1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim
;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup
*target_mem_cgroup
;
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage
:1;
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap
:1;
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap
:1;
92 * Cgroup memory below memory.low is protected as long as we
93 * don't threaten to OOM. If any cgroup is reclaimed at
94 * reduced force or passed over entirely due to its memory.low
95 * setting (memcg_low_skipped), and nothing is reclaimed as a
96 * result, then go back for one more cycle that reclaims the protected
97 * memory (memcg_low_reclaim) to avert OOM.
99 unsigned int memcg_low_reclaim
:1;
100 unsigned int memcg_low_skipped
:1;
102 unsigned int hibernation_mode
:1;
104 /* One of the zones is ready for compaction */
105 unsigned int compaction_ready
:1;
107 /* Allocation order */
110 /* Scan (total_size >> priority) pages at once */
113 /* The highest zone to isolate pages for reclaim from */
116 /* This context's GFP mask */
119 /* Incremented by the number of inactive pages that were scanned */
120 unsigned long nr_scanned
;
122 /* Number of pages freed so far during a call to shrink_zones() */
123 unsigned long nr_reclaimed
;
127 unsigned int unqueued_dirty
;
128 unsigned int congested
;
129 unsigned int writeback
;
130 unsigned int immediate
;
131 unsigned int file_taken
;
135 /* for recording the reclaimed slab by now */
136 struct reclaim_state reclaim_state
;
139 #ifdef ARCH_HAS_PREFETCH
140 #define prefetch_prev_lru_page(_page, _base, _field) \
142 if ((_page)->lru.prev != _base) { \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetch(&prev->_field); \
150 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
153 #ifdef ARCH_HAS_PREFETCHW
154 #define prefetchw_prev_lru_page(_page, _base, _field) \
156 if ((_page)->lru.prev != _base) { \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
164 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168 * From 0 .. 100. Higher means more swappy.
170 int vm_swappiness
= 60;
172 * The total number of pages which are beyond the high watermark within all
175 unsigned long vm_total_pages
;
177 static void set_task_reclaim_state(struct task_struct
*task
,
178 struct reclaim_state
*rs
)
180 /* Check for an overwrite */
181 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
183 /* Check for the nulling of an already-nulled member */
184 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
186 task
->reclaim_state
= rs
;
189 static LIST_HEAD(shrinker_list
);
190 static DECLARE_RWSEM(shrinker_rwsem
);
194 * We allow subsystems to populate their shrinker-related
195 * LRU lists before register_shrinker_prepared() is called
196 * for the shrinker, since we don't want to impose
197 * restrictions on their internal registration order.
198 * In this case shrink_slab_memcg() may find corresponding
199 * bit is set in the shrinkers map.
201 * This value is used by the function to detect registering
202 * shrinkers and to skip do_shrink_slab() calls for them.
204 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
206 static DEFINE_IDR(shrinker_idr
);
207 static int shrinker_nr_max
;
209 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
211 int id
, ret
= -ENOMEM
;
213 down_write(&shrinker_rwsem
);
214 /* This may call shrinker, so it must use down_read_trylock() */
215 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
219 if (id
>= shrinker_nr_max
) {
220 if (memcg_expand_shrinker_maps(id
)) {
221 idr_remove(&shrinker_idr
, id
);
225 shrinker_nr_max
= id
+ 1;
230 up_write(&shrinker_rwsem
);
234 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
236 int id
= shrinker
->id
;
240 down_write(&shrinker_rwsem
);
241 idr_remove(&shrinker_idr
, id
);
242 up_write(&shrinker_rwsem
);
245 static bool global_reclaim(struct scan_control
*sc
)
247 return !sc
->target_mem_cgroup
;
251 * sane_reclaim - is the usual dirty throttling mechanism operational?
252 * @sc: scan_control in question
254 * The normal page dirty throttling mechanism in balance_dirty_pages() is
255 * completely broken with the legacy memcg and direct stalling in
256 * shrink_page_list() is used for throttling instead, which lacks all the
257 * niceties such as fairness, adaptive pausing, bandwidth proportional
258 * allocation and configurability.
260 * This function tests whether the vmscan currently in progress can assume
261 * that the normal dirty throttling mechanism is operational.
263 static bool sane_reclaim(struct scan_control
*sc
)
265 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
269 #ifdef CONFIG_CGROUP_WRITEBACK
270 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
276 static void set_memcg_congestion(pg_data_t
*pgdat
,
277 struct mem_cgroup
*memcg
,
280 struct mem_cgroup_per_node
*mn
;
285 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
286 WRITE_ONCE(mn
->congested
, congested
);
289 static bool memcg_congested(pg_data_t
*pgdat
,
290 struct mem_cgroup
*memcg
)
292 struct mem_cgroup_per_node
*mn
;
294 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
295 return READ_ONCE(mn
->congested
);
299 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
304 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
308 static bool global_reclaim(struct scan_control
*sc
)
313 static bool sane_reclaim(struct scan_control
*sc
)
318 static inline void set_memcg_congestion(struct pglist_data
*pgdat
,
319 struct mem_cgroup
*memcg
, bool congested
)
323 static inline bool memcg_congested(struct pglist_data
*pgdat
,
324 struct mem_cgroup
*memcg
)
332 * This misses isolated pages which are not accounted for to save counters.
333 * As the data only determines if reclaim or compaction continues, it is
334 * not expected that isolated pages will be a dominating factor.
336 unsigned long zone_reclaimable_pages(struct zone
*zone
)
340 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
341 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
342 if (get_nr_swap_pages() > 0)
343 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
344 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
350 * lruvec_lru_size - Returns the number of pages on the given LRU list.
351 * @lruvec: lru vector
353 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
355 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
357 unsigned long lru_size
= 0;
360 if (!mem_cgroup_disabled()) {
361 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
362 lru_size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
364 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
366 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
367 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
370 if (!managed_zone(zone
))
373 if (!mem_cgroup_disabled())
374 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
376 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
377 NR_ZONE_LRU_BASE
+ lru
);
378 lru_size
-= min(size
, lru_size
);
386 * Add a shrinker callback to be called from the vm.
388 int prealloc_shrinker(struct shrinker
*shrinker
)
390 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
392 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
395 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
396 if (!shrinker
->nr_deferred
)
399 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
400 if (prealloc_memcg_shrinker(shrinker
))
407 kfree(shrinker
->nr_deferred
);
408 shrinker
->nr_deferred
= NULL
;
412 void free_prealloced_shrinker(struct shrinker
*shrinker
)
414 if (!shrinker
->nr_deferred
)
417 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
418 unregister_memcg_shrinker(shrinker
);
420 kfree(shrinker
->nr_deferred
);
421 shrinker
->nr_deferred
= NULL
;
424 void register_shrinker_prepared(struct shrinker
*shrinker
)
426 down_write(&shrinker_rwsem
);
427 list_add_tail(&shrinker
->list
, &shrinker_list
);
429 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
430 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
432 up_write(&shrinker_rwsem
);
435 int register_shrinker(struct shrinker
*shrinker
)
437 int err
= prealloc_shrinker(shrinker
);
441 register_shrinker_prepared(shrinker
);
444 EXPORT_SYMBOL(register_shrinker
);
449 void unregister_shrinker(struct shrinker
*shrinker
)
451 if (!shrinker
->nr_deferred
)
453 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
454 unregister_memcg_shrinker(shrinker
);
455 down_write(&shrinker_rwsem
);
456 list_del(&shrinker
->list
);
457 up_write(&shrinker_rwsem
);
458 kfree(shrinker
->nr_deferred
);
459 shrinker
->nr_deferred
= NULL
;
461 EXPORT_SYMBOL(unregister_shrinker
);
463 #define SHRINK_BATCH 128
465 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
466 struct shrinker
*shrinker
, int priority
)
468 unsigned long freed
= 0;
469 unsigned long long delta
;
474 int nid
= shrinkctl
->nid
;
475 long batch_size
= shrinker
->batch
? shrinker
->batch
477 long scanned
= 0, next_deferred
;
479 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
482 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
483 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
487 * copy the current shrinker scan count into a local variable
488 * and zero it so that other concurrent shrinker invocations
489 * don't also do this scanning work.
491 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
494 if (shrinker
->seeks
) {
495 delta
= freeable
>> priority
;
497 do_div(delta
, shrinker
->seeks
);
500 * These objects don't require any IO to create. Trim
501 * them aggressively under memory pressure to keep
502 * them from causing refetches in the IO caches.
504 delta
= freeable
/ 2;
508 if (total_scan
< 0) {
509 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
510 shrinker
->scan_objects
, total_scan
);
511 total_scan
= freeable
;
514 next_deferred
= total_scan
;
517 * We need to avoid excessive windup on filesystem shrinkers
518 * due to large numbers of GFP_NOFS allocations causing the
519 * shrinkers to return -1 all the time. This results in a large
520 * nr being built up so when a shrink that can do some work
521 * comes along it empties the entire cache due to nr >>>
522 * freeable. This is bad for sustaining a working set in
525 * Hence only allow the shrinker to scan the entire cache when
526 * a large delta change is calculated directly.
528 if (delta
< freeable
/ 4)
529 total_scan
= min(total_scan
, freeable
/ 2);
532 * Avoid risking looping forever due to too large nr value:
533 * never try to free more than twice the estimate number of
536 if (total_scan
> freeable
* 2)
537 total_scan
= freeable
* 2;
539 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
540 freeable
, delta
, total_scan
, priority
);
543 * Normally, we should not scan less than batch_size objects in one
544 * pass to avoid too frequent shrinker calls, but if the slab has less
545 * than batch_size objects in total and we are really tight on memory,
546 * we will try to reclaim all available objects, otherwise we can end
547 * up failing allocations although there are plenty of reclaimable
548 * objects spread over several slabs with usage less than the
551 * We detect the "tight on memory" situations by looking at the total
552 * number of objects we want to scan (total_scan). If it is greater
553 * than the total number of objects on slab (freeable), we must be
554 * scanning at high prio and therefore should try to reclaim as much as
557 while (total_scan
>= batch_size
||
558 total_scan
>= freeable
) {
560 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
562 shrinkctl
->nr_to_scan
= nr_to_scan
;
563 shrinkctl
->nr_scanned
= nr_to_scan
;
564 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
565 if (ret
== SHRINK_STOP
)
569 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
570 total_scan
-= shrinkctl
->nr_scanned
;
571 scanned
+= shrinkctl
->nr_scanned
;
576 if (next_deferred
>= scanned
)
577 next_deferred
-= scanned
;
581 * move the unused scan count back into the shrinker in a
582 * manner that handles concurrent updates. If we exhausted the
583 * scan, there is no need to do an update.
585 if (next_deferred
> 0)
586 new_nr
= atomic_long_add_return(next_deferred
,
587 &shrinker
->nr_deferred
[nid
]);
589 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
591 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
596 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
597 struct mem_cgroup
*memcg
, int priority
)
599 struct memcg_shrinker_map
*map
;
600 unsigned long ret
, freed
= 0;
603 if (!mem_cgroup_online(memcg
))
606 if (!down_read_trylock(&shrinker_rwsem
))
609 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
614 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
615 struct shrink_control sc
= {
616 .gfp_mask
= gfp_mask
,
620 struct shrinker
*shrinker
;
622 shrinker
= idr_find(&shrinker_idr
, i
);
623 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
625 clear_bit(i
, map
->map
);
629 /* Call non-slab shrinkers even though kmem is disabled */
630 if (!memcg_kmem_enabled() &&
631 !(shrinker
->flags
& SHRINKER_NONSLAB
))
634 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
635 if (ret
== SHRINK_EMPTY
) {
636 clear_bit(i
, map
->map
);
638 * After the shrinker reported that it had no objects to
639 * free, but before we cleared the corresponding bit in
640 * the memcg shrinker map, a new object might have been
641 * added. To make sure, we have the bit set in this
642 * case, we invoke the shrinker one more time and reset
643 * the bit if it reports that it is not empty anymore.
644 * The memory barrier here pairs with the barrier in
645 * memcg_set_shrinker_bit():
647 * list_lru_add() shrink_slab_memcg()
648 * list_add_tail() clear_bit()
650 * set_bit() do_shrink_slab()
652 smp_mb__after_atomic();
653 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
654 if (ret
== SHRINK_EMPTY
)
657 memcg_set_shrinker_bit(memcg
, nid
, i
);
661 if (rwsem_is_contended(&shrinker_rwsem
)) {
667 up_read(&shrinker_rwsem
);
670 #else /* CONFIG_MEMCG */
671 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
672 struct mem_cgroup
*memcg
, int priority
)
676 #endif /* CONFIG_MEMCG */
679 * shrink_slab - shrink slab caches
680 * @gfp_mask: allocation context
681 * @nid: node whose slab caches to target
682 * @memcg: memory cgroup whose slab caches to target
683 * @priority: the reclaim priority
685 * Call the shrink functions to age shrinkable caches.
687 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
688 * unaware shrinkers will receive a node id of 0 instead.
690 * @memcg specifies the memory cgroup to target. Unaware shrinkers
691 * are called only if it is the root cgroup.
693 * @priority is sc->priority, we take the number of objects and >> by priority
694 * in order to get the scan target.
696 * Returns the number of reclaimed slab objects.
698 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
699 struct mem_cgroup
*memcg
,
702 unsigned long ret
, freed
= 0;
703 struct shrinker
*shrinker
;
706 * The root memcg might be allocated even though memcg is disabled
707 * via "cgroup_disable=memory" boot parameter. This could make
708 * mem_cgroup_is_root() return false, then just run memcg slab
709 * shrink, but skip global shrink. This may result in premature
712 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
713 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
715 if (!down_read_trylock(&shrinker_rwsem
))
718 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
719 struct shrink_control sc
= {
720 .gfp_mask
= gfp_mask
,
725 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
726 if (ret
== SHRINK_EMPTY
)
730 * Bail out if someone want to register a new shrinker to
731 * prevent the regsitration from being stalled for long periods
732 * by parallel ongoing shrinking.
734 if (rwsem_is_contended(&shrinker_rwsem
)) {
740 up_read(&shrinker_rwsem
);
746 void drop_slab_node(int nid
)
751 struct mem_cgroup
*memcg
= NULL
;
754 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
756 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
757 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
758 } while (freed
> 10);
765 for_each_online_node(nid
)
769 static inline int is_page_cache_freeable(struct page
*page
)
772 * A freeable page cache page is referenced only by the caller
773 * that isolated the page, the page cache and optional buffer
774 * heads at page->private.
776 int page_cache_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
778 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
781 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
783 if (current
->flags
& PF_SWAPWRITE
)
785 if (!inode_write_congested(inode
))
787 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
793 * We detected a synchronous write error writing a page out. Probably
794 * -ENOSPC. We need to propagate that into the address_space for a subsequent
795 * fsync(), msync() or close().
797 * The tricky part is that after writepage we cannot touch the mapping: nothing
798 * prevents it from being freed up. But we have a ref on the page and once
799 * that page is locked, the mapping is pinned.
801 * We're allowed to run sleeping lock_page() here because we know the caller has
804 static void handle_write_error(struct address_space
*mapping
,
805 struct page
*page
, int error
)
808 if (page_mapping(page
) == mapping
)
809 mapping_set_error(mapping
, error
);
813 /* possible outcome of pageout() */
815 /* failed to write page out, page is locked */
817 /* move page to the active list, page is locked */
819 /* page has been sent to the disk successfully, page is unlocked */
821 /* page is clean and locked */
826 * pageout is called by shrink_page_list() for each dirty page.
827 * Calls ->writepage().
829 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
830 struct scan_control
*sc
)
833 * If the page is dirty, only perform writeback if that write
834 * will be non-blocking. To prevent this allocation from being
835 * stalled by pagecache activity. But note that there may be
836 * stalls if we need to run get_block(). We could test
837 * PagePrivate for that.
839 * If this process is currently in __generic_file_write_iter() against
840 * this page's queue, we can perform writeback even if that
843 * If the page is swapcache, write it back even if that would
844 * block, for some throttling. This happens by accident, because
845 * swap_backing_dev_info is bust: it doesn't reflect the
846 * congestion state of the swapdevs. Easy to fix, if needed.
848 if (!is_page_cache_freeable(page
))
852 * Some data journaling orphaned pages can have
853 * page->mapping == NULL while being dirty with clean buffers.
855 if (page_has_private(page
)) {
856 if (try_to_free_buffers(page
)) {
857 ClearPageDirty(page
);
858 pr_info("%s: orphaned page\n", __func__
);
864 if (mapping
->a_ops
->writepage
== NULL
)
865 return PAGE_ACTIVATE
;
866 if (!may_write_to_inode(mapping
->host
, sc
))
869 if (clear_page_dirty_for_io(page
)) {
871 struct writeback_control wbc
= {
872 .sync_mode
= WB_SYNC_NONE
,
873 .nr_to_write
= SWAP_CLUSTER_MAX
,
875 .range_end
= LLONG_MAX
,
879 SetPageReclaim(page
);
880 res
= mapping
->a_ops
->writepage(page
, &wbc
);
882 handle_write_error(mapping
, page
, res
);
883 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
884 ClearPageReclaim(page
);
885 return PAGE_ACTIVATE
;
888 if (!PageWriteback(page
)) {
889 /* synchronous write or broken a_ops? */
890 ClearPageReclaim(page
);
892 trace_mm_vmscan_writepage(page
);
893 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
901 * Same as remove_mapping, but if the page is removed from the mapping, it
902 * gets returned with a refcount of 0.
904 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
910 BUG_ON(!PageLocked(page
));
911 BUG_ON(mapping
!= page_mapping(page
));
913 xa_lock_irqsave(&mapping
->i_pages
, flags
);
915 * The non racy check for a busy page.
917 * Must be careful with the order of the tests. When someone has
918 * a ref to the page, it may be possible that they dirty it then
919 * drop the reference. So if PageDirty is tested before page_count
920 * here, then the following race may occur:
922 * get_user_pages(&page);
923 * [user mapping goes away]
925 * !PageDirty(page) [good]
926 * SetPageDirty(page);
928 * !page_count(page) [good, discard it]
930 * [oops, our write_to data is lost]
932 * Reversing the order of the tests ensures such a situation cannot
933 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
934 * load is not satisfied before that of page->_refcount.
936 * Note that if SetPageDirty is always performed via set_page_dirty,
937 * and thus under the i_pages lock, then this ordering is not required.
939 refcount
= 1 + compound_nr(page
);
940 if (!page_ref_freeze(page
, refcount
))
942 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
943 if (unlikely(PageDirty(page
))) {
944 page_ref_unfreeze(page
, refcount
);
948 if (PageSwapCache(page
)) {
949 swp_entry_t swap
= { .val
= page_private(page
) };
950 mem_cgroup_swapout(page
, swap
);
951 __delete_from_swap_cache(page
, swap
);
952 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
953 put_swap_page(page
, swap
);
955 void (*freepage
)(struct page
*);
958 freepage
= mapping
->a_ops
->freepage
;
960 * Remember a shadow entry for reclaimed file cache in
961 * order to detect refaults, thus thrashing, later on.
963 * But don't store shadows in an address space that is
964 * already exiting. This is not just an optizimation,
965 * inode reclaim needs to empty out the radix tree or
966 * the nodes are lost. Don't plant shadows behind its
969 * We also don't store shadows for DAX mappings because the
970 * only page cache pages found in these are zero pages
971 * covering holes, and because we don't want to mix DAX
972 * exceptional entries and shadow exceptional entries in the
973 * same address_space.
975 if (reclaimed
&& page_is_file_cache(page
) &&
976 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
977 shadow
= workingset_eviction(page
);
978 __delete_from_page_cache(page
, shadow
);
979 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
981 if (freepage
!= NULL
)
988 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
993 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
994 * someone else has a ref on the page, abort and return 0. If it was
995 * successfully detached, return 1. Assumes the caller has a single ref on
998 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
1000 if (__remove_mapping(mapping
, page
, false)) {
1002 * Unfreezing the refcount with 1 rather than 2 effectively
1003 * drops the pagecache ref for us without requiring another
1006 page_ref_unfreeze(page
, 1);
1013 * putback_lru_page - put previously isolated page onto appropriate LRU list
1014 * @page: page to be put back to appropriate lru list
1016 * Add previously isolated @page to appropriate LRU list.
1017 * Page may still be unevictable for other reasons.
1019 * lru_lock must not be held, interrupts must be enabled.
1021 void putback_lru_page(struct page
*page
)
1023 lru_cache_add(page
);
1024 put_page(page
); /* drop ref from isolate */
1027 enum page_references
{
1029 PAGEREF_RECLAIM_CLEAN
,
1034 static enum page_references
page_check_references(struct page
*page
,
1035 struct scan_control
*sc
)
1037 int referenced_ptes
, referenced_page
;
1038 unsigned long vm_flags
;
1040 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1042 referenced_page
= TestClearPageReferenced(page
);
1045 * Mlock lost the isolation race with us. Let try_to_unmap()
1046 * move the page to the unevictable list.
1048 if (vm_flags
& VM_LOCKED
)
1049 return PAGEREF_RECLAIM
;
1051 if (referenced_ptes
) {
1052 if (PageSwapBacked(page
))
1053 return PAGEREF_ACTIVATE
;
1055 * All mapped pages start out with page table
1056 * references from the instantiating fault, so we need
1057 * to look twice if a mapped file page is used more
1060 * Mark it and spare it for another trip around the
1061 * inactive list. Another page table reference will
1062 * lead to its activation.
1064 * Note: the mark is set for activated pages as well
1065 * so that recently deactivated but used pages are
1066 * quickly recovered.
1068 SetPageReferenced(page
);
1070 if (referenced_page
|| referenced_ptes
> 1)
1071 return PAGEREF_ACTIVATE
;
1074 * Activate file-backed executable pages after first usage.
1076 if (vm_flags
& VM_EXEC
)
1077 return PAGEREF_ACTIVATE
;
1079 return PAGEREF_KEEP
;
1082 /* Reclaim if clean, defer dirty pages to writeback */
1083 if (referenced_page
&& !PageSwapBacked(page
))
1084 return PAGEREF_RECLAIM_CLEAN
;
1086 return PAGEREF_RECLAIM
;
1089 /* Check if a page is dirty or under writeback */
1090 static void page_check_dirty_writeback(struct page
*page
,
1091 bool *dirty
, bool *writeback
)
1093 struct address_space
*mapping
;
1096 * Anonymous pages are not handled by flushers and must be written
1097 * from reclaim context. Do not stall reclaim based on them
1099 if (!page_is_file_cache(page
) ||
1100 (PageAnon(page
) && !PageSwapBacked(page
))) {
1106 /* By default assume that the page flags are accurate */
1107 *dirty
= PageDirty(page
);
1108 *writeback
= PageWriteback(page
);
1110 /* Verify dirty/writeback state if the filesystem supports it */
1111 if (!page_has_private(page
))
1114 mapping
= page_mapping(page
);
1115 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1116 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1120 * shrink_page_list() returns the number of reclaimed pages
1122 static unsigned long shrink_page_list(struct list_head
*page_list
,
1123 struct pglist_data
*pgdat
,
1124 struct scan_control
*sc
,
1125 enum ttu_flags ttu_flags
,
1126 struct reclaim_stat
*stat
,
1127 bool ignore_references
)
1129 LIST_HEAD(ret_pages
);
1130 LIST_HEAD(free_pages
);
1131 unsigned nr_reclaimed
= 0;
1132 unsigned pgactivate
= 0;
1134 memset(stat
, 0, sizeof(*stat
));
1137 while (!list_empty(page_list
)) {
1138 struct address_space
*mapping
;
1141 enum page_references references
= PAGEREF_RECLAIM
;
1142 bool dirty
, writeback
;
1143 unsigned int nr_pages
;
1147 page
= lru_to_page(page_list
);
1148 list_del(&page
->lru
);
1150 if (!trylock_page(page
))
1153 VM_BUG_ON_PAGE(PageActive(page
), page
);
1155 nr_pages
= compound_nr(page
);
1157 /* Account the number of base pages even though THP */
1158 sc
->nr_scanned
+= nr_pages
;
1160 if (unlikely(!page_evictable(page
)))
1161 goto activate_locked
;
1163 if (!sc
->may_unmap
&& page_mapped(page
))
1166 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1167 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1170 * The number of dirty pages determines if a node is marked
1171 * reclaim_congested which affects wait_iff_congested. kswapd
1172 * will stall and start writing pages if the tail of the LRU
1173 * is all dirty unqueued pages.
1175 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1176 if (dirty
|| writeback
)
1179 if (dirty
&& !writeback
)
1180 stat
->nr_unqueued_dirty
++;
1183 * Treat this page as congested if the underlying BDI is or if
1184 * pages are cycling through the LRU so quickly that the
1185 * pages marked for immediate reclaim are making it to the
1186 * end of the LRU a second time.
1188 mapping
= page_mapping(page
);
1189 if (((dirty
|| writeback
) && mapping
&&
1190 inode_write_congested(mapping
->host
)) ||
1191 (writeback
&& PageReclaim(page
)))
1192 stat
->nr_congested
++;
1195 * If a page at the tail of the LRU is under writeback, there
1196 * are three cases to consider.
1198 * 1) If reclaim is encountering an excessive number of pages
1199 * under writeback and this page is both under writeback and
1200 * PageReclaim then it indicates that pages are being queued
1201 * for IO but are being recycled through the LRU before the
1202 * IO can complete. Waiting on the page itself risks an
1203 * indefinite stall if it is impossible to writeback the
1204 * page due to IO error or disconnected storage so instead
1205 * note that the LRU is being scanned too quickly and the
1206 * caller can stall after page list has been processed.
1208 * 2) Global or new memcg reclaim encounters a page that is
1209 * not marked for immediate reclaim, or the caller does not
1210 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1211 * not to fs). In this case mark the page for immediate
1212 * reclaim and continue scanning.
1214 * Require may_enter_fs because we would wait on fs, which
1215 * may not have submitted IO yet. And the loop driver might
1216 * enter reclaim, and deadlock if it waits on a page for
1217 * which it is needed to do the write (loop masks off
1218 * __GFP_IO|__GFP_FS for this reason); but more thought
1219 * would probably show more reasons.
1221 * 3) Legacy memcg encounters a page that is already marked
1222 * PageReclaim. memcg does not have any dirty pages
1223 * throttling so we could easily OOM just because too many
1224 * pages are in writeback and there is nothing else to
1225 * reclaim. Wait for the writeback to complete.
1227 * In cases 1) and 2) we activate the pages to get them out of
1228 * the way while we continue scanning for clean pages on the
1229 * inactive list and refilling from the active list. The
1230 * observation here is that waiting for disk writes is more
1231 * expensive than potentially causing reloads down the line.
1232 * Since they're marked for immediate reclaim, they won't put
1233 * memory pressure on the cache working set any longer than it
1234 * takes to write them to disk.
1236 if (PageWriteback(page
)) {
1238 if (current_is_kswapd() &&
1239 PageReclaim(page
) &&
1240 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1241 stat
->nr_immediate
++;
1242 goto activate_locked
;
1245 } else if (sane_reclaim(sc
) ||
1246 !PageReclaim(page
) || !may_enter_fs
) {
1248 * This is slightly racy - end_page_writeback()
1249 * might have just cleared PageReclaim, then
1250 * setting PageReclaim here end up interpreted
1251 * as PageReadahead - but that does not matter
1252 * enough to care. What we do want is for this
1253 * page to have PageReclaim set next time memcg
1254 * reclaim reaches the tests above, so it will
1255 * then wait_on_page_writeback() to avoid OOM;
1256 * and it's also appropriate in global reclaim.
1258 SetPageReclaim(page
);
1259 stat
->nr_writeback
++;
1260 goto activate_locked
;
1265 wait_on_page_writeback(page
);
1266 /* then go back and try same page again */
1267 list_add_tail(&page
->lru
, page_list
);
1272 if (!ignore_references
)
1273 references
= page_check_references(page
, sc
);
1275 switch (references
) {
1276 case PAGEREF_ACTIVATE
:
1277 goto activate_locked
;
1279 stat
->nr_ref_keep
+= nr_pages
;
1281 case PAGEREF_RECLAIM
:
1282 case PAGEREF_RECLAIM_CLEAN
:
1283 ; /* try to reclaim the page below */
1287 * Anonymous process memory has backing store?
1288 * Try to allocate it some swap space here.
1289 * Lazyfree page could be freed directly
1291 if (PageAnon(page
) && PageSwapBacked(page
)) {
1292 if (!PageSwapCache(page
)) {
1293 if (!(sc
->gfp_mask
& __GFP_IO
))
1295 if (PageTransHuge(page
)) {
1296 /* cannot split THP, skip it */
1297 if (!can_split_huge_page(page
, NULL
))
1298 goto activate_locked
;
1300 * Split pages without a PMD map right
1301 * away. Chances are some or all of the
1302 * tail pages can be freed without IO.
1304 if (!compound_mapcount(page
) &&
1305 split_huge_page_to_list(page
,
1307 goto activate_locked
;
1309 if (!add_to_swap(page
)) {
1310 if (!PageTransHuge(page
))
1311 goto activate_locked_split
;
1312 /* Fallback to swap normal pages */
1313 if (split_huge_page_to_list(page
,
1315 goto activate_locked
;
1316 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1317 count_vm_event(THP_SWPOUT_FALLBACK
);
1319 if (!add_to_swap(page
))
1320 goto activate_locked_split
;
1325 /* Adding to swap updated mapping */
1326 mapping
= page_mapping(page
);
1328 } else if (unlikely(PageTransHuge(page
))) {
1329 /* Split file THP */
1330 if (split_huge_page_to_list(page
, page_list
))
1335 * THP may get split above, need minus tail pages and update
1336 * nr_pages to avoid accounting tail pages twice.
1338 * The tail pages that are added into swap cache successfully
1341 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1342 sc
->nr_scanned
-= (nr_pages
- 1);
1347 * The page is mapped into the page tables of one or more
1348 * processes. Try to unmap it here.
1350 if (page_mapped(page
)) {
1351 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1353 if (unlikely(PageTransHuge(page
)))
1354 flags
|= TTU_SPLIT_HUGE_PMD
;
1355 if (!try_to_unmap(page
, flags
)) {
1356 stat
->nr_unmap_fail
+= nr_pages
;
1357 goto activate_locked
;
1361 if (PageDirty(page
)) {
1363 * Only kswapd can writeback filesystem pages
1364 * to avoid risk of stack overflow. But avoid
1365 * injecting inefficient single-page IO into
1366 * flusher writeback as much as possible: only
1367 * write pages when we've encountered many
1368 * dirty pages, and when we've already scanned
1369 * the rest of the LRU for clean pages and see
1370 * the same dirty pages again (PageReclaim).
1372 if (page_is_file_cache(page
) &&
1373 (!current_is_kswapd() || !PageReclaim(page
) ||
1374 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1376 * Immediately reclaim when written back.
1377 * Similar in principal to deactivate_page()
1378 * except we already have the page isolated
1379 * and know it's dirty
1381 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1382 SetPageReclaim(page
);
1384 goto activate_locked
;
1387 if (references
== PAGEREF_RECLAIM_CLEAN
)
1391 if (!sc
->may_writepage
)
1395 * Page is dirty. Flush the TLB if a writable entry
1396 * potentially exists to avoid CPU writes after IO
1397 * starts and then write it out here.
1399 try_to_unmap_flush_dirty();
1400 switch (pageout(page
, mapping
, sc
)) {
1404 goto activate_locked
;
1406 if (PageWriteback(page
))
1408 if (PageDirty(page
))
1412 * A synchronous write - probably a ramdisk. Go
1413 * ahead and try to reclaim the page.
1415 if (!trylock_page(page
))
1417 if (PageDirty(page
) || PageWriteback(page
))
1419 mapping
= page_mapping(page
);
1421 ; /* try to free the page below */
1426 * If the page has buffers, try to free the buffer mappings
1427 * associated with this page. If we succeed we try to free
1430 * We do this even if the page is PageDirty().
1431 * try_to_release_page() does not perform I/O, but it is
1432 * possible for a page to have PageDirty set, but it is actually
1433 * clean (all its buffers are clean). This happens if the
1434 * buffers were written out directly, with submit_bh(). ext3
1435 * will do this, as well as the blockdev mapping.
1436 * try_to_release_page() will discover that cleanness and will
1437 * drop the buffers and mark the page clean - it can be freed.
1439 * Rarely, pages can have buffers and no ->mapping. These are
1440 * the pages which were not successfully invalidated in
1441 * truncate_complete_page(). We try to drop those buffers here
1442 * and if that worked, and the page is no longer mapped into
1443 * process address space (page_count == 1) it can be freed.
1444 * Otherwise, leave the page on the LRU so it is swappable.
1446 if (page_has_private(page
)) {
1447 if (!try_to_release_page(page
, sc
->gfp_mask
))
1448 goto activate_locked
;
1449 if (!mapping
&& page_count(page
) == 1) {
1451 if (put_page_testzero(page
))
1455 * rare race with speculative reference.
1456 * the speculative reference will free
1457 * this page shortly, so we may
1458 * increment nr_reclaimed here (and
1459 * leave it off the LRU).
1467 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1468 /* follow __remove_mapping for reference */
1469 if (!page_ref_freeze(page
, 1))
1471 if (PageDirty(page
)) {
1472 page_ref_unfreeze(page
, 1);
1476 count_vm_event(PGLAZYFREED
);
1477 count_memcg_page_event(page
, PGLAZYFREED
);
1478 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1484 * THP may get swapped out in a whole, need account
1487 nr_reclaimed
+= nr_pages
;
1490 * Is there need to periodically free_page_list? It would
1491 * appear not as the counts should be low
1493 if (unlikely(PageTransHuge(page
)))
1494 (*get_compound_page_dtor(page
))(page
);
1496 list_add(&page
->lru
, &free_pages
);
1499 activate_locked_split
:
1501 * The tail pages that are failed to add into swap cache
1502 * reach here. Fixup nr_scanned and nr_pages.
1505 sc
->nr_scanned
-= (nr_pages
- 1);
1509 /* Not a candidate for swapping, so reclaim swap space. */
1510 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1512 try_to_free_swap(page
);
1513 VM_BUG_ON_PAGE(PageActive(page
), page
);
1514 if (!PageMlocked(page
)) {
1515 int type
= page_is_file_cache(page
);
1516 SetPageActive(page
);
1517 stat
->nr_activate
[type
] += nr_pages
;
1518 count_memcg_page_event(page
, PGACTIVATE
);
1523 list_add(&page
->lru
, &ret_pages
);
1524 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1527 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1529 mem_cgroup_uncharge_list(&free_pages
);
1530 try_to_unmap_flush();
1531 free_unref_page_list(&free_pages
);
1533 list_splice(&ret_pages
, page_list
);
1534 count_vm_events(PGACTIVATE
, pgactivate
);
1536 return nr_reclaimed
;
1539 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1540 struct list_head
*page_list
)
1542 struct scan_control sc
= {
1543 .gfp_mask
= GFP_KERNEL
,
1544 .priority
= DEF_PRIORITY
,
1547 struct reclaim_stat dummy_stat
;
1549 struct page
*page
, *next
;
1550 LIST_HEAD(clean_pages
);
1552 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1553 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1554 !__PageMovable(page
) && !PageUnevictable(page
)) {
1555 ClearPageActive(page
);
1556 list_move(&page
->lru
, &clean_pages
);
1560 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1561 TTU_IGNORE_ACCESS
, &dummy_stat
, true);
1562 list_splice(&clean_pages
, page_list
);
1563 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1568 * Attempt to remove the specified page from its LRU. Only take this page
1569 * if it is of the appropriate PageActive status. Pages which are being
1570 * freed elsewhere are also ignored.
1572 * page: page to consider
1573 * mode: one of the LRU isolation modes defined above
1575 * returns 0 on success, -ve errno on failure.
1577 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1581 /* Only take pages on the LRU. */
1585 /* Compaction should not handle unevictable pages but CMA can do so */
1586 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1592 * To minimise LRU disruption, the caller can indicate that it only
1593 * wants to isolate pages it will be able to operate on without
1594 * blocking - clean pages for the most part.
1596 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1597 * that it is possible to migrate without blocking
1599 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1600 /* All the caller can do on PageWriteback is block */
1601 if (PageWriteback(page
))
1604 if (PageDirty(page
)) {
1605 struct address_space
*mapping
;
1609 * Only pages without mappings or that have a
1610 * ->migratepage callback are possible to migrate
1611 * without blocking. However, we can be racing with
1612 * truncation so it's necessary to lock the page
1613 * to stabilise the mapping as truncation holds
1614 * the page lock until after the page is removed
1615 * from the page cache.
1617 if (!trylock_page(page
))
1620 mapping
= page_mapping(page
);
1621 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1628 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1631 if (likely(get_page_unless_zero(page
))) {
1633 * Be careful not to clear PageLRU until after we're
1634 * sure the page is not being freed elsewhere -- the
1635 * page release code relies on it.
1646 * Update LRU sizes after isolating pages. The LRU size updates must
1647 * be complete before mem_cgroup_update_lru_size due to a santity check.
1649 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1650 enum lru_list lru
, unsigned long *nr_zone_taken
)
1654 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1655 if (!nr_zone_taken
[zid
])
1658 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1660 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1667 * pgdat->lru_lock is heavily contended. Some of the functions that
1668 * shrink the lists perform better by taking out a batch of pages
1669 * and working on them outside the LRU lock.
1671 * For pagecache intensive workloads, this function is the hottest
1672 * spot in the kernel (apart from copy_*_user functions).
1674 * Appropriate locks must be held before calling this function.
1676 * @nr_to_scan: The number of eligible pages to look through on the list.
1677 * @lruvec: The LRU vector to pull pages from.
1678 * @dst: The temp list to put pages on to.
1679 * @nr_scanned: The number of pages that were scanned.
1680 * @sc: The scan_control struct for this reclaim session
1681 * @mode: One of the LRU isolation modes
1682 * @lru: LRU list id for isolating
1684 * returns how many pages were moved onto *@dst.
1686 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1687 struct lruvec
*lruvec
, struct list_head
*dst
,
1688 unsigned long *nr_scanned
, struct scan_control
*sc
,
1691 struct list_head
*src
= &lruvec
->lists
[lru
];
1692 unsigned long nr_taken
= 0;
1693 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1694 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1695 unsigned long skipped
= 0;
1696 unsigned long scan
, total_scan
, nr_pages
;
1697 LIST_HEAD(pages_skipped
);
1698 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1702 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1705 page
= lru_to_page(src
);
1706 prefetchw_prev_lru_page(page
, src
, flags
);
1708 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1710 nr_pages
= compound_nr(page
);
1711 total_scan
+= nr_pages
;
1713 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1714 list_move(&page
->lru
, &pages_skipped
);
1715 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1720 * Do not count skipped pages because that makes the function
1721 * return with no isolated pages if the LRU mostly contains
1722 * ineligible pages. This causes the VM to not reclaim any
1723 * pages, triggering a premature OOM.
1725 * Account all tail pages of THP. This would not cause
1726 * premature OOM since __isolate_lru_page() returns -EBUSY
1727 * only when the page is being freed somewhere else.
1730 switch (__isolate_lru_page(page
, mode
)) {
1732 nr_taken
+= nr_pages
;
1733 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1734 list_move(&page
->lru
, dst
);
1738 /* else it is being freed elsewhere */
1739 list_move(&page
->lru
, src
);
1748 * Splice any skipped pages to the start of the LRU list. Note that
1749 * this disrupts the LRU order when reclaiming for lower zones but
1750 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1751 * scanning would soon rescan the same pages to skip and put the
1752 * system at risk of premature OOM.
1754 if (!list_empty(&pages_skipped
)) {
1757 list_splice(&pages_skipped
, src
);
1758 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1759 if (!nr_skipped
[zid
])
1762 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1763 skipped
+= nr_skipped
[zid
];
1766 *nr_scanned
= total_scan
;
1767 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1768 total_scan
, skipped
, nr_taken
, mode
, lru
);
1769 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1774 * isolate_lru_page - tries to isolate a page from its LRU list
1775 * @page: page to isolate from its LRU list
1777 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1778 * vmstat statistic corresponding to whatever LRU list the page was on.
1780 * Returns 0 if the page was removed from an LRU list.
1781 * Returns -EBUSY if the page was not on an LRU list.
1783 * The returned page will have PageLRU() cleared. If it was found on
1784 * the active list, it will have PageActive set. If it was found on
1785 * the unevictable list, it will have the PageUnevictable bit set. That flag
1786 * may need to be cleared by the caller before letting the page go.
1788 * The vmstat statistic corresponding to the list on which the page was
1789 * found will be decremented.
1793 * (1) Must be called with an elevated refcount on the page. This is a
1794 * fundamentnal difference from isolate_lru_pages (which is called
1795 * without a stable reference).
1796 * (2) the lru_lock must not be held.
1797 * (3) interrupts must be enabled.
1799 int isolate_lru_page(struct page
*page
)
1803 VM_BUG_ON_PAGE(!page_count(page
), page
);
1804 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1806 if (PageLRU(page
)) {
1807 pg_data_t
*pgdat
= page_pgdat(page
);
1808 struct lruvec
*lruvec
;
1810 spin_lock_irq(&pgdat
->lru_lock
);
1811 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1812 if (PageLRU(page
)) {
1813 int lru
= page_lru(page
);
1816 del_page_from_lru_list(page
, lruvec
, lru
);
1819 spin_unlock_irq(&pgdat
->lru_lock
);
1825 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1826 * then get resheduled. When there are massive number of tasks doing page
1827 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1828 * the LRU list will go small and be scanned faster than necessary, leading to
1829 * unnecessary swapping, thrashing and OOM.
1831 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1832 struct scan_control
*sc
)
1834 unsigned long inactive
, isolated
;
1836 if (current_is_kswapd())
1839 if (!sane_reclaim(sc
))
1843 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1844 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1846 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1847 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1851 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1852 * won't get blocked by normal direct-reclaimers, forming a circular
1855 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1858 return isolated
> inactive
;
1862 * This moves pages from @list to corresponding LRU list.
1864 * We move them the other way if the page is referenced by one or more
1865 * processes, from rmap.
1867 * If the pages are mostly unmapped, the processing is fast and it is
1868 * appropriate to hold zone_lru_lock across the whole operation. But if
1869 * the pages are mapped, the processing is slow (page_referenced()) so we
1870 * should drop zone_lru_lock around each page. It's impossible to balance
1871 * this, so instead we remove the pages from the LRU while processing them.
1872 * It is safe to rely on PG_active against the non-LRU pages in here because
1873 * nobody will play with that bit on a non-LRU page.
1875 * The downside is that we have to touch page->_refcount against each page.
1876 * But we had to alter page->flags anyway.
1878 * Returns the number of pages moved to the given lruvec.
1881 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1882 struct list_head
*list
)
1884 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1885 int nr_pages
, nr_moved
= 0;
1886 LIST_HEAD(pages_to_free
);
1890 while (!list_empty(list
)) {
1891 page
= lru_to_page(list
);
1892 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1893 if (unlikely(!page_evictable(page
))) {
1894 list_del(&page
->lru
);
1895 spin_unlock_irq(&pgdat
->lru_lock
);
1896 putback_lru_page(page
);
1897 spin_lock_irq(&pgdat
->lru_lock
);
1900 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1903 lru
= page_lru(page
);
1905 nr_pages
= hpage_nr_pages(page
);
1906 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1907 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1909 if (put_page_testzero(page
)) {
1910 __ClearPageLRU(page
);
1911 __ClearPageActive(page
);
1912 del_page_from_lru_list(page
, lruvec
, lru
);
1914 if (unlikely(PageCompound(page
))) {
1915 spin_unlock_irq(&pgdat
->lru_lock
);
1916 (*get_compound_page_dtor(page
))(page
);
1917 spin_lock_irq(&pgdat
->lru_lock
);
1919 list_add(&page
->lru
, &pages_to_free
);
1921 nr_moved
+= nr_pages
;
1926 * To save our caller's stack, now use input list for pages to free.
1928 list_splice(&pages_to_free
, list
);
1934 * If a kernel thread (such as nfsd for loop-back mounts) services
1935 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1936 * In that case we should only throttle if the backing device it is
1937 * writing to is congested. In other cases it is safe to throttle.
1939 static int current_may_throttle(void)
1941 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1942 current
->backing_dev_info
== NULL
||
1943 bdi_write_congested(current
->backing_dev_info
);
1947 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1948 * of reclaimed pages
1950 static noinline_for_stack
unsigned long
1951 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1952 struct scan_control
*sc
, enum lru_list lru
)
1954 LIST_HEAD(page_list
);
1955 unsigned long nr_scanned
;
1956 unsigned long nr_reclaimed
= 0;
1957 unsigned long nr_taken
;
1958 struct reclaim_stat stat
;
1959 int file
= is_file_lru(lru
);
1960 enum vm_event_item item
;
1961 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1962 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1963 bool stalled
= false;
1965 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1969 /* wait a bit for the reclaimer. */
1973 /* We are about to die and free our memory. Return now. */
1974 if (fatal_signal_pending(current
))
1975 return SWAP_CLUSTER_MAX
;
1980 spin_lock_irq(&pgdat
->lru_lock
);
1982 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1983 &nr_scanned
, sc
, lru
);
1985 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1986 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1988 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1989 if (global_reclaim(sc
))
1990 __count_vm_events(item
, nr_scanned
);
1991 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1992 spin_unlock_irq(&pgdat
->lru_lock
);
1997 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
2000 spin_lock_irq(&pgdat
->lru_lock
);
2002 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2003 if (global_reclaim(sc
))
2004 __count_vm_events(item
, nr_reclaimed
);
2005 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2006 reclaim_stat
->recent_rotated
[0] += stat
.nr_activate
[0];
2007 reclaim_stat
->recent_rotated
[1] += stat
.nr_activate
[1];
2009 move_pages_to_lru(lruvec
, &page_list
);
2011 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2013 spin_unlock_irq(&pgdat
->lru_lock
);
2015 mem_cgroup_uncharge_list(&page_list
);
2016 free_unref_page_list(&page_list
);
2019 * If dirty pages are scanned that are not queued for IO, it
2020 * implies that flushers are not doing their job. This can
2021 * happen when memory pressure pushes dirty pages to the end of
2022 * the LRU before the dirty limits are breached and the dirty
2023 * data has expired. It can also happen when the proportion of
2024 * dirty pages grows not through writes but through memory
2025 * pressure reclaiming all the clean cache. And in some cases,
2026 * the flushers simply cannot keep up with the allocation
2027 * rate. Nudge the flusher threads in case they are asleep.
2029 if (stat
.nr_unqueued_dirty
== nr_taken
)
2030 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2032 sc
->nr
.dirty
+= stat
.nr_dirty
;
2033 sc
->nr
.congested
+= stat
.nr_congested
;
2034 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2035 sc
->nr
.writeback
+= stat
.nr_writeback
;
2036 sc
->nr
.immediate
+= stat
.nr_immediate
;
2037 sc
->nr
.taken
+= nr_taken
;
2039 sc
->nr
.file_taken
+= nr_taken
;
2041 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2042 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2043 return nr_reclaimed
;
2046 static void shrink_active_list(unsigned long nr_to_scan
,
2047 struct lruvec
*lruvec
,
2048 struct scan_control
*sc
,
2051 unsigned long nr_taken
;
2052 unsigned long nr_scanned
;
2053 unsigned long vm_flags
;
2054 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2055 LIST_HEAD(l_active
);
2056 LIST_HEAD(l_inactive
);
2058 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2059 unsigned nr_deactivate
, nr_activate
;
2060 unsigned nr_rotated
= 0;
2061 int file
= is_file_lru(lru
);
2062 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2066 spin_lock_irq(&pgdat
->lru_lock
);
2068 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2069 &nr_scanned
, sc
, lru
);
2071 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2072 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2074 __count_vm_events(PGREFILL
, nr_scanned
);
2075 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2077 spin_unlock_irq(&pgdat
->lru_lock
);
2079 while (!list_empty(&l_hold
)) {
2081 page
= lru_to_page(&l_hold
);
2082 list_del(&page
->lru
);
2084 if (unlikely(!page_evictable(page
))) {
2085 putback_lru_page(page
);
2089 if (unlikely(buffer_heads_over_limit
)) {
2090 if (page_has_private(page
) && trylock_page(page
)) {
2091 if (page_has_private(page
))
2092 try_to_release_page(page
, 0);
2097 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2099 nr_rotated
+= hpage_nr_pages(page
);
2101 * Identify referenced, file-backed active pages and
2102 * give them one more trip around the active list. So
2103 * that executable code get better chances to stay in
2104 * memory under moderate memory pressure. Anon pages
2105 * are not likely to be evicted by use-once streaming
2106 * IO, plus JVM can create lots of anon VM_EXEC pages,
2107 * so we ignore them here.
2109 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2110 list_add(&page
->lru
, &l_active
);
2115 ClearPageActive(page
); /* we are de-activating */
2116 SetPageWorkingset(page
);
2117 list_add(&page
->lru
, &l_inactive
);
2121 * Move pages back to the lru list.
2123 spin_lock_irq(&pgdat
->lru_lock
);
2125 * Count referenced pages from currently used mappings as rotated,
2126 * even though only some of them are actually re-activated. This
2127 * helps balance scan pressure between file and anonymous pages in
2130 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2132 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2133 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2134 /* Keep all free pages in l_active list */
2135 list_splice(&l_inactive
, &l_active
);
2137 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2138 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2140 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2141 spin_unlock_irq(&pgdat
->lru_lock
);
2143 mem_cgroup_uncharge_list(&l_active
);
2144 free_unref_page_list(&l_active
);
2145 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2146 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2149 unsigned long reclaim_pages(struct list_head
*page_list
)
2152 unsigned long nr_reclaimed
= 0;
2153 LIST_HEAD(node_page_list
);
2154 struct reclaim_stat dummy_stat
;
2156 struct scan_control sc
= {
2157 .gfp_mask
= GFP_KERNEL
,
2158 .priority
= DEF_PRIORITY
,
2164 while (!list_empty(page_list
)) {
2165 page
= lru_to_page(page_list
);
2167 nid
= page_to_nid(page
);
2168 INIT_LIST_HEAD(&node_page_list
);
2171 if (nid
== page_to_nid(page
)) {
2172 ClearPageActive(page
);
2173 list_move(&page
->lru
, &node_page_list
);
2177 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2180 &dummy_stat
, false);
2181 while (!list_empty(&node_page_list
)) {
2182 page
= lru_to_page(&node_page_list
);
2183 list_del(&page
->lru
);
2184 putback_lru_page(page
);
2190 if (!list_empty(&node_page_list
)) {
2191 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2194 &dummy_stat
, false);
2195 while (!list_empty(&node_page_list
)) {
2196 page
= lru_to_page(&node_page_list
);
2197 list_del(&page
->lru
);
2198 putback_lru_page(page
);
2202 return nr_reclaimed
;
2206 * The inactive anon list should be small enough that the VM never has
2207 * to do too much work.
2209 * The inactive file list should be small enough to leave most memory
2210 * to the established workingset on the scan-resistant active list,
2211 * but large enough to avoid thrashing the aggregate readahead window.
2213 * Both inactive lists should also be large enough that each inactive
2214 * page has a chance to be referenced again before it is reclaimed.
2216 * If that fails and refaulting is observed, the inactive list grows.
2218 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2219 * on this LRU, maintained by the pageout code. An inactive_ratio
2220 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2223 * memory ratio inactive
2224 * -------------------------------------
2233 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2234 struct scan_control
*sc
, bool trace
)
2236 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2237 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2238 enum lru_list inactive_lru
= file
* LRU_FILE
;
2239 unsigned long inactive
, active
;
2240 unsigned long inactive_ratio
;
2241 unsigned long refaults
;
2245 * If we don't have swap space, anonymous page deactivation
2248 if (!file
&& !total_swap_pages
)
2251 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2252 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2255 * When refaults are being observed, it means a new workingset
2256 * is being established. Disable active list protection to get
2257 * rid of the stale workingset quickly.
2259 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
2260 if (file
&& lruvec
->refaults
!= refaults
) {
2263 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2265 inactive_ratio
= int_sqrt(10 * gb
);
2271 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2272 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2273 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2274 inactive_ratio
, file
);
2276 return inactive
* inactive_ratio
< active
;
2279 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2280 struct lruvec
*lruvec
, struct scan_control
*sc
)
2282 if (is_active_lru(lru
)) {
2283 if (inactive_list_is_low(lruvec
, is_file_lru(lru
), sc
, true))
2284 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2288 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2299 * Determine how aggressively the anon and file LRU lists should be
2300 * scanned. The relative value of each set of LRU lists is determined
2301 * by looking at the fraction of the pages scanned we did rotate back
2302 * onto the active list instead of evict.
2304 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2305 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2307 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2308 struct scan_control
*sc
, unsigned long *nr
,
2309 unsigned long *lru_pages
)
2311 int swappiness
= mem_cgroup_swappiness(memcg
);
2312 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2314 u64 denominator
= 0; /* gcc */
2315 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2316 unsigned long anon_prio
, file_prio
;
2317 enum scan_balance scan_balance
;
2318 unsigned long anon
, file
;
2319 unsigned long ap
, fp
;
2322 /* If we have no swap space, do not bother scanning anon pages. */
2323 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2324 scan_balance
= SCAN_FILE
;
2329 * Global reclaim will swap to prevent OOM even with no
2330 * swappiness, but memcg users want to use this knob to
2331 * disable swapping for individual groups completely when
2332 * using the memory controller's swap limit feature would be
2335 if (!global_reclaim(sc
) && !swappiness
) {
2336 scan_balance
= SCAN_FILE
;
2341 * Do not apply any pressure balancing cleverness when the
2342 * system is close to OOM, scan both anon and file equally
2343 * (unless the swappiness setting disagrees with swapping).
2345 if (!sc
->priority
&& swappiness
) {
2346 scan_balance
= SCAN_EQUAL
;
2351 * Prevent the reclaimer from falling into the cache trap: as
2352 * cache pages start out inactive, every cache fault will tip
2353 * the scan balance towards the file LRU. And as the file LRU
2354 * shrinks, so does the window for rotation from references.
2355 * This means we have a runaway feedback loop where a tiny
2356 * thrashing file LRU becomes infinitely more attractive than
2357 * anon pages. Try to detect this based on file LRU size.
2359 if (global_reclaim(sc
)) {
2360 unsigned long pgdatfile
;
2361 unsigned long pgdatfree
;
2363 unsigned long total_high_wmark
= 0;
2365 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2366 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2367 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2369 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2370 struct zone
*zone
= &pgdat
->node_zones
[z
];
2371 if (!managed_zone(zone
))
2374 total_high_wmark
+= high_wmark_pages(zone
);
2377 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2379 * Force SCAN_ANON if there are enough inactive
2380 * anonymous pages on the LRU in eligible zones.
2381 * Otherwise, the small LRU gets thrashed.
2383 if (!inactive_list_is_low(lruvec
, false, sc
, false) &&
2384 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2386 scan_balance
= SCAN_ANON
;
2393 * If there is enough inactive page cache, i.e. if the size of the
2394 * inactive list is greater than that of the active list *and* the
2395 * inactive list actually has some pages to scan on this priority, we
2396 * do not reclaim anything from the anonymous working set right now.
2397 * Without the second condition we could end up never scanning an
2398 * lruvec even if it has plenty of old anonymous pages unless the
2399 * system is under heavy pressure.
2401 if (!inactive_list_is_low(lruvec
, true, sc
, false) &&
2402 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2403 scan_balance
= SCAN_FILE
;
2407 scan_balance
= SCAN_FRACT
;
2410 * With swappiness at 100, anonymous and file have the same priority.
2411 * This scanning priority is essentially the inverse of IO cost.
2413 anon_prio
= swappiness
;
2414 file_prio
= 200 - anon_prio
;
2417 * OK, so we have swap space and a fair amount of page cache
2418 * pages. We use the recently rotated / recently scanned
2419 * ratios to determine how valuable each cache is.
2421 * Because workloads change over time (and to avoid overflow)
2422 * we keep these statistics as a floating average, which ends
2423 * up weighing recent references more than old ones.
2425 * anon in [0], file in [1]
2428 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2429 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2430 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2431 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2433 spin_lock_irq(&pgdat
->lru_lock
);
2434 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2435 reclaim_stat
->recent_scanned
[0] /= 2;
2436 reclaim_stat
->recent_rotated
[0] /= 2;
2439 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2440 reclaim_stat
->recent_scanned
[1] /= 2;
2441 reclaim_stat
->recent_rotated
[1] /= 2;
2445 * The amount of pressure on anon vs file pages is inversely
2446 * proportional to the fraction of recently scanned pages on
2447 * each list that were recently referenced and in active use.
2449 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2450 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2452 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2453 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2454 spin_unlock_irq(&pgdat
->lru_lock
);
2458 denominator
= ap
+ fp
+ 1;
2461 for_each_evictable_lru(lru
) {
2462 int file
= is_file_lru(lru
);
2463 unsigned long lruvec_size
;
2464 unsigned long low
, min
;
2467 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2468 mem_cgroup_protection(sc
->target_mem_cgroup
, memcg
,
2473 * Scale a cgroup's reclaim pressure by proportioning
2474 * its current usage to its memory.low or memory.min
2477 * This is important, as otherwise scanning aggression
2478 * becomes extremely binary -- from nothing as we
2479 * approach the memory protection threshold, to totally
2480 * nominal as we exceed it. This results in requiring
2481 * setting extremely liberal protection thresholds. It
2482 * also means we simply get no protection at all if we
2483 * set it too low, which is not ideal.
2485 * If there is any protection in place, we reduce scan
2486 * pressure by how much of the total memory used is
2487 * within protection thresholds.
2489 * There is one special case: in the first reclaim pass,
2490 * we skip over all groups that are within their low
2491 * protection. If that fails to reclaim enough pages to
2492 * satisfy the reclaim goal, we come back and override
2493 * the best-effort low protection. However, we still
2494 * ideally want to honor how well-behaved groups are in
2495 * that case instead of simply punishing them all
2496 * equally. As such, we reclaim them based on how much
2497 * memory they are using, reducing the scan pressure
2498 * again by how much of the total memory used is under
2501 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2502 unsigned long protection
;
2504 /* memory.low scaling, make sure we retry before OOM */
2505 if (!sc
->memcg_low_reclaim
&& low
> min
) {
2507 sc
->memcg_low_skipped
= 1;
2512 /* Avoid TOCTOU with earlier protection check */
2513 cgroup_size
= max(cgroup_size
, protection
);
2515 scan
= lruvec_size
- lruvec_size
* protection
/
2519 * Minimally target SWAP_CLUSTER_MAX pages to keep
2520 * reclaim moving forwards, avoiding decremeting
2521 * sc->priority further than desirable.
2523 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2528 scan
>>= sc
->priority
;
2531 * If the cgroup's already been deleted, make sure to
2532 * scrape out the remaining cache.
2534 if (!scan
&& !mem_cgroup_online(memcg
))
2535 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2537 switch (scan_balance
) {
2539 /* Scan lists relative to size */
2543 * Scan types proportional to swappiness and
2544 * their relative recent reclaim efficiency.
2545 * Make sure we don't miss the last page on
2546 * the offlined memory cgroups because of a
2549 scan
= mem_cgroup_online(memcg
) ?
2550 div64_u64(scan
* fraction
[file
], denominator
) :
2551 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2556 /* Scan one type exclusively */
2557 if ((scan_balance
== SCAN_FILE
) != file
) {
2563 /* Look ma, no brain */
2567 *lru_pages
+= lruvec_size
;
2573 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2575 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2576 struct scan_control
*sc
, unsigned long *lru_pages
)
2578 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2579 unsigned long nr
[NR_LRU_LISTS
];
2580 unsigned long targets
[NR_LRU_LISTS
];
2581 unsigned long nr_to_scan
;
2583 unsigned long nr_reclaimed
= 0;
2584 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2585 struct blk_plug plug
;
2588 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2590 /* Record the original scan target for proportional adjustments later */
2591 memcpy(targets
, nr
, sizeof(nr
));
2594 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2595 * event that can occur when there is little memory pressure e.g.
2596 * multiple streaming readers/writers. Hence, we do not abort scanning
2597 * when the requested number of pages are reclaimed when scanning at
2598 * DEF_PRIORITY on the assumption that the fact we are direct
2599 * reclaiming implies that kswapd is not keeping up and it is best to
2600 * do a batch of work at once. For memcg reclaim one check is made to
2601 * abort proportional reclaim if either the file or anon lru has already
2602 * dropped to zero at the first pass.
2604 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2605 sc
->priority
== DEF_PRIORITY
);
2607 blk_start_plug(&plug
);
2608 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2609 nr
[LRU_INACTIVE_FILE
]) {
2610 unsigned long nr_anon
, nr_file
, percentage
;
2611 unsigned long nr_scanned
;
2613 for_each_evictable_lru(lru
) {
2615 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2616 nr
[lru
] -= nr_to_scan
;
2618 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2625 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2629 * For kswapd and memcg, reclaim at least the number of pages
2630 * requested. Ensure that the anon and file LRUs are scanned
2631 * proportionally what was requested by get_scan_count(). We
2632 * stop reclaiming one LRU and reduce the amount scanning
2633 * proportional to the original scan target.
2635 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2636 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2639 * It's just vindictive to attack the larger once the smaller
2640 * has gone to zero. And given the way we stop scanning the
2641 * smaller below, this makes sure that we only make one nudge
2642 * towards proportionality once we've got nr_to_reclaim.
2644 if (!nr_file
|| !nr_anon
)
2647 if (nr_file
> nr_anon
) {
2648 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2649 targets
[LRU_ACTIVE_ANON
] + 1;
2651 percentage
= nr_anon
* 100 / scan_target
;
2653 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2654 targets
[LRU_ACTIVE_FILE
] + 1;
2656 percentage
= nr_file
* 100 / scan_target
;
2659 /* Stop scanning the smaller of the LRU */
2661 nr
[lru
+ LRU_ACTIVE
] = 0;
2664 * Recalculate the other LRU scan count based on its original
2665 * scan target and the percentage scanning already complete
2667 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2668 nr_scanned
= targets
[lru
] - nr
[lru
];
2669 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2670 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2673 nr_scanned
= targets
[lru
] - nr
[lru
];
2674 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2675 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2677 scan_adjusted
= true;
2679 blk_finish_plug(&plug
);
2680 sc
->nr_reclaimed
+= nr_reclaimed
;
2683 * Even if we did not try to evict anon pages at all, we want to
2684 * rebalance the anon lru active/inactive ratio.
2686 if (inactive_list_is_low(lruvec
, false, sc
, true))
2687 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2688 sc
, LRU_ACTIVE_ANON
);
2691 /* Use reclaim/compaction for costly allocs or under memory pressure */
2692 static bool in_reclaim_compaction(struct scan_control
*sc
)
2694 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2695 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2696 sc
->priority
< DEF_PRIORITY
- 2))
2703 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2704 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2705 * true if more pages should be reclaimed such that when the page allocator
2706 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2707 * It will give up earlier than that if there is difficulty reclaiming pages.
2709 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2710 unsigned long nr_reclaimed
,
2711 struct scan_control
*sc
)
2713 unsigned long pages_for_compaction
;
2714 unsigned long inactive_lru_pages
;
2717 /* If not in reclaim/compaction mode, stop */
2718 if (!in_reclaim_compaction(sc
))
2722 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2723 * number of pages that were scanned. This will return to the caller
2724 * with the risk reclaim/compaction and the resulting allocation attempt
2725 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2726 * allocations through requiring that the full LRU list has been scanned
2727 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2728 * scan, but that approximation was wrong, and there were corner cases
2729 * where always a non-zero amount of pages were scanned.
2734 /* If compaction would go ahead or the allocation would succeed, stop */
2735 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2736 struct zone
*zone
= &pgdat
->node_zones
[z
];
2737 if (!managed_zone(zone
))
2740 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2741 case COMPACT_SUCCESS
:
2742 case COMPACT_CONTINUE
:
2745 /* check next zone */
2751 * If we have not reclaimed enough pages for compaction and the
2752 * inactive lists are large enough, continue reclaiming
2754 pages_for_compaction
= compact_gap(sc
->order
);
2755 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2756 if (get_nr_swap_pages() > 0)
2757 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2759 return inactive_lru_pages
> pages_for_compaction
;
2762 static bool pgdat_memcg_congested(pg_data_t
*pgdat
, struct mem_cgroup
*memcg
)
2764 return test_bit(PGDAT_CONGESTED
, &pgdat
->flags
) ||
2765 (memcg
&& memcg_congested(pgdat
, memcg
));
2768 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2770 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2771 unsigned long nr_reclaimed
, nr_scanned
;
2772 bool reclaimable
= false;
2775 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2776 unsigned long node_lru_pages
= 0;
2777 struct mem_cgroup
*memcg
;
2779 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2781 nr_reclaimed
= sc
->nr_reclaimed
;
2782 nr_scanned
= sc
->nr_scanned
;
2784 memcg
= mem_cgroup_iter(root
, NULL
, NULL
);
2786 unsigned long lru_pages
;
2787 unsigned long reclaimed
;
2788 unsigned long scanned
;
2791 * This loop can become CPU-bound when target memcgs
2792 * aren't eligible for reclaim - either because they
2793 * don't have any reclaimable pages, or because their
2794 * memory is explicitly protected. Avoid soft lockups.
2798 switch (mem_cgroup_protected(root
, memcg
)) {
2799 case MEMCG_PROT_MIN
:
2802 * If there is no reclaimable memory, OOM.
2805 case MEMCG_PROT_LOW
:
2808 * Respect the protection only as long as
2809 * there is an unprotected supply
2810 * of reclaimable memory from other cgroups.
2812 if (!sc
->memcg_low_reclaim
) {
2813 sc
->memcg_low_skipped
= 1;
2816 memcg_memory_event(memcg
, MEMCG_LOW
);
2818 case MEMCG_PROT_NONE
:
2820 * All protection thresholds breached. We may
2821 * still choose to vary the scan pressure
2822 * applied based on by how much the cgroup in
2823 * question has exceeded its protection
2824 * thresholds (see get_scan_count).
2829 reclaimed
= sc
->nr_reclaimed
;
2830 scanned
= sc
->nr_scanned
;
2831 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2832 node_lru_pages
+= lru_pages
;
2834 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2837 /* Record the group's reclaim efficiency */
2838 vmpressure(sc
->gfp_mask
, memcg
, false,
2839 sc
->nr_scanned
- scanned
,
2840 sc
->nr_reclaimed
- reclaimed
);
2842 } while ((memcg
= mem_cgroup_iter(root
, memcg
, NULL
)));
2844 if (reclaim_state
) {
2845 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2846 reclaim_state
->reclaimed_slab
= 0;
2849 /* Record the subtree's reclaim efficiency */
2850 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2851 sc
->nr_scanned
- nr_scanned
,
2852 sc
->nr_reclaimed
- nr_reclaimed
);
2854 if (sc
->nr_reclaimed
- nr_reclaimed
)
2857 if (current_is_kswapd()) {
2859 * If reclaim is isolating dirty pages under writeback,
2860 * it implies that the long-lived page allocation rate
2861 * is exceeding the page laundering rate. Either the
2862 * global limits are not being effective at throttling
2863 * processes due to the page distribution throughout
2864 * zones or there is heavy usage of a slow backing
2865 * device. The only option is to throttle from reclaim
2866 * context which is not ideal as there is no guarantee
2867 * the dirtying process is throttled in the same way
2868 * balance_dirty_pages() manages.
2870 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2871 * count the number of pages under pages flagged for
2872 * immediate reclaim and stall if any are encountered
2873 * in the nr_immediate check below.
2875 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2876 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2879 * Tag a node as congested if all the dirty pages
2880 * scanned were backed by a congested BDI and
2881 * wait_iff_congested will stall.
2883 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2884 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
2886 /* Allow kswapd to start writing pages during reclaim.*/
2887 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2888 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2891 * If kswapd scans pages marked marked for immediate
2892 * reclaim and under writeback (nr_immediate), it
2893 * implies that pages are cycling through the LRU
2894 * faster than they are written so also forcibly stall.
2896 if (sc
->nr
.immediate
)
2897 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2901 * Legacy memcg will stall in page writeback so avoid forcibly
2902 * stalling in wait_iff_congested().
2904 if (!global_reclaim(sc
) && sane_reclaim(sc
) &&
2905 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2906 set_memcg_congestion(pgdat
, root
, true);
2909 * Stall direct reclaim for IO completions if underlying BDIs
2910 * and node is congested. Allow kswapd to continue until it
2911 * starts encountering unqueued dirty pages or cycling through
2912 * the LRU too quickly.
2914 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
2915 current_may_throttle() && pgdat_memcg_congested(pgdat
, root
))
2916 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2918 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2922 * Kswapd gives up on balancing particular nodes after too
2923 * many failures to reclaim anything from them and goes to
2924 * sleep. On reclaim progress, reset the failure counter. A
2925 * successful direct reclaim run will revive a dormant kswapd.
2928 pgdat
->kswapd_failures
= 0;
2934 * Returns true if compaction should go ahead for a costly-order request, or
2935 * the allocation would already succeed without compaction. Return false if we
2936 * should reclaim first.
2938 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2940 unsigned long watermark
;
2941 enum compact_result suitable
;
2943 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2944 if (suitable
== COMPACT_SUCCESS
)
2945 /* Allocation should succeed already. Don't reclaim. */
2947 if (suitable
== COMPACT_SKIPPED
)
2948 /* Compaction cannot yet proceed. Do reclaim. */
2952 * Compaction is already possible, but it takes time to run and there
2953 * are potentially other callers using the pages just freed. So proceed
2954 * with reclaim to make a buffer of free pages available to give
2955 * compaction a reasonable chance of completing and allocating the page.
2956 * Note that we won't actually reclaim the whole buffer in one attempt
2957 * as the target watermark in should_continue_reclaim() is lower. But if
2958 * we are already above the high+gap watermark, don't reclaim at all.
2960 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2962 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2966 * This is the direct reclaim path, for page-allocating processes. We only
2967 * try to reclaim pages from zones which will satisfy the caller's allocation
2970 * If a zone is deemed to be full of pinned pages then just give it a light
2971 * scan then give up on it.
2973 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2977 unsigned long nr_soft_reclaimed
;
2978 unsigned long nr_soft_scanned
;
2980 pg_data_t
*last_pgdat
= NULL
;
2983 * If the number of buffer_heads in the machine exceeds the maximum
2984 * allowed level, force direct reclaim to scan the highmem zone as
2985 * highmem pages could be pinning lowmem pages storing buffer_heads
2987 orig_mask
= sc
->gfp_mask
;
2988 if (buffer_heads_over_limit
) {
2989 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2990 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2993 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2994 sc
->reclaim_idx
, sc
->nodemask
) {
2996 * Take care memory controller reclaiming has small influence
2999 if (global_reclaim(sc
)) {
3000 if (!cpuset_zone_allowed(zone
,
3001 GFP_KERNEL
| __GFP_HARDWALL
))
3005 * If we already have plenty of memory free for
3006 * compaction in this zone, don't free any more.
3007 * Even though compaction is invoked for any
3008 * non-zero order, only frequent costly order
3009 * reclamation is disruptive enough to become a
3010 * noticeable problem, like transparent huge
3013 if (IS_ENABLED(CONFIG_COMPACTION
) &&
3014 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
3015 compaction_ready(zone
, sc
)) {
3016 sc
->compaction_ready
= true;
3021 * Shrink each node in the zonelist once. If the
3022 * zonelist is ordered by zone (not the default) then a
3023 * node may be shrunk multiple times but in that case
3024 * the user prefers lower zones being preserved.
3026 if (zone
->zone_pgdat
== last_pgdat
)
3030 * This steals pages from memory cgroups over softlimit
3031 * and returns the number of reclaimed pages and
3032 * scanned pages. This works for global memory pressure
3033 * and balancing, not for a memcg's limit.
3035 nr_soft_scanned
= 0;
3036 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
3037 sc
->order
, sc
->gfp_mask
,
3039 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
3040 sc
->nr_scanned
+= nr_soft_scanned
;
3041 /* need some check for avoid more shrink_zone() */
3044 /* See comment about same check for global reclaim above */
3045 if (zone
->zone_pgdat
== last_pgdat
)
3047 last_pgdat
= zone
->zone_pgdat
;
3048 shrink_node(zone
->zone_pgdat
, sc
);
3052 * Restore to original mask to avoid the impact on the caller if we
3053 * promoted it to __GFP_HIGHMEM.
3055 sc
->gfp_mask
= orig_mask
;
3058 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
3060 struct mem_cgroup
*memcg
;
3062 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
3064 unsigned long refaults
;
3065 struct lruvec
*lruvec
;
3067 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3068 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
3069 lruvec
->refaults
= refaults
;
3070 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
3074 * This is the main entry point to direct page reclaim.
3076 * If a full scan of the inactive list fails to free enough memory then we
3077 * are "out of memory" and something needs to be killed.
3079 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3080 * high - the zone may be full of dirty or under-writeback pages, which this
3081 * caller can't do much about. We kick the writeback threads and take explicit
3082 * naps in the hope that some of these pages can be written. But if the
3083 * allocating task holds filesystem locks which prevent writeout this might not
3084 * work, and the allocation attempt will fail.
3086 * returns: 0, if no pages reclaimed
3087 * else, the number of pages reclaimed
3089 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3090 struct scan_control
*sc
)
3092 int initial_priority
= sc
->priority
;
3093 pg_data_t
*last_pgdat
;
3097 delayacct_freepages_start();
3099 if (global_reclaim(sc
))
3100 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3103 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3106 shrink_zones(zonelist
, sc
);
3108 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3111 if (sc
->compaction_ready
)
3115 * If we're getting trouble reclaiming, start doing
3116 * writepage even in laptop mode.
3118 if (sc
->priority
< DEF_PRIORITY
- 2)
3119 sc
->may_writepage
= 1;
3120 } while (--sc
->priority
>= 0);
3123 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3125 if (zone
->zone_pgdat
== last_pgdat
)
3127 last_pgdat
= zone
->zone_pgdat
;
3128 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3129 set_memcg_congestion(last_pgdat
, sc
->target_mem_cgroup
, false);
3132 delayacct_freepages_end();
3134 if (sc
->nr_reclaimed
)
3135 return sc
->nr_reclaimed
;
3137 /* Aborted reclaim to try compaction? don't OOM, then */
3138 if (sc
->compaction_ready
)
3141 /* Untapped cgroup reserves? Don't OOM, retry. */
3142 if (sc
->memcg_low_skipped
) {
3143 sc
->priority
= initial_priority
;
3144 sc
->memcg_low_reclaim
= 1;
3145 sc
->memcg_low_skipped
= 0;
3152 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3155 unsigned long pfmemalloc_reserve
= 0;
3156 unsigned long free_pages
= 0;
3160 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3163 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3164 zone
= &pgdat
->node_zones
[i
];
3165 if (!managed_zone(zone
))
3168 if (!zone_reclaimable_pages(zone
))
3171 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3172 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3175 /* If there are no reserves (unexpected config) then do not throttle */
3176 if (!pfmemalloc_reserve
)
3179 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3181 /* kswapd must be awake if processes are being throttled */
3182 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3183 if (READ_ONCE(pgdat
->kswapd_classzone_idx
) > ZONE_NORMAL
)
3184 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, ZONE_NORMAL
);
3186 wake_up_interruptible(&pgdat
->kswapd_wait
);
3193 * Throttle direct reclaimers if backing storage is backed by the network
3194 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3195 * depleted. kswapd will continue to make progress and wake the processes
3196 * when the low watermark is reached.
3198 * Returns true if a fatal signal was delivered during throttling. If this
3199 * happens, the page allocator should not consider triggering the OOM killer.
3201 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3202 nodemask_t
*nodemask
)
3206 pg_data_t
*pgdat
= NULL
;
3209 * Kernel threads should not be throttled as they may be indirectly
3210 * responsible for cleaning pages necessary for reclaim to make forward
3211 * progress. kjournald for example may enter direct reclaim while
3212 * committing a transaction where throttling it could forcing other
3213 * processes to block on log_wait_commit().
3215 if (current
->flags
& PF_KTHREAD
)
3219 * If a fatal signal is pending, this process should not throttle.
3220 * It should return quickly so it can exit and free its memory
3222 if (fatal_signal_pending(current
))
3226 * Check if the pfmemalloc reserves are ok by finding the first node
3227 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3228 * GFP_KERNEL will be required for allocating network buffers when
3229 * swapping over the network so ZONE_HIGHMEM is unusable.
3231 * Throttling is based on the first usable node and throttled processes
3232 * wait on a queue until kswapd makes progress and wakes them. There
3233 * is an affinity then between processes waking up and where reclaim
3234 * progress has been made assuming the process wakes on the same node.
3235 * More importantly, processes running on remote nodes will not compete
3236 * for remote pfmemalloc reserves and processes on different nodes
3237 * should make reasonable progress.
3239 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3240 gfp_zone(gfp_mask
), nodemask
) {
3241 if (zone_idx(zone
) > ZONE_NORMAL
)
3244 /* Throttle based on the first usable node */
3245 pgdat
= zone
->zone_pgdat
;
3246 if (allow_direct_reclaim(pgdat
))
3251 /* If no zone was usable by the allocation flags then do not throttle */
3255 /* Account for the throttling */
3256 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3259 * If the caller cannot enter the filesystem, it's possible that it
3260 * is due to the caller holding an FS lock or performing a journal
3261 * transaction in the case of a filesystem like ext[3|4]. In this case,
3262 * it is not safe to block on pfmemalloc_wait as kswapd could be
3263 * blocked waiting on the same lock. Instead, throttle for up to a
3264 * second before continuing.
3266 if (!(gfp_mask
& __GFP_FS
)) {
3267 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3268 allow_direct_reclaim(pgdat
), HZ
);
3273 /* Throttle until kswapd wakes the process */
3274 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3275 allow_direct_reclaim(pgdat
));
3278 if (fatal_signal_pending(current
))
3285 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3286 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3288 unsigned long nr_reclaimed
;
3289 struct scan_control sc
= {
3290 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3291 .gfp_mask
= current_gfp_context(gfp_mask
),
3292 .reclaim_idx
= gfp_zone(gfp_mask
),
3294 .nodemask
= nodemask
,
3295 .priority
= DEF_PRIORITY
,
3296 .may_writepage
= !laptop_mode
,
3302 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3303 * Confirm they are large enough for max values.
3305 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3306 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3307 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3310 * Do not enter reclaim if fatal signal was delivered while throttled.
3311 * 1 is returned so that the page allocator does not OOM kill at this
3314 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3317 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3318 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3320 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3322 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3323 set_task_reclaim_state(current
, NULL
);
3325 return nr_reclaimed
;
3330 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3331 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3332 gfp_t gfp_mask
, bool noswap
,
3334 unsigned long *nr_scanned
)
3336 struct scan_control sc
= {
3337 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3338 .target_mem_cgroup
= memcg
,
3339 .may_writepage
= !laptop_mode
,
3341 .reclaim_idx
= MAX_NR_ZONES
- 1,
3342 .may_swap
= !noswap
,
3344 unsigned long lru_pages
;
3346 WARN_ON_ONCE(!current
->reclaim_state
);
3348 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3349 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3351 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3355 * NOTE: Although we can get the priority field, using it
3356 * here is not a good idea, since it limits the pages we can scan.
3357 * if we don't reclaim here, the shrink_node from balance_pgdat
3358 * will pick up pages from other mem cgroup's as well. We hack
3359 * the priority and make it zero.
3361 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3363 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3365 *nr_scanned
= sc
.nr_scanned
;
3367 return sc
.nr_reclaimed
;
3370 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3371 unsigned long nr_pages
,
3375 struct zonelist
*zonelist
;
3376 unsigned long nr_reclaimed
;
3377 unsigned long pflags
;
3379 unsigned int noreclaim_flag
;
3380 struct scan_control sc
= {
3381 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3382 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3383 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3384 .reclaim_idx
= MAX_NR_ZONES
- 1,
3385 .target_mem_cgroup
= memcg
,
3386 .priority
= DEF_PRIORITY
,
3387 .may_writepage
= !laptop_mode
,
3389 .may_swap
= may_swap
,
3392 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3394 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3395 * take care of from where we get pages. So the node where we start the
3396 * scan does not need to be the current node.
3398 nid
= mem_cgroup_select_victim_node(memcg
);
3400 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3402 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3404 psi_memstall_enter(&pflags
);
3405 noreclaim_flag
= memalloc_noreclaim_save();
3407 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3409 memalloc_noreclaim_restore(noreclaim_flag
);
3410 psi_memstall_leave(&pflags
);
3412 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3413 set_task_reclaim_state(current
, NULL
);
3415 return nr_reclaimed
;
3419 static void age_active_anon(struct pglist_data
*pgdat
,
3420 struct scan_control
*sc
)
3422 struct mem_cgroup
*memcg
;
3424 if (!total_swap_pages
)
3427 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3429 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3431 if (inactive_list_is_low(lruvec
, false, sc
, true))
3432 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3433 sc
, LRU_ACTIVE_ANON
);
3435 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3439 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int classzone_idx
)
3445 * Check for watermark boosts top-down as the higher zones
3446 * are more likely to be boosted. Both watermarks and boosts
3447 * should not be checked at the time time as reclaim would
3448 * start prematurely when there is no boosting and a lower
3451 for (i
= classzone_idx
; i
>= 0; i
--) {
3452 zone
= pgdat
->node_zones
+ i
;
3453 if (!managed_zone(zone
))
3456 if (zone
->watermark_boost
)
3464 * Returns true if there is an eligible zone balanced for the request order
3467 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3470 unsigned long mark
= -1;
3474 * Check watermarks bottom-up as lower zones are more likely to
3477 for (i
= 0; i
<= classzone_idx
; i
++) {
3478 zone
= pgdat
->node_zones
+ i
;
3480 if (!managed_zone(zone
))
3483 mark
= high_wmark_pages(zone
);
3484 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3489 * If a node has no populated zone within classzone_idx, it does not
3490 * need balancing by definition. This can happen if a zone-restricted
3491 * allocation tries to wake a remote kswapd.
3499 /* Clear pgdat state for congested, dirty or under writeback. */
3500 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3502 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3503 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3504 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3508 * Prepare kswapd for sleeping. This verifies that there are no processes
3509 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3511 * Returns true if kswapd is ready to sleep
3513 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3516 * The throttled processes are normally woken up in balance_pgdat() as
3517 * soon as allow_direct_reclaim() is true. But there is a potential
3518 * race between when kswapd checks the watermarks and a process gets
3519 * throttled. There is also a potential race if processes get
3520 * throttled, kswapd wakes, a large process exits thereby balancing the
3521 * zones, which causes kswapd to exit balance_pgdat() before reaching
3522 * the wake up checks. If kswapd is going to sleep, no process should
3523 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3524 * the wake up is premature, processes will wake kswapd and get
3525 * throttled again. The difference from wake ups in balance_pgdat() is
3526 * that here we are under prepare_to_wait().
3528 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3529 wake_up_all(&pgdat
->pfmemalloc_wait
);
3531 /* Hopeless node, leave it to direct reclaim */
3532 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3535 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3536 clear_pgdat_congested(pgdat
);
3544 * kswapd shrinks a node of pages that are at or below the highest usable
3545 * zone that is currently unbalanced.
3547 * Returns true if kswapd scanned at least the requested number of pages to
3548 * reclaim or if the lack of progress was due to pages under writeback.
3549 * This is used to determine if the scanning priority needs to be raised.
3551 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3552 struct scan_control
*sc
)
3557 /* Reclaim a number of pages proportional to the number of zones */
3558 sc
->nr_to_reclaim
= 0;
3559 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3560 zone
= pgdat
->node_zones
+ z
;
3561 if (!managed_zone(zone
))
3564 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3568 * Historically care was taken to put equal pressure on all zones but
3569 * now pressure is applied based on node LRU order.
3571 shrink_node(pgdat
, sc
);
3574 * Fragmentation may mean that the system cannot be rebalanced for
3575 * high-order allocations. If twice the allocation size has been
3576 * reclaimed then recheck watermarks only at order-0 to prevent
3577 * excessive reclaim. Assume that a process requested a high-order
3578 * can direct reclaim/compact.
3580 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3583 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3587 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3588 * that are eligible for use by the caller until at least one zone is
3591 * Returns the order kswapd finished reclaiming at.
3593 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3594 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3595 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3596 * or lower is eligible for reclaim until at least one usable zone is
3599 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3602 unsigned long nr_soft_reclaimed
;
3603 unsigned long nr_soft_scanned
;
3604 unsigned long pflags
;
3605 unsigned long nr_boost_reclaim
;
3606 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
3609 struct scan_control sc
= {
3610 .gfp_mask
= GFP_KERNEL
,
3615 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3616 psi_memstall_enter(&pflags
);
3617 __fs_reclaim_acquire();
3619 count_vm_event(PAGEOUTRUN
);
3622 * Account for the reclaim boost. Note that the zone boost is left in
3623 * place so that parallel allocations that are near the watermark will
3624 * stall or direct reclaim until kswapd is finished.
3626 nr_boost_reclaim
= 0;
3627 for (i
= 0; i
<= classzone_idx
; i
++) {
3628 zone
= pgdat
->node_zones
+ i
;
3629 if (!managed_zone(zone
))
3632 nr_boost_reclaim
+= zone
->watermark_boost
;
3633 zone_boosts
[i
] = zone
->watermark_boost
;
3635 boosted
= nr_boost_reclaim
;
3638 sc
.priority
= DEF_PRIORITY
;
3640 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3641 bool raise_priority
= true;
3645 sc
.reclaim_idx
= classzone_idx
;
3648 * If the number of buffer_heads exceeds the maximum allowed
3649 * then consider reclaiming from all zones. This has a dual
3650 * purpose -- on 64-bit systems it is expected that
3651 * buffer_heads are stripped during active rotation. On 32-bit
3652 * systems, highmem pages can pin lowmem memory and shrinking
3653 * buffers can relieve lowmem pressure. Reclaim may still not
3654 * go ahead if all eligible zones for the original allocation
3655 * request are balanced to avoid excessive reclaim from kswapd.
3657 if (buffer_heads_over_limit
) {
3658 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3659 zone
= pgdat
->node_zones
+ i
;
3660 if (!managed_zone(zone
))
3669 * If the pgdat is imbalanced then ignore boosting and preserve
3670 * the watermarks for a later time and restart. Note that the
3671 * zone watermarks will be still reset at the end of balancing
3672 * on the grounds that the normal reclaim should be enough to
3673 * re-evaluate if boosting is required when kswapd next wakes.
3675 balanced
= pgdat_balanced(pgdat
, sc
.order
, classzone_idx
);
3676 if (!balanced
&& nr_boost_reclaim
) {
3677 nr_boost_reclaim
= 0;
3682 * If boosting is not active then only reclaim if there are no
3683 * eligible zones. Note that sc.reclaim_idx is not used as
3684 * buffer_heads_over_limit may have adjusted it.
3686 if (!nr_boost_reclaim
&& balanced
)
3689 /* Limit the priority of boosting to avoid reclaim writeback */
3690 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
3691 raise_priority
= false;
3694 * Do not writeback or swap pages for boosted reclaim. The
3695 * intent is to relieve pressure not issue sub-optimal IO
3696 * from reclaim context. If no pages are reclaimed, the
3697 * reclaim will be aborted.
3699 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
3700 sc
.may_swap
= !nr_boost_reclaim
;
3703 * Do some background aging of the anon list, to give
3704 * pages a chance to be referenced before reclaiming. All
3705 * pages are rotated regardless of classzone as this is
3706 * about consistent aging.
3708 age_active_anon(pgdat
, &sc
);
3711 * If we're getting trouble reclaiming, start doing writepage
3712 * even in laptop mode.
3714 if (sc
.priority
< DEF_PRIORITY
- 2)
3715 sc
.may_writepage
= 1;
3717 /* Call soft limit reclaim before calling shrink_node. */
3719 nr_soft_scanned
= 0;
3720 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3721 sc
.gfp_mask
, &nr_soft_scanned
);
3722 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3725 * There should be no need to raise the scanning priority if
3726 * enough pages are already being scanned that that high
3727 * watermark would be met at 100% efficiency.
3729 if (kswapd_shrink_node(pgdat
, &sc
))
3730 raise_priority
= false;
3733 * If the low watermark is met there is no need for processes
3734 * to be throttled on pfmemalloc_wait as they should not be
3735 * able to safely make forward progress. Wake them
3737 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3738 allow_direct_reclaim(pgdat
))
3739 wake_up_all(&pgdat
->pfmemalloc_wait
);
3741 /* Check if kswapd should be suspending */
3742 __fs_reclaim_release();
3743 ret
= try_to_freeze();
3744 __fs_reclaim_acquire();
3745 if (ret
|| kthread_should_stop())
3749 * Raise priority if scanning rate is too low or there was no
3750 * progress in reclaiming pages
3752 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3753 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
3756 * If reclaim made no progress for a boost, stop reclaim as
3757 * IO cannot be queued and it could be an infinite loop in
3758 * extreme circumstances.
3760 if (nr_boost_reclaim
&& !nr_reclaimed
)
3763 if (raise_priority
|| !nr_reclaimed
)
3765 } while (sc
.priority
>= 1);
3767 if (!sc
.nr_reclaimed
)
3768 pgdat
->kswapd_failures
++;
3771 /* If reclaim was boosted, account for the reclaim done in this pass */
3773 unsigned long flags
;
3775 for (i
= 0; i
<= classzone_idx
; i
++) {
3776 if (!zone_boosts
[i
])
3779 /* Increments are under the zone lock */
3780 zone
= pgdat
->node_zones
+ i
;
3781 spin_lock_irqsave(&zone
->lock
, flags
);
3782 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
3783 spin_unlock_irqrestore(&zone
->lock
, flags
);
3787 * As there is now likely space, wakeup kcompact to defragment
3790 wakeup_kcompactd(pgdat
, pageblock_order
, classzone_idx
);
3793 snapshot_refaults(NULL
, pgdat
);
3794 __fs_reclaim_release();
3795 psi_memstall_leave(&pflags
);
3796 set_task_reclaim_state(current
, NULL
);
3799 * Return the order kswapd stopped reclaiming at as
3800 * prepare_kswapd_sleep() takes it into account. If another caller
3801 * entered the allocator slow path while kswapd was awake, order will
3802 * remain at the higher level.
3808 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3809 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3810 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3811 * after previous reclaim attempt (node is still unbalanced). In that case
3812 * return the zone index of the previous kswapd reclaim cycle.
3814 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3815 enum zone_type prev_classzone_idx
)
3817 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
3819 return curr_idx
== MAX_NR_ZONES
? prev_classzone_idx
: curr_idx
;
3822 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3823 unsigned int classzone_idx
)
3828 if (freezing(current
) || kthread_should_stop())
3831 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3834 * Try to sleep for a short interval. Note that kcompactd will only be
3835 * woken if it is possible to sleep for a short interval. This is
3836 * deliberate on the assumption that if reclaim cannot keep an
3837 * eligible zone balanced that it's also unlikely that compaction will
3840 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3842 * Compaction records what page blocks it recently failed to
3843 * isolate pages from and skips them in the future scanning.
3844 * When kswapd is going to sleep, it is reasonable to assume
3845 * that pages and compaction may succeed so reset the cache.
3847 reset_isolation_suitable(pgdat
);
3850 * We have freed the memory, now we should compact it to make
3851 * allocation of the requested order possible.
3853 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3855 remaining
= schedule_timeout(HZ
/10);
3858 * If woken prematurely then reset kswapd_classzone_idx and
3859 * order. The values will either be from a wakeup request or
3860 * the previous request that slept prematurely.
3863 WRITE_ONCE(pgdat
->kswapd_classzone_idx
,
3864 kswapd_classzone_idx(pgdat
, classzone_idx
));
3866 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
3867 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
3870 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3871 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3875 * After a short sleep, check if it was a premature sleep. If not, then
3876 * go fully to sleep until explicitly woken up.
3879 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3880 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3883 * vmstat counters are not perfectly accurate and the estimated
3884 * value for counters such as NR_FREE_PAGES can deviate from the
3885 * true value by nr_online_cpus * threshold. To avoid the zone
3886 * watermarks being breached while under pressure, we reduce the
3887 * per-cpu vmstat threshold while kswapd is awake and restore
3888 * them before going back to sleep.
3890 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3892 if (!kthread_should_stop())
3895 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3898 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3900 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3902 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3906 * The background pageout daemon, started as a kernel thread
3907 * from the init process.
3909 * This basically trickles out pages so that we have _some_
3910 * free memory available even if there is no other activity
3911 * that frees anything up. This is needed for things like routing
3912 * etc, where we otherwise might have all activity going on in
3913 * asynchronous contexts that cannot page things out.
3915 * If there are applications that are active memory-allocators
3916 * (most normal use), this basically shouldn't matter.
3918 static int kswapd(void *p
)
3920 unsigned int alloc_order
, reclaim_order
;
3921 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3922 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3923 struct task_struct
*tsk
= current
;
3924 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3926 if (!cpumask_empty(cpumask
))
3927 set_cpus_allowed_ptr(tsk
, cpumask
);
3930 * Tell the memory management that we're a "memory allocator",
3931 * and that if we need more memory we should get access to it
3932 * regardless (see "__alloc_pages()"). "kswapd" should
3933 * never get caught in the normal page freeing logic.
3935 * (Kswapd normally doesn't need memory anyway, but sometimes
3936 * you need a small amount of memory in order to be able to
3937 * page out something else, and this flag essentially protects
3938 * us from recursively trying to free more memory as we're
3939 * trying to free the first piece of memory in the first place).
3941 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3944 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3945 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3949 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3950 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3953 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3956 /* Read the new order and classzone_idx */
3957 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3958 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3959 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3960 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3962 ret
= try_to_freeze();
3963 if (kthread_should_stop())
3967 * We can speed up thawing tasks if we don't call balance_pgdat
3968 * after returning from the refrigerator
3974 * Reclaim begins at the requested order but if a high-order
3975 * reclaim fails then kswapd falls back to reclaiming for
3976 * order-0. If that happens, kswapd will consider sleeping
3977 * for the order it finished reclaiming at (reclaim_order)
3978 * but kcompactd is woken to compact for the original
3979 * request (alloc_order).
3981 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3983 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3984 if (reclaim_order
< alloc_order
)
3985 goto kswapd_try_sleep
;
3988 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3994 * A zone is low on free memory or too fragmented for high-order memory. If
3995 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3996 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3997 * has failed or is not needed, still wake up kcompactd if only compaction is
4000 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
4001 enum zone_type classzone_idx
)
4004 enum zone_type curr_idx
;
4006 if (!managed_zone(zone
))
4009 if (!cpuset_zone_allowed(zone
, gfp_flags
))
4012 pgdat
= zone
->zone_pgdat
;
4013 curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
4015 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< classzone_idx
)
4016 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, classzone_idx
);
4018 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
4019 WRITE_ONCE(pgdat
->kswapd_order
, order
);
4021 if (!waitqueue_active(&pgdat
->kswapd_wait
))
4024 /* Hopeless node, leave it to direct reclaim if possible */
4025 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
4026 (pgdat_balanced(pgdat
, order
, classzone_idx
) &&
4027 !pgdat_watermark_boosted(pgdat
, classzone_idx
))) {
4029 * There may be plenty of free memory available, but it's too
4030 * fragmented for high-order allocations. Wake up kcompactd
4031 * and rely on compaction_suitable() to determine if it's
4032 * needed. If it fails, it will defer subsequent attempts to
4033 * ratelimit its work.
4035 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
4036 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
4040 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
4042 wake_up_interruptible(&pgdat
->kswapd_wait
);
4045 #ifdef CONFIG_HIBERNATION
4047 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4050 * Rather than trying to age LRUs the aim is to preserve the overall
4051 * LRU order by reclaiming preferentially
4052 * inactive > active > active referenced > active mapped
4054 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4056 struct scan_control sc
= {
4057 .nr_to_reclaim
= nr_to_reclaim
,
4058 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4059 .reclaim_idx
= MAX_NR_ZONES
- 1,
4060 .priority
= DEF_PRIORITY
,
4064 .hibernation_mode
= 1,
4066 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4067 unsigned long nr_reclaimed
;
4068 unsigned int noreclaim_flag
;
4070 fs_reclaim_acquire(sc
.gfp_mask
);
4071 noreclaim_flag
= memalloc_noreclaim_save();
4072 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4074 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4076 set_task_reclaim_state(current
, NULL
);
4077 memalloc_noreclaim_restore(noreclaim_flag
);
4078 fs_reclaim_release(sc
.gfp_mask
);
4080 return nr_reclaimed
;
4082 #endif /* CONFIG_HIBERNATION */
4084 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4085 not required for correctness. So if the last cpu in a node goes
4086 away, we get changed to run anywhere: as the first one comes back,
4087 restore their cpu bindings. */
4088 static int kswapd_cpu_online(unsigned int cpu
)
4092 for_each_node_state(nid
, N_MEMORY
) {
4093 pg_data_t
*pgdat
= NODE_DATA(nid
);
4094 const struct cpumask
*mask
;
4096 mask
= cpumask_of_node(pgdat
->node_id
);
4098 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
4099 /* One of our CPUs online: restore mask */
4100 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
4106 * This kswapd start function will be called by init and node-hot-add.
4107 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4109 int kswapd_run(int nid
)
4111 pg_data_t
*pgdat
= NODE_DATA(nid
);
4117 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4118 if (IS_ERR(pgdat
->kswapd
)) {
4119 /* failure at boot is fatal */
4120 BUG_ON(system_state
< SYSTEM_RUNNING
);
4121 pr_err("Failed to start kswapd on node %d\n", nid
);
4122 ret
= PTR_ERR(pgdat
->kswapd
);
4123 pgdat
->kswapd
= NULL
;
4129 * Called by memory hotplug when all memory in a node is offlined. Caller must
4130 * hold mem_hotplug_begin/end().
4132 void kswapd_stop(int nid
)
4134 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4137 kthread_stop(kswapd
);
4138 NODE_DATA(nid
)->kswapd
= NULL
;
4142 static int __init
kswapd_init(void)
4147 for_each_node_state(nid
, N_MEMORY
)
4149 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
4150 "mm/vmscan:online", kswapd_cpu_online
,
4156 module_init(kswapd_init
)
4162 * If non-zero call node_reclaim when the number of free pages falls below
4165 int node_reclaim_mode __read_mostly
;
4167 #define RECLAIM_OFF 0
4168 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4169 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4170 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4173 * Priority for NODE_RECLAIM. This determines the fraction of pages
4174 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4177 #define NODE_RECLAIM_PRIORITY 4
4180 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4183 int sysctl_min_unmapped_ratio
= 1;
4186 * If the number of slab pages in a zone grows beyond this percentage then
4187 * slab reclaim needs to occur.
4189 int sysctl_min_slab_ratio
= 5;
4191 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4193 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4194 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4195 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4198 * It's possible for there to be more file mapped pages than
4199 * accounted for by the pages on the file LRU lists because
4200 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4202 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4205 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4206 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4208 unsigned long nr_pagecache_reclaimable
;
4209 unsigned long delta
= 0;
4212 * If RECLAIM_UNMAP is set, then all file pages are considered
4213 * potentially reclaimable. Otherwise, we have to worry about
4214 * pages like swapcache and node_unmapped_file_pages() provides
4217 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4218 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4220 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4222 /* If we can't clean pages, remove dirty pages from consideration */
4223 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4224 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4226 /* Watch for any possible underflows due to delta */
4227 if (unlikely(delta
> nr_pagecache_reclaimable
))
4228 delta
= nr_pagecache_reclaimable
;
4230 return nr_pagecache_reclaimable
- delta
;
4234 * Try to free up some pages from this node through reclaim.
4236 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4238 /* Minimum pages needed in order to stay on node */
4239 const unsigned long nr_pages
= 1 << order
;
4240 struct task_struct
*p
= current
;
4241 unsigned int noreclaim_flag
;
4242 struct scan_control sc
= {
4243 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4244 .gfp_mask
= current_gfp_context(gfp_mask
),
4246 .priority
= NODE_RECLAIM_PRIORITY
,
4247 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4248 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4250 .reclaim_idx
= gfp_zone(gfp_mask
),
4253 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4257 fs_reclaim_acquire(sc
.gfp_mask
);
4259 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4260 * and we also need to be able to write out pages for RECLAIM_WRITE
4261 * and RECLAIM_UNMAP.
4263 noreclaim_flag
= memalloc_noreclaim_save();
4264 p
->flags
|= PF_SWAPWRITE
;
4265 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4267 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4269 * Free memory by calling shrink node with increasing
4270 * priorities until we have enough memory freed.
4273 shrink_node(pgdat
, &sc
);
4274 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4277 set_task_reclaim_state(p
, NULL
);
4278 current
->flags
&= ~PF_SWAPWRITE
;
4279 memalloc_noreclaim_restore(noreclaim_flag
);
4280 fs_reclaim_release(sc
.gfp_mask
);
4282 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4284 return sc
.nr_reclaimed
>= nr_pages
;
4287 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4292 * Node reclaim reclaims unmapped file backed pages and
4293 * slab pages if we are over the defined limits.
4295 * A small portion of unmapped file backed pages is needed for
4296 * file I/O otherwise pages read by file I/O will be immediately
4297 * thrown out if the node is overallocated. So we do not reclaim
4298 * if less than a specified percentage of the node is used by
4299 * unmapped file backed pages.
4301 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4302 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4303 return NODE_RECLAIM_FULL
;
4306 * Do not scan if the allocation should not be delayed.
4308 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4309 return NODE_RECLAIM_NOSCAN
;
4312 * Only run node reclaim on the local node or on nodes that do not
4313 * have associated processors. This will favor the local processor
4314 * over remote processors and spread off node memory allocations
4315 * as wide as possible.
4317 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4318 return NODE_RECLAIM_NOSCAN
;
4320 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4321 return NODE_RECLAIM_NOSCAN
;
4323 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4324 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4327 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4334 * page_evictable - test whether a page is evictable
4335 * @page: the page to test
4337 * Test whether page is evictable--i.e., should be placed on active/inactive
4338 * lists vs unevictable list.
4340 * Reasons page might not be evictable:
4341 * (1) page's mapping marked unevictable
4342 * (2) page is part of an mlocked VMA
4345 int page_evictable(struct page
*page
)
4349 /* Prevent address_space of inode and swap cache from being freed */
4351 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4357 * check_move_unevictable_pages - check pages for evictability and move to
4358 * appropriate zone lru list
4359 * @pvec: pagevec with lru pages to check
4361 * Checks pages for evictability, if an evictable page is in the unevictable
4362 * lru list, moves it to the appropriate evictable lru list. This function
4363 * should be only used for lru pages.
4365 void check_move_unevictable_pages(struct pagevec
*pvec
)
4367 struct lruvec
*lruvec
;
4368 struct pglist_data
*pgdat
= NULL
;
4373 for (i
= 0; i
< pvec
->nr
; i
++) {
4374 struct page
*page
= pvec
->pages
[i
];
4375 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4378 if (pagepgdat
!= pgdat
) {
4380 spin_unlock_irq(&pgdat
->lru_lock
);
4382 spin_lock_irq(&pgdat
->lru_lock
);
4384 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4386 if (!PageLRU(page
) || !PageUnevictable(page
))
4389 if (page_evictable(page
)) {
4390 enum lru_list lru
= page_lru_base_type(page
);
4392 VM_BUG_ON_PAGE(PageActive(page
), page
);
4393 ClearPageUnevictable(page
);
4394 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4395 add_page_to_lru_list(page
, lruvec
, lru
);
4401 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4402 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4403 spin_unlock_irq(&pgdat
->lru_lock
);
4406 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);