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>
59 #include <linux/sched/sysctl.h>
64 #define CREATE_TRACE_POINTS
65 #include <trace/events/vmscan.h>
68 /* How many pages shrink_list() should reclaim */
69 unsigned long nr_to_reclaim
;
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
81 struct mem_cgroup
*target_mem_cgroup
;
84 * Scan pressure balancing between anon and file LRUs
86 unsigned long anon_cost
;
87 unsigned long file_cost
;
89 /* Can active pages be deactivated as part of reclaim? */
90 #define DEACTIVATE_ANON 1
91 #define DEACTIVATE_FILE 2
92 unsigned int may_deactivate
:2;
93 unsigned int force_deactivate
:1;
94 unsigned int skipped_deactivate
:1;
96 /* Writepage batching in laptop mode; RECLAIM_WRITE */
97 unsigned int may_writepage
:1;
99 /* Can mapped pages be reclaimed? */
100 unsigned int may_unmap
:1;
102 /* Can pages be swapped as part of reclaim? */
103 unsigned int may_swap
:1;
106 * Cgroup memory below memory.low is protected as long as we
107 * don't threaten to OOM. If any cgroup is reclaimed at
108 * reduced force or passed over entirely due to its memory.low
109 * setting (memcg_low_skipped), and nothing is reclaimed as a
110 * result, then go back for one more cycle that reclaims the protected
111 * memory (memcg_low_reclaim) to avert OOM.
113 unsigned int memcg_low_reclaim
:1;
114 unsigned int memcg_low_skipped
:1;
116 unsigned int hibernation_mode
:1;
118 /* One of the zones is ready for compaction */
119 unsigned int compaction_ready
:1;
121 /* There is easily reclaimable cold cache in the current node */
122 unsigned int cache_trim_mode
:1;
124 /* The file pages on the current node are dangerously low */
125 unsigned int file_is_tiny
:1;
127 /* Always discard instead of demoting to lower tier memory */
128 unsigned int no_demotion
:1;
130 /* Allocation order */
133 /* Scan (total_size >> priority) pages at once */
136 /* The highest zone to isolate pages for reclaim from */
139 /* This context's GFP mask */
142 /* Incremented by the number of inactive pages that were scanned */
143 unsigned long nr_scanned
;
145 /* Number of pages freed so far during a call to shrink_zones() */
146 unsigned long nr_reclaimed
;
150 unsigned int unqueued_dirty
;
151 unsigned int congested
;
152 unsigned int writeback
;
153 unsigned int immediate
;
154 unsigned int file_taken
;
158 /* for recording the reclaimed slab by now */
159 struct reclaim_state reclaim_state
;
162 #ifdef ARCH_HAS_PREFETCHW
163 #define prefetchw_prev_lru_page(_page, _base, _field) \
165 if ((_page)->lru.prev != _base) { \
168 prev = lru_to_page(&(_page->lru)); \
169 prefetchw(&prev->_field); \
173 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
177 * From 0 .. 200. Higher means more swappy.
179 int vm_swappiness
= 60;
181 static void set_task_reclaim_state(struct task_struct
*task
,
182 struct reclaim_state
*rs
)
184 /* Check for an overwrite */
185 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
187 /* Check for the nulling of an already-nulled member */
188 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
190 task
->reclaim_state
= rs
;
193 static LIST_HEAD(shrinker_list
);
194 static DECLARE_RWSEM(shrinker_rwsem
);
197 static int shrinker_nr_max
;
199 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
200 static inline int shrinker_map_size(int nr_items
)
202 return (DIV_ROUND_UP(nr_items
, BITS_PER_LONG
) * sizeof(unsigned long));
205 static inline int shrinker_defer_size(int nr_items
)
207 return (round_up(nr_items
, BITS_PER_LONG
) * sizeof(atomic_long_t
));
210 static struct shrinker_info
*shrinker_info_protected(struct mem_cgroup
*memcg
,
213 return rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_info
,
214 lockdep_is_held(&shrinker_rwsem
));
217 static int expand_one_shrinker_info(struct mem_cgroup
*memcg
,
218 int map_size
, int defer_size
,
219 int old_map_size
, int old_defer_size
)
221 struct shrinker_info
*new, *old
;
222 struct mem_cgroup_per_node
*pn
;
224 int size
= map_size
+ defer_size
;
227 pn
= memcg
->nodeinfo
[nid
];
228 old
= shrinker_info_protected(memcg
, nid
);
229 /* Not yet online memcg */
233 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
237 new->nr_deferred
= (atomic_long_t
*)(new + 1);
238 new->map
= (void *)new->nr_deferred
+ defer_size
;
240 /* map: set all old bits, clear all new bits */
241 memset(new->map
, (int)0xff, old_map_size
);
242 memset((void *)new->map
+ old_map_size
, 0, map_size
- old_map_size
);
243 /* nr_deferred: copy old values, clear all new values */
244 memcpy(new->nr_deferred
, old
->nr_deferred
, old_defer_size
);
245 memset((void *)new->nr_deferred
+ old_defer_size
, 0,
246 defer_size
- old_defer_size
);
248 rcu_assign_pointer(pn
->shrinker_info
, new);
249 kvfree_rcu(old
, rcu
);
255 void free_shrinker_info(struct mem_cgroup
*memcg
)
257 struct mem_cgroup_per_node
*pn
;
258 struct shrinker_info
*info
;
262 pn
= memcg
->nodeinfo
[nid
];
263 info
= rcu_dereference_protected(pn
->shrinker_info
, true);
265 rcu_assign_pointer(pn
->shrinker_info
, NULL
);
269 int alloc_shrinker_info(struct mem_cgroup
*memcg
)
271 struct shrinker_info
*info
;
272 int nid
, size
, ret
= 0;
273 int map_size
, defer_size
= 0;
275 down_write(&shrinker_rwsem
);
276 map_size
= shrinker_map_size(shrinker_nr_max
);
277 defer_size
= shrinker_defer_size(shrinker_nr_max
);
278 size
= map_size
+ defer_size
;
280 info
= kvzalloc_node(sizeof(*info
) + size
, GFP_KERNEL
, nid
);
282 free_shrinker_info(memcg
);
286 info
->nr_deferred
= (atomic_long_t
*)(info
+ 1);
287 info
->map
= (void *)info
->nr_deferred
+ defer_size
;
288 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_info
, info
);
290 up_write(&shrinker_rwsem
);
295 static inline bool need_expand(int nr_max
)
297 return round_up(nr_max
, BITS_PER_LONG
) >
298 round_up(shrinker_nr_max
, BITS_PER_LONG
);
301 static int expand_shrinker_info(int new_id
)
304 int new_nr_max
= new_id
+ 1;
305 int map_size
, defer_size
= 0;
306 int old_map_size
, old_defer_size
= 0;
307 struct mem_cgroup
*memcg
;
309 if (!need_expand(new_nr_max
))
312 if (!root_mem_cgroup
)
315 lockdep_assert_held(&shrinker_rwsem
);
317 map_size
= shrinker_map_size(new_nr_max
);
318 defer_size
= shrinker_defer_size(new_nr_max
);
319 old_map_size
= shrinker_map_size(shrinker_nr_max
);
320 old_defer_size
= shrinker_defer_size(shrinker_nr_max
);
322 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
324 ret
= expand_one_shrinker_info(memcg
, map_size
, defer_size
,
325 old_map_size
, old_defer_size
);
327 mem_cgroup_iter_break(NULL
, memcg
);
330 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
333 shrinker_nr_max
= new_nr_max
;
338 void set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
340 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
341 struct shrinker_info
*info
;
344 info
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_info
);
345 /* Pairs with smp mb in shrink_slab() */
346 smp_mb__before_atomic();
347 set_bit(shrinker_id
, info
->map
);
352 static DEFINE_IDR(shrinker_idr
);
354 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
356 int id
, ret
= -ENOMEM
;
358 if (mem_cgroup_disabled())
361 down_write(&shrinker_rwsem
);
362 /* This may call shrinker, so it must use down_read_trylock() */
363 id
= idr_alloc(&shrinker_idr
, shrinker
, 0, 0, GFP_KERNEL
);
367 if (id
>= shrinker_nr_max
) {
368 if (expand_shrinker_info(id
)) {
369 idr_remove(&shrinker_idr
, id
);
376 up_write(&shrinker_rwsem
);
380 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
382 int id
= shrinker
->id
;
386 lockdep_assert_held(&shrinker_rwsem
);
388 idr_remove(&shrinker_idr
, id
);
391 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
392 struct mem_cgroup
*memcg
)
394 struct shrinker_info
*info
;
396 info
= shrinker_info_protected(memcg
, nid
);
397 return atomic_long_xchg(&info
->nr_deferred
[shrinker
->id
], 0);
400 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
401 struct mem_cgroup
*memcg
)
403 struct shrinker_info
*info
;
405 info
= shrinker_info_protected(memcg
, nid
);
406 return atomic_long_add_return(nr
, &info
->nr_deferred
[shrinker
->id
]);
409 void reparent_shrinker_deferred(struct mem_cgroup
*memcg
)
413 struct mem_cgroup
*parent
;
414 struct shrinker_info
*child_info
, *parent_info
;
416 parent
= parent_mem_cgroup(memcg
);
418 parent
= root_mem_cgroup
;
420 /* Prevent from concurrent shrinker_info expand */
421 down_read(&shrinker_rwsem
);
423 child_info
= shrinker_info_protected(memcg
, nid
);
424 parent_info
= shrinker_info_protected(parent
, nid
);
425 for (i
= 0; i
< shrinker_nr_max
; i
++) {
426 nr
= atomic_long_read(&child_info
->nr_deferred
[i
]);
427 atomic_long_add(nr
, &parent_info
->nr_deferred
[i
]);
430 up_read(&shrinker_rwsem
);
433 static bool cgroup_reclaim(struct scan_control
*sc
)
435 return sc
->target_mem_cgroup
;
439 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
440 * @sc: scan_control in question
442 * The normal page dirty throttling mechanism in balance_dirty_pages() is
443 * completely broken with the legacy memcg and direct stalling in
444 * shrink_page_list() is used for throttling instead, which lacks all the
445 * niceties such as fairness, adaptive pausing, bandwidth proportional
446 * allocation and configurability.
448 * This function tests whether the vmscan currently in progress can assume
449 * that the normal dirty throttling mechanism is operational.
451 static bool writeback_throttling_sane(struct scan_control
*sc
)
453 if (!cgroup_reclaim(sc
))
455 #ifdef CONFIG_CGROUP_WRITEBACK
456 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
462 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
467 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
471 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
472 struct mem_cgroup
*memcg
)
477 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
478 struct mem_cgroup
*memcg
)
483 static bool cgroup_reclaim(struct scan_control
*sc
)
488 static bool writeback_throttling_sane(struct scan_control
*sc
)
494 static long xchg_nr_deferred(struct shrinker
*shrinker
,
495 struct shrink_control
*sc
)
499 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
503 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
504 return xchg_nr_deferred_memcg(nid
, shrinker
,
507 return atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
511 static long add_nr_deferred(long nr
, struct shrinker
*shrinker
,
512 struct shrink_control
*sc
)
516 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
520 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
521 return add_nr_deferred_memcg(nr
, nid
, shrinker
,
524 return atomic_long_add_return(nr
, &shrinker
->nr_deferred
[nid
]);
527 static bool can_demote(int nid
, struct scan_control
*sc
)
529 if (!numa_demotion_enabled
)
534 /* It is pointless to do demotion in memcg reclaim */
535 if (cgroup_reclaim(sc
))
538 if (next_demotion_node(nid
) == NUMA_NO_NODE
)
544 static inline bool can_reclaim_anon_pages(struct mem_cgroup
*memcg
,
546 struct scan_control
*sc
)
550 * For non-memcg reclaim, is there
551 * space in any swap device?
553 if (get_nr_swap_pages() > 0)
556 /* Is the memcg below its swap limit? */
557 if (mem_cgroup_get_nr_swap_pages(memcg
) > 0)
562 * The page can not be swapped.
564 * Can it be reclaimed from this node via demotion?
566 return can_demote(nid
, sc
);
570 * This misses isolated pages which are not accounted for to save counters.
571 * As the data only determines if reclaim or compaction continues, it is
572 * not expected that isolated pages will be a dominating factor.
574 unsigned long zone_reclaimable_pages(struct zone
*zone
)
578 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
579 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
580 if (can_reclaim_anon_pages(NULL
, zone_to_nid(zone
), NULL
))
581 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
582 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
588 * lruvec_lru_size - Returns the number of pages on the given LRU list.
589 * @lruvec: lru vector
591 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
593 static unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
596 unsigned long size
= 0;
599 for (zid
= 0; zid
<= zone_idx
; zid
++) {
600 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
602 if (!managed_zone(zone
))
605 if (!mem_cgroup_disabled())
606 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
608 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
614 * Add a shrinker callback to be called from the vm.
616 int prealloc_shrinker(struct shrinker
*shrinker
)
621 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
622 err
= prealloc_memcg_shrinker(shrinker
);
626 shrinker
->flags
&= ~SHRINKER_MEMCG_AWARE
;
629 size
= sizeof(*shrinker
->nr_deferred
);
630 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
633 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
634 if (!shrinker
->nr_deferred
)
640 void free_prealloced_shrinker(struct shrinker
*shrinker
)
642 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
643 down_write(&shrinker_rwsem
);
644 unregister_memcg_shrinker(shrinker
);
645 up_write(&shrinker_rwsem
);
649 kfree(shrinker
->nr_deferred
);
650 shrinker
->nr_deferred
= NULL
;
653 void register_shrinker_prepared(struct shrinker
*shrinker
)
655 down_write(&shrinker_rwsem
);
656 list_add_tail(&shrinker
->list
, &shrinker_list
);
657 shrinker
->flags
|= SHRINKER_REGISTERED
;
658 up_write(&shrinker_rwsem
);
661 int register_shrinker(struct shrinker
*shrinker
)
663 int err
= prealloc_shrinker(shrinker
);
667 register_shrinker_prepared(shrinker
);
670 EXPORT_SYMBOL(register_shrinker
);
675 void unregister_shrinker(struct shrinker
*shrinker
)
677 if (!(shrinker
->flags
& SHRINKER_REGISTERED
))
680 down_write(&shrinker_rwsem
);
681 list_del(&shrinker
->list
);
682 shrinker
->flags
&= ~SHRINKER_REGISTERED
;
683 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
684 unregister_memcg_shrinker(shrinker
);
685 up_write(&shrinker_rwsem
);
687 kfree(shrinker
->nr_deferred
);
688 shrinker
->nr_deferred
= NULL
;
690 EXPORT_SYMBOL(unregister_shrinker
);
693 * synchronize_shrinkers - Wait for all running shrinkers to complete.
695 * This is equivalent to calling unregister_shrink() and register_shrinker(),
696 * but atomically and with less overhead. This is useful to guarantee that all
697 * shrinker invocations have seen an update, before freeing memory, similar to
700 void synchronize_shrinkers(void)
702 down_write(&shrinker_rwsem
);
703 up_write(&shrinker_rwsem
);
705 EXPORT_SYMBOL(synchronize_shrinkers
);
707 #define SHRINK_BATCH 128
709 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
710 struct shrinker
*shrinker
, int priority
)
712 unsigned long freed
= 0;
713 unsigned long long delta
;
718 long batch_size
= shrinker
->batch
? shrinker
->batch
720 long scanned
= 0, next_deferred
;
722 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
723 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
727 * copy the current shrinker scan count into a local variable
728 * and zero it so that other concurrent shrinker invocations
729 * don't also do this scanning work.
731 nr
= xchg_nr_deferred(shrinker
, shrinkctl
);
733 if (shrinker
->seeks
) {
734 delta
= freeable
>> priority
;
736 do_div(delta
, shrinker
->seeks
);
739 * These objects don't require any IO to create. Trim
740 * them aggressively under memory pressure to keep
741 * them from causing refetches in the IO caches.
743 delta
= freeable
/ 2;
746 total_scan
= nr
>> priority
;
748 total_scan
= min(total_scan
, (2 * freeable
));
750 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
751 freeable
, delta
, total_scan
, priority
);
754 * Normally, we should not scan less than batch_size objects in one
755 * pass to avoid too frequent shrinker calls, but if the slab has less
756 * than batch_size objects in total and we are really tight on memory,
757 * we will try to reclaim all available objects, otherwise we can end
758 * up failing allocations although there are plenty of reclaimable
759 * objects spread over several slabs with usage less than the
762 * We detect the "tight on memory" situations by looking at the total
763 * number of objects we want to scan (total_scan). If it is greater
764 * than the total number of objects on slab (freeable), we must be
765 * scanning at high prio and therefore should try to reclaim as much as
768 while (total_scan
>= batch_size
||
769 total_scan
>= freeable
) {
771 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
773 shrinkctl
->nr_to_scan
= nr_to_scan
;
774 shrinkctl
->nr_scanned
= nr_to_scan
;
775 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
776 if (ret
== SHRINK_STOP
)
780 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
781 total_scan
-= shrinkctl
->nr_scanned
;
782 scanned
+= shrinkctl
->nr_scanned
;
788 * The deferred work is increased by any new work (delta) that wasn't
789 * done, decreased by old deferred work that was done now.
791 * And it is capped to two times of the freeable items.
793 next_deferred
= max_t(long, (nr
+ delta
- scanned
), 0);
794 next_deferred
= min(next_deferred
, (2 * freeable
));
797 * move the unused scan count back into the shrinker in a
798 * manner that handles concurrent updates.
800 new_nr
= add_nr_deferred(next_deferred
, shrinker
, shrinkctl
);
802 trace_mm_shrink_slab_end(shrinker
, shrinkctl
->nid
, freed
, nr
, new_nr
, total_scan
);
807 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
808 struct mem_cgroup
*memcg
, int priority
)
810 struct shrinker_info
*info
;
811 unsigned long ret
, freed
= 0;
814 if (!mem_cgroup_online(memcg
))
817 if (!down_read_trylock(&shrinker_rwsem
))
820 info
= shrinker_info_protected(memcg
, nid
);
824 for_each_set_bit(i
, info
->map
, shrinker_nr_max
) {
825 struct shrink_control sc
= {
826 .gfp_mask
= gfp_mask
,
830 struct shrinker
*shrinker
;
832 shrinker
= idr_find(&shrinker_idr
, i
);
833 if (unlikely(!shrinker
|| !(shrinker
->flags
& SHRINKER_REGISTERED
))) {
835 clear_bit(i
, info
->map
);
839 /* Call non-slab shrinkers even though kmem is disabled */
840 if (!memcg_kmem_enabled() &&
841 !(shrinker
->flags
& SHRINKER_NONSLAB
))
844 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
845 if (ret
== SHRINK_EMPTY
) {
846 clear_bit(i
, info
->map
);
848 * After the shrinker reported that it had no objects to
849 * free, but before we cleared the corresponding bit in
850 * the memcg shrinker map, a new object might have been
851 * added. To make sure, we have the bit set in this
852 * case, we invoke the shrinker one more time and reset
853 * the bit if it reports that it is not empty anymore.
854 * The memory barrier here pairs with the barrier in
855 * set_shrinker_bit():
857 * list_lru_add() shrink_slab_memcg()
858 * list_add_tail() clear_bit()
860 * set_bit() do_shrink_slab()
862 smp_mb__after_atomic();
863 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
864 if (ret
== SHRINK_EMPTY
)
867 set_shrinker_bit(memcg
, nid
, i
);
871 if (rwsem_is_contended(&shrinker_rwsem
)) {
877 up_read(&shrinker_rwsem
);
880 #else /* CONFIG_MEMCG */
881 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
882 struct mem_cgroup
*memcg
, int priority
)
886 #endif /* CONFIG_MEMCG */
889 * shrink_slab - shrink slab caches
890 * @gfp_mask: allocation context
891 * @nid: node whose slab caches to target
892 * @memcg: memory cgroup whose slab caches to target
893 * @priority: the reclaim priority
895 * Call the shrink functions to age shrinkable caches.
897 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
898 * unaware shrinkers will receive a node id of 0 instead.
900 * @memcg specifies the memory cgroup to target. Unaware shrinkers
901 * are called only if it is the root cgroup.
903 * @priority is sc->priority, we take the number of objects and >> by priority
904 * in order to get the scan target.
906 * Returns the number of reclaimed slab objects.
908 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
909 struct mem_cgroup
*memcg
,
912 unsigned long ret
, freed
= 0;
913 struct shrinker
*shrinker
;
916 * The root memcg might be allocated even though memcg is disabled
917 * via "cgroup_disable=memory" boot parameter. This could make
918 * mem_cgroup_is_root() return false, then just run memcg slab
919 * shrink, but skip global shrink. This may result in premature
922 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
923 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
925 if (!down_read_trylock(&shrinker_rwsem
))
928 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
929 struct shrink_control sc
= {
930 .gfp_mask
= gfp_mask
,
935 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
936 if (ret
== SHRINK_EMPTY
)
940 * Bail out if someone want to register a new shrinker to
941 * prevent the registration from being stalled for long periods
942 * by parallel ongoing shrinking.
944 if (rwsem_is_contended(&shrinker_rwsem
)) {
950 up_read(&shrinker_rwsem
);
956 static void drop_slab_node(int nid
)
962 struct mem_cgroup
*memcg
= NULL
;
964 if (fatal_signal_pending(current
))
968 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
970 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
971 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
972 } while ((freed
>> shift
++) > 1);
979 for_each_online_node(nid
)
983 static inline int is_page_cache_freeable(struct folio
*folio
)
986 * A freeable page cache page is referenced only by the caller
987 * that isolated the page, the page cache and optional buffer
988 * heads at page->private.
990 return folio_ref_count(folio
) - folio_test_private(folio
) ==
991 1 + folio_nr_pages(folio
);
995 * We detected a synchronous write error writing a folio out. Probably
996 * -ENOSPC. We need to propagate that into the address_space for a subsequent
997 * fsync(), msync() or close().
999 * The tricky part is that after writepage we cannot touch the mapping: nothing
1000 * prevents it from being freed up. But we have a ref on the folio and once
1001 * that folio is locked, the mapping is pinned.
1003 * We're allowed to run sleeping folio_lock() here because we know the caller has
1006 static void handle_write_error(struct address_space
*mapping
,
1007 struct folio
*folio
, int error
)
1010 if (folio_mapping(folio
) == mapping
)
1011 mapping_set_error(mapping
, error
);
1012 folio_unlock(folio
);
1015 static bool skip_throttle_noprogress(pg_data_t
*pgdat
)
1017 int reclaimable
= 0, write_pending
= 0;
1021 * If kswapd is disabled, reschedule if necessary but do not
1022 * throttle as the system is likely near OOM.
1024 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
1028 * If there are a lot of dirty/writeback pages then do not
1029 * throttle as throttling will occur when the pages cycle
1030 * towards the end of the LRU if still under writeback.
1032 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1033 struct zone
*zone
= pgdat
->node_zones
+ i
;
1035 if (!managed_zone(zone
))
1038 reclaimable
+= zone_reclaimable_pages(zone
);
1039 write_pending
+= zone_page_state_snapshot(zone
,
1040 NR_ZONE_WRITE_PENDING
);
1042 if (2 * write_pending
<= reclaimable
)
1048 void reclaim_throttle(pg_data_t
*pgdat
, enum vmscan_throttle_state reason
)
1050 wait_queue_head_t
*wqh
= &pgdat
->reclaim_wait
[reason
];
1055 * Do not throttle IO workers, kthreads other than kswapd or
1056 * workqueues. They may be required for reclaim to make
1057 * forward progress (e.g. journalling workqueues or kthreads).
1059 if (!current_is_kswapd() &&
1060 current
->flags
& (PF_IO_WORKER
|PF_KTHREAD
)) {
1066 * These figures are pulled out of thin air.
1067 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1068 * parallel reclaimers which is a short-lived event so the timeout is
1069 * short. Failing to make progress or waiting on writeback are
1070 * potentially long-lived events so use a longer timeout. This is shaky
1071 * logic as a failure to make progress could be due to anything from
1072 * writeback to a slow device to excessive references pages at the tail
1073 * of the inactive LRU.
1076 case VMSCAN_THROTTLE_WRITEBACK
:
1079 if (atomic_inc_return(&pgdat
->nr_writeback_throttled
) == 1) {
1080 WRITE_ONCE(pgdat
->nr_reclaim_start
,
1081 node_page_state(pgdat
, NR_THROTTLED_WRITTEN
));
1085 case VMSCAN_THROTTLE_CONGESTED
:
1087 case VMSCAN_THROTTLE_NOPROGRESS
:
1088 if (skip_throttle_noprogress(pgdat
)) {
1096 case VMSCAN_THROTTLE_ISOLATED
:
1105 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1106 ret
= schedule_timeout(timeout
);
1107 finish_wait(wqh
, &wait
);
1109 if (reason
== VMSCAN_THROTTLE_WRITEBACK
)
1110 atomic_dec(&pgdat
->nr_writeback_throttled
);
1112 trace_mm_vmscan_throttled(pgdat
->node_id
, jiffies_to_usecs(timeout
),
1113 jiffies_to_usecs(timeout
- ret
),
1118 * Account for pages written if tasks are throttled waiting on dirty
1119 * pages to clean. If enough pages have been cleaned since throttling
1120 * started then wakeup the throttled tasks.
1122 void __acct_reclaim_writeback(pg_data_t
*pgdat
, struct folio
*folio
,
1125 unsigned long nr_written
;
1127 node_stat_add_folio(folio
, NR_THROTTLED_WRITTEN
);
1130 * This is an inaccurate read as the per-cpu deltas may not
1131 * be synchronised. However, given that the system is
1132 * writeback throttled, it is not worth taking the penalty
1133 * of getting an accurate count. At worst, the throttle
1134 * timeout guarantees forward progress.
1136 nr_written
= node_page_state(pgdat
, NR_THROTTLED_WRITTEN
) -
1137 READ_ONCE(pgdat
->nr_reclaim_start
);
1139 if (nr_written
> SWAP_CLUSTER_MAX
* nr_throttled
)
1140 wake_up(&pgdat
->reclaim_wait
[VMSCAN_THROTTLE_WRITEBACK
]);
1143 /* possible outcome of pageout() */
1145 /* failed to write page out, page is locked */
1147 /* move page to the active list, page is locked */
1149 /* page has been sent to the disk successfully, page is unlocked */
1151 /* page is clean and locked */
1156 * pageout is called by shrink_page_list() for each dirty page.
1157 * Calls ->writepage().
1159 static pageout_t
pageout(struct folio
*folio
, struct address_space
*mapping
,
1160 struct swap_iocb
**plug
)
1163 * If the folio is dirty, only perform writeback if that write
1164 * will be non-blocking. To prevent this allocation from being
1165 * stalled by pagecache activity. But note that there may be
1166 * stalls if we need to run get_block(). We could test
1167 * PagePrivate for that.
1169 * If this process is currently in __generic_file_write_iter() against
1170 * this folio's queue, we can perform writeback even if that
1173 * If the folio is swapcache, write it back even if that would
1174 * block, for some throttling. This happens by accident, because
1175 * swap_backing_dev_info is bust: it doesn't reflect the
1176 * congestion state of the swapdevs. Easy to fix, if needed.
1178 if (!is_page_cache_freeable(folio
))
1182 * Some data journaling orphaned folios can have
1183 * folio->mapping == NULL while being dirty with clean buffers.
1185 if (folio_test_private(folio
)) {
1186 if (try_to_free_buffers(&folio
->page
)) {
1187 folio_clear_dirty(folio
);
1188 pr_info("%s: orphaned folio\n", __func__
);
1194 if (mapping
->a_ops
->writepage
== NULL
)
1195 return PAGE_ACTIVATE
;
1197 if (folio_clear_dirty_for_io(folio
)) {
1199 struct writeback_control wbc
= {
1200 .sync_mode
= WB_SYNC_NONE
,
1201 .nr_to_write
= SWAP_CLUSTER_MAX
,
1203 .range_end
= LLONG_MAX
,
1208 folio_set_reclaim(folio
);
1209 res
= mapping
->a_ops
->writepage(&folio
->page
, &wbc
);
1211 handle_write_error(mapping
, folio
, res
);
1212 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
1213 folio_clear_reclaim(folio
);
1214 return PAGE_ACTIVATE
;
1217 if (!folio_test_writeback(folio
)) {
1218 /* synchronous write or broken a_ops? */
1219 folio_clear_reclaim(folio
);
1221 trace_mm_vmscan_write_folio(folio
);
1222 node_stat_add_folio(folio
, NR_VMSCAN_WRITE
);
1223 return PAGE_SUCCESS
;
1230 * Same as remove_mapping, but if the page is removed from the mapping, it
1231 * gets returned with a refcount of 0.
1233 static int __remove_mapping(struct address_space
*mapping
, struct folio
*folio
,
1234 bool reclaimed
, struct mem_cgroup
*target_memcg
)
1237 void *shadow
= NULL
;
1239 BUG_ON(!folio_test_locked(folio
));
1240 BUG_ON(mapping
!= folio_mapping(folio
));
1242 if (!folio_test_swapcache(folio
))
1243 spin_lock(&mapping
->host
->i_lock
);
1244 xa_lock_irq(&mapping
->i_pages
);
1246 * The non racy check for a busy page.
1248 * Must be careful with the order of the tests. When someone has
1249 * a ref to the page, it may be possible that they dirty it then
1250 * drop the reference. So if PageDirty is tested before page_count
1251 * here, then the following race may occur:
1253 * get_user_pages(&page);
1254 * [user mapping goes away]
1256 * !PageDirty(page) [good]
1257 * SetPageDirty(page);
1259 * !page_count(page) [good, discard it]
1261 * [oops, our write_to data is lost]
1263 * Reversing the order of the tests ensures such a situation cannot
1264 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1265 * load is not satisfied before that of page->_refcount.
1267 * Note that if SetPageDirty is always performed via set_page_dirty,
1268 * and thus under the i_pages lock, then this ordering is not required.
1270 refcount
= 1 + folio_nr_pages(folio
);
1271 if (!folio_ref_freeze(folio
, refcount
))
1273 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1274 if (unlikely(folio_test_dirty(folio
))) {
1275 folio_ref_unfreeze(folio
, refcount
);
1279 if (folio_test_swapcache(folio
)) {
1280 swp_entry_t swap
= folio_swap_entry(folio
);
1281 mem_cgroup_swapout(folio
, swap
);
1282 if (reclaimed
&& !mapping_exiting(mapping
))
1283 shadow
= workingset_eviction(folio
, target_memcg
);
1284 __delete_from_swap_cache(&folio
->page
, swap
, shadow
);
1285 xa_unlock_irq(&mapping
->i_pages
);
1286 put_swap_page(&folio
->page
, swap
);
1288 void (*freepage
)(struct page
*);
1290 freepage
= mapping
->a_ops
->freepage
;
1292 * Remember a shadow entry for reclaimed file cache in
1293 * order to detect refaults, thus thrashing, later on.
1295 * But don't store shadows in an address space that is
1296 * already exiting. This is not just an optimization,
1297 * inode reclaim needs to empty out the radix tree or
1298 * the nodes are lost. Don't plant shadows behind its
1301 * We also don't store shadows for DAX mappings because the
1302 * only page cache pages found in these are zero pages
1303 * covering holes, and because we don't want to mix DAX
1304 * exceptional entries and shadow exceptional entries in the
1305 * same address_space.
1307 if (reclaimed
&& folio_is_file_lru(folio
) &&
1308 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
1309 shadow
= workingset_eviction(folio
, target_memcg
);
1310 __filemap_remove_folio(folio
, shadow
);
1311 xa_unlock_irq(&mapping
->i_pages
);
1312 if (mapping_shrinkable(mapping
))
1313 inode_add_lru(mapping
->host
);
1314 spin_unlock(&mapping
->host
->i_lock
);
1316 if (freepage
!= NULL
)
1317 freepage(&folio
->page
);
1323 xa_unlock_irq(&mapping
->i_pages
);
1324 if (!folio_test_swapcache(folio
))
1325 spin_unlock(&mapping
->host
->i_lock
);
1330 * remove_mapping() - Attempt to remove a folio from its mapping.
1331 * @mapping: The address space.
1332 * @folio: The folio to remove.
1334 * If the folio is dirty, under writeback or if someone else has a ref
1335 * on it, removal will fail.
1336 * Return: The number of pages removed from the mapping. 0 if the folio
1337 * could not be removed.
1338 * Context: The caller should have a single refcount on the folio and
1341 long remove_mapping(struct address_space
*mapping
, struct folio
*folio
)
1343 if (__remove_mapping(mapping
, folio
, false, NULL
)) {
1345 * Unfreezing the refcount with 1 effectively
1346 * drops the pagecache ref for us without requiring another
1349 folio_ref_unfreeze(folio
, 1);
1350 return folio_nr_pages(folio
);
1356 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1357 * @folio: Folio to be returned to an LRU list.
1359 * Add previously isolated @folio to appropriate LRU list.
1360 * The folio may still be unevictable for other reasons.
1362 * Context: lru_lock must not be held, interrupts must be enabled.
1364 void folio_putback_lru(struct folio
*folio
)
1366 folio_add_lru(folio
);
1367 folio_put(folio
); /* drop ref from isolate */
1370 enum page_references
{
1372 PAGEREF_RECLAIM_CLEAN
,
1377 static enum page_references
folio_check_references(struct folio
*folio
,
1378 struct scan_control
*sc
)
1380 int referenced_ptes
, referenced_folio
;
1381 unsigned long vm_flags
;
1383 referenced_ptes
= folio_referenced(folio
, 1, sc
->target_mem_cgroup
,
1385 referenced_folio
= folio_test_clear_referenced(folio
);
1388 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1389 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1391 if (vm_flags
& VM_LOCKED
)
1392 return PAGEREF_ACTIVATE
;
1394 if (referenced_ptes
) {
1396 * All mapped folios start out with page table
1397 * references from the instantiating fault, so we need
1398 * to look twice if a mapped file/anon folio is used more
1401 * Mark it and spare it for another trip around the
1402 * inactive list. Another page table reference will
1403 * lead to its activation.
1405 * Note: the mark is set for activated folios as well
1406 * so that recently deactivated but used folios are
1407 * quickly recovered.
1409 folio_set_referenced(folio
);
1411 if (referenced_folio
|| referenced_ptes
> 1)
1412 return PAGEREF_ACTIVATE
;
1415 * Activate file-backed executable folios after first usage.
1417 if ((vm_flags
& VM_EXEC
) && !folio_test_swapbacked(folio
))
1418 return PAGEREF_ACTIVATE
;
1420 return PAGEREF_KEEP
;
1423 /* Reclaim if clean, defer dirty folios to writeback */
1424 if (referenced_folio
&& !folio_test_swapbacked(folio
))
1425 return PAGEREF_RECLAIM_CLEAN
;
1427 return PAGEREF_RECLAIM
;
1430 /* Check if a page is dirty or under writeback */
1431 static void folio_check_dirty_writeback(struct folio
*folio
,
1432 bool *dirty
, bool *writeback
)
1434 struct address_space
*mapping
;
1437 * Anonymous pages are not handled by flushers and must be written
1438 * from reclaim context. Do not stall reclaim based on them
1440 if (!folio_is_file_lru(folio
) ||
1441 (folio_test_anon(folio
) && !folio_test_swapbacked(folio
))) {
1447 /* By default assume that the folio flags are accurate */
1448 *dirty
= folio_test_dirty(folio
);
1449 *writeback
= folio_test_writeback(folio
);
1451 /* Verify dirty/writeback state if the filesystem supports it */
1452 if (!folio_test_private(folio
))
1455 mapping
= folio_mapping(folio
);
1456 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1457 mapping
->a_ops
->is_dirty_writeback(&folio
->page
, dirty
, writeback
);
1460 static struct page
*alloc_demote_page(struct page
*page
, unsigned long node
)
1462 struct migration_target_control mtc
= {
1464 * Allocate from 'node', or fail quickly and quietly.
1465 * When this happens, 'page' will likely just be discarded
1466 * instead of migrated.
1468 .gfp_mask
= (GFP_HIGHUSER_MOVABLE
& ~__GFP_RECLAIM
) |
1469 __GFP_THISNODE
| __GFP_NOWARN
|
1470 __GFP_NOMEMALLOC
| GFP_NOWAIT
,
1474 return alloc_migration_target(page
, (unsigned long)&mtc
);
1478 * Take pages on @demote_list and attempt to demote them to
1479 * another node. Pages which are not demoted are left on
1482 static unsigned int demote_page_list(struct list_head
*demote_pages
,
1483 struct pglist_data
*pgdat
)
1485 int target_nid
= next_demotion_node(pgdat
->node_id
);
1486 unsigned int nr_succeeded
;
1488 if (list_empty(demote_pages
))
1491 if (target_nid
== NUMA_NO_NODE
)
1494 /* Demotion ignores all cpuset and mempolicy settings */
1495 migrate_pages(demote_pages
, alloc_demote_page
, NULL
,
1496 target_nid
, MIGRATE_ASYNC
, MR_DEMOTION
,
1499 if (current_is_kswapd())
1500 __count_vm_events(PGDEMOTE_KSWAPD
, nr_succeeded
);
1502 __count_vm_events(PGDEMOTE_DIRECT
, nr_succeeded
);
1504 return nr_succeeded
;
1507 static bool may_enter_fs(struct page
*page
, gfp_t gfp_mask
)
1509 if (gfp_mask
& __GFP_FS
)
1511 if (!PageSwapCache(page
) || !(gfp_mask
& __GFP_IO
))
1514 * We can "enter_fs" for swap-cache with only __GFP_IO
1515 * providing this isn't SWP_FS_OPS.
1516 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1517 * but that will never affect SWP_FS_OPS, so the data_race
1520 return !data_race(page_swap_flags(page
) & SWP_FS_OPS
);
1524 * shrink_page_list() returns the number of reclaimed pages
1526 static unsigned int shrink_page_list(struct list_head
*page_list
,
1527 struct pglist_data
*pgdat
,
1528 struct scan_control
*sc
,
1529 struct reclaim_stat
*stat
,
1530 bool ignore_references
)
1532 LIST_HEAD(ret_pages
);
1533 LIST_HEAD(free_pages
);
1534 LIST_HEAD(demote_pages
);
1535 unsigned int nr_reclaimed
= 0;
1536 unsigned int pgactivate
= 0;
1537 bool do_demote_pass
;
1538 struct swap_iocb
*plug
= NULL
;
1540 memset(stat
, 0, sizeof(*stat
));
1542 do_demote_pass
= can_demote(pgdat
->node_id
, sc
);
1545 while (!list_empty(page_list
)) {
1546 struct address_space
*mapping
;
1548 struct folio
*folio
;
1549 enum page_references references
= PAGEREF_RECLAIM
;
1550 bool dirty
, writeback
;
1551 unsigned int nr_pages
;
1555 folio
= lru_to_folio(page_list
);
1556 list_del(&folio
->lru
);
1557 page
= &folio
->page
;
1559 if (!trylock_page(page
))
1562 VM_BUG_ON_PAGE(PageActive(page
), page
);
1564 nr_pages
= compound_nr(page
);
1566 /* Account the number of base pages even though THP */
1567 sc
->nr_scanned
+= nr_pages
;
1569 if (unlikely(!page_evictable(page
)))
1570 goto activate_locked
;
1572 if (!sc
->may_unmap
&& page_mapped(page
))
1576 * The number of dirty pages determines if a node is marked
1577 * reclaim_congested. kswapd will stall and start writing
1578 * pages if the tail of the LRU is all dirty unqueued pages.
1580 folio_check_dirty_writeback(folio
, &dirty
, &writeback
);
1581 if (dirty
|| writeback
)
1582 stat
->nr_dirty
+= nr_pages
;
1584 if (dirty
&& !writeback
)
1585 stat
->nr_unqueued_dirty
+= nr_pages
;
1588 * Treat this page as congested if the underlying BDI is or if
1589 * pages are cycling through the LRU so quickly that the
1590 * pages marked for immediate reclaim are making it to the
1591 * end of the LRU a second time.
1593 mapping
= page_mapping(page
);
1594 if (writeback
&& PageReclaim(page
))
1595 stat
->nr_congested
+= nr_pages
;
1598 * If a page at the tail of the LRU is under writeback, there
1599 * are three cases to consider.
1601 * 1) If reclaim is encountering an excessive number of pages
1602 * under writeback and this page is both under writeback and
1603 * PageReclaim then it indicates that pages are being queued
1604 * for IO but are being recycled through the LRU before the
1605 * IO can complete. Waiting on the page itself risks an
1606 * indefinite stall if it is impossible to writeback the
1607 * page due to IO error or disconnected storage so instead
1608 * note that the LRU is being scanned too quickly and the
1609 * caller can stall after page list has been processed.
1611 * 2) Global or new memcg reclaim encounters a page that is
1612 * not marked for immediate reclaim, or the caller does not
1613 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1614 * not to fs). In this case mark the page for immediate
1615 * reclaim and continue scanning.
1617 * Require may_enter_fs() because we would wait on fs, which
1618 * may not have submitted IO yet. And the loop driver might
1619 * enter reclaim, and deadlock if it waits on a page for
1620 * which it is needed to do the write (loop masks off
1621 * __GFP_IO|__GFP_FS for this reason); but more thought
1622 * would probably show more reasons.
1624 * 3) Legacy memcg encounters a page that is already marked
1625 * PageReclaim. memcg does not have any dirty pages
1626 * throttling so we could easily OOM just because too many
1627 * pages are in writeback and there is nothing else to
1628 * reclaim. Wait for the writeback to complete.
1630 * In cases 1) and 2) we activate the pages to get them out of
1631 * the way while we continue scanning for clean pages on the
1632 * inactive list and refilling from the active list. The
1633 * observation here is that waiting for disk writes is more
1634 * expensive than potentially causing reloads down the line.
1635 * Since they're marked for immediate reclaim, they won't put
1636 * memory pressure on the cache working set any longer than it
1637 * takes to write them to disk.
1639 if (PageWriteback(page
)) {
1641 if (current_is_kswapd() &&
1642 PageReclaim(page
) &&
1643 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1644 stat
->nr_immediate
+= nr_pages
;
1645 goto activate_locked
;
1648 } else if (writeback_throttling_sane(sc
) ||
1649 !PageReclaim(page
) || !may_enter_fs(page
, sc
->gfp_mask
)) {
1651 * This is slightly racy - end_page_writeback()
1652 * might have just cleared PageReclaim, then
1653 * setting PageReclaim here end up interpreted
1654 * as PageReadahead - but that does not matter
1655 * enough to care. What we do want is for this
1656 * page to have PageReclaim set next time memcg
1657 * reclaim reaches the tests above, so it will
1658 * then wait_on_page_writeback() to avoid OOM;
1659 * and it's also appropriate in global reclaim.
1661 SetPageReclaim(page
);
1662 stat
->nr_writeback
+= nr_pages
;
1663 goto activate_locked
;
1668 wait_on_page_writeback(page
);
1669 /* then go back and try same page again */
1670 list_add_tail(&page
->lru
, page_list
);
1675 if (!ignore_references
)
1676 references
= folio_check_references(folio
, sc
);
1678 switch (references
) {
1679 case PAGEREF_ACTIVATE
:
1680 goto activate_locked
;
1682 stat
->nr_ref_keep
+= nr_pages
;
1684 case PAGEREF_RECLAIM
:
1685 case PAGEREF_RECLAIM_CLEAN
:
1686 ; /* try to reclaim the page below */
1690 * Before reclaiming the page, try to relocate
1691 * its contents to another node.
1693 if (do_demote_pass
&&
1694 (thp_migration_supported() || !PageTransHuge(page
))) {
1695 list_add(&page
->lru
, &demote_pages
);
1701 * Anonymous process memory has backing store?
1702 * Try to allocate it some swap space here.
1703 * Lazyfree page could be freed directly
1705 if (PageAnon(page
) && PageSwapBacked(page
)) {
1706 if (!PageSwapCache(page
)) {
1707 if (!(sc
->gfp_mask
& __GFP_IO
))
1709 if (folio_maybe_dma_pinned(folio
))
1711 if (PageTransHuge(page
)) {
1712 /* cannot split THP, skip it */
1713 if (!can_split_folio(folio
, NULL
))
1714 goto activate_locked
;
1716 * Split pages without a PMD map right
1717 * away. Chances are some or all of the
1718 * tail pages can be freed without IO.
1720 if (!folio_entire_mapcount(folio
) &&
1721 split_folio_to_list(folio
,
1723 goto activate_locked
;
1725 if (!add_to_swap(page
)) {
1726 if (!PageTransHuge(page
))
1727 goto activate_locked_split
;
1728 /* Fallback to swap normal pages */
1729 if (split_folio_to_list(folio
,
1731 goto activate_locked
;
1732 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1733 count_vm_event(THP_SWPOUT_FALLBACK
);
1735 if (!add_to_swap(page
))
1736 goto activate_locked_split
;
1739 /* Adding to swap updated mapping */
1740 mapping
= page_mapping(page
);
1742 } else if (PageSwapBacked(page
) && PageTransHuge(page
)) {
1743 /* Split shmem THP */
1744 if (split_folio_to_list(folio
, page_list
))
1749 * THP may get split above, need minus tail pages and update
1750 * nr_pages to avoid accounting tail pages twice.
1752 * The tail pages that are added into swap cache successfully
1755 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1756 sc
->nr_scanned
-= (nr_pages
- 1);
1761 * The page is mapped into the page tables of one or more
1762 * processes. Try to unmap it here.
1764 if (page_mapped(page
)) {
1765 enum ttu_flags flags
= TTU_BATCH_FLUSH
;
1766 bool was_swapbacked
= PageSwapBacked(page
);
1768 if (PageTransHuge(page
) &&
1769 thp_order(page
) >= HPAGE_PMD_ORDER
)
1770 flags
|= TTU_SPLIT_HUGE_PMD
;
1772 try_to_unmap(folio
, flags
);
1773 if (page_mapped(page
)) {
1774 stat
->nr_unmap_fail
+= nr_pages
;
1775 if (!was_swapbacked
&& PageSwapBacked(page
))
1776 stat
->nr_lazyfree_fail
+= nr_pages
;
1777 goto activate_locked
;
1781 if (PageDirty(page
)) {
1783 * Only kswapd can writeback filesystem pages
1784 * to avoid risk of stack overflow. But avoid
1785 * injecting inefficient single-page IO into
1786 * flusher writeback as much as possible: only
1787 * write pages when we've encountered many
1788 * dirty pages, and when we've already scanned
1789 * the rest of the LRU for clean pages and see
1790 * the same dirty pages again (PageReclaim).
1792 if (page_is_file_lru(page
) &&
1793 (!current_is_kswapd() || !PageReclaim(page
) ||
1794 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1796 * Immediately reclaim when written back.
1797 * Similar in principal to deactivate_page()
1798 * except we already have the page isolated
1799 * and know it's dirty
1801 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1802 SetPageReclaim(page
);
1804 goto activate_locked
;
1807 if (references
== PAGEREF_RECLAIM_CLEAN
)
1809 if (!may_enter_fs(page
, sc
->gfp_mask
))
1811 if (!sc
->may_writepage
)
1815 * Page is dirty. Flush the TLB if a writable entry
1816 * potentially exists to avoid CPU writes after IO
1817 * starts and then write it out here.
1819 try_to_unmap_flush_dirty();
1820 switch (pageout(folio
, mapping
, &plug
)) {
1824 goto activate_locked
;
1826 stat
->nr_pageout
+= nr_pages
;
1828 if (PageWriteback(page
))
1830 if (PageDirty(page
))
1834 * A synchronous write - probably a ramdisk. Go
1835 * ahead and try to reclaim the page.
1837 if (!trylock_page(page
))
1839 if (PageDirty(page
) || PageWriteback(page
))
1841 mapping
= page_mapping(page
);
1844 ; /* try to free the page below */
1849 * If the page has buffers, try to free the buffer mappings
1850 * associated with this page. If we succeed we try to free
1853 * We do this even if the page is PageDirty().
1854 * try_to_release_page() does not perform I/O, but it is
1855 * possible for a page to have PageDirty set, but it is actually
1856 * clean (all its buffers are clean). This happens if the
1857 * buffers were written out directly, with submit_bh(). ext3
1858 * will do this, as well as the blockdev mapping.
1859 * try_to_release_page() will discover that cleanness and will
1860 * drop the buffers and mark the page clean - it can be freed.
1862 * Rarely, pages can have buffers and no ->mapping. These are
1863 * the pages which were not successfully invalidated in
1864 * truncate_cleanup_page(). We try to drop those buffers here
1865 * and if that worked, and the page is no longer mapped into
1866 * process address space (page_count == 1) it can be freed.
1867 * Otherwise, leave the page on the LRU so it is swappable.
1869 if (page_has_private(page
)) {
1870 if (!try_to_release_page(page
, sc
->gfp_mask
))
1871 goto activate_locked
;
1872 if (!mapping
&& page_count(page
) == 1) {
1874 if (put_page_testzero(page
))
1878 * rare race with speculative reference.
1879 * the speculative reference will free
1880 * this page shortly, so we may
1881 * increment nr_reclaimed here (and
1882 * leave it off the LRU).
1890 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1891 /* follow __remove_mapping for reference */
1892 if (!page_ref_freeze(page
, 1))
1895 * The page has only one reference left, which is
1896 * from the isolation. After the caller puts the
1897 * page back on lru and drops the reference, the
1898 * page will be freed anyway. It doesn't matter
1899 * which lru it goes. So we don't bother checking
1902 count_vm_event(PGLAZYFREED
);
1903 count_memcg_page_event(page
, PGLAZYFREED
);
1904 } else if (!mapping
|| !__remove_mapping(mapping
, folio
, true,
1905 sc
->target_mem_cgroup
))
1911 * THP may get swapped out in a whole, need account
1914 nr_reclaimed
+= nr_pages
;
1917 * Is there need to periodically free_page_list? It would
1918 * appear not as the counts should be low
1920 if (unlikely(PageTransHuge(page
)))
1921 destroy_compound_page(page
);
1923 list_add(&page
->lru
, &free_pages
);
1926 activate_locked_split
:
1928 * The tail pages that are failed to add into swap cache
1929 * reach here. Fixup nr_scanned and nr_pages.
1932 sc
->nr_scanned
-= (nr_pages
- 1);
1936 /* Not a candidate for swapping, so reclaim swap space. */
1937 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1939 try_to_free_swap(page
);
1940 VM_BUG_ON_PAGE(PageActive(page
), page
);
1941 if (!PageMlocked(page
)) {
1942 int type
= page_is_file_lru(page
);
1943 SetPageActive(page
);
1944 stat
->nr_activate
[type
] += nr_pages
;
1945 count_memcg_page_event(page
, PGACTIVATE
);
1950 list_add(&page
->lru
, &ret_pages
);
1951 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1953 /* 'page_list' is always empty here */
1955 /* Migrate pages selected for demotion */
1956 nr_reclaimed
+= demote_page_list(&demote_pages
, pgdat
);
1957 /* Pages that could not be demoted are still in @demote_pages */
1958 if (!list_empty(&demote_pages
)) {
1959 /* Pages which failed to demoted go back on @page_list for retry: */
1960 list_splice_init(&demote_pages
, page_list
);
1961 do_demote_pass
= false;
1965 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1967 mem_cgroup_uncharge_list(&free_pages
);
1968 try_to_unmap_flush();
1969 free_unref_page_list(&free_pages
);
1971 list_splice(&ret_pages
, page_list
);
1972 count_vm_events(PGACTIVATE
, pgactivate
);
1975 swap_write_unplug(plug
);
1976 return nr_reclaimed
;
1979 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1980 struct list_head
*page_list
)
1982 struct scan_control sc
= {
1983 .gfp_mask
= GFP_KERNEL
,
1986 struct reclaim_stat stat
;
1987 unsigned int nr_reclaimed
;
1988 struct page
*page
, *next
;
1989 LIST_HEAD(clean_pages
);
1990 unsigned int noreclaim_flag
;
1992 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1993 if (!PageHuge(page
) && page_is_file_lru(page
) &&
1994 !PageDirty(page
) && !__PageMovable(page
) &&
1995 !PageUnevictable(page
)) {
1996 ClearPageActive(page
);
1997 list_move(&page
->lru
, &clean_pages
);
2002 * We should be safe here since we are only dealing with file pages and
2003 * we are not kswapd and therefore cannot write dirty file pages. But
2004 * call memalloc_noreclaim_save() anyway, just in case these conditions
2005 * change in the future.
2007 noreclaim_flag
= memalloc_noreclaim_save();
2008 nr_reclaimed
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
2010 memalloc_noreclaim_restore(noreclaim_flag
);
2012 list_splice(&clean_pages
, page_list
);
2013 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
2014 -(long)nr_reclaimed
);
2016 * Since lazyfree pages are isolated from file LRU from the beginning,
2017 * they will rotate back to anonymous LRU in the end if it failed to
2018 * discard so isolated count will be mismatched.
2019 * Compensate the isolated count for both LRU lists.
2021 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
2022 stat
.nr_lazyfree_fail
);
2023 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
2024 -(long)stat
.nr_lazyfree_fail
);
2025 return nr_reclaimed
;
2029 * Update LRU sizes after isolating pages. The LRU size updates must
2030 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2032 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
2033 enum lru_list lru
, unsigned long *nr_zone_taken
)
2037 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2038 if (!nr_zone_taken
[zid
])
2041 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
2047 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2049 * lruvec->lru_lock is heavily contended. Some of the functions that
2050 * shrink the lists perform better by taking out a batch of pages
2051 * and working on them outside the LRU lock.
2053 * For pagecache intensive workloads, this function is the hottest
2054 * spot in the kernel (apart from copy_*_user functions).
2056 * Lru_lock must be held before calling this function.
2058 * @nr_to_scan: The number of eligible pages to look through on the list.
2059 * @lruvec: The LRU vector to pull pages from.
2060 * @dst: The temp list to put pages on to.
2061 * @nr_scanned: The number of pages that were scanned.
2062 * @sc: The scan_control struct for this reclaim session
2063 * @lru: LRU list id for isolating
2065 * returns how many pages were moved onto *@dst.
2067 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
2068 struct lruvec
*lruvec
, struct list_head
*dst
,
2069 unsigned long *nr_scanned
, struct scan_control
*sc
,
2072 struct list_head
*src
= &lruvec
->lists
[lru
];
2073 unsigned long nr_taken
= 0;
2074 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
2075 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
2076 unsigned long skipped
= 0;
2077 unsigned long scan
, total_scan
, nr_pages
;
2078 LIST_HEAD(pages_skipped
);
2082 while (scan
< nr_to_scan
&& !list_empty(src
)) {
2083 struct list_head
*move_to
= src
;
2086 page
= lru_to_page(src
);
2087 prefetchw_prev_lru_page(page
, src
, flags
);
2089 nr_pages
= compound_nr(page
);
2090 total_scan
+= nr_pages
;
2092 if (page_zonenum(page
) > sc
->reclaim_idx
) {
2093 nr_skipped
[page_zonenum(page
)] += nr_pages
;
2094 move_to
= &pages_skipped
;
2099 * Do not count skipped pages because that makes the function
2100 * return with no isolated pages if the LRU mostly contains
2101 * ineligible pages. This causes the VM to not reclaim any
2102 * pages, triggering a premature OOM.
2103 * Account all tail pages of THP.
2109 if (!sc
->may_unmap
&& page_mapped(page
))
2113 * Be careful not to clear PageLRU until after we're
2114 * sure the page is not being freed elsewhere -- the
2115 * page release code relies on it.
2117 if (unlikely(!get_page_unless_zero(page
)))
2120 if (!TestClearPageLRU(page
)) {
2121 /* Another thread is already isolating this page */
2126 nr_taken
+= nr_pages
;
2127 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
2130 list_move(&page
->lru
, move_to
);
2134 * Splice any skipped pages to the start of the LRU list. Note that
2135 * this disrupts the LRU order when reclaiming for lower zones but
2136 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2137 * scanning would soon rescan the same pages to skip and waste lots
2140 if (!list_empty(&pages_skipped
)) {
2143 list_splice(&pages_skipped
, src
);
2144 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2145 if (!nr_skipped
[zid
])
2148 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
2149 skipped
+= nr_skipped
[zid
];
2152 *nr_scanned
= total_scan
;
2153 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
2154 total_scan
, skipped
, nr_taken
,
2155 sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
, lru
);
2156 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
2161 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2162 * @folio: Folio to isolate from its LRU list.
2164 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2165 * corresponding to whatever LRU list the folio was on.
2167 * The folio will have its LRU flag cleared. If it was found on the
2168 * active list, it will have the Active flag set. If it was found on the
2169 * unevictable list, it will have the Unevictable flag set. These flags
2170 * may need to be cleared by the caller before letting the page go.
2174 * (1) Must be called with an elevated refcount on the page. This is a
2175 * fundamental difference from isolate_lru_pages() (which is called
2176 * without a stable reference).
2177 * (2) The lru_lock must not be held.
2178 * (3) Interrupts must be enabled.
2180 * Return: 0 if the folio was removed from an LRU list.
2181 * -EBUSY if the folio was not on an LRU list.
2183 int folio_isolate_lru(struct folio
*folio
)
2187 VM_BUG_ON_FOLIO(!folio_ref_count(folio
), folio
);
2189 if (folio_test_clear_lru(folio
)) {
2190 struct lruvec
*lruvec
;
2193 lruvec
= folio_lruvec_lock_irq(folio
);
2194 lruvec_del_folio(lruvec
, folio
);
2195 unlock_page_lruvec_irq(lruvec
);
2203 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2204 * then get rescheduled. When there are massive number of tasks doing page
2205 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2206 * the LRU list will go small and be scanned faster than necessary, leading to
2207 * unnecessary swapping, thrashing and OOM.
2209 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
2210 struct scan_control
*sc
)
2212 unsigned long inactive
, isolated
;
2215 if (current_is_kswapd())
2218 if (!writeback_throttling_sane(sc
))
2222 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2223 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
2225 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2226 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
2230 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2231 * won't get blocked by normal direct-reclaimers, forming a circular
2234 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
2237 too_many
= isolated
> inactive
;
2239 /* Wake up tasks throttled due to too_many_isolated. */
2241 wake_throttle_isolated(pgdat
);
2247 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2248 * On return, @list is reused as a list of pages to be freed by the caller.
2250 * Returns the number of pages moved to the given lruvec.
2252 static unsigned int move_pages_to_lru(struct lruvec
*lruvec
,
2253 struct list_head
*list
)
2255 int nr_pages
, nr_moved
= 0;
2256 LIST_HEAD(pages_to_free
);
2259 while (!list_empty(list
)) {
2260 page
= lru_to_page(list
);
2261 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2262 list_del(&page
->lru
);
2263 if (unlikely(!page_evictable(page
))) {
2264 spin_unlock_irq(&lruvec
->lru_lock
);
2265 putback_lru_page(page
);
2266 spin_lock_irq(&lruvec
->lru_lock
);
2271 * The SetPageLRU needs to be kept here for list integrity.
2273 * #0 move_pages_to_lru #1 release_pages
2274 * if !put_page_testzero
2275 * if (put_page_testzero())
2276 * !PageLRU //skip lru_lock
2278 * list_add(&page->lru,)
2279 * list_add(&page->lru,)
2283 if (unlikely(put_page_testzero(page
))) {
2284 __clear_page_lru_flags(page
);
2286 if (unlikely(PageCompound(page
))) {
2287 spin_unlock_irq(&lruvec
->lru_lock
);
2288 destroy_compound_page(page
);
2289 spin_lock_irq(&lruvec
->lru_lock
);
2291 list_add(&page
->lru
, &pages_to_free
);
2297 * All pages were isolated from the same lruvec (and isolation
2298 * inhibits memcg migration).
2300 VM_BUG_ON_PAGE(!folio_matches_lruvec(page_folio(page
), lruvec
), page
);
2301 add_page_to_lru_list(page
, lruvec
);
2302 nr_pages
= thp_nr_pages(page
);
2303 nr_moved
+= nr_pages
;
2304 if (PageActive(page
))
2305 workingset_age_nonresident(lruvec
, nr_pages
);
2309 * To save our caller's stack, now use input list for pages to free.
2311 list_splice(&pages_to_free
, list
);
2317 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2318 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2319 * we should not throttle. Otherwise it is safe to do so.
2321 static int current_may_throttle(void)
2323 return !(current
->flags
& PF_LOCAL_THROTTLE
);
2327 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2328 * of reclaimed pages
2330 static unsigned long
2331 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
2332 struct scan_control
*sc
, enum lru_list lru
)
2334 LIST_HEAD(page_list
);
2335 unsigned long nr_scanned
;
2336 unsigned int nr_reclaimed
= 0;
2337 unsigned long nr_taken
;
2338 struct reclaim_stat stat
;
2339 bool file
= is_file_lru(lru
);
2340 enum vm_event_item item
;
2341 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2342 bool stalled
= false;
2344 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
2348 /* wait a bit for the reclaimer. */
2350 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_ISOLATED
);
2352 /* We are about to die and free our memory. Return now. */
2353 if (fatal_signal_pending(current
))
2354 return SWAP_CLUSTER_MAX
;
2359 spin_lock_irq(&lruvec
->lru_lock
);
2361 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
2362 &nr_scanned
, sc
, lru
);
2364 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2365 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
2366 if (!cgroup_reclaim(sc
))
2367 __count_vm_events(item
, nr_scanned
);
2368 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
2369 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
2371 spin_unlock_irq(&lruvec
->lru_lock
);
2376 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, &stat
, false);
2378 spin_lock_irq(&lruvec
->lru_lock
);
2379 move_pages_to_lru(lruvec
, &page_list
);
2381 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2382 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2383 if (!cgroup_reclaim(sc
))
2384 __count_vm_events(item
, nr_reclaimed
);
2385 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2386 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
2387 spin_unlock_irq(&lruvec
->lru_lock
);
2389 lru_note_cost(lruvec
, file
, stat
.nr_pageout
);
2390 mem_cgroup_uncharge_list(&page_list
);
2391 free_unref_page_list(&page_list
);
2394 * If dirty pages are scanned that are not queued for IO, it
2395 * implies that flushers are not doing their job. This can
2396 * happen when memory pressure pushes dirty pages to the end of
2397 * the LRU before the dirty limits are breached and the dirty
2398 * data has expired. It can also happen when the proportion of
2399 * dirty pages grows not through writes but through memory
2400 * pressure reclaiming all the clean cache. And in some cases,
2401 * the flushers simply cannot keep up with the allocation
2402 * rate. Nudge the flusher threads in case they are asleep.
2404 if (stat
.nr_unqueued_dirty
== nr_taken
)
2405 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2407 sc
->nr
.dirty
+= stat
.nr_dirty
;
2408 sc
->nr
.congested
+= stat
.nr_congested
;
2409 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2410 sc
->nr
.writeback
+= stat
.nr_writeback
;
2411 sc
->nr
.immediate
+= stat
.nr_immediate
;
2412 sc
->nr
.taken
+= nr_taken
;
2414 sc
->nr
.file_taken
+= nr_taken
;
2416 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2417 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2418 return nr_reclaimed
;
2422 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2424 * We move them the other way if the page is referenced by one or more
2427 * If the pages are mostly unmapped, the processing is fast and it is
2428 * appropriate to hold lru_lock across the whole operation. But if
2429 * the pages are mapped, the processing is slow (folio_referenced()), so
2430 * we should drop lru_lock around each page. It's impossible to balance
2431 * this, so instead we remove the pages from the LRU while processing them.
2432 * It is safe to rely on PG_active against the non-LRU pages in here because
2433 * nobody will play with that bit on a non-LRU page.
2435 * The downside is that we have to touch page->_refcount against each page.
2436 * But we had to alter page->flags anyway.
2438 static void shrink_active_list(unsigned long nr_to_scan
,
2439 struct lruvec
*lruvec
,
2440 struct scan_control
*sc
,
2443 unsigned long nr_taken
;
2444 unsigned long nr_scanned
;
2445 unsigned long vm_flags
;
2446 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2447 LIST_HEAD(l_active
);
2448 LIST_HEAD(l_inactive
);
2449 unsigned nr_deactivate
, nr_activate
;
2450 unsigned nr_rotated
= 0;
2451 int file
= is_file_lru(lru
);
2452 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2456 spin_lock_irq(&lruvec
->lru_lock
);
2458 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2459 &nr_scanned
, sc
, lru
);
2461 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2463 if (!cgroup_reclaim(sc
))
2464 __count_vm_events(PGREFILL
, nr_scanned
);
2465 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2467 spin_unlock_irq(&lruvec
->lru_lock
);
2469 while (!list_empty(&l_hold
)) {
2470 struct folio
*folio
;
2474 folio
= lru_to_folio(&l_hold
);
2475 list_del(&folio
->lru
);
2476 page
= &folio
->page
;
2478 if (unlikely(!page_evictable(page
))) {
2479 putback_lru_page(page
);
2483 if (unlikely(buffer_heads_over_limit
)) {
2484 if (page_has_private(page
) && trylock_page(page
)) {
2485 if (page_has_private(page
))
2486 try_to_release_page(page
, 0);
2491 if (folio_referenced(folio
, 0, sc
->target_mem_cgroup
,
2494 * Identify referenced, file-backed active pages and
2495 * give them one more trip around the active list. So
2496 * that executable code get better chances to stay in
2497 * memory under moderate memory pressure. Anon pages
2498 * are not likely to be evicted by use-once streaming
2499 * IO, plus JVM can create lots of anon VM_EXEC pages,
2500 * so we ignore them here.
2502 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2503 nr_rotated
+= thp_nr_pages(page
);
2504 list_add(&page
->lru
, &l_active
);
2509 ClearPageActive(page
); /* we are de-activating */
2510 SetPageWorkingset(page
);
2511 list_add(&page
->lru
, &l_inactive
);
2515 * Move pages back to the lru list.
2517 spin_lock_irq(&lruvec
->lru_lock
);
2519 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2520 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2521 /* Keep all free pages in l_active list */
2522 list_splice(&l_inactive
, &l_active
);
2524 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2525 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2527 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2528 spin_unlock_irq(&lruvec
->lru_lock
);
2530 mem_cgroup_uncharge_list(&l_active
);
2531 free_unref_page_list(&l_active
);
2532 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2533 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2536 unsigned long reclaim_pages(struct list_head
*page_list
)
2538 int nid
= NUMA_NO_NODE
;
2539 unsigned int nr_reclaimed
= 0;
2540 LIST_HEAD(node_page_list
);
2541 struct reclaim_stat dummy_stat
;
2543 unsigned int noreclaim_flag
;
2544 struct scan_control sc
= {
2545 .gfp_mask
= GFP_KERNEL
,
2552 noreclaim_flag
= memalloc_noreclaim_save();
2554 while (!list_empty(page_list
)) {
2555 page
= lru_to_page(page_list
);
2556 if (nid
== NUMA_NO_NODE
) {
2557 nid
= page_to_nid(page
);
2558 INIT_LIST_HEAD(&node_page_list
);
2561 if (nid
== page_to_nid(page
)) {
2562 ClearPageActive(page
);
2563 list_move(&page
->lru
, &node_page_list
);
2567 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2569 &sc
, &dummy_stat
, false);
2570 while (!list_empty(&node_page_list
)) {
2571 page
= lru_to_page(&node_page_list
);
2572 list_del(&page
->lru
);
2573 putback_lru_page(page
);
2579 if (!list_empty(&node_page_list
)) {
2580 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2582 &sc
, &dummy_stat
, false);
2583 while (!list_empty(&node_page_list
)) {
2584 page
= lru_to_page(&node_page_list
);
2585 list_del(&page
->lru
);
2586 putback_lru_page(page
);
2590 memalloc_noreclaim_restore(noreclaim_flag
);
2592 return nr_reclaimed
;
2595 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2596 struct lruvec
*lruvec
, struct scan_control
*sc
)
2598 if (is_active_lru(lru
)) {
2599 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2600 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2602 sc
->skipped_deactivate
= 1;
2606 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2610 * The inactive anon list should be small enough that the VM never has
2611 * to do too much work.
2613 * The inactive file list should be small enough to leave most memory
2614 * to the established workingset on the scan-resistant active list,
2615 * but large enough to avoid thrashing the aggregate readahead window.
2617 * Both inactive lists should also be large enough that each inactive
2618 * page has a chance to be referenced again before it is reclaimed.
2620 * If that fails and refaulting is observed, the inactive list grows.
2622 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2623 * on this LRU, maintained by the pageout code. An inactive_ratio
2624 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2627 * memory ratio inactive
2628 * -------------------------------------
2637 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2639 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2640 unsigned long inactive
, active
;
2641 unsigned long inactive_ratio
;
2644 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2645 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2647 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2649 inactive_ratio
= int_sqrt(10 * gb
);
2653 return inactive
* inactive_ratio
< active
;
2664 * Determine how aggressively the anon and file LRU lists should be
2667 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2668 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2670 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2673 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2674 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2675 unsigned long anon_cost
, file_cost
, total_cost
;
2676 int swappiness
= mem_cgroup_swappiness(memcg
);
2677 u64 fraction
[ANON_AND_FILE
];
2678 u64 denominator
= 0; /* gcc */
2679 enum scan_balance scan_balance
;
2680 unsigned long ap
, fp
;
2683 /* If we have no swap space, do not bother scanning anon pages. */
2684 if (!sc
->may_swap
|| !can_reclaim_anon_pages(memcg
, pgdat
->node_id
, sc
)) {
2685 scan_balance
= SCAN_FILE
;
2690 * Global reclaim will swap to prevent OOM even with no
2691 * swappiness, but memcg users want to use this knob to
2692 * disable swapping for individual groups completely when
2693 * using the memory controller's swap limit feature would be
2696 if (cgroup_reclaim(sc
) && !swappiness
) {
2697 scan_balance
= SCAN_FILE
;
2702 * Do not apply any pressure balancing cleverness when the
2703 * system is close to OOM, scan both anon and file equally
2704 * (unless the swappiness setting disagrees with swapping).
2706 if (!sc
->priority
&& swappiness
) {
2707 scan_balance
= SCAN_EQUAL
;
2712 * If the system is almost out of file pages, force-scan anon.
2714 if (sc
->file_is_tiny
) {
2715 scan_balance
= SCAN_ANON
;
2720 * If there is enough inactive page cache, we do not reclaim
2721 * anything from the anonymous working right now.
2723 if (sc
->cache_trim_mode
) {
2724 scan_balance
= SCAN_FILE
;
2728 scan_balance
= SCAN_FRACT
;
2730 * Calculate the pressure balance between anon and file pages.
2732 * The amount of pressure we put on each LRU is inversely
2733 * proportional to the cost of reclaiming each list, as
2734 * determined by the share of pages that are refaulting, times
2735 * the relative IO cost of bringing back a swapped out
2736 * anonymous page vs reloading a filesystem page (swappiness).
2738 * Although we limit that influence to ensure no list gets
2739 * left behind completely: at least a third of the pressure is
2740 * applied, before swappiness.
2742 * With swappiness at 100, anon and file have equal IO cost.
2744 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2745 anon_cost
= total_cost
+ sc
->anon_cost
;
2746 file_cost
= total_cost
+ sc
->file_cost
;
2747 total_cost
= anon_cost
+ file_cost
;
2749 ap
= swappiness
* (total_cost
+ 1);
2750 ap
/= anon_cost
+ 1;
2752 fp
= (200 - swappiness
) * (total_cost
+ 1);
2753 fp
/= file_cost
+ 1;
2757 denominator
= ap
+ fp
;
2759 for_each_evictable_lru(lru
) {
2760 int file
= is_file_lru(lru
);
2761 unsigned long lruvec_size
;
2762 unsigned long low
, min
;
2765 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2766 mem_cgroup_protection(sc
->target_mem_cgroup
, memcg
,
2771 * Scale a cgroup's reclaim pressure by proportioning
2772 * its current usage to its memory.low or memory.min
2775 * This is important, as otherwise scanning aggression
2776 * becomes extremely binary -- from nothing as we
2777 * approach the memory protection threshold, to totally
2778 * nominal as we exceed it. This results in requiring
2779 * setting extremely liberal protection thresholds. It
2780 * also means we simply get no protection at all if we
2781 * set it too low, which is not ideal.
2783 * If there is any protection in place, we reduce scan
2784 * pressure by how much of the total memory used is
2785 * within protection thresholds.
2787 * There is one special case: in the first reclaim pass,
2788 * we skip over all groups that are within their low
2789 * protection. If that fails to reclaim enough pages to
2790 * satisfy the reclaim goal, we come back and override
2791 * the best-effort low protection. However, we still
2792 * ideally want to honor how well-behaved groups are in
2793 * that case instead of simply punishing them all
2794 * equally. As such, we reclaim them based on how much
2795 * memory they are using, reducing the scan pressure
2796 * again by how much of the total memory used is under
2799 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2800 unsigned long protection
;
2802 /* memory.low scaling, make sure we retry before OOM */
2803 if (!sc
->memcg_low_reclaim
&& low
> min
) {
2805 sc
->memcg_low_skipped
= 1;
2810 /* Avoid TOCTOU with earlier protection check */
2811 cgroup_size
= max(cgroup_size
, protection
);
2813 scan
= lruvec_size
- lruvec_size
* protection
/
2817 * Minimally target SWAP_CLUSTER_MAX pages to keep
2818 * reclaim moving forwards, avoiding decrementing
2819 * sc->priority further than desirable.
2821 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2826 scan
>>= sc
->priority
;
2829 * If the cgroup's already been deleted, make sure to
2830 * scrape out the remaining cache.
2832 if (!scan
&& !mem_cgroup_online(memcg
))
2833 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2835 switch (scan_balance
) {
2837 /* Scan lists relative to size */
2841 * Scan types proportional to swappiness and
2842 * their relative recent reclaim efficiency.
2843 * Make sure we don't miss the last page on
2844 * the offlined memory cgroups because of a
2847 scan
= mem_cgroup_online(memcg
) ?
2848 div64_u64(scan
* fraction
[file
], denominator
) :
2849 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2854 /* Scan one type exclusively */
2855 if ((scan_balance
== SCAN_FILE
) != file
)
2859 /* Look ma, no brain */
2868 * Anonymous LRU management is a waste if there is
2869 * ultimately no way to reclaim the memory.
2871 static bool can_age_anon_pages(struct pglist_data
*pgdat
,
2872 struct scan_control
*sc
)
2874 /* Aging the anon LRU is valuable if swap is present: */
2875 if (total_swap_pages
> 0)
2878 /* Also valuable if anon pages can be demoted: */
2879 return can_demote(pgdat
->node_id
, sc
);
2882 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2884 unsigned long nr
[NR_LRU_LISTS
];
2885 unsigned long targets
[NR_LRU_LISTS
];
2886 unsigned long nr_to_scan
;
2888 unsigned long nr_reclaimed
= 0;
2889 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2890 struct blk_plug plug
;
2893 get_scan_count(lruvec
, sc
, nr
);
2895 /* Record the original scan target for proportional adjustments later */
2896 memcpy(targets
, nr
, sizeof(nr
));
2899 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2900 * event that can occur when there is little memory pressure e.g.
2901 * multiple streaming readers/writers. Hence, we do not abort scanning
2902 * when the requested number of pages are reclaimed when scanning at
2903 * DEF_PRIORITY on the assumption that the fact we are direct
2904 * reclaiming implies that kswapd is not keeping up and it is best to
2905 * do a batch of work at once. For memcg reclaim one check is made to
2906 * abort proportional reclaim if either the file or anon lru has already
2907 * dropped to zero at the first pass.
2909 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2910 sc
->priority
== DEF_PRIORITY
);
2912 blk_start_plug(&plug
);
2913 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2914 nr
[LRU_INACTIVE_FILE
]) {
2915 unsigned long nr_anon
, nr_file
, percentage
;
2916 unsigned long nr_scanned
;
2918 for_each_evictable_lru(lru
) {
2920 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2921 nr
[lru
] -= nr_to_scan
;
2923 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2930 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2934 * For kswapd and memcg, reclaim at least the number of pages
2935 * requested. Ensure that the anon and file LRUs are scanned
2936 * proportionally what was requested by get_scan_count(). We
2937 * stop reclaiming one LRU and reduce the amount scanning
2938 * proportional to the original scan target.
2940 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2941 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2944 * It's just vindictive to attack the larger once the smaller
2945 * has gone to zero. And given the way we stop scanning the
2946 * smaller below, this makes sure that we only make one nudge
2947 * towards proportionality once we've got nr_to_reclaim.
2949 if (!nr_file
|| !nr_anon
)
2952 if (nr_file
> nr_anon
) {
2953 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2954 targets
[LRU_ACTIVE_ANON
] + 1;
2956 percentage
= nr_anon
* 100 / scan_target
;
2958 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2959 targets
[LRU_ACTIVE_FILE
] + 1;
2961 percentage
= nr_file
* 100 / scan_target
;
2964 /* Stop scanning the smaller of the LRU */
2966 nr
[lru
+ LRU_ACTIVE
] = 0;
2969 * Recalculate the other LRU scan count based on its original
2970 * scan target and the percentage scanning already complete
2972 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2973 nr_scanned
= targets
[lru
] - nr
[lru
];
2974 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2975 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2978 nr_scanned
= targets
[lru
] - nr
[lru
];
2979 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2980 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2982 scan_adjusted
= true;
2984 blk_finish_plug(&plug
);
2985 sc
->nr_reclaimed
+= nr_reclaimed
;
2988 * Even if we did not try to evict anon pages at all, we want to
2989 * rebalance the anon lru active/inactive ratio.
2991 if (can_age_anon_pages(lruvec_pgdat(lruvec
), sc
) &&
2992 inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
2993 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2994 sc
, LRU_ACTIVE_ANON
);
2997 /* Use reclaim/compaction for costly allocs or under memory pressure */
2998 static bool in_reclaim_compaction(struct scan_control
*sc
)
3000 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
3001 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
3002 sc
->priority
< DEF_PRIORITY
- 2))
3009 * Reclaim/compaction is used for high-order allocation requests. It reclaims
3010 * order-0 pages before compacting the zone. should_continue_reclaim() returns
3011 * true if more pages should be reclaimed such that when the page allocator
3012 * calls try_to_compact_pages() that it will have enough free pages to succeed.
3013 * It will give up earlier than that if there is difficulty reclaiming pages.
3015 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
3016 unsigned long nr_reclaimed
,
3017 struct scan_control
*sc
)
3019 unsigned long pages_for_compaction
;
3020 unsigned long inactive_lru_pages
;
3023 /* If not in reclaim/compaction mode, stop */
3024 if (!in_reclaim_compaction(sc
))
3028 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
3029 * number of pages that were scanned. This will return to the caller
3030 * with the risk reclaim/compaction and the resulting allocation attempt
3031 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
3032 * allocations through requiring that the full LRU list has been scanned
3033 * first, by assuming that zero delta of sc->nr_scanned means full LRU
3034 * scan, but that approximation was wrong, and there were corner cases
3035 * where always a non-zero amount of pages were scanned.
3040 /* If compaction would go ahead or the allocation would succeed, stop */
3041 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3042 struct zone
*zone
= &pgdat
->node_zones
[z
];
3043 if (!managed_zone(zone
))
3046 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
3047 case COMPACT_SUCCESS
:
3048 case COMPACT_CONTINUE
:
3051 /* check next zone */
3057 * If we have not reclaimed enough pages for compaction and the
3058 * inactive lists are large enough, continue reclaiming
3060 pages_for_compaction
= compact_gap(sc
->order
);
3061 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
3062 if (can_reclaim_anon_pages(NULL
, pgdat
->node_id
, sc
))
3063 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3065 return inactive_lru_pages
> pages_for_compaction
;
3068 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
3070 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
3071 struct mem_cgroup
*memcg
;
3073 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
3075 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3076 unsigned long reclaimed
;
3077 unsigned long scanned
;
3080 * This loop can become CPU-bound when target memcgs
3081 * aren't eligible for reclaim - either because they
3082 * don't have any reclaimable pages, or because their
3083 * memory is explicitly protected. Avoid soft lockups.
3087 mem_cgroup_calculate_protection(target_memcg
, memcg
);
3089 if (mem_cgroup_below_min(memcg
)) {
3092 * If there is no reclaimable memory, OOM.
3095 } else if (mem_cgroup_below_low(memcg
)) {
3098 * Respect the protection only as long as
3099 * there is an unprotected supply
3100 * of reclaimable memory from other cgroups.
3102 if (!sc
->memcg_low_reclaim
) {
3103 sc
->memcg_low_skipped
= 1;
3106 memcg_memory_event(memcg
, MEMCG_LOW
);
3109 reclaimed
= sc
->nr_reclaimed
;
3110 scanned
= sc
->nr_scanned
;
3112 shrink_lruvec(lruvec
, sc
);
3114 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
3117 /* Record the group's reclaim efficiency */
3118 vmpressure(sc
->gfp_mask
, memcg
, false,
3119 sc
->nr_scanned
- scanned
,
3120 sc
->nr_reclaimed
- reclaimed
);
3122 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
3125 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
3127 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
3128 unsigned long nr_reclaimed
, nr_scanned
;
3129 struct lruvec
*target_lruvec
;
3130 bool reclaimable
= false;
3133 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
3137 * Flush the memory cgroup stats, so that we read accurate per-memcg
3138 * lruvec stats for heuristics.
3140 mem_cgroup_flush_stats();
3142 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
3144 nr_reclaimed
= sc
->nr_reclaimed
;
3145 nr_scanned
= sc
->nr_scanned
;
3148 * Determine the scan balance between anon and file LRUs.
3150 spin_lock_irq(&target_lruvec
->lru_lock
);
3151 sc
->anon_cost
= target_lruvec
->anon_cost
;
3152 sc
->file_cost
= target_lruvec
->file_cost
;
3153 spin_unlock_irq(&target_lruvec
->lru_lock
);
3156 * Target desirable inactive:active list ratios for the anon
3157 * and file LRU lists.
3159 if (!sc
->force_deactivate
) {
3160 unsigned long refaults
;
3162 refaults
= lruvec_page_state(target_lruvec
,
3163 WORKINGSET_ACTIVATE_ANON
);
3164 if (refaults
!= target_lruvec
->refaults
[0] ||
3165 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
3166 sc
->may_deactivate
|= DEACTIVATE_ANON
;
3168 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
3171 * When refaults are being observed, it means a new
3172 * workingset is being established. Deactivate to get
3173 * rid of any stale active pages quickly.
3175 refaults
= lruvec_page_state(target_lruvec
,
3176 WORKINGSET_ACTIVATE_FILE
);
3177 if (refaults
!= target_lruvec
->refaults
[1] ||
3178 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
3179 sc
->may_deactivate
|= DEACTIVATE_FILE
;
3181 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
3183 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
3186 * If we have plenty of inactive file pages that aren't
3187 * thrashing, try to reclaim those first before touching
3190 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
3191 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
3192 sc
->cache_trim_mode
= 1;
3194 sc
->cache_trim_mode
= 0;
3197 * Prevent the reclaimer from falling into the cache trap: as
3198 * cache pages start out inactive, every cache fault will tip
3199 * the scan balance towards the file LRU. And as the file LRU
3200 * shrinks, so does the window for rotation from references.
3201 * This means we have a runaway feedback loop where a tiny
3202 * thrashing file LRU becomes infinitely more attractive than
3203 * anon pages. Try to detect this based on file LRU size.
3205 if (!cgroup_reclaim(sc
)) {
3206 unsigned long total_high_wmark
= 0;
3207 unsigned long free
, anon
;
3210 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
3211 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
3212 node_page_state(pgdat
, NR_INACTIVE_FILE
);
3214 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
3215 struct zone
*zone
= &pgdat
->node_zones
[z
];
3216 if (!managed_zone(zone
))
3219 total_high_wmark
+= high_wmark_pages(zone
);
3223 * Consider anon: if that's low too, this isn't a
3224 * runaway file reclaim problem, but rather just
3225 * extreme pressure. Reclaim as per usual then.
3227 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3230 file
+ free
<= total_high_wmark
&&
3231 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
3232 anon
>> sc
->priority
;
3235 shrink_node_memcgs(pgdat
, sc
);
3237 if (reclaim_state
) {
3238 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
3239 reclaim_state
->reclaimed_slab
= 0;
3242 /* Record the subtree's reclaim efficiency */
3243 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
3244 sc
->nr_scanned
- nr_scanned
,
3245 sc
->nr_reclaimed
- nr_reclaimed
);
3247 if (sc
->nr_reclaimed
- nr_reclaimed
)
3250 if (current_is_kswapd()) {
3252 * If reclaim is isolating dirty pages under writeback,
3253 * it implies that the long-lived page allocation rate
3254 * is exceeding the page laundering rate. Either the
3255 * global limits are not being effective at throttling
3256 * processes due to the page distribution throughout
3257 * zones or there is heavy usage of a slow backing
3258 * device. The only option is to throttle from reclaim
3259 * context which is not ideal as there is no guarantee
3260 * the dirtying process is throttled in the same way
3261 * balance_dirty_pages() manages.
3263 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3264 * count the number of pages under pages flagged for
3265 * immediate reclaim and stall if any are encountered
3266 * in the nr_immediate check below.
3268 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
3269 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3271 /* Allow kswapd to start writing pages during reclaim.*/
3272 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
3273 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3276 * If kswapd scans pages marked for immediate
3277 * reclaim and under writeback (nr_immediate), it
3278 * implies that pages are cycling through the LRU
3279 * faster than they are written so forcibly stall
3280 * until some pages complete writeback.
3282 if (sc
->nr
.immediate
)
3283 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_WRITEBACK
);
3287 * Tag a node/memcg as congested if all the dirty pages were marked
3288 * for writeback and immediate reclaim (counted in nr.congested).
3290 * Legacy memcg will stall in page writeback so avoid forcibly
3291 * stalling in reclaim_throttle().
3293 if ((current_is_kswapd() ||
3294 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
3295 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
3296 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
3299 * Stall direct reclaim for IO completions if the lruvec is
3300 * node is congested. Allow kswapd to continue until it
3301 * starts encountering unqueued dirty pages or cycling through
3302 * the LRU too quickly.
3304 if (!current_is_kswapd() && current_may_throttle() &&
3305 !sc
->hibernation_mode
&&
3306 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
3307 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_CONGESTED
);
3309 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
3314 * Kswapd gives up on balancing particular nodes after too
3315 * many failures to reclaim anything from them and goes to
3316 * sleep. On reclaim progress, reset the failure counter. A
3317 * successful direct reclaim run will revive a dormant kswapd.
3320 pgdat
->kswapd_failures
= 0;
3324 * Returns true if compaction should go ahead for a costly-order request, or
3325 * the allocation would already succeed without compaction. Return false if we
3326 * should reclaim first.
3328 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
3330 unsigned long watermark
;
3331 enum compact_result suitable
;
3333 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
3334 if (suitable
== COMPACT_SUCCESS
)
3335 /* Allocation should succeed already. Don't reclaim. */
3337 if (suitable
== COMPACT_SKIPPED
)
3338 /* Compaction cannot yet proceed. Do reclaim. */
3342 * Compaction is already possible, but it takes time to run and there
3343 * are potentially other callers using the pages just freed. So proceed
3344 * with reclaim to make a buffer of free pages available to give
3345 * compaction a reasonable chance of completing and allocating the page.
3346 * Note that we won't actually reclaim the whole buffer in one attempt
3347 * as the target watermark in should_continue_reclaim() is lower. But if
3348 * we are already above the high+gap watermark, don't reclaim at all.
3350 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
3352 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
3355 static void consider_reclaim_throttle(pg_data_t
*pgdat
, struct scan_control
*sc
)
3358 * If reclaim is making progress greater than 12% efficiency then
3359 * wake all the NOPROGRESS throttled tasks.
3361 if (sc
->nr_reclaimed
> (sc
->nr_scanned
>> 3)) {
3362 wait_queue_head_t
*wqh
;
3364 wqh
= &pgdat
->reclaim_wait
[VMSCAN_THROTTLE_NOPROGRESS
];
3365 if (waitqueue_active(wqh
))
3372 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
3373 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
3374 * under writeback and marked for immediate reclaim at the tail of the
3377 if (current_is_kswapd() || cgroup_reclaim(sc
))
3380 /* Throttle if making no progress at high prioities. */
3381 if (sc
->priority
== 1 && !sc
->nr_reclaimed
)
3382 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_NOPROGRESS
);
3386 * This is the direct reclaim path, for page-allocating processes. We only
3387 * try to reclaim pages from zones which will satisfy the caller's allocation
3390 * If a zone is deemed to be full of pinned pages then just give it a light
3391 * scan then give up on it.
3393 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
3397 unsigned long nr_soft_reclaimed
;
3398 unsigned long nr_soft_scanned
;
3400 pg_data_t
*last_pgdat
= NULL
;
3401 pg_data_t
*first_pgdat
= NULL
;
3404 * If the number of buffer_heads in the machine exceeds the maximum
3405 * allowed level, force direct reclaim to scan the highmem zone as
3406 * highmem pages could be pinning lowmem pages storing buffer_heads
3408 orig_mask
= sc
->gfp_mask
;
3409 if (buffer_heads_over_limit
) {
3410 sc
->gfp_mask
|= __GFP_HIGHMEM
;
3411 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
3414 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3415 sc
->reclaim_idx
, sc
->nodemask
) {
3417 * Take care memory controller reclaiming has small influence
3420 if (!cgroup_reclaim(sc
)) {
3421 if (!cpuset_zone_allowed(zone
,
3422 GFP_KERNEL
| __GFP_HARDWALL
))
3426 * If we already have plenty of memory free for
3427 * compaction in this zone, don't free any more.
3428 * Even though compaction is invoked for any
3429 * non-zero order, only frequent costly order
3430 * reclamation is disruptive enough to become a
3431 * noticeable problem, like transparent huge
3434 if (IS_ENABLED(CONFIG_COMPACTION
) &&
3435 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
3436 compaction_ready(zone
, sc
)) {
3437 sc
->compaction_ready
= true;
3442 * Shrink each node in the zonelist once. If the
3443 * zonelist is ordered by zone (not the default) then a
3444 * node may be shrunk multiple times but in that case
3445 * the user prefers lower zones being preserved.
3447 if (zone
->zone_pgdat
== last_pgdat
)
3451 * This steals pages from memory cgroups over softlimit
3452 * and returns the number of reclaimed pages and
3453 * scanned pages. This works for global memory pressure
3454 * and balancing, not for a memcg's limit.
3456 nr_soft_scanned
= 0;
3457 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
3458 sc
->order
, sc
->gfp_mask
,
3460 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
3461 sc
->nr_scanned
+= nr_soft_scanned
;
3462 /* need some check for avoid more shrink_zone() */
3466 first_pgdat
= zone
->zone_pgdat
;
3468 /* See comment about same check for global reclaim above */
3469 if (zone
->zone_pgdat
== last_pgdat
)
3471 last_pgdat
= zone
->zone_pgdat
;
3472 shrink_node(zone
->zone_pgdat
, sc
);
3476 consider_reclaim_throttle(first_pgdat
, sc
);
3479 * Restore to original mask to avoid the impact on the caller if we
3480 * promoted it to __GFP_HIGHMEM.
3482 sc
->gfp_mask
= orig_mask
;
3485 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
3487 struct lruvec
*target_lruvec
;
3488 unsigned long refaults
;
3490 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
3491 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
3492 target_lruvec
->refaults
[0] = refaults
;
3493 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
3494 target_lruvec
->refaults
[1] = refaults
;
3498 * This is the main entry point to direct page reclaim.
3500 * If a full scan of the inactive list fails to free enough memory then we
3501 * are "out of memory" and something needs to be killed.
3503 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3504 * high - the zone may be full of dirty or under-writeback pages, which this
3505 * caller can't do much about. We kick the writeback threads and take explicit
3506 * naps in the hope that some of these pages can be written. But if the
3507 * allocating task holds filesystem locks which prevent writeout this might not
3508 * work, and the allocation attempt will fail.
3510 * returns: 0, if no pages reclaimed
3511 * else, the number of pages reclaimed
3513 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3514 struct scan_control
*sc
)
3516 int initial_priority
= sc
->priority
;
3517 pg_data_t
*last_pgdat
;
3521 delayacct_freepages_start();
3523 if (!cgroup_reclaim(sc
))
3524 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3527 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3530 shrink_zones(zonelist
, sc
);
3532 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3535 if (sc
->compaction_ready
)
3539 * If we're getting trouble reclaiming, start doing
3540 * writepage even in laptop mode.
3542 if (sc
->priority
< DEF_PRIORITY
- 2)
3543 sc
->may_writepage
= 1;
3544 } while (--sc
->priority
>= 0);
3547 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3549 if (zone
->zone_pgdat
== last_pgdat
)
3551 last_pgdat
= zone
->zone_pgdat
;
3553 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3555 if (cgroup_reclaim(sc
)) {
3556 struct lruvec
*lruvec
;
3558 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3560 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3564 delayacct_freepages_end();
3566 if (sc
->nr_reclaimed
)
3567 return sc
->nr_reclaimed
;
3569 /* Aborted reclaim to try compaction? don't OOM, then */
3570 if (sc
->compaction_ready
)
3574 * We make inactive:active ratio decisions based on the node's
3575 * composition of memory, but a restrictive reclaim_idx or a
3576 * memory.low cgroup setting can exempt large amounts of
3577 * memory from reclaim. Neither of which are very common, so
3578 * instead of doing costly eligibility calculations of the
3579 * entire cgroup subtree up front, we assume the estimates are
3580 * good, and retry with forcible deactivation if that fails.
3582 if (sc
->skipped_deactivate
) {
3583 sc
->priority
= initial_priority
;
3584 sc
->force_deactivate
= 1;
3585 sc
->skipped_deactivate
= 0;
3589 /* Untapped cgroup reserves? Don't OOM, retry. */
3590 if (sc
->memcg_low_skipped
) {
3591 sc
->priority
= initial_priority
;
3592 sc
->force_deactivate
= 0;
3593 sc
->memcg_low_reclaim
= 1;
3594 sc
->memcg_low_skipped
= 0;
3601 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3604 unsigned long pfmemalloc_reserve
= 0;
3605 unsigned long free_pages
= 0;
3609 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3612 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3613 zone
= &pgdat
->node_zones
[i
];
3614 if (!managed_zone(zone
))
3617 if (!zone_reclaimable_pages(zone
))
3620 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3621 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3624 /* If there are no reserves (unexpected config) then do not throttle */
3625 if (!pfmemalloc_reserve
)
3628 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3630 /* kswapd must be awake if processes are being throttled */
3631 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3632 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
3633 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
3635 wake_up_interruptible(&pgdat
->kswapd_wait
);
3642 * Throttle direct reclaimers if backing storage is backed by the network
3643 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3644 * depleted. kswapd will continue to make progress and wake the processes
3645 * when the low watermark is reached.
3647 * Returns true if a fatal signal was delivered during throttling. If this
3648 * happens, the page allocator should not consider triggering the OOM killer.
3650 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3651 nodemask_t
*nodemask
)
3655 pg_data_t
*pgdat
= NULL
;
3658 * Kernel threads should not be throttled as they may be indirectly
3659 * responsible for cleaning pages necessary for reclaim to make forward
3660 * progress. kjournald for example may enter direct reclaim while
3661 * committing a transaction where throttling it could forcing other
3662 * processes to block on log_wait_commit().
3664 if (current
->flags
& PF_KTHREAD
)
3668 * If a fatal signal is pending, this process should not throttle.
3669 * It should return quickly so it can exit and free its memory
3671 if (fatal_signal_pending(current
))
3675 * Check if the pfmemalloc reserves are ok by finding the first node
3676 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3677 * GFP_KERNEL will be required for allocating network buffers when
3678 * swapping over the network so ZONE_HIGHMEM is unusable.
3680 * Throttling is based on the first usable node and throttled processes
3681 * wait on a queue until kswapd makes progress and wakes them. There
3682 * is an affinity then between processes waking up and where reclaim
3683 * progress has been made assuming the process wakes on the same node.
3684 * More importantly, processes running on remote nodes will not compete
3685 * for remote pfmemalloc reserves and processes on different nodes
3686 * should make reasonable progress.
3688 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3689 gfp_zone(gfp_mask
), nodemask
) {
3690 if (zone_idx(zone
) > ZONE_NORMAL
)
3693 /* Throttle based on the first usable node */
3694 pgdat
= zone
->zone_pgdat
;
3695 if (allow_direct_reclaim(pgdat
))
3700 /* If no zone was usable by the allocation flags then do not throttle */
3704 /* Account for the throttling */
3705 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3708 * If the caller cannot enter the filesystem, it's possible that it
3709 * is due to the caller holding an FS lock or performing a journal
3710 * transaction in the case of a filesystem like ext[3|4]. In this case,
3711 * it is not safe to block on pfmemalloc_wait as kswapd could be
3712 * blocked waiting on the same lock. Instead, throttle for up to a
3713 * second before continuing.
3715 if (!(gfp_mask
& __GFP_FS
))
3716 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3717 allow_direct_reclaim(pgdat
), HZ
);
3719 /* Throttle until kswapd wakes the process */
3720 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3721 allow_direct_reclaim(pgdat
));
3723 if (fatal_signal_pending(current
))
3730 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3731 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3733 unsigned long nr_reclaimed
;
3734 struct scan_control sc
= {
3735 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3736 .gfp_mask
= current_gfp_context(gfp_mask
),
3737 .reclaim_idx
= gfp_zone(gfp_mask
),
3739 .nodemask
= nodemask
,
3740 .priority
= DEF_PRIORITY
,
3741 .may_writepage
= !laptop_mode
,
3747 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3748 * Confirm they are large enough for max values.
3750 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3751 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3752 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3755 * Do not enter reclaim if fatal signal was delivered while throttled.
3756 * 1 is returned so that the page allocator does not OOM kill at this
3759 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3762 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3763 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3765 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3767 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3768 set_task_reclaim_state(current
, NULL
);
3770 return nr_reclaimed
;
3775 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3776 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3777 gfp_t gfp_mask
, bool noswap
,
3779 unsigned long *nr_scanned
)
3781 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3782 struct scan_control sc
= {
3783 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3784 .target_mem_cgroup
= memcg
,
3785 .may_writepage
= !laptop_mode
,
3787 .reclaim_idx
= MAX_NR_ZONES
- 1,
3788 .may_swap
= !noswap
,
3791 WARN_ON_ONCE(!current
->reclaim_state
);
3793 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3794 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3796 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3800 * NOTE: Although we can get the priority field, using it
3801 * here is not a good idea, since it limits the pages we can scan.
3802 * if we don't reclaim here, the shrink_node from balance_pgdat
3803 * will pick up pages from other mem cgroup's as well. We hack
3804 * the priority and make it zero.
3806 shrink_lruvec(lruvec
, &sc
);
3808 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3810 *nr_scanned
= sc
.nr_scanned
;
3812 return sc
.nr_reclaimed
;
3815 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3816 unsigned long nr_pages
,
3820 unsigned long nr_reclaimed
;
3821 unsigned int noreclaim_flag
;
3822 struct scan_control sc
= {
3823 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3824 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3825 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3826 .reclaim_idx
= MAX_NR_ZONES
- 1,
3827 .target_mem_cgroup
= memcg
,
3828 .priority
= DEF_PRIORITY
,
3829 .may_writepage
= !laptop_mode
,
3831 .may_swap
= may_swap
,
3834 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3835 * equal pressure on all the nodes. This is based on the assumption that
3836 * the reclaim does not bail out early.
3838 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3840 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3841 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3842 noreclaim_flag
= memalloc_noreclaim_save();
3844 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3846 memalloc_noreclaim_restore(noreclaim_flag
);
3847 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3848 set_task_reclaim_state(current
, NULL
);
3850 return nr_reclaimed
;
3854 static void age_active_anon(struct pglist_data
*pgdat
,
3855 struct scan_control
*sc
)
3857 struct mem_cgroup
*memcg
;
3858 struct lruvec
*lruvec
;
3860 if (!can_age_anon_pages(pgdat
, sc
))
3863 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3864 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3867 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3869 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3870 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3871 sc
, LRU_ACTIVE_ANON
);
3872 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3876 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
3882 * Check for watermark boosts top-down as the higher zones
3883 * are more likely to be boosted. Both watermarks and boosts
3884 * should not be checked at the same time as reclaim would
3885 * start prematurely when there is no boosting and a lower
3888 for (i
= highest_zoneidx
; i
>= 0; i
--) {
3889 zone
= pgdat
->node_zones
+ i
;
3890 if (!managed_zone(zone
))
3893 if (zone
->watermark_boost
)
3901 * Returns true if there is an eligible zone balanced for the request order
3902 * and highest_zoneidx
3904 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3907 unsigned long mark
= -1;
3911 * Check watermarks bottom-up as lower zones are more likely to
3914 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3915 zone
= pgdat
->node_zones
+ i
;
3917 if (!managed_zone(zone
))
3920 if (sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
)
3921 mark
= wmark_pages(zone
, WMARK_PROMO
);
3923 mark
= high_wmark_pages(zone
);
3924 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
3929 * If a node has no managed zone within highest_zoneidx, it does not
3930 * need balancing by definition. This can happen if a zone-restricted
3931 * allocation tries to wake a remote kswapd.
3939 /* Clear pgdat state for congested, dirty or under writeback. */
3940 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3942 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3944 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3945 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3946 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3950 * Prepare kswapd for sleeping. This verifies that there are no processes
3951 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3953 * Returns true if kswapd is ready to sleep
3955 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
3956 int highest_zoneidx
)
3959 * The throttled processes are normally woken up in balance_pgdat() as
3960 * soon as allow_direct_reclaim() is true. But there is a potential
3961 * race between when kswapd checks the watermarks and a process gets
3962 * throttled. There is also a potential race if processes get
3963 * throttled, kswapd wakes, a large process exits thereby balancing the
3964 * zones, which causes kswapd to exit balance_pgdat() before reaching
3965 * the wake up checks. If kswapd is going to sleep, no process should
3966 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3967 * the wake up is premature, processes will wake kswapd and get
3968 * throttled again. The difference from wake ups in balance_pgdat() is
3969 * that here we are under prepare_to_wait().
3971 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3972 wake_up_all(&pgdat
->pfmemalloc_wait
);
3974 /* Hopeless node, leave it to direct reclaim */
3975 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3978 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
3979 clear_pgdat_congested(pgdat
);
3987 * kswapd shrinks a node of pages that are at or below the highest usable
3988 * zone that is currently unbalanced.
3990 * Returns true if kswapd scanned at least the requested number of pages to
3991 * reclaim or if the lack of progress was due to pages under writeback.
3992 * This is used to determine if the scanning priority needs to be raised.
3994 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3995 struct scan_control
*sc
)
4000 /* Reclaim a number of pages proportional to the number of zones */
4001 sc
->nr_to_reclaim
= 0;
4002 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
4003 zone
= pgdat
->node_zones
+ z
;
4004 if (!managed_zone(zone
))
4007 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
4011 * Historically care was taken to put equal pressure on all zones but
4012 * now pressure is applied based on node LRU order.
4014 shrink_node(pgdat
, sc
);
4017 * Fragmentation may mean that the system cannot be rebalanced for
4018 * high-order allocations. If twice the allocation size has been
4019 * reclaimed then recheck watermarks only at order-0 to prevent
4020 * excessive reclaim. Assume that a process requested a high-order
4021 * can direct reclaim/compact.
4023 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
4026 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
4029 /* Page allocator PCP high watermark is lowered if reclaim is active. */
4031 update_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
, bool active
)
4036 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4037 zone
= pgdat
->node_zones
+ i
;
4039 if (!managed_zone(zone
))
4043 set_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4045 clear_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4050 set_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4052 update_reclaim_active(pgdat
, highest_zoneidx
, true);
4056 clear_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4058 update_reclaim_active(pgdat
, highest_zoneidx
, false);
4062 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
4063 * that are eligible for use by the caller until at least one zone is
4066 * Returns the order kswapd finished reclaiming at.
4068 * kswapd scans the zones in the highmem->normal->dma direction. It skips
4069 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
4070 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
4071 * or lower is eligible for reclaim until at least one usable zone is
4074 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
4077 unsigned long nr_soft_reclaimed
;
4078 unsigned long nr_soft_scanned
;
4079 unsigned long pflags
;
4080 unsigned long nr_boost_reclaim
;
4081 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
4084 struct scan_control sc
= {
4085 .gfp_mask
= GFP_KERNEL
,
4090 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4091 psi_memstall_enter(&pflags
);
4092 __fs_reclaim_acquire(_THIS_IP_
);
4094 count_vm_event(PAGEOUTRUN
);
4097 * Account for the reclaim boost. Note that the zone boost is left in
4098 * place so that parallel allocations that are near the watermark will
4099 * stall or direct reclaim until kswapd is finished.
4101 nr_boost_reclaim
= 0;
4102 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4103 zone
= pgdat
->node_zones
+ i
;
4104 if (!managed_zone(zone
))
4107 nr_boost_reclaim
+= zone
->watermark_boost
;
4108 zone_boosts
[i
] = zone
->watermark_boost
;
4110 boosted
= nr_boost_reclaim
;
4113 set_reclaim_active(pgdat
, highest_zoneidx
);
4114 sc
.priority
= DEF_PRIORITY
;
4116 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
4117 bool raise_priority
= true;
4121 sc
.reclaim_idx
= highest_zoneidx
;
4124 * If the number of buffer_heads exceeds the maximum allowed
4125 * then consider reclaiming from all zones. This has a dual
4126 * purpose -- on 64-bit systems it is expected that
4127 * buffer_heads are stripped during active rotation. On 32-bit
4128 * systems, highmem pages can pin lowmem memory and shrinking
4129 * buffers can relieve lowmem pressure. Reclaim may still not
4130 * go ahead if all eligible zones for the original allocation
4131 * request are balanced to avoid excessive reclaim from kswapd.
4133 if (buffer_heads_over_limit
) {
4134 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
4135 zone
= pgdat
->node_zones
+ i
;
4136 if (!managed_zone(zone
))
4145 * If the pgdat is imbalanced then ignore boosting and preserve
4146 * the watermarks for a later time and restart. Note that the
4147 * zone watermarks will be still reset at the end of balancing
4148 * on the grounds that the normal reclaim should be enough to
4149 * re-evaluate if boosting is required when kswapd next wakes.
4151 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
4152 if (!balanced
&& nr_boost_reclaim
) {
4153 nr_boost_reclaim
= 0;
4158 * If boosting is not active then only reclaim if there are no
4159 * eligible zones. Note that sc.reclaim_idx is not used as
4160 * buffer_heads_over_limit may have adjusted it.
4162 if (!nr_boost_reclaim
&& balanced
)
4165 /* Limit the priority of boosting to avoid reclaim writeback */
4166 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
4167 raise_priority
= false;
4170 * Do not writeback or swap pages for boosted reclaim. The
4171 * intent is to relieve pressure not issue sub-optimal IO
4172 * from reclaim context. If no pages are reclaimed, the
4173 * reclaim will be aborted.
4175 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
4176 sc
.may_swap
= !nr_boost_reclaim
;
4179 * Do some background aging of the anon list, to give
4180 * pages a chance to be referenced before reclaiming. All
4181 * pages are rotated regardless of classzone as this is
4182 * about consistent aging.
4184 age_active_anon(pgdat
, &sc
);
4187 * If we're getting trouble reclaiming, start doing writepage
4188 * even in laptop mode.
4190 if (sc
.priority
< DEF_PRIORITY
- 2)
4191 sc
.may_writepage
= 1;
4193 /* Call soft limit reclaim before calling shrink_node. */
4195 nr_soft_scanned
= 0;
4196 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
4197 sc
.gfp_mask
, &nr_soft_scanned
);
4198 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
4201 * There should be no need to raise the scanning priority if
4202 * enough pages are already being scanned that that high
4203 * watermark would be met at 100% efficiency.
4205 if (kswapd_shrink_node(pgdat
, &sc
))
4206 raise_priority
= false;
4209 * If the low watermark is met there is no need for processes
4210 * to be throttled on pfmemalloc_wait as they should not be
4211 * able to safely make forward progress. Wake them
4213 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
4214 allow_direct_reclaim(pgdat
))
4215 wake_up_all(&pgdat
->pfmemalloc_wait
);
4217 /* Check if kswapd should be suspending */
4218 __fs_reclaim_release(_THIS_IP_
);
4219 ret
= try_to_freeze();
4220 __fs_reclaim_acquire(_THIS_IP_
);
4221 if (ret
|| kthread_should_stop())
4225 * Raise priority if scanning rate is too low or there was no
4226 * progress in reclaiming pages
4228 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
4229 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
4232 * If reclaim made no progress for a boost, stop reclaim as
4233 * IO cannot be queued and it could be an infinite loop in
4234 * extreme circumstances.
4236 if (nr_boost_reclaim
&& !nr_reclaimed
)
4239 if (raise_priority
|| !nr_reclaimed
)
4241 } while (sc
.priority
>= 1);
4243 if (!sc
.nr_reclaimed
)
4244 pgdat
->kswapd_failures
++;
4247 clear_reclaim_active(pgdat
, highest_zoneidx
);
4249 /* If reclaim was boosted, account for the reclaim done in this pass */
4251 unsigned long flags
;
4253 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4254 if (!zone_boosts
[i
])
4257 /* Increments are under the zone lock */
4258 zone
= pgdat
->node_zones
+ i
;
4259 spin_lock_irqsave(&zone
->lock
, flags
);
4260 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
4261 spin_unlock_irqrestore(&zone
->lock
, flags
);
4265 * As there is now likely space, wakeup kcompact to defragment
4268 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
4271 snapshot_refaults(NULL
, pgdat
);
4272 __fs_reclaim_release(_THIS_IP_
);
4273 psi_memstall_leave(&pflags
);
4274 set_task_reclaim_state(current
, NULL
);
4277 * Return the order kswapd stopped reclaiming at as
4278 * prepare_kswapd_sleep() takes it into account. If another caller
4279 * entered the allocator slow path while kswapd was awake, order will
4280 * remain at the higher level.
4286 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4287 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4288 * not a valid index then either kswapd runs for first time or kswapd couldn't
4289 * sleep after previous reclaim attempt (node is still unbalanced). In that
4290 * case return the zone index of the previous kswapd reclaim cycle.
4292 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
4293 enum zone_type prev_highest_zoneidx
)
4295 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4297 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
4300 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
4301 unsigned int highest_zoneidx
)
4306 if (freezing(current
) || kthread_should_stop())
4309 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4312 * Try to sleep for a short interval. Note that kcompactd will only be
4313 * woken if it is possible to sleep for a short interval. This is
4314 * deliberate on the assumption that if reclaim cannot keep an
4315 * eligible zone balanced that it's also unlikely that compaction will
4318 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4320 * Compaction records what page blocks it recently failed to
4321 * isolate pages from and skips them in the future scanning.
4322 * When kswapd is going to sleep, it is reasonable to assume
4323 * that pages and compaction may succeed so reset the cache.
4325 reset_isolation_suitable(pgdat
);
4328 * We have freed the memory, now we should compact it to make
4329 * allocation of the requested order possible.
4331 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
4333 remaining
= schedule_timeout(HZ
/10);
4336 * If woken prematurely then reset kswapd_highest_zoneidx and
4337 * order. The values will either be from a wakeup request or
4338 * the previous request that slept prematurely.
4341 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
4342 kswapd_highest_zoneidx(pgdat
,
4345 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
4346 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
4349 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4350 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4354 * After a short sleep, check if it was a premature sleep. If not, then
4355 * go fully to sleep until explicitly woken up.
4358 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4359 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
4362 * vmstat counters are not perfectly accurate and the estimated
4363 * value for counters such as NR_FREE_PAGES can deviate from the
4364 * true value by nr_online_cpus * threshold. To avoid the zone
4365 * watermarks being breached while under pressure, we reduce the
4366 * per-cpu vmstat threshold while kswapd is awake and restore
4367 * them before going back to sleep.
4369 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
4371 if (!kthread_should_stop())
4374 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
4377 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
4379 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
4381 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4385 * The background pageout daemon, started as a kernel thread
4386 * from the init process.
4388 * This basically trickles out pages so that we have _some_
4389 * free memory available even if there is no other activity
4390 * that frees anything up. This is needed for things like routing
4391 * etc, where we otherwise might have all activity going on in
4392 * asynchronous contexts that cannot page things out.
4394 * If there are applications that are active memory-allocators
4395 * (most normal use), this basically shouldn't matter.
4397 static int kswapd(void *p
)
4399 unsigned int alloc_order
, reclaim_order
;
4400 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
4401 pg_data_t
*pgdat
= (pg_data_t
*)p
;
4402 struct task_struct
*tsk
= current
;
4403 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
4405 if (!cpumask_empty(cpumask
))
4406 set_cpus_allowed_ptr(tsk
, cpumask
);
4409 * Tell the memory management that we're a "memory allocator",
4410 * and that if we need more memory we should get access to it
4411 * regardless (see "__alloc_pages()"). "kswapd" should
4412 * never get caught in the normal page freeing logic.
4414 * (Kswapd normally doesn't need memory anyway, but sometimes
4415 * you need a small amount of memory in order to be able to
4416 * page out something else, and this flag essentially protects
4417 * us from recursively trying to free more memory as we're
4418 * trying to free the first piece of memory in the first place).
4420 tsk
->flags
|= PF_MEMALLOC
| PF_KSWAPD
;
4423 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4424 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4425 atomic_set(&pgdat
->nr_writeback_throttled
, 0);
4429 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
4430 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4434 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
4437 /* Read the new order and highest_zoneidx */
4438 alloc_order
= READ_ONCE(pgdat
->kswapd_order
);
4439 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4441 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4442 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4444 ret
= try_to_freeze();
4445 if (kthread_should_stop())
4449 * We can speed up thawing tasks if we don't call balance_pgdat
4450 * after returning from the refrigerator
4456 * Reclaim begins at the requested order but if a high-order
4457 * reclaim fails then kswapd falls back to reclaiming for
4458 * order-0. If that happens, kswapd will consider sleeping
4459 * for the order it finished reclaiming at (reclaim_order)
4460 * but kcompactd is woken to compact for the original
4461 * request (alloc_order).
4463 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
4465 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
4467 if (reclaim_order
< alloc_order
)
4468 goto kswapd_try_sleep
;
4471 tsk
->flags
&= ~(PF_MEMALLOC
| PF_KSWAPD
);
4477 * A zone is low on free memory or too fragmented for high-order memory. If
4478 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4479 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4480 * has failed or is not needed, still wake up kcompactd if only compaction is
4483 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
4484 enum zone_type highest_zoneidx
)
4487 enum zone_type curr_idx
;
4489 if (!managed_zone(zone
))
4492 if (!cpuset_zone_allowed(zone
, gfp_flags
))
4495 pgdat
= zone
->zone_pgdat
;
4496 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4498 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
4499 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
4501 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
4502 WRITE_ONCE(pgdat
->kswapd_order
, order
);
4504 if (!waitqueue_active(&pgdat
->kswapd_wait
))
4507 /* Hopeless node, leave it to direct reclaim if possible */
4508 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
4509 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
4510 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
4512 * There may be plenty of free memory available, but it's too
4513 * fragmented for high-order allocations. Wake up kcompactd
4514 * and rely on compaction_suitable() to determine if it's
4515 * needed. If it fails, it will defer subsequent attempts to
4516 * ratelimit its work.
4518 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
4519 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
4523 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
4525 wake_up_interruptible(&pgdat
->kswapd_wait
);
4528 #ifdef CONFIG_HIBERNATION
4530 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4533 * Rather than trying to age LRUs the aim is to preserve the overall
4534 * LRU order by reclaiming preferentially
4535 * inactive > active > active referenced > active mapped
4537 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4539 struct scan_control sc
= {
4540 .nr_to_reclaim
= nr_to_reclaim
,
4541 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4542 .reclaim_idx
= MAX_NR_ZONES
- 1,
4543 .priority
= DEF_PRIORITY
,
4547 .hibernation_mode
= 1,
4549 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4550 unsigned long nr_reclaimed
;
4551 unsigned int noreclaim_flag
;
4553 fs_reclaim_acquire(sc
.gfp_mask
);
4554 noreclaim_flag
= memalloc_noreclaim_save();
4555 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4557 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4559 set_task_reclaim_state(current
, NULL
);
4560 memalloc_noreclaim_restore(noreclaim_flag
);
4561 fs_reclaim_release(sc
.gfp_mask
);
4563 return nr_reclaimed
;
4565 #endif /* CONFIG_HIBERNATION */
4568 * This kswapd start function will be called by init and node-hot-add.
4569 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4571 void kswapd_run(int nid
)
4573 pg_data_t
*pgdat
= NODE_DATA(nid
);
4578 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4579 if (IS_ERR(pgdat
->kswapd
)) {
4580 /* failure at boot is fatal */
4581 BUG_ON(system_state
< SYSTEM_RUNNING
);
4582 pr_err("Failed to start kswapd on node %d\n", nid
);
4583 pgdat
->kswapd
= NULL
;
4588 * Called by memory hotplug when all memory in a node is offlined. Caller must
4589 * hold mem_hotplug_begin/end().
4591 void kswapd_stop(int nid
)
4593 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4596 kthread_stop(kswapd
);
4597 NODE_DATA(nid
)->kswapd
= NULL
;
4601 static int __init
kswapd_init(void)
4606 for_each_node_state(nid
, N_MEMORY
)
4611 module_init(kswapd_init
)
4617 * If non-zero call node_reclaim when the number of free pages falls below
4620 int node_reclaim_mode __read_mostly
;
4623 * Priority for NODE_RECLAIM. This determines the fraction of pages
4624 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4627 #define NODE_RECLAIM_PRIORITY 4
4630 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4633 int sysctl_min_unmapped_ratio
= 1;
4636 * If the number of slab pages in a zone grows beyond this percentage then
4637 * slab reclaim needs to occur.
4639 int sysctl_min_slab_ratio
= 5;
4641 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4643 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4644 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4645 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4648 * It's possible for there to be more file mapped pages than
4649 * accounted for by the pages on the file LRU lists because
4650 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4652 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4655 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4656 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4658 unsigned long nr_pagecache_reclaimable
;
4659 unsigned long delta
= 0;
4662 * If RECLAIM_UNMAP is set, then all file pages are considered
4663 * potentially reclaimable. Otherwise, we have to worry about
4664 * pages like swapcache and node_unmapped_file_pages() provides
4667 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4668 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4670 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4672 /* If we can't clean pages, remove dirty pages from consideration */
4673 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4674 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4676 /* Watch for any possible underflows due to delta */
4677 if (unlikely(delta
> nr_pagecache_reclaimable
))
4678 delta
= nr_pagecache_reclaimable
;
4680 return nr_pagecache_reclaimable
- delta
;
4684 * Try to free up some pages from this node through reclaim.
4686 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4688 /* Minimum pages needed in order to stay on node */
4689 const unsigned long nr_pages
= 1 << order
;
4690 struct task_struct
*p
= current
;
4691 unsigned int noreclaim_flag
;
4692 struct scan_control sc
= {
4693 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4694 .gfp_mask
= current_gfp_context(gfp_mask
),
4696 .priority
= NODE_RECLAIM_PRIORITY
,
4697 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4698 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4700 .reclaim_idx
= gfp_zone(gfp_mask
),
4702 unsigned long pflags
;
4704 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4708 psi_memstall_enter(&pflags
);
4709 fs_reclaim_acquire(sc
.gfp_mask
);
4711 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4713 noreclaim_flag
= memalloc_noreclaim_save();
4714 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4716 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4718 * Free memory by calling shrink node with increasing
4719 * priorities until we have enough memory freed.
4722 shrink_node(pgdat
, &sc
);
4723 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4726 set_task_reclaim_state(p
, NULL
);
4727 memalloc_noreclaim_restore(noreclaim_flag
);
4728 fs_reclaim_release(sc
.gfp_mask
);
4729 psi_memstall_leave(&pflags
);
4731 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4733 return sc
.nr_reclaimed
>= nr_pages
;
4736 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4741 * Node reclaim reclaims unmapped file backed pages and
4742 * slab pages if we are over the defined limits.
4744 * A small portion of unmapped file backed pages is needed for
4745 * file I/O otherwise pages read by file I/O will be immediately
4746 * thrown out if the node is overallocated. So we do not reclaim
4747 * if less than a specified percentage of the node is used by
4748 * unmapped file backed pages.
4750 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4751 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
4752 pgdat
->min_slab_pages
)
4753 return NODE_RECLAIM_FULL
;
4756 * Do not scan if the allocation should not be delayed.
4758 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4759 return NODE_RECLAIM_NOSCAN
;
4762 * Only run node reclaim on the local node or on nodes that do not
4763 * have associated processors. This will favor the local processor
4764 * over remote processors and spread off node memory allocations
4765 * as wide as possible.
4767 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4768 return NODE_RECLAIM_NOSCAN
;
4770 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4771 return NODE_RECLAIM_NOSCAN
;
4773 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4774 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4777 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4784 * check_move_unevictable_pages - check pages for evictability and move to
4785 * appropriate zone lru list
4786 * @pvec: pagevec with lru pages to check
4788 * Checks pages for evictability, if an evictable page is in the unevictable
4789 * lru list, moves it to the appropriate evictable lru list. This function
4790 * should be only used for lru pages.
4792 void check_move_unevictable_pages(struct pagevec
*pvec
)
4794 struct lruvec
*lruvec
= NULL
;
4799 for (i
= 0; i
< pvec
->nr
; i
++) {
4800 struct page
*page
= pvec
->pages
[i
];
4801 struct folio
*folio
= page_folio(page
);
4804 if (PageTransTail(page
))
4807 nr_pages
= thp_nr_pages(page
);
4808 pgscanned
+= nr_pages
;
4810 /* block memcg migration during page moving between lru */
4811 if (!TestClearPageLRU(page
))
4814 lruvec
= folio_lruvec_relock_irq(folio
, lruvec
);
4815 if (page_evictable(page
) && PageUnevictable(page
)) {
4816 del_page_from_lru_list(page
, lruvec
);
4817 ClearPageUnevictable(page
);
4818 add_page_to_lru_list(page
, lruvec
);
4819 pgrescued
+= nr_pages
;
4825 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4826 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4827 unlock_page_lruvec_irq(lruvec
);
4828 } else if (pgscanned
) {
4829 count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4832 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);