1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/migrate.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/pagevec.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52 #include <linux/psi.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
57 #include <linux/swapops.h>
58 #include <linux/balloon_compaction.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/vmscan.h>
66 /* How many pages shrink_list() should reclaim */
67 unsigned long nr_to_reclaim
;
70 * Nodemask of nodes allowed by the caller. If NULL, all nodes
76 * The memory cgroup that hit its limit and as a result is the
77 * primary target of this reclaim invocation.
79 struct mem_cgroup
*target_mem_cgroup
;
82 * Scan pressure balancing between anon and file LRUs
84 unsigned long anon_cost
;
85 unsigned long file_cost
;
87 /* Can active pages be deactivated as part of reclaim? */
88 #define DEACTIVATE_ANON 1
89 #define DEACTIVATE_FILE 2
90 unsigned int may_deactivate
:2;
91 unsigned int force_deactivate
:1;
92 unsigned int skipped_deactivate
:1;
94 /* Writepage batching in laptop mode; RECLAIM_WRITE */
95 unsigned int may_writepage
:1;
97 /* Can mapped pages be reclaimed? */
98 unsigned int may_unmap
:1;
100 /* Can pages be swapped as part of reclaim? */
101 unsigned int may_swap
:1;
104 * Cgroup memory below memory.low is protected as long as we
105 * don't threaten to OOM. If any cgroup is reclaimed at
106 * reduced force or passed over entirely due to its memory.low
107 * setting (memcg_low_skipped), and nothing is reclaimed as a
108 * result, then go back for one more cycle that reclaims the protected
109 * memory (memcg_low_reclaim) to avert OOM.
111 unsigned int memcg_low_reclaim
:1;
112 unsigned int memcg_low_skipped
:1;
114 unsigned int hibernation_mode
:1;
116 /* One of the zones is ready for compaction */
117 unsigned int compaction_ready
:1;
119 /* There is easily reclaimable cold cache in the current node */
120 unsigned int cache_trim_mode
:1;
122 /* The file pages on the current node are dangerously low */
123 unsigned int file_is_tiny
:1;
125 /* Always discard instead of demoting to lower tier memory */
126 unsigned int no_demotion
:1;
128 /* Allocation order */
131 /* Scan (total_size >> priority) pages at once */
134 /* The highest zone to isolate pages for reclaim from */
137 /* This context's GFP mask */
140 /* Incremented by the number of inactive pages that were scanned */
141 unsigned long nr_scanned
;
143 /* Number of pages freed so far during a call to shrink_zones() */
144 unsigned long nr_reclaimed
;
148 unsigned int unqueued_dirty
;
149 unsigned int congested
;
150 unsigned int writeback
;
151 unsigned int immediate
;
152 unsigned int file_taken
;
156 /* for recording the reclaimed slab by now */
157 struct reclaim_state reclaim_state
;
160 #ifdef ARCH_HAS_PREFETCHW
161 #define prefetchw_prev_lru_page(_page, _base, _field) \
163 if ((_page)->lru.prev != _base) { \
166 prev = lru_to_page(&(_page->lru)); \
167 prefetchw(&prev->_field); \
171 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
175 * From 0 .. 200. Higher means more swappy.
177 int vm_swappiness
= 60;
179 static void set_task_reclaim_state(struct task_struct
*task
,
180 struct reclaim_state
*rs
)
182 /* Check for an overwrite */
183 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
185 /* Check for the nulling of an already-nulled member */
186 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
188 task
->reclaim_state
= rs
;
191 static LIST_HEAD(shrinker_list
);
192 static DECLARE_RWSEM(shrinker_rwsem
);
195 static int shrinker_nr_max
;
197 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
198 static inline int shrinker_map_size(int nr_items
)
200 return (DIV_ROUND_UP(nr_items
, BITS_PER_LONG
) * sizeof(unsigned long));
203 static inline int shrinker_defer_size(int nr_items
)
205 return (round_up(nr_items
, BITS_PER_LONG
) * sizeof(atomic_long_t
));
208 static struct shrinker_info
*shrinker_info_protected(struct mem_cgroup
*memcg
,
211 return rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_info
,
212 lockdep_is_held(&shrinker_rwsem
));
215 static int expand_one_shrinker_info(struct mem_cgroup
*memcg
,
216 int map_size
, int defer_size
,
217 int old_map_size
, int old_defer_size
)
219 struct shrinker_info
*new, *old
;
220 struct mem_cgroup_per_node
*pn
;
222 int size
= map_size
+ defer_size
;
225 pn
= memcg
->nodeinfo
[nid
];
226 old
= shrinker_info_protected(memcg
, nid
);
227 /* Not yet online memcg */
231 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
235 new->nr_deferred
= (atomic_long_t
*)(new + 1);
236 new->map
= (void *)new->nr_deferred
+ defer_size
;
238 /* map: set all old bits, clear all new bits */
239 memset(new->map
, (int)0xff, old_map_size
);
240 memset((void *)new->map
+ old_map_size
, 0, map_size
- old_map_size
);
241 /* nr_deferred: copy old values, clear all new values */
242 memcpy(new->nr_deferred
, old
->nr_deferred
, old_defer_size
);
243 memset((void *)new->nr_deferred
+ old_defer_size
, 0,
244 defer_size
- old_defer_size
);
246 rcu_assign_pointer(pn
->shrinker_info
, new);
247 kvfree_rcu(old
, rcu
);
253 void free_shrinker_info(struct mem_cgroup
*memcg
)
255 struct mem_cgroup_per_node
*pn
;
256 struct shrinker_info
*info
;
260 pn
= memcg
->nodeinfo
[nid
];
261 info
= rcu_dereference_protected(pn
->shrinker_info
, true);
263 rcu_assign_pointer(pn
->shrinker_info
, NULL
);
267 int alloc_shrinker_info(struct mem_cgroup
*memcg
)
269 struct shrinker_info
*info
;
270 int nid
, size
, ret
= 0;
271 int map_size
, defer_size
= 0;
273 down_write(&shrinker_rwsem
);
274 map_size
= shrinker_map_size(shrinker_nr_max
);
275 defer_size
= shrinker_defer_size(shrinker_nr_max
);
276 size
= map_size
+ defer_size
;
278 info
= kvzalloc_node(sizeof(*info
) + size
, GFP_KERNEL
, nid
);
280 free_shrinker_info(memcg
);
284 info
->nr_deferred
= (atomic_long_t
*)(info
+ 1);
285 info
->map
= (void *)info
->nr_deferred
+ defer_size
;
286 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_info
, info
);
288 up_write(&shrinker_rwsem
);
293 static inline bool need_expand(int nr_max
)
295 return round_up(nr_max
, BITS_PER_LONG
) >
296 round_up(shrinker_nr_max
, BITS_PER_LONG
);
299 static int expand_shrinker_info(int new_id
)
302 int new_nr_max
= new_id
+ 1;
303 int map_size
, defer_size
= 0;
304 int old_map_size
, old_defer_size
= 0;
305 struct mem_cgroup
*memcg
;
307 if (!need_expand(new_nr_max
))
310 if (!root_mem_cgroup
)
313 lockdep_assert_held(&shrinker_rwsem
);
315 map_size
= shrinker_map_size(new_nr_max
);
316 defer_size
= shrinker_defer_size(new_nr_max
);
317 old_map_size
= shrinker_map_size(shrinker_nr_max
);
318 old_defer_size
= shrinker_defer_size(shrinker_nr_max
);
320 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
322 ret
= expand_one_shrinker_info(memcg
, map_size
, defer_size
,
323 old_map_size
, old_defer_size
);
325 mem_cgroup_iter_break(NULL
, memcg
);
328 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
331 shrinker_nr_max
= new_nr_max
;
336 void set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
338 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
339 struct shrinker_info
*info
;
342 info
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_info
);
343 /* Pairs with smp mb in shrink_slab() */
344 smp_mb__before_atomic();
345 set_bit(shrinker_id
, info
->map
);
350 static DEFINE_IDR(shrinker_idr
);
352 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
354 int id
, ret
= -ENOMEM
;
356 if (mem_cgroup_disabled())
359 down_write(&shrinker_rwsem
);
360 /* This may call shrinker, so it must use down_read_trylock() */
361 id
= idr_alloc(&shrinker_idr
, shrinker
, 0, 0, GFP_KERNEL
);
365 if (id
>= shrinker_nr_max
) {
366 if (expand_shrinker_info(id
)) {
367 idr_remove(&shrinker_idr
, id
);
374 up_write(&shrinker_rwsem
);
378 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
380 int id
= shrinker
->id
;
384 lockdep_assert_held(&shrinker_rwsem
);
386 idr_remove(&shrinker_idr
, id
);
389 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
390 struct mem_cgroup
*memcg
)
392 struct shrinker_info
*info
;
394 info
= shrinker_info_protected(memcg
, nid
);
395 return atomic_long_xchg(&info
->nr_deferred
[shrinker
->id
], 0);
398 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
399 struct mem_cgroup
*memcg
)
401 struct shrinker_info
*info
;
403 info
= shrinker_info_protected(memcg
, nid
);
404 return atomic_long_add_return(nr
, &info
->nr_deferred
[shrinker
->id
]);
407 void reparent_shrinker_deferred(struct mem_cgroup
*memcg
)
411 struct mem_cgroup
*parent
;
412 struct shrinker_info
*child_info
, *parent_info
;
414 parent
= parent_mem_cgroup(memcg
);
416 parent
= root_mem_cgroup
;
418 /* Prevent from concurrent shrinker_info expand */
419 down_read(&shrinker_rwsem
);
421 child_info
= shrinker_info_protected(memcg
, nid
);
422 parent_info
= shrinker_info_protected(parent
, nid
);
423 for (i
= 0; i
< shrinker_nr_max
; i
++) {
424 nr
= atomic_long_read(&child_info
->nr_deferred
[i
]);
425 atomic_long_add(nr
, &parent_info
->nr_deferred
[i
]);
428 up_read(&shrinker_rwsem
);
431 static bool cgroup_reclaim(struct scan_control
*sc
)
433 return sc
->target_mem_cgroup
;
437 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
438 * @sc: scan_control in question
440 * The normal page dirty throttling mechanism in balance_dirty_pages() is
441 * completely broken with the legacy memcg and direct stalling in
442 * shrink_page_list() is used for throttling instead, which lacks all the
443 * niceties such as fairness, adaptive pausing, bandwidth proportional
444 * allocation and configurability.
446 * This function tests whether the vmscan currently in progress can assume
447 * that the normal dirty throttling mechanism is operational.
449 static bool writeback_throttling_sane(struct scan_control
*sc
)
451 if (!cgroup_reclaim(sc
))
453 #ifdef CONFIG_CGROUP_WRITEBACK
454 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
460 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
465 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
469 static long xchg_nr_deferred_memcg(int nid
, struct shrinker
*shrinker
,
470 struct mem_cgroup
*memcg
)
475 static long add_nr_deferred_memcg(long nr
, int nid
, struct shrinker
*shrinker
,
476 struct mem_cgroup
*memcg
)
481 static bool cgroup_reclaim(struct scan_control
*sc
)
486 static bool writeback_throttling_sane(struct scan_control
*sc
)
492 static long xchg_nr_deferred(struct shrinker
*shrinker
,
493 struct shrink_control
*sc
)
497 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
501 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
502 return xchg_nr_deferred_memcg(nid
, shrinker
,
505 return atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
509 static long add_nr_deferred(long nr
, struct shrinker
*shrinker
,
510 struct shrink_control
*sc
)
514 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
518 (shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
519 return add_nr_deferred_memcg(nr
, nid
, shrinker
,
522 return atomic_long_add_return(nr
, &shrinker
->nr_deferred
[nid
]);
525 static bool can_demote(int nid
, struct scan_control
*sc
)
527 if (!numa_demotion_enabled
)
532 /* It is pointless to do demotion in memcg reclaim */
533 if (cgroup_reclaim(sc
))
536 if (next_demotion_node(nid
) == NUMA_NO_NODE
)
542 static inline bool can_reclaim_anon_pages(struct mem_cgroup
*memcg
,
544 struct scan_control
*sc
)
548 * For non-memcg reclaim, is there
549 * space in any swap device?
551 if (get_nr_swap_pages() > 0)
554 /* Is the memcg below its swap limit? */
555 if (mem_cgroup_get_nr_swap_pages(memcg
) > 0)
560 * The page can not be swapped.
562 * Can it be reclaimed from this node via demotion?
564 return can_demote(nid
, sc
);
568 * This misses isolated pages which are not accounted for to save counters.
569 * As the data only determines if reclaim or compaction continues, it is
570 * not expected that isolated pages will be a dominating factor.
572 unsigned long zone_reclaimable_pages(struct zone
*zone
)
576 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
577 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
578 if (can_reclaim_anon_pages(NULL
, zone_to_nid(zone
), NULL
))
579 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
580 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
586 * lruvec_lru_size - Returns the number of pages on the given LRU list.
587 * @lruvec: lru vector
589 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
591 static unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
594 unsigned long size
= 0;
597 for (zid
= 0; zid
<= zone_idx
&& zid
< MAX_NR_ZONES
; zid
++) {
598 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
600 if (!managed_zone(zone
))
603 if (!mem_cgroup_disabled())
604 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
606 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
612 * Add a shrinker callback to be called from the vm.
614 int prealloc_shrinker(struct shrinker
*shrinker
)
619 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
620 err
= prealloc_memcg_shrinker(shrinker
);
624 shrinker
->flags
&= ~SHRINKER_MEMCG_AWARE
;
627 size
= sizeof(*shrinker
->nr_deferred
);
628 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
631 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
632 if (!shrinker
->nr_deferred
)
638 void free_prealloced_shrinker(struct shrinker
*shrinker
)
640 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
641 down_write(&shrinker_rwsem
);
642 unregister_memcg_shrinker(shrinker
);
643 up_write(&shrinker_rwsem
);
647 kfree(shrinker
->nr_deferred
);
648 shrinker
->nr_deferred
= NULL
;
651 void register_shrinker_prepared(struct shrinker
*shrinker
)
653 down_write(&shrinker_rwsem
);
654 list_add_tail(&shrinker
->list
, &shrinker_list
);
655 shrinker
->flags
|= SHRINKER_REGISTERED
;
656 up_write(&shrinker_rwsem
);
659 int register_shrinker(struct shrinker
*shrinker
)
661 int err
= prealloc_shrinker(shrinker
);
665 register_shrinker_prepared(shrinker
);
668 EXPORT_SYMBOL(register_shrinker
);
673 void unregister_shrinker(struct shrinker
*shrinker
)
675 if (!(shrinker
->flags
& SHRINKER_REGISTERED
))
678 down_write(&shrinker_rwsem
);
679 list_del(&shrinker
->list
);
680 shrinker
->flags
&= ~SHRINKER_REGISTERED
;
681 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
682 unregister_memcg_shrinker(shrinker
);
683 up_write(&shrinker_rwsem
);
685 kfree(shrinker
->nr_deferred
);
686 shrinker
->nr_deferred
= NULL
;
688 EXPORT_SYMBOL(unregister_shrinker
);
691 * synchronize_shrinkers - Wait for all running shrinkers to complete.
693 * This is equivalent to calling unregister_shrink() and register_shrinker(),
694 * but atomically and with less overhead. This is useful to guarantee that all
695 * shrinker invocations have seen an update, before freeing memory, similar to
698 void synchronize_shrinkers(void)
700 down_write(&shrinker_rwsem
);
701 up_write(&shrinker_rwsem
);
703 EXPORT_SYMBOL(synchronize_shrinkers
);
705 #define SHRINK_BATCH 128
707 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
708 struct shrinker
*shrinker
, int priority
)
710 unsigned long freed
= 0;
711 unsigned long long delta
;
716 long batch_size
= shrinker
->batch
? shrinker
->batch
718 long scanned
= 0, next_deferred
;
720 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
721 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
725 * copy the current shrinker scan count into a local variable
726 * and zero it so that other concurrent shrinker invocations
727 * don't also do this scanning work.
729 nr
= xchg_nr_deferred(shrinker
, shrinkctl
);
731 if (shrinker
->seeks
) {
732 delta
= freeable
>> priority
;
734 do_div(delta
, shrinker
->seeks
);
737 * These objects don't require any IO to create. Trim
738 * them aggressively under memory pressure to keep
739 * them from causing refetches in the IO caches.
741 delta
= freeable
/ 2;
744 total_scan
= nr
>> priority
;
746 total_scan
= min(total_scan
, (2 * freeable
));
748 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
749 freeable
, delta
, total_scan
, priority
);
752 * Normally, we should not scan less than batch_size objects in one
753 * pass to avoid too frequent shrinker calls, but if the slab has less
754 * than batch_size objects in total and we are really tight on memory,
755 * we will try to reclaim all available objects, otherwise we can end
756 * up failing allocations although there are plenty of reclaimable
757 * objects spread over several slabs with usage less than the
760 * We detect the "tight on memory" situations by looking at the total
761 * number of objects we want to scan (total_scan). If it is greater
762 * than the total number of objects on slab (freeable), we must be
763 * scanning at high prio and therefore should try to reclaim as much as
766 while (total_scan
>= batch_size
||
767 total_scan
>= freeable
) {
769 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
771 shrinkctl
->nr_to_scan
= nr_to_scan
;
772 shrinkctl
->nr_scanned
= nr_to_scan
;
773 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
774 if (ret
== SHRINK_STOP
)
778 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
779 total_scan
-= shrinkctl
->nr_scanned
;
780 scanned
+= shrinkctl
->nr_scanned
;
786 * The deferred work is increased by any new work (delta) that wasn't
787 * done, decreased by old deferred work that was done now.
789 * And it is capped to two times of the freeable items.
791 next_deferred
= max_t(long, (nr
+ delta
- scanned
), 0);
792 next_deferred
= min(next_deferred
, (2 * freeable
));
795 * move the unused scan count back into the shrinker in a
796 * manner that handles concurrent updates.
798 new_nr
= add_nr_deferred(next_deferred
, shrinker
, shrinkctl
);
800 trace_mm_shrink_slab_end(shrinker
, shrinkctl
->nid
, freed
, nr
, new_nr
, total_scan
);
805 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
806 struct mem_cgroup
*memcg
, int priority
)
808 struct shrinker_info
*info
;
809 unsigned long ret
, freed
= 0;
812 if (!mem_cgroup_online(memcg
))
815 if (!down_read_trylock(&shrinker_rwsem
))
818 info
= shrinker_info_protected(memcg
, nid
);
822 for_each_set_bit(i
, info
->map
, shrinker_nr_max
) {
823 struct shrink_control sc
= {
824 .gfp_mask
= gfp_mask
,
828 struct shrinker
*shrinker
;
830 shrinker
= idr_find(&shrinker_idr
, i
);
831 if (unlikely(!shrinker
|| !(shrinker
->flags
& SHRINKER_REGISTERED
))) {
833 clear_bit(i
, info
->map
);
837 /* Call non-slab shrinkers even though kmem is disabled */
838 if (!memcg_kmem_enabled() &&
839 !(shrinker
->flags
& SHRINKER_NONSLAB
))
842 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
843 if (ret
== SHRINK_EMPTY
) {
844 clear_bit(i
, info
->map
);
846 * After the shrinker reported that it had no objects to
847 * free, but before we cleared the corresponding bit in
848 * the memcg shrinker map, a new object might have been
849 * added. To make sure, we have the bit set in this
850 * case, we invoke the shrinker one more time and reset
851 * the bit if it reports that it is not empty anymore.
852 * The memory barrier here pairs with the barrier in
853 * set_shrinker_bit():
855 * list_lru_add() shrink_slab_memcg()
856 * list_add_tail() clear_bit()
858 * set_bit() do_shrink_slab()
860 smp_mb__after_atomic();
861 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
862 if (ret
== SHRINK_EMPTY
)
865 set_shrinker_bit(memcg
, nid
, i
);
869 if (rwsem_is_contended(&shrinker_rwsem
)) {
875 up_read(&shrinker_rwsem
);
878 #else /* CONFIG_MEMCG */
879 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
880 struct mem_cgroup
*memcg
, int priority
)
884 #endif /* CONFIG_MEMCG */
887 * shrink_slab - shrink slab caches
888 * @gfp_mask: allocation context
889 * @nid: node whose slab caches to target
890 * @memcg: memory cgroup whose slab caches to target
891 * @priority: the reclaim priority
893 * Call the shrink functions to age shrinkable caches.
895 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
896 * unaware shrinkers will receive a node id of 0 instead.
898 * @memcg specifies the memory cgroup to target. Unaware shrinkers
899 * are called only if it is the root cgroup.
901 * @priority is sc->priority, we take the number of objects and >> by priority
902 * in order to get the scan target.
904 * Returns the number of reclaimed slab objects.
906 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
907 struct mem_cgroup
*memcg
,
910 unsigned long ret
, freed
= 0;
911 struct shrinker
*shrinker
;
914 * The root memcg might be allocated even though memcg is disabled
915 * via "cgroup_disable=memory" boot parameter. This could make
916 * mem_cgroup_is_root() return false, then just run memcg slab
917 * shrink, but skip global shrink. This may result in premature
920 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
921 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
923 if (!down_read_trylock(&shrinker_rwsem
))
926 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
927 struct shrink_control sc
= {
928 .gfp_mask
= gfp_mask
,
933 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
934 if (ret
== SHRINK_EMPTY
)
938 * Bail out if someone want to register a new shrinker to
939 * prevent the registration from being stalled for long periods
940 * by parallel ongoing shrinking.
942 if (rwsem_is_contended(&shrinker_rwsem
)) {
948 up_read(&shrinker_rwsem
);
954 static void drop_slab_node(int nid
)
960 struct mem_cgroup
*memcg
= NULL
;
962 if (fatal_signal_pending(current
))
966 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
968 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
969 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
970 } while ((freed
>> shift
++) > 1);
977 for_each_online_node(nid
)
981 static inline int is_page_cache_freeable(struct page
*page
)
984 * A freeable page cache page is referenced only by the caller
985 * that isolated the page, the page cache and optional buffer
986 * heads at page->private.
988 int page_cache_pins
= thp_nr_pages(page
);
989 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
993 * We detected a synchronous write error writing a page out. Probably
994 * -ENOSPC. We need to propagate that into the address_space for a subsequent
995 * fsync(), msync() or close().
997 * The tricky part is that after writepage we cannot touch the mapping: nothing
998 * prevents it from being freed up. But we have a ref on the page and once
999 * that page is locked, the mapping is pinned.
1001 * We're allowed to run sleeping lock_page() here because we know the caller has
1004 static void handle_write_error(struct address_space
*mapping
,
1005 struct page
*page
, int error
)
1008 if (page_mapping(page
) == mapping
)
1009 mapping_set_error(mapping
, error
);
1013 static bool skip_throttle_noprogress(pg_data_t
*pgdat
)
1015 int reclaimable
= 0, write_pending
= 0;
1019 * If kswapd is disabled, reschedule if necessary but do not
1020 * throttle as the system is likely near OOM.
1022 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
1026 * If there are a lot of dirty/writeback pages then do not
1027 * throttle as throttling will occur when the pages cycle
1028 * towards the end of the LRU if still under writeback.
1030 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
1031 struct zone
*zone
= pgdat
->node_zones
+ i
;
1033 if (!populated_zone(zone
))
1036 reclaimable
+= zone_reclaimable_pages(zone
);
1037 write_pending
+= zone_page_state_snapshot(zone
,
1038 NR_ZONE_WRITE_PENDING
);
1040 if (2 * write_pending
<= reclaimable
)
1046 void reclaim_throttle(pg_data_t
*pgdat
, enum vmscan_throttle_state reason
)
1048 wait_queue_head_t
*wqh
= &pgdat
->reclaim_wait
[reason
];
1053 * Do not throttle IO workers, kthreads other than kswapd or
1054 * workqueues. They may be required for reclaim to make
1055 * forward progress (e.g. journalling workqueues or kthreads).
1057 if (!current_is_kswapd() &&
1058 current
->flags
& (PF_IO_WORKER
|PF_KTHREAD
)) {
1064 * These figures are pulled out of thin air.
1065 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1066 * parallel reclaimers which is a short-lived event so the timeout is
1067 * short. Failing to make progress or waiting on writeback are
1068 * potentially long-lived events so use a longer timeout. This is shaky
1069 * logic as a failure to make progress could be due to anything from
1070 * writeback to a slow device to excessive references pages at the tail
1071 * of the inactive LRU.
1074 case VMSCAN_THROTTLE_WRITEBACK
:
1077 if (atomic_inc_return(&pgdat
->nr_writeback_throttled
) == 1) {
1078 WRITE_ONCE(pgdat
->nr_reclaim_start
,
1079 node_page_state(pgdat
, NR_THROTTLED_WRITTEN
));
1083 case VMSCAN_THROTTLE_CONGESTED
:
1085 case VMSCAN_THROTTLE_NOPROGRESS
:
1086 if (skip_throttle_noprogress(pgdat
)) {
1094 case VMSCAN_THROTTLE_ISOLATED
:
1103 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1104 ret
= schedule_timeout(timeout
);
1105 finish_wait(wqh
, &wait
);
1107 if (reason
== VMSCAN_THROTTLE_WRITEBACK
)
1108 atomic_dec(&pgdat
->nr_writeback_throttled
);
1110 trace_mm_vmscan_throttled(pgdat
->node_id
, jiffies_to_usecs(timeout
),
1111 jiffies_to_usecs(timeout
- ret
),
1116 * Account for pages written if tasks are throttled waiting on dirty
1117 * pages to clean. If enough pages have been cleaned since throttling
1118 * started then wakeup the throttled tasks.
1120 void __acct_reclaim_writeback(pg_data_t
*pgdat
, struct folio
*folio
,
1123 unsigned long nr_written
;
1125 node_stat_add_folio(folio
, NR_THROTTLED_WRITTEN
);
1128 * This is an inaccurate read as the per-cpu deltas may not
1129 * be synchronised. However, given that the system is
1130 * writeback throttled, it is not worth taking the penalty
1131 * of getting an accurate count. At worst, the throttle
1132 * timeout guarantees forward progress.
1134 nr_written
= node_page_state(pgdat
, NR_THROTTLED_WRITTEN
) -
1135 READ_ONCE(pgdat
->nr_reclaim_start
);
1137 if (nr_written
> SWAP_CLUSTER_MAX
* nr_throttled
)
1138 wake_up(&pgdat
->reclaim_wait
[VMSCAN_THROTTLE_WRITEBACK
]);
1141 /* possible outcome of pageout() */
1143 /* failed to write page out, page is locked */
1145 /* move page to the active list, page is locked */
1147 /* page has been sent to the disk successfully, page is unlocked */
1149 /* page is clean and locked */
1154 * pageout is called by shrink_page_list() for each dirty page.
1155 * Calls ->writepage().
1157 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
1160 * If the page is dirty, only perform writeback if that write
1161 * will be non-blocking. To prevent this allocation from being
1162 * stalled by pagecache activity. But note that there may be
1163 * stalls if we need to run get_block(). We could test
1164 * PagePrivate for that.
1166 * If this process is currently in __generic_file_write_iter() against
1167 * this page's queue, we can perform writeback even if that
1170 * If the page is swapcache, write it back even if that would
1171 * block, for some throttling. This happens by accident, because
1172 * swap_backing_dev_info is bust: it doesn't reflect the
1173 * congestion state of the swapdevs. Easy to fix, if needed.
1175 if (!is_page_cache_freeable(page
))
1179 * Some data journaling orphaned pages can have
1180 * page->mapping == NULL while being dirty with clean buffers.
1182 if (page_has_private(page
)) {
1183 if (try_to_free_buffers(page
)) {
1184 ClearPageDirty(page
);
1185 pr_info("%s: orphaned page\n", __func__
);
1191 if (mapping
->a_ops
->writepage
== NULL
)
1192 return PAGE_ACTIVATE
;
1194 if (clear_page_dirty_for_io(page
)) {
1196 struct writeback_control wbc
= {
1197 .sync_mode
= WB_SYNC_NONE
,
1198 .nr_to_write
= SWAP_CLUSTER_MAX
,
1200 .range_end
= LLONG_MAX
,
1204 SetPageReclaim(page
);
1205 res
= mapping
->a_ops
->writepage(page
, &wbc
);
1207 handle_write_error(mapping
, page
, res
);
1208 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
1209 ClearPageReclaim(page
);
1210 return PAGE_ACTIVATE
;
1213 if (!PageWriteback(page
)) {
1214 /* synchronous write or broken a_ops? */
1215 ClearPageReclaim(page
);
1217 trace_mm_vmscan_writepage(page
);
1218 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
1219 return PAGE_SUCCESS
;
1226 * Same as remove_mapping, but if the page is removed from the mapping, it
1227 * gets returned with a refcount of 0.
1229 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
1230 bool reclaimed
, struct mem_cgroup
*target_memcg
)
1233 void *shadow
= NULL
;
1235 BUG_ON(!PageLocked(page
));
1236 BUG_ON(mapping
!= page_mapping(page
));
1238 if (!PageSwapCache(page
))
1239 spin_lock(&mapping
->host
->i_lock
);
1240 xa_lock_irq(&mapping
->i_pages
);
1242 * The non racy check for a busy page.
1244 * Must be careful with the order of the tests. When someone has
1245 * a ref to the page, it may be possible that they dirty it then
1246 * drop the reference. So if PageDirty is tested before page_count
1247 * here, then the following race may occur:
1249 * get_user_pages(&page);
1250 * [user mapping goes away]
1252 * !PageDirty(page) [good]
1253 * SetPageDirty(page);
1255 * !page_count(page) [good, discard it]
1257 * [oops, our write_to data is lost]
1259 * Reversing the order of the tests ensures such a situation cannot
1260 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1261 * load is not satisfied before that of page->_refcount.
1263 * Note that if SetPageDirty is always performed via set_page_dirty,
1264 * and thus under the i_pages lock, then this ordering is not required.
1266 refcount
= 1 + compound_nr(page
);
1267 if (!page_ref_freeze(page
, refcount
))
1269 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1270 if (unlikely(PageDirty(page
))) {
1271 page_ref_unfreeze(page
, refcount
);
1275 if (PageSwapCache(page
)) {
1276 swp_entry_t swap
= { .val
= page_private(page
) };
1277 mem_cgroup_swapout(page
, swap
);
1278 if (reclaimed
&& !mapping_exiting(mapping
))
1279 shadow
= workingset_eviction(page
, target_memcg
);
1280 __delete_from_swap_cache(page
, swap
, shadow
);
1281 xa_unlock_irq(&mapping
->i_pages
);
1282 put_swap_page(page
, swap
);
1284 void (*freepage
)(struct page
*);
1286 freepage
= mapping
->a_ops
->freepage
;
1288 * Remember a shadow entry for reclaimed file cache in
1289 * order to detect refaults, thus thrashing, later on.
1291 * But don't store shadows in an address space that is
1292 * already exiting. This is not just an optimization,
1293 * inode reclaim needs to empty out the radix tree or
1294 * the nodes are lost. Don't plant shadows behind its
1297 * We also don't store shadows for DAX mappings because the
1298 * only page cache pages found in these are zero pages
1299 * covering holes, and because we don't want to mix DAX
1300 * exceptional entries and shadow exceptional entries in the
1301 * same address_space.
1303 if (reclaimed
&& page_is_file_lru(page
) &&
1304 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
1305 shadow
= workingset_eviction(page
, target_memcg
);
1306 __delete_from_page_cache(page
, shadow
);
1307 xa_unlock_irq(&mapping
->i_pages
);
1308 if (mapping_shrinkable(mapping
))
1309 inode_add_lru(mapping
->host
);
1310 spin_unlock(&mapping
->host
->i_lock
);
1312 if (freepage
!= NULL
)
1319 xa_unlock_irq(&mapping
->i_pages
);
1320 if (!PageSwapCache(page
))
1321 spin_unlock(&mapping
->host
->i_lock
);
1326 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1327 * someone else has a ref on the page, abort and return 0. If it was
1328 * successfully detached, return 1. Assumes the caller has a single ref on
1331 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
1333 if (__remove_mapping(mapping
, page
, false, NULL
)) {
1335 * Unfreezing the refcount with 1 rather than 2 effectively
1336 * drops the pagecache ref for us without requiring another
1339 page_ref_unfreeze(page
, 1);
1346 * putback_lru_page - put previously isolated page onto appropriate LRU list
1347 * @page: page to be put back to appropriate lru list
1349 * Add previously isolated @page to appropriate LRU list.
1350 * Page may still be unevictable for other reasons.
1352 * lru_lock must not be held, interrupts must be enabled.
1354 void putback_lru_page(struct page
*page
)
1356 lru_cache_add(page
);
1357 put_page(page
); /* drop ref from isolate */
1360 enum page_references
{
1362 PAGEREF_RECLAIM_CLEAN
,
1367 static enum page_references
page_check_references(struct page
*page
,
1368 struct scan_control
*sc
)
1370 int referenced_ptes
, referenced_page
;
1371 unsigned long vm_flags
;
1373 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1375 referenced_page
= TestClearPageReferenced(page
);
1378 * Mlock lost the isolation race with us. Let try_to_unmap()
1379 * move the page to the unevictable list.
1381 if (vm_flags
& VM_LOCKED
)
1382 return PAGEREF_RECLAIM
;
1384 if (referenced_ptes
) {
1386 * All mapped pages start out with page table
1387 * references from the instantiating fault, so we need
1388 * to look twice if a mapped file page is used more
1391 * Mark it and spare it for another trip around the
1392 * inactive list. Another page table reference will
1393 * lead to its activation.
1395 * Note: the mark is set for activated pages as well
1396 * so that recently deactivated but used pages are
1397 * quickly recovered.
1399 SetPageReferenced(page
);
1401 if (referenced_page
|| referenced_ptes
> 1)
1402 return PAGEREF_ACTIVATE
;
1405 * Activate file-backed executable pages after first usage.
1407 if ((vm_flags
& VM_EXEC
) && !PageSwapBacked(page
))
1408 return PAGEREF_ACTIVATE
;
1410 return PAGEREF_KEEP
;
1413 /* Reclaim if clean, defer dirty pages to writeback */
1414 if (referenced_page
&& !PageSwapBacked(page
))
1415 return PAGEREF_RECLAIM_CLEAN
;
1417 return PAGEREF_RECLAIM
;
1420 /* Check if a page is dirty or under writeback */
1421 static void page_check_dirty_writeback(struct page
*page
,
1422 bool *dirty
, bool *writeback
)
1424 struct address_space
*mapping
;
1427 * Anonymous pages are not handled by flushers and must be written
1428 * from reclaim context. Do not stall reclaim based on them
1430 if (!page_is_file_lru(page
) ||
1431 (PageAnon(page
) && !PageSwapBacked(page
))) {
1437 /* By default assume that the page flags are accurate */
1438 *dirty
= PageDirty(page
);
1439 *writeback
= PageWriteback(page
);
1441 /* Verify dirty/writeback state if the filesystem supports it */
1442 if (!page_has_private(page
))
1445 mapping
= page_mapping(page
);
1446 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1447 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1450 static struct page
*alloc_demote_page(struct page
*page
, unsigned long node
)
1452 struct migration_target_control mtc
= {
1454 * Allocate from 'node', or fail quickly and quietly.
1455 * When this happens, 'page' will likely just be discarded
1456 * instead of migrated.
1458 .gfp_mask
= (GFP_HIGHUSER_MOVABLE
& ~__GFP_RECLAIM
) |
1459 __GFP_THISNODE
| __GFP_NOWARN
|
1460 __GFP_NOMEMALLOC
| GFP_NOWAIT
,
1464 return alloc_migration_target(page
, (unsigned long)&mtc
);
1468 * Take pages on @demote_list and attempt to demote them to
1469 * another node. Pages which are not demoted are left on
1472 static unsigned int demote_page_list(struct list_head
*demote_pages
,
1473 struct pglist_data
*pgdat
)
1475 int target_nid
= next_demotion_node(pgdat
->node_id
);
1476 unsigned int nr_succeeded
;
1478 if (list_empty(demote_pages
))
1481 if (target_nid
== NUMA_NO_NODE
)
1484 /* Demotion ignores all cpuset and mempolicy settings */
1485 migrate_pages(demote_pages
, alloc_demote_page
, NULL
,
1486 target_nid
, MIGRATE_ASYNC
, MR_DEMOTION
,
1489 if (current_is_kswapd())
1490 __count_vm_events(PGDEMOTE_KSWAPD
, nr_succeeded
);
1492 __count_vm_events(PGDEMOTE_DIRECT
, nr_succeeded
);
1494 return nr_succeeded
;
1498 * shrink_page_list() returns the number of reclaimed pages
1500 static unsigned int shrink_page_list(struct list_head
*page_list
,
1501 struct pglist_data
*pgdat
,
1502 struct scan_control
*sc
,
1503 struct reclaim_stat
*stat
,
1504 bool ignore_references
)
1506 LIST_HEAD(ret_pages
);
1507 LIST_HEAD(free_pages
);
1508 LIST_HEAD(demote_pages
);
1509 unsigned int nr_reclaimed
= 0;
1510 unsigned int pgactivate
= 0;
1511 bool do_demote_pass
;
1513 memset(stat
, 0, sizeof(*stat
));
1515 do_demote_pass
= can_demote(pgdat
->node_id
, sc
);
1518 while (!list_empty(page_list
)) {
1519 struct address_space
*mapping
;
1521 enum page_references references
= PAGEREF_RECLAIM
;
1522 bool dirty
, writeback
, may_enter_fs
;
1523 unsigned int nr_pages
;
1527 page
= lru_to_page(page_list
);
1528 list_del(&page
->lru
);
1530 if (!trylock_page(page
))
1533 VM_BUG_ON_PAGE(PageActive(page
), page
);
1535 nr_pages
= compound_nr(page
);
1537 /* Account the number of base pages even though THP */
1538 sc
->nr_scanned
+= nr_pages
;
1540 if (unlikely(!page_evictable(page
)))
1541 goto activate_locked
;
1543 if (!sc
->may_unmap
&& page_mapped(page
))
1546 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1547 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1550 * The number of dirty pages determines if a node is marked
1551 * reclaim_congested. kswapd will stall and start writing
1552 * pages if the tail of the LRU is all dirty unqueued pages.
1554 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1555 if (dirty
|| writeback
)
1558 if (dirty
&& !writeback
)
1559 stat
->nr_unqueued_dirty
++;
1562 * Treat this page as congested if the underlying BDI is or if
1563 * pages are cycling through the LRU so quickly that the
1564 * pages marked for immediate reclaim are making it to the
1565 * end of the LRU a second time.
1567 mapping
= page_mapping(page
);
1568 if (writeback
&& PageReclaim(page
))
1569 stat
->nr_congested
++;
1572 * If a page at the tail of the LRU is under writeback, there
1573 * are three cases to consider.
1575 * 1) If reclaim is encountering an excessive number of pages
1576 * under writeback and this page is both under writeback and
1577 * PageReclaim then it indicates that pages are being queued
1578 * for IO but are being recycled through the LRU before the
1579 * IO can complete. Waiting on the page itself risks an
1580 * indefinite stall if it is impossible to writeback the
1581 * page due to IO error or disconnected storage so instead
1582 * note that the LRU is being scanned too quickly and the
1583 * caller can stall after page list has been processed.
1585 * 2) Global or new memcg reclaim encounters a page that is
1586 * not marked for immediate reclaim, or the caller does not
1587 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1588 * not to fs). In this case mark the page for immediate
1589 * reclaim and continue scanning.
1591 * Require may_enter_fs because we would wait on fs, which
1592 * may not have submitted IO yet. And the loop driver might
1593 * enter reclaim, and deadlock if it waits on a page for
1594 * which it is needed to do the write (loop masks off
1595 * __GFP_IO|__GFP_FS for this reason); but more thought
1596 * would probably show more reasons.
1598 * 3) Legacy memcg encounters a page that is already marked
1599 * PageReclaim. memcg does not have any dirty pages
1600 * throttling so we could easily OOM just because too many
1601 * pages are in writeback and there is nothing else to
1602 * reclaim. Wait for the writeback to complete.
1604 * In cases 1) and 2) we activate the pages to get them out of
1605 * the way while we continue scanning for clean pages on the
1606 * inactive list and refilling from the active list. The
1607 * observation here is that waiting for disk writes is more
1608 * expensive than potentially causing reloads down the line.
1609 * Since they're marked for immediate reclaim, they won't put
1610 * memory pressure on the cache working set any longer than it
1611 * takes to write them to disk.
1613 if (PageWriteback(page
)) {
1615 if (current_is_kswapd() &&
1616 PageReclaim(page
) &&
1617 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1618 stat
->nr_immediate
++;
1619 goto activate_locked
;
1622 } else if (writeback_throttling_sane(sc
) ||
1623 !PageReclaim(page
) || !may_enter_fs
) {
1625 * This is slightly racy - end_page_writeback()
1626 * might have just cleared PageReclaim, then
1627 * setting PageReclaim here end up interpreted
1628 * as PageReadahead - but that does not matter
1629 * enough to care. What we do want is for this
1630 * page to have PageReclaim set next time memcg
1631 * reclaim reaches the tests above, so it will
1632 * then wait_on_page_writeback() to avoid OOM;
1633 * and it's also appropriate in global reclaim.
1635 SetPageReclaim(page
);
1636 stat
->nr_writeback
++;
1637 goto activate_locked
;
1642 wait_on_page_writeback(page
);
1643 /* then go back and try same page again */
1644 list_add_tail(&page
->lru
, page_list
);
1649 if (!ignore_references
)
1650 references
= page_check_references(page
, sc
);
1652 switch (references
) {
1653 case PAGEREF_ACTIVATE
:
1654 goto activate_locked
;
1656 stat
->nr_ref_keep
+= nr_pages
;
1658 case PAGEREF_RECLAIM
:
1659 case PAGEREF_RECLAIM_CLEAN
:
1660 ; /* try to reclaim the page below */
1664 * Before reclaiming the page, try to relocate
1665 * its contents to another node.
1667 if (do_demote_pass
&&
1668 (thp_migration_supported() || !PageTransHuge(page
))) {
1669 list_add(&page
->lru
, &demote_pages
);
1675 * Anonymous process memory has backing store?
1676 * Try to allocate it some swap space here.
1677 * Lazyfree page could be freed directly
1679 if (PageAnon(page
) && PageSwapBacked(page
)) {
1680 if (!PageSwapCache(page
)) {
1681 if (!(sc
->gfp_mask
& __GFP_IO
))
1683 if (page_maybe_dma_pinned(page
))
1685 if (PageTransHuge(page
)) {
1686 /* cannot split THP, skip it */
1687 if (!can_split_huge_page(page
, NULL
))
1688 goto activate_locked
;
1690 * Split pages without a PMD map right
1691 * away. Chances are some or all of the
1692 * tail pages can be freed without IO.
1694 if (!compound_mapcount(page
) &&
1695 split_huge_page_to_list(page
,
1697 goto activate_locked
;
1699 if (!add_to_swap(page
)) {
1700 if (!PageTransHuge(page
))
1701 goto activate_locked_split
;
1702 /* Fallback to swap normal pages */
1703 if (split_huge_page_to_list(page
,
1705 goto activate_locked
;
1706 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1707 count_vm_event(THP_SWPOUT_FALLBACK
);
1709 if (!add_to_swap(page
))
1710 goto activate_locked_split
;
1713 may_enter_fs
= true;
1715 /* Adding to swap updated mapping */
1716 mapping
= page_mapping(page
);
1718 } else if (unlikely(PageTransHuge(page
))) {
1719 /* Split file THP */
1720 if (split_huge_page_to_list(page
, page_list
))
1725 * THP may get split above, need minus tail pages and update
1726 * nr_pages to avoid accounting tail pages twice.
1728 * The tail pages that are added into swap cache successfully
1731 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1732 sc
->nr_scanned
-= (nr_pages
- 1);
1737 * The page is mapped into the page tables of one or more
1738 * processes. Try to unmap it here.
1740 if (page_mapped(page
)) {
1741 enum ttu_flags flags
= TTU_BATCH_FLUSH
;
1742 bool was_swapbacked
= PageSwapBacked(page
);
1744 if (unlikely(PageTransHuge(page
)))
1745 flags
|= TTU_SPLIT_HUGE_PMD
;
1747 try_to_unmap(page
, flags
);
1748 if (page_mapped(page
)) {
1749 stat
->nr_unmap_fail
+= nr_pages
;
1750 if (!was_swapbacked
&& PageSwapBacked(page
))
1751 stat
->nr_lazyfree_fail
+= nr_pages
;
1752 goto activate_locked
;
1756 if (PageDirty(page
)) {
1758 * Only kswapd can writeback filesystem pages
1759 * to avoid risk of stack overflow. But avoid
1760 * injecting inefficient single-page IO into
1761 * flusher writeback as much as possible: only
1762 * write pages when we've encountered many
1763 * dirty pages, and when we've already scanned
1764 * the rest of the LRU for clean pages and see
1765 * the same dirty pages again (PageReclaim).
1767 if (page_is_file_lru(page
) &&
1768 (!current_is_kswapd() || !PageReclaim(page
) ||
1769 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1771 * Immediately reclaim when written back.
1772 * Similar in principal to deactivate_page()
1773 * except we already have the page isolated
1774 * and know it's dirty
1776 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1777 SetPageReclaim(page
);
1779 goto activate_locked
;
1782 if (references
== PAGEREF_RECLAIM_CLEAN
)
1786 if (!sc
->may_writepage
)
1790 * Page is dirty. Flush the TLB if a writable entry
1791 * potentially exists to avoid CPU writes after IO
1792 * starts and then write it out here.
1794 try_to_unmap_flush_dirty();
1795 switch (pageout(page
, mapping
)) {
1799 goto activate_locked
;
1801 stat
->nr_pageout
+= thp_nr_pages(page
);
1803 if (PageWriteback(page
))
1805 if (PageDirty(page
))
1809 * A synchronous write - probably a ramdisk. Go
1810 * ahead and try to reclaim the page.
1812 if (!trylock_page(page
))
1814 if (PageDirty(page
) || PageWriteback(page
))
1816 mapping
= page_mapping(page
);
1819 ; /* try to free the page below */
1824 * If the page has buffers, try to free the buffer mappings
1825 * associated with this page. If we succeed we try to free
1828 * We do this even if the page is PageDirty().
1829 * try_to_release_page() does not perform I/O, but it is
1830 * possible for a page to have PageDirty set, but it is actually
1831 * clean (all its buffers are clean). This happens if the
1832 * buffers were written out directly, with submit_bh(). ext3
1833 * will do this, as well as the blockdev mapping.
1834 * try_to_release_page() will discover that cleanness and will
1835 * drop the buffers and mark the page clean - it can be freed.
1837 * Rarely, pages can have buffers and no ->mapping. These are
1838 * the pages which were not successfully invalidated in
1839 * truncate_cleanup_page(). We try to drop those buffers here
1840 * and if that worked, and the page is no longer mapped into
1841 * process address space (page_count == 1) it can be freed.
1842 * Otherwise, leave the page on the LRU so it is swappable.
1844 if (page_has_private(page
)) {
1845 if (!try_to_release_page(page
, sc
->gfp_mask
))
1846 goto activate_locked
;
1847 if (!mapping
&& page_count(page
) == 1) {
1849 if (put_page_testzero(page
))
1853 * rare race with speculative reference.
1854 * the speculative reference will free
1855 * this page shortly, so we may
1856 * increment nr_reclaimed here (and
1857 * leave it off the LRU).
1865 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1866 /* follow __remove_mapping for reference */
1867 if (!page_ref_freeze(page
, 1))
1870 * The page has only one reference left, which is
1871 * from the isolation. After the caller puts the
1872 * page back on lru and drops the reference, the
1873 * page will be freed anyway. It doesn't matter
1874 * which lru it goes. So we don't bother checking
1877 count_vm_event(PGLAZYFREED
);
1878 count_memcg_page_event(page
, PGLAZYFREED
);
1879 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true,
1880 sc
->target_mem_cgroup
))
1886 * THP may get swapped out in a whole, need account
1889 nr_reclaimed
+= nr_pages
;
1892 * Is there need to periodically free_page_list? It would
1893 * appear not as the counts should be low
1895 if (unlikely(PageTransHuge(page
)))
1896 destroy_compound_page(page
);
1898 list_add(&page
->lru
, &free_pages
);
1901 activate_locked_split
:
1903 * The tail pages that are failed to add into swap cache
1904 * reach here. Fixup nr_scanned and nr_pages.
1907 sc
->nr_scanned
-= (nr_pages
- 1);
1911 /* Not a candidate for swapping, so reclaim swap space. */
1912 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1914 try_to_free_swap(page
);
1915 VM_BUG_ON_PAGE(PageActive(page
), page
);
1916 if (!PageMlocked(page
)) {
1917 int type
= page_is_file_lru(page
);
1918 SetPageActive(page
);
1919 stat
->nr_activate
[type
] += nr_pages
;
1920 count_memcg_page_event(page
, PGACTIVATE
);
1925 list_add(&page
->lru
, &ret_pages
);
1926 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1928 /* 'page_list' is always empty here */
1930 /* Migrate pages selected for demotion */
1931 nr_reclaimed
+= demote_page_list(&demote_pages
, pgdat
);
1932 /* Pages that could not be demoted are still in @demote_pages */
1933 if (!list_empty(&demote_pages
)) {
1934 /* Pages which failed to demoted go back on @page_list for retry: */
1935 list_splice_init(&demote_pages
, page_list
);
1936 do_demote_pass
= false;
1940 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1942 mem_cgroup_uncharge_list(&free_pages
);
1943 try_to_unmap_flush();
1944 free_unref_page_list(&free_pages
);
1946 list_splice(&ret_pages
, page_list
);
1947 count_vm_events(PGACTIVATE
, pgactivate
);
1949 return nr_reclaimed
;
1952 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1953 struct list_head
*page_list
)
1955 struct scan_control sc
= {
1956 .gfp_mask
= GFP_KERNEL
,
1959 struct reclaim_stat stat
;
1960 unsigned int nr_reclaimed
;
1961 struct page
*page
, *next
;
1962 LIST_HEAD(clean_pages
);
1963 unsigned int noreclaim_flag
;
1965 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1966 if (!PageHuge(page
) && page_is_file_lru(page
) &&
1967 !PageDirty(page
) && !__PageMovable(page
) &&
1968 !PageUnevictable(page
)) {
1969 ClearPageActive(page
);
1970 list_move(&page
->lru
, &clean_pages
);
1975 * We should be safe here since we are only dealing with file pages and
1976 * we are not kswapd and therefore cannot write dirty file pages. But
1977 * call memalloc_noreclaim_save() anyway, just in case these conditions
1978 * change in the future.
1980 noreclaim_flag
= memalloc_noreclaim_save();
1981 nr_reclaimed
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1983 memalloc_noreclaim_restore(noreclaim_flag
);
1985 list_splice(&clean_pages
, page_list
);
1986 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1987 -(long)nr_reclaimed
);
1989 * Since lazyfree pages are isolated from file LRU from the beginning,
1990 * they will rotate back to anonymous LRU in the end if it failed to
1991 * discard so isolated count will be mismatched.
1992 * Compensate the isolated count for both LRU lists.
1994 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
1995 stat
.nr_lazyfree_fail
);
1996 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1997 -(long)stat
.nr_lazyfree_fail
);
1998 return nr_reclaimed
;
2002 * Attempt to remove the specified page from its LRU. Only take this page
2003 * if it is of the appropriate PageActive status. Pages which are being
2004 * freed elsewhere are also ignored.
2006 * page: page to consider
2007 * mode: one of the LRU isolation modes defined above
2009 * returns true on success, false on failure.
2011 bool __isolate_lru_page_prepare(struct page
*page
, isolate_mode_t mode
)
2013 /* Only take pages on the LRU. */
2017 /* Compaction should not handle unevictable pages but CMA can do so */
2018 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
2022 * To minimise LRU disruption, the caller can indicate that it only
2023 * wants to isolate pages it will be able to operate on without
2024 * blocking - clean pages for the most part.
2026 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
2027 * that it is possible to migrate without blocking
2029 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
2030 /* All the caller can do on PageWriteback is block */
2031 if (PageWriteback(page
))
2034 if (PageDirty(page
)) {
2035 struct address_space
*mapping
;
2039 * Only pages without mappings or that have a
2040 * ->migratepage callback are possible to migrate
2041 * without blocking. However, we can be racing with
2042 * truncation so it's necessary to lock the page
2043 * to stabilise the mapping as truncation holds
2044 * the page lock until after the page is removed
2045 * from the page cache.
2047 if (!trylock_page(page
))
2050 mapping
= page_mapping(page
);
2051 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
2058 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
2065 * Update LRU sizes after isolating pages. The LRU size updates must
2066 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2068 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
2069 enum lru_list lru
, unsigned long *nr_zone_taken
)
2073 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2074 if (!nr_zone_taken
[zid
])
2077 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
2083 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2085 * lruvec->lru_lock is heavily contended. Some of the functions that
2086 * shrink the lists perform better by taking out a batch of pages
2087 * and working on them outside the LRU lock.
2089 * For pagecache intensive workloads, this function is the hottest
2090 * spot in the kernel (apart from copy_*_user functions).
2092 * Lru_lock must be held before calling this function.
2094 * @nr_to_scan: The number of eligible pages to look through on the list.
2095 * @lruvec: The LRU vector to pull pages from.
2096 * @dst: The temp list to put pages on to.
2097 * @nr_scanned: The number of pages that were scanned.
2098 * @sc: The scan_control struct for this reclaim session
2099 * @lru: LRU list id for isolating
2101 * returns how many pages were moved onto *@dst.
2103 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
2104 struct lruvec
*lruvec
, struct list_head
*dst
,
2105 unsigned long *nr_scanned
, struct scan_control
*sc
,
2108 struct list_head
*src
= &lruvec
->lists
[lru
];
2109 unsigned long nr_taken
= 0;
2110 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
2111 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
2112 unsigned long skipped
= 0;
2113 unsigned long scan
, total_scan
, nr_pages
;
2114 LIST_HEAD(pages_skipped
);
2115 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
2119 while (scan
< nr_to_scan
&& !list_empty(src
)) {
2122 page
= lru_to_page(src
);
2123 prefetchw_prev_lru_page(page
, src
, flags
);
2125 nr_pages
= compound_nr(page
);
2126 total_scan
+= nr_pages
;
2128 if (page_zonenum(page
) > sc
->reclaim_idx
) {
2129 list_move(&page
->lru
, &pages_skipped
);
2130 nr_skipped
[page_zonenum(page
)] += nr_pages
;
2135 * Do not count skipped pages because that makes the function
2136 * return with no isolated pages if the LRU mostly contains
2137 * ineligible pages. This causes the VM to not reclaim any
2138 * pages, triggering a premature OOM.
2140 * Account all tail pages of THP. This would not cause
2141 * premature OOM since __isolate_lru_page() returns -EBUSY
2142 * only when the page is being freed somewhere else.
2145 if (!__isolate_lru_page_prepare(page
, mode
)) {
2146 /* It is being freed elsewhere */
2147 list_move(&page
->lru
, src
);
2151 * Be careful not to clear PageLRU until after we're
2152 * sure the page is not being freed elsewhere -- the
2153 * page release code relies on it.
2155 if (unlikely(!get_page_unless_zero(page
))) {
2156 list_move(&page
->lru
, src
);
2160 if (!TestClearPageLRU(page
)) {
2161 /* Another thread is already isolating this page */
2163 list_move(&page
->lru
, src
);
2167 nr_taken
+= nr_pages
;
2168 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
2169 list_move(&page
->lru
, dst
);
2173 * Splice any skipped pages to the start of the LRU list. Note that
2174 * this disrupts the LRU order when reclaiming for lower zones but
2175 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2176 * scanning would soon rescan the same pages to skip and put the
2177 * system at risk of premature OOM.
2179 if (!list_empty(&pages_skipped
)) {
2182 list_splice(&pages_skipped
, src
);
2183 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
2184 if (!nr_skipped
[zid
])
2187 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
2188 skipped
+= nr_skipped
[zid
];
2191 *nr_scanned
= total_scan
;
2192 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
2193 total_scan
, skipped
, nr_taken
, mode
, lru
);
2194 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
2199 * isolate_lru_page - tries to isolate a page from its LRU list
2200 * @page: page to isolate from its LRU list
2202 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
2203 * vmstat statistic corresponding to whatever LRU list the page was on.
2205 * Returns 0 if the page was removed from an LRU list.
2206 * Returns -EBUSY if the page was not on an LRU list.
2208 * The returned page will have PageLRU() cleared. If it was found on
2209 * the active list, it will have PageActive set. If it was found on
2210 * the unevictable list, it will have the PageUnevictable bit set. That flag
2211 * may need to be cleared by the caller before letting the page go.
2213 * The vmstat statistic corresponding to the list on which the page was
2214 * found will be decremented.
2218 * (1) Must be called with an elevated refcount on the page. This is a
2219 * fundamental difference from isolate_lru_pages (which is called
2220 * without a stable reference).
2221 * (2) the lru_lock must not be held.
2222 * (3) interrupts must be enabled.
2224 int isolate_lru_page(struct page
*page
)
2226 struct folio
*folio
= page_folio(page
);
2229 VM_BUG_ON_PAGE(!page_count(page
), page
);
2230 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
2232 if (TestClearPageLRU(page
)) {
2233 struct lruvec
*lruvec
;
2236 lruvec
= folio_lruvec_lock_irq(folio
);
2237 del_page_from_lru_list(page
, lruvec
);
2238 unlock_page_lruvec_irq(lruvec
);
2246 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2247 * then get rescheduled. When there are massive number of tasks doing page
2248 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2249 * the LRU list will go small and be scanned faster than necessary, leading to
2250 * unnecessary swapping, thrashing and OOM.
2252 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
2253 struct scan_control
*sc
)
2255 unsigned long inactive
, isolated
;
2258 if (current_is_kswapd())
2261 if (!writeback_throttling_sane(sc
))
2265 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2266 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
2268 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2269 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
2273 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2274 * won't get blocked by normal direct-reclaimers, forming a circular
2277 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
2280 too_many
= isolated
> inactive
;
2282 /* Wake up tasks throttled due to too_many_isolated. */
2284 wake_throttle_isolated(pgdat
);
2290 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2291 * On return, @list is reused as a list of pages to be freed by the caller.
2293 * Returns the number of pages moved to the given lruvec.
2295 static unsigned int move_pages_to_lru(struct lruvec
*lruvec
,
2296 struct list_head
*list
)
2298 int nr_pages
, nr_moved
= 0;
2299 LIST_HEAD(pages_to_free
);
2302 while (!list_empty(list
)) {
2303 page
= lru_to_page(list
);
2304 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2305 list_del(&page
->lru
);
2306 if (unlikely(!page_evictable(page
))) {
2307 spin_unlock_irq(&lruvec
->lru_lock
);
2308 putback_lru_page(page
);
2309 spin_lock_irq(&lruvec
->lru_lock
);
2314 * The SetPageLRU needs to be kept here for list integrity.
2316 * #0 move_pages_to_lru #1 release_pages
2317 * if !put_page_testzero
2318 * if (put_page_testzero())
2319 * !PageLRU //skip lru_lock
2321 * list_add(&page->lru,)
2322 * list_add(&page->lru,)
2326 if (unlikely(put_page_testzero(page
))) {
2327 __clear_page_lru_flags(page
);
2329 if (unlikely(PageCompound(page
))) {
2330 spin_unlock_irq(&lruvec
->lru_lock
);
2331 destroy_compound_page(page
);
2332 spin_lock_irq(&lruvec
->lru_lock
);
2334 list_add(&page
->lru
, &pages_to_free
);
2340 * All pages were isolated from the same lruvec (and isolation
2341 * inhibits memcg migration).
2343 VM_BUG_ON_PAGE(!folio_matches_lruvec(page_folio(page
), lruvec
), page
);
2344 add_page_to_lru_list(page
, lruvec
);
2345 nr_pages
= thp_nr_pages(page
);
2346 nr_moved
+= nr_pages
;
2347 if (PageActive(page
))
2348 workingset_age_nonresident(lruvec
, nr_pages
);
2352 * To save our caller's stack, now use input list for pages to free.
2354 list_splice(&pages_to_free
, list
);
2360 * If a kernel thread (such as nfsd for loop-back mounts) services
2361 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2362 * In that case we should only throttle if the backing device it is
2363 * writing to is congested. In other cases it is safe to throttle.
2365 static int current_may_throttle(void)
2367 return !(current
->flags
& PF_LOCAL_THROTTLE
);
2371 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2372 * of reclaimed pages
2374 static unsigned long
2375 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
2376 struct scan_control
*sc
, enum lru_list lru
)
2378 LIST_HEAD(page_list
);
2379 unsigned long nr_scanned
;
2380 unsigned int nr_reclaimed
= 0;
2381 unsigned long nr_taken
;
2382 struct reclaim_stat stat
;
2383 bool file
= is_file_lru(lru
);
2384 enum vm_event_item item
;
2385 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2386 bool stalled
= false;
2388 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
2392 /* wait a bit for the reclaimer. */
2394 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_ISOLATED
);
2396 /* We are about to die and free our memory. Return now. */
2397 if (fatal_signal_pending(current
))
2398 return SWAP_CLUSTER_MAX
;
2403 spin_lock_irq(&lruvec
->lru_lock
);
2405 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
2406 &nr_scanned
, sc
, lru
);
2408 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2409 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
2410 if (!cgroup_reclaim(sc
))
2411 __count_vm_events(item
, nr_scanned
);
2412 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
2413 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
2415 spin_unlock_irq(&lruvec
->lru_lock
);
2420 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, &stat
, false);
2422 spin_lock_irq(&lruvec
->lru_lock
);
2423 move_pages_to_lru(lruvec
, &page_list
);
2425 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2426 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2427 if (!cgroup_reclaim(sc
))
2428 __count_vm_events(item
, nr_reclaimed
);
2429 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2430 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
2431 spin_unlock_irq(&lruvec
->lru_lock
);
2433 lru_note_cost(lruvec
, file
, stat
.nr_pageout
);
2434 mem_cgroup_uncharge_list(&page_list
);
2435 free_unref_page_list(&page_list
);
2438 * If dirty pages are scanned that are not queued for IO, it
2439 * implies that flushers are not doing their job. This can
2440 * happen when memory pressure pushes dirty pages to the end of
2441 * the LRU before the dirty limits are breached and the dirty
2442 * data has expired. It can also happen when the proportion of
2443 * dirty pages grows not through writes but through memory
2444 * pressure reclaiming all the clean cache. And in some cases,
2445 * the flushers simply cannot keep up with the allocation
2446 * rate. Nudge the flusher threads in case they are asleep.
2448 if (stat
.nr_unqueued_dirty
== nr_taken
)
2449 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2451 sc
->nr
.dirty
+= stat
.nr_dirty
;
2452 sc
->nr
.congested
+= stat
.nr_congested
;
2453 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2454 sc
->nr
.writeback
+= stat
.nr_writeback
;
2455 sc
->nr
.immediate
+= stat
.nr_immediate
;
2456 sc
->nr
.taken
+= nr_taken
;
2458 sc
->nr
.file_taken
+= nr_taken
;
2460 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2461 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2462 return nr_reclaimed
;
2466 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2468 * We move them the other way if the page is referenced by one or more
2471 * If the pages are mostly unmapped, the processing is fast and it is
2472 * appropriate to hold lru_lock across the whole operation. But if
2473 * the pages are mapped, the processing is slow (page_referenced()), so
2474 * we should drop lru_lock around each page. It's impossible to balance
2475 * this, so instead we remove the pages from the LRU while processing them.
2476 * It is safe to rely on PG_active against the non-LRU pages in here because
2477 * nobody will play with that bit on a non-LRU page.
2479 * The downside is that we have to touch page->_refcount against each page.
2480 * But we had to alter page->flags anyway.
2482 static void shrink_active_list(unsigned long nr_to_scan
,
2483 struct lruvec
*lruvec
,
2484 struct scan_control
*sc
,
2487 unsigned long nr_taken
;
2488 unsigned long nr_scanned
;
2489 unsigned long vm_flags
;
2490 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2491 LIST_HEAD(l_active
);
2492 LIST_HEAD(l_inactive
);
2494 unsigned nr_deactivate
, nr_activate
;
2495 unsigned nr_rotated
= 0;
2496 int file
= is_file_lru(lru
);
2497 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2501 spin_lock_irq(&lruvec
->lru_lock
);
2503 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2504 &nr_scanned
, sc
, lru
);
2506 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2508 if (!cgroup_reclaim(sc
))
2509 __count_vm_events(PGREFILL
, nr_scanned
);
2510 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2512 spin_unlock_irq(&lruvec
->lru_lock
);
2514 while (!list_empty(&l_hold
)) {
2516 page
= lru_to_page(&l_hold
);
2517 list_del(&page
->lru
);
2519 if (unlikely(!page_evictable(page
))) {
2520 putback_lru_page(page
);
2524 if (unlikely(buffer_heads_over_limit
)) {
2525 if (page_has_private(page
) && trylock_page(page
)) {
2526 if (page_has_private(page
))
2527 try_to_release_page(page
, 0);
2532 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2535 * Identify referenced, file-backed active pages and
2536 * give them one more trip around the active list. So
2537 * that executable code get better chances to stay in
2538 * memory under moderate memory pressure. Anon pages
2539 * are not likely to be evicted by use-once streaming
2540 * IO, plus JVM can create lots of anon VM_EXEC pages,
2541 * so we ignore them here.
2543 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2544 nr_rotated
+= thp_nr_pages(page
);
2545 list_add(&page
->lru
, &l_active
);
2550 ClearPageActive(page
); /* we are de-activating */
2551 SetPageWorkingset(page
);
2552 list_add(&page
->lru
, &l_inactive
);
2556 * Move pages back to the lru list.
2558 spin_lock_irq(&lruvec
->lru_lock
);
2560 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2561 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2562 /* Keep all free pages in l_active list */
2563 list_splice(&l_inactive
, &l_active
);
2565 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2566 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2568 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2569 spin_unlock_irq(&lruvec
->lru_lock
);
2571 mem_cgroup_uncharge_list(&l_active
);
2572 free_unref_page_list(&l_active
);
2573 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2574 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2577 unsigned long reclaim_pages(struct list_head
*page_list
)
2579 int nid
= NUMA_NO_NODE
;
2580 unsigned int nr_reclaimed
= 0;
2581 LIST_HEAD(node_page_list
);
2582 struct reclaim_stat dummy_stat
;
2584 unsigned int noreclaim_flag
;
2585 struct scan_control sc
= {
2586 .gfp_mask
= GFP_KERNEL
,
2593 noreclaim_flag
= memalloc_noreclaim_save();
2595 while (!list_empty(page_list
)) {
2596 page
= lru_to_page(page_list
);
2597 if (nid
== NUMA_NO_NODE
) {
2598 nid
= page_to_nid(page
);
2599 INIT_LIST_HEAD(&node_page_list
);
2602 if (nid
== page_to_nid(page
)) {
2603 ClearPageActive(page
);
2604 list_move(&page
->lru
, &node_page_list
);
2608 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2610 &sc
, &dummy_stat
, false);
2611 while (!list_empty(&node_page_list
)) {
2612 page
= lru_to_page(&node_page_list
);
2613 list_del(&page
->lru
);
2614 putback_lru_page(page
);
2620 if (!list_empty(&node_page_list
)) {
2621 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2623 &sc
, &dummy_stat
, false);
2624 while (!list_empty(&node_page_list
)) {
2625 page
= lru_to_page(&node_page_list
);
2626 list_del(&page
->lru
);
2627 putback_lru_page(page
);
2631 memalloc_noreclaim_restore(noreclaim_flag
);
2633 return nr_reclaimed
;
2636 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2637 struct lruvec
*lruvec
, struct scan_control
*sc
)
2639 if (is_active_lru(lru
)) {
2640 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2641 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2643 sc
->skipped_deactivate
= 1;
2647 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2651 * The inactive anon list should be small enough that the VM never has
2652 * to do too much work.
2654 * The inactive file list should be small enough to leave most memory
2655 * to the established workingset on the scan-resistant active list,
2656 * but large enough to avoid thrashing the aggregate readahead window.
2658 * Both inactive lists should also be large enough that each inactive
2659 * page has a chance to be referenced again before it is reclaimed.
2661 * If that fails and refaulting is observed, the inactive list grows.
2663 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2664 * on this LRU, maintained by the pageout code. An inactive_ratio
2665 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2668 * memory ratio inactive
2669 * -------------------------------------
2678 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2680 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2681 unsigned long inactive
, active
;
2682 unsigned long inactive_ratio
;
2685 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2686 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2688 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2690 inactive_ratio
= int_sqrt(10 * gb
);
2694 return inactive
* inactive_ratio
< active
;
2705 * Determine how aggressively the anon and file LRU lists should be
2706 * scanned. The relative value of each set of LRU lists is determined
2707 * by looking at the fraction of the pages scanned we did rotate back
2708 * onto the active list instead of evict.
2710 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2711 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2713 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2716 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2717 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2718 unsigned long anon_cost
, file_cost
, total_cost
;
2719 int swappiness
= mem_cgroup_swappiness(memcg
);
2720 u64 fraction
[ANON_AND_FILE
];
2721 u64 denominator
= 0; /* gcc */
2722 enum scan_balance scan_balance
;
2723 unsigned long ap
, fp
;
2726 /* If we have no swap space, do not bother scanning anon pages. */
2727 if (!sc
->may_swap
|| !can_reclaim_anon_pages(memcg
, pgdat
->node_id
, sc
)) {
2728 scan_balance
= SCAN_FILE
;
2733 * Global reclaim will swap to prevent OOM even with no
2734 * swappiness, but memcg users want to use this knob to
2735 * disable swapping for individual groups completely when
2736 * using the memory controller's swap limit feature would be
2739 if (cgroup_reclaim(sc
) && !swappiness
) {
2740 scan_balance
= SCAN_FILE
;
2745 * Do not apply any pressure balancing cleverness when the
2746 * system is close to OOM, scan both anon and file equally
2747 * (unless the swappiness setting disagrees with swapping).
2749 if (!sc
->priority
&& swappiness
) {
2750 scan_balance
= SCAN_EQUAL
;
2755 * If the system is almost out of file pages, force-scan anon.
2757 if (sc
->file_is_tiny
) {
2758 scan_balance
= SCAN_ANON
;
2763 * If there is enough inactive page cache, we do not reclaim
2764 * anything from the anonymous working right now.
2766 if (sc
->cache_trim_mode
) {
2767 scan_balance
= SCAN_FILE
;
2771 scan_balance
= SCAN_FRACT
;
2773 * Calculate the pressure balance between anon and file pages.
2775 * The amount of pressure we put on each LRU is inversely
2776 * proportional to the cost of reclaiming each list, as
2777 * determined by the share of pages that are refaulting, times
2778 * the relative IO cost of bringing back a swapped out
2779 * anonymous page vs reloading a filesystem page (swappiness).
2781 * Although we limit that influence to ensure no list gets
2782 * left behind completely: at least a third of the pressure is
2783 * applied, before swappiness.
2785 * With swappiness at 100, anon and file have equal IO cost.
2787 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2788 anon_cost
= total_cost
+ sc
->anon_cost
;
2789 file_cost
= total_cost
+ sc
->file_cost
;
2790 total_cost
= anon_cost
+ file_cost
;
2792 ap
= swappiness
* (total_cost
+ 1);
2793 ap
/= anon_cost
+ 1;
2795 fp
= (200 - swappiness
) * (total_cost
+ 1);
2796 fp
/= file_cost
+ 1;
2800 denominator
= ap
+ fp
;
2802 for_each_evictable_lru(lru
) {
2803 int file
= is_file_lru(lru
);
2804 unsigned long lruvec_size
;
2805 unsigned long low
, min
;
2808 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2809 mem_cgroup_protection(sc
->target_mem_cgroup
, memcg
,
2814 * Scale a cgroup's reclaim pressure by proportioning
2815 * its current usage to its memory.low or memory.min
2818 * This is important, as otherwise scanning aggression
2819 * becomes extremely binary -- from nothing as we
2820 * approach the memory protection threshold, to totally
2821 * nominal as we exceed it. This results in requiring
2822 * setting extremely liberal protection thresholds. It
2823 * also means we simply get no protection at all if we
2824 * set it too low, which is not ideal.
2826 * If there is any protection in place, we reduce scan
2827 * pressure by how much of the total memory used is
2828 * within protection thresholds.
2830 * There is one special case: in the first reclaim pass,
2831 * we skip over all groups that are within their low
2832 * protection. If that fails to reclaim enough pages to
2833 * satisfy the reclaim goal, we come back and override
2834 * the best-effort low protection. However, we still
2835 * ideally want to honor how well-behaved groups are in
2836 * that case instead of simply punishing them all
2837 * equally. As such, we reclaim them based on how much
2838 * memory they are using, reducing the scan pressure
2839 * again by how much of the total memory used is under
2842 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2843 unsigned long protection
;
2845 /* memory.low scaling, make sure we retry before OOM */
2846 if (!sc
->memcg_low_reclaim
&& low
> min
) {
2848 sc
->memcg_low_skipped
= 1;
2853 /* Avoid TOCTOU with earlier protection check */
2854 cgroup_size
= max(cgroup_size
, protection
);
2856 scan
= lruvec_size
- lruvec_size
* protection
/
2860 * Minimally target SWAP_CLUSTER_MAX pages to keep
2861 * reclaim moving forwards, avoiding decrementing
2862 * sc->priority further than desirable.
2864 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2869 scan
>>= sc
->priority
;
2872 * If the cgroup's already been deleted, make sure to
2873 * scrape out the remaining cache.
2875 if (!scan
&& !mem_cgroup_online(memcg
))
2876 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2878 switch (scan_balance
) {
2880 /* Scan lists relative to size */
2884 * Scan types proportional to swappiness and
2885 * their relative recent reclaim efficiency.
2886 * Make sure we don't miss the last page on
2887 * the offlined memory cgroups because of a
2890 scan
= mem_cgroup_online(memcg
) ?
2891 div64_u64(scan
* fraction
[file
], denominator
) :
2892 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2897 /* Scan one type exclusively */
2898 if ((scan_balance
== SCAN_FILE
) != file
)
2902 /* Look ma, no brain */
2911 * Anonymous LRU management is a waste if there is
2912 * ultimately no way to reclaim the memory.
2914 static bool can_age_anon_pages(struct pglist_data
*pgdat
,
2915 struct scan_control
*sc
)
2917 /* Aging the anon LRU is valuable if swap is present: */
2918 if (total_swap_pages
> 0)
2921 /* Also valuable if anon pages can be demoted: */
2922 return can_demote(pgdat
->node_id
, sc
);
2925 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2927 unsigned long nr
[NR_LRU_LISTS
];
2928 unsigned long targets
[NR_LRU_LISTS
];
2929 unsigned long nr_to_scan
;
2931 unsigned long nr_reclaimed
= 0;
2932 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2933 struct blk_plug plug
;
2936 get_scan_count(lruvec
, sc
, nr
);
2938 /* Record the original scan target for proportional adjustments later */
2939 memcpy(targets
, nr
, sizeof(nr
));
2942 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2943 * event that can occur when there is little memory pressure e.g.
2944 * multiple streaming readers/writers. Hence, we do not abort scanning
2945 * when the requested number of pages are reclaimed when scanning at
2946 * DEF_PRIORITY on the assumption that the fact we are direct
2947 * reclaiming implies that kswapd is not keeping up and it is best to
2948 * do a batch of work at once. For memcg reclaim one check is made to
2949 * abort proportional reclaim if either the file or anon lru has already
2950 * dropped to zero at the first pass.
2952 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2953 sc
->priority
== DEF_PRIORITY
);
2955 blk_start_plug(&plug
);
2956 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2957 nr
[LRU_INACTIVE_FILE
]) {
2958 unsigned long nr_anon
, nr_file
, percentage
;
2959 unsigned long nr_scanned
;
2961 for_each_evictable_lru(lru
) {
2963 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2964 nr
[lru
] -= nr_to_scan
;
2966 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2973 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2977 * For kswapd and memcg, reclaim at least the number of pages
2978 * requested. Ensure that the anon and file LRUs are scanned
2979 * proportionally what was requested by get_scan_count(). We
2980 * stop reclaiming one LRU and reduce the amount scanning
2981 * proportional to the original scan target.
2983 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2984 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2987 * It's just vindictive to attack the larger once the smaller
2988 * has gone to zero. And given the way we stop scanning the
2989 * smaller below, this makes sure that we only make one nudge
2990 * towards proportionality once we've got nr_to_reclaim.
2992 if (!nr_file
|| !nr_anon
)
2995 if (nr_file
> nr_anon
) {
2996 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2997 targets
[LRU_ACTIVE_ANON
] + 1;
2999 percentage
= nr_anon
* 100 / scan_target
;
3001 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
3002 targets
[LRU_ACTIVE_FILE
] + 1;
3004 percentage
= nr_file
* 100 / scan_target
;
3007 /* Stop scanning the smaller of the LRU */
3009 nr
[lru
+ LRU_ACTIVE
] = 0;
3012 * Recalculate the other LRU scan count based on its original
3013 * scan target and the percentage scanning already complete
3015 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
3016 nr_scanned
= targets
[lru
] - nr
[lru
];
3017 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
3018 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
3021 nr_scanned
= targets
[lru
] - nr
[lru
];
3022 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
3023 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
3025 scan_adjusted
= true;
3027 blk_finish_plug(&plug
);
3028 sc
->nr_reclaimed
+= nr_reclaimed
;
3031 * Even if we did not try to evict anon pages at all, we want to
3032 * rebalance the anon lru active/inactive ratio.
3034 if (can_age_anon_pages(lruvec_pgdat(lruvec
), sc
) &&
3035 inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3036 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3037 sc
, LRU_ACTIVE_ANON
);
3040 /* Use reclaim/compaction for costly allocs or under memory pressure */
3041 static bool in_reclaim_compaction(struct scan_control
*sc
)
3043 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
3044 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
3045 sc
->priority
< DEF_PRIORITY
- 2))
3052 * Reclaim/compaction is used for high-order allocation requests. It reclaims
3053 * order-0 pages before compacting the zone. should_continue_reclaim() returns
3054 * true if more pages should be reclaimed such that when the page allocator
3055 * calls try_to_compact_pages() that it will have enough free pages to succeed.
3056 * It will give up earlier than that if there is difficulty reclaiming pages.
3058 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
3059 unsigned long nr_reclaimed
,
3060 struct scan_control
*sc
)
3062 unsigned long pages_for_compaction
;
3063 unsigned long inactive_lru_pages
;
3066 /* If not in reclaim/compaction mode, stop */
3067 if (!in_reclaim_compaction(sc
))
3071 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
3072 * number of pages that were scanned. This will return to the caller
3073 * with the risk reclaim/compaction and the resulting allocation attempt
3074 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
3075 * allocations through requiring that the full LRU list has been scanned
3076 * first, by assuming that zero delta of sc->nr_scanned means full LRU
3077 * scan, but that approximation was wrong, and there were corner cases
3078 * where always a non-zero amount of pages were scanned.
3083 /* If compaction would go ahead or the allocation would succeed, stop */
3084 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3085 struct zone
*zone
= &pgdat
->node_zones
[z
];
3086 if (!managed_zone(zone
))
3089 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
3090 case COMPACT_SUCCESS
:
3091 case COMPACT_CONTINUE
:
3094 /* check next zone */
3100 * If we have not reclaimed enough pages for compaction and the
3101 * inactive lists are large enough, continue reclaiming
3103 pages_for_compaction
= compact_gap(sc
->order
);
3104 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
3105 if (can_reclaim_anon_pages(NULL
, pgdat
->node_id
, sc
))
3106 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3108 return inactive_lru_pages
> pages_for_compaction
;
3111 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
3113 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
3114 struct mem_cgroup
*memcg
;
3116 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
3118 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3119 unsigned long reclaimed
;
3120 unsigned long scanned
;
3123 * This loop can become CPU-bound when target memcgs
3124 * aren't eligible for reclaim - either because they
3125 * don't have any reclaimable pages, or because their
3126 * memory is explicitly protected. Avoid soft lockups.
3130 mem_cgroup_calculate_protection(target_memcg
, memcg
);
3132 if (mem_cgroup_below_min(memcg
)) {
3135 * If there is no reclaimable memory, OOM.
3138 } else if (mem_cgroup_below_low(memcg
)) {
3141 * Respect the protection only as long as
3142 * there is an unprotected supply
3143 * of reclaimable memory from other cgroups.
3145 if (!sc
->memcg_low_reclaim
) {
3146 sc
->memcg_low_skipped
= 1;
3149 memcg_memory_event(memcg
, MEMCG_LOW
);
3152 reclaimed
= sc
->nr_reclaimed
;
3153 scanned
= sc
->nr_scanned
;
3155 shrink_lruvec(lruvec
, sc
);
3157 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
3160 /* Record the group's reclaim efficiency */
3161 vmpressure(sc
->gfp_mask
, memcg
, false,
3162 sc
->nr_scanned
- scanned
,
3163 sc
->nr_reclaimed
- reclaimed
);
3165 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
3168 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
3170 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
3171 unsigned long nr_reclaimed
, nr_scanned
;
3172 struct lruvec
*target_lruvec
;
3173 bool reclaimable
= false;
3176 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
3180 * Flush the memory cgroup stats, so that we read accurate per-memcg
3181 * lruvec stats for heuristics.
3183 mem_cgroup_flush_stats();
3185 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
3187 nr_reclaimed
= sc
->nr_reclaimed
;
3188 nr_scanned
= sc
->nr_scanned
;
3191 * Determine the scan balance between anon and file LRUs.
3193 spin_lock_irq(&target_lruvec
->lru_lock
);
3194 sc
->anon_cost
= target_lruvec
->anon_cost
;
3195 sc
->file_cost
= target_lruvec
->file_cost
;
3196 spin_unlock_irq(&target_lruvec
->lru_lock
);
3199 * Target desirable inactive:active list ratios for the anon
3200 * and file LRU lists.
3202 if (!sc
->force_deactivate
) {
3203 unsigned long refaults
;
3205 refaults
= lruvec_page_state(target_lruvec
,
3206 WORKINGSET_ACTIVATE_ANON
);
3207 if (refaults
!= target_lruvec
->refaults
[0] ||
3208 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
3209 sc
->may_deactivate
|= DEACTIVATE_ANON
;
3211 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
3214 * When refaults are being observed, it means a new
3215 * workingset is being established. Deactivate to get
3216 * rid of any stale active pages quickly.
3218 refaults
= lruvec_page_state(target_lruvec
,
3219 WORKINGSET_ACTIVATE_FILE
);
3220 if (refaults
!= target_lruvec
->refaults
[1] ||
3221 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
3222 sc
->may_deactivate
|= DEACTIVATE_FILE
;
3224 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
3226 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
3229 * If we have plenty of inactive file pages that aren't
3230 * thrashing, try to reclaim those first before touching
3233 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
3234 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
3235 sc
->cache_trim_mode
= 1;
3237 sc
->cache_trim_mode
= 0;
3240 * Prevent the reclaimer from falling into the cache trap: as
3241 * cache pages start out inactive, every cache fault will tip
3242 * the scan balance towards the file LRU. And as the file LRU
3243 * shrinks, so does the window for rotation from references.
3244 * This means we have a runaway feedback loop where a tiny
3245 * thrashing file LRU becomes infinitely more attractive than
3246 * anon pages. Try to detect this based on file LRU size.
3248 if (!cgroup_reclaim(sc
)) {
3249 unsigned long total_high_wmark
= 0;
3250 unsigned long free
, anon
;
3253 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
3254 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
3255 node_page_state(pgdat
, NR_INACTIVE_FILE
);
3257 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
3258 struct zone
*zone
= &pgdat
->node_zones
[z
];
3259 if (!managed_zone(zone
))
3262 total_high_wmark
+= high_wmark_pages(zone
);
3266 * Consider anon: if that's low too, this isn't a
3267 * runaway file reclaim problem, but rather just
3268 * extreme pressure. Reclaim as per usual then.
3270 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
3273 file
+ free
<= total_high_wmark
&&
3274 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
3275 anon
>> sc
->priority
;
3278 shrink_node_memcgs(pgdat
, sc
);
3280 if (reclaim_state
) {
3281 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
3282 reclaim_state
->reclaimed_slab
= 0;
3285 /* Record the subtree's reclaim efficiency */
3286 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
3287 sc
->nr_scanned
- nr_scanned
,
3288 sc
->nr_reclaimed
- nr_reclaimed
);
3290 if (sc
->nr_reclaimed
- nr_reclaimed
)
3293 if (current_is_kswapd()) {
3295 * If reclaim is isolating dirty pages under writeback,
3296 * it implies that the long-lived page allocation rate
3297 * is exceeding the page laundering rate. Either the
3298 * global limits are not being effective at throttling
3299 * processes due to the page distribution throughout
3300 * zones or there is heavy usage of a slow backing
3301 * device. The only option is to throttle from reclaim
3302 * context which is not ideal as there is no guarantee
3303 * the dirtying process is throttled in the same way
3304 * balance_dirty_pages() manages.
3306 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3307 * count the number of pages under pages flagged for
3308 * immediate reclaim and stall if any are encountered
3309 * in the nr_immediate check below.
3311 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
3312 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3314 /* Allow kswapd to start writing pages during reclaim.*/
3315 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
3316 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3319 * If kswapd scans pages marked for immediate
3320 * reclaim and under writeback (nr_immediate), it
3321 * implies that pages are cycling through the LRU
3322 * faster than they are written so forcibly stall
3323 * until some pages complete writeback.
3325 if (sc
->nr
.immediate
)
3326 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_WRITEBACK
);
3330 * Tag a node/memcg as congested if all the dirty pages were marked
3331 * for writeback and immediate reclaim (counted in nr.congested).
3333 * Legacy memcg will stall in page writeback so avoid forcibly
3334 * stalling in reclaim_throttle().
3336 if ((current_is_kswapd() ||
3337 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
3338 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
3339 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
3342 * Stall direct reclaim for IO completions if the lruvec is
3343 * node is congested. Allow kswapd to continue until it
3344 * starts encountering unqueued dirty pages or cycling through
3345 * the LRU too quickly.
3347 if (!current_is_kswapd() && current_may_throttle() &&
3348 !sc
->hibernation_mode
&&
3349 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
3350 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_CONGESTED
);
3352 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
3357 * Kswapd gives up on balancing particular nodes after too
3358 * many failures to reclaim anything from them and goes to
3359 * sleep. On reclaim progress, reset the failure counter. A
3360 * successful direct reclaim run will revive a dormant kswapd.
3363 pgdat
->kswapd_failures
= 0;
3367 * Returns true if compaction should go ahead for a costly-order request, or
3368 * the allocation would already succeed without compaction. Return false if we
3369 * should reclaim first.
3371 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
3373 unsigned long watermark
;
3374 enum compact_result suitable
;
3376 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
3377 if (suitable
== COMPACT_SUCCESS
)
3378 /* Allocation should succeed already. Don't reclaim. */
3380 if (suitable
== COMPACT_SKIPPED
)
3381 /* Compaction cannot yet proceed. Do reclaim. */
3385 * Compaction is already possible, but it takes time to run and there
3386 * are potentially other callers using the pages just freed. So proceed
3387 * with reclaim to make a buffer of free pages available to give
3388 * compaction a reasonable chance of completing and allocating the page.
3389 * Note that we won't actually reclaim the whole buffer in one attempt
3390 * as the target watermark in should_continue_reclaim() is lower. But if
3391 * we are already above the high+gap watermark, don't reclaim at all.
3393 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
3395 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
3398 static void consider_reclaim_throttle(pg_data_t
*pgdat
, struct scan_control
*sc
)
3401 * If reclaim is making progress greater than 12% efficiency then
3402 * wake all the NOPROGRESS throttled tasks.
3404 if (sc
->nr_reclaimed
> (sc
->nr_scanned
>> 3)) {
3405 wait_queue_head_t
*wqh
;
3407 wqh
= &pgdat
->reclaim_wait
[VMSCAN_THROTTLE_NOPROGRESS
];
3408 if (waitqueue_active(wqh
))
3415 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
3416 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
3417 * under writeback and marked for immediate reclaim at the tail of the
3420 if (current_is_kswapd() || cgroup_reclaim(sc
))
3423 /* Throttle if making no progress at high prioities. */
3424 if (sc
->priority
== 1 && !sc
->nr_reclaimed
)
3425 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_NOPROGRESS
);
3429 * This is the direct reclaim path, for page-allocating processes. We only
3430 * try to reclaim pages from zones which will satisfy the caller's allocation
3433 * If a zone is deemed to be full of pinned pages then just give it a light
3434 * scan then give up on it.
3436 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
3440 unsigned long nr_soft_reclaimed
;
3441 unsigned long nr_soft_scanned
;
3443 pg_data_t
*last_pgdat
= NULL
;
3444 pg_data_t
*first_pgdat
= NULL
;
3447 * If the number of buffer_heads in the machine exceeds the maximum
3448 * allowed level, force direct reclaim to scan the highmem zone as
3449 * highmem pages could be pinning lowmem pages storing buffer_heads
3451 orig_mask
= sc
->gfp_mask
;
3452 if (buffer_heads_over_limit
) {
3453 sc
->gfp_mask
|= __GFP_HIGHMEM
;
3454 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
3457 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3458 sc
->reclaim_idx
, sc
->nodemask
) {
3460 * Take care memory controller reclaiming has small influence
3463 if (!cgroup_reclaim(sc
)) {
3464 if (!cpuset_zone_allowed(zone
,
3465 GFP_KERNEL
| __GFP_HARDWALL
))
3469 * If we already have plenty of memory free for
3470 * compaction in this zone, don't free any more.
3471 * Even though compaction is invoked for any
3472 * non-zero order, only frequent costly order
3473 * reclamation is disruptive enough to become a
3474 * noticeable problem, like transparent huge
3477 if (IS_ENABLED(CONFIG_COMPACTION
) &&
3478 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
3479 compaction_ready(zone
, sc
)) {
3480 sc
->compaction_ready
= true;
3485 * Shrink each node in the zonelist once. If the
3486 * zonelist is ordered by zone (not the default) then a
3487 * node may be shrunk multiple times but in that case
3488 * the user prefers lower zones being preserved.
3490 if (zone
->zone_pgdat
== last_pgdat
)
3494 * This steals pages from memory cgroups over softlimit
3495 * and returns the number of reclaimed pages and
3496 * scanned pages. This works for global memory pressure
3497 * and balancing, not for a memcg's limit.
3499 nr_soft_scanned
= 0;
3500 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
3501 sc
->order
, sc
->gfp_mask
,
3503 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
3504 sc
->nr_scanned
+= nr_soft_scanned
;
3505 /* need some check for avoid more shrink_zone() */
3509 first_pgdat
= zone
->zone_pgdat
;
3511 /* See comment about same check for global reclaim above */
3512 if (zone
->zone_pgdat
== last_pgdat
)
3514 last_pgdat
= zone
->zone_pgdat
;
3515 shrink_node(zone
->zone_pgdat
, sc
);
3519 consider_reclaim_throttle(first_pgdat
, sc
);
3522 * Restore to original mask to avoid the impact on the caller if we
3523 * promoted it to __GFP_HIGHMEM.
3525 sc
->gfp_mask
= orig_mask
;
3528 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
3530 struct lruvec
*target_lruvec
;
3531 unsigned long refaults
;
3533 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
3534 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
3535 target_lruvec
->refaults
[0] = refaults
;
3536 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
3537 target_lruvec
->refaults
[1] = refaults
;
3541 * This is the main entry point to direct page reclaim.
3543 * If a full scan of the inactive list fails to free enough memory then we
3544 * are "out of memory" and something needs to be killed.
3546 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3547 * high - the zone may be full of dirty or under-writeback pages, which this
3548 * caller can't do much about. We kick the writeback threads and take explicit
3549 * naps in the hope that some of these pages can be written. But if the
3550 * allocating task holds filesystem locks which prevent writeout this might not
3551 * work, and the allocation attempt will fail.
3553 * returns: 0, if no pages reclaimed
3554 * else, the number of pages reclaimed
3556 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3557 struct scan_control
*sc
)
3559 int initial_priority
= sc
->priority
;
3560 pg_data_t
*last_pgdat
;
3564 delayacct_freepages_start();
3566 if (!cgroup_reclaim(sc
))
3567 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3570 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3573 shrink_zones(zonelist
, sc
);
3575 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3578 if (sc
->compaction_ready
)
3582 * If we're getting trouble reclaiming, start doing
3583 * writepage even in laptop mode.
3585 if (sc
->priority
< DEF_PRIORITY
- 2)
3586 sc
->may_writepage
= 1;
3587 } while (--sc
->priority
>= 0);
3590 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3592 if (zone
->zone_pgdat
== last_pgdat
)
3594 last_pgdat
= zone
->zone_pgdat
;
3596 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3598 if (cgroup_reclaim(sc
)) {
3599 struct lruvec
*lruvec
;
3601 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3603 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3607 delayacct_freepages_end();
3609 if (sc
->nr_reclaimed
)
3610 return sc
->nr_reclaimed
;
3612 /* Aborted reclaim to try compaction? don't OOM, then */
3613 if (sc
->compaction_ready
)
3617 * We make inactive:active ratio decisions based on the node's
3618 * composition of memory, but a restrictive reclaim_idx or a
3619 * memory.low cgroup setting can exempt large amounts of
3620 * memory from reclaim. Neither of which are very common, so
3621 * instead of doing costly eligibility calculations of the
3622 * entire cgroup subtree up front, we assume the estimates are
3623 * good, and retry with forcible deactivation if that fails.
3625 if (sc
->skipped_deactivate
) {
3626 sc
->priority
= initial_priority
;
3627 sc
->force_deactivate
= 1;
3628 sc
->skipped_deactivate
= 0;
3632 /* Untapped cgroup reserves? Don't OOM, retry. */
3633 if (sc
->memcg_low_skipped
) {
3634 sc
->priority
= initial_priority
;
3635 sc
->force_deactivate
= 0;
3636 sc
->memcg_low_reclaim
= 1;
3637 sc
->memcg_low_skipped
= 0;
3644 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3647 unsigned long pfmemalloc_reserve
= 0;
3648 unsigned long free_pages
= 0;
3652 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3655 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3656 zone
= &pgdat
->node_zones
[i
];
3657 if (!managed_zone(zone
))
3660 if (!zone_reclaimable_pages(zone
))
3663 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3664 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3667 /* If there are no reserves (unexpected config) then do not throttle */
3668 if (!pfmemalloc_reserve
)
3671 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3673 /* kswapd must be awake if processes are being throttled */
3674 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3675 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
3676 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
3678 wake_up_interruptible(&pgdat
->kswapd_wait
);
3685 * Throttle direct reclaimers if backing storage is backed by the network
3686 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3687 * depleted. kswapd will continue to make progress and wake the processes
3688 * when the low watermark is reached.
3690 * Returns true if a fatal signal was delivered during throttling. If this
3691 * happens, the page allocator should not consider triggering the OOM killer.
3693 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3694 nodemask_t
*nodemask
)
3698 pg_data_t
*pgdat
= NULL
;
3701 * Kernel threads should not be throttled as they may be indirectly
3702 * responsible for cleaning pages necessary for reclaim to make forward
3703 * progress. kjournald for example may enter direct reclaim while
3704 * committing a transaction where throttling it could forcing other
3705 * processes to block on log_wait_commit().
3707 if (current
->flags
& PF_KTHREAD
)
3711 * If a fatal signal is pending, this process should not throttle.
3712 * It should return quickly so it can exit and free its memory
3714 if (fatal_signal_pending(current
))
3718 * Check if the pfmemalloc reserves are ok by finding the first node
3719 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3720 * GFP_KERNEL will be required for allocating network buffers when
3721 * swapping over the network so ZONE_HIGHMEM is unusable.
3723 * Throttling is based on the first usable node and throttled processes
3724 * wait on a queue until kswapd makes progress and wakes them. There
3725 * is an affinity then between processes waking up and where reclaim
3726 * progress has been made assuming the process wakes on the same node.
3727 * More importantly, processes running on remote nodes will not compete
3728 * for remote pfmemalloc reserves and processes on different nodes
3729 * should make reasonable progress.
3731 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3732 gfp_zone(gfp_mask
), nodemask
) {
3733 if (zone_idx(zone
) > ZONE_NORMAL
)
3736 /* Throttle based on the first usable node */
3737 pgdat
= zone
->zone_pgdat
;
3738 if (allow_direct_reclaim(pgdat
))
3743 /* If no zone was usable by the allocation flags then do not throttle */
3747 /* Account for the throttling */
3748 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3751 * If the caller cannot enter the filesystem, it's possible that it
3752 * is due to the caller holding an FS lock or performing a journal
3753 * transaction in the case of a filesystem like ext[3|4]. In this case,
3754 * it is not safe to block on pfmemalloc_wait as kswapd could be
3755 * blocked waiting on the same lock. Instead, throttle for up to a
3756 * second before continuing.
3758 if (!(gfp_mask
& __GFP_FS
))
3759 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3760 allow_direct_reclaim(pgdat
), HZ
);
3762 /* Throttle until kswapd wakes the process */
3763 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3764 allow_direct_reclaim(pgdat
));
3766 if (fatal_signal_pending(current
))
3773 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3774 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3776 unsigned long nr_reclaimed
;
3777 struct scan_control sc
= {
3778 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3779 .gfp_mask
= current_gfp_context(gfp_mask
),
3780 .reclaim_idx
= gfp_zone(gfp_mask
),
3782 .nodemask
= nodemask
,
3783 .priority
= DEF_PRIORITY
,
3784 .may_writepage
= !laptop_mode
,
3790 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3791 * Confirm they are large enough for max values.
3793 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3794 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3795 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3798 * Do not enter reclaim if fatal signal was delivered while throttled.
3799 * 1 is returned so that the page allocator does not OOM kill at this
3802 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3805 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3806 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3808 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3810 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3811 set_task_reclaim_state(current
, NULL
);
3813 return nr_reclaimed
;
3818 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3819 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3820 gfp_t gfp_mask
, bool noswap
,
3822 unsigned long *nr_scanned
)
3824 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3825 struct scan_control sc
= {
3826 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3827 .target_mem_cgroup
= memcg
,
3828 .may_writepage
= !laptop_mode
,
3830 .reclaim_idx
= MAX_NR_ZONES
- 1,
3831 .may_swap
= !noswap
,
3834 WARN_ON_ONCE(!current
->reclaim_state
);
3836 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3837 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3839 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3843 * NOTE: Although we can get the priority field, using it
3844 * here is not a good idea, since it limits the pages we can scan.
3845 * if we don't reclaim here, the shrink_node from balance_pgdat
3846 * will pick up pages from other mem cgroup's as well. We hack
3847 * the priority and make it zero.
3849 shrink_lruvec(lruvec
, &sc
);
3851 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3853 *nr_scanned
= sc
.nr_scanned
;
3855 return sc
.nr_reclaimed
;
3858 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3859 unsigned long nr_pages
,
3863 unsigned long nr_reclaimed
;
3864 unsigned int noreclaim_flag
;
3865 struct scan_control sc
= {
3866 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3867 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3868 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3869 .reclaim_idx
= MAX_NR_ZONES
- 1,
3870 .target_mem_cgroup
= memcg
,
3871 .priority
= DEF_PRIORITY
,
3872 .may_writepage
= !laptop_mode
,
3874 .may_swap
= may_swap
,
3877 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3878 * equal pressure on all the nodes. This is based on the assumption that
3879 * the reclaim does not bail out early.
3881 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3883 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3884 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3885 noreclaim_flag
= memalloc_noreclaim_save();
3887 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3889 memalloc_noreclaim_restore(noreclaim_flag
);
3890 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3891 set_task_reclaim_state(current
, NULL
);
3893 return nr_reclaimed
;
3897 static void age_active_anon(struct pglist_data
*pgdat
,
3898 struct scan_control
*sc
)
3900 struct mem_cgroup
*memcg
;
3901 struct lruvec
*lruvec
;
3903 if (!can_age_anon_pages(pgdat
, sc
))
3906 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3907 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3910 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3912 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3913 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3914 sc
, LRU_ACTIVE_ANON
);
3915 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3919 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
3925 * Check for watermark boosts top-down as the higher zones
3926 * are more likely to be boosted. Both watermarks and boosts
3927 * should not be checked at the same time as reclaim would
3928 * start prematurely when there is no boosting and a lower
3931 for (i
= highest_zoneidx
; i
>= 0; i
--) {
3932 zone
= pgdat
->node_zones
+ i
;
3933 if (!managed_zone(zone
))
3936 if (zone
->watermark_boost
)
3944 * Returns true if there is an eligible zone balanced for the request order
3945 * and highest_zoneidx
3947 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3950 unsigned long mark
= -1;
3954 * Check watermarks bottom-up as lower zones are more likely to
3957 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3958 zone
= pgdat
->node_zones
+ i
;
3960 if (!managed_zone(zone
))
3963 mark
= high_wmark_pages(zone
);
3964 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
3969 * If a node has no populated zone within highest_zoneidx, it does not
3970 * need balancing by definition. This can happen if a zone-restricted
3971 * allocation tries to wake a remote kswapd.
3979 /* Clear pgdat state for congested, dirty or under writeback. */
3980 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3982 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3984 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3985 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3986 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3990 * Prepare kswapd for sleeping. This verifies that there are no processes
3991 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3993 * Returns true if kswapd is ready to sleep
3995 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
3996 int highest_zoneidx
)
3999 * The throttled processes are normally woken up in balance_pgdat() as
4000 * soon as allow_direct_reclaim() is true. But there is a potential
4001 * race between when kswapd checks the watermarks and a process gets
4002 * throttled. There is also a potential race if processes get
4003 * throttled, kswapd wakes, a large process exits thereby balancing the
4004 * zones, which causes kswapd to exit balance_pgdat() before reaching
4005 * the wake up checks. If kswapd is going to sleep, no process should
4006 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
4007 * the wake up is premature, processes will wake kswapd and get
4008 * throttled again. The difference from wake ups in balance_pgdat() is
4009 * that here we are under prepare_to_wait().
4011 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
4012 wake_up_all(&pgdat
->pfmemalloc_wait
);
4014 /* Hopeless node, leave it to direct reclaim */
4015 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
4018 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
4019 clear_pgdat_congested(pgdat
);
4027 * kswapd shrinks a node of pages that are at or below the highest usable
4028 * zone that is currently unbalanced.
4030 * Returns true if kswapd scanned at least the requested number of pages to
4031 * reclaim or if the lack of progress was due to pages under writeback.
4032 * This is used to determine if the scanning priority needs to be raised.
4034 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
4035 struct scan_control
*sc
)
4040 /* Reclaim a number of pages proportional to the number of zones */
4041 sc
->nr_to_reclaim
= 0;
4042 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
4043 zone
= pgdat
->node_zones
+ z
;
4044 if (!managed_zone(zone
))
4047 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
4051 * Historically care was taken to put equal pressure on all zones but
4052 * now pressure is applied based on node LRU order.
4054 shrink_node(pgdat
, sc
);
4057 * Fragmentation may mean that the system cannot be rebalanced for
4058 * high-order allocations. If twice the allocation size has been
4059 * reclaimed then recheck watermarks only at order-0 to prevent
4060 * excessive reclaim. Assume that a process requested a high-order
4061 * can direct reclaim/compact.
4063 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
4066 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
4069 /* Page allocator PCP high watermark is lowered if reclaim is active. */
4071 update_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
, bool active
)
4076 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4077 zone
= pgdat
->node_zones
+ i
;
4079 if (!managed_zone(zone
))
4083 set_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4085 clear_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
4090 set_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4092 update_reclaim_active(pgdat
, highest_zoneidx
, true);
4096 clear_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
4098 update_reclaim_active(pgdat
, highest_zoneidx
, false);
4102 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
4103 * that are eligible for use by the caller until at least one zone is
4106 * Returns the order kswapd finished reclaiming at.
4108 * kswapd scans the zones in the highmem->normal->dma direction. It skips
4109 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
4110 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
4111 * or lower is eligible for reclaim until at least one usable zone is
4114 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
4117 unsigned long nr_soft_reclaimed
;
4118 unsigned long nr_soft_scanned
;
4119 unsigned long pflags
;
4120 unsigned long nr_boost_reclaim
;
4121 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
4124 struct scan_control sc
= {
4125 .gfp_mask
= GFP_KERNEL
,
4130 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4131 psi_memstall_enter(&pflags
);
4132 __fs_reclaim_acquire(_THIS_IP_
);
4134 count_vm_event(PAGEOUTRUN
);
4137 * Account for the reclaim boost. Note that the zone boost is left in
4138 * place so that parallel allocations that are near the watermark will
4139 * stall or direct reclaim until kswapd is finished.
4141 nr_boost_reclaim
= 0;
4142 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4143 zone
= pgdat
->node_zones
+ i
;
4144 if (!managed_zone(zone
))
4147 nr_boost_reclaim
+= zone
->watermark_boost
;
4148 zone_boosts
[i
] = zone
->watermark_boost
;
4150 boosted
= nr_boost_reclaim
;
4153 set_reclaim_active(pgdat
, highest_zoneidx
);
4154 sc
.priority
= DEF_PRIORITY
;
4156 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
4157 bool raise_priority
= true;
4161 sc
.reclaim_idx
= highest_zoneidx
;
4164 * If the number of buffer_heads exceeds the maximum allowed
4165 * then consider reclaiming from all zones. This has a dual
4166 * purpose -- on 64-bit systems it is expected that
4167 * buffer_heads are stripped during active rotation. On 32-bit
4168 * systems, highmem pages can pin lowmem memory and shrinking
4169 * buffers can relieve lowmem pressure. Reclaim may still not
4170 * go ahead if all eligible zones for the original allocation
4171 * request are balanced to avoid excessive reclaim from kswapd.
4173 if (buffer_heads_over_limit
) {
4174 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
4175 zone
= pgdat
->node_zones
+ i
;
4176 if (!managed_zone(zone
))
4185 * If the pgdat is imbalanced then ignore boosting and preserve
4186 * the watermarks for a later time and restart. Note that the
4187 * zone watermarks will be still reset at the end of balancing
4188 * on the grounds that the normal reclaim should be enough to
4189 * re-evaluate if boosting is required when kswapd next wakes.
4191 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
4192 if (!balanced
&& nr_boost_reclaim
) {
4193 nr_boost_reclaim
= 0;
4198 * If boosting is not active then only reclaim if there are no
4199 * eligible zones. Note that sc.reclaim_idx is not used as
4200 * buffer_heads_over_limit may have adjusted it.
4202 if (!nr_boost_reclaim
&& balanced
)
4205 /* Limit the priority of boosting to avoid reclaim writeback */
4206 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
4207 raise_priority
= false;
4210 * Do not writeback or swap pages for boosted reclaim. The
4211 * intent is to relieve pressure not issue sub-optimal IO
4212 * from reclaim context. If no pages are reclaimed, the
4213 * reclaim will be aborted.
4215 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
4216 sc
.may_swap
= !nr_boost_reclaim
;
4219 * Do some background aging of the anon list, to give
4220 * pages a chance to be referenced before reclaiming. All
4221 * pages are rotated regardless of classzone as this is
4222 * about consistent aging.
4224 age_active_anon(pgdat
, &sc
);
4227 * If we're getting trouble reclaiming, start doing writepage
4228 * even in laptop mode.
4230 if (sc
.priority
< DEF_PRIORITY
- 2)
4231 sc
.may_writepage
= 1;
4233 /* Call soft limit reclaim before calling shrink_node. */
4235 nr_soft_scanned
= 0;
4236 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
4237 sc
.gfp_mask
, &nr_soft_scanned
);
4238 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
4241 * There should be no need to raise the scanning priority if
4242 * enough pages are already being scanned that that high
4243 * watermark would be met at 100% efficiency.
4245 if (kswapd_shrink_node(pgdat
, &sc
))
4246 raise_priority
= false;
4249 * If the low watermark is met there is no need for processes
4250 * to be throttled on pfmemalloc_wait as they should not be
4251 * able to safely make forward progress. Wake them
4253 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
4254 allow_direct_reclaim(pgdat
))
4255 wake_up_all(&pgdat
->pfmemalloc_wait
);
4257 /* Check if kswapd should be suspending */
4258 __fs_reclaim_release(_THIS_IP_
);
4259 ret
= try_to_freeze();
4260 __fs_reclaim_acquire(_THIS_IP_
);
4261 if (ret
|| kthread_should_stop())
4265 * Raise priority if scanning rate is too low or there was no
4266 * progress in reclaiming pages
4268 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
4269 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
4272 * If reclaim made no progress for a boost, stop reclaim as
4273 * IO cannot be queued and it could be an infinite loop in
4274 * extreme circumstances.
4276 if (nr_boost_reclaim
&& !nr_reclaimed
)
4279 if (raise_priority
|| !nr_reclaimed
)
4281 } while (sc
.priority
>= 1);
4283 if (!sc
.nr_reclaimed
)
4284 pgdat
->kswapd_failures
++;
4287 clear_reclaim_active(pgdat
, highest_zoneidx
);
4289 /* If reclaim was boosted, account for the reclaim done in this pass */
4291 unsigned long flags
;
4293 for (i
= 0; i
<= highest_zoneidx
; i
++) {
4294 if (!zone_boosts
[i
])
4297 /* Increments are under the zone lock */
4298 zone
= pgdat
->node_zones
+ i
;
4299 spin_lock_irqsave(&zone
->lock
, flags
);
4300 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
4301 spin_unlock_irqrestore(&zone
->lock
, flags
);
4305 * As there is now likely space, wakeup kcompact to defragment
4308 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
4311 snapshot_refaults(NULL
, pgdat
);
4312 __fs_reclaim_release(_THIS_IP_
);
4313 psi_memstall_leave(&pflags
);
4314 set_task_reclaim_state(current
, NULL
);
4317 * Return the order kswapd stopped reclaiming at as
4318 * prepare_kswapd_sleep() takes it into account. If another caller
4319 * entered the allocator slow path while kswapd was awake, order will
4320 * remain at the higher level.
4326 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4327 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4328 * not a valid index then either kswapd runs for first time or kswapd couldn't
4329 * sleep after previous reclaim attempt (node is still unbalanced). In that
4330 * case return the zone index of the previous kswapd reclaim cycle.
4332 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
4333 enum zone_type prev_highest_zoneidx
)
4335 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4337 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
4340 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
4341 unsigned int highest_zoneidx
)
4346 if (freezing(current
) || kthread_should_stop())
4349 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4352 * Try to sleep for a short interval. Note that kcompactd will only be
4353 * woken if it is possible to sleep for a short interval. This is
4354 * deliberate on the assumption that if reclaim cannot keep an
4355 * eligible zone balanced that it's also unlikely that compaction will
4358 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4360 * Compaction records what page blocks it recently failed to
4361 * isolate pages from and skips them in the future scanning.
4362 * When kswapd is going to sleep, it is reasonable to assume
4363 * that pages and compaction may succeed so reset the cache.
4365 reset_isolation_suitable(pgdat
);
4368 * We have freed the memory, now we should compact it to make
4369 * allocation of the requested order possible.
4371 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
4373 remaining
= schedule_timeout(HZ
/10);
4376 * If woken prematurely then reset kswapd_highest_zoneidx and
4377 * order. The values will either be from a wakeup request or
4378 * the previous request that slept prematurely.
4381 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
4382 kswapd_highest_zoneidx(pgdat
,
4385 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
4386 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
4389 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4390 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
4394 * After a short sleep, check if it was a premature sleep. If not, then
4395 * go fully to sleep until explicitly woken up.
4398 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
4399 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
4402 * vmstat counters are not perfectly accurate and the estimated
4403 * value for counters such as NR_FREE_PAGES can deviate from the
4404 * true value by nr_online_cpus * threshold. To avoid the zone
4405 * watermarks being breached while under pressure, we reduce the
4406 * per-cpu vmstat threshold while kswapd is awake and restore
4407 * them before going back to sleep.
4409 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
4411 if (!kthread_should_stop())
4414 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
4417 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
4419 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
4421 finish_wait(&pgdat
->kswapd_wait
, &wait
);
4425 * The background pageout daemon, started as a kernel thread
4426 * from the init process.
4428 * This basically trickles out pages so that we have _some_
4429 * free memory available even if there is no other activity
4430 * that frees anything up. This is needed for things like routing
4431 * etc, where we otherwise might have all activity going on in
4432 * asynchronous contexts that cannot page things out.
4434 * If there are applications that are active memory-allocators
4435 * (most normal use), this basically shouldn't matter.
4437 static int kswapd(void *p
)
4439 unsigned int alloc_order
, reclaim_order
;
4440 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
4441 pg_data_t
*pgdat
= (pg_data_t
*)p
;
4442 struct task_struct
*tsk
= current
;
4443 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
4445 if (!cpumask_empty(cpumask
))
4446 set_cpus_allowed_ptr(tsk
, cpumask
);
4449 * Tell the memory management that we're a "memory allocator",
4450 * and that if we need more memory we should get access to it
4451 * regardless (see "__alloc_pages()"). "kswapd" should
4452 * never get caught in the normal page freeing logic.
4454 * (Kswapd normally doesn't need memory anyway, but sometimes
4455 * you need a small amount of memory in order to be able to
4456 * page out something else, and this flag essentially protects
4457 * us from recursively trying to free more memory as we're
4458 * trying to free the first piece of memory in the first place).
4460 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
4463 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4464 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4465 atomic_set(&pgdat
->nr_writeback_throttled
, 0);
4469 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
4470 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4474 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
4477 /* Read the new order and highest_zoneidx */
4478 alloc_order
= READ_ONCE(pgdat
->kswapd_order
);
4479 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
4481 WRITE_ONCE(pgdat
->kswapd_order
, 0);
4482 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
4484 ret
= try_to_freeze();
4485 if (kthread_should_stop())
4489 * We can speed up thawing tasks if we don't call balance_pgdat
4490 * after returning from the refrigerator
4496 * Reclaim begins at the requested order but if a high-order
4497 * reclaim fails then kswapd falls back to reclaiming for
4498 * order-0. If that happens, kswapd will consider sleeping
4499 * for the order it finished reclaiming at (reclaim_order)
4500 * but kcompactd is woken to compact for the original
4501 * request (alloc_order).
4503 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
4505 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
4507 if (reclaim_order
< alloc_order
)
4508 goto kswapd_try_sleep
;
4511 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
4517 * A zone is low on free memory or too fragmented for high-order memory. If
4518 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4519 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4520 * has failed or is not needed, still wake up kcompactd if only compaction is
4523 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
4524 enum zone_type highest_zoneidx
)
4527 enum zone_type curr_idx
;
4529 if (!managed_zone(zone
))
4532 if (!cpuset_zone_allowed(zone
, gfp_flags
))
4535 pgdat
= zone
->zone_pgdat
;
4536 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
4538 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
4539 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
4541 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
4542 WRITE_ONCE(pgdat
->kswapd_order
, order
);
4544 if (!waitqueue_active(&pgdat
->kswapd_wait
))
4547 /* Hopeless node, leave it to direct reclaim if possible */
4548 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
4549 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
4550 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
4552 * There may be plenty of free memory available, but it's too
4553 * fragmented for high-order allocations. Wake up kcompactd
4554 * and rely on compaction_suitable() to determine if it's
4555 * needed. If it fails, it will defer subsequent attempts to
4556 * ratelimit its work.
4558 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
4559 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
4563 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
4565 wake_up_interruptible(&pgdat
->kswapd_wait
);
4568 #ifdef CONFIG_HIBERNATION
4570 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4573 * Rather than trying to age LRUs the aim is to preserve the overall
4574 * LRU order by reclaiming preferentially
4575 * inactive > active > active referenced > active mapped
4577 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4579 struct scan_control sc
= {
4580 .nr_to_reclaim
= nr_to_reclaim
,
4581 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4582 .reclaim_idx
= MAX_NR_ZONES
- 1,
4583 .priority
= DEF_PRIORITY
,
4587 .hibernation_mode
= 1,
4589 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4590 unsigned long nr_reclaimed
;
4591 unsigned int noreclaim_flag
;
4593 fs_reclaim_acquire(sc
.gfp_mask
);
4594 noreclaim_flag
= memalloc_noreclaim_save();
4595 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4597 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4599 set_task_reclaim_state(current
, NULL
);
4600 memalloc_noreclaim_restore(noreclaim_flag
);
4601 fs_reclaim_release(sc
.gfp_mask
);
4603 return nr_reclaimed
;
4605 #endif /* CONFIG_HIBERNATION */
4608 * This kswapd start function will be called by init and node-hot-add.
4609 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4611 void kswapd_run(int nid
)
4613 pg_data_t
*pgdat
= NODE_DATA(nid
);
4618 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4619 if (IS_ERR(pgdat
->kswapd
)) {
4620 /* failure at boot is fatal */
4621 BUG_ON(system_state
< SYSTEM_RUNNING
);
4622 pr_err("Failed to start kswapd on node %d\n", nid
);
4623 pgdat
->kswapd
= NULL
;
4628 * Called by memory hotplug when all memory in a node is offlined. Caller must
4629 * hold mem_hotplug_begin/end().
4631 void kswapd_stop(int nid
)
4633 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4636 kthread_stop(kswapd
);
4637 NODE_DATA(nid
)->kswapd
= NULL
;
4641 static int __init
kswapd_init(void)
4646 for_each_node_state(nid
, N_MEMORY
)
4651 module_init(kswapd_init
)
4657 * If non-zero call node_reclaim when the number of free pages falls below
4660 int node_reclaim_mode __read_mostly
;
4663 * Priority for NODE_RECLAIM. This determines the fraction of pages
4664 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4667 #define NODE_RECLAIM_PRIORITY 4
4670 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4673 int sysctl_min_unmapped_ratio
= 1;
4676 * If the number of slab pages in a zone grows beyond this percentage then
4677 * slab reclaim needs to occur.
4679 int sysctl_min_slab_ratio
= 5;
4681 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4683 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4684 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4685 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4688 * It's possible for there to be more file mapped pages than
4689 * accounted for by the pages on the file LRU lists because
4690 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4692 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4695 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4696 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4698 unsigned long nr_pagecache_reclaimable
;
4699 unsigned long delta
= 0;
4702 * If RECLAIM_UNMAP is set, then all file pages are considered
4703 * potentially reclaimable. Otherwise, we have to worry about
4704 * pages like swapcache and node_unmapped_file_pages() provides
4707 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4708 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4710 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4712 /* If we can't clean pages, remove dirty pages from consideration */
4713 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4714 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4716 /* Watch for any possible underflows due to delta */
4717 if (unlikely(delta
> nr_pagecache_reclaimable
))
4718 delta
= nr_pagecache_reclaimable
;
4720 return nr_pagecache_reclaimable
- delta
;
4724 * Try to free up some pages from this node through reclaim.
4726 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4728 /* Minimum pages needed in order to stay on node */
4729 const unsigned long nr_pages
= 1 << order
;
4730 struct task_struct
*p
= current
;
4731 unsigned int noreclaim_flag
;
4732 struct scan_control sc
= {
4733 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4734 .gfp_mask
= current_gfp_context(gfp_mask
),
4736 .priority
= NODE_RECLAIM_PRIORITY
,
4737 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4738 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4740 .reclaim_idx
= gfp_zone(gfp_mask
),
4742 unsigned long pflags
;
4744 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4748 psi_memstall_enter(&pflags
);
4749 fs_reclaim_acquire(sc
.gfp_mask
);
4751 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4752 * and we also need to be able to write out pages for RECLAIM_WRITE
4753 * and RECLAIM_UNMAP.
4755 noreclaim_flag
= memalloc_noreclaim_save();
4756 p
->flags
|= PF_SWAPWRITE
;
4757 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4759 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4761 * Free memory by calling shrink node with increasing
4762 * priorities until we have enough memory freed.
4765 shrink_node(pgdat
, &sc
);
4766 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4769 set_task_reclaim_state(p
, NULL
);
4770 current
->flags
&= ~PF_SWAPWRITE
;
4771 memalloc_noreclaim_restore(noreclaim_flag
);
4772 fs_reclaim_release(sc
.gfp_mask
);
4773 psi_memstall_leave(&pflags
);
4775 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4777 return sc
.nr_reclaimed
>= nr_pages
;
4780 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4785 * Node reclaim reclaims unmapped file backed pages and
4786 * slab pages if we are over the defined limits.
4788 * A small portion of unmapped file backed pages is needed for
4789 * file I/O otherwise pages read by file I/O will be immediately
4790 * thrown out if the node is overallocated. So we do not reclaim
4791 * if less than a specified percentage of the node is used by
4792 * unmapped file backed pages.
4794 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4795 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
4796 pgdat
->min_slab_pages
)
4797 return NODE_RECLAIM_FULL
;
4800 * Do not scan if the allocation should not be delayed.
4802 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4803 return NODE_RECLAIM_NOSCAN
;
4806 * Only run node reclaim on the local node or on nodes that do not
4807 * have associated processors. This will favor the local processor
4808 * over remote processors and spread off node memory allocations
4809 * as wide as possible.
4811 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4812 return NODE_RECLAIM_NOSCAN
;
4814 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4815 return NODE_RECLAIM_NOSCAN
;
4817 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4818 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4821 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4828 * check_move_unevictable_pages - check pages for evictability and move to
4829 * appropriate zone lru list
4830 * @pvec: pagevec with lru pages to check
4832 * Checks pages for evictability, if an evictable page is in the unevictable
4833 * lru list, moves it to the appropriate evictable lru list. This function
4834 * should be only used for lru pages.
4836 void check_move_unevictable_pages(struct pagevec
*pvec
)
4838 struct lruvec
*lruvec
= NULL
;
4843 for (i
= 0; i
< pvec
->nr
; i
++) {
4844 struct page
*page
= pvec
->pages
[i
];
4845 struct folio
*folio
= page_folio(page
);
4848 if (PageTransTail(page
))
4851 nr_pages
= thp_nr_pages(page
);
4852 pgscanned
+= nr_pages
;
4854 /* block memcg migration during page moving between lru */
4855 if (!TestClearPageLRU(page
))
4858 lruvec
= folio_lruvec_relock_irq(folio
, lruvec
);
4859 if (page_evictable(page
) && PageUnevictable(page
)) {
4860 del_page_from_lru_list(page
, lruvec
);
4861 ClearPageUnevictable(page
);
4862 add_page_to_lru_list(page
, lruvec
);
4863 pgrescued
+= nr_pages
;
4869 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4870 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4871 unlock_page_lruvec_irq(lruvec
);
4872 } else if (pgscanned
) {
4873 count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4876 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);