1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/migrate.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/pagevec.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52 #include <linux/psi.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
57 #include <linux/swapops.h>
58 #include <linux/balloon_compaction.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/vmscan.h>
66 /* How many pages shrink_list() should reclaim */
67 unsigned long nr_to_reclaim
;
70 * Nodemask of nodes allowed by the caller. If NULL, all nodes
76 * The memory cgroup that hit its limit and as a result is the
77 * primary target of this reclaim invocation.
79 struct mem_cgroup
*target_mem_cgroup
;
82 * Scan pressure balancing between anon and file LRUs
84 unsigned long anon_cost
;
85 unsigned long file_cost
;
87 /* Can active pages be deactivated as part of reclaim? */
88 #define DEACTIVATE_ANON 1
89 #define DEACTIVATE_FILE 2
90 unsigned int may_deactivate
:2;
91 unsigned int force_deactivate
:1;
92 unsigned int skipped_deactivate
:1;
94 /* Writepage batching in laptop mode; RECLAIM_WRITE */
95 unsigned int may_writepage
:1;
97 /* Can mapped pages be reclaimed? */
98 unsigned int may_unmap
:1;
100 /* Can pages be swapped as part of reclaim? */
101 unsigned int may_swap
:1;
104 * Cgroup memory below memory.low is protected as long as we
105 * don't threaten to OOM. If any cgroup is reclaimed at
106 * reduced force or passed over entirely due to its memory.low
107 * setting (memcg_low_skipped), and nothing is reclaimed as a
108 * result, then go back for one more cycle that reclaims the protected
109 * memory (memcg_low_reclaim) to avert OOM.
111 unsigned int memcg_low_reclaim
:1;
112 unsigned int memcg_low_skipped
:1;
114 unsigned int hibernation_mode
:1;
116 /* One of the zones is ready for compaction */
117 unsigned int compaction_ready
:1;
119 /* There is easily reclaimable cold cache in the current node */
120 unsigned int cache_trim_mode
:1;
122 /* The file pages on the current node are dangerously low */
123 unsigned int file_is_tiny
:1;
125 /* Always discard instead of demoting to lower tier memory */
126 unsigned int no_demotion
:1;
128 /* Allocation order */
131 /* Scan (total_size >> priority) pages at once */
134 /* The highest zone to isolate pages for reclaim from */
137 /* This context's GFP mask */
140 /* Incremented by the number of inactive pages that were scanned */
141 unsigned long nr_scanned
;
143 /* Number of pages freed so far during a call to shrink_zones() */
144 unsigned long nr_reclaimed
;
148 unsigned int unqueued_dirty
;
149 unsigned int congested
;
150 unsigned int writeback
;
151 unsigned int immediate
;
152 unsigned int file_taken
;
156 /* for recording the reclaimed slab by now */
157 struct reclaim_state reclaim_state
;
160 #ifdef ARCH_HAS_PREFETCHW
161 #define prefetchw_prev_lru_page(_page, _base, _field) \
163 if ((_page)->lru.prev != _base) { \
166 prev = lru_to_page(&(_page->lru)); \
167 prefetchw(&prev->_field); \
171 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
175 * From 0 .. 200. Higher means more swappy.
177 int vm_swappiness
= 60;
179 static void set_task_reclaim_state(struct task_struct
*task
,
180 struct reclaim_state
*rs
)
182 /* Check for an overwrite */
183 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
185 /* Check for the nulling of an already-nulled member */
186 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
188 task
->reclaim_state
= rs
;
191 static LIST_HEAD(shrinker_list
);
192 static DECLARE_RWSEM(shrinker_rwsem
);
195 static int shrinker_nr_max
;
197 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
198 static inline int shrinker_map_size(int nr_items
)
200 return (DIV_ROUND_UP(nr_items
, BITS_PER_LONG
) * sizeof(unsigned long));
203 static inline int shrinker_defer_size(int nr_items
)
205 return (round_up(nr_items
, BITS_PER_LONG
) * sizeof(atomic_long_t
));
208 static struct shrinker_info
*shrinker_info_protected(struct mem_cgroup
*memcg
,
211 return rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_info
,
212 lockdep_is_held(&shrinker_rwsem
));
215 static int expand_one_shrinker_info(struct mem_cgroup
*memcg
,
216 int map_size
, int defer_size
,
217 int old_map_size
, int old_defer_size
)
219 struct shrinker_info
*new, *old
;
220 struct mem_cgroup_per_node
*pn
;
222 int size
= map_size
+ defer_size
;
225 pn
= memcg
->nodeinfo
[nid
];
226 old
= shrinker_info_protected(memcg
, nid
);
227 /* Not yet online memcg */
231 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
235 new->nr_deferred
= (atomic_long_t
*)(new + 1);
236 new->map
= (void *)new->nr_deferred
+ defer_size
;
238 /* map: set all old bits, clear all new bits */
239 memset(new->map
, (int)0xff, old_map_size
);
240 memset((void *)new->map
+ old_map_size
, 0, map_size
- old_map_size
);
241 /* nr_deferred: copy old values, clear all new values */
242 memcpy(new->nr_deferred
, old
->nr_deferred
, old_defer_size
);
243 memset((void *)new->nr_deferred
+ old_defer_size
, 0,
244 defer_size
- old_defer_size
);
246 rcu_assign_pointer(pn
->shrinker_info
, new);
247 kvfree_rcu(old
, rcu
);
253 void free_shrinker_info(struct mem_cgroup
*memcg
)
255 struct mem_cgroup_per_node
*pn
;
256 struct shrinker_info
*info
;
260 pn
= memcg
->nodeinfo
[nid
];
261 info
= rcu_dereference_protected(pn
->shrinker_info
, true);
263 rcu_assign_pointer(pn
->shrinker_info
, NULL
);
267 int alloc_shrinker_info(struct mem_cgroup
*memcg
)
269 struct shrinker_info
*info
;
270 int nid
, size
, ret
= 0;
271 int map_size
, defer_size
= 0;
273 down_write(&shrinker_rwsem
);
274 map_size
= shrinker_map_size(shrinker_nr_max
);
275 defer_size
= shrinker_defer_size(shrinker_nr_max
);
276 size
= map_size
+ defer_size
;
278 info
= kvzalloc_node(sizeof(*info
) + size
, GFP_KERNEL
, nid
);
280 free_shrinker_info(memcg
);
284 info
->nr_deferred
= (atomic_long_t
*)(info
+ 1);
285 info
->map
= (void *)info
->nr_deferred
+ defer_size
;
286 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_info
, info
);
288 up_write(&shrinker_rwsem
);
293 static inline bool need_expand(int nr_max
)
295 return round_up(nr_max
, BITS_PER_LONG
) >
296 round_up(shrinker_nr_max
, BITS_PER_LONG
);
299 static int expand_shrinker_info(int new_id
)
302 int new_nr_max
= new_id
+ 1;
303 int map_size
, defer_size
= 0;
304 int old_map_size
, old_defer_size
= 0;
305 struct mem_cgroup
*memcg
;
307 if (!need_expand(new_nr_max
))
310 if (!root_mem_cgroup
)
313 lockdep_assert_held(&shrinker_rwsem
);
315 map_size
= shrinker_map_size(new_nr_max
);
316 defer_size
= shrinker_defer_size(new_nr_max
);
317 old_map_size
= shrinker_map_size(shrinker_nr_max
);
318 old_defer_size
= shrinker_defer_size(shrinker_nr_max
);
320 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
322 ret
= expand_one_shrinker_info(memcg
, map_size
, defer_size
,
323 old_map_size
, old_defer_size
);
325 mem_cgroup_iter_break(NULL
, memcg
);
328 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
331 shrinker_nr_max
= new_nr_max
;
336 void set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
338 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
339 struct shrinker_info
*info
;
342 info
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_info
);
343 /* Pairs with smp mb in shrink_slab() */
344 smp_mb__before_atomic();
345 set_bit(shrinker_id
, info
->map
);
350 static DEFINE_IDR(shrinker_idr
);
352 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
354 int id
, ret
= -ENOMEM
;
356 if (mem_cgroup_disabled())
359 down_write(&shrinker_rwsem
);
360 /* This may call shrinker, so it must use down_read_trylock() */
361 id
= idr_alloc(&shrinker_idr
, shrinker
, 0, 0, GFP_KERNEL
);
365 if (id
>= shrinker_nr_max
) {
366 if (expand_shrinker_info(id
)) {
367 idr_remove(&shrinker_idr
, id
);
374 up_write(&shrinker_rwsem
);
378 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
380 int id
= shrinker
->id
;
384 lockdep_assert_held(&shrinker_rwsem
);
386 idr_remove(&shrinker_idr
, id
);
389 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
390 struct mem_cgroup
*memcg
)
392 struct shrinker_info
*info
;
394 info
= shrinker_info_protected(memcg
, nid
);
395 return atomic_long_xchg(&info
->nr_deferred
[shrinker
->id
], 0);
398 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
399 struct mem_cgroup
*memcg
)
401 struct shrinker_info
*info
;
403 info
= shrinker_info_protected(memcg
, nid
);
404 return atomic_long_add_return(nr
, &info
->nr_deferred
[shrinker
->id
]);
407 void reparent_shrinker_deferred(struct mem_cgroup
*memcg
)
411 struct mem_cgroup
*parent
;
412 struct shrinker_info
*child_info
, *parent_info
;
414 parent
= parent_mem_cgroup(memcg
);
416 parent
= root_mem_cgroup
;
418 /* Prevent from concurrent shrinker_info expand */
419 down_read(&shrinker_rwsem
);
421 child_info
= shrinker_info_protected(memcg
, nid
);
422 parent_info
= shrinker_info_protected(parent
, nid
);
423 for (i
= 0; i
< shrinker_nr_max
; i
++) {
424 nr
= atomic_long_read(&child_info
->nr_deferred
[i
]);
425 atomic_long_add(nr
, &parent_info
->nr_deferred
[i
]);
428 up_read(&shrinker_rwsem
);
431 static bool cgroup_reclaim(struct scan_control
*sc
)
433 return sc
->target_mem_cgroup
;
437 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
438 * @sc: scan_control in question
440 * The normal page dirty throttling mechanism in balance_dirty_pages() is
441 * completely broken with the legacy memcg and direct stalling in
442 * shrink_page_list() is used for throttling instead, which lacks all the
443 * niceties such as fairness, adaptive pausing, bandwidth proportional
444 * allocation and configurability.
446 * This function tests whether the vmscan currently in progress can assume
447 * that the normal dirty throttling mechanism is operational.
449 static bool writeback_throttling_sane(struct scan_control
*sc
)
451 if (!cgroup_reclaim(sc
))
453 #ifdef CONFIG_CGROUP_WRITEBACK
454 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
460 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
465 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
469 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
470 struct mem_cgroup
*memcg
)
475 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
476 struct mem_cgroup
*memcg
)
481 static bool cgroup_reclaim(struct scan_control
*sc
)
486 static bool writeback_throttling_sane(struct scan_control
*sc
)
492 static long xchg_nr_deferred(struct shrinker
*shrinker
,
493 struct shrink_control
*sc
)
497 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
501 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
502 return xchg_nr_deferred_memcg(nid
, shrinker
,
505 return atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
509 static long add_nr_deferred(long nr
, struct shrinker
*shrinker
,
510 struct shrink_control
*sc
)
514 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
518 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
519 return add_nr_deferred_memcg(nr
, nid
, shrinker
,
522 return atomic_long_add_return(nr
, &shrinker
->nr_deferred
[nid
]);
525 static bool can_demote(int nid
, struct scan_control
*sc
)
527 if (!numa_demotion_enabled
)
532 /* It is pointless to do demotion in memcg reclaim */
533 if (cgroup_reclaim(sc
))
536 if (next_demotion_node(nid
) == NUMA_NO_NODE
)
542 static inline bool can_reclaim_anon_pages(struct mem_cgroup
*memcg
,
544 struct scan_control
*sc
)
548 * For non-memcg reclaim, is there
549 * space in any swap device?
551 if (get_nr_swap_pages() > 0)
554 /* Is the memcg below its swap limit? */
555 if (mem_cgroup_get_nr_swap_pages(memcg
) > 0)
560 * The page can not be swapped.
562 * Can it be reclaimed from this node via demotion?
564 return can_demote(nid
, sc
);
568 * This misses isolated pages which are not accounted for to save counters.
569 * As the data only determines if reclaim or compaction continues, it is
570 * not expected that isolated pages will be a dominating factor.
572 unsigned long zone_reclaimable_pages(struct zone
*zone
)
576 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
577 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
578 if (can_reclaim_anon_pages(NULL
, zone_to_nid(zone
), NULL
))
579 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
580 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
586 * lruvec_lru_size - Returns the number of pages on the given LRU list.
587 * @lruvec: lru vector
589 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
591 static unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
594 unsigned long size
= 0;
597 for (zid
= 0; zid
<= zone_idx
&& zid
< MAX_NR_ZONES
; zid
++) {
598 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
600 if (!managed_zone(zone
))
603 if (!mem_cgroup_disabled())
604 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
606 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
612 * Add a shrinker callback to be called from the vm.
614 int prealloc_shrinker(struct shrinker
*shrinker
)
619 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
620 err
= prealloc_memcg_shrinker(shrinker
);
624 shrinker
->flags
&= ~SHRINKER_MEMCG_AWARE
;
627 size
= sizeof(*shrinker
->nr_deferred
);
628 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
631 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
632 if (!shrinker
->nr_deferred
)
638 void free_prealloced_shrinker(struct shrinker
*shrinker
)
640 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
641 down_write(&shrinker_rwsem
);
642 unregister_memcg_shrinker(shrinker
);
643 up_write(&shrinker_rwsem
);
647 kfree(shrinker
->nr_deferred
);
648 shrinker
->nr_deferred
= NULL
;
651 void register_shrinker_prepared(struct shrinker
*shrinker
)
653 down_write(&shrinker_rwsem
);
654 list_add_tail(&shrinker
->list
, &shrinker_list
);
655 shrinker
->flags
|= SHRINKER_REGISTERED
;
656 up_write(&shrinker_rwsem
);
659 int register_shrinker(struct shrinker
*shrinker
)
661 int err
= prealloc_shrinker(shrinker
);
665 register_shrinker_prepared(shrinker
);
668 EXPORT_SYMBOL(register_shrinker
);
673 void unregister_shrinker(struct shrinker
*shrinker
)
675 if (!(shrinker
->flags
& SHRINKER_REGISTERED
))
678 down_write(&shrinker_rwsem
);
679 list_del(&shrinker
->list
);
680 shrinker
->flags
&= ~SHRINKER_REGISTERED
;
681 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
682 unregister_memcg_shrinker(shrinker
);
683 up_write(&shrinker_rwsem
);
685 kfree(shrinker
->nr_deferred
);
686 shrinker
->nr_deferred
= NULL
;
688 EXPORT_SYMBOL(unregister_shrinker
);
691 * synchronize_shrinkers - Wait for all running shrinkers to complete.
693 * This is equivalent to calling unregister_shrink() and register_shrinker(),
694 * but atomically and with less overhead. This is useful to guarantee that all
695 * shrinker invocations have seen an update, before freeing memory, similar to
698 void synchronize_shrinkers(void)
700 down_write(&shrinker_rwsem
);
701 up_write(&shrinker_rwsem
);
703 EXPORT_SYMBOL(synchronize_shrinkers
);
705 #define SHRINK_BATCH 128
707 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
708 struct shrinker
*shrinker
, int priority
)
710 unsigned long freed
= 0;
711 unsigned long long delta
;
716 long batch_size
= shrinker
->batch
? shrinker
->batch
718 long scanned
= 0, next_deferred
;
720 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
721 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
725 * copy the current shrinker scan count into a local variable
726 * and zero it so that other concurrent shrinker invocations
727 * don't also do this scanning work.
729 nr
= xchg_nr_deferred(shrinker
, shrinkctl
);
731 if (shrinker
->seeks
) {
732 delta
= freeable
>> priority
;
734 do_div(delta
, shrinker
->seeks
);
737 * These objects don't require any IO to create. Trim
738 * them aggressively under memory pressure to keep
739 * them from causing refetches in the IO caches.
741 delta
= freeable
/ 2;
744 total_scan
= nr
>> priority
;
746 total_scan
= min(total_scan
, (2 * freeable
));
748 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
749 freeable
, delta
, total_scan
, priority
);
752 * Normally, we should not scan less than batch_size objects in one
753 * pass to avoid too frequent shrinker calls, but if the slab has less
754 * than batch_size objects in total and we are really tight on memory,
755 * we will try to reclaim all available objects, otherwise we can end
756 * up failing allocations although there are plenty of reclaimable
757 * objects spread over several slabs with usage less than the
760 * We detect the "tight on memory" situations by looking at the total
761 * number of objects we want to scan (total_scan). If it is greater
762 * than the total number of objects on slab (freeable), we must be
763 * scanning at high prio and therefore should try to reclaim as much as
766 while (total_scan
>= batch_size
||
767 total_scan
>= freeable
) {
769 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
771 shrinkctl
->nr_to_scan
= nr_to_scan
;
772 shrinkctl
->nr_scanned
= nr_to_scan
;
773 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
774 if (ret
== SHRINK_STOP
)
778 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
779 total_scan
-= shrinkctl
->nr_scanned
;
780 scanned
+= shrinkctl
->nr_scanned
;
786 * The deferred work is increased by any new work (delta) that wasn't
787 * done, decreased by old deferred work that was done now.
789 * And it is capped to two times of the freeable items.
791 next_deferred
= max_t(long, (nr
+ delta
- scanned
), 0);
792 next_deferred
= min(next_deferred
, (2 * freeable
));
795 * move the unused scan count back into the shrinker in a
796 * manner that handles concurrent updates.
798 new_nr
= add_nr_deferred(next_deferred
, shrinker
, shrinkctl
);
800 trace_mm_shrink_slab_end(shrinker
, shrinkctl
->nid
, freed
, nr
, new_nr
, total_scan
);
805 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
806 struct mem_cgroup
*memcg
, int priority
)
808 struct shrinker_info
*info
;
809 unsigned long ret
, freed
= 0;
812 if (!mem_cgroup_online(memcg
))
815 if (!down_read_trylock(&shrinker_rwsem
))
818 info
= shrinker_info_protected(memcg
, nid
);
822 for_each_set_bit(i
, info
->map
, shrinker_nr_max
) {
823 struct shrink_control sc
= {
824 .gfp_mask
= gfp_mask
,
828 struct shrinker
*shrinker
;
830 shrinker
= idr_find(&shrinker_idr
, i
);
831 if (unlikely(!shrinker
|| !(shrinker
->flags
& SHRINKER_REGISTERED
))) {
833 clear_bit(i
, info
->map
);
837 /* Call non-slab shrinkers even though kmem is disabled */
838 if (!memcg_kmem_enabled() &&
839 !(shrinker
->flags
& SHRINKER_NONSLAB
))
842 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
843 if (ret
== SHRINK_EMPTY
) {
844 clear_bit(i
, info
->map
);
846 * After the shrinker reported that it had no objects to
847 * free, but before we cleared the corresponding bit in
848 * the memcg shrinker map, a new object might have been
849 * added. To make sure, we have the bit set in this
850 * case, we invoke the shrinker one more time and reset
851 * the bit if it reports that it is not empty anymore.
852 * The memory barrier here pairs with the barrier in
853 * set_shrinker_bit():
855 * list_lru_add() shrink_slab_memcg()
856 * list_add_tail() clear_bit()
858 * set_bit() do_shrink_slab()
860 smp_mb__after_atomic();
861 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
862 if (ret
== SHRINK_EMPTY
)
865 set_shrinker_bit(memcg
, nid
, i
);
869 if (rwsem_is_contended(&shrinker_rwsem
)) {
875 up_read(&shrinker_rwsem
);
878 #else /* CONFIG_MEMCG */
879 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
880 struct mem_cgroup
*memcg
, int priority
)
884 #endif /* CONFIG_MEMCG */
887 * shrink_slab - shrink slab caches
888 * @gfp_mask: allocation context
889 * @nid: node whose slab caches to target
890 * @memcg: memory cgroup whose slab caches to target
891 * @priority: the reclaim priority
893 * Call the shrink functions to age shrinkable caches.
895 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
896 * unaware shrinkers will receive a node id of 0 instead.
898 * @memcg specifies the memory cgroup to target. Unaware shrinkers
899 * are called only if it is the root cgroup.
901 * @priority is sc->priority, we take the number of objects and >> by priority
902 * in order to get the scan target.
904 * Returns the number of reclaimed slab objects.
906 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
907 struct mem_cgroup
*memcg
,
910 unsigned long ret
, freed
= 0;
911 struct shrinker
*shrinker
;
914 * The root memcg might be allocated even though memcg is disabled
915 * via "cgroup_disable=memory" boot parameter. This could make
916 * mem_cgroup_is_root() return false, then just run memcg slab
917 * shrink, but skip global shrink. This may result in premature
920 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
921 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
923 if (!down_read_trylock(&shrinker_rwsem
))
926 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
927 struct shrink_control sc
= {
928 .gfp_mask
= gfp_mask
,
933 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
934 if (ret
== SHRINK_EMPTY
)
938 * Bail out if someone want to register a new shrinker to
939 * prevent the registration from being stalled for long periods
940 * by parallel ongoing shrinking.
942 if (rwsem_is_contended(&shrinker_rwsem
)) {
948 up_read(&shrinker_rwsem
);
954 static void drop_slab_node(int nid
)
960 struct mem_cgroup
*memcg
= NULL
;
962 if (fatal_signal_pending(current
))
966 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
968 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
969 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
970 } while ((freed
>> shift
++) > 1);
977 for_each_online_node(nid
)
981 static inline int is_page_cache_freeable(struct page
*page
)
984 * A freeable page cache page is referenced only by the caller
985 * that isolated the page, the page cache and optional buffer
986 * heads at page->private.
988 int page_cache_pins
= thp_nr_pages(page
);
989 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
992 static int may_write_to_inode(struct inode
*inode
)
994 if (current
->flags
& PF_SWAPWRITE
)
996 if (!inode_write_congested(inode
))
998 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
1004 * We detected a synchronous write error writing a page out. Probably
1005 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1006 * fsync(), msync() or close().
1008 * The tricky part is that after writepage we cannot touch the mapping: nothing
1009 * prevents it from being freed up. But we have a ref on the page and once
1010 * that page is locked, the mapping is pinned.
1012 * We're allowed to run sleeping lock_page() here because we know the caller has
1015 static void handle_write_error(struct address_space
*mapping
,
1016 struct page
*page
, int error
)
1019 if (page_mapping(page
) == mapping
)
1020 mapping_set_error(mapping
, error
);
1024 static bool skip_throttle_noprogress(pg_data_t
*pgdat
)
1026 int reclaimable
= 0, write_pending
= 0;
1030 * If kswapd is disabled, reschedule if necessary but do not
1031 * throttle as the system is likely near OOM.
1033 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
1037 * If there are a lot of dirty/writeback pages then do not
1038 * throttle as throttling will occur when the pages cycle
1039 * towards the end of the LRU if still under writeback.
1041 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1042 struct zone
*zone
= pgdat
->node_zones
+ i
;
1044 if (!populated_zone(zone
))
1047 reclaimable
+= zone_reclaimable_pages(zone
);
1048 write_pending
+= zone_page_state_snapshot(zone
,
1049 NR_ZONE_WRITE_PENDING
);
1051 if (2 * write_pending
<= reclaimable
)
1057 void reclaim_throttle(pg_data_t
*pgdat
, enum vmscan_throttle_state reason
)
1059 wait_queue_head_t
*wqh
= &pgdat
->reclaim_wait
[reason
];
1064 * Do not throttle IO workers, kthreads other than kswapd or
1065 * workqueues. They may be required for reclaim to make
1066 * forward progress (e.g. journalling workqueues or kthreads).
1068 if (!current_is_kswapd() &&
1069 current
->flags
& (PF_IO_WORKER
|PF_KTHREAD
)) {
1075 * These figures are pulled out of thin air.
1076 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1077 * parallel reclaimers which is a short-lived event so the timeout is
1078 * short. Failing to make progress or waiting on writeback are
1079 * potentially long-lived events so use a longer timeout. This is shaky
1080 * logic as a failure to make progress could be due to anything from
1081 * writeback to a slow device to excessive references pages at the tail
1082 * of the inactive LRU.
1085 case VMSCAN_THROTTLE_WRITEBACK
:
1088 if (atomic_inc_return(&pgdat
->nr_writeback_throttled
) == 1) {
1089 WRITE_ONCE(pgdat
->nr_reclaim_start
,
1090 node_page_state(pgdat
, NR_THROTTLED_WRITTEN
));
1094 case VMSCAN_THROTTLE_CONGESTED
:
1096 case VMSCAN_THROTTLE_NOPROGRESS
:
1097 if (skip_throttle_noprogress(pgdat
)) {
1105 case VMSCAN_THROTTLE_ISOLATED
:
1114 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1115 ret
= schedule_timeout(timeout
);
1116 finish_wait(wqh
, &wait
);
1118 if (reason
== VMSCAN_THROTTLE_WRITEBACK
)
1119 atomic_dec(&pgdat
->nr_writeback_throttled
);
1121 trace_mm_vmscan_throttled(pgdat
->node_id
, jiffies_to_usecs(timeout
),
1122 jiffies_to_usecs(timeout
- ret
),
1127 * Account for pages written if tasks are throttled waiting on dirty
1128 * pages to clean. If enough pages have been cleaned since throttling
1129 * started then wakeup the throttled tasks.
1131 void __acct_reclaim_writeback(pg_data_t
*pgdat
, struct folio
*folio
,
1134 unsigned long nr_written
;
1136 node_stat_add_folio(folio
, NR_THROTTLED_WRITTEN
);
1139 * This is an inaccurate read as the per-cpu deltas may not
1140 * be synchronised. However, given that the system is
1141 * writeback throttled, it is not worth taking the penalty
1142 * of getting an accurate count. At worst, the throttle
1143 * timeout guarantees forward progress.
1145 nr_written
= node_page_state(pgdat
, NR_THROTTLED_WRITTEN
) -
1146 READ_ONCE(pgdat
->nr_reclaim_start
);
1148 if (nr_written
> SWAP_CLUSTER_MAX
* nr_throttled
)
1149 wake_up(&pgdat
->reclaim_wait
[VMSCAN_THROTTLE_WRITEBACK
]);
1152 /* possible outcome of pageout() */
1154 /* failed to write page out, page is locked */
1156 /* move page to the active list, page is locked */
1158 /* page has been sent to the disk successfully, page is unlocked */
1160 /* page is clean and locked */
1165 * pageout is called by shrink_page_list() for each dirty page.
1166 * Calls ->writepage().
1168 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
1171 * If the page is dirty, only perform writeback if that write
1172 * will be non-blocking. To prevent this allocation from being
1173 * stalled by pagecache activity. But note that there may be
1174 * stalls if we need to run get_block(). We could test
1175 * PagePrivate for that.
1177 * If this process is currently in __generic_file_write_iter() against
1178 * this page's queue, we can perform writeback even if that
1181 * If the page is swapcache, write it back even if that would
1182 * block, for some throttling. This happens by accident, because
1183 * swap_backing_dev_info is bust: it doesn't reflect the
1184 * congestion state of the swapdevs. Easy to fix, if needed.
1186 if (!is_page_cache_freeable(page
))
1190 * Some data journaling orphaned pages can have
1191 * page->mapping == NULL while being dirty with clean buffers.
1193 if (page_has_private(page
)) {
1194 if (try_to_free_buffers(page
)) {
1195 ClearPageDirty(page
);
1196 pr_info("%s: orphaned page\n", __func__
);
1202 if (mapping
->a_ops
->writepage
== NULL
)
1203 return PAGE_ACTIVATE
;
1204 if (!may_write_to_inode(mapping
->host
))
1207 if (clear_page_dirty_for_io(page
)) {
1209 struct writeback_control wbc
= {
1210 .sync_mode
= WB_SYNC_NONE
,
1211 .nr_to_write
= SWAP_CLUSTER_MAX
,
1213 .range_end
= LLONG_MAX
,
1217 SetPageReclaim(page
);
1218 res
= mapping
->a_ops
->writepage(page
, &wbc
);
1220 handle_write_error(mapping
, page
, res
);
1221 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
1222 ClearPageReclaim(page
);
1223 return PAGE_ACTIVATE
;
1226 if (!PageWriteback(page
)) {
1227 /* synchronous write or broken a_ops? */
1228 ClearPageReclaim(page
);
1230 trace_mm_vmscan_writepage(page
);
1231 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
1232 return PAGE_SUCCESS
;
1239 * Same as remove_mapping, but if the page is removed from the mapping, it
1240 * gets returned with a refcount of 0.
1242 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
1243 bool reclaimed
, struct mem_cgroup
*target_memcg
)
1246 void *shadow
= NULL
;
1248 BUG_ON(!PageLocked(page
));
1249 BUG_ON(mapping
!= page_mapping(page
));
1251 if (!PageSwapCache(page
))
1252 spin_lock(&mapping
->host
->i_lock
);
1253 xa_lock_irq(&mapping
->i_pages
);
1255 * The non racy check for a busy page.
1257 * Must be careful with the order of the tests. When someone has
1258 * a ref to the page, it may be possible that they dirty it then
1259 * drop the reference. So if PageDirty is tested before page_count
1260 * here, then the following race may occur:
1262 * get_user_pages(&page);
1263 * [user mapping goes away]
1265 * !PageDirty(page) [good]
1266 * SetPageDirty(page);
1268 * !page_count(page) [good, discard it]
1270 * [oops, our write_to data is lost]
1272 * Reversing the order of the tests ensures such a situation cannot
1273 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1274 * load is not satisfied before that of page->_refcount.
1276 * Note that if SetPageDirty is always performed via set_page_dirty,
1277 * and thus under the i_pages lock, then this ordering is not required.
1279 refcount
= 1 + compound_nr(page
);
1280 if (!page_ref_freeze(page
, refcount
))
1282 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1283 if (unlikely(PageDirty(page
))) {
1284 page_ref_unfreeze(page
, refcount
);
1288 if (PageSwapCache(page
)) {
1289 swp_entry_t swap
= { .val
= page_private(page
) };
1290 mem_cgroup_swapout(page
, swap
);
1291 if (reclaimed
&& !mapping_exiting(mapping
))
1292 shadow
= workingset_eviction(page
, target_memcg
);
1293 __delete_from_swap_cache(page
, swap
, shadow
);
1294 xa_unlock_irq(&mapping
->i_pages
);
1295 put_swap_page(page
, swap
);
1297 void (*freepage
)(struct page
*);
1299 freepage
= mapping
->a_ops
->freepage
;
1301 * Remember a shadow entry for reclaimed file cache in
1302 * order to detect refaults, thus thrashing, later on.
1304 * But don't store shadows in an address space that is
1305 * already exiting. This is not just an optimization,
1306 * inode reclaim needs to empty out the radix tree or
1307 * the nodes are lost. Don't plant shadows behind its
1310 * We also don't store shadows for DAX mappings because the
1311 * only page cache pages found in these are zero pages
1312 * covering holes, and because we don't want to mix DAX
1313 * exceptional entries and shadow exceptional entries in the
1314 * same address_space.
1316 if (reclaimed
&& page_is_file_lru(page
) &&
1317 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
1318 shadow
= workingset_eviction(page
, target_memcg
);
1319 __delete_from_page_cache(page
, shadow
);
1320 xa_unlock_irq(&mapping
->i_pages
);
1321 if (mapping_shrinkable(mapping
))
1322 inode_add_lru(mapping
->host
);
1323 spin_unlock(&mapping
->host
->i_lock
);
1325 if (freepage
!= NULL
)
1332 xa_unlock_irq(&mapping
->i_pages
);
1333 if (!PageSwapCache(page
))
1334 spin_unlock(&mapping
->host
->i_lock
);
1339 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1340 * someone else has a ref on the page, abort and return 0. If it was
1341 * successfully detached, return 1. Assumes the caller has a single ref on
1344 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
1346 if (__remove_mapping(mapping
, page
, false, NULL
)) {
1348 * Unfreezing the refcount with 1 rather than 2 effectively
1349 * drops the pagecache ref for us without requiring another
1352 page_ref_unfreeze(page
, 1);
1359 * putback_lru_page - put previously isolated page onto appropriate LRU list
1360 * @page: page to be put back to appropriate lru list
1362 * Add previously isolated @page to appropriate LRU list.
1363 * Page may still be unevictable for other reasons.
1365 * lru_lock must not be held, interrupts must be enabled.
1367 void putback_lru_page(struct page
*page
)
1369 lru_cache_add(page
);
1370 put_page(page
); /* drop ref from isolate */
1373 enum page_references
{
1375 PAGEREF_RECLAIM_CLEAN
,
1380 static enum page_references
page_check_references(struct page
*page
,
1381 struct scan_control
*sc
)
1383 int referenced_ptes
, referenced_page
;
1384 unsigned long vm_flags
;
1386 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1388 referenced_page
= TestClearPageReferenced(page
);
1391 * Mlock lost the isolation race with us. Let try_to_unmap()
1392 * move the page to the unevictable list.
1394 if (vm_flags
& VM_LOCKED
)
1395 return PAGEREF_RECLAIM
;
1397 if (referenced_ptes
) {
1399 * All mapped pages start out with page table
1400 * references from the instantiating fault, so we need
1401 * to look twice if a mapped file page is used more
1404 * Mark it and spare it for another trip around the
1405 * inactive list. Another page table reference will
1406 * lead to its activation.
1408 * Note: the mark is set for activated pages as well
1409 * so that recently deactivated but used pages are
1410 * quickly recovered.
1412 SetPageReferenced(page
);
1414 if (referenced_page
|| referenced_ptes
> 1)
1415 return PAGEREF_ACTIVATE
;
1418 * Activate file-backed executable pages after first usage.
1420 if ((vm_flags
& VM_EXEC
) && !PageSwapBacked(page
))
1421 return PAGEREF_ACTIVATE
;
1423 return PAGEREF_KEEP
;
1426 /* Reclaim if clean, defer dirty pages to writeback */
1427 if (referenced_page
&& !PageSwapBacked(page
))
1428 return PAGEREF_RECLAIM_CLEAN
;
1430 return PAGEREF_RECLAIM
;
1433 /* Check if a page is dirty or under writeback */
1434 static void page_check_dirty_writeback(struct page
*page
,
1435 bool *dirty
, bool *writeback
)
1437 struct address_space
*mapping
;
1440 * Anonymous pages are not handled by flushers and must be written
1441 * from reclaim context. Do not stall reclaim based on them
1443 if (!page_is_file_lru(page
) ||
1444 (PageAnon(page
) && !PageSwapBacked(page
))) {
1450 /* By default assume that the page flags are accurate */
1451 *dirty
= PageDirty(page
);
1452 *writeback
= PageWriteback(page
);
1454 /* Verify dirty/writeback state if the filesystem supports it */
1455 if (!page_has_private(page
))
1458 mapping
= page_mapping(page
);
1459 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1460 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1463 static struct page
*alloc_demote_page(struct page
*page
, unsigned long node
)
1465 struct migration_target_control mtc
= {
1467 * Allocate from 'node', or fail quickly and quietly.
1468 * When this happens, 'page' will likely just be discarded
1469 * instead of migrated.
1471 .gfp_mask
= (GFP_HIGHUSER_MOVABLE
& ~__GFP_RECLAIM
) |
1472 __GFP_THISNODE
| __GFP_NOWARN
|
1473 __GFP_NOMEMALLOC
| GFP_NOWAIT
,
1477 return alloc_migration_target(page
, (unsigned long)&mtc
);
1481 * Take pages on @demote_list and attempt to demote them to
1482 * another node. Pages which are not demoted are left on
1485 static unsigned int demote_page_list(struct list_head
*demote_pages
,
1486 struct pglist_data
*pgdat
)
1488 int target_nid
= next_demotion_node(pgdat
->node_id
);
1489 unsigned int nr_succeeded
;
1491 if (list_empty(demote_pages
))
1494 if (target_nid
== NUMA_NO_NODE
)
1497 /* Demotion ignores all cpuset and mempolicy settings */
1498 migrate_pages(demote_pages
, alloc_demote_page
, NULL
,
1499 target_nid
, MIGRATE_ASYNC
, MR_DEMOTION
,
1502 if (current_is_kswapd())
1503 __count_vm_events(PGDEMOTE_KSWAPD
, nr_succeeded
);
1505 __count_vm_events(PGDEMOTE_DIRECT
, nr_succeeded
);
1507 return nr_succeeded
;
1511 * shrink_page_list() returns the number of reclaimed pages
1513 static unsigned int shrink_page_list(struct list_head
*page_list
,
1514 struct pglist_data
*pgdat
,
1515 struct scan_control
*sc
,
1516 struct reclaim_stat
*stat
,
1517 bool ignore_references
)
1519 LIST_HEAD(ret_pages
);
1520 LIST_HEAD(free_pages
);
1521 LIST_HEAD(demote_pages
);
1522 unsigned int nr_reclaimed
= 0;
1523 unsigned int pgactivate
= 0;
1524 bool do_demote_pass
;
1526 memset(stat
, 0, sizeof(*stat
));
1528 do_demote_pass
= can_demote(pgdat
->node_id
, sc
);
1531 while (!list_empty(page_list
)) {
1532 struct address_space
*mapping
;
1534 enum page_references references
= PAGEREF_RECLAIM
;
1535 bool dirty
, writeback
, may_enter_fs
;
1536 unsigned int nr_pages
;
1540 page
= lru_to_page(page_list
);
1541 list_del(&page
->lru
);
1543 if (!trylock_page(page
))
1546 VM_BUG_ON_PAGE(PageActive(page
), page
);
1548 nr_pages
= compound_nr(page
);
1550 /* Account the number of base pages even though THP */
1551 sc
->nr_scanned
+= nr_pages
;
1553 if (unlikely(!page_evictable(page
)))
1554 goto activate_locked
;
1556 if (!sc
->may_unmap
&& page_mapped(page
))
1559 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1560 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1563 * The number of dirty pages determines if a node is marked
1564 * reclaim_congested. kswapd will stall and start writing
1565 * pages if the tail of the LRU is all dirty unqueued pages.
1567 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1568 if (dirty
|| writeback
)
1571 if (dirty
&& !writeback
)
1572 stat
->nr_unqueued_dirty
++;
1575 * Treat this page as congested if the underlying BDI is or if
1576 * pages are cycling through the LRU so quickly that the
1577 * pages marked for immediate reclaim are making it to the
1578 * end of the LRU a second time.
1580 mapping
= page_mapping(page
);
1581 if (((dirty
|| writeback
) && mapping
&&
1582 inode_write_congested(mapping
->host
)) ||
1583 (writeback
&& PageReclaim(page
)))
1584 stat
->nr_congested
++;
1587 * If a page at the tail of the LRU is under writeback, there
1588 * are three cases to consider.
1590 * 1) If reclaim is encountering an excessive number of pages
1591 * under writeback and this page is both under writeback and
1592 * PageReclaim then it indicates that pages are being queued
1593 * for IO but are being recycled through the LRU before the
1594 * IO can complete. Waiting on the page itself risks an
1595 * indefinite stall if it is impossible to writeback the
1596 * page due to IO error or disconnected storage so instead
1597 * note that the LRU is being scanned too quickly and the
1598 * caller can stall after page list has been processed.
1600 * 2) Global or new memcg reclaim encounters a page that is
1601 * not marked for immediate reclaim, or the caller does not
1602 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1603 * not to fs). In this case mark the page for immediate
1604 * reclaim and continue scanning.
1606 * Require may_enter_fs because we would wait on fs, which
1607 * may not have submitted IO yet. And the loop driver might
1608 * enter reclaim, and deadlock if it waits on a page for
1609 * which it is needed to do the write (loop masks off
1610 * __GFP_IO|__GFP_FS for this reason); but more thought
1611 * would probably show more reasons.
1613 * 3) Legacy memcg encounters a page that is already marked
1614 * PageReclaim. memcg does not have any dirty pages
1615 * throttling so we could easily OOM just because too many
1616 * pages are in writeback and there is nothing else to
1617 * reclaim. Wait for the writeback to complete.
1619 * In cases 1) and 2) we activate the pages to get them out of
1620 * the way while we continue scanning for clean pages on the
1621 * inactive list and refilling from the active list. The
1622 * observation here is that waiting for disk writes is more
1623 * expensive than potentially causing reloads down the line.
1624 * Since they're marked for immediate reclaim, they won't put
1625 * memory pressure on the cache working set any longer than it
1626 * takes to write them to disk.
1628 if (PageWriteback(page
)) {
1630 if (current_is_kswapd() &&
1631 PageReclaim(page
) &&
1632 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1633 stat
->nr_immediate
++;
1634 goto activate_locked
;
1637 } else if (writeback_throttling_sane(sc
) ||
1638 !PageReclaim(page
) || !may_enter_fs
) {
1640 * This is slightly racy - end_page_writeback()
1641 * might have just cleared PageReclaim, then
1642 * setting PageReclaim here end up interpreted
1643 * as PageReadahead - but that does not matter
1644 * enough to care. What we do want is for this
1645 * page to have PageReclaim set next time memcg
1646 * reclaim reaches the tests above, so it will
1647 * then wait_on_page_writeback() to avoid OOM;
1648 * and it's also appropriate in global reclaim.
1650 SetPageReclaim(page
);
1651 stat
->nr_writeback
++;
1652 goto activate_locked
;
1657 wait_on_page_writeback(page
);
1658 /* then go back and try same page again */
1659 list_add_tail(&page
->lru
, page_list
);
1664 if (!ignore_references
)
1665 references
= page_check_references(page
, sc
);
1667 switch (references
) {
1668 case PAGEREF_ACTIVATE
:
1669 goto activate_locked
;
1671 stat
->nr_ref_keep
+= nr_pages
;
1673 case PAGEREF_RECLAIM
:
1674 case PAGEREF_RECLAIM_CLEAN
:
1675 ; /* try to reclaim the page below */
1679 * Before reclaiming the page, try to relocate
1680 * its contents to another node.
1682 if (do_demote_pass
&&
1683 (thp_migration_supported() || !PageTransHuge(page
))) {
1684 list_add(&page
->lru
, &demote_pages
);
1690 * Anonymous process memory has backing store?
1691 * Try to allocate it some swap space here.
1692 * Lazyfree page could be freed directly
1694 if (PageAnon(page
) && PageSwapBacked(page
)) {
1695 if (!PageSwapCache(page
)) {
1696 if (!(sc
->gfp_mask
& __GFP_IO
))
1698 if (page_maybe_dma_pinned(page
))
1700 if (PageTransHuge(page
)) {
1701 /* cannot split THP, skip it */
1702 if (!can_split_huge_page(page
, NULL
))
1703 goto activate_locked
;
1705 * Split pages without a PMD map right
1706 * away. Chances are some or all of the
1707 * tail pages can be freed without IO.
1709 if (!compound_mapcount(page
) &&
1710 split_huge_page_to_list(page
,
1712 goto activate_locked
;
1714 if (!add_to_swap(page
)) {
1715 if (!PageTransHuge(page
))
1716 goto activate_locked_split
;
1717 /* Fallback to swap normal pages */
1718 if (split_huge_page_to_list(page
,
1720 goto activate_locked
;
1721 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1722 count_vm_event(THP_SWPOUT_FALLBACK
);
1724 if (!add_to_swap(page
))
1725 goto activate_locked_split
;
1728 may_enter_fs
= true;
1730 /* Adding to swap updated mapping */
1731 mapping
= page_mapping(page
);
1733 } else if (unlikely(PageTransHuge(page
))) {
1734 /* Split file THP */
1735 if (split_huge_page_to_list(page
, page_list
))
1740 * THP may get split above, need minus tail pages and update
1741 * nr_pages to avoid accounting tail pages twice.
1743 * The tail pages that are added into swap cache successfully
1746 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1747 sc
->nr_scanned
-= (nr_pages
- 1);
1752 * The page is mapped into the page tables of one or more
1753 * processes. Try to unmap it here.
1755 if (page_mapped(page
)) {
1756 enum ttu_flags flags
= TTU_BATCH_FLUSH
;
1757 bool was_swapbacked
= PageSwapBacked(page
);
1759 if (unlikely(PageTransHuge(page
)))
1760 flags
|= TTU_SPLIT_HUGE_PMD
;
1762 try_to_unmap(page
, flags
);
1763 if (page_mapped(page
)) {
1764 stat
->nr_unmap_fail
+= nr_pages
;
1765 if (!was_swapbacked
&& PageSwapBacked(page
))
1766 stat
->nr_lazyfree_fail
+= nr_pages
;
1767 goto activate_locked
;
1771 if (PageDirty(page
)) {
1773 * Only kswapd can writeback filesystem pages
1774 * to avoid risk of stack overflow. But avoid
1775 * injecting inefficient single-page IO into
1776 * flusher writeback as much as possible: only
1777 * write pages when we've encountered many
1778 * dirty pages, and when we've already scanned
1779 * the rest of the LRU for clean pages and see
1780 * the same dirty pages again (PageReclaim).
1782 if (page_is_file_lru(page
) &&
1783 (!current_is_kswapd() || !PageReclaim(page
) ||
1784 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1786 * Immediately reclaim when written back.
1787 * Similar in principal to deactivate_page()
1788 * except we already have the page isolated
1789 * and know it's dirty
1791 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1792 SetPageReclaim(page
);
1794 goto activate_locked
;
1797 if (references
== PAGEREF_RECLAIM_CLEAN
)
1801 if (!sc
->may_writepage
)
1805 * Page is dirty. Flush the TLB if a writable entry
1806 * potentially exists to avoid CPU writes after IO
1807 * starts and then write it out here.
1809 try_to_unmap_flush_dirty();
1810 switch (pageout(page
, mapping
)) {
1814 goto activate_locked
;
1816 stat
->nr_pageout
+= thp_nr_pages(page
);
1818 if (PageWriteback(page
))
1820 if (PageDirty(page
))
1824 * A synchronous write - probably a ramdisk. Go
1825 * ahead and try to reclaim the page.
1827 if (!trylock_page(page
))
1829 if (PageDirty(page
) || PageWriteback(page
))
1831 mapping
= page_mapping(page
);
1834 ; /* try to free the page below */
1839 * If the page has buffers, try to free the buffer mappings
1840 * associated with this page. If we succeed we try to free
1843 * We do this even if the page is PageDirty().
1844 * try_to_release_page() does not perform I/O, but it is
1845 * possible for a page to have PageDirty set, but it is actually
1846 * clean (all its buffers are clean). This happens if the
1847 * buffers were written out directly, with submit_bh(). ext3
1848 * will do this, as well as the blockdev mapping.
1849 * try_to_release_page() will discover that cleanness and will
1850 * drop the buffers and mark the page clean - it can be freed.
1852 * Rarely, pages can have buffers and no ->mapping. These are
1853 * the pages which were not successfully invalidated in
1854 * truncate_cleanup_page(). We try to drop those buffers here
1855 * and if that worked, and the page is no longer mapped into
1856 * process address space (page_count == 1) it can be freed.
1857 * Otherwise, leave the page on the LRU so it is swappable.
1859 if (page_has_private(page
)) {
1860 if (!try_to_release_page(page
, sc
->gfp_mask
))
1861 goto activate_locked
;
1862 if (!mapping
&& page_count(page
) == 1) {
1864 if (put_page_testzero(page
))
1868 * rare race with speculative reference.
1869 * the speculative reference will free
1870 * this page shortly, so we may
1871 * increment nr_reclaimed here (and
1872 * leave it off the LRU).
1880 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1881 /* follow __remove_mapping for reference */
1882 if (!page_ref_freeze(page
, 1))
1885 * The page has only one reference left, which is
1886 * from the isolation. After the caller puts the
1887 * page back on lru and drops the reference, the
1888 * page will be freed anyway. It doesn't matter
1889 * which lru it goes. So we don't bother checking
1892 count_vm_event(PGLAZYFREED
);
1893 count_memcg_page_event(page
, PGLAZYFREED
);
1894 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true,
1895 sc
->target_mem_cgroup
))
1901 * THP may get swapped out in a whole, need account
1904 nr_reclaimed
+= nr_pages
;
1907 * Is there need to periodically free_page_list? It would
1908 * appear not as the counts should be low
1910 if (unlikely(PageTransHuge(page
)))
1911 destroy_compound_page(page
);
1913 list_add(&page
->lru
, &free_pages
);
1916 activate_locked_split
:
1918 * The tail pages that are failed to add into swap cache
1919 * reach here. Fixup nr_scanned and nr_pages.
1922 sc
->nr_scanned
-= (nr_pages
- 1);
1926 /* Not a candidate for swapping, so reclaim swap space. */
1927 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1929 try_to_free_swap(page
);
1930 VM_BUG_ON_PAGE(PageActive(page
), page
);
1931 if (!PageMlocked(page
)) {
1932 int type
= page_is_file_lru(page
);
1933 SetPageActive(page
);
1934 stat
->nr_activate
[type
] += nr_pages
;
1935 count_memcg_page_event(page
, PGACTIVATE
);
1940 list_add(&page
->lru
, &ret_pages
);
1941 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1943 /* 'page_list' is always empty here */
1945 /* Migrate pages selected for demotion */
1946 nr_reclaimed
+= demote_page_list(&demote_pages
, pgdat
);
1947 /* Pages that could not be demoted are still in @demote_pages */
1948 if (!list_empty(&demote_pages
)) {
1949 /* Pages which failed to demoted go back on @page_list for retry: */
1950 list_splice_init(&demote_pages
, page_list
);
1951 do_demote_pass
= false;
1955 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1957 mem_cgroup_uncharge_list(&free_pages
);
1958 try_to_unmap_flush();
1959 free_unref_page_list(&free_pages
);
1961 list_splice(&ret_pages
, page_list
);
1962 count_vm_events(PGACTIVATE
, pgactivate
);
1964 return nr_reclaimed
;
1967 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1968 struct list_head
*page_list
)
1970 struct scan_control sc
= {
1971 .gfp_mask
= GFP_KERNEL
,
1974 struct reclaim_stat stat
;
1975 unsigned int nr_reclaimed
;
1976 struct page
*page
, *next
;
1977 LIST_HEAD(clean_pages
);
1978 unsigned int noreclaim_flag
;
1980 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1981 if (!PageHuge(page
) && page_is_file_lru(page
) &&
1982 !PageDirty(page
) && !__PageMovable(page
) &&
1983 !PageUnevictable(page
)) {
1984 ClearPageActive(page
);
1985 list_move(&page
->lru
, &clean_pages
);
1990 * We should be safe here since we are only dealing with file pages and
1991 * we are not kswapd and therefore cannot write dirty file pages. But
1992 * call memalloc_noreclaim_save() anyway, just in case these conditions
1993 * change in the future.
1995 noreclaim_flag
= memalloc_noreclaim_save();
1996 nr_reclaimed
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1998 memalloc_noreclaim_restore(noreclaim_flag
);
2000 list_splice(&clean_pages
, page_list
);
2001 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
2002 -(long)nr_reclaimed
);
2004 * Since lazyfree pages are isolated from file LRU from the beginning,
2005 * they will rotate back to anonymous LRU in the end if it failed to
2006 * discard so isolated count will be mismatched.
2007 * Compensate the isolated count for both LRU lists.
2009 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
2010 stat
.nr_lazyfree_fail
);
2011 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
2012 -(long)stat
.nr_lazyfree_fail
);
2013 return nr_reclaimed
;
2017 * Attempt to remove the specified page from its LRU. Only take this page
2018 * if it is of the appropriate PageActive status. Pages which are being
2019 * freed elsewhere are also ignored.
2021 * page: page to consider
2022 * mode: one of the LRU isolation modes defined above
2024 * returns true on success, false on failure.
2026 bool __isolate_lru_page_prepare(struct page
*page
, isolate_mode_t mode
)
2028 /* Only take pages on the LRU. */
2032 /* Compaction should not handle unevictable pages but CMA can do so */
2033 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
2037 * To minimise LRU disruption, the caller can indicate that it only
2038 * wants to isolate pages it will be able to operate on without
2039 * blocking - clean pages for the most part.
2041 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
2042 * that it is possible to migrate without blocking
2044 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
2045 /* All the caller can do on PageWriteback is block */
2046 if (PageWriteback(page
))
2049 if (PageDirty(page
)) {
2050 struct address_space
*mapping
;
2054 * Only pages without mappings or that have a
2055 * ->migratepage callback are possible to migrate
2056 * without blocking. However, we can be racing with
2057 * truncation so it's necessary to lock the page
2058 * to stabilise the mapping as truncation holds
2059 * the page lock until after the page is removed
2060 * from the page cache.
2062 if (!trylock_page(page
))
2065 mapping
= page_mapping(page
);
2066 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
2073 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
2080 * Update LRU sizes after isolating pages. The LRU size updates must
2081 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2083 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
2084 enum lru_list lru
, unsigned long *nr_zone_taken
)
2088 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2089 if (!nr_zone_taken
[zid
])
2092 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
2098 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2100 * lruvec->lru_lock is heavily contended. Some of the functions that
2101 * shrink the lists perform better by taking out a batch of pages
2102 * and working on them outside the LRU lock.
2104 * For pagecache intensive workloads, this function is the hottest
2105 * spot in the kernel (apart from copy_*_user functions).
2107 * Lru_lock must be held before calling this function.
2109 * @nr_to_scan: The number of eligible pages to look through on the list.
2110 * @lruvec: The LRU vector to pull pages from.
2111 * @dst: The temp list to put pages on to.
2112 * @nr_scanned: The number of pages that were scanned.
2113 * @sc: The scan_control struct for this reclaim session
2114 * @lru: LRU list id for isolating
2116 * returns how many pages were moved onto *@dst.
2118 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
2119 struct lruvec
*lruvec
, struct list_head
*dst
,
2120 unsigned long *nr_scanned
, struct scan_control
*sc
,
2123 struct list_head
*src
= &lruvec
->lists
[lru
];
2124 unsigned long nr_taken
= 0;
2125 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
2126 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
2127 unsigned long skipped
= 0;
2128 unsigned long scan
, total_scan
, nr_pages
;
2129 LIST_HEAD(pages_skipped
);
2130 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
2134 while (scan
< nr_to_scan
&& !list_empty(src
)) {
2137 page
= lru_to_page(src
);
2138 prefetchw_prev_lru_page(page
, src
, flags
);
2140 nr_pages
= compound_nr(page
);
2141 total_scan
+= nr_pages
;
2143 if (page_zonenum(page
) > sc
->reclaim_idx
) {
2144 list_move(&page
->lru
, &pages_skipped
);
2145 nr_skipped
[page_zonenum(page
)] += nr_pages
;
2150 * Do not count skipped pages because that makes the function
2151 * return with no isolated pages if the LRU mostly contains
2152 * ineligible pages. This causes the VM to not reclaim any
2153 * pages, triggering a premature OOM.
2155 * Account all tail pages of THP. This would not cause
2156 * premature OOM since __isolate_lru_page() returns -EBUSY
2157 * only when the page is being freed somewhere else.
2160 if (!__isolate_lru_page_prepare(page
, mode
)) {
2161 /* It is being freed elsewhere */
2162 list_move(&page
->lru
, src
);
2166 * Be careful not to clear PageLRU until after we're
2167 * sure the page is not being freed elsewhere -- the
2168 * page release code relies on it.
2170 if (unlikely(!get_page_unless_zero(page
))) {
2171 list_move(&page
->lru
, src
);
2175 if (!TestClearPageLRU(page
)) {
2176 /* Another thread is already isolating this page */
2178 list_move(&page
->lru
, src
);
2182 nr_taken
+= nr_pages
;
2183 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
2184 list_move(&page
->lru
, dst
);
2188 * Splice any skipped pages to the start of the LRU list. Note that
2189 * this disrupts the LRU order when reclaiming for lower zones but
2190 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2191 * scanning would soon rescan the same pages to skip and put the
2192 * system at risk of premature OOM.
2194 if (!list_empty(&pages_skipped
)) {
2197 list_splice(&pages_skipped
, src
);
2198 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2199 if (!nr_skipped
[zid
])
2202 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
2203 skipped
+= nr_skipped
[zid
];
2206 *nr_scanned
= total_scan
;
2207 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
2208 total_scan
, skipped
, nr_taken
, mode
, lru
);
2209 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
2214 * isolate_lru_page - tries to isolate a page from its LRU list
2215 * @page: page to isolate from its LRU list
2217 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
2218 * vmstat statistic corresponding to whatever LRU list the page was on.
2220 * Returns 0 if the page was removed from an LRU list.
2221 * Returns -EBUSY if the page was not on an LRU list.
2223 * The returned page will have PageLRU() cleared. If it was found on
2224 * the active list, it will have PageActive set. If it was found on
2225 * the unevictable list, it will have the PageUnevictable bit set. That flag
2226 * may need to be cleared by the caller before letting the page go.
2228 * The vmstat statistic corresponding to the list on which the page was
2229 * found will be decremented.
2233 * (1) Must be called with an elevated refcount on the page. This is a
2234 * fundamental difference from isolate_lru_pages (which is called
2235 * without a stable reference).
2236 * (2) the lru_lock must not be held.
2237 * (3) interrupts must be enabled.
2239 int isolate_lru_page(struct page
*page
)
2241 struct folio
*folio
= page_folio(page
);
2244 VM_BUG_ON_PAGE(!page_count(page
), page
);
2245 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
2247 if (TestClearPageLRU(page
)) {
2248 struct lruvec
*lruvec
;
2251 lruvec
= folio_lruvec_lock_irq(folio
);
2252 del_page_from_lru_list(page
, lruvec
);
2253 unlock_page_lruvec_irq(lruvec
);
2261 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2262 * then get rescheduled. When there are massive number of tasks doing page
2263 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2264 * the LRU list will go small and be scanned faster than necessary, leading to
2265 * unnecessary swapping, thrashing and OOM.
2267 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
2268 struct scan_control
*sc
)
2270 unsigned long inactive
, isolated
;
2273 if (current_is_kswapd())
2276 if (!writeback_throttling_sane(sc
))
2280 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2281 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
2283 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2284 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
2288 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2289 * won't get blocked by normal direct-reclaimers, forming a circular
2292 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
2295 too_many
= isolated
> inactive
;
2297 /* Wake up tasks throttled due to too_many_isolated. */
2299 wake_throttle_isolated(pgdat
);
2305 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2306 * On return, @list is reused as a list of pages to be freed by the caller.
2308 * Returns the number of pages moved to the given lruvec.
2310 static unsigned int move_pages_to_lru(struct lruvec
*lruvec
,
2311 struct list_head
*list
)
2313 int nr_pages
, nr_moved
= 0;
2314 LIST_HEAD(pages_to_free
);
2317 while (!list_empty(list
)) {
2318 page
= lru_to_page(list
);
2319 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2320 list_del(&page
->lru
);
2321 if (unlikely(!page_evictable(page
))) {
2322 spin_unlock_irq(&lruvec
->lru_lock
);
2323 putback_lru_page(page
);
2324 spin_lock_irq(&lruvec
->lru_lock
);
2329 * The SetPageLRU needs to be kept here for list integrity.
2331 * #0 move_pages_to_lru #1 release_pages
2332 * if !put_page_testzero
2333 * if (put_page_testzero())
2334 * !PageLRU //skip lru_lock
2336 * list_add(&page->lru,)
2337 * list_add(&page->lru,)
2341 if (unlikely(put_page_testzero(page
))) {
2342 __clear_page_lru_flags(page
);
2344 if (unlikely(PageCompound(page
))) {
2345 spin_unlock_irq(&lruvec
->lru_lock
);
2346 destroy_compound_page(page
);
2347 spin_lock_irq(&lruvec
->lru_lock
);
2349 list_add(&page
->lru
, &pages_to_free
);
2355 * All pages were isolated from the same lruvec (and isolation
2356 * inhibits memcg migration).
2358 VM_BUG_ON_PAGE(!folio_matches_lruvec(page_folio(page
), lruvec
), page
);
2359 add_page_to_lru_list(page
, lruvec
);
2360 nr_pages
= thp_nr_pages(page
);
2361 nr_moved
+= nr_pages
;
2362 if (PageActive(page
))
2363 workingset_age_nonresident(lruvec
, nr_pages
);
2367 * To save our caller's stack, now use input list for pages to free.
2369 list_splice(&pages_to_free
, list
);
2375 * If a kernel thread (such as nfsd for loop-back mounts) services
2376 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2377 * In that case we should only throttle if the backing device it is
2378 * writing to is congested. In other cases it is safe to throttle.
2380 static int current_may_throttle(void)
2382 return !(current
->flags
& PF_LOCAL_THROTTLE
) ||
2383 current
->backing_dev_info
== NULL
||
2384 bdi_write_congested(current
->backing_dev_info
);
2388 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2389 * of reclaimed pages
2391 static unsigned long
2392 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
2393 struct scan_control
*sc
, enum lru_list lru
)
2395 LIST_HEAD(page_list
);
2396 unsigned long nr_scanned
;
2397 unsigned int nr_reclaimed
= 0;
2398 unsigned long nr_taken
;
2399 struct reclaim_stat stat
;
2400 bool file
= is_file_lru(lru
);
2401 enum vm_event_item item
;
2402 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2403 bool stalled
= false;
2405 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
2409 /* wait a bit for the reclaimer. */
2411 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_ISOLATED
);
2413 /* We are about to die and free our memory. Return now. */
2414 if (fatal_signal_pending(current
))
2415 return SWAP_CLUSTER_MAX
;
2420 spin_lock_irq(&lruvec
->lru_lock
);
2422 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
2423 &nr_scanned
, sc
, lru
);
2425 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2426 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
2427 if (!cgroup_reclaim(sc
))
2428 __count_vm_events(item
, nr_scanned
);
2429 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
2430 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
2432 spin_unlock_irq(&lruvec
->lru_lock
);
2437 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, &stat
, false);
2439 spin_lock_irq(&lruvec
->lru_lock
);
2440 move_pages_to_lru(lruvec
, &page_list
);
2442 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2443 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2444 if (!cgroup_reclaim(sc
))
2445 __count_vm_events(item
, nr_reclaimed
);
2446 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2447 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
2448 spin_unlock_irq(&lruvec
->lru_lock
);
2450 lru_note_cost(lruvec
, file
, stat
.nr_pageout
);
2451 mem_cgroup_uncharge_list(&page_list
);
2452 free_unref_page_list(&page_list
);
2455 * If dirty pages are scanned that are not queued for IO, it
2456 * implies that flushers are not doing their job. This can
2457 * happen when memory pressure pushes dirty pages to the end of
2458 * the LRU before the dirty limits are breached and the dirty
2459 * data has expired. It can also happen when the proportion of
2460 * dirty pages grows not through writes but through memory
2461 * pressure reclaiming all the clean cache. And in some cases,
2462 * the flushers simply cannot keep up with the allocation
2463 * rate. Nudge the flusher threads in case they are asleep.
2465 if (stat
.nr_unqueued_dirty
== nr_taken
)
2466 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2468 sc
->nr
.dirty
+= stat
.nr_dirty
;
2469 sc
->nr
.congested
+= stat
.nr_congested
;
2470 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2471 sc
->nr
.writeback
+= stat
.nr_writeback
;
2472 sc
->nr
.immediate
+= stat
.nr_immediate
;
2473 sc
->nr
.taken
+= nr_taken
;
2475 sc
->nr
.file_taken
+= nr_taken
;
2477 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2478 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2479 return nr_reclaimed
;
2483 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2485 * We move them the other way if the page is referenced by one or more
2488 * If the pages are mostly unmapped, the processing is fast and it is
2489 * appropriate to hold lru_lock across the whole operation. But if
2490 * the pages are mapped, the processing is slow (page_referenced()), so
2491 * we should drop lru_lock around each page. It's impossible to balance
2492 * this, so instead we remove the pages from the LRU while processing them.
2493 * It is safe to rely on PG_active against the non-LRU pages in here because
2494 * nobody will play with that bit on a non-LRU page.
2496 * The downside is that we have to touch page->_refcount against each page.
2497 * But we had to alter page->flags anyway.
2499 static void shrink_active_list(unsigned long nr_to_scan
,
2500 struct lruvec
*lruvec
,
2501 struct scan_control
*sc
,
2504 unsigned long nr_taken
;
2505 unsigned long nr_scanned
;
2506 unsigned long vm_flags
;
2507 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2508 LIST_HEAD(l_active
);
2509 LIST_HEAD(l_inactive
);
2511 unsigned nr_deactivate
, nr_activate
;
2512 unsigned nr_rotated
= 0;
2513 int file
= is_file_lru(lru
);
2514 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2518 spin_lock_irq(&lruvec
->lru_lock
);
2520 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2521 &nr_scanned
, sc
, lru
);
2523 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2525 if (!cgroup_reclaim(sc
))
2526 __count_vm_events(PGREFILL
, nr_scanned
);
2527 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2529 spin_unlock_irq(&lruvec
->lru_lock
);
2531 while (!list_empty(&l_hold
)) {
2533 page
= lru_to_page(&l_hold
);
2534 list_del(&page
->lru
);
2536 if (unlikely(!page_evictable(page
))) {
2537 putback_lru_page(page
);
2541 if (unlikely(buffer_heads_over_limit
)) {
2542 if (page_has_private(page
) && trylock_page(page
)) {
2543 if (page_has_private(page
))
2544 try_to_release_page(page
, 0);
2549 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2552 * Identify referenced, file-backed active pages and
2553 * give them one more trip around the active list. So
2554 * that executable code get better chances to stay in
2555 * memory under moderate memory pressure. Anon pages
2556 * are not likely to be evicted by use-once streaming
2557 * IO, plus JVM can create lots of anon VM_EXEC pages,
2558 * so we ignore them here.
2560 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2561 nr_rotated
+= thp_nr_pages(page
);
2562 list_add(&page
->lru
, &l_active
);
2567 ClearPageActive(page
); /* we are de-activating */
2568 SetPageWorkingset(page
);
2569 list_add(&page
->lru
, &l_inactive
);
2573 * Move pages back to the lru list.
2575 spin_lock_irq(&lruvec
->lru_lock
);
2577 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2578 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2579 /* Keep all free pages in l_active list */
2580 list_splice(&l_inactive
, &l_active
);
2582 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2583 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2585 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2586 spin_unlock_irq(&lruvec
->lru_lock
);
2588 mem_cgroup_uncharge_list(&l_active
);
2589 free_unref_page_list(&l_active
);
2590 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2591 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2594 unsigned long reclaim_pages(struct list_head
*page_list
)
2596 int nid
= NUMA_NO_NODE
;
2597 unsigned int nr_reclaimed
= 0;
2598 LIST_HEAD(node_page_list
);
2599 struct reclaim_stat dummy_stat
;
2601 unsigned int noreclaim_flag
;
2602 struct scan_control sc
= {
2603 .gfp_mask
= GFP_KERNEL
,
2610 noreclaim_flag
= memalloc_noreclaim_save();
2612 while (!list_empty(page_list
)) {
2613 page
= lru_to_page(page_list
);
2614 if (nid
== NUMA_NO_NODE
) {
2615 nid
= page_to_nid(page
);
2616 INIT_LIST_HEAD(&node_page_list
);
2619 if (nid
== page_to_nid(page
)) {
2620 ClearPageActive(page
);
2621 list_move(&page
->lru
, &node_page_list
);
2625 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2627 &sc
, &dummy_stat
, false);
2628 while (!list_empty(&node_page_list
)) {
2629 page
= lru_to_page(&node_page_list
);
2630 list_del(&page
->lru
);
2631 putback_lru_page(page
);
2637 if (!list_empty(&node_page_list
)) {
2638 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2640 &sc
, &dummy_stat
, false);
2641 while (!list_empty(&node_page_list
)) {
2642 page
= lru_to_page(&node_page_list
);
2643 list_del(&page
->lru
);
2644 putback_lru_page(page
);
2648 memalloc_noreclaim_restore(noreclaim_flag
);
2650 return nr_reclaimed
;
2653 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2654 struct lruvec
*lruvec
, struct scan_control
*sc
)
2656 if (is_active_lru(lru
)) {
2657 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2658 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2660 sc
->skipped_deactivate
= 1;
2664 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2668 * The inactive anon list should be small enough that the VM never has
2669 * to do too much work.
2671 * The inactive file list should be small enough to leave most memory
2672 * to the established workingset on the scan-resistant active list,
2673 * but large enough to avoid thrashing the aggregate readahead window.
2675 * Both inactive lists should also be large enough that each inactive
2676 * page has a chance to be referenced again before it is reclaimed.
2678 * If that fails and refaulting is observed, the inactive list grows.
2680 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2681 * on this LRU, maintained by the pageout code. An inactive_ratio
2682 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2685 * memory ratio inactive
2686 * -------------------------------------
2695 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2697 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2698 unsigned long inactive
, active
;
2699 unsigned long inactive_ratio
;
2702 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2703 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2705 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2707 inactive_ratio
= int_sqrt(10 * gb
);
2711 return inactive
* inactive_ratio
< active
;
2722 * Determine how aggressively the anon and file LRU lists should be
2723 * scanned. The relative value of each set of LRU lists is determined
2724 * by looking at the fraction of the pages scanned we did rotate back
2725 * onto the active list instead of evict.
2727 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2728 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2730 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2733 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2734 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2735 unsigned long anon_cost
, file_cost
, total_cost
;
2736 int swappiness
= mem_cgroup_swappiness(memcg
);
2737 u64 fraction
[ANON_AND_FILE
];
2738 u64 denominator
= 0; /* gcc */
2739 enum scan_balance scan_balance
;
2740 unsigned long ap
, fp
;
2743 /* If we have no swap space, do not bother scanning anon pages. */
2744 if (!sc
->may_swap
|| !can_reclaim_anon_pages(memcg
, pgdat
->node_id
, sc
)) {
2745 scan_balance
= SCAN_FILE
;
2750 * Global reclaim will swap to prevent OOM even with no
2751 * swappiness, but memcg users want to use this knob to
2752 * disable swapping for individual groups completely when
2753 * using the memory controller's swap limit feature would be
2756 if (cgroup_reclaim(sc
) && !swappiness
) {
2757 scan_balance
= SCAN_FILE
;
2762 * Do not apply any pressure balancing cleverness when the
2763 * system is close to OOM, scan both anon and file equally
2764 * (unless the swappiness setting disagrees with swapping).
2766 if (!sc
->priority
&& swappiness
) {
2767 scan_balance
= SCAN_EQUAL
;
2772 * If the system is almost out of file pages, force-scan anon.
2774 if (sc
->file_is_tiny
) {
2775 scan_balance
= SCAN_ANON
;
2780 * If there is enough inactive page cache, we do not reclaim
2781 * anything from the anonymous working right now.
2783 if (sc
->cache_trim_mode
) {
2784 scan_balance
= SCAN_FILE
;
2788 scan_balance
= SCAN_FRACT
;
2790 * Calculate the pressure balance between anon and file pages.
2792 * The amount of pressure we put on each LRU is inversely
2793 * proportional to the cost of reclaiming each list, as
2794 * determined by the share of pages that are refaulting, times
2795 * the relative IO cost of bringing back a swapped out
2796 * anonymous page vs reloading a filesystem page (swappiness).
2798 * Although we limit that influence to ensure no list gets
2799 * left behind completely: at least a third of the pressure is
2800 * applied, before swappiness.
2802 * With swappiness at 100, anon and file have equal IO cost.
2804 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2805 anon_cost
= total_cost
+ sc
->anon_cost
;
2806 file_cost
= total_cost
+ sc
->file_cost
;
2807 total_cost
= anon_cost
+ file_cost
;
2809 ap
= swappiness
* (total_cost
+ 1);
2810 ap
/= anon_cost
+ 1;
2812 fp
= (200 - swappiness
) * (total_cost
+ 1);
2813 fp
/= file_cost
+ 1;
2817 denominator
= ap
+ fp
;
2819 for_each_evictable_lru(lru
) {
2820 int file
= is_file_lru(lru
);
2821 unsigned long lruvec_size
;
2822 unsigned long low
, min
;
2825 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2826 mem_cgroup_protection(sc
->target_mem_cgroup
, memcg
,
2831 * Scale a cgroup's reclaim pressure by proportioning
2832 * its current usage to its memory.low or memory.min
2835 * This is important, as otherwise scanning aggression
2836 * becomes extremely binary -- from nothing as we
2837 * approach the memory protection threshold, to totally
2838 * nominal as we exceed it. This results in requiring
2839 * setting extremely liberal protection thresholds. It
2840 * also means we simply get no protection at all if we
2841 * set it too low, which is not ideal.
2843 * If there is any protection in place, we reduce scan
2844 * pressure by how much of the total memory used is
2845 * within protection thresholds.
2847 * There is one special case: in the first reclaim pass,
2848 * we skip over all groups that are within their low
2849 * protection. If that fails to reclaim enough pages to
2850 * satisfy the reclaim goal, we come back and override
2851 * the best-effort low protection. However, we still
2852 * ideally want to honor how well-behaved groups are in
2853 * that case instead of simply punishing them all
2854 * equally. As such, we reclaim them based on how much
2855 * memory they are using, reducing the scan pressure
2856 * again by how much of the total memory used is under
2859 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2860 unsigned long protection
;
2862 /* memory.low scaling, make sure we retry before OOM */
2863 if (!sc
->memcg_low_reclaim
&& low
> min
) {
2865 sc
->memcg_low_skipped
= 1;
2870 /* Avoid TOCTOU with earlier protection check */
2871 cgroup_size
= max(cgroup_size
, protection
);
2873 scan
= lruvec_size
- lruvec_size
* protection
/
2877 * Minimally target SWAP_CLUSTER_MAX pages to keep
2878 * reclaim moving forwards, avoiding decrementing
2879 * sc->priority further than desirable.
2881 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2886 scan
>>= sc
->priority
;
2889 * If the cgroup's already been deleted, make sure to
2890 * scrape out the remaining cache.
2892 if (!scan
&& !mem_cgroup_online(memcg
))
2893 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2895 switch (scan_balance
) {
2897 /* Scan lists relative to size */
2901 * Scan types proportional to swappiness and
2902 * their relative recent reclaim efficiency.
2903 * Make sure we don't miss the last page on
2904 * the offlined memory cgroups because of a
2907 scan
= mem_cgroup_online(memcg
) ?
2908 div64_u64(scan
* fraction
[file
], denominator
) :
2909 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2914 /* Scan one type exclusively */
2915 if ((scan_balance
== SCAN_FILE
) != file
)
2919 /* Look ma, no brain */
2928 * Anonymous LRU management is a waste if there is
2929 * ultimately no way to reclaim the memory.
2931 static bool can_age_anon_pages(struct pglist_data
*pgdat
,
2932 struct scan_control
*sc
)
2934 /* Aging the anon LRU is valuable if swap is present: */
2935 if (total_swap_pages
> 0)
2938 /* Also valuable if anon pages can be demoted: */
2939 return can_demote(pgdat
->node_id
, sc
);
2942 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2944 unsigned long nr
[NR_LRU_LISTS
];
2945 unsigned long targets
[NR_LRU_LISTS
];
2946 unsigned long nr_to_scan
;
2948 unsigned long nr_reclaimed
= 0;
2949 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2950 struct blk_plug plug
;
2953 get_scan_count(lruvec
, sc
, nr
);
2955 /* Record the original scan target for proportional adjustments later */
2956 memcpy(targets
, nr
, sizeof(nr
));
2959 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2960 * event that can occur when there is little memory pressure e.g.
2961 * multiple streaming readers/writers. Hence, we do not abort scanning
2962 * when the requested number of pages are reclaimed when scanning at
2963 * DEF_PRIORITY on the assumption that the fact we are direct
2964 * reclaiming implies that kswapd is not keeping up and it is best to
2965 * do a batch of work at once. For memcg reclaim one check is made to
2966 * abort proportional reclaim if either the file or anon lru has already
2967 * dropped to zero at the first pass.
2969 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2970 sc
->priority
== DEF_PRIORITY
);
2972 blk_start_plug(&plug
);
2973 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2974 nr
[LRU_INACTIVE_FILE
]) {
2975 unsigned long nr_anon
, nr_file
, percentage
;
2976 unsigned long nr_scanned
;
2978 for_each_evictable_lru(lru
) {
2980 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2981 nr
[lru
] -= nr_to_scan
;
2983 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2990 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2994 * For kswapd and memcg, reclaim at least the number of pages
2995 * requested. Ensure that the anon and file LRUs are scanned
2996 * proportionally what was requested by get_scan_count(). We
2997 * stop reclaiming one LRU and reduce the amount scanning
2998 * proportional to the original scan target.
3000 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
3001 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
3004 * It's just vindictive to attack the larger once the smaller
3005 * has gone to zero. And given the way we stop scanning the
3006 * smaller below, this makes sure that we only make one nudge
3007 * towards proportionality once we've got nr_to_reclaim.
3009 if (!nr_file
|| !nr_anon
)
3012 if (nr_file
> nr_anon
) {
3013 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
3014 targets
[LRU_ACTIVE_ANON
] + 1;
3016 percentage
= nr_anon
* 100 / scan_target
;
3018 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
3019 targets
[LRU_ACTIVE_FILE
] + 1;
3021 percentage
= nr_file
* 100 / scan_target
;
3024 /* Stop scanning the smaller of the LRU */
3026 nr
[lru
+ LRU_ACTIVE
] = 0;
3029 * Recalculate the other LRU scan count based on its original
3030 * scan target and the percentage scanning already complete
3032 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
3033 nr_scanned
= targets
[lru
] - nr
[lru
];
3034 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
3035 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
3038 nr_scanned
= targets
[lru
] - nr
[lru
];
3039 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
3040 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
3042 scan_adjusted
= true;
3044 blk_finish_plug(&plug
);
3045 sc
->nr_reclaimed
+= nr_reclaimed
;
3048 * Even if we did not try to evict anon pages at all, we want to
3049 * rebalance the anon lru active/inactive ratio.
3051 if (can_age_anon_pages(lruvec_pgdat(lruvec
), sc
) &&
3052 inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3053 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3054 sc
, LRU_ACTIVE_ANON
);
3057 /* Use reclaim/compaction for costly allocs or under memory pressure */
3058 static bool in_reclaim_compaction(struct scan_control
*sc
)
3060 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
3061 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
3062 sc
->priority
< DEF_PRIORITY
- 2))
3069 * Reclaim/compaction is used for high-order allocation requests. It reclaims
3070 * order-0 pages before compacting the zone. should_continue_reclaim() returns
3071 * true if more pages should be reclaimed such that when the page allocator
3072 * calls try_to_compact_pages() that it will have enough free pages to succeed.
3073 * It will give up earlier than that if there is difficulty reclaiming pages.
3075 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
3076 unsigned long nr_reclaimed
,
3077 struct scan_control
*sc
)
3079 unsigned long pages_for_compaction
;
3080 unsigned long inactive_lru_pages
;
3083 /* If not in reclaim/compaction mode, stop */
3084 if (!in_reclaim_compaction(sc
))
3088 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
3089 * number of pages that were scanned. This will return to the caller
3090 * with the risk reclaim/compaction and the resulting allocation attempt
3091 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
3092 * allocations through requiring that the full LRU list has been scanned
3093 * first, by assuming that zero delta of sc->nr_scanned means full LRU
3094 * scan, but that approximation was wrong, and there were corner cases
3095 * where always a non-zero amount of pages were scanned.
3100 /* If compaction would go ahead or the allocation would succeed, stop */
3101 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3102 struct zone
*zone
= &pgdat
->node_zones
[z
];
3103 if (!managed_zone(zone
))
3106 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
3107 case COMPACT_SUCCESS
:
3108 case COMPACT_CONTINUE
:
3111 /* check next zone */
3117 * If we have not reclaimed enough pages for compaction and the
3118 * inactive lists are large enough, continue reclaiming
3120 pages_for_compaction
= compact_gap(sc
->order
);
3121 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
3122 if (can_reclaim_anon_pages(NULL
, pgdat
->node_id
, sc
))
3123 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3125 return inactive_lru_pages
> pages_for_compaction
;
3128 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
3130 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
3131 struct mem_cgroup
*memcg
;
3133 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
3135 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3136 unsigned long reclaimed
;
3137 unsigned long scanned
;
3140 * This loop can become CPU-bound when target memcgs
3141 * aren't eligible for reclaim - either because they
3142 * don't have any reclaimable pages, or because their
3143 * memory is explicitly protected. Avoid soft lockups.
3147 mem_cgroup_calculate_protection(target_memcg
, memcg
);
3149 if (mem_cgroup_below_min(memcg
)) {
3152 * If there is no reclaimable memory, OOM.
3155 } else if (mem_cgroup_below_low(memcg
)) {
3158 * Respect the protection only as long as
3159 * there is an unprotected supply
3160 * of reclaimable memory from other cgroups.
3162 if (!sc
->memcg_low_reclaim
) {
3163 sc
->memcg_low_skipped
= 1;
3166 memcg_memory_event(memcg
, MEMCG_LOW
);
3169 reclaimed
= sc
->nr_reclaimed
;
3170 scanned
= sc
->nr_scanned
;
3172 shrink_lruvec(lruvec
, sc
);
3174 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
3177 /* Record the group's reclaim efficiency */
3178 vmpressure(sc
->gfp_mask
, memcg
, false,
3179 sc
->nr_scanned
- scanned
,
3180 sc
->nr_reclaimed
- reclaimed
);
3182 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
3185 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
3187 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
3188 unsigned long nr_reclaimed
, nr_scanned
;
3189 struct lruvec
*target_lruvec
;
3190 bool reclaimable
= false;
3193 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
3197 * Flush the memory cgroup stats, so that we read accurate per-memcg
3198 * lruvec stats for heuristics.
3200 mem_cgroup_flush_stats();
3202 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
3204 nr_reclaimed
= sc
->nr_reclaimed
;
3205 nr_scanned
= sc
->nr_scanned
;
3208 * Determine the scan balance between anon and file LRUs.
3210 spin_lock_irq(&target_lruvec
->lru_lock
);
3211 sc
->anon_cost
= target_lruvec
->anon_cost
;
3212 sc
->file_cost
= target_lruvec
->file_cost
;
3213 spin_unlock_irq(&target_lruvec
->lru_lock
);
3216 * Target desirable inactive:active list ratios for the anon
3217 * and file LRU lists.
3219 if (!sc
->force_deactivate
) {
3220 unsigned long refaults
;
3222 refaults
= lruvec_page_state(target_lruvec
,
3223 WORKINGSET_ACTIVATE_ANON
);
3224 if (refaults
!= target_lruvec
->refaults
[0] ||
3225 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
3226 sc
->may_deactivate
|= DEACTIVATE_ANON
;
3228 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
3231 * When refaults are being observed, it means a new
3232 * workingset is being established. Deactivate to get
3233 * rid of any stale active pages quickly.
3235 refaults
= lruvec_page_state(target_lruvec
,
3236 WORKINGSET_ACTIVATE_FILE
);
3237 if (refaults
!= target_lruvec
->refaults
[1] ||
3238 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
3239 sc
->may_deactivate
|= DEACTIVATE_FILE
;
3241 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
3243 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
3246 * If we have plenty of inactive file pages that aren't
3247 * thrashing, try to reclaim those first before touching
3250 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
3251 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
3252 sc
->cache_trim_mode
= 1;
3254 sc
->cache_trim_mode
= 0;
3257 * Prevent the reclaimer from falling into the cache trap: as
3258 * cache pages start out inactive, every cache fault will tip
3259 * the scan balance towards the file LRU. And as the file LRU
3260 * shrinks, so does the window for rotation from references.
3261 * This means we have a runaway feedback loop where a tiny
3262 * thrashing file LRU becomes infinitely more attractive than
3263 * anon pages. Try to detect this based on file LRU size.
3265 if (!cgroup_reclaim(sc
)) {
3266 unsigned long total_high_wmark
= 0;
3267 unsigned long free
, anon
;
3270 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
3271 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
3272 node_page_state(pgdat
, NR_INACTIVE_FILE
);
3274 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
3275 struct zone
*zone
= &pgdat
->node_zones
[z
];
3276 if (!managed_zone(zone
))
3279 total_high_wmark
+= high_wmark_pages(zone
);
3283 * Consider anon: if that's low too, this isn't a
3284 * runaway file reclaim problem, but rather just
3285 * extreme pressure. Reclaim as per usual then.
3287 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3290 file
+ free
<= total_high_wmark
&&
3291 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
3292 anon
>> sc
->priority
;
3295 shrink_node_memcgs(pgdat
, sc
);
3297 if (reclaim_state
) {
3298 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
3299 reclaim_state
->reclaimed_slab
= 0;
3302 /* Record the subtree's reclaim efficiency */
3303 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
3304 sc
->nr_scanned
- nr_scanned
,
3305 sc
->nr_reclaimed
- nr_reclaimed
);
3307 if (sc
->nr_reclaimed
- nr_reclaimed
)
3310 if (current_is_kswapd()) {
3312 * If reclaim is isolating dirty pages under writeback,
3313 * it implies that the long-lived page allocation rate
3314 * is exceeding the page laundering rate. Either the
3315 * global limits are not being effective at throttling
3316 * processes due to the page distribution throughout
3317 * zones or there is heavy usage of a slow backing
3318 * device. The only option is to throttle from reclaim
3319 * context which is not ideal as there is no guarantee
3320 * the dirtying process is throttled in the same way
3321 * balance_dirty_pages() manages.
3323 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3324 * count the number of pages under pages flagged for
3325 * immediate reclaim and stall if any are encountered
3326 * in the nr_immediate check below.
3328 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
3329 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3331 /* Allow kswapd to start writing pages during reclaim.*/
3332 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
3333 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3336 * If kswapd scans pages marked for immediate
3337 * reclaim and under writeback (nr_immediate), it
3338 * implies that pages are cycling through the LRU
3339 * faster than they are written so forcibly stall
3340 * until some pages complete writeback.
3342 if (sc
->nr
.immediate
)
3343 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_WRITEBACK
);
3347 * Tag a node/memcg as congested if all the dirty pages were marked
3348 * for writeback and immediate reclaim (counted in nr.congested).
3350 * Legacy memcg will stall in page writeback so avoid forcibly
3351 * stalling in reclaim_throttle().
3353 if ((current_is_kswapd() ||
3354 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
3355 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
3356 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
3359 * Stall direct reclaim for IO completions if the lruvec is
3360 * node is congested. Allow kswapd to continue until it
3361 * starts encountering unqueued dirty pages or cycling through
3362 * the LRU too quickly.
3364 if (!current_is_kswapd() && current_may_throttle() &&
3365 !sc
->hibernation_mode
&&
3366 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
3367 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_CONGESTED
);
3369 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
3374 * Kswapd gives up on balancing particular nodes after too
3375 * many failures to reclaim anything from them and goes to
3376 * sleep. On reclaim progress, reset the failure counter. A
3377 * successful direct reclaim run will revive a dormant kswapd.
3380 pgdat
->kswapd_failures
= 0;
3384 * Returns true if compaction should go ahead for a costly-order request, or
3385 * the allocation would already succeed without compaction. Return false if we
3386 * should reclaim first.
3388 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
3390 unsigned long watermark
;
3391 enum compact_result suitable
;
3393 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
3394 if (suitable
== COMPACT_SUCCESS
)
3395 /* Allocation should succeed already. Don't reclaim. */
3397 if (suitable
== COMPACT_SKIPPED
)
3398 /* Compaction cannot yet proceed. Do reclaim. */
3402 * Compaction is already possible, but it takes time to run and there
3403 * are potentially other callers using the pages just freed. So proceed
3404 * with reclaim to make a buffer of free pages available to give
3405 * compaction a reasonable chance of completing and allocating the page.
3406 * Note that we won't actually reclaim the whole buffer in one attempt
3407 * as the target watermark in should_continue_reclaim() is lower. But if
3408 * we are already above the high+gap watermark, don't reclaim at all.
3410 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
3412 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
3415 static void consider_reclaim_throttle(pg_data_t
*pgdat
, struct scan_control
*sc
)
3418 * If reclaim is making progress greater than 12% efficiency then
3419 * wake all the NOPROGRESS throttled tasks.
3421 if (sc
->nr_reclaimed
> (sc
->nr_scanned
>> 3)) {
3422 wait_queue_head_t
*wqh
;
3424 wqh
= &pgdat
->reclaim_wait
[VMSCAN_THROTTLE_NOPROGRESS
];
3425 if (waitqueue_active(wqh
))
3432 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
3433 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
3434 * under writeback and marked for immediate reclaim at the tail of the
3437 if (current_is_kswapd() || cgroup_reclaim(sc
))
3440 /* Throttle if making no progress at high prioities. */
3441 if (sc
->priority
== 1 && !sc
->nr_reclaimed
)
3442 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_NOPROGRESS
);
3446 * This is the direct reclaim path, for page-allocating processes. We only
3447 * try to reclaim pages from zones which will satisfy the caller's allocation
3450 * If a zone is deemed to be full of pinned pages then just give it a light
3451 * scan then give up on it.
3453 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
3457 unsigned long nr_soft_reclaimed
;
3458 unsigned long nr_soft_scanned
;
3460 pg_data_t
*last_pgdat
= NULL
;
3461 pg_data_t
*first_pgdat
= NULL
;
3464 * If the number of buffer_heads in the machine exceeds the maximum
3465 * allowed level, force direct reclaim to scan the highmem zone as
3466 * highmem pages could be pinning lowmem pages storing buffer_heads
3468 orig_mask
= sc
->gfp_mask
;
3469 if (buffer_heads_over_limit
) {
3470 sc
->gfp_mask
|= __GFP_HIGHMEM
;
3471 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
3474 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3475 sc
->reclaim_idx
, sc
->nodemask
) {
3477 * Take care memory controller reclaiming has small influence
3480 if (!cgroup_reclaim(sc
)) {
3481 if (!cpuset_zone_allowed(zone
,
3482 GFP_KERNEL
| __GFP_HARDWALL
))
3486 * If we already have plenty of memory free for
3487 * compaction in this zone, don't free any more.
3488 * Even though compaction is invoked for any
3489 * non-zero order, only frequent costly order
3490 * reclamation is disruptive enough to become a
3491 * noticeable problem, like transparent huge
3494 if (IS_ENABLED(CONFIG_COMPACTION
) &&
3495 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
3496 compaction_ready(zone
, sc
)) {
3497 sc
->compaction_ready
= true;
3502 * Shrink each node in the zonelist once. If the
3503 * zonelist is ordered by zone (not the default) then a
3504 * node may be shrunk multiple times but in that case
3505 * the user prefers lower zones being preserved.
3507 if (zone
->zone_pgdat
== last_pgdat
)
3511 * This steals pages from memory cgroups over softlimit
3512 * and returns the number of reclaimed pages and
3513 * scanned pages. This works for global memory pressure
3514 * and balancing, not for a memcg's limit.
3516 nr_soft_scanned
= 0;
3517 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
3518 sc
->order
, sc
->gfp_mask
,
3520 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
3521 sc
->nr_scanned
+= nr_soft_scanned
;
3522 /* need some check for avoid more shrink_zone() */
3526 first_pgdat
= zone
->zone_pgdat
;
3528 /* See comment about same check for global reclaim above */
3529 if (zone
->zone_pgdat
== last_pgdat
)
3531 last_pgdat
= zone
->zone_pgdat
;
3532 shrink_node(zone
->zone_pgdat
, sc
);
3536 consider_reclaim_throttle(first_pgdat
, sc
);
3539 * Restore to original mask to avoid the impact on the caller if we
3540 * promoted it to __GFP_HIGHMEM.
3542 sc
->gfp_mask
= orig_mask
;
3545 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
3547 struct lruvec
*target_lruvec
;
3548 unsigned long refaults
;
3550 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
3551 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
3552 target_lruvec
->refaults
[0] = refaults
;
3553 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
3554 target_lruvec
->refaults
[1] = refaults
;
3558 * This is the main entry point to direct page reclaim.
3560 * If a full scan of the inactive list fails to free enough memory then we
3561 * are "out of memory" and something needs to be killed.
3563 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3564 * high - the zone may be full of dirty or under-writeback pages, which this
3565 * caller can't do much about. We kick the writeback threads and take explicit
3566 * naps in the hope that some of these pages can be written. But if the
3567 * allocating task holds filesystem locks which prevent writeout this might not
3568 * work, and the allocation attempt will fail.
3570 * returns: 0, if no pages reclaimed
3571 * else, the number of pages reclaimed
3573 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3574 struct scan_control
*sc
)
3576 int initial_priority
= sc
->priority
;
3577 pg_data_t
*last_pgdat
;
3581 delayacct_freepages_start();
3583 if (!cgroup_reclaim(sc
))
3584 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3587 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3590 shrink_zones(zonelist
, sc
);
3592 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3595 if (sc
->compaction_ready
)
3599 * If we're getting trouble reclaiming, start doing
3600 * writepage even in laptop mode.
3602 if (sc
->priority
< DEF_PRIORITY
- 2)
3603 sc
->may_writepage
= 1;
3604 } while (--sc
->priority
>= 0);
3607 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3609 if (zone
->zone_pgdat
== last_pgdat
)
3611 last_pgdat
= zone
->zone_pgdat
;
3613 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3615 if (cgroup_reclaim(sc
)) {
3616 struct lruvec
*lruvec
;
3618 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3620 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3624 delayacct_freepages_end();
3626 if (sc
->nr_reclaimed
)
3627 return sc
->nr_reclaimed
;
3629 /* Aborted reclaim to try compaction? don't OOM, then */
3630 if (sc
->compaction_ready
)
3634 * We make inactive:active ratio decisions based on the node's
3635 * composition of memory, but a restrictive reclaim_idx or a
3636 * memory.low cgroup setting can exempt large amounts of
3637 * memory from reclaim. Neither of which are very common, so
3638 * instead of doing costly eligibility calculations of the
3639 * entire cgroup subtree up front, we assume the estimates are
3640 * good, and retry with forcible deactivation if that fails.
3642 if (sc
->skipped_deactivate
) {
3643 sc
->priority
= initial_priority
;
3644 sc
->force_deactivate
= 1;
3645 sc
->skipped_deactivate
= 0;
3649 /* Untapped cgroup reserves? Don't OOM, retry. */
3650 if (sc
->memcg_low_skipped
) {
3651 sc
->priority
= initial_priority
;
3652 sc
->force_deactivate
= 0;
3653 sc
->memcg_low_reclaim
= 1;
3654 sc
->memcg_low_skipped
= 0;
3661 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3664 unsigned long pfmemalloc_reserve
= 0;
3665 unsigned long free_pages
= 0;
3669 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3672 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3673 zone
= &pgdat
->node_zones
[i
];
3674 if (!managed_zone(zone
))
3677 if (!zone_reclaimable_pages(zone
))
3680 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3681 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3684 /* If there are no reserves (unexpected config) then do not throttle */
3685 if (!pfmemalloc_reserve
)
3688 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3690 /* kswapd must be awake if processes are being throttled */
3691 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3692 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
3693 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
3695 wake_up_interruptible(&pgdat
->kswapd_wait
);
3702 * Throttle direct reclaimers if backing storage is backed by the network
3703 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3704 * depleted. kswapd will continue to make progress and wake the processes
3705 * when the low watermark is reached.
3707 * Returns true if a fatal signal was delivered during throttling. If this
3708 * happens, the page allocator should not consider triggering the OOM killer.
3710 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3711 nodemask_t
*nodemask
)
3715 pg_data_t
*pgdat
= NULL
;
3718 * Kernel threads should not be throttled as they may be indirectly
3719 * responsible for cleaning pages necessary for reclaim to make forward
3720 * progress. kjournald for example may enter direct reclaim while
3721 * committing a transaction where throttling it could forcing other
3722 * processes to block on log_wait_commit().
3724 if (current
->flags
& PF_KTHREAD
)
3728 * If a fatal signal is pending, this process should not throttle.
3729 * It should return quickly so it can exit and free its memory
3731 if (fatal_signal_pending(current
))
3735 * Check if the pfmemalloc reserves are ok by finding the first node
3736 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3737 * GFP_KERNEL will be required for allocating network buffers when
3738 * swapping over the network so ZONE_HIGHMEM is unusable.
3740 * Throttling is based on the first usable node and throttled processes
3741 * wait on a queue until kswapd makes progress and wakes them. There
3742 * is an affinity then between processes waking up and where reclaim
3743 * progress has been made assuming the process wakes on the same node.
3744 * More importantly, processes running on remote nodes will not compete
3745 * for remote pfmemalloc reserves and processes on different nodes
3746 * should make reasonable progress.
3748 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3749 gfp_zone(gfp_mask
), nodemask
) {
3750 if (zone_idx(zone
) > ZONE_NORMAL
)
3753 /* Throttle based on the first usable node */
3754 pgdat
= zone
->zone_pgdat
;
3755 if (allow_direct_reclaim(pgdat
))
3760 /* If no zone was usable by the allocation flags then do not throttle */
3764 /* Account for the throttling */
3765 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3768 * If the caller cannot enter the filesystem, it's possible that it
3769 * is due to the caller holding an FS lock or performing a journal
3770 * transaction in the case of a filesystem like ext[3|4]. In this case,
3771 * it is not safe to block on pfmemalloc_wait as kswapd could be
3772 * blocked waiting on the same lock. Instead, throttle for up to a
3773 * second before continuing.
3775 if (!(gfp_mask
& __GFP_FS
))
3776 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3777 allow_direct_reclaim(pgdat
), HZ
);
3779 /* Throttle until kswapd wakes the process */
3780 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3781 allow_direct_reclaim(pgdat
));
3783 if (fatal_signal_pending(current
))
3790 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3791 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3793 unsigned long nr_reclaimed
;
3794 struct scan_control sc
= {
3795 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3796 .gfp_mask
= current_gfp_context(gfp_mask
),
3797 .reclaim_idx
= gfp_zone(gfp_mask
),
3799 .nodemask
= nodemask
,
3800 .priority
= DEF_PRIORITY
,
3801 .may_writepage
= !laptop_mode
,
3807 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3808 * Confirm they are large enough for max values.
3810 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3811 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3812 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3815 * Do not enter reclaim if fatal signal was delivered while throttled.
3816 * 1 is returned so that the page allocator does not OOM kill at this
3819 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3822 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3823 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3825 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3827 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3828 set_task_reclaim_state(current
, NULL
);
3830 return nr_reclaimed
;
3835 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3836 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3837 gfp_t gfp_mask
, bool noswap
,
3839 unsigned long *nr_scanned
)
3841 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3842 struct scan_control sc
= {
3843 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3844 .target_mem_cgroup
= memcg
,
3845 .may_writepage
= !laptop_mode
,
3847 .reclaim_idx
= MAX_NR_ZONES
- 1,
3848 .may_swap
= !noswap
,
3851 WARN_ON_ONCE(!current
->reclaim_state
);
3853 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3854 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3856 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3860 * NOTE: Although we can get the priority field, using it
3861 * here is not a good idea, since it limits the pages we can scan.
3862 * if we don't reclaim here, the shrink_node from balance_pgdat
3863 * will pick up pages from other mem cgroup's as well. We hack
3864 * the priority and make it zero.
3866 shrink_lruvec(lruvec
, &sc
);
3868 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3870 *nr_scanned
= sc
.nr_scanned
;
3872 return sc
.nr_reclaimed
;
3875 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3876 unsigned long nr_pages
,
3880 unsigned long nr_reclaimed
;
3881 unsigned int noreclaim_flag
;
3882 struct scan_control sc
= {
3883 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3884 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3885 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3886 .reclaim_idx
= MAX_NR_ZONES
- 1,
3887 .target_mem_cgroup
= memcg
,
3888 .priority
= DEF_PRIORITY
,
3889 .may_writepage
= !laptop_mode
,
3891 .may_swap
= may_swap
,
3894 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3895 * equal pressure on all the nodes. This is based on the assumption that
3896 * the reclaim does not bail out early.
3898 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3900 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3901 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3902 noreclaim_flag
= memalloc_noreclaim_save();
3904 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3906 memalloc_noreclaim_restore(noreclaim_flag
);
3907 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3908 set_task_reclaim_state(current
, NULL
);
3910 return nr_reclaimed
;
3914 static void age_active_anon(struct pglist_data
*pgdat
,
3915 struct scan_control
*sc
)
3917 struct mem_cgroup
*memcg
;
3918 struct lruvec
*lruvec
;
3920 if (!can_age_anon_pages(pgdat
, sc
))
3923 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3924 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3927 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3929 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3930 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3931 sc
, LRU_ACTIVE_ANON
);
3932 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3936 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
3942 * Check for watermark boosts top-down as the higher zones
3943 * are more likely to be boosted. Both watermarks and boosts
3944 * should not be checked at the same time as reclaim would
3945 * start prematurely when there is no boosting and a lower
3948 for (i
= highest_zoneidx
; i
>= 0; i
--) {
3949 zone
= pgdat
->node_zones
+ i
;
3950 if (!managed_zone(zone
))
3953 if (zone
->watermark_boost
)
3961 * Returns true if there is an eligible zone balanced for the request order
3962 * and highest_zoneidx
3964 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3967 unsigned long mark
= -1;
3971 * Check watermarks bottom-up as lower zones are more likely to
3974 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3975 zone
= pgdat
->node_zones
+ i
;
3977 if (!managed_zone(zone
))
3980 mark
= high_wmark_pages(zone
);
3981 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
3986 * If a node has no populated zone within highest_zoneidx, it does not
3987 * need balancing by definition. This can happen if a zone-restricted
3988 * allocation tries to wake a remote kswapd.
3996 /* Clear pgdat state for congested, dirty or under writeback. */
3997 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3999 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
4001 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
4002 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
4003 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
4007 * Prepare kswapd for sleeping. This verifies that there are no processes
4008 * waiting in throttle_direct_reclaim() and that watermarks have been met.
4010 * Returns true if kswapd is ready to sleep
4012 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
4013 int highest_zoneidx
)
4016 * The throttled processes are normally woken up in balance_pgdat() as
4017 * soon as allow_direct_reclaim() is true. But there is a potential
4018 * race between when kswapd checks the watermarks and a process gets
4019 * throttled. There is also a potential race if processes get
4020 * throttled, kswapd wakes, a large process exits thereby balancing the
4021 * zones, which causes kswapd to exit balance_pgdat() before reaching
4022 * the wake up checks. If kswapd is going to sleep, no process should
4023 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
4024 * the wake up is premature, processes will wake kswapd and get
4025 * throttled again. The difference from wake ups in balance_pgdat() is
4026 * that here we are under prepare_to_wait().
4028 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
4029 wake_up_all(&pgdat
->pfmemalloc_wait
);
4031 /* Hopeless node, leave it to direct reclaim */
4032 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
4035 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
4036 clear_pgdat_congested(pgdat
);
4044 * kswapd shrinks a node of pages that are at or below the highest usable
4045 * zone that is currently unbalanced.
4047 * Returns true if kswapd scanned at least the requested number of pages to
4048 * reclaim or if the lack of progress was due to pages under writeback.
4049 * This is used to determine if the scanning priority needs to be raised.
4051 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
4052 struct scan_control
*sc
)
4057 /* Reclaim a number of pages proportional to the number of zones */
4058 sc
->nr_to_reclaim
= 0;
4059 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
4060 zone
= pgdat
->node_zones
+ z
;
4061 if (!managed_zone(zone
))
4064 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
4068 * Historically care was taken to put equal pressure on all zones but
4069 * now pressure is applied based on node LRU order.
4071 shrink_node(pgdat
, sc
);
4074 * Fragmentation may mean that the system cannot be rebalanced for
4075 * high-order allocations. If twice the allocation size has been
4076 * reclaimed then recheck watermarks only at order-0 to prevent
4077 * excessive reclaim. Assume that a process requested a high-order
4078 * can direct reclaim/compact.
4080 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
4083 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
4086 /* Page allocator PCP high watermark is lowered if reclaim is active. */
4088 update_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
, bool active
)
4093 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4094 zone
= pgdat
->node_zones
+ i
;
4096 if (!managed_zone(zone
))
4100 set_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4102 clear_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4107 set_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4109 update_reclaim_active(pgdat
, highest_zoneidx
, true);
4113 clear_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4115 update_reclaim_active(pgdat
, highest_zoneidx
, false);
4119 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
4120 * that are eligible for use by the caller until at least one zone is
4123 * Returns the order kswapd finished reclaiming at.
4125 * kswapd scans the zones in the highmem->normal->dma direction. It skips
4126 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
4127 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
4128 * or lower is eligible for reclaim until at least one usable zone is
4131 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
4134 unsigned long nr_soft_reclaimed
;
4135 unsigned long nr_soft_scanned
;
4136 unsigned long pflags
;
4137 unsigned long nr_boost_reclaim
;
4138 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
4141 struct scan_control sc
= {
4142 .gfp_mask
= GFP_KERNEL
,
4147 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4148 psi_memstall_enter(&pflags
);
4149 __fs_reclaim_acquire(_THIS_IP_
);
4151 count_vm_event(PAGEOUTRUN
);
4154 * Account for the reclaim boost. Note that the zone boost is left in
4155 * place so that parallel allocations that are near the watermark will
4156 * stall or direct reclaim until kswapd is finished.
4158 nr_boost_reclaim
= 0;
4159 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4160 zone
= pgdat
->node_zones
+ i
;
4161 if (!managed_zone(zone
))
4164 nr_boost_reclaim
+= zone
->watermark_boost
;
4165 zone_boosts
[i
] = zone
->watermark_boost
;
4167 boosted
= nr_boost_reclaim
;
4170 set_reclaim_active(pgdat
, highest_zoneidx
);
4171 sc
.priority
= DEF_PRIORITY
;
4173 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
4174 bool raise_priority
= true;
4178 sc
.reclaim_idx
= highest_zoneidx
;
4181 * If the number of buffer_heads exceeds the maximum allowed
4182 * then consider reclaiming from all zones. This has a dual
4183 * purpose -- on 64-bit systems it is expected that
4184 * buffer_heads are stripped during active rotation. On 32-bit
4185 * systems, highmem pages can pin lowmem memory and shrinking
4186 * buffers can relieve lowmem pressure. Reclaim may still not
4187 * go ahead if all eligible zones for the original allocation
4188 * request are balanced to avoid excessive reclaim from kswapd.
4190 if (buffer_heads_over_limit
) {
4191 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
4192 zone
= pgdat
->node_zones
+ i
;
4193 if (!managed_zone(zone
))
4202 * If the pgdat is imbalanced then ignore boosting and preserve
4203 * the watermarks for a later time and restart. Note that the
4204 * zone watermarks will be still reset at the end of balancing
4205 * on the grounds that the normal reclaim should be enough to
4206 * re-evaluate if boosting is required when kswapd next wakes.
4208 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
4209 if (!balanced
&& nr_boost_reclaim
) {
4210 nr_boost_reclaim
= 0;
4215 * If boosting is not active then only reclaim if there are no
4216 * eligible zones. Note that sc.reclaim_idx is not used as
4217 * buffer_heads_over_limit may have adjusted it.
4219 if (!nr_boost_reclaim
&& balanced
)
4222 /* Limit the priority of boosting to avoid reclaim writeback */
4223 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
4224 raise_priority
= false;
4227 * Do not writeback or swap pages for boosted reclaim. The
4228 * intent is to relieve pressure not issue sub-optimal IO
4229 * from reclaim context. If no pages are reclaimed, the
4230 * reclaim will be aborted.
4232 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
4233 sc
.may_swap
= !nr_boost_reclaim
;
4236 * Do some background aging of the anon list, to give
4237 * pages a chance to be referenced before reclaiming. All
4238 * pages are rotated regardless of classzone as this is
4239 * about consistent aging.
4241 age_active_anon(pgdat
, &sc
);
4244 * If we're getting trouble reclaiming, start doing writepage
4245 * even in laptop mode.
4247 if (sc
.priority
< DEF_PRIORITY
- 2)
4248 sc
.may_writepage
= 1;
4250 /* Call soft limit reclaim before calling shrink_node. */
4252 nr_soft_scanned
= 0;
4253 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
4254 sc
.gfp_mask
, &nr_soft_scanned
);
4255 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
4258 * There should be no need to raise the scanning priority if
4259 * enough pages are already being scanned that that high
4260 * watermark would be met at 100% efficiency.
4262 if (kswapd_shrink_node(pgdat
, &sc
))
4263 raise_priority
= false;
4266 * If the low watermark is met there is no need for processes
4267 * to be throttled on pfmemalloc_wait as they should not be
4268 * able to safely make forward progress. Wake them
4270 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
4271 allow_direct_reclaim(pgdat
))
4272 wake_up_all(&pgdat
->pfmemalloc_wait
);
4274 /* Check if kswapd should be suspending */
4275 __fs_reclaim_release(_THIS_IP_
);
4276 ret
= try_to_freeze();
4277 __fs_reclaim_acquire(_THIS_IP_
);
4278 if (ret
|| kthread_should_stop())
4282 * Raise priority if scanning rate is too low or there was no
4283 * progress in reclaiming pages
4285 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
4286 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
4289 * If reclaim made no progress for a boost, stop reclaim as
4290 * IO cannot be queued and it could be an infinite loop in
4291 * extreme circumstances.
4293 if (nr_boost_reclaim
&& !nr_reclaimed
)
4296 if (raise_priority
|| !nr_reclaimed
)
4298 } while (sc
.priority
>= 1);
4300 if (!sc
.nr_reclaimed
)
4301 pgdat
->kswapd_failures
++;
4304 clear_reclaim_active(pgdat
, highest_zoneidx
);
4306 /* If reclaim was boosted, account for the reclaim done in this pass */
4308 unsigned long flags
;
4310 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4311 if (!zone_boosts
[i
])
4314 /* Increments are under the zone lock */
4315 zone
= pgdat
->node_zones
+ i
;
4316 spin_lock_irqsave(&zone
->lock
, flags
);
4317 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
4318 spin_unlock_irqrestore(&zone
->lock
, flags
);
4322 * As there is now likely space, wakeup kcompact to defragment
4325 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
4328 snapshot_refaults(NULL
, pgdat
);
4329 __fs_reclaim_release(_THIS_IP_
);
4330 psi_memstall_leave(&pflags
);
4331 set_task_reclaim_state(current
, NULL
);
4334 * Return the order kswapd stopped reclaiming at as
4335 * prepare_kswapd_sleep() takes it into account. If another caller
4336 * entered the allocator slow path while kswapd was awake, order will
4337 * remain at the higher level.
4343 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4344 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4345 * not a valid index then either kswapd runs for first time or kswapd couldn't
4346 * sleep after previous reclaim attempt (node is still unbalanced). In that
4347 * case return the zone index of the previous kswapd reclaim cycle.
4349 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
4350 enum zone_type prev_highest_zoneidx
)
4352 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4354 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
4357 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
4358 unsigned int highest_zoneidx
)
4363 if (freezing(current
) || kthread_should_stop())
4366 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4369 * Try to sleep for a short interval. Note that kcompactd will only be
4370 * woken if it is possible to sleep for a short interval. This is
4371 * deliberate on the assumption that if reclaim cannot keep an
4372 * eligible zone balanced that it's also unlikely that compaction will
4375 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4377 * Compaction records what page blocks it recently failed to
4378 * isolate pages from and skips them in the future scanning.
4379 * When kswapd is going to sleep, it is reasonable to assume
4380 * that pages and compaction may succeed so reset the cache.
4382 reset_isolation_suitable(pgdat
);
4385 * We have freed the memory, now we should compact it to make
4386 * allocation of the requested order possible.
4388 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
4390 remaining
= schedule_timeout(HZ
/10);
4393 * If woken prematurely then reset kswapd_highest_zoneidx and
4394 * order. The values will either be from a wakeup request or
4395 * the previous request that slept prematurely.
4398 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
4399 kswapd_highest_zoneidx(pgdat
,
4402 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
4403 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
4406 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4407 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4411 * After a short sleep, check if it was a premature sleep. If not, then
4412 * go fully to sleep until explicitly woken up.
4415 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4416 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
4419 * vmstat counters are not perfectly accurate and the estimated
4420 * value for counters such as NR_FREE_PAGES can deviate from the
4421 * true value by nr_online_cpus * threshold. To avoid the zone
4422 * watermarks being breached while under pressure, we reduce the
4423 * per-cpu vmstat threshold while kswapd is awake and restore
4424 * them before going back to sleep.
4426 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
4428 if (!kthread_should_stop())
4431 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
4434 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
4436 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
4438 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4442 * The background pageout daemon, started as a kernel thread
4443 * from the init process.
4445 * This basically trickles out pages so that we have _some_
4446 * free memory available even if there is no other activity
4447 * that frees anything up. This is needed for things like routing
4448 * etc, where we otherwise might have all activity going on in
4449 * asynchronous contexts that cannot page things out.
4451 * If there are applications that are active memory-allocators
4452 * (most normal use), this basically shouldn't matter.
4454 static int kswapd(void *p
)
4456 unsigned int alloc_order
, reclaim_order
;
4457 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
4458 pg_data_t
*pgdat
= (pg_data_t
*)p
;
4459 struct task_struct
*tsk
= current
;
4460 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
4462 if (!cpumask_empty(cpumask
))
4463 set_cpus_allowed_ptr(tsk
, cpumask
);
4466 * Tell the memory management that we're a "memory allocator",
4467 * and that if we need more memory we should get access to it
4468 * regardless (see "__alloc_pages()"). "kswapd" should
4469 * never get caught in the normal page freeing logic.
4471 * (Kswapd normally doesn't need memory anyway, but sometimes
4472 * you need a small amount of memory in order to be able to
4473 * page out something else, and this flag essentially protects
4474 * us from recursively trying to free more memory as we're
4475 * trying to free the first piece of memory in the first place).
4477 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
4480 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4481 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4482 atomic_set(&pgdat
->nr_writeback_throttled
, 0);
4486 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
4487 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4491 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
4494 /* Read the new order and highest_zoneidx */
4495 alloc_order
= READ_ONCE(pgdat
->kswapd_order
);
4496 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4498 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4499 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4501 ret
= try_to_freeze();
4502 if (kthread_should_stop())
4506 * We can speed up thawing tasks if we don't call balance_pgdat
4507 * after returning from the refrigerator
4513 * Reclaim begins at the requested order but if a high-order
4514 * reclaim fails then kswapd falls back to reclaiming for
4515 * order-0. If that happens, kswapd will consider sleeping
4516 * for the order it finished reclaiming at (reclaim_order)
4517 * but kcompactd is woken to compact for the original
4518 * request (alloc_order).
4520 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
4522 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
4524 if (reclaim_order
< alloc_order
)
4525 goto kswapd_try_sleep
;
4528 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
4534 * A zone is low on free memory or too fragmented for high-order memory. If
4535 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4536 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4537 * has failed or is not needed, still wake up kcompactd if only compaction is
4540 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
4541 enum zone_type highest_zoneidx
)
4544 enum zone_type curr_idx
;
4546 if (!managed_zone(zone
))
4549 if (!cpuset_zone_allowed(zone
, gfp_flags
))
4552 pgdat
= zone
->zone_pgdat
;
4553 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4555 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
4556 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
4558 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
4559 WRITE_ONCE(pgdat
->kswapd_order
, order
);
4561 if (!waitqueue_active(&pgdat
->kswapd_wait
))
4564 /* Hopeless node, leave it to direct reclaim if possible */
4565 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
4566 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
4567 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
4569 * There may be plenty of free memory available, but it's too
4570 * fragmented for high-order allocations. Wake up kcompactd
4571 * and rely on compaction_suitable() to determine if it's
4572 * needed. If it fails, it will defer subsequent attempts to
4573 * ratelimit its work.
4575 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
4576 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
4580 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
4582 wake_up_interruptible(&pgdat
->kswapd_wait
);
4585 #ifdef CONFIG_HIBERNATION
4587 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4590 * Rather than trying to age LRUs the aim is to preserve the overall
4591 * LRU order by reclaiming preferentially
4592 * inactive > active > active referenced > active mapped
4594 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4596 struct scan_control sc
= {
4597 .nr_to_reclaim
= nr_to_reclaim
,
4598 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4599 .reclaim_idx
= MAX_NR_ZONES
- 1,
4600 .priority
= DEF_PRIORITY
,
4604 .hibernation_mode
= 1,
4606 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4607 unsigned long nr_reclaimed
;
4608 unsigned int noreclaim_flag
;
4610 fs_reclaim_acquire(sc
.gfp_mask
);
4611 noreclaim_flag
= memalloc_noreclaim_save();
4612 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4614 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4616 set_task_reclaim_state(current
, NULL
);
4617 memalloc_noreclaim_restore(noreclaim_flag
);
4618 fs_reclaim_release(sc
.gfp_mask
);
4620 return nr_reclaimed
;
4622 #endif /* CONFIG_HIBERNATION */
4625 * This kswapd start function will be called by init and node-hot-add.
4626 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4628 void kswapd_run(int nid
)
4630 pg_data_t
*pgdat
= NODE_DATA(nid
);
4635 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4636 if (IS_ERR(pgdat
->kswapd
)) {
4637 /* failure at boot is fatal */
4638 BUG_ON(system_state
< SYSTEM_RUNNING
);
4639 pr_err("Failed to start kswapd on node %d\n", nid
);
4640 pgdat
->kswapd
= NULL
;
4645 * Called by memory hotplug when all memory in a node is offlined. Caller must
4646 * hold mem_hotplug_begin/end().
4648 void kswapd_stop(int nid
)
4650 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4653 kthread_stop(kswapd
);
4654 NODE_DATA(nid
)->kswapd
= NULL
;
4658 static int __init
kswapd_init(void)
4663 for_each_node_state(nid
, N_MEMORY
)
4668 module_init(kswapd_init
)
4674 * If non-zero call node_reclaim when the number of free pages falls below
4677 int node_reclaim_mode __read_mostly
;
4680 * Priority for NODE_RECLAIM. This determines the fraction of pages
4681 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4684 #define NODE_RECLAIM_PRIORITY 4
4687 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4690 int sysctl_min_unmapped_ratio
= 1;
4693 * If the number of slab pages in a zone grows beyond this percentage then
4694 * slab reclaim needs to occur.
4696 int sysctl_min_slab_ratio
= 5;
4698 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4700 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4701 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4702 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4705 * It's possible for there to be more file mapped pages than
4706 * accounted for by the pages on the file LRU lists because
4707 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4709 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4712 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4713 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4715 unsigned long nr_pagecache_reclaimable
;
4716 unsigned long delta
= 0;
4719 * If RECLAIM_UNMAP is set, then all file pages are considered
4720 * potentially reclaimable. Otherwise, we have to worry about
4721 * pages like swapcache and node_unmapped_file_pages() provides
4724 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4725 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4727 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4729 /* If we can't clean pages, remove dirty pages from consideration */
4730 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4731 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4733 /* Watch for any possible underflows due to delta */
4734 if (unlikely(delta
> nr_pagecache_reclaimable
))
4735 delta
= nr_pagecache_reclaimable
;
4737 return nr_pagecache_reclaimable
- delta
;
4741 * Try to free up some pages from this node through reclaim.
4743 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4745 /* Minimum pages needed in order to stay on node */
4746 const unsigned long nr_pages
= 1 << order
;
4747 struct task_struct
*p
= current
;
4748 unsigned int noreclaim_flag
;
4749 struct scan_control sc
= {
4750 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4751 .gfp_mask
= current_gfp_context(gfp_mask
),
4753 .priority
= NODE_RECLAIM_PRIORITY
,
4754 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4755 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4757 .reclaim_idx
= gfp_zone(gfp_mask
),
4759 unsigned long pflags
;
4761 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4765 psi_memstall_enter(&pflags
);
4766 fs_reclaim_acquire(sc
.gfp_mask
);
4768 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4769 * and we also need to be able to write out pages for RECLAIM_WRITE
4770 * and RECLAIM_UNMAP.
4772 noreclaim_flag
= memalloc_noreclaim_save();
4773 p
->flags
|= PF_SWAPWRITE
;
4774 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4776 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4778 * Free memory by calling shrink node with increasing
4779 * priorities until we have enough memory freed.
4782 shrink_node(pgdat
, &sc
);
4783 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4786 set_task_reclaim_state(p
, NULL
);
4787 current
->flags
&= ~PF_SWAPWRITE
;
4788 memalloc_noreclaim_restore(noreclaim_flag
);
4789 fs_reclaim_release(sc
.gfp_mask
);
4790 psi_memstall_leave(&pflags
);
4792 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4794 return sc
.nr_reclaimed
>= nr_pages
;
4797 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4802 * Node reclaim reclaims unmapped file backed pages and
4803 * slab pages if we are over the defined limits.
4805 * A small portion of unmapped file backed pages is needed for
4806 * file I/O otherwise pages read by file I/O will be immediately
4807 * thrown out if the node is overallocated. So we do not reclaim
4808 * if less than a specified percentage of the node is used by
4809 * unmapped file backed pages.
4811 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4812 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
4813 pgdat
->min_slab_pages
)
4814 return NODE_RECLAIM_FULL
;
4817 * Do not scan if the allocation should not be delayed.
4819 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4820 return NODE_RECLAIM_NOSCAN
;
4823 * Only run node reclaim on the local node or on nodes that do not
4824 * have associated processors. This will favor the local processor
4825 * over remote processors and spread off node memory allocations
4826 * as wide as possible.
4828 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4829 return NODE_RECLAIM_NOSCAN
;
4831 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4832 return NODE_RECLAIM_NOSCAN
;
4834 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4835 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4838 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4845 * check_move_unevictable_pages - check pages for evictability and move to
4846 * appropriate zone lru list
4847 * @pvec: pagevec with lru pages to check
4849 * Checks pages for evictability, if an evictable page is in the unevictable
4850 * lru list, moves it to the appropriate evictable lru list. This function
4851 * should be only used for lru pages.
4853 void check_move_unevictable_pages(struct pagevec
*pvec
)
4855 struct lruvec
*lruvec
= NULL
;
4860 for (i
= 0; i
< pvec
->nr
; i
++) {
4861 struct page
*page
= pvec
->pages
[i
];
4862 struct folio
*folio
= page_folio(page
);
4865 if (PageTransTail(page
))
4868 nr_pages
= thp_nr_pages(page
);
4869 pgscanned
+= nr_pages
;
4871 /* block memcg migration during page moving between lru */
4872 if (!TestClearPageLRU(page
))
4875 lruvec
= folio_lruvec_relock_irq(folio
, lruvec
);
4876 if (page_evictable(page
) && PageUnevictable(page
)) {
4877 del_page_from_lru_list(page
, lruvec
);
4878 ClearPageUnevictable(page
);
4879 add_page_to_lru_list(page
, lruvec
);
4880 pgrescued
+= nr_pages
;
4886 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4887 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4888 unlock_page_lruvec_irq(lruvec
);
4889 } else if (pgscanned
) {
4890 count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4893 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);