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 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
96 unsigned int memcg_low_reclaim
:1;
97 unsigned int memcg_low_skipped
:1;
99 unsigned int hibernation_mode
:1;
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready
:1;
104 /* Allocation order */
107 /* Scan (total_size >> priority) pages at once */
110 /* The highest zone to isolate pages for reclaim from */
113 /* This context's GFP mask */
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned
;
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed
;
124 unsigned int unqueued_dirty
;
125 unsigned int congested
;
126 unsigned int writeback
;
127 unsigned int immediate
;
128 unsigned int file_taken
;
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state
;
136 #ifdef ARCH_HAS_PREFETCH
137 #define prefetch_prev_lru_page(_page, _base, _field) \
139 if ((_page)->lru.prev != _base) { \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
147 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
150 #ifdef ARCH_HAS_PREFETCHW
151 #define prefetchw_prev_lru_page(_page, _base, _field) \
153 if ((_page)->lru.prev != _base) { \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
161 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165 * From 0 .. 100. Higher means more swappy.
167 int vm_swappiness
= 60;
169 * The total number of pages which are beyond the high watermark within all
172 unsigned long vm_total_pages
;
174 static void set_task_reclaim_state(struct task_struct
*task
,
175 struct reclaim_state
*rs
)
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
183 task
->reclaim_state
= rs
;
186 static LIST_HEAD(shrinker_list
);
187 static DECLARE_RWSEM(shrinker_rwsem
);
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
201 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
203 static DEFINE_IDR(shrinker_idr
);
204 static int shrinker_nr_max
;
206 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
208 int id
, ret
= -ENOMEM
;
210 down_write(&shrinker_rwsem
);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
216 if (id
>= shrinker_nr_max
) {
217 if (memcg_expand_shrinker_maps(id
)) {
218 idr_remove(&shrinker_idr
, id
);
222 shrinker_nr_max
= id
+ 1;
227 up_write(&shrinker_rwsem
);
231 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
233 int id
= shrinker
->id
;
237 down_write(&shrinker_rwsem
);
238 idr_remove(&shrinker_idr
, id
);
239 up_write(&shrinker_rwsem
);
242 static bool global_reclaim(struct scan_control
*sc
)
244 return !sc
->target_mem_cgroup
;
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
260 static bool sane_reclaim(struct scan_control
*sc
)
262 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
266 #ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
273 static void set_memcg_congestion(pg_data_t
*pgdat
,
274 struct mem_cgroup
*memcg
,
277 struct mem_cgroup_per_node
*mn
;
282 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
283 WRITE_ONCE(mn
->congested
, congested
);
286 static bool memcg_congested(pg_data_t
*pgdat
,
287 struct mem_cgroup
*memcg
)
289 struct mem_cgroup_per_node
*mn
;
291 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
292 return READ_ONCE(mn
->congested
);
296 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
301 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
305 static bool global_reclaim(struct scan_control
*sc
)
310 static bool sane_reclaim(struct scan_control
*sc
)
315 static inline void set_memcg_congestion(struct pglist_data
*pgdat
,
316 struct mem_cgroup
*memcg
, bool congested
)
320 static inline bool memcg_congested(struct pglist_data
*pgdat
,
321 struct mem_cgroup
*memcg
)
329 * This misses isolated pages which are not accounted for to save counters.
330 * As the data only determines if reclaim or compaction continues, it is
331 * not expected that isolated pages will be a dominating factor.
333 unsigned long zone_reclaimable_pages(struct zone
*zone
)
337 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
338 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
339 if (get_nr_swap_pages() > 0)
340 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
341 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
347 * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 * @lruvec: lru vector
350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
352 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
354 unsigned long lru_size
= 0;
357 if (!mem_cgroup_disabled()) {
358 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
359 lru_size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
361 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
363 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
364 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
367 if (!managed_zone(zone
))
370 if (!mem_cgroup_disabled())
371 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
373 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
374 NR_ZONE_LRU_BASE
+ lru
);
375 lru_size
-= min(size
, lru_size
);
383 * Add a shrinker callback to be called from the vm.
385 int prealloc_shrinker(struct shrinker
*shrinker
)
387 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
389 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
392 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
393 if (!shrinker
->nr_deferred
)
396 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
397 if (prealloc_memcg_shrinker(shrinker
))
404 kfree(shrinker
->nr_deferred
);
405 shrinker
->nr_deferred
= NULL
;
409 void free_prealloced_shrinker(struct shrinker
*shrinker
)
411 if (!shrinker
->nr_deferred
)
414 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
415 unregister_memcg_shrinker(shrinker
);
417 kfree(shrinker
->nr_deferred
);
418 shrinker
->nr_deferred
= NULL
;
421 void register_shrinker_prepared(struct shrinker
*shrinker
)
423 down_write(&shrinker_rwsem
);
424 list_add_tail(&shrinker
->list
, &shrinker_list
);
426 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
427 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
429 up_write(&shrinker_rwsem
);
432 int register_shrinker(struct shrinker
*shrinker
)
434 int err
= prealloc_shrinker(shrinker
);
438 register_shrinker_prepared(shrinker
);
441 EXPORT_SYMBOL(register_shrinker
);
446 void unregister_shrinker(struct shrinker
*shrinker
)
448 if (!shrinker
->nr_deferred
)
450 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
451 unregister_memcg_shrinker(shrinker
);
452 down_write(&shrinker_rwsem
);
453 list_del(&shrinker
->list
);
454 up_write(&shrinker_rwsem
);
455 kfree(shrinker
->nr_deferred
);
456 shrinker
->nr_deferred
= NULL
;
458 EXPORT_SYMBOL(unregister_shrinker
);
460 #define SHRINK_BATCH 128
462 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
463 struct shrinker
*shrinker
, int priority
)
465 unsigned long freed
= 0;
466 unsigned long long delta
;
471 int nid
= shrinkctl
->nid
;
472 long batch_size
= shrinker
->batch
? shrinker
->batch
474 long scanned
= 0, next_deferred
;
476 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
479 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
480 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
484 * copy the current shrinker scan count into a local variable
485 * and zero it so that other concurrent shrinker invocations
486 * don't also do this scanning work.
488 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
491 if (shrinker
->seeks
) {
492 delta
= freeable
>> priority
;
494 do_div(delta
, shrinker
->seeks
);
497 * These objects don't require any IO to create. Trim
498 * them aggressively under memory pressure to keep
499 * them from causing refetches in the IO caches.
501 delta
= freeable
/ 2;
505 if (total_scan
< 0) {
506 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 shrinker
->scan_objects
, total_scan
);
508 total_scan
= freeable
;
511 next_deferred
= total_scan
;
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
525 if (delta
< freeable
/ 4)
526 total_scan
= min(total_scan
, freeable
/ 2);
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
533 if (total_scan
> freeable
* 2)
534 total_scan
= freeable
* 2;
536 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
537 freeable
, delta
, total_scan
, priority
);
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
554 while (total_scan
>= batch_size
||
555 total_scan
>= freeable
) {
557 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
559 shrinkctl
->nr_to_scan
= nr_to_scan
;
560 shrinkctl
->nr_scanned
= nr_to_scan
;
561 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
562 if (ret
== SHRINK_STOP
)
566 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
567 total_scan
-= shrinkctl
->nr_scanned
;
568 scanned
+= shrinkctl
->nr_scanned
;
573 if (next_deferred
>= scanned
)
574 next_deferred
-= scanned
;
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
582 if (next_deferred
> 0)
583 new_nr
= atomic_long_add_return(next_deferred
,
584 &shrinker
->nr_deferred
[nid
]);
586 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
588 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
593 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
594 struct mem_cgroup
*memcg
, int priority
)
596 struct memcg_shrinker_map
*map
;
597 unsigned long ret
, freed
= 0;
600 if (!mem_cgroup_online(memcg
))
603 if (!down_read_trylock(&shrinker_rwsem
))
606 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
611 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
612 struct shrink_control sc
= {
613 .gfp_mask
= gfp_mask
,
617 struct shrinker
*shrinker
;
619 shrinker
= idr_find(&shrinker_idr
, i
);
620 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
622 clear_bit(i
, map
->map
);
626 /* Call non-slab shrinkers even though kmem is disabled */
627 if (!memcg_kmem_enabled() &&
628 !(shrinker
->flags
& SHRINKER_NONSLAB
))
631 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
632 if (ret
== SHRINK_EMPTY
) {
633 clear_bit(i
, map
->map
);
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
647 * set_bit() do_shrink_slab()
649 smp_mb__after_atomic();
650 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
651 if (ret
== SHRINK_EMPTY
)
654 memcg_set_shrinker_bit(memcg
, nid
, i
);
658 if (rwsem_is_contended(&shrinker_rwsem
)) {
664 up_read(&shrinker_rwsem
);
667 #else /* CONFIG_MEMCG */
668 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
669 struct mem_cgroup
*memcg
, int priority
)
673 #endif /* CONFIG_MEMCG */
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
682 * Call the shrink functions to age shrinkable caches.
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
693 * Returns the number of reclaimed slab objects.
695 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
696 struct mem_cgroup
*memcg
,
699 unsigned long ret
, freed
= 0;
700 struct shrinker
*shrinker
;
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
710 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
712 if (!down_read_trylock(&shrinker_rwsem
))
715 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
716 struct shrink_control sc
= {
717 .gfp_mask
= gfp_mask
,
722 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
723 if (ret
== SHRINK_EMPTY
)
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
731 if (rwsem_is_contended(&shrinker_rwsem
)) {
737 up_read(&shrinker_rwsem
);
743 void drop_slab_node(int nid
)
748 struct mem_cgroup
*memcg
= NULL
;
751 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
753 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
754 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
755 } while (freed
> 10);
762 for_each_online_node(nid
)
766 static inline int is_page_cache_freeable(struct page
*page
)
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
773 int page_cache_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
775 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
778 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
780 if (current
->flags
& PF_SWAPWRITE
)
782 if (!inode_write_congested(inode
))
784 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
798 * We're allowed to run sleeping lock_page() here because we know the caller has
801 static void handle_write_error(struct address_space
*mapping
,
802 struct page
*page
, int error
)
805 if (page_mapping(page
) == mapping
)
806 mapping_set_error(mapping
, error
);
810 /* possible outcome of pageout() */
812 /* failed to write page out, page is locked */
814 /* move page to the active list, page is locked */
816 /* page has been sent to the disk successfully, page is unlocked */
818 /* page is clean and locked */
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
826 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
827 struct scan_control
*sc
)
830 * If the page is dirty, only perform writeback if that write
831 * will be non-blocking. To prevent this allocation from being
832 * stalled by pagecache activity. But note that there may be
833 * stalls if we need to run get_block(). We could test
834 * PagePrivate for that.
836 * If this process is currently in __generic_file_write_iter() against
837 * this page's queue, we can perform writeback even if that
840 * If the page is swapcache, write it back even if that would
841 * block, for some throttling. This happens by accident, because
842 * swap_backing_dev_info is bust: it doesn't reflect the
843 * congestion state of the swapdevs. Easy to fix, if needed.
845 if (!is_page_cache_freeable(page
))
849 * Some data journaling orphaned pages can have
850 * page->mapping == NULL while being dirty with clean buffers.
852 if (page_has_private(page
)) {
853 if (try_to_free_buffers(page
)) {
854 ClearPageDirty(page
);
855 pr_info("%s: orphaned page\n", __func__
);
861 if (mapping
->a_ops
->writepage
== NULL
)
862 return PAGE_ACTIVATE
;
863 if (!may_write_to_inode(mapping
->host
, sc
))
866 if (clear_page_dirty_for_io(page
)) {
868 struct writeback_control wbc
= {
869 .sync_mode
= WB_SYNC_NONE
,
870 .nr_to_write
= SWAP_CLUSTER_MAX
,
872 .range_end
= LLONG_MAX
,
876 SetPageReclaim(page
);
877 res
= mapping
->a_ops
->writepage(page
, &wbc
);
879 handle_write_error(mapping
, page
, res
);
880 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
881 ClearPageReclaim(page
);
882 return PAGE_ACTIVATE
;
885 if (!PageWriteback(page
)) {
886 /* synchronous write or broken a_ops? */
887 ClearPageReclaim(page
);
889 trace_mm_vmscan_writepage(page
);
890 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
898 * Same as remove_mapping, but if the page is removed from the mapping, it
899 * gets returned with a refcount of 0.
901 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
907 BUG_ON(!PageLocked(page
));
908 BUG_ON(mapping
!= page_mapping(page
));
910 xa_lock_irqsave(&mapping
->i_pages
, flags
);
912 * The non racy check for a busy page.
914 * Must be careful with the order of the tests. When someone has
915 * a ref to the page, it may be possible that they dirty it then
916 * drop the reference. So if PageDirty is tested before page_count
917 * here, then the following race may occur:
919 * get_user_pages(&page);
920 * [user mapping goes away]
922 * !PageDirty(page) [good]
923 * SetPageDirty(page);
925 * !page_count(page) [good, discard it]
927 * [oops, our write_to data is lost]
929 * Reversing the order of the tests ensures such a situation cannot
930 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 * load is not satisfied before that of page->_refcount.
933 * Note that if SetPageDirty is always performed via set_page_dirty,
934 * and thus under the i_pages lock, then this ordering is not required.
936 refcount
= 1 + compound_nr(page
);
937 if (!page_ref_freeze(page
, refcount
))
939 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 if (unlikely(PageDirty(page
))) {
941 page_ref_unfreeze(page
, refcount
);
945 if (PageSwapCache(page
)) {
946 swp_entry_t swap
= { .val
= page_private(page
) };
947 mem_cgroup_swapout(page
, swap
);
948 __delete_from_swap_cache(page
, swap
);
949 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
950 put_swap_page(page
, swap
);
952 void (*freepage
)(struct page
*);
955 freepage
= mapping
->a_ops
->freepage
;
957 * Remember a shadow entry for reclaimed file cache in
958 * order to detect refaults, thus thrashing, later on.
960 * But don't store shadows in an address space that is
961 * already exiting. This is not just an optizimation,
962 * inode reclaim needs to empty out the radix tree or
963 * the nodes are lost. Don't plant shadows behind its
966 * We also don't store shadows for DAX mappings because the
967 * only page cache pages found in these are zero pages
968 * covering holes, and because we don't want to mix DAX
969 * exceptional entries and shadow exceptional entries in the
970 * same address_space.
972 if (reclaimed
&& page_is_file_cache(page
) &&
973 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
974 shadow
= workingset_eviction(page
);
975 __delete_from_page_cache(page
, shadow
);
976 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
978 if (freepage
!= NULL
)
985 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
990 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 * someone else has a ref on the page, abort and return 0. If it was
992 * successfully detached, return 1. Assumes the caller has a single ref on
995 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
997 if (__remove_mapping(mapping
, page
, false)) {
999 * Unfreezing the refcount with 1 rather than 2 effectively
1000 * drops the pagecache ref for us without requiring another
1003 page_ref_unfreeze(page
, 1);
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1016 * lru_lock must not be held, interrupts must be enabled.
1018 void putback_lru_page(struct page
*page
)
1020 lru_cache_add(page
);
1021 put_page(page
); /* drop ref from isolate */
1024 enum page_references
{
1026 PAGEREF_RECLAIM_CLEAN
,
1031 static enum page_references
page_check_references(struct page
*page
,
1032 struct scan_control
*sc
)
1034 int referenced_ptes
, referenced_page
;
1035 unsigned long vm_flags
;
1037 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1039 referenced_page
= TestClearPageReferenced(page
);
1042 * Mlock lost the isolation race with us. Let try_to_unmap()
1043 * move the page to the unevictable list.
1045 if (vm_flags
& VM_LOCKED
)
1046 return PAGEREF_RECLAIM
;
1048 if (referenced_ptes
) {
1049 if (PageSwapBacked(page
))
1050 return PAGEREF_ACTIVATE
;
1052 * All mapped pages start out with page table
1053 * references from the instantiating fault, so we need
1054 * to look twice if a mapped file page is used more
1057 * Mark it and spare it for another trip around the
1058 * inactive list. Another page table reference will
1059 * lead to its activation.
1061 * Note: the mark is set for activated pages as well
1062 * so that recently deactivated but used pages are
1063 * quickly recovered.
1065 SetPageReferenced(page
);
1067 if (referenced_page
|| referenced_ptes
> 1)
1068 return PAGEREF_ACTIVATE
;
1071 * Activate file-backed executable pages after first usage.
1073 if (vm_flags
& VM_EXEC
)
1074 return PAGEREF_ACTIVATE
;
1076 return PAGEREF_KEEP
;
1079 /* Reclaim if clean, defer dirty pages to writeback */
1080 if (referenced_page
&& !PageSwapBacked(page
))
1081 return PAGEREF_RECLAIM_CLEAN
;
1083 return PAGEREF_RECLAIM
;
1086 /* Check if a page is dirty or under writeback */
1087 static void page_check_dirty_writeback(struct page
*page
,
1088 bool *dirty
, bool *writeback
)
1090 struct address_space
*mapping
;
1093 * Anonymous pages are not handled by flushers and must be written
1094 * from reclaim context. Do not stall reclaim based on them
1096 if (!page_is_file_cache(page
) ||
1097 (PageAnon(page
) && !PageSwapBacked(page
))) {
1103 /* By default assume that the page flags are accurate */
1104 *dirty
= PageDirty(page
);
1105 *writeback
= PageWriteback(page
);
1107 /* Verify dirty/writeback state if the filesystem supports it */
1108 if (!page_has_private(page
))
1111 mapping
= page_mapping(page
);
1112 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1113 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1117 * shrink_page_list() returns the number of reclaimed pages
1119 static unsigned long shrink_page_list(struct list_head
*page_list
,
1120 struct pglist_data
*pgdat
,
1121 struct scan_control
*sc
,
1122 enum ttu_flags ttu_flags
,
1123 struct reclaim_stat
*stat
,
1124 bool ignore_references
)
1126 LIST_HEAD(ret_pages
);
1127 LIST_HEAD(free_pages
);
1128 unsigned nr_reclaimed
= 0;
1129 unsigned pgactivate
= 0;
1131 memset(stat
, 0, sizeof(*stat
));
1134 while (!list_empty(page_list
)) {
1135 struct address_space
*mapping
;
1138 enum page_references references
= PAGEREF_RECLAIM
;
1139 bool dirty
, writeback
;
1140 unsigned int nr_pages
;
1144 page
= lru_to_page(page_list
);
1145 list_del(&page
->lru
);
1147 if (!trylock_page(page
))
1150 VM_BUG_ON_PAGE(PageActive(page
), page
);
1152 nr_pages
= compound_nr(page
);
1154 /* Account the number of base pages even though THP */
1155 sc
->nr_scanned
+= nr_pages
;
1157 if (unlikely(!page_evictable(page
)))
1158 goto activate_locked
;
1160 if (!sc
->may_unmap
&& page_mapped(page
))
1163 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1164 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1167 * The number of dirty pages determines if a node is marked
1168 * reclaim_congested which affects wait_iff_congested. kswapd
1169 * will stall and start writing pages if the tail of the LRU
1170 * is all dirty unqueued pages.
1172 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1173 if (dirty
|| writeback
)
1176 if (dirty
&& !writeback
)
1177 stat
->nr_unqueued_dirty
++;
1180 * Treat this page as congested if the underlying BDI is or if
1181 * pages are cycling through the LRU so quickly that the
1182 * pages marked for immediate reclaim are making it to the
1183 * end of the LRU a second time.
1185 mapping
= page_mapping(page
);
1186 if (((dirty
|| writeback
) && mapping
&&
1187 inode_write_congested(mapping
->host
)) ||
1188 (writeback
&& PageReclaim(page
)))
1189 stat
->nr_congested
++;
1192 * If a page at the tail of the LRU is under writeback, there
1193 * are three cases to consider.
1195 * 1) If reclaim is encountering an excessive number of pages
1196 * under writeback and this page is both under writeback and
1197 * PageReclaim then it indicates that pages are being queued
1198 * for IO but are being recycled through the LRU before the
1199 * IO can complete. Waiting on the page itself risks an
1200 * indefinite stall if it is impossible to writeback the
1201 * page due to IO error or disconnected storage so instead
1202 * note that the LRU is being scanned too quickly and the
1203 * caller can stall after page list has been processed.
1205 * 2) Global or new memcg reclaim encounters a page that is
1206 * not marked for immediate reclaim, or the caller does not
1207 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 * not to fs). In this case mark the page for immediate
1209 * reclaim and continue scanning.
1211 * Require may_enter_fs because we would wait on fs, which
1212 * may not have submitted IO yet. And the loop driver might
1213 * enter reclaim, and deadlock if it waits on a page for
1214 * which it is needed to do the write (loop masks off
1215 * __GFP_IO|__GFP_FS for this reason); but more thought
1216 * would probably show more reasons.
1218 * 3) Legacy memcg encounters a page that is already marked
1219 * PageReclaim. memcg does not have any dirty pages
1220 * throttling so we could easily OOM just because too many
1221 * pages are in writeback and there is nothing else to
1222 * reclaim. Wait for the writeback to complete.
1224 * In cases 1) and 2) we activate the pages to get them out of
1225 * the way while we continue scanning for clean pages on the
1226 * inactive list and refilling from the active list. The
1227 * observation here is that waiting for disk writes is more
1228 * expensive than potentially causing reloads down the line.
1229 * Since they're marked for immediate reclaim, they won't put
1230 * memory pressure on the cache working set any longer than it
1231 * takes to write them to disk.
1233 if (PageWriteback(page
)) {
1235 if (current_is_kswapd() &&
1236 PageReclaim(page
) &&
1237 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1238 stat
->nr_immediate
++;
1239 goto activate_locked
;
1242 } else if (sane_reclaim(sc
) ||
1243 !PageReclaim(page
) || !may_enter_fs
) {
1245 * This is slightly racy - end_page_writeback()
1246 * might have just cleared PageReclaim, then
1247 * setting PageReclaim here end up interpreted
1248 * as PageReadahead - but that does not matter
1249 * enough to care. What we do want is for this
1250 * page to have PageReclaim set next time memcg
1251 * reclaim reaches the tests above, so it will
1252 * then wait_on_page_writeback() to avoid OOM;
1253 * and it's also appropriate in global reclaim.
1255 SetPageReclaim(page
);
1256 stat
->nr_writeback
++;
1257 goto activate_locked
;
1262 wait_on_page_writeback(page
);
1263 /* then go back and try same page again */
1264 list_add_tail(&page
->lru
, page_list
);
1269 if (!ignore_references
)
1270 references
= page_check_references(page
, sc
);
1272 switch (references
) {
1273 case PAGEREF_ACTIVATE
:
1274 goto activate_locked
;
1276 stat
->nr_ref_keep
+= nr_pages
;
1278 case PAGEREF_RECLAIM
:
1279 case PAGEREF_RECLAIM_CLEAN
:
1280 ; /* try to reclaim the page below */
1284 * Anonymous process memory has backing store?
1285 * Try to allocate it some swap space here.
1286 * Lazyfree page could be freed directly
1288 if (PageAnon(page
) && PageSwapBacked(page
)) {
1289 if (!PageSwapCache(page
)) {
1290 if (!(sc
->gfp_mask
& __GFP_IO
))
1292 if (PageTransHuge(page
)) {
1293 /* cannot split THP, skip it */
1294 if (!can_split_huge_page(page
, NULL
))
1295 goto activate_locked
;
1297 * Split pages without a PMD map right
1298 * away. Chances are some or all of the
1299 * tail pages can be freed without IO.
1301 if (!compound_mapcount(page
) &&
1302 split_huge_page_to_list(page
,
1304 goto activate_locked
;
1306 if (!add_to_swap(page
)) {
1307 if (!PageTransHuge(page
))
1308 goto activate_locked_split
;
1309 /* Fallback to swap normal pages */
1310 if (split_huge_page_to_list(page
,
1312 goto activate_locked
;
1313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 count_vm_event(THP_SWPOUT_FALLBACK
);
1316 if (!add_to_swap(page
))
1317 goto activate_locked_split
;
1322 /* Adding to swap updated mapping */
1323 mapping
= page_mapping(page
);
1325 } else if (unlikely(PageTransHuge(page
))) {
1326 /* Split file THP */
1327 if (split_huge_page_to_list(page
, page_list
))
1332 * THP may get split above, need minus tail pages and update
1333 * nr_pages to avoid accounting tail pages twice.
1335 * The tail pages that are added into swap cache successfully
1338 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1339 sc
->nr_scanned
-= (nr_pages
- 1);
1344 * The page is mapped into the page tables of one or more
1345 * processes. Try to unmap it here.
1347 if (page_mapped(page
)) {
1348 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1350 if (unlikely(PageTransHuge(page
)))
1351 flags
|= TTU_SPLIT_HUGE_PMD
;
1352 if (!try_to_unmap(page
, flags
)) {
1353 stat
->nr_unmap_fail
+= nr_pages
;
1354 goto activate_locked
;
1358 if (PageDirty(page
)) {
1360 * Only kswapd can writeback filesystem pages
1361 * to avoid risk of stack overflow. But avoid
1362 * injecting inefficient single-page IO into
1363 * flusher writeback as much as possible: only
1364 * write pages when we've encountered many
1365 * dirty pages, and when we've already scanned
1366 * the rest of the LRU for clean pages and see
1367 * the same dirty pages again (PageReclaim).
1369 if (page_is_file_cache(page
) &&
1370 (!current_is_kswapd() || !PageReclaim(page
) ||
1371 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1373 * Immediately reclaim when written back.
1374 * Similar in principal to deactivate_page()
1375 * except we already have the page isolated
1376 * and know it's dirty
1378 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1379 SetPageReclaim(page
);
1381 goto activate_locked
;
1384 if (references
== PAGEREF_RECLAIM_CLEAN
)
1388 if (!sc
->may_writepage
)
1392 * Page is dirty. Flush the TLB if a writable entry
1393 * potentially exists to avoid CPU writes after IO
1394 * starts and then write it out here.
1396 try_to_unmap_flush_dirty();
1397 switch (pageout(page
, mapping
, sc
)) {
1401 goto activate_locked
;
1403 if (PageWriteback(page
))
1405 if (PageDirty(page
))
1409 * A synchronous write - probably a ramdisk. Go
1410 * ahead and try to reclaim the page.
1412 if (!trylock_page(page
))
1414 if (PageDirty(page
) || PageWriteback(page
))
1416 mapping
= page_mapping(page
);
1418 ; /* try to free the page below */
1423 * If the page has buffers, try to free the buffer mappings
1424 * associated with this page. If we succeed we try to free
1427 * We do this even if the page is PageDirty().
1428 * try_to_release_page() does not perform I/O, but it is
1429 * possible for a page to have PageDirty set, but it is actually
1430 * clean (all its buffers are clean). This happens if the
1431 * buffers were written out directly, with submit_bh(). ext3
1432 * will do this, as well as the blockdev mapping.
1433 * try_to_release_page() will discover that cleanness and will
1434 * drop the buffers and mark the page clean - it can be freed.
1436 * Rarely, pages can have buffers and no ->mapping. These are
1437 * the pages which were not successfully invalidated in
1438 * truncate_complete_page(). We try to drop those buffers here
1439 * and if that worked, and the page is no longer mapped into
1440 * process address space (page_count == 1) it can be freed.
1441 * Otherwise, leave the page on the LRU so it is swappable.
1443 if (page_has_private(page
)) {
1444 if (!try_to_release_page(page
, sc
->gfp_mask
))
1445 goto activate_locked
;
1446 if (!mapping
&& page_count(page
) == 1) {
1448 if (put_page_testzero(page
))
1452 * rare race with speculative reference.
1453 * the speculative reference will free
1454 * this page shortly, so we may
1455 * increment nr_reclaimed here (and
1456 * leave it off the LRU).
1464 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1465 /* follow __remove_mapping for reference */
1466 if (!page_ref_freeze(page
, 1))
1468 if (PageDirty(page
)) {
1469 page_ref_unfreeze(page
, 1);
1473 count_vm_event(PGLAZYFREED
);
1474 count_memcg_page_event(page
, PGLAZYFREED
);
1475 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1481 * THP may get swapped out in a whole, need account
1484 nr_reclaimed
+= nr_pages
;
1487 * Is there need to periodically free_page_list? It would
1488 * appear not as the counts should be low
1490 if (unlikely(PageTransHuge(page
)))
1491 (*get_compound_page_dtor(page
))(page
);
1493 list_add(&page
->lru
, &free_pages
);
1496 activate_locked_split
:
1498 * The tail pages that are failed to add into swap cache
1499 * reach here. Fixup nr_scanned and nr_pages.
1502 sc
->nr_scanned
-= (nr_pages
- 1);
1506 /* Not a candidate for swapping, so reclaim swap space. */
1507 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1509 try_to_free_swap(page
);
1510 VM_BUG_ON_PAGE(PageActive(page
), page
);
1511 if (!PageMlocked(page
)) {
1512 int type
= page_is_file_cache(page
);
1513 SetPageActive(page
);
1514 stat
->nr_activate
[type
] += nr_pages
;
1515 count_memcg_page_event(page
, PGACTIVATE
);
1520 list_add(&page
->lru
, &ret_pages
);
1521 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1524 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1526 mem_cgroup_uncharge_list(&free_pages
);
1527 try_to_unmap_flush();
1528 free_unref_page_list(&free_pages
);
1530 list_splice(&ret_pages
, page_list
);
1531 count_vm_events(PGACTIVATE
, pgactivate
);
1533 return nr_reclaimed
;
1536 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1537 struct list_head
*page_list
)
1539 struct scan_control sc
= {
1540 .gfp_mask
= GFP_KERNEL
,
1541 .priority
= DEF_PRIORITY
,
1544 struct reclaim_stat dummy_stat
;
1546 struct page
*page
, *next
;
1547 LIST_HEAD(clean_pages
);
1549 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1550 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1551 !__PageMovable(page
) && !PageUnevictable(page
)) {
1552 ClearPageActive(page
);
1553 list_move(&page
->lru
, &clean_pages
);
1557 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1558 TTU_IGNORE_ACCESS
, &dummy_stat
, true);
1559 list_splice(&clean_pages
, page_list
);
1560 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1565 * Attempt to remove the specified page from its LRU. Only take this page
1566 * if it is of the appropriate PageActive status. Pages which are being
1567 * freed elsewhere are also ignored.
1569 * page: page to consider
1570 * mode: one of the LRU isolation modes defined above
1572 * returns 0 on success, -ve errno on failure.
1574 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1578 /* Only take pages on the LRU. */
1582 /* Compaction should not handle unevictable pages but CMA can do so */
1583 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1589 * To minimise LRU disruption, the caller can indicate that it only
1590 * wants to isolate pages it will be able to operate on without
1591 * blocking - clean pages for the most part.
1593 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 * that it is possible to migrate without blocking
1596 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1597 /* All the caller can do on PageWriteback is block */
1598 if (PageWriteback(page
))
1601 if (PageDirty(page
)) {
1602 struct address_space
*mapping
;
1606 * Only pages without mappings or that have a
1607 * ->migratepage callback are possible to migrate
1608 * without blocking. However, we can be racing with
1609 * truncation so it's necessary to lock the page
1610 * to stabilise the mapping as truncation holds
1611 * the page lock until after the page is removed
1612 * from the page cache.
1614 if (!trylock_page(page
))
1617 mapping
= page_mapping(page
);
1618 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1625 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1628 if (likely(get_page_unless_zero(page
))) {
1630 * Be careful not to clear PageLRU until after we're
1631 * sure the page is not being freed elsewhere -- the
1632 * page release code relies on it.
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1646 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1647 enum lru_list lru
, unsigned long *nr_zone_taken
)
1651 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1652 if (!nr_zone_taken
[zid
])
1655 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1657 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1664 * pgdat->lru_lock is heavily contended. Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1671 * Appropriate locks must be held before calling this function.
1673 * @nr_to_scan: The number of eligible pages to look through on the list.
1674 * @lruvec: The LRU vector to pull pages from.
1675 * @dst: The temp list to put pages on to.
1676 * @nr_scanned: The number of pages that were scanned.
1677 * @sc: The scan_control struct for this reclaim session
1678 * @mode: One of the LRU isolation modes
1679 * @lru: LRU list id for isolating
1681 * returns how many pages were moved onto *@dst.
1683 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1684 struct lruvec
*lruvec
, struct list_head
*dst
,
1685 unsigned long *nr_scanned
, struct scan_control
*sc
,
1688 struct list_head
*src
= &lruvec
->lists
[lru
];
1689 unsigned long nr_taken
= 0;
1690 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1691 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1692 unsigned long skipped
= 0;
1693 unsigned long scan
, total_scan
, nr_pages
;
1694 LIST_HEAD(pages_skipped
);
1695 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1699 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1702 page
= lru_to_page(src
);
1703 prefetchw_prev_lru_page(page
, src
, flags
);
1705 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1707 nr_pages
= compound_nr(page
);
1708 total_scan
+= nr_pages
;
1710 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1711 list_move(&page
->lru
, &pages_skipped
);
1712 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1717 * Do not count skipped pages because that makes the function
1718 * return with no isolated pages if the LRU mostly contains
1719 * ineligible pages. This causes the VM to not reclaim any
1720 * pages, triggering a premature OOM.
1722 * Account all tail pages of THP. This would not cause
1723 * premature OOM since __isolate_lru_page() returns -EBUSY
1724 * only when the page is being freed somewhere else.
1727 switch (__isolate_lru_page(page
, mode
)) {
1729 nr_taken
+= nr_pages
;
1730 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1731 list_move(&page
->lru
, dst
);
1735 /* else it is being freed elsewhere */
1736 list_move(&page
->lru
, src
);
1745 * Splice any skipped pages to the start of the LRU list. Note that
1746 * this disrupts the LRU order when reclaiming for lower zones but
1747 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 * scanning would soon rescan the same pages to skip and put the
1749 * system at risk of premature OOM.
1751 if (!list_empty(&pages_skipped
)) {
1754 list_splice(&pages_skipped
, src
);
1755 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1756 if (!nr_skipped
[zid
])
1759 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1760 skipped
+= nr_skipped
[zid
];
1763 *nr_scanned
= total_scan
;
1764 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1765 total_scan
, skipped
, nr_taken
, mode
, lru
);
1766 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1780 * The returned page will have PageLRU() cleared. If it was found on
1781 * the active list, it will have PageActive set. If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 * fundamentnal difference from isolate_lru_pages (which is called
1792 * without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1796 int isolate_lru_page(struct page
*page
)
1800 VM_BUG_ON_PAGE(!page_count(page
), page
);
1801 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1803 if (PageLRU(page
)) {
1804 pg_data_t
*pgdat
= page_pgdat(page
);
1805 struct lruvec
*lruvec
;
1807 spin_lock_irq(&pgdat
->lru_lock
);
1808 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1809 if (PageLRU(page
)) {
1810 int lru
= page_lru(page
);
1813 del_page_from_lru_list(page
, lruvec
, lru
);
1816 spin_unlock_irq(&pgdat
->lru_lock
);
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1828 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1829 struct scan_control
*sc
)
1831 unsigned long inactive
, isolated
;
1833 if (current_is_kswapd())
1836 if (!sane_reclaim(sc
))
1840 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1841 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1843 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1844 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1848 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 * won't get blocked by normal direct-reclaimers, forming a circular
1852 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1855 return isolated
> inactive
;
1859 * This moves pages from @list to corresponding LRU list.
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation. But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page. It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1875 * Returns the number of pages moved to the given lruvec.
1878 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1879 struct list_head
*list
)
1881 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1882 int nr_pages
, nr_moved
= 0;
1883 LIST_HEAD(pages_to_free
);
1887 while (!list_empty(list
)) {
1888 page
= lru_to_page(list
);
1889 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1890 if (unlikely(!page_evictable(page
))) {
1891 list_del(&page
->lru
);
1892 spin_unlock_irq(&pgdat
->lru_lock
);
1893 putback_lru_page(page
);
1894 spin_lock_irq(&pgdat
->lru_lock
);
1897 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1900 lru
= page_lru(page
);
1902 nr_pages
= hpage_nr_pages(page
);
1903 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1904 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1906 if (put_page_testzero(page
)) {
1907 __ClearPageLRU(page
);
1908 __ClearPageActive(page
);
1909 del_page_from_lru_list(page
, lruvec
, lru
);
1911 if (unlikely(PageCompound(page
))) {
1912 spin_unlock_irq(&pgdat
->lru_lock
);
1913 (*get_compound_page_dtor(page
))(page
);
1914 spin_lock_irq(&pgdat
->lru_lock
);
1916 list_add(&page
->lru
, &pages_to_free
);
1918 nr_moved
+= nr_pages
;
1923 * To save our caller's stack, now use input list for pages to free.
1925 list_splice(&pages_to_free
, list
);
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested. In other cases it is safe to throttle.
1936 static int current_may_throttle(void)
1938 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1939 current
->backing_dev_info
== NULL
||
1940 bdi_write_congested(current
->backing_dev_info
);
1944 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 * of reclaimed pages
1947 static noinline_for_stack
unsigned long
1948 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1949 struct scan_control
*sc
, enum lru_list lru
)
1951 LIST_HEAD(page_list
);
1952 unsigned long nr_scanned
;
1953 unsigned long nr_reclaimed
= 0;
1954 unsigned long nr_taken
;
1955 struct reclaim_stat stat
;
1956 int file
= is_file_lru(lru
);
1957 enum vm_event_item item
;
1958 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1959 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1960 bool stalled
= false;
1962 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1966 /* wait a bit for the reclaimer. */
1970 /* We are about to die and free our memory. Return now. */
1971 if (fatal_signal_pending(current
))
1972 return SWAP_CLUSTER_MAX
;
1977 spin_lock_irq(&pgdat
->lru_lock
);
1979 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1980 &nr_scanned
, sc
, lru
);
1982 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1983 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1985 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1986 if (global_reclaim(sc
))
1987 __count_vm_events(item
, nr_scanned
);
1988 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1989 spin_unlock_irq(&pgdat
->lru_lock
);
1994 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1997 spin_lock_irq(&pgdat
->lru_lock
);
1999 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2000 if (global_reclaim(sc
))
2001 __count_vm_events(item
, nr_reclaimed
);
2002 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2003 reclaim_stat
->recent_rotated
[0] += stat
.nr_activate
[0];
2004 reclaim_stat
->recent_rotated
[1] += stat
.nr_activate
[1];
2006 move_pages_to_lru(lruvec
, &page_list
);
2008 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2010 spin_unlock_irq(&pgdat
->lru_lock
);
2012 mem_cgroup_uncharge_list(&page_list
);
2013 free_unref_page_list(&page_list
);
2016 * If dirty pages are scanned that are not queued for IO, it
2017 * implies that flushers are not doing their job. This can
2018 * happen when memory pressure pushes dirty pages to the end of
2019 * the LRU before the dirty limits are breached and the dirty
2020 * data has expired. It can also happen when the proportion of
2021 * dirty pages grows not through writes but through memory
2022 * pressure reclaiming all the clean cache. And in some cases,
2023 * the flushers simply cannot keep up with the allocation
2024 * rate. Nudge the flusher threads in case they are asleep.
2026 if (stat
.nr_unqueued_dirty
== nr_taken
)
2027 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2029 sc
->nr
.dirty
+= stat
.nr_dirty
;
2030 sc
->nr
.congested
+= stat
.nr_congested
;
2031 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2032 sc
->nr
.writeback
+= stat
.nr_writeback
;
2033 sc
->nr
.immediate
+= stat
.nr_immediate
;
2034 sc
->nr
.taken
+= nr_taken
;
2036 sc
->nr
.file_taken
+= nr_taken
;
2038 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2039 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2040 return nr_reclaimed
;
2043 static void shrink_active_list(unsigned long nr_to_scan
,
2044 struct lruvec
*lruvec
,
2045 struct scan_control
*sc
,
2048 unsigned long nr_taken
;
2049 unsigned long nr_scanned
;
2050 unsigned long vm_flags
;
2051 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2052 LIST_HEAD(l_active
);
2053 LIST_HEAD(l_inactive
);
2055 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2056 unsigned nr_deactivate
, nr_activate
;
2057 unsigned nr_rotated
= 0;
2058 int file
= is_file_lru(lru
);
2059 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2063 spin_lock_irq(&pgdat
->lru_lock
);
2065 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2066 &nr_scanned
, sc
, lru
);
2068 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2069 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2071 __count_vm_events(PGREFILL
, nr_scanned
);
2072 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2074 spin_unlock_irq(&pgdat
->lru_lock
);
2076 while (!list_empty(&l_hold
)) {
2078 page
= lru_to_page(&l_hold
);
2079 list_del(&page
->lru
);
2081 if (unlikely(!page_evictable(page
))) {
2082 putback_lru_page(page
);
2086 if (unlikely(buffer_heads_over_limit
)) {
2087 if (page_has_private(page
) && trylock_page(page
)) {
2088 if (page_has_private(page
))
2089 try_to_release_page(page
, 0);
2094 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2096 nr_rotated
+= hpage_nr_pages(page
);
2098 * Identify referenced, file-backed active pages and
2099 * give them one more trip around the active list. So
2100 * that executable code get better chances to stay in
2101 * memory under moderate memory pressure. Anon pages
2102 * are not likely to be evicted by use-once streaming
2103 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 * so we ignore them here.
2106 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2107 list_add(&page
->lru
, &l_active
);
2112 ClearPageActive(page
); /* we are de-activating */
2113 SetPageWorkingset(page
);
2114 list_add(&page
->lru
, &l_inactive
);
2118 * Move pages back to the lru list.
2120 spin_lock_irq(&pgdat
->lru_lock
);
2122 * Count referenced pages from currently used mappings as rotated,
2123 * even though only some of them are actually re-activated. This
2124 * helps balance scan pressure between file and anonymous pages in
2127 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2129 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2130 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2131 /* Keep all free pages in l_active list */
2132 list_splice(&l_inactive
, &l_active
);
2134 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2135 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2137 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2138 spin_unlock_irq(&pgdat
->lru_lock
);
2140 mem_cgroup_uncharge_list(&l_active
);
2141 free_unref_page_list(&l_active
);
2142 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2143 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2146 unsigned long reclaim_pages(struct list_head
*page_list
)
2149 unsigned long nr_reclaimed
= 0;
2150 LIST_HEAD(node_page_list
);
2151 struct reclaim_stat dummy_stat
;
2153 struct scan_control sc
= {
2154 .gfp_mask
= GFP_KERNEL
,
2155 .priority
= DEF_PRIORITY
,
2161 while (!list_empty(page_list
)) {
2162 page
= lru_to_page(page_list
);
2164 nid
= page_to_nid(page
);
2165 INIT_LIST_HEAD(&node_page_list
);
2168 if (nid
== page_to_nid(page
)) {
2169 ClearPageActive(page
);
2170 list_move(&page
->lru
, &node_page_list
);
2174 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2177 &dummy_stat
, false);
2178 while (!list_empty(&node_page_list
)) {
2179 page
= lru_to_page(&node_page_list
);
2180 list_del(&page
->lru
);
2181 putback_lru_page(page
);
2187 if (!list_empty(&node_page_list
)) {
2188 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2191 &dummy_stat
, false);
2192 while (!list_empty(&node_page_list
)) {
2193 page
= lru_to_page(&node_page_list
);
2194 list_del(&page
->lru
);
2195 putback_lru_page(page
);
2199 return nr_reclaimed
;
2203 * The inactive anon list should be small enough that the VM never has
2204 * to do too much work.
2206 * The inactive file list should be small enough to leave most memory
2207 * to the established workingset on the scan-resistant active list,
2208 * but large enough to avoid thrashing the aggregate readahead window.
2210 * Both inactive lists should also be large enough that each inactive
2211 * page has a chance to be referenced again before it is reclaimed.
2213 * If that fails and refaulting is observed, the inactive list grows.
2215 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 * on this LRU, maintained by the pageout code. An inactive_ratio
2217 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2220 * memory ratio inactive
2221 * -------------------------------------
2230 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2231 struct scan_control
*sc
, bool trace
)
2233 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2234 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2235 enum lru_list inactive_lru
= file
* LRU_FILE
;
2236 unsigned long inactive
, active
;
2237 unsigned long inactive_ratio
;
2238 unsigned long refaults
;
2242 * If we don't have swap space, anonymous page deactivation
2245 if (!file
&& !total_swap_pages
)
2248 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2249 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2252 * When refaults are being observed, it means a new workingset
2253 * is being established. Disable active list protection to get
2254 * rid of the stale workingset quickly.
2256 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
2257 if (file
&& lruvec
->refaults
!= refaults
) {
2260 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2262 inactive_ratio
= int_sqrt(10 * gb
);
2268 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2269 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2270 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2271 inactive_ratio
, file
);
2273 return inactive
* inactive_ratio
< active
;
2276 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2277 struct lruvec
*lruvec
, struct scan_control
*sc
)
2279 if (is_active_lru(lru
)) {
2280 if (inactive_list_is_low(lruvec
, is_file_lru(lru
), sc
, true))
2281 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2285 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2296 * Determine how aggressively the anon and file LRU lists should be
2297 * scanned. The relative value of each set of LRU lists is determined
2298 * by looking at the fraction of the pages scanned we did rotate back
2299 * onto the active list instead of evict.
2301 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2304 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2305 struct scan_control
*sc
, unsigned long *nr
,
2306 unsigned long *lru_pages
)
2308 int swappiness
= mem_cgroup_swappiness(memcg
);
2309 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2311 u64 denominator
= 0; /* gcc */
2312 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2313 unsigned long anon_prio
, file_prio
;
2314 enum scan_balance scan_balance
;
2315 unsigned long anon
, file
;
2316 unsigned long ap
, fp
;
2319 /* If we have no swap space, do not bother scanning anon pages. */
2320 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2321 scan_balance
= SCAN_FILE
;
2326 * Global reclaim will swap to prevent OOM even with no
2327 * swappiness, but memcg users want to use this knob to
2328 * disable swapping for individual groups completely when
2329 * using the memory controller's swap limit feature would be
2332 if (!global_reclaim(sc
) && !swappiness
) {
2333 scan_balance
= SCAN_FILE
;
2338 * Do not apply any pressure balancing cleverness when the
2339 * system is close to OOM, scan both anon and file equally
2340 * (unless the swappiness setting disagrees with swapping).
2342 if (!sc
->priority
&& swappiness
) {
2343 scan_balance
= SCAN_EQUAL
;
2348 * Prevent the reclaimer from falling into the cache trap: as
2349 * cache pages start out inactive, every cache fault will tip
2350 * the scan balance towards the file LRU. And as the file LRU
2351 * shrinks, so does the window for rotation from references.
2352 * This means we have a runaway feedback loop where a tiny
2353 * thrashing file LRU becomes infinitely more attractive than
2354 * anon pages. Try to detect this based on file LRU size.
2356 if (global_reclaim(sc
)) {
2357 unsigned long pgdatfile
;
2358 unsigned long pgdatfree
;
2360 unsigned long total_high_wmark
= 0;
2362 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2363 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2364 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2366 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2367 struct zone
*zone
= &pgdat
->node_zones
[z
];
2368 if (!managed_zone(zone
))
2371 total_high_wmark
+= high_wmark_pages(zone
);
2374 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2376 * Force SCAN_ANON if there are enough inactive
2377 * anonymous pages on the LRU in eligible zones.
2378 * Otherwise, the small LRU gets thrashed.
2380 if (!inactive_list_is_low(lruvec
, false, sc
, false) &&
2381 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2383 scan_balance
= SCAN_ANON
;
2390 * If there is enough inactive page cache, i.e. if the size of the
2391 * inactive list is greater than that of the active list *and* the
2392 * inactive list actually has some pages to scan on this priority, we
2393 * do not reclaim anything from the anonymous working set right now.
2394 * Without the second condition we could end up never scanning an
2395 * lruvec even if it has plenty of old anonymous pages unless the
2396 * system is under heavy pressure.
2398 if (!inactive_list_is_low(lruvec
, true, sc
, false) &&
2399 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2400 scan_balance
= SCAN_FILE
;
2404 scan_balance
= SCAN_FRACT
;
2407 * With swappiness at 100, anonymous and file have the same priority.
2408 * This scanning priority is essentially the inverse of IO cost.
2410 anon_prio
= swappiness
;
2411 file_prio
= 200 - anon_prio
;
2414 * OK, so we have swap space and a fair amount of page cache
2415 * pages. We use the recently rotated / recently scanned
2416 * ratios to determine how valuable each cache is.
2418 * Because workloads change over time (and to avoid overflow)
2419 * we keep these statistics as a floating average, which ends
2420 * up weighing recent references more than old ones.
2422 * anon in [0], file in [1]
2425 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2426 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2427 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2428 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2430 spin_lock_irq(&pgdat
->lru_lock
);
2431 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2432 reclaim_stat
->recent_scanned
[0] /= 2;
2433 reclaim_stat
->recent_rotated
[0] /= 2;
2436 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2437 reclaim_stat
->recent_scanned
[1] /= 2;
2438 reclaim_stat
->recent_rotated
[1] /= 2;
2442 * The amount of pressure on anon vs file pages is inversely
2443 * proportional to the fraction of recently scanned pages on
2444 * each list that were recently referenced and in active use.
2446 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2447 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2449 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2450 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2451 spin_unlock_irq(&pgdat
->lru_lock
);
2455 denominator
= ap
+ fp
+ 1;
2458 for_each_evictable_lru(lru
) {
2459 int file
= is_file_lru(lru
);
2460 unsigned long lruvec_size
;
2462 unsigned long protection
;
2464 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2465 protection
= mem_cgroup_protection(memcg
,
2466 sc
->memcg_low_reclaim
);
2470 * Scale a cgroup's reclaim pressure by proportioning
2471 * its current usage to its memory.low or memory.min
2474 * This is important, as otherwise scanning aggression
2475 * becomes extremely binary -- from nothing as we
2476 * approach the memory protection threshold, to totally
2477 * nominal as we exceed it. This results in requiring
2478 * setting extremely liberal protection thresholds. It
2479 * also means we simply get no protection at all if we
2480 * set it too low, which is not ideal.
2482 * If there is any protection in place, we reduce scan
2483 * pressure by how much of the total memory used is
2484 * within protection thresholds.
2486 * There is one special case: in the first reclaim pass,
2487 * we skip over all groups that are within their low
2488 * protection. If that fails to reclaim enough pages to
2489 * satisfy the reclaim goal, we come back and override
2490 * the best-effort low protection. However, we still
2491 * ideally want to honor how well-behaved groups are in
2492 * that case instead of simply punishing them all
2493 * equally. As such, we reclaim them based on how much
2494 * memory they are using, reducing the scan pressure
2495 * again by how much of the total memory used is under
2498 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2500 /* Avoid TOCTOU with earlier protection check */
2501 cgroup_size
= max(cgroup_size
, protection
);
2503 scan
= lruvec_size
- lruvec_size
* protection
/
2507 * Minimally target SWAP_CLUSTER_MAX pages to keep
2508 * reclaim moving forwards, avoiding decremeting
2509 * sc->priority further than desirable.
2511 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2516 scan
>>= sc
->priority
;
2519 * If the cgroup's already been deleted, make sure to
2520 * scrape out the remaining cache.
2522 if (!scan
&& !mem_cgroup_online(memcg
))
2523 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2525 switch (scan_balance
) {
2527 /* Scan lists relative to size */
2531 * Scan types proportional to swappiness and
2532 * their relative recent reclaim efficiency.
2533 * Make sure we don't miss the last page on
2534 * the offlined memory cgroups because of a
2537 scan
= mem_cgroup_online(memcg
) ?
2538 div64_u64(scan
* fraction
[file
], denominator
) :
2539 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2544 /* Scan one type exclusively */
2545 if ((scan_balance
== SCAN_FILE
) != file
) {
2551 /* Look ma, no brain */
2555 *lru_pages
+= lruvec_size
;
2561 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2563 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2564 struct scan_control
*sc
, unsigned long *lru_pages
)
2566 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2567 unsigned long nr
[NR_LRU_LISTS
];
2568 unsigned long targets
[NR_LRU_LISTS
];
2569 unsigned long nr_to_scan
;
2571 unsigned long nr_reclaimed
= 0;
2572 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2573 struct blk_plug plug
;
2576 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2578 /* Record the original scan target for proportional adjustments later */
2579 memcpy(targets
, nr
, sizeof(nr
));
2582 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2583 * event that can occur when there is little memory pressure e.g.
2584 * multiple streaming readers/writers. Hence, we do not abort scanning
2585 * when the requested number of pages are reclaimed when scanning at
2586 * DEF_PRIORITY on the assumption that the fact we are direct
2587 * reclaiming implies that kswapd is not keeping up and it is best to
2588 * do a batch of work at once. For memcg reclaim one check is made to
2589 * abort proportional reclaim if either the file or anon lru has already
2590 * dropped to zero at the first pass.
2592 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2593 sc
->priority
== DEF_PRIORITY
);
2595 blk_start_plug(&plug
);
2596 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2597 nr
[LRU_INACTIVE_FILE
]) {
2598 unsigned long nr_anon
, nr_file
, percentage
;
2599 unsigned long nr_scanned
;
2601 for_each_evictable_lru(lru
) {
2603 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2604 nr
[lru
] -= nr_to_scan
;
2606 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2613 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2617 * For kswapd and memcg, reclaim at least the number of pages
2618 * requested. Ensure that the anon and file LRUs are scanned
2619 * proportionally what was requested by get_scan_count(). We
2620 * stop reclaiming one LRU and reduce the amount scanning
2621 * proportional to the original scan target.
2623 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2624 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2627 * It's just vindictive to attack the larger once the smaller
2628 * has gone to zero. And given the way we stop scanning the
2629 * smaller below, this makes sure that we only make one nudge
2630 * towards proportionality once we've got nr_to_reclaim.
2632 if (!nr_file
|| !nr_anon
)
2635 if (nr_file
> nr_anon
) {
2636 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2637 targets
[LRU_ACTIVE_ANON
] + 1;
2639 percentage
= nr_anon
* 100 / scan_target
;
2641 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2642 targets
[LRU_ACTIVE_FILE
] + 1;
2644 percentage
= nr_file
* 100 / scan_target
;
2647 /* Stop scanning the smaller of the LRU */
2649 nr
[lru
+ LRU_ACTIVE
] = 0;
2652 * Recalculate the other LRU scan count based on its original
2653 * scan target and the percentage scanning already complete
2655 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2656 nr_scanned
= targets
[lru
] - nr
[lru
];
2657 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2658 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2661 nr_scanned
= targets
[lru
] - nr
[lru
];
2662 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2663 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2665 scan_adjusted
= true;
2667 blk_finish_plug(&plug
);
2668 sc
->nr_reclaimed
+= nr_reclaimed
;
2671 * Even if we did not try to evict anon pages at all, we want to
2672 * rebalance the anon lru active/inactive ratio.
2674 if (inactive_list_is_low(lruvec
, false, sc
, true))
2675 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2676 sc
, LRU_ACTIVE_ANON
);
2679 /* Use reclaim/compaction for costly allocs or under memory pressure */
2680 static bool in_reclaim_compaction(struct scan_control
*sc
)
2682 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2683 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2684 sc
->priority
< DEF_PRIORITY
- 2))
2691 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2692 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2693 * true if more pages should be reclaimed such that when the page allocator
2694 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2695 * It will give up earlier than that if there is difficulty reclaiming pages.
2697 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2698 unsigned long nr_reclaimed
,
2699 struct scan_control
*sc
)
2701 unsigned long pages_for_compaction
;
2702 unsigned long inactive_lru_pages
;
2705 /* If not in reclaim/compaction mode, stop */
2706 if (!in_reclaim_compaction(sc
))
2710 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2711 * number of pages that were scanned. This will return to the caller
2712 * with the risk reclaim/compaction and the resulting allocation attempt
2713 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2714 * allocations through requiring that the full LRU list has been scanned
2715 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2716 * scan, but that approximation was wrong, and there were corner cases
2717 * where always a non-zero amount of pages were scanned.
2722 /* If compaction would go ahead or the allocation would succeed, stop */
2723 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2724 struct zone
*zone
= &pgdat
->node_zones
[z
];
2725 if (!managed_zone(zone
))
2728 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2729 case COMPACT_SUCCESS
:
2730 case COMPACT_CONTINUE
:
2733 /* check next zone */
2739 * If we have not reclaimed enough pages for compaction and the
2740 * inactive lists are large enough, continue reclaiming
2742 pages_for_compaction
= compact_gap(sc
->order
);
2743 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2744 if (get_nr_swap_pages() > 0)
2745 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2747 return inactive_lru_pages
> pages_for_compaction
;
2750 static bool pgdat_memcg_congested(pg_data_t
*pgdat
, struct mem_cgroup
*memcg
)
2752 return test_bit(PGDAT_CONGESTED
, &pgdat
->flags
) ||
2753 (memcg
&& memcg_congested(pgdat
, memcg
));
2756 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2758 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2759 unsigned long nr_reclaimed
, nr_scanned
;
2760 bool reclaimable
= false;
2763 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2764 unsigned long node_lru_pages
= 0;
2765 struct mem_cgroup
*memcg
;
2767 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2769 nr_reclaimed
= sc
->nr_reclaimed
;
2770 nr_scanned
= sc
->nr_scanned
;
2772 memcg
= mem_cgroup_iter(root
, NULL
, NULL
);
2774 unsigned long lru_pages
;
2775 unsigned long reclaimed
;
2776 unsigned long scanned
;
2779 * This loop can become CPU-bound when target memcgs
2780 * aren't eligible for reclaim - either because they
2781 * don't have any reclaimable pages, or because their
2782 * memory is explicitly protected. Avoid soft lockups.
2786 switch (mem_cgroup_protected(root
, memcg
)) {
2787 case MEMCG_PROT_MIN
:
2790 * If there is no reclaimable memory, OOM.
2793 case MEMCG_PROT_LOW
:
2796 * Respect the protection only as long as
2797 * there is an unprotected supply
2798 * of reclaimable memory from other cgroups.
2800 if (!sc
->memcg_low_reclaim
) {
2801 sc
->memcg_low_skipped
= 1;
2804 memcg_memory_event(memcg
, MEMCG_LOW
);
2806 case MEMCG_PROT_NONE
:
2808 * All protection thresholds breached. We may
2809 * still choose to vary the scan pressure
2810 * applied based on by how much the cgroup in
2811 * question has exceeded its protection
2812 * thresholds (see get_scan_count).
2817 reclaimed
= sc
->nr_reclaimed
;
2818 scanned
= sc
->nr_scanned
;
2819 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2820 node_lru_pages
+= lru_pages
;
2822 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2825 /* Record the group's reclaim efficiency */
2826 vmpressure(sc
->gfp_mask
, memcg
, false,
2827 sc
->nr_scanned
- scanned
,
2828 sc
->nr_reclaimed
- reclaimed
);
2830 } while ((memcg
= mem_cgroup_iter(root
, memcg
, NULL
)));
2832 if (reclaim_state
) {
2833 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2834 reclaim_state
->reclaimed_slab
= 0;
2837 /* Record the subtree's reclaim efficiency */
2838 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2839 sc
->nr_scanned
- nr_scanned
,
2840 sc
->nr_reclaimed
- nr_reclaimed
);
2842 if (sc
->nr_reclaimed
- nr_reclaimed
)
2845 if (current_is_kswapd()) {
2847 * If reclaim is isolating dirty pages under writeback,
2848 * it implies that the long-lived page allocation rate
2849 * is exceeding the page laundering rate. Either the
2850 * global limits are not being effective at throttling
2851 * processes due to the page distribution throughout
2852 * zones or there is heavy usage of a slow backing
2853 * device. The only option is to throttle from reclaim
2854 * context which is not ideal as there is no guarantee
2855 * the dirtying process is throttled in the same way
2856 * balance_dirty_pages() manages.
2858 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2859 * count the number of pages under pages flagged for
2860 * immediate reclaim and stall if any are encountered
2861 * in the nr_immediate check below.
2863 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2864 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2867 * Tag a node as congested if all the dirty pages
2868 * scanned were backed by a congested BDI and
2869 * wait_iff_congested will stall.
2871 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2872 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
2874 /* Allow kswapd to start writing pages during reclaim.*/
2875 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2876 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2879 * If kswapd scans pages marked marked for immediate
2880 * reclaim and under writeback (nr_immediate), it
2881 * implies that pages are cycling through the LRU
2882 * faster than they are written so also forcibly stall.
2884 if (sc
->nr
.immediate
)
2885 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2889 * Legacy memcg will stall in page writeback so avoid forcibly
2890 * stalling in wait_iff_congested().
2892 if (!global_reclaim(sc
) && sane_reclaim(sc
) &&
2893 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2894 set_memcg_congestion(pgdat
, root
, true);
2897 * Stall direct reclaim for IO completions if underlying BDIs
2898 * and node is congested. Allow kswapd to continue until it
2899 * starts encountering unqueued dirty pages or cycling through
2900 * the LRU too quickly.
2902 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
2903 current_may_throttle() && pgdat_memcg_congested(pgdat
, root
))
2904 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2906 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2910 * Kswapd gives up on balancing particular nodes after too
2911 * many failures to reclaim anything from them and goes to
2912 * sleep. On reclaim progress, reset the failure counter. A
2913 * successful direct reclaim run will revive a dormant kswapd.
2916 pgdat
->kswapd_failures
= 0;
2922 * Returns true if compaction should go ahead for a costly-order request, or
2923 * the allocation would already succeed without compaction. Return false if we
2924 * should reclaim first.
2926 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2928 unsigned long watermark
;
2929 enum compact_result suitable
;
2931 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2932 if (suitable
== COMPACT_SUCCESS
)
2933 /* Allocation should succeed already. Don't reclaim. */
2935 if (suitable
== COMPACT_SKIPPED
)
2936 /* Compaction cannot yet proceed. Do reclaim. */
2940 * Compaction is already possible, but it takes time to run and there
2941 * are potentially other callers using the pages just freed. So proceed
2942 * with reclaim to make a buffer of free pages available to give
2943 * compaction a reasonable chance of completing and allocating the page.
2944 * Note that we won't actually reclaim the whole buffer in one attempt
2945 * as the target watermark in should_continue_reclaim() is lower. But if
2946 * we are already above the high+gap watermark, don't reclaim at all.
2948 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2950 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2954 * This is the direct reclaim path, for page-allocating processes. We only
2955 * try to reclaim pages from zones which will satisfy the caller's allocation
2958 * If a zone is deemed to be full of pinned pages then just give it a light
2959 * scan then give up on it.
2961 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2965 unsigned long nr_soft_reclaimed
;
2966 unsigned long nr_soft_scanned
;
2968 pg_data_t
*last_pgdat
= NULL
;
2971 * If the number of buffer_heads in the machine exceeds the maximum
2972 * allowed level, force direct reclaim to scan the highmem zone as
2973 * highmem pages could be pinning lowmem pages storing buffer_heads
2975 orig_mask
= sc
->gfp_mask
;
2976 if (buffer_heads_over_limit
) {
2977 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2978 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2981 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2982 sc
->reclaim_idx
, sc
->nodemask
) {
2984 * Take care memory controller reclaiming has small influence
2987 if (global_reclaim(sc
)) {
2988 if (!cpuset_zone_allowed(zone
,
2989 GFP_KERNEL
| __GFP_HARDWALL
))
2993 * If we already have plenty of memory free for
2994 * compaction in this zone, don't free any more.
2995 * Even though compaction is invoked for any
2996 * non-zero order, only frequent costly order
2997 * reclamation is disruptive enough to become a
2998 * noticeable problem, like transparent huge
3001 if (IS_ENABLED(CONFIG_COMPACTION
) &&
3002 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
3003 compaction_ready(zone
, sc
)) {
3004 sc
->compaction_ready
= true;
3009 * Shrink each node in the zonelist once. If the
3010 * zonelist is ordered by zone (not the default) then a
3011 * node may be shrunk multiple times but in that case
3012 * the user prefers lower zones being preserved.
3014 if (zone
->zone_pgdat
== last_pgdat
)
3018 * This steals pages from memory cgroups over softlimit
3019 * and returns the number of reclaimed pages and
3020 * scanned pages. This works for global memory pressure
3021 * and balancing, not for a memcg's limit.
3023 nr_soft_scanned
= 0;
3024 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
3025 sc
->order
, sc
->gfp_mask
,
3027 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
3028 sc
->nr_scanned
+= nr_soft_scanned
;
3029 /* need some check for avoid more shrink_zone() */
3032 /* See comment about same check for global reclaim above */
3033 if (zone
->zone_pgdat
== last_pgdat
)
3035 last_pgdat
= zone
->zone_pgdat
;
3036 shrink_node(zone
->zone_pgdat
, sc
);
3040 * Restore to original mask to avoid the impact on the caller if we
3041 * promoted it to __GFP_HIGHMEM.
3043 sc
->gfp_mask
= orig_mask
;
3046 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
3048 struct mem_cgroup
*memcg
;
3050 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
3052 unsigned long refaults
;
3053 struct lruvec
*lruvec
;
3055 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3056 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
3057 lruvec
->refaults
= refaults
;
3058 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
3062 * This is the main entry point to direct page reclaim.
3064 * If a full scan of the inactive list fails to free enough memory then we
3065 * are "out of memory" and something needs to be killed.
3067 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3068 * high - the zone may be full of dirty or under-writeback pages, which this
3069 * caller can't do much about. We kick the writeback threads and take explicit
3070 * naps in the hope that some of these pages can be written. But if the
3071 * allocating task holds filesystem locks which prevent writeout this might not
3072 * work, and the allocation attempt will fail.
3074 * returns: 0, if no pages reclaimed
3075 * else, the number of pages reclaimed
3077 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3078 struct scan_control
*sc
)
3080 int initial_priority
= sc
->priority
;
3081 pg_data_t
*last_pgdat
;
3085 delayacct_freepages_start();
3087 if (global_reclaim(sc
))
3088 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3091 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3094 shrink_zones(zonelist
, sc
);
3096 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3099 if (sc
->compaction_ready
)
3103 * If we're getting trouble reclaiming, start doing
3104 * writepage even in laptop mode.
3106 if (sc
->priority
< DEF_PRIORITY
- 2)
3107 sc
->may_writepage
= 1;
3108 } while (--sc
->priority
>= 0);
3111 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3113 if (zone
->zone_pgdat
== last_pgdat
)
3115 last_pgdat
= zone
->zone_pgdat
;
3116 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3117 set_memcg_congestion(last_pgdat
, sc
->target_mem_cgroup
, false);
3120 delayacct_freepages_end();
3122 if (sc
->nr_reclaimed
)
3123 return sc
->nr_reclaimed
;
3125 /* Aborted reclaim to try compaction? don't OOM, then */
3126 if (sc
->compaction_ready
)
3129 /* Untapped cgroup reserves? Don't OOM, retry. */
3130 if (sc
->memcg_low_skipped
) {
3131 sc
->priority
= initial_priority
;
3132 sc
->memcg_low_reclaim
= 1;
3133 sc
->memcg_low_skipped
= 0;
3140 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3143 unsigned long pfmemalloc_reserve
= 0;
3144 unsigned long free_pages
= 0;
3148 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3151 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3152 zone
= &pgdat
->node_zones
[i
];
3153 if (!managed_zone(zone
))
3156 if (!zone_reclaimable_pages(zone
))
3159 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3160 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3163 /* If there are no reserves (unexpected config) then do not throttle */
3164 if (!pfmemalloc_reserve
)
3167 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3169 /* kswapd must be awake if processes are being throttled */
3170 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3171 if (READ_ONCE(pgdat
->kswapd_classzone_idx
) > ZONE_NORMAL
)
3172 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, ZONE_NORMAL
);
3174 wake_up_interruptible(&pgdat
->kswapd_wait
);
3181 * Throttle direct reclaimers if backing storage is backed by the network
3182 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3183 * depleted. kswapd will continue to make progress and wake the processes
3184 * when the low watermark is reached.
3186 * Returns true if a fatal signal was delivered during throttling. If this
3187 * happens, the page allocator should not consider triggering the OOM killer.
3189 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3190 nodemask_t
*nodemask
)
3194 pg_data_t
*pgdat
= NULL
;
3197 * Kernel threads should not be throttled as they may be indirectly
3198 * responsible for cleaning pages necessary for reclaim to make forward
3199 * progress. kjournald for example may enter direct reclaim while
3200 * committing a transaction where throttling it could forcing other
3201 * processes to block on log_wait_commit().
3203 if (current
->flags
& PF_KTHREAD
)
3207 * If a fatal signal is pending, this process should not throttle.
3208 * It should return quickly so it can exit and free its memory
3210 if (fatal_signal_pending(current
))
3214 * Check if the pfmemalloc reserves are ok by finding the first node
3215 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3216 * GFP_KERNEL will be required for allocating network buffers when
3217 * swapping over the network so ZONE_HIGHMEM is unusable.
3219 * Throttling is based on the first usable node and throttled processes
3220 * wait on a queue until kswapd makes progress and wakes them. There
3221 * is an affinity then between processes waking up and where reclaim
3222 * progress has been made assuming the process wakes on the same node.
3223 * More importantly, processes running on remote nodes will not compete
3224 * for remote pfmemalloc reserves and processes on different nodes
3225 * should make reasonable progress.
3227 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3228 gfp_zone(gfp_mask
), nodemask
) {
3229 if (zone_idx(zone
) > ZONE_NORMAL
)
3232 /* Throttle based on the first usable node */
3233 pgdat
= zone
->zone_pgdat
;
3234 if (allow_direct_reclaim(pgdat
))
3239 /* If no zone was usable by the allocation flags then do not throttle */
3243 /* Account for the throttling */
3244 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3247 * If the caller cannot enter the filesystem, it's possible that it
3248 * is due to the caller holding an FS lock or performing a journal
3249 * transaction in the case of a filesystem like ext[3|4]. In this case,
3250 * it is not safe to block on pfmemalloc_wait as kswapd could be
3251 * blocked waiting on the same lock. Instead, throttle for up to a
3252 * second before continuing.
3254 if (!(gfp_mask
& __GFP_FS
)) {
3255 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3256 allow_direct_reclaim(pgdat
), HZ
);
3261 /* Throttle until kswapd wakes the process */
3262 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3263 allow_direct_reclaim(pgdat
));
3266 if (fatal_signal_pending(current
))
3273 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3274 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3276 unsigned long nr_reclaimed
;
3277 struct scan_control sc
= {
3278 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3279 .gfp_mask
= current_gfp_context(gfp_mask
),
3280 .reclaim_idx
= gfp_zone(gfp_mask
),
3282 .nodemask
= nodemask
,
3283 .priority
= DEF_PRIORITY
,
3284 .may_writepage
= !laptop_mode
,
3290 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3291 * Confirm they are large enough for max values.
3293 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3294 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3295 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3298 * Do not enter reclaim if fatal signal was delivered while throttled.
3299 * 1 is returned so that the page allocator does not OOM kill at this
3302 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3305 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3306 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3308 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3310 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3311 set_task_reclaim_state(current
, NULL
);
3313 return nr_reclaimed
;
3318 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3319 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3320 gfp_t gfp_mask
, bool noswap
,
3322 unsigned long *nr_scanned
)
3324 struct scan_control sc
= {
3325 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3326 .target_mem_cgroup
= memcg
,
3327 .may_writepage
= !laptop_mode
,
3329 .reclaim_idx
= MAX_NR_ZONES
- 1,
3330 .may_swap
= !noswap
,
3332 unsigned long lru_pages
;
3334 WARN_ON_ONCE(!current
->reclaim_state
);
3336 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3337 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3339 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3343 * NOTE: Although we can get the priority field, using it
3344 * here is not a good idea, since it limits the pages we can scan.
3345 * if we don't reclaim here, the shrink_node from balance_pgdat
3346 * will pick up pages from other mem cgroup's as well. We hack
3347 * the priority and make it zero.
3349 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3351 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3353 *nr_scanned
= sc
.nr_scanned
;
3355 return sc
.nr_reclaimed
;
3358 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3359 unsigned long nr_pages
,
3363 struct zonelist
*zonelist
;
3364 unsigned long nr_reclaimed
;
3365 unsigned long pflags
;
3367 unsigned int noreclaim_flag
;
3368 struct scan_control sc
= {
3369 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3370 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3371 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3372 .reclaim_idx
= MAX_NR_ZONES
- 1,
3373 .target_mem_cgroup
= memcg
,
3374 .priority
= DEF_PRIORITY
,
3375 .may_writepage
= !laptop_mode
,
3377 .may_swap
= may_swap
,
3380 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3382 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3383 * take care of from where we get pages. So the node where we start the
3384 * scan does not need to be the current node.
3386 nid
= mem_cgroup_select_victim_node(memcg
);
3388 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3390 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3392 psi_memstall_enter(&pflags
);
3393 noreclaim_flag
= memalloc_noreclaim_save();
3395 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3397 memalloc_noreclaim_restore(noreclaim_flag
);
3398 psi_memstall_leave(&pflags
);
3400 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3401 set_task_reclaim_state(current
, NULL
);
3403 return nr_reclaimed
;
3407 static void age_active_anon(struct pglist_data
*pgdat
,
3408 struct scan_control
*sc
)
3410 struct mem_cgroup
*memcg
;
3412 if (!total_swap_pages
)
3415 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3417 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3419 if (inactive_list_is_low(lruvec
, false, sc
, true))
3420 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3421 sc
, LRU_ACTIVE_ANON
);
3423 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3427 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int classzone_idx
)
3433 * Check for watermark boosts top-down as the higher zones
3434 * are more likely to be boosted. Both watermarks and boosts
3435 * should not be checked at the time time as reclaim would
3436 * start prematurely when there is no boosting and a lower
3439 for (i
= classzone_idx
; i
>= 0; i
--) {
3440 zone
= pgdat
->node_zones
+ i
;
3441 if (!managed_zone(zone
))
3444 if (zone
->watermark_boost
)
3452 * Returns true if there is an eligible zone balanced for the request order
3455 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3458 unsigned long mark
= -1;
3462 * Check watermarks bottom-up as lower zones are more likely to
3465 for (i
= 0; i
<= classzone_idx
; i
++) {
3466 zone
= pgdat
->node_zones
+ i
;
3468 if (!managed_zone(zone
))
3471 mark
= high_wmark_pages(zone
);
3472 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3477 * If a node has no populated zone within classzone_idx, it does not
3478 * need balancing by definition. This can happen if a zone-restricted
3479 * allocation tries to wake a remote kswapd.
3487 /* Clear pgdat state for congested, dirty or under writeback. */
3488 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3490 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3491 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3492 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3496 * Prepare kswapd for sleeping. This verifies that there are no processes
3497 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3499 * Returns true if kswapd is ready to sleep
3501 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3504 * The throttled processes are normally woken up in balance_pgdat() as
3505 * soon as allow_direct_reclaim() is true. But there is a potential
3506 * race between when kswapd checks the watermarks and a process gets
3507 * throttled. There is also a potential race if processes get
3508 * throttled, kswapd wakes, a large process exits thereby balancing the
3509 * zones, which causes kswapd to exit balance_pgdat() before reaching
3510 * the wake up checks. If kswapd is going to sleep, no process should
3511 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3512 * the wake up is premature, processes will wake kswapd and get
3513 * throttled again. The difference from wake ups in balance_pgdat() is
3514 * that here we are under prepare_to_wait().
3516 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3517 wake_up_all(&pgdat
->pfmemalloc_wait
);
3519 /* Hopeless node, leave it to direct reclaim */
3520 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3523 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3524 clear_pgdat_congested(pgdat
);
3532 * kswapd shrinks a node of pages that are at or below the highest usable
3533 * zone that is currently unbalanced.
3535 * Returns true if kswapd scanned at least the requested number of pages to
3536 * reclaim or if the lack of progress was due to pages under writeback.
3537 * This is used to determine if the scanning priority needs to be raised.
3539 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3540 struct scan_control
*sc
)
3545 /* Reclaim a number of pages proportional to the number of zones */
3546 sc
->nr_to_reclaim
= 0;
3547 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3548 zone
= pgdat
->node_zones
+ z
;
3549 if (!managed_zone(zone
))
3552 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3556 * Historically care was taken to put equal pressure on all zones but
3557 * now pressure is applied based on node LRU order.
3559 shrink_node(pgdat
, sc
);
3562 * Fragmentation may mean that the system cannot be rebalanced for
3563 * high-order allocations. If twice the allocation size has been
3564 * reclaimed then recheck watermarks only at order-0 to prevent
3565 * excessive reclaim. Assume that a process requested a high-order
3566 * can direct reclaim/compact.
3568 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3571 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3575 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3576 * that are eligible for use by the caller until at least one zone is
3579 * Returns the order kswapd finished reclaiming at.
3581 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3582 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3583 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3584 * or lower is eligible for reclaim until at least one usable zone is
3587 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3590 unsigned long nr_soft_reclaimed
;
3591 unsigned long nr_soft_scanned
;
3592 unsigned long pflags
;
3593 unsigned long nr_boost_reclaim
;
3594 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
3597 struct scan_control sc
= {
3598 .gfp_mask
= GFP_KERNEL
,
3603 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3604 psi_memstall_enter(&pflags
);
3605 __fs_reclaim_acquire();
3607 count_vm_event(PAGEOUTRUN
);
3610 * Account for the reclaim boost. Note that the zone boost is left in
3611 * place so that parallel allocations that are near the watermark will
3612 * stall or direct reclaim until kswapd is finished.
3614 nr_boost_reclaim
= 0;
3615 for (i
= 0; i
<= classzone_idx
; i
++) {
3616 zone
= pgdat
->node_zones
+ i
;
3617 if (!managed_zone(zone
))
3620 nr_boost_reclaim
+= zone
->watermark_boost
;
3621 zone_boosts
[i
] = zone
->watermark_boost
;
3623 boosted
= nr_boost_reclaim
;
3626 sc
.priority
= DEF_PRIORITY
;
3628 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3629 bool raise_priority
= true;
3633 sc
.reclaim_idx
= classzone_idx
;
3636 * If the number of buffer_heads exceeds the maximum allowed
3637 * then consider reclaiming from all zones. This has a dual
3638 * purpose -- on 64-bit systems it is expected that
3639 * buffer_heads are stripped during active rotation. On 32-bit
3640 * systems, highmem pages can pin lowmem memory and shrinking
3641 * buffers can relieve lowmem pressure. Reclaim may still not
3642 * go ahead if all eligible zones for the original allocation
3643 * request are balanced to avoid excessive reclaim from kswapd.
3645 if (buffer_heads_over_limit
) {
3646 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3647 zone
= pgdat
->node_zones
+ i
;
3648 if (!managed_zone(zone
))
3657 * If the pgdat is imbalanced then ignore boosting and preserve
3658 * the watermarks for a later time and restart. Note that the
3659 * zone watermarks will be still reset at the end of balancing
3660 * on the grounds that the normal reclaim should be enough to
3661 * re-evaluate if boosting is required when kswapd next wakes.
3663 balanced
= pgdat_balanced(pgdat
, sc
.order
, classzone_idx
);
3664 if (!balanced
&& nr_boost_reclaim
) {
3665 nr_boost_reclaim
= 0;
3670 * If boosting is not active then only reclaim if there are no
3671 * eligible zones. Note that sc.reclaim_idx is not used as
3672 * buffer_heads_over_limit may have adjusted it.
3674 if (!nr_boost_reclaim
&& balanced
)
3677 /* Limit the priority of boosting to avoid reclaim writeback */
3678 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
3679 raise_priority
= false;
3682 * Do not writeback or swap pages for boosted reclaim. The
3683 * intent is to relieve pressure not issue sub-optimal IO
3684 * from reclaim context. If no pages are reclaimed, the
3685 * reclaim will be aborted.
3687 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
3688 sc
.may_swap
= !nr_boost_reclaim
;
3691 * Do some background aging of the anon list, to give
3692 * pages a chance to be referenced before reclaiming. All
3693 * pages are rotated regardless of classzone as this is
3694 * about consistent aging.
3696 age_active_anon(pgdat
, &sc
);
3699 * If we're getting trouble reclaiming, start doing writepage
3700 * even in laptop mode.
3702 if (sc
.priority
< DEF_PRIORITY
- 2)
3703 sc
.may_writepage
= 1;
3705 /* Call soft limit reclaim before calling shrink_node. */
3707 nr_soft_scanned
= 0;
3708 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3709 sc
.gfp_mask
, &nr_soft_scanned
);
3710 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3713 * There should be no need to raise the scanning priority if
3714 * enough pages are already being scanned that that high
3715 * watermark would be met at 100% efficiency.
3717 if (kswapd_shrink_node(pgdat
, &sc
))
3718 raise_priority
= false;
3721 * If the low watermark is met there is no need for processes
3722 * to be throttled on pfmemalloc_wait as they should not be
3723 * able to safely make forward progress. Wake them
3725 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3726 allow_direct_reclaim(pgdat
))
3727 wake_up_all(&pgdat
->pfmemalloc_wait
);
3729 /* Check if kswapd should be suspending */
3730 __fs_reclaim_release();
3731 ret
= try_to_freeze();
3732 __fs_reclaim_acquire();
3733 if (ret
|| kthread_should_stop())
3737 * Raise priority if scanning rate is too low or there was no
3738 * progress in reclaiming pages
3740 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3741 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
3744 * If reclaim made no progress for a boost, stop reclaim as
3745 * IO cannot be queued and it could be an infinite loop in
3746 * extreme circumstances.
3748 if (nr_boost_reclaim
&& !nr_reclaimed
)
3751 if (raise_priority
|| !nr_reclaimed
)
3753 } while (sc
.priority
>= 1);
3755 if (!sc
.nr_reclaimed
)
3756 pgdat
->kswapd_failures
++;
3759 /* If reclaim was boosted, account for the reclaim done in this pass */
3761 unsigned long flags
;
3763 for (i
= 0; i
<= classzone_idx
; i
++) {
3764 if (!zone_boosts
[i
])
3767 /* Increments are under the zone lock */
3768 zone
= pgdat
->node_zones
+ i
;
3769 spin_lock_irqsave(&zone
->lock
, flags
);
3770 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
3771 spin_unlock_irqrestore(&zone
->lock
, flags
);
3775 * As there is now likely space, wakeup kcompact to defragment
3778 wakeup_kcompactd(pgdat
, pageblock_order
, classzone_idx
);
3781 snapshot_refaults(NULL
, pgdat
);
3782 __fs_reclaim_release();
3783 psi_memstall_leave(&pflags
);
3784 set_task_reclaim_state(current
, NULL
);
3787 * Return the order kswapd stopped reclaiming at as
3788 * prepare_kswapd_sleep() takes it into account. If another caller
3789 * entered the allocator slow path while kswapd was awake, order will
3790 * remain at the higher level.
3796 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3797 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3798 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3799 * after previous reclaim attempt (node is still unbalanced). In that case
3800 * return the zone index of the previous kswapd reclaim cycle.
3802 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3803 enum zone_type prev_classzone_idx
)
3805 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
3807 return curr_idx
== MAX_NR_ZONES
? prev_classzone_idx
: curr_idx
;
3810 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3811 unsigned int classzone_idx
)
3816 if (freezing(current
) || kthread_should_stop())
3819 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3822 * Try to sleep for a short interval. Note that kcompactd will only be
3823 * woken if it is possible to sleep for a short interval. This is
3824 * deliberate on the assumption that if reclaim cannot keep an
3825 * eligible zone balanced that it's also unlikely that compaction will
3828 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3830 * Compaction records what page blocks it recently failed to
3831 * isolate pages from and skips them in the future scanning.
3832 * When kswapd is going to sleep, it is reasonable to assume
3833 * that pages and compaction may succeed so reset the cache.
3835 reset_isolation_suitable(pgdat
);
3838 * We have freed the memory, now we should compact it to make
3839 * allocation of the requested order possible.
3841 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3843 remaining
= schedule_timeout(HZ
/10);
3846 * If woken prematurely then reset kswapd_classzone_idx and
3847 * order. The values will either be from a wakeup request or
3848 * the previous request that slept prematurely.
3851 WRITE_ONCE(pgdat
->kswapd_classzone_idx
,
3852 kswapd_classzone_idx(pgdat
, classzone_idx
));
3854 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
3855 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
3858 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3859 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3863 * After a short sleep, check if it was a premature sleep. If not, then
3864 * go fully to sleep until explicitly woken up.
3867 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3868 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3871 * vmstat counters are not perfectly accurate and the estimated
3872 * value for counters such as NR_FREE_PAGES can deviate from the
3873 * true value by nr_online_cpus * threshold. To avoid the zone
3874 * watermarks being breached while under pressure, we reduce the
3875 * per-cpu vmstat threshold while kswapd is awake and restore
3876 * them before going back to sleep.
3878 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3880 if (!kthread_should_stop())
3883 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3886 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3888 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3890 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3894 * The background pageout daemon, started as a kernel thread
3895 * from the init process.
3897 * This basically trickles out pages so that we have _some_
3898 * free memory available even if there is no other activity
3899 * that frees anything up. This is needed for things like routing
3900 * etc, where we otherwise might have all activity going on in
3901 * asynchronous contexts that cannot page things out.
3903 * If there are applications that are active memory-allocators
3904 * (most normal use), this basically shouldn't matter.
3906 static int kswapd(void *p
)
3908 unsigned int alloc_order
, reclaim_order
;
3909 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3910 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3911 struct task_struct
*tsk
= current
;
3912 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3914 if (!cpumask_empty(cpumask
))
3915 set_cpus_allowed_ptr(tsk
, cpumask
);
3918 * Tell the memory management that we're a "memory allocator",
3919 * and that if we need more memory we should get access to it
3920 * regardless (see "__alloc_pages()"). "kswapd" should
3921 * never get caught in the normal page freeing logic.
3923 * (Kswapd normally doesn't need memory anyway, but sometimes
3924 * you need a small amount of memory in order to be able to
3925 * page out something else, and this flag essentially protects
3926 * us from recursively trying to free more memory as we're
3927 * trying to free the first piece of memory in the first place).
3929 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3932 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3933 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3937 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3938 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3941 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3944 /* Read the new order and classzone_idx */
3945 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3946 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3947 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3948 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3950 ret
= try_to_freeze();
3951 if (kthread_should_stop())
3955 * We can speed up thawing tasks if we don't call balance_pgdat
3956 * after returning from the refrigerator
3962 * Reclaim begins at the requested order but if a high-order
3963 * reclaim fails then kswapd falls back to reclaiming for
3964 * order-0. If that happens, kswapd will consider sleeping
3965 * for the order it finished reclaiming at (reclaim_order)
3966 * but kcompactd is woken to compact for the original
3967 * request (alloc_order).
3969 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3971 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3972 if (reclaim_order
< alloc_order
)
3973 goto kswapd_try_sleep
;
3976 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3982 * A zone is low on free memory or too fragmented for high-order memory. If
3983 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3984 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3985 * has failed or is not needed, still wake up kcompactd if only compaction is
3988 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3989 enum zone_type classzone_idx
)
3992 enum zone_type curr_idx
;
3994 if (!managed_zone(zone
))
3997 if (!cpuset_zone_allowed(zone
, gfp_flags
))
4000 pgdat
= zone
->zone_pgdat
;
4001 curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
4003 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< classzone_idx
)
4004 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, classzone_idx
);
4006 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
4007 WRITE_ONCE(pgdat
->kswapd_order
, order
);
4009 if (!waitqueue_active(&pgdat
->kswapd_wait
))
4012 /* Hopeless node, leave it to direct reclaim if possible */
4013 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
4014 (pgdat_balanced(pgdat
, order
, classzone_idx
) &&
4015 !pgdat_watermark_boosted(pgdat
, classzone_idx
))) {
4017 * There may be plenty of free memory available, but it's too
4018 * fragmented for high-order allocations. Wake up kcompactd
4019 * and rely on compaction_suitable() to determine if it's
4020 * needed. If it fails, it will defer subsequent attempts to
4021 * ratelimit its work.
4023 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
4024 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
4028 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
4030 wake_up_interruptible(&pgdat
->kswapd_wait
);
4033 #ifdef CONFIG_HIBERNATION
4035 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4038 * Rather than trying to age LRUs the aim is to preserve the overall
4039 * LRU order by reclaiming preferentially
4040 * inactive > active > active referenced > active mapped
4042 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4044 struct scan_control sc
= {
4045 .nr_to_reclaim
= nr_to_reclaim
,
4046 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4047 .reclaim_idx
= MAX_NR_ZONES
- 1,
4048 .priority
= DEF_PRIORITY
,
4052 .hibernation_mode
= 1,
4054 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4055 unsigned long nr_reclaimed
;
4056 unsigned int noreclaim_flag
;
4058 fs_reclaim_acquire(sc
.gfp_mask
);
4059 noreclaim_flag
= memalloc_noreclaim_save();
4060 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4062 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4064 set_task_reclaim_state(current
, NULL
);
4065 memalloc_noreclaim_restore(noreclaim_flag
);
4066 fs_reclaim_release(sc
.gfp_mask
);
4068 return nr_reclaimed
;
4070 #endif /* CONFIG_HIBERNATION */
4072 /* It's optimal to keep kswapds on the same CPUs as their memory, but
4073 not required for correctness. So if the last cpu in a node goes
4074 away, we get changed to run anywhere: as the first one comes back,
4075 restore their cpu bindings. */
4076 static int kswapd_cpu_online(unsigned int cpu
)
4080 for_each_node_state(nid
, N_MEMORY
) {
4081 pg_data_t
*pgdat
= NODE_DATA(nid
);
4082 const struct cpumask
*mask
;
4084 mask
= cpumask_of_node(pgdat
->node_id
);
4086 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
4087 /* One of our CPUs online: restore mask */
4088 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
4094 * This kswapd start function will be called by init and node-hot-add.
4095 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4097 int kswapd_run(int nid
)
4099 pg_data_t
*pgdat
= NODE_DATA(nid
);
4105 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4106 if (IS_ERR(pgdat
->kswapd
)) {
4107 /* failure at boot is fatal */
4108 BUG_ON(system_state
< SYSTEM_RUNNING
);
4109 pr_err("Failed to start kswapd on node %d\n", nid
);
4110 ret
= PTR_ERR(pgdat
->kswapd
);
4111 pgdat
->kswapd
= NULL
;
4117 * Called by memory hotplug when all memory in a node is offlined. Caller must
4118 * hold mem_hotplug_begin/end().
4120 void kswapd_stop(int nid
)
4122 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4125 kthread_stop(kswapd
);
4126 NODE_DATA(nid
)->kswapd
= NULL
;
4130 static int __init
kswapd_init(void)
4135 for_each_node_state(nid
, N_MEMORY
)
4137 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
4138 "mm/vmscan:online", kswapd_cpu_online
,
4144 module_init(kswapd_init
)
4150 * If non-zero call node_reclaim when the number of free pages falls below
4153 int node_reclaim_mode __read_mostly
;
4155 #define RECLAIM_OFF 0
4156 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4157 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4158 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4161 * Priority for NODE_RECLAIM. This determines the fraction of pages
4162 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4165 #define NODE_RECLAIM_PRIORITY 4
4168 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4171 int sysctl_min_unmapped_ratio
= 1;
4174 * If the number of slab pages in a zone grows beyond this percentage then
4175 * slab reclaim needs to occur.
4177 int sysctl_min_slab_ratio
= 5;
4179 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4181 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4182 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4183 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4186 * It's possible for there to be more file mapped pages than
4187 * accounted for by the pages on the file LRU lists because
4188 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4190 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4193 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4194 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4196 unsigned long nr_pagecache_reclaimable
;
4197 unsigned long delta
= 0;
4200 * If RECLAIM_UNMAP is set, then all file pages are considered
4201 * potentially reclaimable. Otherwise, we have to worry about
4202 * pages like swapcache and node_unmapped_file_pages() provides
4205 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4206 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4208 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4210 /* If we can't clean pages, remove dirty pages from consideration */
4211 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4212 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4214 /* Watch for any possible underflows due to delta */
4215 if (unlikely(delta
> nr_pagecache_reclaimable
))
4216 delta
= nr_pagecache_reclaimable
;
4218 return nr_pagecache_reclaimable
- delta
;
4222 * Try to free up some pages from this node through reclaim.
4224 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4226 /* Minimum pages needed in order to stay on node */
4227 const unsigned long nr_pages
= 1 << order
;
4228 struct task_struct
*p
= current
;
4229 unsigned int noreclaim_flag
;
4230 struct scan_control sc
= {
4231 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4232 .gfp_mask
= current_gfp_context(gfp_mask
),
4234 .priority
= NODE_RECLAIM_PRIORITY
,
4235 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4236 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4238 .reclaim_idx
= gfp_zone(gfp_mask
),
4241 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4245 fs_reclaim_acquire(sc
.gfp_mask
);
4247 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4248 * and we also need to be able to write out pages for RECLAIM_WRITE
4249 * and RECLAIM_UNMAP.
4251 noreclaim_flag
= memalloc_noreclaim_save();
4252 p
->flags
|= PF_SWAPWRITE
;
4253 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4255 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4257 * Free memory by calling shrink node with increasing
4258 * priorities until we have enough memory freed.
4261 shrink_node(pgdat
, &sc
);
4262 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4265 set_task_reclaim_state(p
, NULL
);
4266 current
->flags
&= ~PF_SWAPWRITE
;
4267 memalloc_noreclaim_restore(noreclaim_flag
);
4268 fs_reclaim_release(sc
.gfp_mask
);
4270 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4272 return sc
.nr_reclaimed
>= nr_pages
;
4275 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4280 * Node reclaim reclaims unmapped file backed pages and
4281 * slab pages if we are over the defined limits.
4283 * A small portion of unmapped file backed pages is needed for
4284 * file I/O otherwise pages read by file I/O will be immediately
4285 * thrown out if the node is overallocated. So we do not reclaim
4286 * if less than a specified percentage of the node is used by
4287 * unmapped file backed pages.
4289 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4290 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4291 return NODE_RECLAIM_FULL
;
4294 * Do not scan if the allocation should not be delayed.
4296 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4297 return NODE_RECLAIM_NOSCAN
;
4300 * Only run node reclaim on the local node or on nodes that do not
4301 * have associated processors. This will favor the local processor
4302 * over remote processors and spread off node memory allocations
4303 * as wide as possible.
4305 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4306 return NODE_RECLAIM_NOSCAN
;
4308 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4309 return NODE_RECLAIM_NOSCAN
;
4311 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4312 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4315 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4322 * page_evictable - test whether a page is evictable
4323 * @page: the page to test
4325 * Test whether page is evictable--i.e., should be placed on active/inactive
4326 * lists vs unevictable list.
4328 * Reasons page might not be evictable:
4329 * (1) page's mapping marked unevictable
4330 * (2) page is part of an mlocked VMA
4333 int page_evictable(struct page
*page
)
4337 /* Prevent address_space of inode and swap cache from being freed */
4339 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4345 * check_move_unevictable_pages - check pages for evictability and move to
4346 * appropriate zone lru list
4347 * @pvec: pagevec with lru pages to check
4349 * Checks pages for evictability, if an evictable page is in the unevictable
4350 * lru list, moves it to the appropriate evictable lru list. This function
4351 * should be only used for lru pages.
4353 void check_move_unevictable_pages(struct pagevec
*pvec
)
4355 struct lruvec
*lruvec
;
4356 struct pglist_data
*pgdat
= NULL
;
4361 for (i
= 0; i
< pvec
->nr
; i
++) {
4362 struct page
*page
= pvec
->pages
[i
];
4363 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4366 if (pagepgdat
!= pgdat
) {
4368 spin_unlock_irq(&pgdat
->lru_lock
);
4370 spin_lock_irq(&pgdat
->lru_lock
);
4372 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4374 if (!PageLRU(page
) || !PageUnevictable(page
))
4377 if (page_evictable(page
)) {
4378 enum lru_list lru
= page_lru_base_type(page
);
4380 VM_BUG_ON_PAGE(PageActive(page
), page
);
4381 ClearPageUnevictable(page
);
4382 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4383 add_page_to_lru_list(page
, lruvec
, lru
);
4389 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4390 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4391 spin_unlock_irq(&pgdat
->lru_lock
);
4394 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);