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 LIST_HEAD(shrinker_list
);
175 static DECLARE_RWSEM(shrinker_rwsem
);
177 #ifdef CONFIG_MEMCG_KMEM
180 * We allow subsystems to populate their shrinker-related
181 * LRU lists before register_shrinker_prepared() is called
182 * for the shrinker, since we don't want to impose
183 * restrictions on their internal registration order.
184 * In this case shrink_slab_memcg() may find corresponding
185 * bit is set in the shrinkers map.
187 * This value is used by the function to detect registering
188 * shrinkers and to skip do_shrink_slab() calls for them.
190 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
192 static DEFINE_IDR(shrinker_idr
);
193 static int shrinker_nr_max
;
195 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
197 int id
, ret
= -ENOMEM
;
199 down_write(&shrinker_rwsem
);
200 /* This may call shrinker, so it must use down_read_trylock() */
201 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
205 if (id
>= shrinker_nr_max
) {
206 if (memcg_expand_shrinker_maps(id
)) {
207 idr_remove(&shrinker_idr
, id
);
211 shrinker_nr_max
= id
+ 1;
216 up_write(&shrinker_rwsem
);
220 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
222 int id
= shrinker
->id
;
226 down_write(&shrinker_rwsem
);
227 idr_remove(&shrinker_idr
, id
);
228 up_write(&shrinker_rwsem
);
230 #else /* CONFIG_MEMCG_KMEM */
231 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
236 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
239 #endif /* CONFIG_MEMCG_KMEM */
241 static void set_task_reclaim_state(struct task_struct
*task
,
242 struct reclaim_state
*rs
)
244 /* Check for an overwrite */
245 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
247 /* Check for the nulling of an already-nulled member */
248 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
250 task
->reclaim_state
= rs
;
254 static bool global_reclaim(struct scan_control
*sc
)
256 return !sc
->target_mem_cgroup
;
260 * sane_reclaim - is the usual dirty throttling mechanism operational?
261 * @sc: scan_control in question
263 * The normal page dirty throttling mechanism in balance_dirty_pages() is
264 * completely broken with the legacy memcg and direct stalling in
265 * shrink_page_list() is used for throttling instead, which lacks all the
266 * niceties such as fairness, adaptive pausing, bandwidth proportional
267 * allocation and configurability.
269 * This function tests whether the vmscan currently in progress can assume
270 * that the normal dirty throttling mechanism is operational.
272 static bool sane_reclaim(struct scan_control
*sc
)
274 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
278 #ifdef CONFIG_CGROUP_WRITEBACK
279 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
285 static void set_memcg_congestion(pg_data_t
*pgdat
,
286 struct mem_cgroup
*memcg
,
289 struct mem_cgroup_per_node
*mn
;
294 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
295 WRITE_ONCE(mn
->congested
, congested
);
298 static bool memcg_congested(pg_data_t
*pgdat
,
299 struct mem_cgroup
*memcg
)
301 struct mem_cgroup_per_node
*mn
;
303 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
304 return READ_ONCE(mn
->congested
);
308 static bool global_reclaim(struct scan_control
*sc
)
313 static bool sane_reclaim(struct scan_control
*sc
)
318 static inline void set_memcg_congestion(struct pglist_data
*pgdat
,
319 struct mem_cgroup
*memcg
, bool congested
)
323 static inline bool memcg_congested(struct pglist_data
*pgdat
,
324 struct mem_cgroup
*memcg
)
332 * This misses isolated pages which are not accounted for to save counters.
333 * As the data only determines if reclaim or compaction continues, it is
334 * not expected that isolated pages will be a dominating factor.
336 unsigned long zone_reclaimable_pages(struct zone
*zone
)
340 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
341 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
342 if (get_nr_swap_pages() > 0)
343 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
344 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
350 * lruvec_lru_size - Returns the number of pages on the given LRU list.
351 * @lruvec: lru vector
353 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
355 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
357 unsigned long lru_size
;
360 if (!mem_cgroup_disabled())
361 lru_size
= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
363 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
365 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
366 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
369 if (!managed_zone(zone
))
372 if (!mem_cgroup_disabled())
373 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
375 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
376 NR_ZONE_LRU_BASE
+ lru
);
377 lru_size
-= min(size
, lru_size
);
385 * Add a shrinker callback to be called from the vm.
387 int prealloc_shrinker(struct shrinker
*shrinker
)
389 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
391 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
394 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
395 if (!shrinker
->nr_deferred
)
398 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
399 if (prealloc_memcg_shrinker(shrinker
))
406 kfree(shrinker
->nr_deferred
);
407 shrinker
->nr_deferred
= NULL
;
411 void free_prealloced_shrinker(struct shrinker
*shrinker
)
413 if (!shrinker
->nr_deferred
)
416 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
417 unregister_memcg_shrinker(shrinker
);
419 kfree(shrinker
->nr_deferred
);
420 shrinker
->nr_deferred
= NULL
;
423 void register_shrinker_prepared(struct shrinker
*shrinker
)
425 down_write(&shrinker_rwsem
);
426 list_add_tail(&shrinker
->list
, &shrinker_list
);
427 #ifdef CONFIG_MEMCG_KMEM
428 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
429 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
431 up_write(&shrinker_rwsem
);
434 int register_shrinker(struct shrinker
*shrinker
)
436 int err
= prealloc_shrinker(shrinker
);
440 register_shrinker_prepared(shrinker
);
443 EXPORT_SYMBOL(register_shrinker
);
448 void unregister_shrinker(struct shrinker
*shrinker
)
450 if (!shrinker
->nr_deferred
)
452 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
453 unregister_memcg_shrinker(shrinker
);
454 down_write(&shrinker_rwsem
);
455 list_del(&shrinker
->list
);
456 up_write(&shrinker_rwsem
);
457 kfree(shrinker
->nr_deferred
);
458 shrinker
->nr_deferred
= NULL
;
460 EXPORT_SYMBOL(unregister_shrinker
);
462 #define SHRINK_BATCH 128
464 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
465 struct shrinker
*shrinker
, int priority
)
467 unsigned long freed
= 0;
468 unsigned long long delta
;
473 int nid
= shrinkctl
->nid
;
474 long batch_size
= shrinker
->batch
? shrinker
->batch
476 long scanned
= 0, next_deferred
;
478 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
481 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
482 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
486 * copy the current shrinker scan count into a local variable
487 * and zero it so that other concurrent shrinker invocations
488 * don't also do this scanning work.
490 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
493 if (shrinker
->seeks
) {
494 delta
= freeable
>> priority
;
496 do_div(delta
, shrinker
->seeks
);
499 * These objects don't require any IO to create. Trim
500 * them aggressively under memory pressure to keep
501 * them from causing refetches in the IO caches.
503 delta
= freeable
/ 2;
507 if (total_scan
< 0) {
508 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
509 shrinker
->scan_objects
, total_scan
);
510 total_scan
= freeable
;
513 next_deferred
= total_scan
;
516 * We need to avoid excessive windup on filesystem shrinkers
517 * due to large numbers of GFP_NOFS allocations causing the
518 * shrinkers to return -1 all the time. This results in a large
519 * nr being built up so when a shrink that can do some work
520 * comes along it empties the entire cache due to nr >>>
521 * freeable. This is bad for sustaining a working set in
524 * Hence only allow the shrinker to scan the entire cache when
525 * a large delta change is calculated directly.
527 if (delta
< freeable
/ 4)
528 total_scan
= min(total_scan
, freeable
/ 2);
531 * Avoid risking looping forever due to too large nr value:
532 * never try to free more than twice the estimate number of
535 if (total_scan
> freeable
* 2)
536 total_scan
= freeable
* 2;
538 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
539 freeable
, delta
, total_scan
, priority
);
542 * Normally, we should not scan less than batch_size objects in one
543 * pass to avoid too frequent shrinker calls, but if the slab has less
544 * than batch_size objects in total and we are really tight on memory,
545 * we will try to reclaim all available objects, otherwise we can end
546 * up failing allocations although there are plenty of reclaimable
547 * objects spread over several slabs with usage less than the
550 * We detect the "tight on memory" situations by looking at the total
551 * number of objects we want to scan (total_scan). If it is greater
552 * than the total number of objects on slab (freeable), we must be
553 * scanning at high prio and therefore should try to reclaim as much as
556 while (total_scan
>= batch_size
||
557 total_scan
>= freeable
) {
559 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
561 shrinkctl
->nr_to_scan
= nr_to_scan
;
562 shrinkctl
->nr_scanned
= nr_to_scan
;
563 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
564 if (ret
== SHRINK_STOP
)
568 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
569 total_scan
-= shrinkctl
->nr_scanned
;
570 scanned
+= shrinkctl
->nr_scanned
;
575 if (next_deferred
>= scanned
)
576 next_deferred
-= scanned
;
580 * move the unused scan count back into the shrinker in a
581 * manner that handles concurrent updates. If we exhausted the
582 * scan, there is no need to do an update.
584 if (next_deferred
> 0)
585 new_nr
= atomic_long_add_return(next_deferred
,
586 &shrinker
->nr_deferred
[nid
]);
588 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
590 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
594 #ifdef CONFIG_MEMCG_KMEM
595 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
596 struct mem_cgroup
*memcg
, int priority
)
598 struct memcg_shrinker_map
*map
;
599 unsigned long ret
, freed
= 0;
602 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
))
605 if (!down_read_trylock(&shrinker_rwsem
))
608 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
613 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
614 struct shrink_control sc
= {
615 .gfp_mask
= gfp_mask
,
619 struct shrinker
*shrinker
;
621 shrinker
= idr_find(&shrinker_idr
, i
);
622 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
624 clear_bit(i
, map
->map
);
628 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
629 if (ret
== SHRINK_EMPTY
) {
630 clear_bit(i
, map
->map
);
632 * After the shrinker reported that it had no objects to
633 * free, but before we cleared the corresponding bit in
634 * the memcg shrinker map, a new object might have been
635 * added. To make sure, we have the bit set in this
636 * case, we invoke the shrinker one more time and reset
637 * the bit if it reports that it is not empty anymore.
638 * The memory barrier here pairs with the barrier in
639 * memcg_set_shrinker_bit():
641 * list_lru_add() shrink_slab_memcg()
642 * list_add_tail() clear_bit()
644 * set_bit() do_shrink_slab()
646 smp_mb__after_atomic();
647 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
648 if (ret
== SHRINK_EMPTY
)
651 memcg_set_shrinker_bit(memcg
, nid
, i
);
655 if (rwsem_is_contended(&shrinker_rwsem
)) {
661 up_read(&shrinker_rwsem
);
664 #else /* CONFIG_MEMCG_KMEM */
665 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
666 struct mem_cgroup
*memcg
, int priority
)
670 #endif /* CONFIG_MEMCG_KMEM */
673 * shrink_slab - shrink slab caches
674 * @gfp_mask: allocation context
675 * @nid: node whose slab caches to target
676 * @memcg: memory cgroup whose slab caches to target
677 * @priority: the reclaim priority
679 * Call the shrink functions to age shrinkable caches.
681 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
682 * unaware shrinkers will receive a node id of 0 instead.
684 * @memcg specifies the memory cgroup to target. Unaware shrinkers
685 * are called only if it is the root cgroup.
687 * @priority is sc->priority, we take the number of objects and >> by priority
688 * in order to get the scan target.
690 * Returns the number of reclaimed slab objects.
692 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
693 struct mem_cgroup
*memcg
,
696 unsigned long ret
, freed
= 0;
697 struct shrinker
*shrinker
;
700 * The root memcg might be allocated even though memcg is disabled
701 * via "cgroup_disable=memory" boot parameter. This could make
702 * mem_cgroup_is_root() return false, then just run memcg slab
703 * shrink, but skip global shrink. This may result in premature
706 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
707 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
709 if (!down_read_trylock(&shrinker_rwsem
))
712 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
713 struct shrink_control sc
= {
714 .gfp_mask
= gfp_mask
,
719 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
720 if (ret
== SHRINK_EMPTY
)
724 * Bail out if someone want to register a new shrinker to
725 * prevent the regsitration from being stalled for long periods
726 * by parallel ongoing shrinking.
728 if (rwsem_is_contended(&shrinker_rwsem
)) {
734 up_read(&shrinker_rwsem
);
740 void drop_slab_node(int nid
)
745 struct mem_cgroup
*memcg
= NULL
;
748 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
750 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
751 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
752 } while (freed
> 10);
759 for_each_online_node(nid
)
763 static inline int is_page_cache_freeable(struct page
*page
)
766 * A freeable page cache page is referenced only by the caller
767 * that isolated the page, the page cache and optional buffer
768 * heads at page->private.
770 int page_cache_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
772 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
775 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
777 if (current
->flags
& PF_SWAPWRITE
)
779 if (!inode_write_congested(inode
))
781 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
787 * We detected a synchronous write error writing a page out. Probably
788 * -ENOSPC. We need to propagate that into the address_space for a subsequent
789 * fsync(), msync() or close().
791 * The tricky part is that after writepage we cannot touch the mapping: nothing
792 * prevents it from being freed up. But we have a ref on the page and once
793 * that page is locked, the mapping is pinned.
795 * We're allowed to run sleeping lock_page() here because we know the caller has
798 static void handle_write_error(struct address_space
*mapping
,
799 struct page
*page
, int error
)
802 if (page_mapping(page
) == mapping
)
803 mapping_set_error(mapping
, error
);
807 /* possible outcome of pageout() */
809 /* failed to write page out, page is locked */
811 /* move page to the active list, page is locked */
813 /* page has been sent to the disk successfully, page is unlocked */
815 /* page is clean and locked */
820 * pageout is called by shrink_page_list() for each dirty page.
821 * Calls ->writepage().
823 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
824 struct scan_control
*sc
)
827 * If the page is dirty, only perform writeback if that write
828 * will be non-blocking. To prevent this allocation from being
829 * stalled by pagecache activity. But note that there may be
830 * stalls if we need to run get_block(). We could test
831 * PagePrivate for that.
833 * If this process is currently in __generic_file_write_iter() against
834 * this page's queue, we can perform writeback even if that
837 * If the page is swapcache, write it back even if that would
838 * block, for some throttling. This happens by accident, because
839 * swap_backing_dev_info is bust: it doesn't reflect the
840 * congestion state of the swapdevs. Easy to fix, if needed.
842 if (!is_page_cache_freeable(page
))
846 * Some data journaling orphaned pages can have
847 * page->mapping == NULL while being dirty with clean buffers.
849 if (page_has_private(page
)) {
850 if (try_to_free_buffers(page
)) {
851 ClearPageDirty(page
);
852 pr_info("%s: orphaned page\n", __func__
);
858 if (mapping
->a_ops
->writepage
== NULL
)
859 return PAGE_ACTIVATE
;
860 if (!may_write_to_inode(mapping
->host
, sc
))
863 if (clear_page_dirty_for_io(page
)) {
865 struct writeback_control wbc
= {
866 .sync_mode
= WB_SYNC_NONE
,
867 .nr_to_write
= SWAP_CLUSTER_MAX
,
869 .range_end
= LLONG_MAX
,
873 SetPageReclaim(page
);
874 res
= mapping
->a_ops
->writepage(page
, &wbc
);
876 handle_write_error(mapping
, page
, res
);
877 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
878 ClearPageReclaim(page
);
879 return PAGE_ACTIVATE
;
882 if (!PageWriteback(page
)) {
883 /* synchronous write or broken a_ops? */
884 ClearPageReclaim(page
);
886 trace_mm_vmscan_writepage(page
);
887 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
895 * Same as remove_mapping, but if the page is removed from the mapping, it
896 * gets returned with a refcount of 0.
898 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
904 BUG_ON(!PageLocked(page
));
905 BUG_ON(mapping
!= page_mapping(page
));
907 xa_lock_irqsave(&mapping
->i_pages
, flags
);
909 * The non racy check for a busy page.
911 * Must be careful with the order of the tests. When someone has
912 * a ref to the page, it may be possible that they dirty it then
913 * drop the reference. So if PageDirty is tested before page_count
914 * here, then the following race may occur:
916 * get_user_pages(&page);
917 * [user mapping goes away]
919 * !PageDirty(page) [good]
920 * SetPageDirty(page);
922 * !page_count(page) [good, discard it]
924 * [oops, our write_to data is lost]
926 * Reversing the order of the tests ensures such a situation cannot
927 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
928 * load is not satisfied before that of page->_refcount.
930 * Note that if SetPageDirty is always performed via set_page_dirty,
931 * and thus under the i_pages lock, then this ordering is not required.
933 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
934 refcount
= 1 + HPAGE_PMD_NR
;
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
,
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_CLEAN
;
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
= 1 << compound_order(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
);
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 mem_cgroup_uncharge(page
);
1492 (*get_compound_page_dtor(page
))(page
);
1494 list_add(&page
->lru
, &free_pages
);
1497 activate_locked_split
:
1499 * The tail pages that are failed to add into swap cache
1500 * reach here. Fixup nr_scanned and nr_pages.
1503 sc
->nr_scanned
-= (nr_pages
- 1);
1507 /* Not a candidate for swapping, so reclaim swap space. */
1508 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1510 try_to_free_swap(page
);
1511 VM_BUG_ON_PAGE(PageActive(page
), page
);
1512 if (!PageMlocked(page
)) {
1513 int type
= page_is_file_cache(page
);
1514 SetPageActive(page
);
1515 stat
->nr_activate
[type
] += nr_pages
;
1516 count_memcg_page_event(page
, PGACTIVATE
);
1521 list_add(&page
->lru
, &ret_pages
);
1522 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1525 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1527 mem_cgroup_uncharge_list(&free_pages
);
1528 try_to_unmap_flush();
1529 free_unref_page_list(&free_pages
);
1531 list_splice(&ret_pages
, page_list
);
1532 count_vm_events(PGACTIVATE
, pgactivate
);
1534 return nr_reclaimed
;
1537 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1538 struct list_head
*page_list
)
1540 struct scan_control sc
= {
1541 .gfp_mask
= GFP_KERNEL
,
1542 .priority
= DEF_PRIORITY
,
1545 struct reclaim_stat dummy_stat
;
1547 struct page
*page
, *next
;
1548 LIST_HEAD(clean_pages
);
1550 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1551 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1552 !__PageMovable(page
) && !PageUnevictable(page
)) {
1553 ClearPageActive(page
);
1554 list_move(&page
->lru
, &clean_pages
);
1558 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1559 TTU_IGNORE_ACCESS
, &dummy_stat
, true);
1560 list_splice(&clean_pages
, page_list
);
1561 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1566 * Attempt to remove the specified page from its LRU. Only take this page
1567 * if it is of the appropriate PageActive status. Pages which are being
1568 * freed elsewhere are also ignored.
1570 * page: page to consider
1571 * mode: one of the LRU isolation modes defined above
1573 * returns 0 on success, -ve errno on failure.
1575 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1579 /* Only take pages on the LRU. */
1583 /* Compaction should not handle unevictable pages but CMA can do so */
1584 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1590 * To minimise LRU disruption, the caller can indicate that it only
1591 * wants to isolate pages it will be able to operate on without
1592 * blocking - clean pages for the most part.
1594 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1595 * that it is possible to migrate without blocking
1597 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1598 /* All the caller can do on PageWriteback is block */
1599 if (PageWriteback(page
))
1602 if (PageDirty(page
)) {
1603 struct address_space
*mapping
;
1607 * Only pages without mappings or that have a
1608 * ->migratepage callback are possible to migrate
1609 * without blocking. However, we can be racing with
1610 * truncation so it's necessary to lock the page
1611 * to stabilise the mapping as truncation holds
1612 * the page lock until after the page is removed
1613 * from the page cache.
1615 if (!trylock_page(page
))
1618 mapping
= page_mapping(page
);
1619 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1626 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1629 if (likely(get_page_unless_zero(page
))) {
1631 * Be careful not to clear PageLRU until after we're
1632 * sure the page is not being freed elsewhere -- the
1633 * page release code relies on it.
1644 * Update LRU sizes after isolating pages. The LRU size updates must
1645 * be complete before mem_cgroup_update_lru_size due to a santity check.
1647 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1648 enum lru_list lru
, unsigned long *nr_zone_taken
)
1652 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1653 if (!nr_zone_taken
[zid
])
1656 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1658 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1665 * pgdat->lru_lock is heavily contended. Some of the functions that
1666 * shrink the lists perform better by taking out a batch of pages
1667 * and working on them outside the LRU lock.
1669 * For pagecache intensive workloads, this function is the hottest
1670 * spot in the kernel (apart from copy_*_user functions).
1672 * Appropriate locks must be held before calling this function.
1674 * @nr_to_scan: The number of eligible pages to look through on the list.
1675 * @lruvec: The LRU vector to pull pages from.
1676 * @dst: The temp list to put pages on to.
1677 * @nr_scanned: The number of pages that were scanned.
1678 * @sc: The scan_control struct for this reclaim session
1679 * @mode: One of the LRU isolation modes
1680 * @lru: LRU list id for isolating
1682 * returns how many pages were moved onto *@dst.
1684 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1685 struct lruvec
*lruvec
, struct list_head
*dst
,
1686 unsigned long *nr_scanned
, struct scan_control
*sc
,
1689 struct list_head
*src
= &lruvec
->lists
[lru
];
1690 unsigned long nr_taken
= 0;
1691 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1692 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1693 unsigned long skipped
= 0;
1694 unsigned long scan
, total_scan
, nr_pages
;
1695 LIST_HEAD(pages_skipped
);
1696 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1700 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1703 page
= lru_to_page(src
);
1704 prefetchw_prev_lru_page(page
, src
, flags
);
1706 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1708 nr_pages
= 1 << compound_order(page
);
1709 total_scan
+= nr_pages
;
1711 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1712 list_move(&page
->lru
, &pages_skipped
);
1713 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1718 * Do not count skipped pages because that makes the function
1719 * return with no isolated pages if the LRU mostly contains
1720 * ineligible pages. This causes the VM to not reclaim any
1721 * pages, triggering a premature OOM.
1723 * Account all tail pages of THP. This would not cause
1724 * premature OOM since __isolate_lru_page() returns -EBUSY
1725 * only when the page is being freed somewhere else.
1728 switch (__isolate_lru_page(page
, mode
)) {
1730 nr_taken
+= nr_pages
;
1731 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1732 list_move(&page
->lru
, dst
);
1736 /* else it is being freed elsewhere */
1737 list_move(&page
->lru
, src
);
1746 * Splice any skipped pages to the start of the LRU list. Note that
1747 * this disrupts the LRU order when reclaiming for lower zones but
1748 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1749 * scanning would soon rescan the same pages to skip and put the
1750 * system at risk of premature OOM.
1752 if (!list_empty(&pages_skipped
)) {
1755 list_splice(&pages_skipped
, src
);
1756 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1757 if (!nr_skipped
[zid
])
1760 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1761 skipped
+= nr_skipped
[zid
];
1764 *nr_scanned
= total_scan
;
1765 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1766 total_scan
, skipped
, nr_taken
, mode
, lru
);
1767 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1772 * isolate_lru_page - tries to isolate a page from its LRU list
1773 * @page: page to isolate from its LRU list
1775 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1776 * vmstat statistic corresponding to whatever LRU list the page was on.
1778 * Returns 0 if the page was removed from an LRU list.
1779 * Returns -EBUSY if the page was not on an LRU list.
1781 * The returned page will have PageLRU() cleared. If it was found on
1782 * the active list, it will have PageActive set. If it was found on
1783 * the unevictable list, it will have the PageUnevictable bit set. That flag
1784 * may need to be cleared by the caller before letting the page go.
1786 * The vmstat statistic corresponding to the list on which the page was
1787 * found will be decremented.
1791 * (1) Must be called with an elevated refcount on the page. This is a
1792 * fundamentnal difference from isolate_lru_pages (which is called
1793 * without a stable reference).
1794 * (2) the lru_lock must not be held.
1795 * (3) interrupts must be enabled.
1797 int isolate_lru_page(struct page
*page
)
1801 VM_BUG_ON_PAGE(!page_count(page
), page
);
1802 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1804 if (PageLRU(page
)) {
1805 pg_data_t
*pgdat
= page_pgdat(page
);
1806 struct lruvec
*lruvec
;
1808 spin_lock_irq(&pgdat
->lru_lock
);
1809 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1810 if (PageLRU(page
)) {
1811 int lru
= page_lru(page
);
1814 del_page_from_lru_list(page
, lruvec
, lru
);
1817 spin_unlock_irq(&pgdat
->lru_lock
);
1823 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1824 * then get resheduled. When there are massive number of tasks doing page
1825 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1826 * the LRU list will go small and be scanned faster than necessary, leading to
1827 * unnecessary swapping, thrashing and OOM.
1829 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1830 struct scan_control
*sc
)
1832 unsigned long inactive
, isolated
;
1834 if (current_is_kswapd())
1837 if (!sane_reclaim(sc
))
1841 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1842 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1844 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1845 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1849 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1850 * won't get blocked by normal direct-reclaimers, forming a circular
1853 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1856 return isolated
> inactive
;
1860 * This moves pages from @list to corresponding LRU list.
1862 * We move them the other way if the page is referenced by one or more
1863 * processes, from rmap.
1865 * If the pages are mostly unmapped, the processing is fast and it is
1866 * appropriate to hold zone_lru_lock across the whole operation. But if
1867 * the pages are mapped, the processing is slow (page_referenced()) so we
1868 * should drop zone_lru_lock around each page. It's impossible to balance
1869 * this, so instead we remove the pages from the LRU while processing them.
1870 * It is safe to rely on PG_active against the non-LRU pages in here because
1871 * nobody will play with that bit on a non-LRU page.
1873 * The downside is that we have to touch page->_refcount against each page.
1874 * But we had to alter page->flags anyway.
1876 * Returns the number of pages moved to the given lruvec.
1879 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1880 struct list_head
*list
)
1882 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1883 int nr_pages
, nr_moved
= 0;
1884 LIST_HEAD(pages_to_free
);
1888 while (!list_empty(list
)) {
1889 page
= lru_to_page(list
);
1890 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1891 if (unlikely(!page_evictable(page
))) {
1892 list_del(&page
->lru
);
1893 spin_unlock_irq(&pgdat
->lru_lock
);
1894 putback_lru_page(page
);
1895 spin_lock_irq(&pgdat
->lru_lock
);
1898 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1901 lru
= page_lru(page
);
1903 nr_pages
= hpage_nr_pages(page
);
1904 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1905 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1907 if (put_page_testzero(page
)) {
1908 __ClearPageLRU(page
);
1909 __ClearPageActive(page
);
1910 del_page_from_lru_list(page
, lruvec
, lru
);
1912 if (unlikely(PageCompound(page
))) {
1913 spin_unlock_irq(&pgdat
->lru_lock
);
1914 mem_cgroup_uncharge(page
);
1915 (*get_compound_page_dtor(page
))(page
);
1916 spin_lock_irq(&pgdat
->lru_lock
);
1918 list_add(&page
->lru
, &pages_to_free
);
1920 nr_moved
+= nr_pages
;
1925 * To save our caller's stack, now use input list for pages to free.
1927 list_splice(&pages_to_free
, list
);
1933 * If a kernel thread (such as nfsd for loop-back mounts) services
1934 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1935 * In that case we should only throttle if the backing device it is
1936 * writing to is congested. In other cases it is safe to throttle.
1938 static int current_may_throttle(void)
1940 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1941 current
->backing_dev_info
== NULL
||
1942 bdi_write_congested(current
->backing_dev_info
);
1946 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1947 * of reclaimed pages
1949 static noinline_for_stack
unsigned long
1950 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1951 struct scan_control
*sc
, enum lru_list lru
)
1953 LIST_HEAD(page_list
);
1954 unsigned long nr_scanned
;
1955 unsigned long nr_reclaimed
= 0;
1956 unsigned long nr_taken
;
1957 struct reclaim_stat stat
;
1958 int file
= is_file_lru(lru
);
1959 enum vm_event_item item
;
1960 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1961 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1962 bool stalled
= false;
1964 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1968 /* wait a bit for the reclaimer. */
1972 /* We are about to die and free our memory. Return now. */
1973 if (fatal_signal_pending(current
))
1974 return SWAP_CLUSTER_MAX
;
1979 spin_lock_irq(&pgdat
->lru_lock
);
1981 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1982 &nr_scanned
, sc
, lru
);
1984 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1985 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1987 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1988 if (global_reclaim(sc
))
1989 __count_vm_events(item
, nr_scanned
);
1990 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1991 spin_unlock_irq(&pgdat
->lru_lock
);
1996 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1999 spin_lock_irq(&pgdat
->lru_lock
);
2001 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
2002 if (global_reclaim(sc
))
2003 __count_vm_events(item
, nr_reclaimed
);
2004 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
2005 reclaim_stat
->recent_rotated
[0] += stat
.nr_activate
[0];
2006 reclaim_stat
->recent_rotated
[1] += stat
.nr_activate
[1];
2008 move_pages_to_lru(lruvec
, &page_list
);
2010 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2012 spin_unlock_irq(&pgdat
->lru_lock
);
2014 mem_cgroup_uncharge_list(&page_list
);
2015 free_unref_page_list(&page_list
);
2018 * If dirty pages are scanned that are not queued for IO, it
2019 * implies that flushers are not doing their job. This can
2020 * happen when memory pressure pushes dirty pages to the end of
2021 * the LRU before the dirty limits are breached and the dirty
2022 * data has expired. It can also happen when the proportion of
2023 * dirty pages grows not through writes but through memory
2024 * pressure reclaiming all the clean cache. And in some cases,
2025 * the flushers simply cannot keep up with the allocation
2026 * rate. Nudge the flusher threads in case they are asleep.
2028 if (stat
.nr_unqueued_dirty
== nr_taken
)
2029 wakeup_flusher_threads(WB_REASON_VMSCAN
);
2031 sc
->nr
.dirty
+= stat
.nr_dirty
;
2032 sc
->nr
.congested
+= stat
.nr_congested
;
2033 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
2034 sc
->nr
.writeback
+= stat
.nr_writeback
;
2035 sc
->nr
.immediate
+= stat
.nr_immediate
;
2036 sc
->nr
.taken
+= nr_taken
;
2038 sc
->nr
.file_taken
+= nr_taken
;
2040 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2041 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2042 return nr_reclaimed
;
2045 static void shrink_active_list(unsigned long nr_to_scan
,
2046 struct lruvec
*lruvec
,
2047 struct scan_control
*sc
,
2050 unsigned long nr_taken
;
2051 unsigned long nr_scanned
;
2052 unsigned long vm_flags
;
2053 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2054 LIST_HEAD(l_active
);
2055 LIST_HEAD(l_inactive
);
2057 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2058 unsigned nr_deactivate
, nr_activate
;
2059 unsigned nr_rotated
= 0;
2060 int file
= is_file_lru(lru
);
2061 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2065 spin_lock_irq(&pgdat
->lru_lock
);
2067 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2068 &nr_scanned
, sc
, lru
);
2070 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2071 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2073 __count_vm_events(PGREFILL
, nr_scanned
);
2074 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2076 spin_unlock_irq(&pgdat
->lru_lock
);
2078 while (!list_empty(&l_hold
)) {
2080 page
= lru_to_page(&l_hold
);
2081 list_del(&page
->lru
);
2083 if (unlikely(!page_evictable(page
))) {
2084 putback_lru_page(page
);
2088 if (unlikely(buffer_heads_over_limit
)) {
2089 if (page_has_private(page
) && trylock_page(page
)) {
2090 if (page_has_private(page
))
2091 try_to_release_page(page
, 0);
2096 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2098 nr_rotated
+= hpage_nr_pages(page
);
2100 * Identify referenced, file-backed active pages and
2101 * give them one more trip around the active list. So
2102 * that executable code get better chances to stay in
2103 * memory under moderate memory pressure. Anon pages
2104 * are not likely to be evicted by use-once streaming
2105 * IO, plus JVM can create lots of anon VM_EXEC pages,
2106 * so we ignore them here.
2108 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2109 list_add(&page
->lru
, &l_active
);
2114 ClearPageActive(page
); /* we are de-activating */
2115 SetPageWorkingset(page
);
2116 list_add(&page
->lru
, &l_inactive
);
2120 * Move pages back to the lru list.
2122 spin_lock_irq(&pgdat
->lru_lock
);
2124 * Count referenced pages from currently used mappings as rotated,
2125 * even though only some of them are actually re-activated. This
2126 * helps balance scan pressure between file and anonymous pages in
2129 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2131 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2132 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2133 /* Keep all free pages in l_active list */
2134 list_splice(&l_inactive
, &l_active
);
2136 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2137 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2139 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2140 spin_unlock_irq(&pgdat
->lru_lock
);
2142 mem_cgroup_uncharge_list(&l_active
);
2143 free_unref_page_list(&l_active
);
2144 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2145 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2149 * The inactive anon list should be small enough that the VM never has
2150 * to do too much work.
2152 * The inactive file list should be small enough to leave most memory
2153 * to the established workingset on the scan-resistant active list,
2154 * but large enough to avoid thrashing the aggregate readahead window.
2156 * Both inactive lists should also be large enough that each inactive
2157 * page has a chance to be referenced again before it is reclaimed.
2159 * If that fails and refaulting is observed, the inactive list grows.
2161 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2162 * on this LRU, maintained by the pageout code. An inactive_ratio
2163 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2166 * memory ratio inactive
2167 * -------------------------------------
2176 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2177 struct scan_control
*sc
, bool trace
)
2179 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2180 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2181 enum lru_list inactive_lru
= file
* LRU_FILE
;
2182 unsigned long inactive
, active
;
2183 unsigned long inactive_ratio
;
2184 unsigned long refaults
;
2188 * If we don't have swap space, anonymous page deactivation
2191 if (!file
&& !total_swap_pages
)
2194 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2195 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2198 * When refaults are being observed, it means a new workingset
2199 * is being established. Disable active list protection to get
2200 * rid of the stale workingset quickly.
2202 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
2203 if (file
&& lruvec
->refaults
!= refaults
) {
2206 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2208 inactive_ratio
= int_sqrt(10 * gb
);
2214 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2215 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2216 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2217 inactive_ratio
, file
);
2219 return inactive
* inactive_ratio
< active
;
2222 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2223 struct lruvec
*lruvec
, struct scan_control
*sc
)
2225 if (is_active_lru(lru
)) {
2226 if (inactive_list_is_low(lruvec
, is_file_lru(lru
), sc
, true))
2227 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2231 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2242 * Determine how aggressively the anon and file LRU lists should be
2243 * scanned. The relative value of each set of LRU lists is determined
2244 * by looking at the fraction of the pages scanned we did rotate back
2245 * onto the active list instead of evict.
2247 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2248 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2250 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2251 struct scan_control
*sc
, unsigned long *nr
,
2252 unsigned long *lru_pages
)
2254 int swappiness
= mem_cgroup_swappiness(memcg
);
2255 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2257 u64 denominator
= 0; /* gcc */
2258 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2259 unsigned long anon_prio
, file_prio
;
2260 enum scan_balance scan_balance
;
2261 unsigned long anon
, file
;
2262 unsigned long ap
, fp
;
2265 /* If we have no swap space, do not bother scanning anon pages. */
2266 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2267 scan_balance
= SCAN_FILE
;
2272 * Global reclaim will swap to prevent OOM even with no
2273 * swappiness, but memcg users want to use this knob to
2274 * disable swapping for individual groups completely when
2275 * using the memory controller's swap limit feature would be
2278 if (!global_reclaim(sc
) && !swappiness
) {
2279 scan_balance
= SCAN_FILE
;
2284 * Do not apply any pressure balancing cleverness when the
2285 * system is close to OOM, scan both anon and file equally
2286 * (unless the swappiness setting disagrees with swapping).
2288 if (!sc
->priority
&& swappiness
) {
2289 scan_balance
= SCAN_EQUAL
;
2294 * Prevent the reclaimer from falling into the cache trap: as
2295 * cache pages start out inactive, every cache fault will tip
2296 * the scan balance towards the file LRU. And as the file LRU
2297 * shrinks, so does the window for rotation from references.
2298 * This means we have a runaway feedback loop where a tiny
2299 * thrashing file LRU becomes infinitely more attractive than
2300 * anon pages. Try to detect this based on file LRU size.
2302 if (global_reclaim(sc
)) {
2303 unsigned long pgdatfile
;
2304 unsigned long pgdatfree
;
2306 unsigned long total_high_wmark
= 0;
2308 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2309 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2310 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2312 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2313 struct zone
*zone
= &pgdat
->node_zones
[z
];
2314 if (!managed_zone(zone
))
2317 total_high_wmark
+= high_wmark_pages(zone
);
2320 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2322 * Force SCAN_ANON if there are enough inactive
2323 * anonymous pages on the LRU in eligible zones.
2324 * Otherwise, the small LRU gets thrashed.
2326 if (!inactive_list_is_low(lruvec
, false, sc
, false) &&
2327 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2329 scan_balance
= SCAN_ANON
;
2336 * If there is enough inactive page cache, i.e. if the size of the
2337 * inactive list is greater than that of the active list *and* the
2338 * inactive list actually has some pages to scan on this priority, we
2339 * do not reclaim anything from the anonymous working set right now.
2340 * Without the second condition we could end up never scanning an
2341 * lruvec even if it has plenty of old anonymous pages unless the
2342 * system is under heavy pressure.
2344 if (!inactive_list_is_low(lruvec
, true, sc
, false) &&
2345 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2346 scan_balance
= SCAN_FILE
;
2350 scan_balance
= SCAN_FRACT
;
2353 * With swappiness at 100, anonymous and file have the same priority.
2354 * This scanning priority is essentially the inverse of IO cost.
2356 anon_prio
= swappiness
;
2357 file_prio
= 200 - anon_prio
;
2360 * OK, so we have swap space and a fair amount of page cache
2361 * pages. We use the recently rotated / recently scanned
2362 * ratios to determine how valuable each cache is.
2364 * Because workloads change over time (and to avoid overflow)
2365 * we keep these statistics as a floating average, which ends
2366 * up weighing recent references more than old ones.
2368 * anon in [0], file in [1]
2371 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2372 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2373 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2374 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2376 spin_lock_irq(&pgdat
->lru_lock
);
2377 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2378 reclaim_stat
->recent_scanned
[0] /= 2;
2379 reclaim_stat
->recent_rotated
[0] /= 2;
2382 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2383 reclaim_stat
->recent_scanned
[1] /= 2;
2384 reclaim_stat
->recent_rotated
[1] /= 2;
2388 * The amount of pressure on anon vs file pages is inversely
2389 * proportional to the fraction of recently scanned pages on
2390 * each list that were recently referenced and in active use.
2392 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2393 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2395 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2396 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2397 spin_unlock_irq(&pgdat
->lru_lock
);
2401 denominator
= ap
+ fp
+ 1;
2404 for_each_evictable_lru(lru
) {
2405 int file
= is_file_lru(lru
);
2409 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2410 scan
= size
>> sc
->priority
;
2412 * If the cgroup's already been deleted, make sure to
2413 * scrape out the remaining cache.
2415 if (!scan
&& !mem_cgroup_online(memcg
))
2416 scan
= min(size
, SWAP_CLUSTER_MAX
);
2418 switch (scan_balance
) {
2420 /* Scan lists relative to size */
2424 * Scan types proportional to swappiness and
2425 * their relative recent reclaim efficiency.
2426 * Make sure we don't miss the last page
2427 * because of a round-off error.
2429 scan
= DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2434 /* Scan one type exclusively */
2435 if ((scan_balance
== SCAN_FILE
) != file
) {
2441 /* Look ma, no brain */
2451 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2453 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2454 struct scan_control
*sc
, unsigned long *lru_pages
)
2456 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2457 unsigned long nr
[NR_LRU_LISTS
];
2458 unsigned long targets
[NR_LRU_LISTS
];
2459 unsigned long nr_to_scan
;
2461 unsigned long nr_reclaimed
= 0;
2462 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2463 struct blk_plug plug
;
2466 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2468 /* Record the original scan target for proportional adjustments later */
2469 memcpy(targets
, nr
, sizeof(nr
));
2472 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2473 * event that can occur when there is little memory pressure e.g.
2474 * multiple streaming readers/writers. Hence, we do not abort scanning
2475 * when the requested number of pages are reclaimed when scanning at
2476 * DEF_PRIORITY on the assumption that the fact we are direct
2477 * reclaiming implies that kswapd is not keeping up and it is best to
2478 * do a batch of work at once. For memcg reclaim one check is made to
2479 * abort proportional reclaim if either the file or anon lru has already
2480 * dropped to zero at the first pass.
2482 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2483 sc
->priority
== DEF_PRIORITY
);
2485 blk_start_plug(&plug
);
2486 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2487 nr
[LRU_INACTIVE_FILE
]) {
2488 unsigned long nr_anon
, nr_file
, percentage
;
2489 unsigned long nr_scanned
;
2491 for_each_evictable_lru(lru
) {
2493 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2494 nr
[lru
] -= nr_to_scan
;
2496 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2503 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2507 * For kswapd and memcg, reclaim at least the number of pages
2508 * requested. Ensure that the anon and file LRUs are scanned
2509 * proportionally what was requested by get_scan_count(). We
2510 * stop reclaiming one LRU and reduce the amount scanning
2511 * proportional to the original scan target.
2513 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2514 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2517 * It's just vindictive to attack the larger once the smaller
2518 * has gone to zero. And given the way we stop scanning the
2519 * smaller below, this makes sure that we only make one nudge
2520 * towards proportionality once we've got nr_to_reclaim.
2522 if (!nr_file
|| !nr_anon
)
2525 if (nr_file
> nr_anon
) {
2526 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2527 targets
[LRU_ACTIVE_ANON
] + 1;
2529 percentage
= nr_anon
* 100 / scan_target
;
2531 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2532 targets
[LRU_ACTIVE_FILE
] + 1;
2534 percentage
= nr_file
* 100 / scan_target
;
2537 /* Stop scanning the smaller of the LRU */
2539 nr
[lru
+ LRU_ACTIVE
] = 0;
2542 * Recalculate the other LRU scan count based on its original
2543 * scan target and the percentage scanning already complete
2545 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2546 nr_scanned
= targets
[lru
] - nr
[lru
];
2547 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2548 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2551 nr_scanned
= targets
[lru
] - nr
[lru
];
2552 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2553 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2555 scan_adjusted
= true;
2557 blk_finish_plug(&plug
);
2558 sc
->nr_reclaimed
+= nr_reclaimed
;
2561 * Even if we did not try to evict anon pages at all, we want to
2562 * rebalance the anon lru active/inactive ratio.
2564 if (inactive_list_is_low(lruvec
, false, sc
, true))
2565 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2566 sc
, LRU_ACTIVE_ANON
);
2569 /* Use reclaim/compaction for costly allocs or under memory pressure */
2570 static bool in_reclaim_compaction(struct scan_control
*sc
)
2572 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2573 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2574 sc
->priority
< DEF_PRIORITY
- 2))
2581 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2582 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2583 * true if more pages should be reclaimed such that when the page allocator
2584 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2585 * It will give up earlier than that if there is difficulty reclaiming pages.
2587 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2588 unsigned long nr_reclaimed
,
2589 unsigned long nr_scanned
,
2590 struct scan_control
*sc
)
2592 unsigned long pages_for_compaction
;
2593 unsigned long inactive_lru_pages
;
2596 /* If not in reclaim/compaction mode, stop */
2597 if (!in_reclaim_compaction(sc
))
2600 /* Consider stopping depending on scan and reclaim activity */
2601 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2603 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2604 * full LRU list has been scanned and we are still failing
2605 * to reclaim pages. This full LRU scan is potentially
2606 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2608 if (!nr_reclaimed
&& !nr_scanned
)
2612 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2613 * fail without consequence, stop if we failed to reclaim
2614 * any pages from the last SWAP_CLUSTER_MAX number of
2615 * pages that were scanned. This will return to the
2616 * caller faster at the risk reclaim/compaction and
2617 * the resulting allocation attempt fails
2624 * If we have not reclaimed enough pages for compaction and the
2625 * inactive lists are large enough, continue reclaiming
2627 pages_for_compaction
= compact_gap(sc
->order
);
2628 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2629 if (get_nr_swap_pages() > 0)
2630 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2631 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2632 inactive_lru_pages
> pages_for_compaction
)
2635 /* If compaction would go ahead or the allocation would succeed, stop */
2636 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2637 struct zone
*zone
= &pgdat
->node_zones
[z
];
2638 if (!managed_zone(zone
))
2641 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2642 case COMPACT_SUCCESS
:
2643 case COMPACT_CONTINUE
:
2646 /* check next zone */
2653 static bool pgdat_memcg_congested(pg_data_t
*pgdat
, struct mem_cgroup
*memcg
)
2655 return test_bit(PGDAT_CONGESTED
, &pgdat
->flags
) ||
2656 (memcg
&& memcg_congested(pgdat
, memcg
));
2659 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2661 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2662 unsigned long nr_reclaimed
, nr_scanned
;
2663 bool reclaimable
= false;
2666 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2667 struct mem_cgroup_reclaim_cookie reclaim
= {
2669 .priority
= sc
->priority
,
2671 unsigned long node_lru_pages
= 0;
2672 struct mem_cgroup
*memcg
;
2674 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2676 nr_reclaimed
= sc
->nr_reclaimed
;
2677 nr_scanned
= sc
->nr_scanned
;
2679 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2681 unsigned long lru_pages
;
2682 unsigned long reclaimed
;
2683 unsigned long scanned
;
2685 switch (mem_cgroup_protected(root
, memcg
)) {
2686 case MEMCG_PROT_MIN
:
2689 * If there is no reclaimable memory, OOM.
2692 case MEMCG_PROT_LOW
:
2695 * Respect the protection only as long as
2696 * there is an unprotected supply
2697 * of reclaimable memory from other cgroups.
2699 if (!sc
->memcg_low_reclaim
) {
2700 sc
->memcg_low_skipped
= 1;
2703 memcg_memory_event(memcg
, MEMCG_LOW
);
2705 case MEMCG_PROT_NONE
:
2709 reclaimed
= sc
->nr_reclaimed
;
2710 scanned
= sc
->nr_scanned
;
2711 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2712 node_lru_pages
+= lru_pages
;
2714 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2717 /* Record the group's reclaim efficiency */
2718 vmpressure(sc
->gfp_mask
, memcg
, false,
2719 sc
->nr_scanned
- scanned
,
2720 sc
->nr_reclaimed
- reclaimed
);
2723 * Kswapd have to scan all memory cgroups to fulfill
2724 * the overall scan target for the node.
2726 * Limit reclaim, on the other hand, only cares about
2727 * nr_to_reclaim pages to be reclaimed and it will
2728 * retry with decreasing priority if one round over the
2729 * whole hierarchy is not sufficient.
2731 if (!current_is_kswapd() &&
2732 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2733 mem_cgroup_iter_break(root
, memcg
);
2736 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2738 if (reclaim_state
) {
2739 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2740 reclaim_state
->reclaimed_slab
= 0;
2743 /* Record the subtree's reclaim efficiency */
2744 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2745 sc
->nr_scanned
- nr_scanned
,
2746 sc
->nr_reclaimed
- nr_reclaimed
);
2748 if (sc
->nr_reclaimed
- nr_reclaimed
)
2751 if (current_is_kswapd()) {
2753 * If reclaim is isolating dirty pages under writeback,
2754 * it implies that the long-lived page allocation rate
2755 * is exceeding the page laundering rate. Either the
2756 * global limits are not being effective at throttling
2757 * processes due to the page distribution throughout
2758 * zones or there is heavy usage of a slow backing
2759 * device. The only option is to throttle from reclaim
2760 * context which is not ideal as there is no guarantee
2761 * the dirtying process is throttled in the same way
2762 * balance_dirty_pages() manages.
2764 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2765 * count the number of pages under pages flagged for
2766 * immediate reclaim and stall if any are encountered
2767 * in the nr_immediate check below.
2769 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2770 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2773 * Tag a node as congested if all the dirty pages
2774 * scanned were backed by a congested BDI and
2775 * wait_iff_congested will stall.
2777 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2778 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
2780 /* Allow kswapd to start writing pages during reclaim.*/
2781 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2782 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2785 * If kswapd scans pages marked marked for immediate
2786 * reclaim and under writeback (nr_immediate), it
2787 * implies that pages are cycling through the LRU
2788 * faster than they are written so also forcibly stall.
2790 if (sc
->nr
.immediate
)
2791 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2795 * Legacy memcg will stall in page writeback so avoid forcibly
2796 * stalling in wait_iff_congested().
2798 if (!global_reclaim(sc
) && sane_reclaim(sc
) &&
2799 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2800 set_memcg_congestion(pgdat
, root
, true);
2803 * Stall direct reclaim for IO completions if underlying BDIs
2804 * and node is congested. Allow kswapd to continue until it
2805 * starts encountering unqueued dirty pages or cycling through
2806 * the LRU too quickly.
2808 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
2809 current_may_throttle() && pgdat_memcg_congested(pgdat
, root
))
2810 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2812 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2813 sc
->nr_scanned
- nr_scanned
, sc
));
2816 * Kswapd gives up on balancing particular nodes after too
2817 * many failures to reclaim anything from them and goes to
2818 * sleep. On reclaim progress, reset the failure counter. A
2819 * successful direct reclaim run will revive a dormant kswapd.
2822 pgdat
->kswapd_failures
= 0;
2828 * Returns true if compaction should go ahead for a costly-order request, or
2829 * the allocation would already succeed without compaction. Return false if we
2830 * should reclaim first.
2832 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2834 unsigned long watermark
;
2835 enum compact_result suitable
;
2837 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2838 if (suitable
== COMPACT_SUCCESS
)
2839 /* Allocation should succeed already. Don't reclaim. */
2841 if (suitable
== COMPACT_SKIPPED
)
2842 /* Compaction cannot yet proceed. Do reclaim. */
2846 * Compaction is already possible, but it takes time to run and there
2847 * are potentially other callers using the pages just freed. So proceed
2848 * with reclaim to make a buffer of free pages available to give
2849 * compaction a reasonable chance of completing and allocating the page.
2850 * Note that we won't actually reclaim the whole buffer in one attempt
2851 * as the target watermark in should_continue_reclaim() is lower. But if
2852 * we are already above the high+gap watermark, don't reclaim at all.
2854 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2856 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2860 * This is the direct reclaim path, for page-allocating processes. We only
2861 * try to reclaim pages from zones which will satisfy the caller's allocation
2864 * If a zone is deemed to be full of pinned pages then just give it a light
2865 * scan then give up on it.
2867 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2871 unsigned long nr_soft_reclaimed
;
2872 unsigned long nr_soft_scanned
;
2874 pg_data_t
*last_pgdat
= NULL
;
2877 * If the number of buffer_heads in the machine exceeds the maximum
2878 * allowed level, force direct reclaim to scan the highmem zone as
2879 * highmem pages could be pinning lowmem pages storing buffer_heads
2881 orig_mask
= sc
->gfp_mask
;
2882 if (buffer_heads_over_limit
) {
2883 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2884 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2887 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2888 sc
->reclaim_idx
, sc
->nodemask
) {
2890 * Take care memory controller reclaiming has small influence
2893 if (global_reclaim(sc
)) {
2894 if (!cpuset_zone_allowed(zone
,
2895 GFP_KERNEL
| __GFP_HARDWALL
))
2899 * If we already have plenty of memory free for
2900 * compaction in this zone, don't free any more.
2901 * Even though compaction is invoked for any
2902 * non-zero order, only frequent costly order
2903 * reclamation is disruptive enough to become a
2904 * noticeable problem, like transparent huge
2907 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2908 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2909 compaction_ready(zone
, sc
)) {
2910 sc
->compaction_ready
= true;
2915 * Shrink each node in the zonelist once. If the
2916 * zonelist is ordered by zone (not the default) then a
2917 * node may be shrunk multiple times but in that case
2918 * the user prefers lower zones being preserved.
2920 if (zone
->zone_pgdat
== last_pgdat
)
2924 * This steals pages from memory cgroups over softlimit
2925 * and returns the number of reclaimed pages and
2926 * scanned pages. This works for global memory pressure
2927 * and balancing, not for a memcg's limit.
2929 nr_soft_scanned
= 0;
2930 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2931 sc
->order
, sc
->gfp_mask
,
2933 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2934 sc
->nr_scanned
+= nr_soft_scanned
;
2935 /* need some check for avoid more shrink_zone() */
2938 /* See comment about same check for global reclaim above */
2939 if (zone
->zone_pgdat
== last_pgdat
)
2941 last_pgdat
= zone
->zone_pgdat
;
2942 shrink_node(zone
->zone_pgdat
, sc
);
2946 * Restore to original mask to avoid the impact on the caller if we
2947 * promoted it to __GFP_HIGHMEM.
2949 sc
->gfp_mask
= orig_mask
;
2952 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2954 struct mem_cgroup
*memcg
;
2956 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2958 unsigned long refaults
;
2959 struct lruvec
*lruvec
;
2961 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2962 refaults
= lruvec_page_state_local(lruvec
, WORKINGSET_ACTIVATE
);
2963 lruvec
->refaults
= refaults
;
2964 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2968 * This is the main entry point to direct page reclaim.
2970 * If a full scan of the inactive list fails to free enough memory then we
2971 * are "out of memory" and something needs to be killed.
2973 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2974 * high - the zone may be full of dirty or under-writeback pages, which this
2975 * caller can't do much about. We kick the writeback threads and take explicit
2976 * naps in the hope that some of these pages can be written. But if the
2977 * allocating task holds filesystem locks which prevent writeout this might not
2978 * work, and the allocation attempt will fail.
2980 * returns: 0, if no pages reclaimed
2981 * else, the number of pages reclaimed
2983 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2984 struct scan_control
*sc
)
2986 int initial_priority
= sc
->priority
;
2987 pg_data_t
*last_pgdat
;
2991 delayacct_freepages_start();
2993 if (global_reclaim(sc
))
2994 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2997 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3000 shrink_zones(zonelist
, sc
);
3002 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3005 if (sc
->compaction_ready
)
3009 * If we're getting trouble reclaiming, start doing
3010 * writepage even in laptop mode.
3012 if (sc
->priority
< DEF_PRIORITY
- 2)
3013 sc
->may_writepage
= 1;
3014 } while (--sc
->priority
>= 0);
3017 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3019 if (zone
->zone_pgdat
== last_pgdat
)
3021 last_pgdat
= zone
->zone_pgdat
;
3022 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3023 set_memcg_congestion(last_pgdat
, sc
->target_mem_cgroup
, false);
3026 delayacct_freepages_end();
3028 if (sc
->nr_reclaimed
)
3029 return sc
->nr_reclaimed
;
3031 /* Aborted reclaim to try compaction? don't OOM, then */
3032 if (sc
->compaction_ready
)
3035 /* Untapped cgroup reserves? Don't OOM, retry. */
3036 if (sc
->memcg_low_skipped
) {
3037 sc
->priority
= initial_priority
;
3038 sc
->memcg_low_reclaim
= 1;
3039 sc
->memcg_low_skipped
= 0;
3046 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3049 unsigned long pfmemalloc_reserve
= 0;
3050 unsigned long free_pages
= 0;
3054 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3057 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3058 zone
= &pgdat
->node_zones
[i
];
3059 if (!managed_zone(zone
))
3062 if (!zone_reclaimable_pages(zone
))
3065 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3066 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3069 /* If there are no reserves (unexpected config) then do not throttle */
3070 if (!pfmemalloc_reserve
)
3073 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3075 /* kswapd must be awake if processes are being throttled */
3076 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3077 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
3078 (enum zone_type
)ZONE_NORMAL
);
3079 wake_up_interruptible(&pgdat
->kswapd_wait
);
3086 * Throttle direct reclaimers if backing storage is backed by the network
3087 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3088 * depleted. kswapd will continue to make progress and wake the processes
3089 * when the low watermark is reached.
3091 * Returns true if a fatal signal was delivered during throttling. If this
3092 * happens, the page allocator should not consider triggering the OOM killer.
3094 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3095 nodemask_t
*nodemask
)
3099 pg_data_t
*pgdat
= NULL
;
3102 * Kernel threads should not be throttled as they may be indirectly
3103 * responsible for cleaning pages necessary for reclaim to make forward
3104 * progress. kjournald for example may enter direct reclaim while
3105 * committing a transaction where throttling it could forcing other
3106 * processes to block on log_wait_commit().
3108 if (current
->flags
& PF_KTHREAD
)
3112 * If a fatal signal is pending, this process should not throttle.
3113 * It should return quickly so it can exit and free its memory
3115 if (fatal_signal_pending(current
))
3119 * Check if the pfmemalloc reserves are ok by finding the first node
3120 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3121 * GFP_KERNEL will be required for allocating network buffers when
3122 * swapping over the network so ZONE_HIGHMEM is unusable.
3124 * Throttling is based on the first usable node and throttled processes
3125 * wait on a queue until kswapd makes progress and wakes them. There
3126 * is an affinity then between processes waking up and where reclaim
3127 * progress has been made assuming the process wakes on the same node.
3128 * More importantly, processes running on remote nodes will not compete
3129 * for remote pfmemalloc reserves and processes on different nodes
3130 * should make reasonable progress.
3132 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3133 gfp_zone(gfp_mask
), nodemask
) {
3134 if (zone_idx(zone
) > ZONE_NORMAL
)
3137 /* Throttle based on the first usable node */
3138 pgdat
= zone
->zone_pgdat
;
3139 if (allow_direct_reclaim(pgdat
))
3144 /* If no zone was usable by the allocation flags then do not throttle */
3148 /* Account for the throttling */
3149 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3152 * If the caller cannot enter the filesystem, it's possible that it
3153 * is due to the caller holding an FS lock or performing a journal
3154 * transaction in the case of a filesystem like ext[3|4]. In this case,
3155 * it is not safe to block on pfmemalloc_wait as kswapd could be
3156 * blocked waiting on the same lock. Instead, throttle for up to a
3157 * second before continuing.
3159 if (!(gfp_mask
& __GFP_FS
)) {
3160 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3161 allow_direct_reclaim(pgdat
), HZ
);
3166 /* Throttle until kswapd wakes the process */
3167 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3168 allow_direct_reclaim(pgdat
));
3171 if (fatal_signal_pending(current
))
3178 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3179 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3181 unsigned long nr_reclaimed
;
3182 struct scan_control sc
= {
3183 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3184 .gfp_mask
= current_gfp_context(gfp_mask
),
3185 .reclaim_idx
= gfp_zone(gfp_mask
),
3187 .nodemask
= nodemask
,
3188 .priority
= DEF_PRIORITY
,
3189 .may_writepage
= !laptop_mode
,
3195 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3196 * Confirm they are large enough for max values.
3198 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3199 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3200 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3203 * Do not enter reclaim if fatal signal was delivered while throttled.
3204 * 1 is returned so that the page allocator does not OOM kill at this
3207 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3210 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3211 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3213 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3215 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3216 set_task_reclaim_state(current
, NULL
);
3218 return nr_reclaimed
;
3223 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3224 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3225 gfp_t gfp_mask
, bool noswap
,
3227 unsigned long *nr_scanned
)
3229 struct scan_control sc
= {
3230 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3231 .target_mem_cgroup
= memcg
,
3232 .may_writepage
= !laptop_mode
,
3234 .reclaim_idx
= MAX_NR_ZONES
- 1,
3235 .may_swap
= !noswap
,
3237 unsigned long lru_pages
;
3239 WARN_ON_ONCE(!current
->reclaim_state
);
3241 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3242 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3244 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3248 * NOTE: Although we can get the priority field, using it
3249 * here is not a good idea, since it limits the pages we can scan.
3250 * if we don't reclaim here, the shrink_node from balance_pgdat
3251 * will pick up pages from other mem cgroup's as well. We hack
3252 * the priority and make it zero.
3254 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3256 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3258 *nr_scanned
= sc
.nr_scanned
;
3260 return sc
.nr_reclaimed
;
3263 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3264 unsigned long nr_pages
,
3268 struct zonelist
*zonelist
;
3269 unsigned long nr_reclaimed
;
3270 unsigned long pflags
;
3272 unsigned int noreclaim_flag
;
3273 struct scan_control sc
= {
3274 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3275 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3276 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3277 .reclaim_idx
= MAX_NR_ZONES
- 1,
3278 .target_mem_cgroup
= memcg
,
3279 .priority
= DEF_PRIORITY
,
3280 .may_writepage
= !laptop_mode
,
3282 .may_swap
= may_swap
,
3285 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3287 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3288 * take care of from where we get pages. So the node where we start the
3289 * scan does not need to be the current node.
3291 nid
= mem_cgroup_select_victim_node(memcg
);
3293 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3295 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3297 psi_memstall_enter(&pflags
);
3298 noreclaim_flag
= memalloc_noreclaim_save();
3300 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3302 memalloc_noreclaim_restore(noreclaim_flag
);
3303 psi_memstall_leave(&pflags
);
3305 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3306 set_task_reclaim_state(current
, NULL
);
3308 return nr_reclaimed
;
3312 static void age_active_anon(struct pglist_data
*pgdat
,
3313 struct scan_control
*sc
)
3315 struct mem_cgroup
*memcg
;
3317 if (!total_swap_pages
)
3320 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3322 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3324 if (inactive_list_is_low(lruvec
, false, sc
, true))
3325 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3326 sc
, LRU_ACTIVE_ANON
);
3328 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3332 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int classzone_idx
)
3338 * Check for watermark boosts top-down as the higher zones
3339 * are more likely to be boosted. Both watermarks and boosts
3340 * should not be checked at the time time as reclaim would
3341 * start prematurely when there is no boosting and a lower
3344 for (i
= classzone_idx
; i
>= 0; i
--) {
3345 zone
= pgdat
->node_zones
+ i
;
3346 if (!managed_zone(zone
))
3349 if (zone
->watermark_boost
)
3357 * Returns true if there is an eligible zone balanced for the request order
3360 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3363 unsigned long mark
= -1;
3367 * Check watermarks bottom-up as lower zones are more likely to
3370 for (i
= 0; i
<= classzone_idx
; i
++) {
3371 zone
= pgdat
->node_zones
+ i
;
3373 if (!managed_zone(zone
))
3376 mark
= high_wmark_pages(zone
);
3377 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3382 * If a node has no populated zone within classzone_idx, it does not
3383 * need balancing by definition. This can happen if a zone-restricted
3384 * allocation tries to wake a remote kswapd.
3392 /* Clear pgdat state for congested, dirty or under writeback. */
3393 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3395 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3396 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3397 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3401 * Prepare kswapd for sleeping. This verifies that there are no processes
3402 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3404 * Returns true if kswapd is ready to sleep
3406 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3409 * The throttled processes are normally woken up in balance_pgdat() as
3410 * soon as allow_direct_reclaim() is true. But there is a potential
3411 * race between when kswapd checks the watermarks and a process gets
3412 * throttled. There is also a potential race if processes get
3413 * throttled, kswapd wakes, a large process exits thereby balancing the
3414 * zones, which causes kswapd to exit balance_pgdat() before reaching
3415 * the wake up checks. If kswapd is going to sleep, no process should
3416 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3417 * the wake up is premature, processes will wake kswapd and get
3418 * throttled again. The difference from wake ups in balance_pgdat() is
3419 * that here we are under prepare_to_wait().
3421 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3422 wake_up_all(&pgdat
->pfmemalloc_wait
);
3424 /* Hopeless node, leave it to direct reclaim */
3425 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3428 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3429 clear_pgdat_congested(pgdat
);
3437 * kswapd shrinks a node of pages that are at or below the highest usable
3438 * zone that is currently unbalanced.
3440 * Returns true if kswapd scanned at least the requested number of pages to
3441 * reclaim or if the lack of progress was due to pages under writeback.
3442 * This is used to determine if the scanning priority needs to be raised.
3444 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3445 struct scan_control
*sc
)
3450 /* Reclaim a number of pages proportional to the number of zones */
3451 sc
->nr_to_reclaim
= 0;
3452 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3453 zone
= pgdat
->node_zones
+ z
;
3454 if (!managed_zone(zone
))
3457 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3461 * Historically care was taken to put equal pressure on all zones but
3462 * now pressure is applied based on node LRU order.
3464 shrink_node(pgdat
, sc
);
3467 * Fragmentation may mean that the system cannot be rebalanced for
3468 * high-order allocations. If twice the allocation size has been
3469 * reclaimed then recheck watermarks only at order-0 to prevent
3470 * excessive reclaim. Assume that a process requested a high-order
3471 * can direct reclaim/compact.
3473 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3476 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3480 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3481 * that are eligible for use by the caller until at least one zone is
3484 * Returns the order kswapd finished reclaiming at.
3486 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3487 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3488 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3489 * or lower is eligible for reclaim until at least one usable zone is
3492 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3495 unsigned long nr_soft_reclaimed
;
3496 unsigned long nr_soft_scanned
;
3497 unsigned long pflags
;
3498 unsigned long nr_boost_reclaim
;
3499 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
3502 struct scan_control sc
= {
3503 .gfp_mask
= GFP_KERNEL
,
3508 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3509 psi_memstall_enter(&pflags
);
3510 __fs_reclaim_acquire();
3512 count_vm_event(PAGEOUTRUN
);
3515 * Account for the reclaim boost. Note that the zone boost is left in
3516 * place so that parallel allocations that are near the watermark will
3517 * stall or direct reclaim until kswapd is finished.
3519 nr_boost_reclaim
= 0;
3520 for (i
= 0; i
<= classzone_idx
; i
++) {
3521 zone
= pgdat
->node_zones
+ i
;
3522 if (!managed_zone(zone
))
3525 nr_boost_reclaim
+= zone
->watermark_boost
;
3526 zone_boosts
[i
] = zone
->watermark_boost
;
3528 boosted
= nr_boost_reclaim
;
3531 sc
.priority
= DEF_PRIORITY
;
3533 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3534 bool raise_priority
= true;
3538 sc
.reclaim_idx
= classzone_idx
;
3541 * If the number of buffer_heads exceeds the maximum allowed
3542 * then consider reclaiming from all zones. This has a dual
3543 * purpose -- on 64-bit systems it is expected that
3544 * buffer_heads are stripped during active rotation. On 32-bit
3545 * systems, highmem pages can pin lowmem memory and shrinking
3546 * buffers can relieve lowmem pressure. Reclaim may still not
3547 * go ahead if all eligible zones for the original allocation
3548 * request are balanced to avoid excessive reclaim from kswapd.
3550 if (buffer_heads_over_limit
) {
3551 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3552 zone
= pgdat
->node_zones
+ i
;
3553 if (!managed_zone(zone
))
3562 * If the pgdat is imbalanced then ignore boosting and preserve
3563 * the watermarks for a later time and restart. Note that the
3564 * zone watermarks will be still reset at the end of balancing
3565 * on the grounds that the normal reclaim should be enough to
3566 * re-evaluate if boosting is required when kswapd next wakes.
3568 balanced
= pgdat_balanced(pgdat
, sc
.order
, classzone_idx
);
3569 if (!balanced
&& nr_boost_reclaim
) {
3570 nr_boost_reclaim
= 0;
3575 * If boosting is not active then only reclaim if there are no
3576 * eligible zones. Note that sc.reclaim_idx is not used as
3577 * buffer_heads_over_limit may have adjusted it.
3579 if (!nr_boost_reclaim
&& balanced
)
3582 /* Limit the priority of boosting to avoid reclaim writeback */
3583 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
3584 raise_priority
= false;
3587 * Do not writeback or swap pages for boosted reclaim. The
3588 * intent is to relieve pressure not issue sub-optimal IO
3589 * from reclaim context. If no pages are reclaimed, the
3590 * reclaim will be aborted.
3592 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
3593 sc
.may_swap
= !nr_boost_reclaim
;
3596 * Do some background aging of the anon list, to give
3597 * pages a chance to be referenced before reclaiming. All
3598 * pages are rotated regardless of classzone as this is
3599 * about consistent aging.
3601 age_active_anon(pgdat
, &sc
);
3604 * If we're getting trouble reclaiming, start doing writepage
3605 * even in laptop mode.
3607 if (sc
.priority
< DEF_PRIORITY
- 2)
3608 sc
.may_writepage
= 1;
3610 /* Call soft limit reclaim before calling shrink_node. */
3612 nr_soft_scanned
= 0;
3613 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3614 sc
.gfp_mask
, &nr_soft_scanned
);
3615 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3618 * There should be no need to raise the scanning priority if
3619 * enough pages are already being scanned that that high
3620 * watermark would be met at 100% efficiency.
3622 if (kswapd_shrink_node(pgdat
, &sc
))
3623 raise_priority
= false;
3626 * If the low watermark is met there is no need for processes
3627 * to be throttled on pfmemalloc_wait as they should not be
3628 * able to safely make forward progress. Wake them
3630 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3631 allow_direct_reclaim(pgdat
))
3632 wake_up_all(&pgdat
->pfmemalloc_wait
);
3634 /* Check if kswapd should be suspending */
3635 __fs_reclaim_release();
3636 ret
= try_to_freeze();
3637 __fs_reclaim_acquire();
3638 if (ret
|| kthread_should_stop())
3642 * Raise priority if scanning rate is too low or there was no
3643 * progress in reclaiming pages
3645 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3646 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
3649 * If reclaim made no progress for a boost, stop reclaim as
3650 * IO cannot be queued and it could be an infinite loop in
3651 * extreme circumstances.
3653 if (nr_boost_reclaim
&& !nr_reclaimed
)
3656 if (raise_priority
|| !nr_reclaimed
)
3658 } while (sc
.priority
>= 1);
3660 if (!sc
.nr_reclaimed
)
3661 pgdat
->kswapd_failures
++;
3664 /* If reclaim was boosted, account for the reclaim done in this pass */
3666 unsigned long flags
;
3668 for (i
= 0; i
<= classzone_idx
; i
++) {
3669 if (!zone_boosts
[i
])
3672 /* Increments are under the zone lock */
3673 zone
= pgdat
->node_zones
+ i
;
3674 spin_lock_irqsave(&zone
->lock
, flags
);
3675 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
3676 spin_unlock_irqrestore(&zone
->lock
, flags
);
3680 * As there is now likely space, wakeup kcompact to defragment
3683 wakeup_kcompactd(pgdat
, pageblock_order
, classzone_idx
);
3686 snapshot_refaults(NULL
, pgdat
);
3687 __fs_reclaim_release();
3688 psi_memstall_leave(&pflags
);
3689 set_task_reclaim_state(current
, NULL
);
3692 * Return the order kswapd stopped reclaiming at as
3693 * prepare_kswapd_sleep() takes it into account. If another caller
3694 * entered the allocator slow path while kswapd was awake, order will
3695 * remain at the higher level.
3701 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3702 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3703 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3704 * after previous reclaim attempt (node is still unbalanced). In that case
3705 * return the zone index of the previous kswapd reclaim cycle.
3707 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3708 enum zone_type prev_classzone_idx
)
3710 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3711 return prev_classzone_idx
;
3712 return pgdat
->kswapd_classzone_idx
;
3715 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3716 unsigned int classzone_idx
)
3721 if (freezing(current
) || kthread_should_stop())
3724 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3727 * Try to sleep for a short interval. Note that kcompactd will only be
3728 * woken if it is possible to sleep for a short interval. This is
3729 * deliberate on the assumption that if reclaim cannot keep an
3730 * eligible zone balanced that it's also unlikely that compaction will
3733 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3735 * Compaction records what page blocks it recently failed to
3736 * isolate pages from and skips them in the future scanning.
3737 * When kswapd is going to sleep, it is reasonable to assume
3738 * that pages and compaction may succeed so reset the cache.
3740 reset_isolation_suitable(pgdat
);
3743 * We have freed the memory, now we should compact it to make
3744 * allocation of the requested order possible.
3746 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3748 remaining
= schedule_timeout(HZ
/10);
3751 * If woken prematurely then reset kswapd_classzone_idx and
3752 * order. The values will either be from a wakeup request or
3753 * the previous request that slept prematurely.
3756 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3757 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3760 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3761 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3765 * After a short sleep, check if it was a premature sleep. If not, then
3766 * go fully to sleep until explicitly woken up.
3769 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3770 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3773 * vmstat counters are not perfectly accurate and the estimated
3774 * value for counters such as NR_FREE_PAGES can deviate from the
3775 * true value by nr_online_cpus * threshold. To avoid the zone
3776 * watermarks being breached while under pressure, we reduce the
3777 * per-cpu vmstat threshold while kswapd is awake and restore
3778 * them before going back to sleep.
3780 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3782 if (!kthread_should_stop())
3785 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3788 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3790 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3792 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3796 * The background pageout daemon, started as a kernel thread
3797 * from the init process.
3799 * This basically trickles out pages so that we have _some_
3800 * free memory available even if there is no other activity
3801 * that frees anything up. This is needed for things like routing
3802 * etc, where we otherwise might have all activity going on in
3803 * asynchronous contexts that cannot page things out.
3805 * If there are applications that are active memory-allocators
3806 * (most normal use), this basically shouldn't matter.
3808 static int kswapd(void *p
)
3810 unsigned int alloc_order
, reclaim_order
;
3811 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3812 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3813 struct task_struct
*tsk
= current
;
3814 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3816 if (!cpumask_empty(cpumask
))
3817 set_cpus_allowed_ptr(tsk
, cpumask
);
3820 * Tell the memory management that we're a "memory allocator",
3821 * and that if we need more memory we should get access to it
3822 * regardless (see "__alloc_pages()"). "kswapd" should
3823 * never get caught in the normal page freeing logic.
3825 * (Kswapd normally doesn't need memory anyway, but sometimes
3826 * you need a small amount of memory in order to be able to
3827 * page out something else, and this flag essentially protects
3828 * us from recursively trying to free more memory as we're
3829 * trying to free the first piece of memory in the first place).
3831 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3834 pgdat
->kswapd_order
= 0;
3835 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3839 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3840 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3843 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3846 /* Read the new order and classzone_idx */
3847 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3848 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3849 pgdat
->kswapd_order
= 0;
3850 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3852 ret
= try_to_freeze();
3853 if (kthread_should_stop())
3857 * We can speed up thawing tasks if we don't call balance_pgdat
3858 * after returning from the refrigerator
3864 * Reclaim begins at the requested order but if a high-order
3865 * reclaim fails then kswapd falls back to reclaiming for
3866 * order-0. If that happens, kswapd will consider sleeping
3867 * for the order it finished reclaiming at (reclaim_order)
3868 * but kcompactd is woken to compact for the original
3869 * request (alloc_order).
3871 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3873 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3874 if (reclaim_order
< alloc_order
)
3875 goto kswapd_try_sleep
;
3878 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3884 * A zone is low on free memory or too fragmented for high-order memory. If
3885 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3886 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3887 * has failed or is not needed, still wake up kcompactd if only compaction is
3890 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3891 enum zone_type classzone_idx
)
3895 if (!managed_zone(zone
))
3898 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3900 pgdat
= zone
->zone_pgdat
;
3902 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3903 pgdat
->kswapd_classzone_idx
= classzone_idx
;
3905 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
,
3907 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3908 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3911 /* Hopeless node, leave it to direct reclaim if possible */
3912 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3913 (pgdat_balanced(pgdat
, order
, classzone_idx
) &&
3914 !pgdat_watermark_boosted(pgdat
, classzone_idx
))) {
3916 * There may be plenty of free memory available, but it's too
3917 * fragmented for high-order allocations. Wake up kcompactd
3918 * and rely on compaction_suitable() to determine if it's
3919 * needed. If it fails, it will defer subsequent attempts to
3920 * ratelimit its work.
3922 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3923 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3927 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
3929 wake_up_interruptible(&pgdat
->kswapd_wait
);
3932 #ifdef CONFIG_HIBERNATION
3934 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3937 * Rather than trying to age LRUs the aim is to preserve the overall
3938 * LRU order by reclaiming preferentially
3939 * inactive > active > active referenced > active mapped
3941 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3943 struct scan_control sc
= {
3944 .nr_to_reclaim
= nr_to_reclaim
,
3945 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3946 .reclaim_idx
= MAX_NR_ZONES
- 1,
3947 .priority
= DEF_PRIORITY
,
3951 .hibernation_mode
= 1,
3953 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3954 unsigned long nr_reclaimed
;
3955 unsigned int noreclaim_flag
;
3957 fs_reclaim_acquire(sc
.gfp_mask
);
3958 noreclaim_flag
= memalloc_noreclaim_save();
3959 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3961 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3963 set_task_reclaim_state(current
, NULL
);
3964 memalloc_noreclaim_restore(noreclaim_flag
);
3965 fs_reclaim_release(sc
.gfp_mask
);
3967 return nr_reclaimed
;
3969 #endif /* CONFIG_HIBERNATION */
3971 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3972 not required for correctness. So if the last cpu in a node goes
3973 away, we get changed to run anywhere: as the first one comes back,
3974 restore their cpu bindings. */
3975 static int kswapd_cpu_online(unsigned int cpu
)
3979 for_each_node_state(nid
, N_MEMORY
) {
3980 pg_data_t
*pgdat
= NODE_DATA(nid
);
3981 const struct cpumask
*mask
;
3983 mask
= cpumask_of_node(pgdat
->node_id
);
3985 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3986 /* One of our CPUs online: restore mask */
3987 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3993 * This kswapd start function will be called by init and node-hot-add.
3994 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3996 int kswapd_run(int nid
)
3998 pg_data_t
*pgdat
= NODE_DATA(nid
);
4004 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4005 if (IS_ERR(pgdat
->kswapd
)) {
4006 /* failure at boot is fatal */
4007 BUG_ON(system_state
< SYSTEM_RUNNING
);
4008 pr_err("Failed to start kswapd on node %d\n", nid
);
4009 ret
= PTR_ERR(pgdat
->kswapd
);
4010 pgdat
->kswapd
= NULL
;
4016 * Called by memory hotplug when all memory in a node is offlined. Caller must
4017 * hold mem_hotplug_begin/end().
4019 void kswapd_stop(int nid
)
4021 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4024 kthread_stop(kswapd
);
4025 NODE_DATA(nid
)->kswapd
= NULL
;
4029 static int __init
kswapd_init(void)
4034 for_each_node_state(nid
, N_MEMORY
)
4036 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
4037 "mm/vmscan:online", kswapd_cpu_online
,
4043 module_init(kswapd_init
)
4049 * If non-zero call node_reclaim when the number of free pages falls below
4052 int node_reclaim_mode __read_mostly
;
4054 #define RECLAIM_OFF 0
4055 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4056 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4057 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4060 * Priority for NODE_RECLAIM. This determines the fraction of pages
4061 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4064 #define NODE_RECLAIM_PRIORITY 4
4067 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4070 int sysctl_min_unmapped_ratio
= 1;
4073 * If the number of slab pages in a zone grows beyond this percentage then
4074 * slab reclaim needs to occur.
4076 int sysctl_min_slab_ratio
= 5;
4078 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4080 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4081 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4082 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4085 * It's possible for there to be more file mapped pages than
4086 * accounted for by the pages on the file LRU lists because
4087 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4089 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4092 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4093 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4095 unsigned long nr_pagecache_reclaimable
;
4096 unsigned long delta
= 0;
4099 * If RECLAIM_UNMAP is set, then all file pages are considered
4100 * potentially reclaimable. Otherwise, we have to worry about
4101 * pages like swapcache and node_unmapped_file_pages() provides
4104 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4105 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4107 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4109 /* If we can't clean pages, remove dirty pages from consideration */
4110 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4111 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4113 /* Watch for any possible underflows due to delta */
4114 if (unlikely(delta
> nr_pagecache_reclaimable
))
4115 delta
= nr_pagecache_reclaimable
;
4117 return nr_pagecache_reclaimable
- delta
;
4121 * Try to free up some pages from this node through reclaim.
4123 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4125 /* Minimum pages needed in order to stay on node */
4126 const unsigned long nr_pages
= 1 << order
;
4127 struct task_struct
*p
= current
;
4128 unsigned int noreclaim_flag
;
4129 struct scan_control sc
= {
4130 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4131 .gfp_mask
= current_gfp_context(gfp_mask
),
4133 .priority
= NODE_RECLAIM_PRIORITY
,
4134 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4135 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4137 .reclaim_idx
= gfp_zone(gfp_mask
),
4140 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4144 fs_reclaim_acquire(sc
.gfp_mask
);
4146 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4147 * and we also need to be able to write out pages for RECLAIM_WRITE
4148 * and RECLAIM_UNMAP.
4150 noreclaim_flag
= memalloc_noreclaim_save();
4151 p
->flags
|= PF_SWAPWRITE
;
4152 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4154 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4156 * Free memory by calling shrink node with increasing
4157 * priorities until we have enough memory freed.
4160 shrink_node(pgdat
, &sc
);
4161 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4164 set_task_reclaim_state(p
, NULL
);
4165 current
->flags
&= ~PF_SWAPWRITE
;
4166 memalloc_noreclaim_restore(noreclaim_flag
);
4167 fs_reclaim_release(sc
.gfp_mask
);
4169 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4171 return sc
.nr_reclaimed
>= nr_pages
;
4174 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4179 * Node reclaim reclaims unmapped file backed pages and
4180 * slab pages if we are over the defined limits.
4182 * A small portion of unmapped file backed pages is needed for
4183 * file I/O otherwise pages read by file I/O will be immediately
4184 * thrown out if the node is overallocated. So we do not reclaim
4185 * if less than a specified percentage of the node is used by
4186 * unmapped file backed pages.
4188 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4189 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4190 return NODE_RECLAIM_FULL
;
4193 * Do not scan if the allocation should not be delayed.
4195 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4196 return NODE_RECLAIM_NOSCAN
;
4199 * Only run node reclaim on the local node or on nodes that do not
4200 * have associated processors. This will favor the local processor
4201 * over remote processors and spread off node memory allocations
4202 * as wide as possible.
4204 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4205 return NODE_RECLAIM_NOSCAN
;
4207 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4208 return NODE_RECLAIM_NOSCAN
;
4210 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4211 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4214 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4221 * page_evictable - test whether a page is evictable
4222 * @page: the page to test
4224 * Test whether page is evictable--i.e., should be placed on active/inactive
4225 * lists vs unevictable list.
4227 * Reasons page might not be evictable:
4228 * (1) page's mapping marked unevictable
4229 * (2) page is part of an mlocked VMA
4232 int page_evictable(struct page
*page
)
4236 /* Prevent address_space of inode and swap cache from being freed */
4238 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4244 * check_move_unevictable_pages - check pages for evictability and move to
4245 * appropriate zone lru list
4246 * @pvec: pagevec with lru pages to check
4248 * Checks pages for evictability, if an evictable page is in the unevictable
4249 * lru list, moves it to the appropriate evictable lru list. This function
4250 * should be only used for lru pages.
4252 void check_move_unevictable_pages(struct pagevec
*pvec
)
4254 struct lruvec
*lruvec
;
4255 struct pglist_data
*pgdat
= NULL
;
4260 for (i
= 0; i
< pvec
->nr
; i
++) {
4261 struct page
*page
= pvec
->pages
[i
];
4262 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4265 if (pagepgdat
!= pgdat
) {
4267 spin_unlock_irq(&pgdat
->lru_lock
);
4269 spin_lock_irq(&pgdat
->lru_lock
);
4271 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4273 if (!PageLRU(page
) || !PageUnevictable(page
))
4276 if (page_evictable(page
)) {
4277 enum lru_list lru
= page_lru_base_type(page
);
4279 VM_BUG_ON_PAGE(PageActive(page
), page
);
4280 ClearPageUnevictable(page
);
4281 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4282 add_page_to_lru_list(page
, lruvec
, lru
);
4288 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4289 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4290 spin_unlock_irq(&pgdat
->lru_lock
);
4293 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);