgroups over softlimit
2946 * and returns the number of reclaimed pages and
2947 * scanned pages. This works for global memory pressure
2948 * and balancing, not for a memcg's limit.
2950 nr_soft_scanned
= 0;
2951 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2952 sc
->order
, sc
->gfp_mask
,
2954 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2955 sc
->nr_scanned
+= nr_soft_scanned
;
2956 /* need some check for avoid more shrink_zone() */
2959 /* See comment about same check for global reclaim above */
2960 if (zone
->zone_pgdat
== last_pgdat
)
2962 last_pgdat
= zone
->zone_pgdat
;
2963 shrink_node(zone
->zone_pgdat
, sc
);
2967 * Restore to original mask to avoid the impact on the caller if we
2968 * promoted it to __GFP_HIGHMEM.
2970 sc
->gfp_mask
= orig_mask
;
2973 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
2975 struct lruvec
*target_lruvec
;
2976 unsigned long refaults
;
2978 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
2979 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
2980 target_lruvec
->refaults
[0] = refaults
;
2981 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
2982 target_lruvec
->refaults
[1] = refaults
;
2986 * This is the main entry point to direct page reclaim.
2988 * If a full scan of the inactive list fails to free enough memory then we
2989 * are "out of memory" and something needs to be killed.
2991 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2992 * high - the zone may be full of dirty or under-writeback pages, which this
2993 * caller can't do much about. We kick the writeback threads and take explicit
2994 * naps in the hope that some of these pages can be written. But if the
2995 * allocating task holds filesystem locks which prevent writeout this might not
2996 * work, and the allocation attempt will fail.
2998 * returns: 0, if no pages reclaimed
2999 * else, the number of pages reclaimed
3001 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3002 struct scan_control
*sc
)
3004 int initial_priority
= sc
->priority
;
3005 pg_data_t
*last_pgdat
;
3009 delayacct_freepages_start();
3011 if (!cgroup_reclaim(sc
))
3012 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3015 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3018 shrink_zones(zonelist
, sc
);
3020 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3023 if (sc
->compaction_ready
)
3027 * If we're getting trouble reclaiming, start doing
3028 * writepage even in laptop mode.
3030 if (sc
->priority
< DEF_PRIORITY
- 2)
3031 sc
->may_writepage
= 1;
3032 } while (--sc
->priority
>= 0);
3035 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3037 if (zone
->zone_pgdat
== last_pgdat
)
3039 last_pgdat
= zone
->zone_pgdat
;
3041 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3043 if (cgroup_reclaim(sc
)) {
3044 struct lruvec
*lruvec
;
3046 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3048 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3052 delayacct_freepages_end();
3054 if (sc
->nr_reclaimed
)
3055 return sc
->nr_reclaimed
;
3057 /* Aborted reclaim to try compaction? don't OOM, then */
3058 if (sc
->compaction_ready
)
3062 * We make inactive:active ratio decisions based on the node's
3063 * composition of memory, but a restrictive reclaim_idx or a
3064 * memory.low cgroup setting can exempt large amounts of
3065 * memory from reclaim. Neither of which are very common, so
3066 * instead of doing costly eligibility calculations of the
3067 * entire cgroup subtree up front, we assume the estimates are
3068 * good, and retry with forcible deactivation if that fails.
3070 if (sc
->skipped_deactivate
) {
3071 sc
->priority
= initial_priority
;
3072 sc
->force_deactivate
= 1;
3073 sc
->skipped_deactivate
= 0;
3077 /* Untapped cgroup reserves? Don't OOM, retry. */
3078 if (sc
->memcg_low_skipped
) {
3079 sc
->priority
= initial_priority
;
3080 sc
->force_deactivate
= 0;
3081 sc
->memcg_low_reclaim
= 1;
3082 sc
->memcg_low_skipped
= 0;
3089 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3092 unsigned long pfmemalloc_reserve
= 0;
3093 unsigned long free_pages
= 0;
3097 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3100 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3101 zone
= &pgdat
->node_zones
[i
];
3102 if (!managed_zone(zone
))
3105 if (!zone_reclaimable_pages(zone
))
3108 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3109 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3112 /* If there are no reserves (unexpected config) then do not throttle */
3113 if (!pfmemalloc_reserve
)
3116 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3118 /* kswapd must be awake if processes are being throttled */
3119 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3120 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
3121 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
3123 wake_up_interruptible(&pgdat
->kswapd_wait
);
3130 * Throttle direct reclaimers if backing storage is backed by the network
3131 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3132 * depleted. kswapd will continue to make progress and wake the processes
3133 * when the low watermark is reached.
3135 * Returns true if a fatal signal was delivered during throttling. If this
3136 * happens, the page allocator should not consider triggering the OOM killer.
3138 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3139 nodemask_t
*nodemask
)
3143 pg_data_t
*pgdat
= NULL
;
3146 * Kernel threads should not be throttled as they may be indirectly
3147 * responsible for cleaning pages necessary for reclaim to make forward
3148 * progress. kjournald for example may enter direct reclaim while
3149 * committing a transaction where throttling it could forcing other
3150 * processes to block on log_wait_commit().
3152 if (current
->flags
& PF_KTHREAD
)
3156 * If a fatal signal is pending, this process should not throttle.
3157 * It should return quickly so it can exit and free its memory
3159 if (fatal_signal_pending(current
))
3163 * Check if the pfmemalloc reserves are ok by finding the first node
3164 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3165 * GFP_KERNEL will be required for allocating network buffers when
3166 * swapping over the network so ZONE_HIGHMEM is unusable.
3168 * Throttling is based on the first usable node and throttled processes
3169 * wait on a queue until kswapd makes progress and wakes them. There
3170 * is an affinity then between processes waking up and where reclaim
3171 * progress has been made assuming the process wakes on the same node.
3172 * More importantly, processes running on remote nodes will not compete
3173 * for remote pfmemalloc reserves and processes on different nodes
3174 * should make reasonable progress.
3176 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3177 gfp_zone(gfp_mask
), nodemask
) {
3178 if (zone_idx(zone
) > ZONE_NORMAL
)
3181 /* Throttle based on the first usable node */
3182 pgdat
= zone
->zone_pgdat
;
3183 if (allow_direct_reclaim(pgdat
))
3188 /* If no zone was usable by the allocation flags then do not throttle */
3192 /* Account for the throttling */
3193 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3196 * If the caller cannot enter the filesystem, it's possible that it
3197 * is due to the caller holding an FS lock or performing a journal
3198 * transaction in the case of a filesystem like ext[3|4]. In this case,
3199 * it is not safe to block on pfmemalloc_wait as kswapd could be
3200 * blocked waiting on the same lock. Instead, throttle for up to a
3201 * second before continuing.
3203 if (!(gfp_mask
& __GFP_FS
)) {
3204 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3205 allow_direct_reclaim(pgdat
), HZ
);
3210 /* Throttle until kswapd wakes the process */
3211 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3212 allow_direct_reclaim(pgdat
));
3215 if (fatal_signal_pending(current
))
3222 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3223 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3225 unsigned long nr_reclaimed
;
3226 struct scan_control sc
= {
3227 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3228 .gfp_mask
= current_gfp_context(gfp_mask
),
3229 .reclaim_idx
= gfp_zone(gfp_mask
),
3231 .nodemask
= nodemask
,
3232 .priority
= DEF_PRIORITY
,
3233 .may_writepage
= !laptop_mode
,
3239 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3240 * Confirm they are large enough for max values.
3242 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3243 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3244 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3247 * Do not enter reclaim if fatal signal was delivered while throttled.
3248 * 1 is returned so that the page allocator does not OOM kill at this
3251 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3254 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3255 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3257 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3259 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3260 set_task_reclaim_state(current
, NULL
);
3262 return nr_reclaimed
;
3267 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3268 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3269 gfp_t gfp_mask
, bool noswap
,
3271 unsigned long *nr_scanned
)
3273 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3274 struct scan_control sc
= {
3275 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3276 .target_mem_cgroup
= memcg
,
3277 .may_writepage
= !laptop_mode
,
3279 .reclaim_idx
= MAX_NR_ZONES
- 1,
3280 .may_swap
= !noswap
,
3283 WARN_ON_ONCE(!current
->reclaim_state
);
3285 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3286 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3288 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3292 * NOTE: Although we can get the priority field, using it
3293 * here is not a good idea, since it limits the pages we can scan.
3294 * if we don't reclaim here, the shrink_node from balance_pgdat
3295 * will pick up pages from other mem cgroup's as well. We hack
3296 * the priority and make it zero.
3298 shrink_lruvec(lruvec
, &sc
);
3300 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3302 *nr_scanned
= sc
.nr_scanned
;
3304 return sc
.nr_reclaimed
;
3307 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3308 unsigned long nr_pages
,
3312 unsigned long nr_reclaimed
;
3313 unsigned int noreclaim_flag
;
3314 struct scan_control sc
= {
3315 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3316 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3317 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3318 .reclaim_idx
= MAX_NR_ZONES
- 1,
3319 .target_mem_cgroup
= memcg
,
3320 .priority
= DEF_PRIORITY
,
3321 .may_writepage
= !laptop_mode
,
3323 .may_swap
= may_swap
,
3326 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3327 * equal pressure on all the nodes. This is based on the assumption that
3328 * the reclaim does not bail out early.
3330 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3332 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3333 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3334 noreclaim_flag
= memalloc_noreclaim_save();
3336 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3338 memalloc_noreclaim_restore(noreclaim_flag
);
3339 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3340 set_task_reclaim_state(current
, NULL
);
3342 return nr_reclaimed
;
3346 static void age_active_anon(struct pglist_data
*pgdat
,
3347 struct scan_control
*sc
)
3349 struct mem_cgroup
*memcg
;
3350 struct lruvec
*lruvec
;
3352 if (!total_swap_pages
)
3355 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3356 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3359 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3361 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3362 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3363 sc
, LRU_ACTIVE_ANON
);
3364 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3368 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
3374 * Check for watermark boosts top-down as the higher zones
3375 * are more likely to be boosted. Both watermarks and boosts
3376 * should not be checked at the same time as reclaim would
3377 * start prematurely when there is no boosting and a lower
3380 for (i
= highest_zoneidx
; i
>= 0; i
--) {
3381 zone
= pgdat
->node_zones
+ i
;
3382 if (!managed_zone(zone
))
3385 if (zone
->watermark_boost
)
3393 * Returns true if there is an eligible zone balanced for the request order
3394 * and highest_zoneidx
3396 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3399 unsigned long mark
= -1;
3403 * Check watermarks bottom-up as lower zones are more likely to
3406 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3407 zone
= pgdat
->node_zones
+ i
;
3409 if (!managed_zone(zone
))
3412 mark
= high_wmark_pages(zone
);
3413 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
3418 * If a node has no populated zone within highest_zoneidx, it does not
3419 * need balancing by definition. This can happen if a zone-restricted
3420 * allocation tries to wake a remote kswapd.
3428 /* Clear pgdat state for congested, dirty or under writeback. */
3429 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3431 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3433 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3434 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3435 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3439 * Prepare kswapd for sleeping. This verifies that there are no processes
3440 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3442 * Returns true if kswapd is ready to sleep
3444 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
3445 int highest_zoneidx
)
3448 * The throttled processes are normally woken up in balance_pgdat() as
3449 * soon as allow_direct_reclaim() is true. But there is a potential
3450 * race between when kswapd checks the watermarks and a process gets
3451 * throttled. There is also a potential race if processes get
3452 * throttled, kswapd wakes, a large process exits thereby balancing the
3453 * zones, which causes kswapd to exit balance_pgdat() before reaching
3454 * the wake up checks. If kswapd is going to sleep, no process should
3455 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3456 * the wake up is premature, processes will wake kswapd and get
3457 * throttled again. The difference from wake ups in balance_pgdat() is
3458 * that here we are under prepare_to_wait().
3460 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3461 wake_up_all(&pgdat
->pfmemalloc_wait
);
3463 /* Hopeless node, leave it to direct reclaim */
3464 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3467 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
3468 clear_pgdat_congested(pgdat
);
3476 * kswapd shrinks a node of pages that are at or below the highest usable
3477 * zone that is currently unbalanced.
3479 * Returns true if kswapd scanned at least the requested number of pages to
3480 * reclaim or if the lack of progress was due to pages under writeback.
3481 * This is used to determine if the scanning priority needs to be raised.
3483 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3484 struct scan_control
*sc
)
3489 /* Reclaim a number of pages proportional to the number of zones */
3490 sc
->nr_to_reclaim
= 0;
3491 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3492 zone
= pgdat
->node_zones
+ z
;
3493 if (!managed_zone(zone
))
3496 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3500 * Historically care was taken to put equal pressure on all zones but
3501 * now pressure is applied based on node LRU order.
3503 shrink_node(pgdat
, sc
);
3506 * Fragmentation may mean that the system cannot be rebalanced for
3507 * high-order allocations. If twice the allocation size has been
3508 * reclaimed then recheck watermarks only at order-0 to prevent
3509 * excessive reclaim. Assume that a process requested a high-order
3510 * can direct reclaim/compact.
3512 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3515 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3519 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3520 * that are eligible for use by the caller until at least one zone is
3523 * Returns the order kswapd finished reclaiming at.
3525 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3526 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3527 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3528 * or lower is eligible for reclaim until at least one usable zone is
3531 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3534 unsigned long nr_soft_reclaimed
;
3535 unsigned long nr_soft_scanned
;
3536 unsigned long pflags
;
3537 unsigned long nr_boost_reclaim
;
3538 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
3541 struct scan_control sc
= {
3542 .gfp_mask
= GFP_KERNEL
,
3547 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3548 psi_memstall_enter(&pflags
);
3549 __fs_reclaim_acquire();
3551 count_vm_event(PAGEOUTRUN
);
3554 * Account for the reclaim boost. Note that the zone boost is left in
3555 * place so that parallel allocations that are near the watermark will
3556 * stall or direct reclaim until kswapd is finished.
3558 nr_boost_reclaim
= 0;
3559 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3560 zone
= pgdat
->node_zones
+ i
;
3561 if (!managed_zone(zone
))
3564 nr_boost_reclaim
+= zone
->watermark_boost
;
3565 zone_boosts
[i
] = zone
->watermark_boost
;
3567 boosted
= nr_boost_reclaim
;
3570 sc
.priority
= DEF_PRIORITY
;
3572 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3573 bool raise_priority
= true;
3577 sc
.reclaim_idx
= highest_zoneidx
;
3580 * If the number of buffer_heads exceeds the maximum allowed
3581 * then consider reclaiming from all zones. This has a dual
3582 * purpose -- on 64-bit systems it is expected that
3583 * buffer_heads are stripped during active rotation. On 32-bit
3584 * systems, highmem pages can pin lowmem memory and shrinking
3585 * buffers can relieve lowmem pressure. Reclaim may still not
3586 * go ahead if all eligible zones for the original allocation
3587 * request are balanced to avoid excessive reclaim from kswapd.
3589 if (buffer_heads_over_limit
) {
3590 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3591 zone
= pgdat
->node_zones
+ i
;
3592 if (!managed_zone(zone
))
3601 * If the pgdat is imbalanced then ignore boosting and preserve
3602 * the watermarks for a later time and restart. Note that the
3603 * zone watermarks will be still reset at the end of balancing
3604 * on the grounds that the normal reclaim should be enough to
3605 * re-evaluate if boosting is required when kswapd next wakes.
3607 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
3608 if (!balanced
&& nr_boost_reclaim
) {
3609 nr_boost_reclaim
= 0;
3614 * If boosting is not active then only reclaim if there are no
3615 * eligible zones. Note that sc.reclaim_idx is not used as
3616 * buffer_heads_over_limit may have adjusted it.
3618 if (!nr_boost_reclaim
&& balanced
)
3621 /* Limit the priority of boosting to avoid reclaim writeback */
3622 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
3623 raise_priority
= false;
3626 * Do not writeback or swap pages for boosted reclaim. The
3627 * intent is to relieve pressure not issue sub-optimal IO
3628 * from reclaim context. If no pages are reclaimed, the
3629 * reclaim will be aborted.
3631 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
3632 sc
.may_swap
= !nr_boost_reclaim
;
3635 * Do some background aging of the anon list, to give
3636 * pages a chance to be referenced before reclaiming. All
3637 * pages are rotated regardless of classzone as this is
3638 * about consistent aging.
3640 age_active_anon(pgdat
, &sc
);
3643 * If we're getting trouble reclaiming, start doing writepage
3644 * even in laptop mode.
3646 if (sc
.priority
< DEF_PRIORITY
- 2)
3647 sc
.may_writepage
= 1;
3649 /* Call soft limit reclaim before calling shrink_node. */
3651 nr_soft_scanned
= 0;
3652 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3653 sc
.gfp_mask
, &nr_soft_scanned
);
3654 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3657 * There should be no need to raise the scanning priority if
3658 * enough pages are already being scanned that that high
3659 * watermark would be met at 100% efficiency.
3661 if (kswapd_shrink_node(pgdat
, &sc
))
3662 raise_priority
= false;
3665 * If the low watermark is met there is no need for processes
3666 * to be throttled on pfmemalloc_wait as they should not be
3667 * able to safely make forward progress. Wake them
3669 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3670 allow_direct_reclaim(pgdat
))
3671 wake_up_all(&pgdat
->pfmemalloc_wait
);
3673 /* Check if kswapd should be suspending */
3674 __fs_reclaim_release();
3675 ret
= try_to_freeze();
3676 __fs_reclaim_acquire();
3677 if (ret
|| kthread_should_stop())
3681 * Raise priority if scanning rate is too low or there was no
3682 * progress in reclaiming pages
3684 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3685 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
3688 * If reclaim made no progress for a boost, stop reclaim as
3689 * IO cannot be queued and it could be an infinite loop in
3690 * extreme circumstances.
3692 if (nr_boost_reclaim
&& !nr_reclaimed
)
3695 if (raise_priority
|| !nr_reclaimed
)
3697 } while (sc
.priority
>= 1);
3699 if (!sc
.nr_reclaimed
)
3700 pgdat
->kswapd_failures
++;
3703 /* If reclaim was boosted, account for the reclaim done in this pass */
3705 unsigned long flags
;
3707 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3708 if (!zone_boosts
[i
])
3711 /* Increments are under the zone lock */
3712 zone
= pgdat
->node_zones
+ i
;
3713 spin_lock_irqsave(&zone
->lock
, flags
);
3714 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
3715 spin_unlock_irqrestore(&zone
->lock
, flags
);
3719 * As there is now likely space, wakeup kcompact to defragment
3722 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
3725 snapshot_refaults(NULL
, pgdat
);
3726 __fs_reclaim_release();
3727 psi_memstall_leave(&pflags
);
3728 set_task_reclaim_state(current
, NULL
);
3731 * Return the order kswapd stopped reclaiming at as
3732 * prepare_kswapd_sleep() takes it into account. If another caller
3733 * entered the allocator slow path while kswapd was awake, order will
3734 * remain at the higher level.
3740 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3741 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3742 * not a valid index then either kswapd runs for first time or kswapd couldn't
3743 * sleep after previous reclaim attempt (node is still unbalanced). In that
3744 * case return the zone index of the previous kswapd reclaim cycle.
3746 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
3747 enum zone_type prev_highest_zoneidx
)
3749 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
3751 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
3754 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3755 unsigned int highest_zoneidx
)
3760 if (freezing(current
) || kthread_should_stop())
3763 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3766 * Try to sleep for a short interval. Note that kcompactd will only be
3767 * woken if it is possible to sleep for a short interval. This is
3768 * deliberate on the assumption that if reclaim cannot keep an
3769 * eligible zone balanced that it's also unlikely that compaction will
3772 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
3774 * Compaction records what page blocks it recently failed to
3775 * isolate pages from and skips them in the future scanning.
3776 * When kswapd is going to sleep, it is reasonable to assume
3777 * that pages and compaction may succeed so reset the cache.
3779 reset_isolation_suitable(pgdat
);
3782 * We have freed the memory, now we should compact it to make
3783 * allocation of the requested order possible.
3785 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
3787 remaining
= schedule_timeout(HZ
/10);
3790 * If woken prematurely then reset kswapd_highest_zoneidx and
3791 * order. The values will either be from a wakeup request or
3792 * the previous request that slept prematurely.
3795 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
3796 kswapd_highest_zoneidx(pgdat
,
3799 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
3800 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
3803 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3804 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3808 * After a short sleep, check if it was a premature sleep. If not, then
3809 * go fully to sleep until explicitly woken up.
3812 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
3813 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3816 * vmstat counters are not perfectly accurate and the estimated
3817 * value for counters such as NR_FREE_PAGES can deviate from the
3818 * true value by nr_online_cpus * threshold. To avoid the zone
3819 * watermarks being breached while under pressure, we reduce the
3820 * per-cpu vmstat threshold while kswapd is awake and restore
3821 * them before going back to sleep.
3823 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3825 if (!kthread_should_stop())
3828 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3831 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3833 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3835 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3839 * The background pageout daemon, started as a kernel thread
3840 * from the init process.
3842 * This basically trickles out pages so that we have _some_
3843 * free memory available even if there is no other activity
3844 * that frees anything up. This is needed for things like routing
3845 * etc, where we otherwise might have all activity going on in
3846 * asynchronous contexts that cannot page things out.
3848 * If there are applications that are active memory-allocators
3849 * (most normal use), this basically shouldn't matter.
3851 static int kswapd(void *p
)
3853 unsigned int alloc_order
, reclaim_order
;
3854 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
3855 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3856 struct task_struct
*tsk
= current
;
3857 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3859 if (!cpumask_empty(cpumask
))
3860 set_cpus_allowed_ptr(tsk
, cpumask
);
3863 * Tell the memory management that we're a "memory allocator",
3864 * and that if we need more memory we should get access to it
3865 * regardless (see "__alloc_pages()"). "kswapd" should
3866 * never get caught in the normal page freeing logic.
3868 * (Kswapd normally doesn't need memory anyway, but sometimes
3869 * you need a small amount of memory in order to be able to
3870 * page out something else, and this flag essentially protects
3871 * us from recursively trying to free more memory as we're
3872 * trying to free the first piece of memory in the first place).
3874 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3877 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3878 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
3882 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3883 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
3887 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3890 /* Read the new order and highest_zoneidx */
3891 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3892 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
3894 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3895 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
3897 ret
= try_to_freeze();
3898 if (kthread_should_stop())
3902 * We can speed up thawing tasks if we don't call balance_pgdat
3903 * after returning from the refrigerator
3909 * Reclaim begins at the requested order but if a high-order
3910 * reclaim fails then kswapd falls back to reclaiming for
3911 * order-0. If that happens, kswapd will consider sleeping
3912 * for the order it finished reclaiming at (reclaim_order)
3913 * but kcompactd is woken to compact for the original
3914 * request (alloc_order).
3916 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
3918 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
3920 if (reclaim_order
< alloc_order
)
3921 goto kswapd_try_sleep
;
3924 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3930 * A zone is low on free memory or too fragmented for high-order memory. If
3931 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3932 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3933 * has failed or is not needed, still wake up kcompactd if only compaction is
3936 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3937 enum zone_type highest_zoneidx
)
3940 enum zone_type curr_idx
;
3942 if (!managed_zone(zone
))
3945 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3948 pgdat
= zone
->zone_pgdat
;
3949 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
3951 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
3952 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
3954 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
3955 WRITE_ONCE(pgdat
->kswapd_order
, order
);
3957 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3960 /* Hopeless node, leave it to direct reclaim if possible */
3961 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3962 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
3963 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
3965 * There may be plenty of free memory available, but it's too
3966 * fragmented for high-order allocations. Wake up kcompactd
3967 * and rely on compaction_suitable() to determine if it's
3968 * needed. If it fails, it will defer subsequent attempts to
3969 * ratelimit its work.
3971 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3972 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
3976 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
3978 wake_up_interruptible(&pgdat
->kswapd_wait
);
3981 #ifdef CONFIG_HIBERNATION
3983 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3986 * Rather than trying to age LRUs the aim is to preserve the overall
3987 * LRU order by reclaiming preferentially
3988 * inactive > active > active referenced > active mapped
3990 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3992 struct scan_control sc
= {
3993 .nr_to_reclaim
= nr_to_reclaim
,
3994 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3995 .reclaim_idx
= MAX_NR_ZONES
- 1,
3996 .priority
= DEF_PRIORITY
,
4000 .hibernation_mode
= 1,
4002 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4003 unsigned long nr_reclaimed
;
4004 unsigned int noreclaim_flag
;
4006 fs_reclaim_acquire(sc
.gfp_mask
);
4007 noreclaim_flag
= memalloc_noreclaim_save();
4008 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4010 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4012 set_task_reclaim_state(current
, NULL
);
4013 memalloc_noreclaim_restore(noreclaim_flag
);
4014 fs_reclaim_release(sc
.gfp_mask
);
4016 return nr_reclaimed
;
4018 #endif /* CONFIG_HIBERNATION */
4021 * This kswapd start function will be called by init and node-hot-add.
4022 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4024 int kswapd_run(int nid
)
4026 pg_data_t
*pgdat
= NODE_DATA(nid
);
4032 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4033 if (IS_ERR(pgdat
->kswapd
)) {
4034 /* failure at boot is fatal */
4035 BUG_ON(system_state
< SYSTEM_RUNNING
);
4036 pr_err("Failed to start kswapd on node %d\n", nid
);
4037 ret
= PTR_ERR(pgdat
->kswapd
);
4038 pgdat
->kswapd
= NULL
;
4044 * Called by memory hotplug when all memory in a node is offlined. Caller must
4045 * hold mem_hotplug_begin/end().
4047 void kswapd_stop(int nid
)
4049 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4052 kthread_stop(kswapd
);
4053 NODE_DATA(nid
)->kswapd
= NULL
;
4057 static int __init
kswapd_init(void)
4062 for_each_node_state(nid
, N_MEMORY
)
4067 module_init(kswapd_init
)
4073 * If non-zero call node_reclaim when the number of free pages falls below
4076 int node_reclaim_mode __read_mostly
;
4078 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4079 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4082 * Priority for NODE_RECLAIM. This determines the fraction of pages
4083 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4086 #define NODE_RECLAIM_PRIORITY 4
4089 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4092 int sysctl_min_unmapped_ratio
= 1;
4095 * If the number of slab pages in a zone grows beyond this percentage then
4096 * slab reclaim needs to occur.
4098 int sysctl_min_slab_ratio
= 5;
4100 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4102 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4103 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4104 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4107 * It's possible for there to be more file mapped pages than
4108 * accounted for by the pages on the file LRU lists because
4109 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4111 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4114 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4115 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4117 unsigned long nr_pagecache_reclaimable
;
4118 unsigned long delta
= 0;
4121 * If RECLAIM_UNMAP is set, then all file pages are considered
4122 * potentially reclaimable. Otherwise, we have to worry about
4123 * pages like swapcache and node_unmapped_file_pages() provides
4126 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4127 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4129 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4131 /* If we can't clean pages, remove dirty pages from consideration */
4132 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4133 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4135 /* Watch for any possible underflows due to delta */
4136 if (unlikely(delta
> nr_pagecache_reclaimable
))
4137 delta
= nr_pagecache_reclaimable
;
4139 return nr_pagecache_reclaimable
- delta
;
4143 * Try to free up some pages from this node through reclaim.
4145 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4147 /* Minimum pages needed in order to stay on node */
4148 const unsigned long nr_pages
= 1 << order
;
4149 struct task_struct
*p
= current
;
4150 unsigned int noreclaim_flag
;
4151 struct scan_control sc
= {
4152 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4153 .gfp_mask
= current_gfp_context(gfp_mask
),
4155 .priority
= NODE_RECLAIM_PRIORITY
,
4156 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4157 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4159 .reclaim_idx
= gfp_zone(gfp_mask
),
4162 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4166 fs_reclaim_acquire(sc
.gfp_mask
);
4168 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4169 * and we also need to be able to write out pages for RECLAIM_WRITE
4170 * and RECLAIM_UNMAP.
4172 noreclaim_flag
= memalloc_noreclaim_save();
4173 p
->flags
|= PF_SWAPWRITE
;
4174 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4176 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4178 * Free memory by calling shrink node with increasing
4179 * priorities until we have enough memory freed.
4182 shrink_node(pgdat
, &sc
);
4183 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4186 set_task_reclaim_state(p
, NULL
);
4187 current
->flags
&= ~PF_SWAPWRITE
;
4188 memalloc_noreclaim_restore(noreclaim_flag
);
4189 fs_reclaim_release(sc
.gfp_mask
);
4191 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4193 return sc
.nr_reclaimed
>= nr_pages
;
4196 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4201 * Node reclaim reclaims unmapped file backed pages and
4202 * slab pages if we are over the defined limits.
4204 * A small portion of unmapped file backed pages is needed for
4205 * file I/O otherwise pages read by file I/O will be immediately
4206 * thrown out if the node is overallocated. So we do not reclaim
4207 * if less than a specified percentage of the node is used by
4208 * unmapped file backed pages.
4210 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4211 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
4212 pgdat
->min_slab_pages
)
4213 return NODE_RECLAIM_FULL
;
4216 * Do not scan if the allocation should not be delayed.
4218 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4219 return NODE_RECLAIM_NOSCAN
;
4222 * Only run node reclaim on the local node or on nodes that do not
4223 * have associated processors. This will favor the local processor
4224 * over remote processors and spread off node memory allocations
4225 * as wide as possible.
4227 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4228 return NODE_RECLAIM_NOSCAN
;
4230 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4231 return NODE_RECLAIM_NOSCAN
;
4233 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4234 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4237 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4244 * check_move_unevictable_pages - check pages for evictability and move to
4245 * appropriate zone lru list
4246 * @pvec: pagevec with lru pages to check
4248 * Checks pages for evictability, if an evictable page is in the unevictable
4249 * lru list, moves it to the appropriate evictable lru list. This function
4250 * should be only used for lru pages.
4252 void check_move_unevictable_pages(struct pagevec
*pvec
)
4254 struct lruvec
*lruvec
;
4255 struct pglist_data
*pgdat
= NULL
;
4260 for (i
= 0; i
< pvec
->nr
; i
++) {
4261 struct page
*page
= pvec
->pages
[i
];
4262 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4265 if (pagepgdat
!= pgdat
) {
4267 spin_unlock_irq(&pgdat
->lru_lock
);
4269 spin_lock_irq(&pgdat
->lru_lock
);
4271 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4273 if (!PageLRU(page
) || !PageUnevictable(page
))
4276 if (page_evictable(page
)) {
4277 enum lru_list lru
= page_lru_base_type(page
);
4279 VM_BUG_ON_PAGE(PageActive(page
), page
);
4280 ClearPageUnevictable(page
);
4281 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4282 add_page_to_lru_list(page
, lruvec
, lru
);
4288 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4289 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4290 spin_unlock_irq(&pgdat
->lru_lock
);
4293 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);