1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/pagevec.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
51 #include <linux/psi.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim
;
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
78 struct mem_cgroup
*target_mem_cgroup
;
81 * Scan pressure balancing between anon and file LRUs
83 unsigned long anon_cost
;
84 unsigned long file_cost
;
86 /* Can active pages be deactivated as part of reclaim? */
87 #define DEACTIVATE_ANON 1
88 #define DEACTIVATE_FILE 2
89 unsigned int may_deactivate
:2;
90 unsigned int force_deactivate
:1;
91 unsigned int skipped_deactivate
:1;
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage
:1;
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap
:1;
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap
:1;
103 * Cgroups are not reclaimed below their configured memory.low,
104 * unless we threaten to OOM. If any cgroups are skipped due to
105 * memory.low and nothing was reclaimed, go back for memory.low.
107 unsigned int memcg_low_reclaim
:1;
108 unsigned int memcg_low_skipped
:1;
110 unsigned int hibernation_mode
:1;
112 /* One of the zones is ready for compaction */
113 unsigned int compaction_ready
:1;
115 /* There is easily reclaimable cold cache in the current node */
116 unsigned int cache_trim_mode
:1;
118 /* The file pages on the current node are dangerously low */
119 unsigned int file_is_tiny
:1;
121 /* Allocation order */
124 /* Scan (total_size >> priority) pages at once */
127 /* The highest zone to isolate pages for reclaim from */
130 /* This context's GFP mask */
133 /* Incremented by the number of inactive pages that were scanned */
134 unsigned long nr_scanned
;
136 /* Number of pages freed so far during a call to shrink_zones() */
137 unsigned long nr_reclaimed
;
141 unsigned int unqueued_dirty
;
142 unsigned int congested
;
143 unsigned int writeback
;
144 unsigned int immediate
;
145 unsigned int file_taken
;
149 /* for recording the reclaimed slab by now */
150 struct reclaim_state reclaim_state
;
153 #ifdef ARCH_HAS_PREFETCHW
154 #define prefetchw_prev_lru_page(_page, _base, _field) \
156 if ((_page)->lru.prev != _base) { \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
164 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168 * From 0 .. 200. Higher means more swappy.
170 int vm_swappiness
= 60;
172 static void set_task_reclaim_state(struct task_struct
*task
,
173 struct reclaim_state
*rs
)
175 /* Check for an overwrite */
176 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
178 /* Check for the nulling of an already-nulled member */
179 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
181 task
->reclaim_state
= rs
;
184 static LIST_HEAD(shrinker_list
);
185 static DECLARE_RWSEM(shrinker_rwsem
);
189 * We allow subsystems to populate their shrinker-related
190 * LRU lists before register_shrinker_prepared() is called
191 * for the shrinker, since we don't want to impose
192 * restrictions on their internal registration order.
193 * In this case shrink_slab_memcg() may find corresponding
194 * bit is set in the shrinkers map.
196 * This value is used by the function to detect registering
197 * shrinkers and to skip do_shrink_slab() calls for them.
199 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
201 static DEFINE_IDR(shrinker_idr
);
202 static int shrinker_nr_max
;
204 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
206 int id
, ret
= -ENOMEM
;
208 down_write(&shrinker_rwsem
);
209 /* This may call shrinker, so it must use down_read_trylock() */
210 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
214 if (id
>= shrinker_nr_max
) {
215 if (memcg_expand_shrinker_maps(id
)) {
216 idr_remove(&shrinker_idr
, id
);
220 shrinker_nr_max
= id
+ 1;
225 up_write(&shrinker_rwsem
);
229 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
231 int id
= shrinker
->id
;
235 down_write(&shrinker_rwsem
);
236 idr_remove(&shrinker_idr
, id
);
237 up_write(&shrinker_rwsem
);
240 static bool cgroup_reclaim(struct scan_control
*sc
)
242 return sc
->target_mem_cgroup
;
246 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
247 * @sc: scan_control in question
249 * The normal page dirty throttling mechanism in balance_dirty_pages() is
250 * completely broken with the legacy memcg and direct stalling in
251 * shrink_page_list() is used for throttling instead, which lacks all the
252 * niceties such as fairness, adaptive pausing, bandwidth proportional
253 * allocation and configurability.
255 * This function tests whether the vmscan currently in progress can assume
256 * that the normal dirty throttling mechanism is operational.
258 static bool writeback_throttling_sane(struct scan_control
*sc
)
260 if (!cgroup_reclaim(sc
))
262 #ifdef CONFIG_CGROUP_WRITEBACK
263 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
269 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
274 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
278 static bool cgroup_reclaim(struct scan_control
*sc
)
283 static bool writeback_throttling_sane(struct scan_control
*sc
)
290 * This misses isolated pages which are not accounted for to save counters.
291 * As the data only determines if reclaim or compaction continues, it is
292 * not expected that isolated pages will be a dominating factor.
294 unsigned long zone_reclaimable_pages(struct zone
*zone
)
298 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
299 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
300 if (get_nr_swap_pages() > 0)
301 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
302 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
308 * lruvec_lru_size - Returns the number of pages on the given LRU list.
309 * @lruvec: lru vector
311 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
313 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
315 unsigned long size
= 0;
318 for (zid
= 0; zid
<= zone_idx
&& zid
< MAX_NR_ZONES
; zid
++) {
319 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
321 if (!managed_zone(zone
))
324 if (!mem_cgroup_disabled())
325 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
327 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
333 * Add a shrinker callback to be called from the vm.
335 int prealloc_shrinker(struct shrinker
*shrinker
)
337 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
339 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
342 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
343 if (!shrinker
->nr_deferred
)
346 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
347 if (prealloc_memcg_shrinker(shrinker
))
354 kfree(shrinker
->nr_deferred
);
355 shrinker
->nr_deferred
= NULL
;
359 void free_prealloced_shrinker(struct shrinker
*shrinker
)
361 if (!shrinker
->nr_deferred
)
364 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
365 unregister_memcg_shrinker(shrinker
);
367 kfree(shrinker
->nr_deferred
);
368 shrinker
->nr_deferred
= NULL
;
371 void register_shrinker_prepared(struct shrinker
*shrinker
)
373 down_write(&shrinker_rwsem
);
374 list_add_tail(&shrinker
->list
, &shrinker_list
);
376 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
377 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
379 up_write(&shrinker_rwsem
);
382 int register_shrinker(struct shrinker
*shrinker
)
384 int err
= prealloc_shrinker(shrinker
);
388 register_shrinker_prepared(shrinker
);
391 EXPORT_SYMBOL(register_shrinker
);
396 void unregister_shrinker(struct shrinker
*shrinker
)
398 if (!shrinker
->nr_deferred
)
400 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
401 unregister_memcg_shrinker(shrinker
);
402 down_write(&shrinker_rwsem
);
403 list_del(&shrinker
->list
);
404 up_write(&shrinker_rwsem
);
405 kfree(shrinker
->nr_deferred
);
406 shrinker
->nr_deferred
= NULL
;
408 EXPORT_SYMBOL(unregister_shrinker
);
410 #define SHRINK_BATCH 128
412 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
413 struct shrinker
*shrinker
, int priority
)
415 unsigned long freed
= 0;
416 unsigned long long delta
;
421 int nid
= shrinkctl
->nid
;
422 long batch_size
= shrinker
->batch
? shrinker
->batch
424 long scanned
= 0, next_deferred
;
426 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
429 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
430 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
434 * copy the current shrinker scan count into a local variable
435 * and zero it so that other concurrent shrinker invocations
436 * don't also do this scanning work.
438 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
441 if (shrinker
->seeks
) {
442 delta
= freeable
>> priority
;
444 do_div(delta
, shrinker
->seeks
);
447 * These objects don't require any IO to create. Trim
448 * them aggressively under memory pressure to keep
449 * them from causing refetches in the IO caches.
451 delta
= freeable
/ 2;
455 if (total_scan
< 0) {
456 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
457 shrinker
->scan_objects
, total_scan
);
458 total_scan
= freeable
;
461 next_deferred
= total_scan
;
464 * We need to avoid excessive windup on filesystem shrinkers
465 * due to large numbers of GFP_NOFS allocations causing the
466 * shrinkers to return -1 all the time. This results in a large
467 * nr being built up so when a shrink that can do some work
468 * comes along it empties the entire cache due to nr >>>
469 * freeable. This is bad for sustaining a working set in
472 * Hence only allow the shrinker to scan the entire cache when
473 * a large delta change is calculated directly.
475 if (delta
< freeable
/ 4)
476 total_scan
= min(total_scan
, freeable
/ 2);
479 * Avoid risking looping forever due to too large nr value:
480 * never try to free more than twice the estimate number of
483 if (total_scan
> freeable
* 2)
484 total_scan
= freeable
* 2;
486 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
487 freeable
, delta
, total_scan
, priority
);
490 * Normally, we should not scan less than batch_size objects in one
491 * pass to avoid too frequent shrinker calls, but if the slab has less
492 * than batch_size objects in total and we are really tight on memory,
493 * we will try to reclaim all available objects, otherwise we can end
494 * up failing allocations although there are plenty of reclaimable
495 * objects spread over several slabs with usage less than the
498 * We detect the "tight on memory" situations by looking at the total
499 * number of objects we want to scan (total_scan). If it is greater
500 * than the total number of objects on slab (freeable), we must be
501 * scanning at high prio and therefore should try to reclaim as much as
504 while (total_scan
>= batch_size
||
505 total_scan
>= freeable
) {
507 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
509 shrinkctl
->nr_to_scan
= nr_to_scan
;
510 shrinkctl
->nr_scanned
= nr_to_scan
;
511 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
512 if (ret
== SHRINK_STOP
)
516 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
517 total_scan
-= shrinkctl
->nr_scanned
;
518 scanned
+= shrinkctl
->nr_scanned
;
523 if (next_deferred
>= scanned
)
524 next_deferred
-= scanned
;
528 * move the unused scan count back into the shrinker in a
529 * manner that handles concurrent updates. If we exhausted the
530 * scan, there is no need to do an update.
532 if (next_deferred
> 0)
533 new_nr
= atomic_long_add_return(next_deferred
,
534 &shrinker
->nr_deferred
[nid
]);
536 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
538 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
543 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
544 struct mem_cgroup
*memcg
, int priority
)
546 struct memcg_shrinker_map
*map
;
547 unsigned long ret
, freed
= 0;
550 if (!mem_cgroup_online(memcg
))
553 if (!down_read_trylock(&shrinker_rwsem
))
556 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
561 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
562 struct shrink_control sc
= {
563 .gfp_mask
= gfp_mask
,
567 struct shrinker
*shrinker
;
569 shrinker
= idr_find(&shrinker_idr
, i
);
570 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
572 clear_bit(i
, map
->map
);
576 /* Call non-slab shrinkers even though kmem is disabled */
577 if (!memcg_kmem_enabled() &&
578 !(shrinker
->flags
& SHRINKER_NONSLAB
))
581 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
582 if (ret
== SHRINK_EMPTY
) {
583 clear_bit(i
, map
->map
);
585 * After the shrinker reported that it had no objects to
586 * free, but before we cleared the corresponding bit in
587 * the memcg shrinker map, a new object might have been
588 * added. To make sure, we have the bit set in this
589 * case, we invoke the shrinker one more time and reset
590 * the bit if it reports that it is not empty anymore.
591 * The memory barrier here pairs with the barrier in
592 * memcg_set_shrinker_bit():
594 * list_lru_add() shrink_slab_memcg()
595 * list_add_tail() clear_bit()
597 * set_bit() do_shrink_slab()
599 smp_mb__after_atomic();
600 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
601 if (ret
== SHRINK_EMPTY
)
604 memcg_set_shrinker_bit(memcg
, nid
, i
);
608 if (rwsem_is_contended(&shrinker_rwsem
)) {
614 up_read(&shrinker_rwsem
);
617 #else /* CONFIG_MEMCG */
618 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
619 struct mem_cgroup
*memcg
, int priority
)
623 #endif /* CONFIG_MEMCG */
626 * shrink_slab - shrink slab caches
627 * @gfp_mask: allocation context
628 * @nid: node whose slab caches to target
629 * @memcg: memory cgroup whose slab caches to target
630 * @priority: the reclaim priority
632 * Call the shrink functions to age shrinkable caches.
634 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
635 * unaware shrinkers will receive a node id of 0 instead.
637 * @memcg specifies the memory cgroup to target. Unaware shrinkers
638 * are called only if it is the root cgroup.
640 * @priority is sc->priority, we take the number of objects and >> by priority
641 * in order to get the scan target.
643 * Returns the number of reclaimed slab objects.
645 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
646 struct mem_cgroup
*memcg
,
649 unsigned long ret
, freed
= 0;
650 struct shrinker
*shrinker
;
653 * The root memcg might be allocated even though memcg is disabled
654 * via "cgroup_disable=memory" boot parameter. This could make
655 * mem_cgroup_is_root() return false, then just run memcg slab
656 * shrink, but skip global shrink. This may result in premature
659 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
660 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
662 if (!down_read_trylock(&shrinker_rwsem
))
665 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
666 struct shrink_control sc
= {
667 .gfp_mask
= gfp_mask
,
672 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
673 if (ret
== SHRINK_EMPTY
)
677 * Bail out if someone want to register a new shrinker to
678 * prevent the registration from being stalled for long periods
679 * by parallel ongoing shrinking.
681 if (rwsem_is_contended(&shrinker_rwsem
)) {
687 up_read(&shrinker_rwsem
);
693 void drop_slab_node(int nid
)
698 struct mem_cgroup
*memcg
= NULL
;
700 if (fatal_signal_pending(current
))
704 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
706 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
707 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
708 } while (freed
> 10);
715 for_each_online_node(nid
)
719 static inline int is_page_cache_freeable(struct page
*page
)
722 * A freeable page cache page is referenced only by the caller
723 * that isolated the page, the page cache and optional buffer
724 * heads at page->private.
726 int page_cache_pins
= thp_nr_pages(page
);
727 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
730 static int may_write_to_inode(struct inode
*inode
)
732 if (current
->flags
& PF_SWAPWRITE
)
734 if (!inode_write_congested(inode
))
736 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
742 * We detected a synchronous write error writing a page out. Probably
743 * -ENOSPC. We need to propagate that into the address_space for a subsequent
744 * fsync(), msync() or close().
746 * The tricky part is that after writepage we cannot touch the mapping: nothing
747 * prevents it from being freed up. But we have a ref on the page and once
748 * that page is locked, the mapping is pinned.
750 * We're allowed to run sleeping lock_page() here because we know the caller has
753 static void handle_write_error(struct address_space
*mapping
,
754 struct page
*page
, int error
)
757 if (page_mapping(page
) == mapping
)
758 mapping_set_error(mapping
, error
);
762 /* possible outcome of pageout() */
764 /* failed to write page out, page is locked */
766 /* move page to the active list, page is locked */
768 /* page has been sent to the disk successfully, page is unlocked */
770 /* page is clean and locked */
775 * pageout is called by shrink_page_list() for each dirty page.
776 * Calls ->writepage().
778 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
781 * If the page is dirty, only perform writeback if that write
782 * will be non-blocking. To prevent this allocation from being
783 * stalled by pagecache activity. But note that there may be
784 * stalls if we need to run get_block(). We could test
785 * PagePrivate for that.
787 * If this process is currently in __generic_file_write_iter() against
788 * this page's queue, we can perform writeback even if that
791 * If the page is swapcache, write it back even if that would
792 * block, for some throttling. This happens by accident, because
793 * swap_backing_dev_info is bust: it doesn't reflect the
794 * congestion state of the swapdevs. Easy to fix, if needed.
796 if (!is_page_cache_freeable(page
))
800 * Some data journaling orphaned pages can have
801 * page->mapping == NULL while being dirty with clean buffers.
803 if (page_has_private(page
)) {
804 if (try_to_free_buffers(page
)) {
805 ClearPageDirty(page
);
806 pr_info("%s: orphaned page\n", __func__
);
812 if (mapping
->a_ops
->writepage
== NULL
)
813 return PAGE_ACTIVATE
;
814 if (!may_write_to_inode(mapping
->host
))
817 if (clear_page_dirty_for_io(page
)) {
819 struct writeback_control wbc
= {
820 .sync_mode
= WB_SYNC_NONE
,
821 .nr_to_write
= SWAP_CLUSTER_MAX
,
823 .range_end
= LLONG_MAX
,
827 SetPageReclaim(page
);
828 res
= mapping
->a_ops
->writepage(page
, &wbc
);
830 handle_write_error(mapping
, page
, res
);
831 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
832 ClearPageReclaim(page
);
833 return PAGE_ACTIVATE
;
836 if (!PageWriteback(page
)) {
837 /* synchronous write or broken a_ops? */
838 ClearPageReclaim(page
);
840 trace_mm_vmscan_writepage(page
);
841 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
849 * Same as remove_mapping, but if the page is removed from the mapping, it
850 * gets returned with a refcount of 0.
852 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
853 bool reclaimed
, struct mem_cgroup
*target_memcg
)
859 BUG_ON(!PageLocked(page
));
860 BUG_ON(mapping
!= page_mapping(page
));
862 xa_lock_irqsave(&mapping
->i_pages
, flags
);
864 * The non racy check for a busy page.
866 * Must be careful with the order of the tests. When someone has
867 * a ref to the page, it may be possible that they dirty it then
868 * drop the reference. So if PageDirty is tested before page_count
869 * here, then the following race may occur:
871 * get_user_pages(&page);
872 * [user mapping goes away]
874 * !PageDirty(page) [good]
875 * SetPageDirty(page);
877 * !page_count(page) [good, discard it]
879 * [oops, our write_to data is lost]
881 * Reversing the order of the tests ensures such a situation cannot
882 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
883 * load is not satisfied before that of page->_refcount.
885 * Note that if SetPageDirty is always performed via set_page_dirty,
886 * and thus under the i_pages lock, then this ordering is not required.
888 refcount
= 1 + compound_nr(page
);
889 if (!page_ref_freeze(page
, refcount
))
891 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
892 if (unlikely(PageDirty(page
))) {
893 page_ref_unfreeze(page
, refcount
);
897 if (PageSwapCache(page
)) {
898 swp_entry_t swap
= { .val
= page_private(page
) };
899 mem_cgroup_swapout(page
, swap
);
900 if (reclaimed
&& !mapping_exiting(mapping
))
901 shadow
= workingset_eviction(page
, target_memcg
);
902 __delete_from_swap_cache(page
, swap
, shadow
);
903 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
904 put_swap_page(page
, swap
);
906 void (*freepage
)(struct page
*);
908 freepage
= mapping
->a_ops
->freepage
;
910 * Remember a shadow entry for reclaimed file cache in
911 * order to detect refaults, thus thrashing, later on.
913 * But don't store shadows in an address space that is
914 * already exiting. This is not just an optimization,
915 * inode reclaim needs to empty out the radix tree or
916 * the nodes are lost. Don't plant shadows behind its
919 * We also don't store shadows for DAX mappings because the
920 * only page cache pages found in these are zero pages
921 * covering holes, and because we don't want to mix DAX
922 * exceptional entries and shadow exceptional entries in the
923 * same address_space.
925 if (reclaimed
&& page_is_file_lru(page
) &&
926 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
927 shadow
= workingset_eviction(page
, target_memcg
);
928 __delete_from_page_cache(page
, shadow
);
929 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
931 if (freepage
!= NULL
)
938 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
943 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
944 * someone else has a ref on the page, abort and return 0. If it was
945 * successfully detached, return 1. Assumes the caller has a single ref on
948 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
950 if (__remove_mapping(mapping
, page
, false, NULL
)) {
952 * Unfreezing the refcount with 1 rather than 2 effectively
953 * drops the pagecache ref for us without requiring another
956 page_ref_unfreeze(page
, 1);
963 * putback_lru_page - put previously isolated page onto appropriate LRU list
964 * @page: page to be put back to appropriate lru list
966 * Add previously isolated @page to appropriate LRU list.
967 * Page may still be unevictable for other reasons.
969 * lru_lock must not be held, interrupts must be enabled.
971 void putback_lru_page(struct page
*page
)
974 put_page(page
); /* drop ref from isolate */
977 enum page_references
{
979 PAGEREF_RECLAIM_CLEAN
,
984 static enum page_references
page_check_references(struct page
*page
,
985 struct scan_control
*sc
)
987 int referenced_ptes
, referenced_page
;
988 unsigned long vm_flags
;
990 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
992 referenced_page
= TestClearPageReferenced(page
);
995 * Mlock lost the isolation race with us. Let try_to_unmap()
996 * move the page to the unevictable list.
998 if (vm_flags
& VM_LOCKED
)
999 return PAGEREF_RECLAIM
;
1001 if (referenced_ptes
) {
1003 * All mapped pages start out with page table
1004 * references from the instantiating fault, so we need
1005 * to look twice if a mapped file page is used more
1008 * Mark it and spare it for another trip around the
1009 * inactive list. Another page table reference will
1010 * lead to its activation.
1012 * Note: the mark is set for activated pages as well
1013 * so that recently deactivated but used pages are
1014 * quickly recovered.
1016 SetPageReferenced(page
);
1018 if (referenced_page
|| referenced_ptes
> 1)
1019 return PAGEREF_ACTIVATE
;
1022 * Activate file-backed executable pages after first usage.
1024 if ((vm_flags
& VM_EXEC
) && !PageSwapBacked(page
))
1025 return PAGEREF_ACTIVATE
;
1027 return PAGEREF_KEEP
;
1030 /* Reclaim if clean, defer dirty pages to writeback */
1031 if (referenced_page
&& !PageSwapBacked(page
))
1032 return PAGEREF_RECLAIM_CLEAN
;
1034 return PAGEREF_RECLAIM
;
1037 /* Check if a page is dirty or under writeback */
1038 static void page_check_dirty_writeback(struct page
*page
,
1039 bool *dirty
, bool *writeback
)
1041 struct address_space
*mapping
;
1044 * Anonymous pages are not handled by flushers and must be written
1045 * from reclaim context. Do not stall reclaim based on them
1047 if (!page_is_file_lru(page
) ||
1048 (PageAnon(page
) && !PageSwapBacked(page
))) {
1054 /* By default assume that the page flags are accurate */
1055 *dirty
= PageDirty(page
);
1056 *writeback
= PageWriteback(page
);
1058 /* Verify dirty/writeback state if the filesystem supports it */
1059 if (!page_has_private(page
))
1062 mapping
= page_mapping(page
);
1063 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1064 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1068 * shrink_page_list() returns the number of reclaimed pages
1070 static unsigned int shrink_page_list(struct list_head
*page_list
,
1071 struct pglist_data
*pgdat
,
1072 struct scan_control
*sc
,
1073 struct reclaim_stat
*stat
,
1074 bool ignore_references
)
1076 LIST_HEAD(ret_pages
);
1077 LIST_HEAD(free_pages
);
1078 unsigned int nr_reclaimed
= 0;
1079 unsigned int pgactivate
= 0;
1081 memset(stat
, 0, sizeof(*stat
));
1084 while (!list_empty(page_list
)) {
1085 struct address_space
*mapping
;
1087 enum page_references references
= PAGEREF_RECLAIM
;
1088 bool dirty
, writeback
, may_enter_fs
;
1089 unsigned int nr_pages
;
1093 page
= lru_to_page(page_list
);
1094 list_del(&page
->lru
);
1096 if (!trylock_page(page
))
1099 VM_BUG_ON_PAGE(PageActive(page
), page
);
1101 nr_pages
= compound_nr(page
);
1103 /* Account the number of base pages even though THP */
1104 sc
->nr_scanned
+= nr_pages
;
1106 if (unlikely(!page_evictable(page
)))
1107 goto activate_locked
;
1109 if (!sc
->may_unmap
&& page_mapped(page
))
1112 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1113 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1116 * The number of dirty pages determines if a node is marked
1117 * reclaim_congested which affects wait_iff_congested. kswapd
1118 * will stall and start writing pages if the tail of the LRU
1119 * is all dirty unqueued pages.
1121 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1122 if (dirty
|| writeback
)
1125 if (dirty
&& !writeback
)
1126 stat
->nr_unqueued_dirty
++;
1129 * Treat this page as congested if the underlying BDI is or if
1130 * pages are cycling through the LRU so quickly that the
1131 * pages marked for immediate reclaim are making it to the
1132 * end of the LRU a second time.
1134 mapping
= page_mapping(page
);
1135 if (((dirty
|| writeback
) && mapping
&&
1136 inode_write_congested(mapping
->host
)) ||
1137 (writeback
&& PageReclaim(page
)))
1138 stat
->nr_congested
++;
1141 * If a page at the tail of the LRU is under writeback, there
1142 * are three cases to consider.
1144 * 1) If reclaim is encountering an excessive number of pages
1145 * under writeback and this page is both under writeback and
1146 * PageReclaim then it indicates that pages are being queued
1147 * for IO but are being recycled through the LRU before the
1148 * IO can complete. Waiting on the page itself risks an
1149 * indefinite stall if it is impossible to writeback the
1150 * page due to IO error or disconnected storage so instead
1151 * note that the LRU is being scanned too quickly and the
1152 * caller can stall after page list has been processed.
1154 * 2) Global or new memcg reclaim encounters a page that is
1155 * not marked for immediate reclaim, or the caller does not
1156 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1157 * not to fs). In this case mark the page for immediate
1158 * reclaim and continue scanning.
1160 * Require may_enter_fs because we would wait on fs, which
1161 * may not have submitted IO yet. And the loop driver might
1162 * enter reclaim, and deadlock if it waits on a page for
1163 * which it is needed to do the write (loop masks off
1164 * __GFP_IO|__GFP_FS for this reason); but more thought
1165 * would probably show more reasons.
1167 * 3) Legacy memcg encounters a page that is already marked
1168 * PageReclaim. memcg does not have any dirty pages
1169 * throttling so we could easily OOM just because too many
1170 * pages are in writeback and there is nothing else to
1171 * reclaim. Wait for the writeback to complete.
1173 * In cases 1) and 2) we activate the pages to get them out of
1174 * the way while we continue scanning for clean pages on the
1175 * inactive list and refilling from the active list. The
1176 * observation here is that waiting for disk writes is more
1177 * expensive than potentially causing reloads down the line.
1178 * Since they're marked for immediate reclaim, they won't put
1179 * memory pressure on the cache working set any longer than it
1180 * takes to write them to disk.
1182 if (PageWriteback(page
)) {
1184 if (current_is_kswapd() &&
1185 PageReclaim(page
) &&
1186 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1187 stat
->nr_immediate
++;
1188 goto activate_locked
;
1191 } else if (writeback_throttling_sane(sc
) ||
1192 !PageReclaim(page
) || !may_enter_fs
) {
1194 * This is slightly racy - end_page_writeback()
1195 * might have just cleared PageReclaim, then
1196 * setting PageReclaim here end up interpreted
1197 * as PageReadahead - but that does not matter
1198 * enough to care. What we do want is for this
1199 * page to have PageReclaim set next time memcg
1200 * reclaim reaches the tests above, so it will
1201 * then wait_on_page_writeback() to avoid OOM;
1202 * and it's also appropriate in global reclaim.
1204 SetPageReclaim(page
);
1205 stat
->nr_writeback
++;
1206 goto activate_locked
;
1211 wait_on_page_writeback(page
);
1212 /* then go back and try same page again */
1213 list_add_tail(&page
->lru
, page_list
);
1218 if (!ignore_references
)
1219 references
= page_check_references(page
, sc
);
1221 switch (references
) {
1222 case PAGEREF_ACTIVATE
:
1223 goto activate_locked
;
1225 stat
->nr_ref_keep
+= nr_pages
;
1227 case PAGEREF_RECLAIM
:
1228 case PAGEREF_RECLAIM_CLEAN
:
1229 ; /* try to reclaim the page below */
1233 * Anonymous process memory has backing store?
1234 * Try to allocate it some swap space here.
1235 * Lazyfree page could be freed directly
1237 if (PageAnon(page
) && PageSwapBacked(page
)) {
1238 if (!PageSwapCache(page
)) {
1239 if (!(sc
->gfp_mask
& __GFP_IO
))
1241 if (page_maybe_dma_pinned(page
))
1243 if (PageTransHuge(page
)) {
1244 /* cannot split THP, skip it */
1245 if (!can_split_huge_page(page
, NULL
))
1246 goto activate_locked
;
1248 * Split pages without a PMD map right
1249 * away. Chances are some or all of the
1250 * tail pages can be freed without IO.
1252 if (!compound_mapcount(page
) &&
1253 split_huge_page_to_list(page
,
1255 goto activate_locked
;
1257 if (!add_to_swap(page
)) {
1258 if (!PageTransHuge(page
))
1259 goto activate_locked_split
;
1260 /* Fallback to swap normal pages */
1261 if (split_huge_page_to_list(page
,
1263 goto activate_locked
;
1264 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1265 count_vm_event(THP_SWPOUT_FALLBACK
);
1267 if (!add_to_swap(page
))
1268 goto activate_locked_split
;
1271 may_enter_fs
= true;
1273 /* Adding to swap updated mapping */
1274 mapping
= page_mapping(page
);
1276 } else if (unlikely(PageTransHuge(page
))) {
1277 /* Split file THP */
1278 if (split_huge_page_to_list(page
, page_list
))
1283 * THP may get split above, need minus tail pages and update
1284 * nr_pages to avoid accounting tail pages twice.
1286 * The tail pages that are added into swap cache successfully
1289 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1290 sc
->nr_scanned
-= (nr_pages
- 1);
1295 * The page is mapped into the page tables of one or more
1296 * processes. Try to unmap it here.
1298 if (page_mapped(page
)) {
1299 enum ttu_flags flags
= TTU_BATCH_FLUSH
;
1300 bool was_swapbacked
= PageSwapBacked(page
);
1302 if (unlikely(PageTransHuge(page
)))
1303 flags
|= TTU_SPLIT_HUGE_PMD
;
1305 if (!try_to_unmap(page
, flags
)) {
1306 stat
->nr_unmap_fail
+= nr_pages
;
1307 if (!was_swapbacked
&& PageSwapBacked(page
))
1308 stat
->nr_lazyfree_fail
+= nr_pages
;
1309 goto activate_locked
;
1313 if (PageDirty(page
)) {
1315 * Only kswapd can writeback filesystem pages
1316 * to avoid risk of stack overflow. But avoid
1317 * injecting inefficient single-page IO into
1318 * flusher writeback as much as possible: only
1319 * write pages when we've encountered many
1320 * dirty pages, and when we've already scanned
1321 * the rest of the LRU for clean pages and see
1322 * the same dirty pages again (PageReclaim).
1324 if (page_is_file_lru(page
) &&
1325 (!current_is_kswapd() || !PageReclaim(page
) ||
1326 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1328 * Immediately reclaim when written back.
1329 * Similar in principal to deactivate_page()
1330 * except we already have the page isolated
1331 * and know it's dirty
1333 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1334 SetPageReclaim(page
);
1336 goto activate_locked
;
1339 if (references
== PAGEREF_RECLAIM_CLEAN
)
1343 if (!sc
->may_writepage
)
1347 * Page is dirty. Flush the TLB if a writable entry
1348 * potentially exists to avoid CPU writes after IO
1349 * starts and then write it out here.
1351 try_to_unmap_flush_dirty();
1352 switch (pageout(page
, mapping
)) {
1356 goto activate_locked
;
1358 stat
->nr_pageout
+= thp_nr_pages(page
);
1360 if (PageWriteback(page
))
1362 if (PageDirty(page
))
1366 * A synchronous write - probably a ramdisk. Go
1367 * ahead and try to reclaim the page.
1369 if (!trylock_page(page
))
1371 if (PageDirty(page
) || PageWriteback(page
))
1373 mapping
= page_mapping(page
);
1376 ; /* try to free the page below */
1381 * If the page has buffers, try to free the buffer mappings
1382 * associated with this page. If we succeed we try to free
1385 * We do this even if the page is PageDirty().
1386 * try_to_release_page() does not perform I/O, but it is
1387 * possible for a page to have PageDirty set, but it is actually
1388 * clean (all its buffers are clean). This happens if the
1389 * buffers were written out directly, with submit_bh(). ext3
1390 * will do this, as well as the blockdev mapping.
1391 * try_to_release_page() will discover that cleanness and will
1392 * drop the buffers and mark the page clean - it can be freed.
1394 * Rarely, pages can have buffers and no ->mapping. These are
1395 * the pages which were not successfully invalidated in
1396 * truncate_cleanup_page(). We try to drop those buffers here
1397 * and if that worked, and the page is no longer mapped into
1398 * process address space (page_count == 1) it can be freed.
1399 * Otherwise, leave the page on the LRU so it is swappable.
1401 if (page_has_private(page
)) {
1402 if (!try_to_release_page(page
, sc
->gfp_mask
))
1403 goto activate_locked
;
1404 if (!mapping
&& page_count(page
) == 1) {
1406 if (put_page_testzero(page
))
1410 * rare race with speculative reference.
1411 * the speculative reference will free
1412 * this page shortly, so we may
1413 * increment nr_reclaimed here (and
1414 * leave it off the LRU).
1422 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1423 /* follow __remove_mapping for reference */
1424 if (!page_ref_freeze(page
, 1))
1426 if (PageDirty(page
)) {
1427 page_ref_unfreeze(page
, 1);
1431 count_vm_event(PGLAZYFREED
);
1432 count_memcg_page_event(page
, PGLAZYFREED
);
1433 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true,
1434 sc
->target_mem_cgroup
))
1440 * THP may get swapped out in a whole, need account
1443 nr_reclaimed
+= nr_pages
;
1446 * Is there need to periodically free_page_list? It would
1447 * appear not as the counts should be low
1449 if (unlikely(PageTransHuge(page
)))
1450 destroy_compound_page(page
);
1452 list_add(&page
->lru
, &free_pages
);
1455 activate_locked_split
:
1457 * The tail pages that are failed to add into swap cache
1458 * reach here. Fixup nr_scanned and nr_pages.
1461 sc
->nr_scanned
-= (nr_pages
- 1);
1465 /* Not a candidate for swapping, so reclaim swap space. */
1466 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1468 try_to_free_swap(page
);
1469 VM_BUG_ON_PAGE(PageActive(page
), page
);
1470 if (!PageMlocked(page
)) {
1471 int type
= page_is_file_lru(page
);
1472 SetPageActive(page
);
1473 stat
->nr_activate
[type
] += nr_pages
;
1474 count_memcg_page_event(page
, PGACTIVATE
);
1479 list_add(&page
->lru
, &ret_pages
);
1480 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1483 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1485 mem_cgroup_uncharge_list(&free_pages
);
1486 try_to_unmap_flush();
1487 free_unref_page_list(&free_pages
);
1489 list_splice(&ret_pages
, page_list
);
1490 count_vm_events(PGACTIVATE
, pgactivate
);
1492 return nr_reclaimed
;
1495 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1496 struct list_head
*page_list
)
1498 struct scan_control sc
= {
1499 .gfp_mask
= GFP_KERNEL
,
1500 .priority
= DEF_PRIORITY
,
1503 struct reclaim_stat stat
;
1504 unsigned int nr_reclaimed
;
1505 struct page
*page
, *next
;
1506 LIST_HEAD(clean_pages
);
1508 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1509 if (page_is_file_lru(page
) && !PageDirty(page
) &&
1510 !__PageMovable(page
) && !PageUnevictable(page
)) {
1511 ClearPageActive(page
);
1512 list_move(&page
->lru
, &clean_pages
);
1516 nr_reclaimed
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1518 list_splice(&clean_pages
, page_list
);
1519 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1520 -(long)nr_reclaimed
);
1522 * Since lazyfree pages are isolated from file LRU from the beginning,
1523 * they will rotate back to anonymous LRU in the end if it failed to
1524 * discard so isolated count will be mismatched.
1525 * Compensate the isolated count for both LRU lists.
1527 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
1528 stat
.nr_lazyfree_fail
);
1529 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1530 -(long)stat
.nr_lazyfree_fail
);
1531 return nr_reclaimed
;
1535 * Attempt to remove the specified page from its LRU. Only take this page
1536 * if it is of the appropriate PageActive status. Pages which are being
1537 * freed elsewhere are also ignored.
1539 * page: page to consider
1540 * mode: one of the LRU isolation modes defined above
1542 * returns 0 on success, -ve errno on failure.
1544 int __isolate_lru_page_prepare(struct page
*page
, isolate_mode_t mode
)
1548 /* Only take pages on the LRU. */
1552 /* Compaction should not handle unevictable pages but CMA can do so */
1553 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1557 * To minimise LRU disruption, the caller can indicate that it only
1558 * wants to isolate pages it will be able to operate on without
1559 * blocking - clean pages for the most part.
1561 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1562 * that it is possible to migrate without blocking
1564 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1565 /* All the caller can do on PageWriteback is block */
1566 if (PageWriteback(page
))
1569 if (PageDirty(page
)) {
1570 struct address_space
*mapping
;
1574 * Only pages without mappings or that have a
1575 * ->migratepage callback are possible to migrate
1576 * without blocking. However, we can be racing with
1577 * truncation so it's necessary to lock the page
1578 * to stabilise the mapping as truncation holds
1579 * the page lock until after the page is removed
1580 * from the page cache.
1582 if (!trylock_page(page
))
1585 mapping
= page_mapping(page
);
1586 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1593 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1600 * Update LRU sizes after isolating pages. The LRU size updates must
1601 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1603 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1604 enum lru_list lru
, unsigned long *nr_zone_taken
)
1608 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1609 if (!nr_zone_taken
[zid
])
1612 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1618 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1620 * lruvec->lru_lock is heavily contended. Some of the functions that
1621 * shrink the lists perform better by taking out a batch of pages
1622 * and working on them outside the LRU lock.
1624 * For pagecache intensive workloads, this function is the hottest
1625 * spot in the kernel (apart from copy_*_user functions).
1627 * Lru_lock must be held before calling this function.
1629 * @nr_to_scan: The number of eligible pages to look through on the list.
1630 * @lruvec: The LRU vector to pull pages from.
1631 * @dst: The temp list to put pages on to.
1632 * @nr_scanned: The number of pages that were scanned.
1633 * @sc: The scan_control struct for this reclaim session
1634 * @lru: LRU list id for isolating
1636 * returns how many pages were moved onto *@dst.
1638 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1639 struct lruvec
*lruvec
, struct list_head
*dst
,
1640 unsigned long *nr_scanned
, struct scan_control
*sc
,
1643 struct list_head
*src
= &lruvec
->lists
[lru
];
1644 unsigned long nr_taken
= 0;
1645 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1646 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1647 unsigned long skipped
= 0;
1648 unsigned long scan
, total_scan
, nr_pages
;
1649 LIST_HEAD(pages_skipped
);
1650 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1654 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1657 page
= lru_to_page(src
);
1658 prefetchw_prev_lru_page(page
, src
, flags
);
1660 nr_pages
= compound_nr(page
);
1661 total_scan
+= nr_pages
;
1663 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1664 list_move(&page
->lru
, &pages_skipped
);
1665 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1670 * Do not count skipped pages because that makes the function
1671 * return with no isolated pages if the LRU mostly contains
1672 * ineligible pages. This causes the VM to not reclaim any
1673 * pages, triggering a premature OOM.
1675 * Account all tail pages of THP. This would not cause
1676 * premature OOM since __isolate_lru_page() returns -EBUSY
1677 * only when the page is being freed somewhere else.
1680 switch (__isolate_lru_page_prepare(page
, mode
)) {
1683 * Be careful not to clear PageLRU until after we're
1684 * sure the page is not being freed elsewhere -- the
1685 * page release code relies on it.
1687 if (unlikely(!get_page_unless_zero(page
)))
1690 if (!TestClearPageLRU(page
)) {
1692 * This page may in other isolation path,
1693 * but we still hold lru_lock.
1699 nr_taken
+= nr_pages
;
1700 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1701 list_move(&page
->lru
, dst
);
1706 /* else it is being freed elsewhere */
1707 list_move(&page
->lru
, src
);
1712 * Splice any skipped pages to the start of the LRU list. Note that
1713 * this disrupts the LRU order when reclaiming for lower zones but
1714 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1715 * scanning would soon rescan the same pages to skip and put the
1716 * system at risk of premature OOM.
1718 if (!list_empty(&pages_skipped
)) {
1721 list_splice(&pages_skipped
, src
);
1722 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1723 if (!nr_skipped
[zid
])
1726 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1727 skipped
+= nr_skipped
[zid
];
1730 *nr_scanned
= total_scan
;
1731 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1732 total_scan
, skipped
, nr_taken
, mode
, lru
);
1733 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1738 * isolate_lru_page - tries to isolate a page from its LRU list
1739 * @page: page to isolate from its LRU list
1741 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1742 * vmstat statistic corresponding to whatever LRU list the page was on.
1744 * Returns 0 if the page was removed from an LRU list.
1745 * Returns -EBUSY if the page was not on an LRU list.
1747 * The returned page will have PageLRU() cleared. If it was found on
1748 * the active list, it will have PageActive set. If it was found on
1749 * the unevictable list, it will have the PageUnevictable bit set. That flag
1750 * may need to be cleared by the caller before letting the page go.
1752 * The vmstat statistic corresponding to the list on which the page was
1753 * found will be decremented.
1757 * (1) Must be called with an elevated refcount on the page. This is a
1758 * fundamental difference from isolate_lru_pages (which is called
1759 * without a stable reference).
1760 * (2) the lru_lock must not be held.
1761 * (3) interrupts must be enabled.
1763 int isolate_lru_page(struct page
*page
)
1767 VM_BUG_ON_PAGE(!page_count(page
), page
);
1768 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1770 if (TestClearPageLRU(page
)) {
1771 struct lruvec
*lruvec
;
1774 lruvec
= lock_page_lruvec_irq(page
);
1775 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
1776 unlock_page_lruvec_irq(lruvec
);
1784 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1785 * then get rescheduled. When there are massive number of tasks doing page
1786 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1787 * the LRU list will go small and be scanned faster than necessary, leading to
1788 * unnecessary swapping, thrashing and OOM.
1790 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1791 struct scan_control
*sc
)
1793 unsigned long inactive
, isolated
;
1795 if (current_is_kswapd())
1798 if (!writeback_throttling_sane(sc
))
1802 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1803 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1805 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1806 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1810 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1811 * won't get blocked by normal direct-reclaimers, forming a circular
1814 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1817 return isolated
> inactive
;
1821 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
1822 * On return, @list is reused as a list of pages to be freed by the caller.
1824 * Returns the number of pages moved to the given lruvec.
1826 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1827 struct list_head
*list
)
1829 int nr_pages
, nr_moved
= 0;
1830 LIST_HEAD(pages_to_free
);
1834 while (!list_empty(list
)) {
1835 page
= lru_to_page(list
);
1836 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1837 list_del(&page
->lru
);
1838 if (unlikely(!page_evictable(page
))) {
1839 spin_unlock_irq(&lruvec
->lru_lock
);
1840 putback_lru_page(page
);
1841 spin_lock_irq(&lruvec
->lru_lock
);
1846 * The SetPageLRU needs to be kept here for list integrity.
1848 * #0 move_pages_to_lru #1 release_pages
1849 * if !put_page_testzero
1850 * if (put_page_testzero())
1851 * !PageLRU //skip lru_lock
1853 * list_add(&page->lru,)
1854 * list_add(&page->lru,)
1858 if (unlikely(put_page_testzero(page
))) {
1859 __ClearPageLRU(page
);
1860 __ClearPageActive(page
);
1862 if (unlikely(PageCompound(page
))) {
1863 spin_unlock_irq(&lruvec
->lru_lock
);
1864 destroy_compound_page(page
);
1865 spin_lock_irq(&lruvec
->lru_lock
);
1867 list_add(&page
->lru
, &pages_to_free
);
1873 * All pages were isolated from the same lruvec (and isolation
1874 * inhibits memcg migration).
1876 VM_BUG_ON_PAGE(!lruvec_holds_page_lru_lock(page
, lruvec
), page
);
1877 lru
= page_lru(page
);
1878 nr_pages
= thp_nr_pages(page
);
1880 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1881 list_add(&page
->lru
, &lruvec
->lists
[lru
]);
1882 nr_moved
+= nr_pages
;
1883 if (PageActive(page
))
1884 workingset_age_nonresident(lruvec
, nr_pages
);
1888 * To save our caller's stack, now use input list for pages to free.
1890 list_splice(&pages_to_free
, list
);
1896 * If a kernel thread (such as nfsd for loop-back mounts) services
1897 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1898 * In that case we should only throttle if the backing device it is
1899 * writing to is congested. In other cases it is safe to throttle.
1901 static int current_may_throttle(void)
1903 return !(current
->flags
& PF_LOCAL_THROTTLE
) ||
1904 current
->backing_dev_info
== NULL
||
1905 bdi_write_congested(current
->backing_dev_info
);
1909 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1910 * of reclaimed pages
1912 static noinline_for_stack
unsigned long
1913 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1914 struct scan_control
*sc
, enum lru_list lru
)
1916 LIST_HEAD(page_list
);
1917 unsigned long nr_scanned
;
1918 unsigned int nr_reclaimed
= 0;
1919 unsigned long nr_taken
;
1920 struct reclaim_stat stat
;
1921 bool file
= is_file_lru(lru
);
1922 enum vm_event_item item
;
1923 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1924 bool stalled
= false;
1926 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1930 /* wait a bit for the reclaimer. */
1934 /* We are about to die and free our memory. Return now. */
1935 if (fatal_signal_pending(current
))
1936 return SWAP_CLUSTER_MAX
;
1941 spin_lock_irq(&lruvec
->lru_lock
);
1943 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1944 &nr_scanned
, sc
, lru
);
1946 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1947 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1948 if (!cgroup_reclaim(sc
))
1949 __count_vm_events(item
, nr_scanned
);
1950 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1951 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
1953 spin_unlock_irq(&lruvec
->lru_lock
);
1958 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, &stat
, false);
1960 spin_lock_irq(&lruvec
->lru_lock
);
1961 move_pages_to_lru(lruvec
, &page_list
);
1963 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1964 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
1965 if (!cgroup_reclaim(sc
))
1966 __count_vm_events(item
, nr_reclaimed
);
1967 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
1968 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
1969 spin_unlock_irq(&lruvec
->lru_lock
);
1971 lru_note_cost(lruvec
, file
, stat
.nr_pageout
);
1972 mem_cgroup_uncharge_list(&page_list
);
1973 free_unref_page_list(&page_list
);
1976 * If dirty pages are scanned that are not queued for IO, it
1977 * implies that flushers are not doing their job. This can
1978 * happen when memory pressure pushes dirty pages to the end of
1979 * the LRU before the dirty limits are breached and the dirty
1980 * data has expired. It can also happen when the proportion of
1981 * dirty pages grows not through writes but through memory
1982 * pressure reclaiming all the clean cache. And in some cases,
1983 * the flushers simply cannot keep up with the allocation
1984 * rate. Nudge the flusher threads in case they are asleep.
1986 if (stat
.nr_unqueued_dirty
== nr_taken
)
1987 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1989 sc
->nr
.dirty
+= stat
.nr_dirty
;
1990 sc
->nr
.congested
+= stat
.nr_congested
;
1991 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1992 sc
->nr
.writeback
+= stat
.nr_writeback
;
1993 sc
->nr
.immediate
+= stat
.nr_immediate
;
1994 sc
->nr
.taken
+= nr_taken
;
1996 sc
->nr
.file_taken
+= nr_taken
;
1998 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1999 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2000 return nr_reclaimed
;
2004 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2006 * We move them the other way if the page is referenced by one or more
2009 * If the pages are mostly unmapped, the processing is fast and it is
2010 * appropriate to hold lru_lock across the whole operation. But if
2011 * the pages are mapped, the processing is slow (page_referenced()), so
2012 * we should drop lru_lock around each page. It's impossible to balance
2013 * this, so instead we remove the pages from the LRU while processing them.
2014 * It is safe to rely on PG_active against the non-LRU pages in here because
2015 * nobody will play with that bit on a non-LRU page.
2017 * The downside is that we have to touch page->_refcount against each page.
2018 * But we had to alter page->flags anyway.
2020 static void shrink_active_list(unsigned long nr_to_scan
,
2021 struct lruvec
*lruvec
,
2022 struct scan_control
*sc
,
2025 unsigned long nr_taken
;
2026 unsigned long nr_scanned
;
2027 unsigned long vm_flags
;
2028 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2029 LIST_HEAD(l_active
);
2030 LIST_HEAD(l_inactive
);
2032 unsigned nr_deactivate
, nr_activate
;
2033 unsigned nr_rotated
= 0;
2034 int file
= is_file_lru(lru
);
2035 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2039 spin_lock_irq(&lruvec
->lru_lock
);
2041 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2042 &nr_scanned
, sc
, lru
);
2044 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2046 if (!cgroup_reclaim(sc
))
2047 __count_vm_events(PGREFILL
, nr_scanned
);
2048 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2050 spin_unlock_irq(&lruvec
->lru_lock
);
2052 while (!list_empty(&l_hold
)) {
2054 page
= lru_to_page(&l_hold
);
2055 list_del(&page
->lru
);
2057 if (unlikely(!page_evictable(page
))) {
2058 putback_lru_page(page
);
2062 if (unlikely(buffer_heads_over_limit
)) {
2063 if (page_has_private(page
) && trylock_page(page
)) {
2064 if (page_has_private(page
))
2065 try_to_release_page(page
, 0);
2070 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2073 * Identify referenced, file-backed active pages and
2074 * give them one more trip around the active list. So
2075 * that executable code get better chances to stay in
2076 * memory under moderate memory pressure. Anon pages
2077 * are not likely to be evicted by use-once streaming
2078 * IO, plus JVM can create lots of anon VM_EXEC pages,
2079 * so we ignore them here.
2081 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2082 nr_rotated
+= thp_nr_pages(page
);
2083 list_add(&page
->lru
, &l_active
);
2088 ClearPageActive(page
); /* we are de-activating */
2089 SetPageWorkingset(page
);
2090 list_add(&page
->lru
, &l_inactive
);
2094 * Move pages back to the lru list.
2096 spin_lock_irq(&lruvec
->lru_lock
);
2098 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2099 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2100 /* Keep all free pages in l_active list */
2101 list_splice(&l_inactive
, &l_active
);
2103 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2104 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2106 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2107 spin_unlock_irq(&lruvec
->lru_lock
);
2109 mem_cgroup_uncharge_list(&l_active
);
2110 free_unref_page_list(&l_active
);
2111 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2112 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2115 unsigned long reclaim_pages(struct list_head
*page_list
)
2117 int nid
= NUMA_NO_NODE
;
2118 unsigned int nr_reclaimed
= 0;
2119 LIST_HEAD(node_page_list
);
2120 struct reclaim_stat dummy_stat
;
2122 struct scan_control sc
= {
2123 .gfp_mask
= GFP_KERNEL
,
2124 .priority
= DEF_PRIORITY
,
2130 while (!list_empty(page_list
)) {
2131 page
= lru_to_page(page_list
);
2132 if (nid
== NUMA_NO_NODE
) {
2133 nid
= page_to_nid(page
);
2134 INIT_LIST_HEAD(&node_page_list
);
2137 if (nid
== page_to_nid(page
)) {
2138 ClearPageActive(page
);
2139 list_move(&page
->lru
, &node_page_list
);
2143 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2145 &sc
, &dummy_stat
, false);
2146 while (!list_empty(&node_page_list
)) {
2147 page
= lru_to_page(&node_page_list
);
2148 list_del(&page
->lru
);
2149 putback_lru_page(page
);
2155 if (!list_empty(&node_page_list
)) {
2156 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2158 &sc
, &dummy_stat
, false);
2159 while (!list_empty(&node_page_list
)) {
2160 page
= lru_to_page(&node_page_list
);
2161 list_del(&page
->lru
);
2162 putback_lru_page(page
);
2166 return nr_reclaimed
;
2169 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2170 struct lruvec
*lruvec
, struct scan_control
*sc
)
2172 if (is_active_lru(lru
)) {
2173 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2174 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2176 sc
->skipped_deactivate
= 1;
2180 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2184 * The inactive anon list should be small enough that the VM never has
2185 * to do too much work.
2187 * The inactive file list should be small enough to leave most memory
2188 * to the established workingset on the scan-resistant active list,
2189 * but large enough to avoid thrashing the aggregate readahead window.
2191 * Both inactive lists should also be large enough that each inactive
2192 * page has a chance to be referenced again before it is reclaimed.
2194 * If that fails and refaulting is observed, the inactive list grows.
2196 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2197 * on this LRU, maintained by the pageout code. An inactive_ratio
2198 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2201 * memory ratio inactive
2202 * -------------------------------------
2211 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2213 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2214 unsigned long inactive
, active
;
2215 unsigned long inactive_ratio
;
2218 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2219 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2221 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2223 inactive_ratio
= int_sqrt(10 * gb
);
2227 return inactive
* inactive_ratio
< active
;
2238 * Determine how aggressively the anon and file LRU lists should be
2239 * scanned. The relative value of each set of LRU lists is determined
2240 * by looking at the fraction of the pages scanned we did rotate back
2241 * onto the active list instead of evict.
2243 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2244 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2246 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2249 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2250 unsigned long anon_cost
, file_cost
, total_cost
;
2251 int swappiness
= mem_cgroup_swappiness(memcg
);
2252 u64 fraction
[ANON_AND_FILE
];
2253 u64 denominator
= 0; /* gcc */
2254 enum scan_balance scan_balance
;
2255 unsigned long ap
, fp
;
2258 /* If we have no swap space, do not bother scanning anon pages. */
2259 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2260 scan_balance
= SCAN_FILE
;
2265 * Global reclaim will swap to prevent OOM even with no
2266 * swappiness, but memcg users want to use this knob to
2267 * disable swapping for individual groups completely when
2268 * using the memory controller's swap limit feature would be
2271 if (cgroup_reclaim(sc
) && !swappiness
) {
2272 scan_balance
= SCAN_FILE
;
2277 * Do not apply any pressure balancing cleverness when the
2278 * system is close to OOM, scan both anon and file equally
2279 * (unless the swappiness setting disagrees with swapping).
2281 if (!sc
->priority
&& swappiness
) {
2282 scan_balance
= SCAN_EQUAL
;
2287 * If the system is almost out of file pages, force-scan anon.
2289 if (sc
->file_is_tiny
) {
2290 scan_balance
= SCAN_ANON
;
2295 * If there is enough inactive page cache, we do not reclaim
2296 * anything from the anonymous working right now.
2298 if (sc
->cache_trim_mode
) {
2299 scan_balance
= SCAN_FILE
;
2303 scan_balance
= SCAN_FRACT
;
2305 * Calculate the pressure balance between anon and file pages.
2307 * The amount of pressure we put on each LRU is inversely
2308 * proportional to the cost of reclaiming each list, as
2309 * determined by the share of pages that are refaulting, times
2310 * the relative IO cost of bringing back a swapped out
2311 * anonymous page vs reloading a filesystem page (swappiness).
2313 * Although we limit that influence to ensure no list gets
2314 * left behind completely: at least a third of the pressure is
2315 * applied, before swappiness.
2317 * With swappiness at 100, anon and file have equal IO cost.
2319 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2320 anon_cost
= total_cost
+ sc
->anon_cost
;
2321 file_cost
= total_cost
+ sc
->file_cost
;
2322 total_cost
= anon_cost
+ file_cost
;
2324 ap
= swappiness
* (total_cost
+ 1);
2325 ap
/= anon_cost
+ 1;
2327 fp
= (200 - swappiness
) * (total_cost
+ 1);
2328 fp
/= file_cost
+ 1;
2332 denominator
= ap
+ fp
;
2334 for_each_evictable_lru(lru
) {
2335 int file
= is_file_lru(lru
);
2336 unsigned long lruvec_size
;
2338 unsigned long protection
;
2340 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2341 protection
= mem_cgroup_protection(sc
->target_mem_cgroup
,
2343 sc
->memcg_low_reclaim
);
2347 * Scale a cgroup's reclaim pressure by proportioning
2348 * its current usage to its memory.low or memory.min
2351 * This is important, as otherwise scanning aggression
2352 * becomes extremely binary -- from nothing as we
2353 * approach the memory protection threshold, to totally
2354 * nominal as we exceed it. This results in requiring
2355 * setting extremely liberal protection thresholds. It
2356 * also means we simply get no protection at all if we
2357 * set it too low, which is not ideal.
2359 * If there is any protection in place, we reduce scan
2360 * pressure by how much of the total memory used is
2361 * within protection thresholds.
2363 * There is one special case: in the first reclaim pass,
2364 * we skip over all groups that are within their low
2365 * protection. If that fails to reclaim enough pages to
2366 * satisfy the reclaim goal, we come back and override
2367 * the best-effort low protection. However, we still
2368 * ideally want to honor how well-behaved groups are in
2369 * that case instead of simply punishing them all
2370 * equally. As such, we reclaim them based on how much
2371 * memory they are using, reducing the scan pressure
2372 * again by how much of the total memory used is under
2375 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2377 /* Avoid TOCTOU with earlier protection check */
2378 cgroup_size
= max(cgroup_size
, protection
);
2380 scan
= lruvec_size
- lruvec_size
* protection
/
2384 * Minimally target SWAP_CLUSTER_MAX pages to keep
2385 * reclaim moving forwards, avoiding decrementing
2386 * sc->priority further than desirable.
2388 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2393 scan
>>= sc
->priority
;
2396 * If the cgroup's already been deleted, make sure to
2397 * scrape out the remaining cache.
2399 if (!scan
&& !mem_cgroup_online(memcg
))
2400 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2402 switch (scan_balance
) {
2404 /* Scan lists relative to size */
2408 * Scan types proportional to swappiness and
2409 * their relative recent reclaim efficiency.
2410 * Make sure we don't miss the last page on
2411 * the offlined memory cgroups because of a
2414 scan
= mem_cgroup_online(memcg
) ?
2415 div64_u64(scan
* fraction
[file
], denominator
) :
2416 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2421 /* Scan one type exclusively */
2422 if ((scan_balance
== SCAN_FILE
) != file
)
2426 /* Look ma, no brain */
2434 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2436 unsigned long nr
[NR_LRU_LISTS
];
2437 unsigned long targets
[NR_LRU_LISTS
];
2438 unsigned long nr_to_scan
;
2440 unsigned long nr_reclaimed
= 0;
2441 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2442 struct blk_plug plug
;
2445 get_scan_count(lruvec
, sc
, nr
);
2447 /* Record the original scan target for proportional adjustments later */
2448 memcpy(targets
, nr
, sizeof(nr
));
2451 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2452 * event that can occur when there is little memory pressure e.g.
2453 * multiple streaming readers/writers. Hence, we do not abort scanning
2454 * when the requested number of pages are reclaimed when scanning at
2455 * DEF_PRIORITY on the assumption that the fact we are direct
2456 * reclaiming implies that kswapd is not keeping up and it is best to
2457 * do a batch of work at once. For memcg reclaim one check is made to
2458 * abort proportional reclaim if either the file or anon lru has already
2459 * dropped to zero at the first pass.
2461 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2462 sc
->priority
== DEF_PRIORITY
);
2464 blk_start_plug(&plug
);
2465 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2466 nr
[LRU_INACTIVE_FILE
]) {
2467 unsigned long nr_anon
, nr_file
, percentage
;
2468 unsigned long nr_scanned
;
2470 for_each_evictable_lru(lru
) {
2472 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2473 nr
[lru
] -= nr_to_scan
;
2475 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2482 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2486 * For kswapd and memcg, reclaim at least the number of pages
2487 * requested. Ensure that the anon and file LRUs are scanned
2488 * proportionally what was requested by get_scan_count(). We
2489 * stop reclaiming one LRU and reduce the amount scanning
2490 * proportional to the original scan target.
2492 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2493 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2496 * It's just vindictive to attack the larger once the smaller
2497 * has gone to zero. And given the way we stop scanning the
2498 * smaller below, this makes sure that we only make one nudge
2499 * towards proportionality once we've got nr_to_reclaim.
2501 if (!nr_file
|| !nr_anon
)
2504 if (nr_file
> nr_anon
) {
2505 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2506 targets
[LRU_ACTIVE_ANON
] + 1;
2508 percentage
= nr_anon
* 100 / scan_target
;
2510 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2511 targets
[LRU_ACTIVE_FILE
] + 1;
2513 percentage
= nr_file
* 100 / scan_target
;
2516 /* Stop scanning the smaller of the LRU */
2518 nr
[lru
+ LRU_ACTIVE
] = 0;
2521 * Recalculate the other LRU scan count based on its original
2522 * scan target and the percentage scanning already complete
2524 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2525 nr_scanned
= targets
[lru
] - nr
[lru
];
2526 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2527 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2530 nr_scanned
= targets
[lru
] - nr
[lru
];
2531 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2532 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2534 scan_adjusted
= true;
2536 blk_finish_plug(&plug
);
2537 sc
->nr_reclaimed
+= nr_reclaimed
;
2540 * Even if we did not try to evict anon pages at all, we want to
2541 * rebalance the anon lru active/inactive ratio.
2543 if (total_swap_pages
&& inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
2544 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2545 sc
, LRU_ACTIVE_ANON
);
2548 /* Use reclaim/compaction for costly allocs or under memory pressure */
2549 static bool in_reclaim_compaction(struct scan_control
*sc
)
2551 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2552 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2553 sc
->priority
< DEF_PRIORITY
- 2))
2560 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2561 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2562 * true if more pages should be reclaimed such that when the page allocator
2563 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2564 * It will give up earlier than that if there is difficulty reclaiming pages.
2566 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2567 unsigned long nr_reclaimed
,
2568 struct scan_control
*sc
)
2570 unsigned long pages_for_compaction
;
2571 unsigned long inactive_lru_pages
;
2574 /* If not in reclaim/compaction mode, stop */
2575 if (!in_reclaim_compaction(sc
))
2579 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2580 * number of pages that were scanned. This will return to the caller
2581 * with the risk reclaim/compaction and the resulting allocation attempt
2582 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2583 * allocations through requiring that the full LRU list has been scanned
2584 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2585 * scan, but that approximation was wrong, and there were corner cases
2586 * where always a non-zero amount of pages were scanned.
2591 /* If compaction would go ahead or the allocation would succeed, stop */
2592 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2593 struct zone
*zone
= &pgdat
->node_zones
[z
];
2594 if (!managed_zone(zone
))
2597 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2598 case COMPACT_SUCCESS
:
2599 case COMPACT_CONTINUE
:
2602 /* check next zone */
2608 * If we have not reclaimed enough pages for compaction and the
2609 * inactive lists are large enough, continue reclaiming
2611 pages_for_compaction
= compact_gap(sc
->order
);
2612 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2613 if (get_nr_swap_pages() > 0)
2614 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2616 return inactive_lru_pages
> pages_for_compaction
;
2619 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
2621 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
2622 struct mem_cgroup
*memcg
;
2624 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
2626 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
2627 unsigned long reclaimed
;
2628 unsigned long scanned
;
2631 * This loop can become CPU-bound when target memcgs
2632 * aren't eligible for reclaim - either because they
2633 * don't have any reclaimable pages, or because their
2634 * memory is explicitly protected. Avoid soft lockups.
2638 mem_cgroup_calculate_protection(target_memcg
, memcg
);
2640 if (mem_cgroup_below_min(memcg
)) {
2643 * If there is no reclaimable memory, OOM.
2646 } else if (mem_cgroup_below_low(memcg
)) {
2649 * Respect the protection only as long as
2650 * there is an unprotected supply
2651 * of reclaimable memory from other cgroups.
2653 if (!sc
->memcg_low_reclaim
) {
2654 sc
->memcg_low_skipped
= 1;
2657 memcg_memory_event(memcg
, MEMCG_LOW
);
2660 reclaimed
= sc
->nr_reclaimed
;
2661 scanned
= sc
->nr_scanned
;
2663 shrink_lruvec(lruvec
, sc
);
2665 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2668 /* Record the group's reclaim efficiency */
2669 vmpressure(sc
->gfp_mask
, memcg
, false,
2670 sc
->nr_scanned
- scanned
,
2671 sc
->nr_reclaimed
- reclaimed
);
2673 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
2676 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2678 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2679 unsigned long nr_reclaimed
, nr_scanned
;
2680 struct lruvec
*target_lruvec
;
2681 bool reclaimable
= false;
2684 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
2687 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2689 nr_reclaimed
= sc
->nr_reclaimed
;
2690 nr_scanned
= sc
->nr_scanned
;
2693 * Determine the scan balance between anon and file LRUs.
2695 spin_lock_irq(&target_lruvec
->lru_lock
);
2696 sc
->anon_cost
= target_lruvec
->anon_cost
;
2697 sc
->file_cost
= target_lruvec
->file_cost
;
2698 spin_unlock_irq(&target_lruvec
->lru_lock
);
2701 * Target desirable inactive:active list ratios for the anon
2702 * and file LRU lists.
2704 if (!sc
->force_deactivate
) {
2705 unsigned long refaults
;
2707 refaults
= lruvec_page_state(target_lruvec
,
2708 WORKINGSET_ACTIVATE_ANON
);
2709 if (refaults
!= target_lruvec
->refaults
[0] ||
2710 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
2711 sc
->may_deactivate
|= DEACTIVATE_ANON
;
2713 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
2716 * When refaults are being observed, it means a new
2717 * workingset is being established. Deactivate to get
2718 * rid of any stale active pages quickly.
2720 refaults
= lruvec_page_state(target_lruvec
,
2721 WORKINGSET_ACTIVATE_FILE
);
2722 if (refaults
!= target_lruvec
->refaults
[1] ||
2723 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
2724 sc
->may_deactivate
|= DEACTIVATE_FILE
;
2726 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
2728 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
2731 * If we have plenty of inactive file pages that aren't
2732 * thrashing, try to reclaim those first before touching
2735 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
2736 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
2737 sc
->cache_trim_mode
= 1;
2739 sc
->cache_trim_mode
= 0;
2742 * Prevent the reclaimer from falling into the cache trap: as
2743 * cache pages start out inactive, every cache fault will tip
2744 * the scan balance towards the file LRU. And as the file LRU
2745 * shrinks, so does the window for rotation from references.
2746 * This means we have a runaway feedback loop where a tiny
2747 * thrashing file LRU becomes infinitely more attractive than
2748 * anon pages. Try to detect this based on file LRU size.
2750 if (!cgroup_reclaim(sc
)) {
2751 unsigned long total_high_wmark
= 0;
2752 unsigned long free
, anon
;
2755 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2756 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2757 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2759 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2760 struct zone
*zone
= &pgdat
->node_zones
[z
];
2761 if (!managed_zone(zone
))
2764 total_high_wmark
+= high_wmark_pages(zone
);
2768 * Consider anon: if that's low too, this isn't a
2769 * runaway file reclaim problem, but rather just
2770 * extreme pressure. Reclaim as per usual then.
2772 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2775 file
+ free
<= total_high_wmark
&&
2776 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
2777 anon
>> sc
->priority
;
2780 shrink_node_memcgs(pgdat
, sc
);
2782 if (reclaim_state
) {
2783 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2784 reclaim_state
->reclaimed_slab
= 0;
2787 /* Record the subtree's reclaim efficiency */
2788 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2789 sc
->nr_scanned
- nr_scanned
,
2790 sc
->nr_reclaimed
- nr_reclaimed
);
2792 if (sc
->nr_reclaimed
- nr_reclaimed
)
2795 if (current_is_kswapd()) {
2797 * If reclaim is isolating dirty pages under writeback,
2798 * it implies that the long-lived page allocation rate
2799 * is exceeding the page laundering rate. Either the
2800 * global limits are not being effective at throttling
2801 * processes due to the page distribution throughout
2802 * zones or there is heavy usage of a slow backing
2803 * device. The only option is to throttle from reclaim
2804 * context which is not ideal as there is no guarantee
2805 * the dirtying process is throttled in the same way
2806 * balance_dirty_pages() manages.
2808 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2809 * count the number of pages under pages flagged for
2810 * immediate reclaim and stall if any are encountered
2811 * in the nr_immediate check below.
2813 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2814 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2816 /* Allow kswapd to start writing pages during reclaim.*/
2817 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2818 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2821 * If kswapd scans pages marked for immediate
2822 * reclaim and under writeback (nr_immediate), it
2823 * implies that pages are cycling through the LRU
2824 * faster than they are written so also forcibly stall.
2826 if (sc
->nr
.immediate
)
2827 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2831 * Tag a node/memcg as congested if all the dirty pages
2832 * scanned were backed by a congested BDI and
2833 * wait_iff_congested will stall.
2835 * Legacy memcg will stall in page writeback so avoid forcibly
2836 * stalling in wait_iff_congested().
2838 if ((current_is_kswapd() ||
2839 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
2840 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2841 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
2844 * Stall direct reclaim for IO completions if underlying BDIs
2845 * and node is congested. Allow kswapd to continue until it
2846 * starts encountering unqueued dirty pages or cycling through
2847 * the LRU too quickly.
2849 if (!current_is_kswapd() && current_may_throttle() &&
2850 !sc
->hibernation_mode
&&
2851 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
2852 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2854 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2859 * Kswapd gives up on balancing particular nodes after too
2860 * many failures to reclaim anything from them and goes to
2861 * sleep. On reclaim progress, reset the failure counter. A
2862 * successful direct reclaim run will revive a dormant kswapd.
2865 pgdat
->kswapd_failures
= 0;
2869 * Returns true if compaction should go ahead for a costly-order request, or
2870 * the allocation would already succeed without compaction. Return false if we
2871 * should reclaim first.
2873 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2875 unsigned long watermark
;
2876 enum compact_result suitable
;
2878 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2879 if (suitable
== COMPACT_SUCCESS
)
2880 /* Allocation should succeed already. Don't reclaim. */
2882 if (suitable
== COMPACT_SKIPPED
)
2883 /* Compaction cannot yet proceed. Do reclaim. */
2887 * Compaction is already possible, but it takes time to run and there
2888 * are potentially other callers using the pages just freed. So proceed
2889 * with reclaim to make a buffer of free pages available to give
2890 * compaction a reasonable chance of completing and allocating the page.
2891 * Note that we won't actually reclaim the whole buffer in one attempt
2892 * as the target watermark in should_continue_reclaim() is lower. But if
2893 * we are already above the high+gap watermark, don't reclaim at all.
2895 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2897 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2901 * This is the direct reclaim path, for page-allocating processes. We only
2902 * try to reclaim pages from zones which will satisfy the caller's allocation
2905 * If a zone is deemed to be full of pinned pages then just give it a light
2906 * scan then give up on it.
2908 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2912 unsigned long nr_soft_reclaimed
;
2913 unsigned long nr_soft_scanned
;
2915 pg_data_t
*last_pgdat
= NULL
;
2918 * If the number of buffer_heads in the machine exceeds the maximum
2919 * allowed level, force direct reclaim to scan the highmem zone as
2920 * highmem pages could be pinning lowmem pages storing buffer_heads
2922 orig_mask
= sc
->gfp_mask
;
2923 if (buffer_heads_over_limit
) {
2924 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2925 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2928 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2929 sc
->reclaim_idx
, sc
->nodemask
) {
2931 * Take care memory controller reclaiming has small influence
2934 if (!cgroup_reclaim(sc
)) {
2935 if (!cpuset_zone_allowed(zone
,
2936 GFP_KERNEL
| __GFP_HARDWALL
))
2940 * If we already have plenty of memory free for
2941 * compaction in this zone, don't free any more.
2942 * Even though compaction is invoked for any
2943 * non-zero order, only frequent costly order
2944 * reclamation is disruptive enough to become a
2945 * noticeable problem, like transparent huge
2948 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2949 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2950 compaction_ready(zone
, sc
)) {
2951 sc
->compaction_ready
= true;
2956 * Shrink each node in the zonelist once. If the
2957 * zonelist is ordered by zone (not the default) then a
2958 * node may be shrunk multiple times but in that case
2959 * the user prefers lower zones being preserved.
2961 if (zone
->zone_pgdat
== last_pgdat
)
2965 * This steals pages from memory cgroups over softlimit
2966 * and returns the number of reclaimed pages and
2967 * scanned pages. This works for global memory pressure
2968 * and balancing, not for a memcg's limit.
2970 nr_soft_scanned
= 0;
2971 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2972 sc
->order
, sc
->gfp_mask
,
2974 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2975 sc
->nr_scanned
+= nr_soft_scanned
;
2976 /* need some check for avoid more shrink_zone() */
2979 /* See comment about same check for global reclaim above */
2980 if (zone
->zone_pgdat
== last_pgdat
)
2982 last_pgdat
= zone
->zone_pgdat
;
2983 shrink_node(zone
->zone_pgdat
, sc
);
2987 * Restore to original mask to avoid the impact on the caller if we
2988 * promoted it to __GFP_HIGHMEM.
2990 sc
->gfp_mask
= orig_mask
;
2993 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
2995 struct lruvec
*target_lruvec
;
2996 unsigned long refaults
;
2998 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
2999 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
3000 target_lruvec
->refaults
[0] = refaults
;
3001 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
3002 target_lruvec
->refaults
[1] = refaults
;
3006 * This is the main entry point to direct page reclaim.
3008 * If a full scan of the inactive list fails to free enough memory then we
3009 * are "out of memory" and something needs to be killed.
3011 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3012 * high - the zone may be full of dirty or under-writeback pages, which this
3013 * caller can't do much about. We kick the writeback threads and take explicit
3014 * naps in the hope that some of these pages can be written. But if the
3015 * allocating task holds filesystem locks which prevent writeout this might not
3016 * work, and the allocation attempt will fail.
3018 * returns: 0, if no pages reclaimed
3019 * else, the number of pages reclaimed
3021 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3022 struct scan_control
*sc
)
3024 int initial_priority
= sc
->priority
;
3025 pg_data_t
*last_pgdat
;
3029 delayacct_freepages_start();
3031 if (!cgroup_reclaim(sc
))
3032 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3035 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3038 shrink_zones(zonelist
, sc
);
3040 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3043 if (sc
->compaction_ready
)
3047 * If we're getting trouble reclaiming, start doing
3048 * writepage even in laptop mode.
3050 if (sc
->priority
< DEF_PRIORITY
- 2)
3051 sc
->may_writepage
= 1;
3052 } while (--sc
->priority
>= 0);
3055 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3057 if (zone
->zone_pgdat
== last_pgdat
)
3059 last_pgdat
= zone
->zone_pgdat
;
3061 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3063 if (cgroup_reclaim(sc
)) {
3064 struct lruvec
*lruvec
;
3066 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3068 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3072 delayacct_freepages_end();
3074 if (sc
->nr_reclaimed
)
3075 return sc
->nr_reclaimed
;
3077 /* Aborted reclaim to try compaction? don't OOM, then */
3078 if (sc
->compaction_ready
)
3082 * We make inactive:active ratio decisions based on the node's
3083 * composition of memory, but a restrictive reclaim_idx or a
3084 * memory.low cgroup setting can exempt large amounts of
3085 * memory from reclaim. Neither of which are very common, so
3086 * instead of doing costly eligibility calculations of the
3087 * entire cgroup subtree up front, we assume the estimates are
3088 * good, and retry with forcible deactivation if that fails.
3090 if (sc
->skipped_deactivate
) {
3091 sc
->priority
= initial_priority
;
3092 sc
->force_deactivate
= 1;
3093 sc
->skipped_deactivate
= 0;
3097 /* Untapped cgroup reserves? Don't OOM, retry. */
3098 if (sc
->memcg_low_skipped
) {
3099 sc
->priority
= initial_priority
;
3100 sc
->force_deactivate
= 0;
3101 sc
->memcg_low_reclaim
= 1;
3102 sc
->memcg_low_skipped
= 0;
3109 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3112 unsigned long pfmemalloc_reserve
= 0;
3113 unsigned long free_pages
= 0;
3117 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3120 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3121 zone
= &pgdat
->node_zones
[i
];
3122 if (!managed_zone(zone
))
3125 if (!zone_reclaimable_pages(zone
))
3128 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3129 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3132 /* If there are no reserves (unexpected config) then do not throttle */
3133 if (!pfmemalloc_reserve
)
3136 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3138 /* kswapd must be awake if processes are being throttled */
3139 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3140 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
3141 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
3143 wake_up_interruptible(&pgdat
->kswapd_wait
);
3150 * Throttle direct reclaimers if backing storage is backed by the network
3151 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3152 * depleted. kswapd will continue to make progress and wake the processes
3153 * when the low watermark is reached.
3155 * Returns true if a fatal signal was delivered during throttling. If this
3156 * happens, the page allocator should not consider triggering the OOM killer.
3158 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3159 nodemask_t
*nodemask
)
3163 pg_data_t
*pgdat
= NULL
;
3166 * Kernel threads should not be throttled as they may be indirectly
3167 * responsible for cleaning pages necessary for reclaim to make forward
3168 * progress. kjournald for example may enter direct reclaim while
3169 * committing a transaction where throttling it could forcing other
3170 * processes to block on log_wait_commit().
3172 if (current
->flags
& PF_KTHREAD
)
3176 * If a fatal signal is pending, this process should not throttle.
3177 * It should return quickly so it can exit and free its memory
3179 if (fatal_signal_pending(current
))
3183 * Check if the pfmemalloc reserves are ok by finding the first node
3184 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3185 * GFP_KERNEL will be required for allocating network buffers when
3186 * swapping over the network so ZONE_HIGHMEM is unusable.
3188 * Throttling is based on the first usable node and throttled processes
3189 * wait on a queue until kswapd makes progress and wakes them. There
3190 * is an affinity then between processes waking up and where reclaim
3191 * progress has been made assuming the process wakes on the same node.
3192 * More importantly, processes running on remote nodes will not compete
3193 * for remote pfmemalloc reserves and processes on different nodes
3194 * should make reasonable progress.
3196 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3197 gfp_zone(gfp_mask
), nodemask
) {
3198 if (zone_idx(zone
) > ZONE_NORMAL
)
3201 /* Throttle based on the first usable node */
3202 pgdat
= zone
->zone_pgdat
;
3203 if (allow_direct_reclaim(pgdat
))
3208 /* If no zone was usable by the allocation flags then do not throttle */
3212 /* Account for the throttling */
3213 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3216 * If the caller cannot enter the filesystem, it's possible that it
3217 * is due to the caller holding an FS lock or performing a journal
3218 * transaction in the case of a filesystem like ext[3|4]. In this case,
3219 * it is not safe to block on pfmemalloc_wait as kswapd could be
3220 * blocked waiting on the same lock. Instead, throttle for up to a
3221 * second before continuing.
3223 if (!(gfp_mask
& __GFP_FS
)) {
3224 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3225 allow_direct_reclaim(pgdat
), HZ
);
3230 /* Throttle until kswapd wakes the process */
3231 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3232 allow_direct_reclaim(pgdat
));
3235 if (fatal_signal_pending(current
))
3242 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3243 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3245 unsigned long nr_reclaimed
;
3246 struct scan_control sc
= {
3247 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3248 .gfp_mask
= current_gfp_context(gfp_mask
),
3249 .reclaim_idx
= gfp_zone(gfp_mask
),
3251 .nodemask
= nodemask
,
3252 .priority
= DEF_PRIORITY
,
3253 .may_writepage
= !laptop_mode
,
3259 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3260 * Confirm they are large enough for max values.
3262 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3263 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3264 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3267 * Do not enter reclaim if fatal signal was delivered while throttled.
3268 * 1 is returned so that the page allocator does not OOM kill at this
3271 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3274 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3275 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3277 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3279 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3280 set_task_reclaim_state(current
, NULL
);
3282 return nr_reclaimed
;
3287 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3288 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3289 gfp_t gfp_mask
, bool noswap
,
3291 unsigned long *nr_scanned
)
3293 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3294 struct scan_control sc
= {
3295 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3296 .target_mem_cgroup
= memcg
,
3297 .may_writepage
= !laptop_mode
,
3299 .reclaim_idx
= MAX_NR_ZONES
- 1,
3300 .may_swap
= !noswap
,
3303 WARN_ON_ONCE(!current
->reclaim_state
);
3305 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3306 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3308 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3312 * NOTE: Although we can get the priority field, using it
3313 * here is not a good idea, since it limits the pages we can scan.
3314 * if we don't reclaim here, the shrink_node from balance_pgdat
3315 * will pick up pages from other mem cgroup's as well. We hack
3316 * the priority and make it zero.
3318 shrink_lruvec(lruvec
, &sc
);
3320 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3322 *nr_scanned
= sc
.nr_scanned
;
3324 return sc
.nr_reclaimed
;
3327 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3328 unsigned long nr_pages
,
3332 unsigned long nr_reclaimed
;
3333 unsigned int noreclaim_flag
;
3334 struct scan_control sc
= {
3335 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3336 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3337 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3338 .reclaim_idx
= MAX_NR_ZONES
- 1,
3339 .target_mem_cgroup
= memcg
,
3340 .priority
= DEF_PRIORITY
,
3341 .may_writepage
= !laptop_mode
,
3343 .may_swap
= may_swap
,
3346 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3347 * equal pressure on all the nodes. This is based on the assumption that
3348 * the reclaim does not bail out early.
3350 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3352 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3353 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3354 noreclaim_flag
= memalloc_noreclaim_save();
3356 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3358 memalloc_noreclaim_restore(noreclaim_flag
);
3359 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3360 set_task_reclaim_state(current
, NULL
);
3362 return nr_reclaimed
;
3366 static void age_active_anon(struct pglist_data
*pgdat
,
3367 struct scan_control
*sc
)
3369 struct mem_cgroup
*memcg
;
3370 struct lruvec
*lruvec
;
3372 if (!total_swap_pages
)
3375 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3376 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3379 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3381 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3382 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3383 sc
, LRU_ACTIVE_ANON
);
3384 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3388 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
3394 * Check for watermark boosts top-down as the higher zones
3395 * are more likely to be boosted. Both watermarks and boosts
3396 * should not be checked at the same time as reclaim would
3397 * start prematurely when there is no boosting and a lower
3400 for (i
= highest_zoneidx
; i
>= 0; i
--) {
3401 zone
= pgdat
->node_zones
+ i
;
3402 if (!managed_zone(zone
))
3405 if (zone
->watermark_boost
)
3413 * Returns true if there is an eligible zone balanced for the request order
3414 * and highest_zoneidx
3416 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3419 unsigned long mark
= -1;
3423 * Check watermarks bottom-up as lower zones are more likely to
3426 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3427 zone
= pgdat
->node_zones
+ i
;
3429 if (!managed_zone(zone
))
3432 mark
= high_wmark_pages(zone
);
3433 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
3438 * If a node has no populated zone within highest_zoneidx, it does not
3439 * need balancing by definition. This can happen if a zone-restricted
3440 * allocation tries to wake a remote kswapd.
3448 /* Clear pgdat state for congested, dirty or under writeback. */
3449 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3451 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3453 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3454 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3455 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3459 * Prepare kswapd for sleeping. This verifies that there are no processes
3460 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3462 * Returns true if kswapd is ready to sleep
3464 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
3465 int highest_zoneidx
)
3468 * The throttled processes are normally woken up in balance_pgdat() as
3469 * soon as allow_direct_reclaim() is true. But there is a potential
3470 * race between when kswapd checks the watermarks and a process gets
3471 * throttled. There is also a potential race if processes get
3472 * throttled, kswapd wakes, a large process exits thereby balancing the
3473 * zones, which causes kswapd to exit balance_pgdat() before reaching
3474 * the wake up checks. If kswapd is going to sleep, no process should
3475 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3476 * the wake up is premature, processes will wake kswapd and get
3477 * throttled again. The difference from wake ups in balance_pgdat() is
3478 * that here we are under prepare_to_wait().
3480 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3481 wake_up_all(&pgdat
->pfmemalloc_wait
);
3483 /* Hopeless node, leave it to direct reclaim */
3484 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3487 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
3488 clear_pgdat_congested(pgdat
);
3496 * kswapd shrinks a node of pages that are at or below the highest usable
3497 * zone that is currently unbalanced.
3499 * Returns true if kswapd scanned at least the requested number of pages to
3500 * reclaim or if the lack of progress was due to pages under writeback.
3501 * This is used to determine if the scanning priority needs to be raised.
3503 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3504 struct scan_control
*sc
)
3509 /* Reclaim a number of pages proportional to the number of zones */
3510 sc
->nr_to_reclaim
= 0;
3511 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3512 zone
= pgdat
->node_zones
+ z
;
3513 if (!managed_zone(zone
))
3516 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3520 * Historically care was taken to put equal pressure on all zones but
3521 * now pressure is applied based on node LRU order.
3523 shrink_node(pgdat
, sc
);
3526 * Fragmentation may mean that the system cannot be rebalanced for
3527 * high-order allocations. If twice the allocation size has been
3528 * reclaimed then recheck watermarks only at order-0 to prevent
3529 * excessive reclaim. Assume that a process requested a high-order
3530 * can direct reclaim/compact.
3532 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3535 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3539 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3540 * that are eligible for use by the caller until at least one zone is
3543 * Returns the order kswapd finished reclaiming at.
3545 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3546 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3547 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3548 * or lower is eligible for reclaim until at least one usable zone is
3551 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
3554 unsigned long nr_soft_reclaimed
;
3555 unsigned long nr_soft_scanned
;
3556 unsigned long pflags
;
3557 unsigned long nr_boost_reclaim
;
3558 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
3561 struct scan_control sc
= {
3562 .gfp_mask
= GFP_KERNEL
,
3567 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3568 psi_memstall_enter(&pflags
);
3569 __fs_reclaim_acquire();
3571 count_vm_event(PAGEOUTRUN
);
3574 * Account for the reclaim boost. Note that the zone boost is left in
3575 * place so that parallel allocations that are near the watermark will
3576 * stall or direct reclaim until kswapd is finished.
3578 nr_boost_reclaim
= 0;
3579 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3580 zone
= pgdat
->node_zones
+ i
;
3581 if (!managed_zone(zone
))
3584 nr_boost_reclaim
+= zone
->watermark_boost
;
3585 zone_boosts
[i
] = zone
->watermark_boost
;
3587 boosted
= nr_boost_reclaim
;
3590 sc
.priority
= DEF_PRIORITY
;
3592 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3593 bool raise_priority
= true;
3597 sc
.reclaim_idx
= highest_zoneidx
;
3600 * If the number of buffer_heads exceeds the maximum allowed
3601 * then consider reclaiming from all zones. This has a dual
3602 * purpose -- on 64-bit systems it is expected that
3603 * buffer_heads are stripped during active rotation. On 32-bit
3604 * systems, highmem pages can pin lowmem memory and shrinking
3605 * buffers can relieve lowmem pressure. Reclaim may still not
3606 * go ahead if all eligible zones for the original allocation
3607 * request are balanced to avoid excessive reclaim from kswapd.
3609 if (buffer_heads_over_limit
) {
3610 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3611 zone
= pgdat
->node_zones
+ i
;
3612 if (!managed_zone(zone
))
3621 * If the pgdat is imbalanced then ignore boosting and preserve
3622 * the watermarks for a later time and restart. Note that the
3623 * zone watermarks will be still reset at the end of balancing
3624 * on the grounds that the normal reclaim should be enough to
3625 * re-evaluate if boosting is required when kswapd next wakes.
3627 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
3628 if (!balanced
&& nr_boost_reclaim
) {
3629 nr_boost_reclaim
= 0;
3634 * If boosting is not active then only reclaim if there are no
3635 * eligible zones. Note that sc.reclaim_idx is not used as
3636 * buffer_heads_over_limit may have adjusted it.
3638 if (!nr_boost_reclaim
&& balanced
)
3641 /* Limit the priority of boosting to avoid reclaim writeback */
3642 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
3643 raise_priority
= false;
3646 * Do not writeback or swap pages for boosted reclaim. The
3647 * intent is to relieve pressure not issue sub-optimal IO
3648 * from reclaim context. If no pages are reclaimed, the
3649 * reclaim will be aborted.
3651 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
3652 sc
.may_swap
= !nr_boost_reclaim
;
3655 * Do some background aging of the anon list, to give
3656 * pages a chance to be referenced before reclaiming. All
3657 * pages are rotated regardless of classzone as this is
3658 * about consistent aging.
3660 age_active_anon(pgdat
, &sc
);
3663 * If we're getting trouble reclaiming, start doing writepage
3664 * even in laptop mode.
3666 if (sc
.priority
< DEF_PRIORITY
- 2)
3667 sc
.may_writepage
= 1;
3669 /* Call soft limit reclaim before calling shrink_node. */
3671 nr_soft_scanned
= 0;
3672 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3673 sc
.gfp_mask
, &nr_soft_scanned
);
3674 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3677 * There should be no need to raise the scanning priority if
3678 * enough pages are already being scanned that that high
3679 * watermark would be met at 100% efficiency.
3681 if (kswapd_shrink_node(pgdat
, &sc
))
3682 raise_priority
= false;
3685 * If the low watermark is met there is no need for processes
3686 * to be throttled on pfmemalloc_wait as they should not be
3687 * able to safely make forward progress. Wake them
3689 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3690 allow_direct_reclaim(pgdat
))
3691 wake_up_all(&pgdat
->pfmemalloc_wait
);
3693 /* Check if kswapd should be suspending */
3694 __fs_reclaim_release();
3695 ret
= try_to_freeze();
3696 __fs_reclaim_acquire();
3697 if (ret
|| kthread_should_stop())
3701 * Raise priority if scanning rate is too low or there was no
3702 * progress in reclaiming pages
3704 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3705 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
3708 * If reclaim made no progress for a boost, stop reclaim as
3709 * IO cannot be queued and it could be an infinite loop in
3710 * extreme circumstances.
3712 if (nr_boost_reclaim
&& !nr_reclaimed
)
3715 if (raise_priority
|| !nr_reclaimed
)
3717 } while (sc
.priority
>= 1);
3719 if (!sc
.nr_reclaimed
)
3720 pgdat
->kswapd_failures
++;
3723 /* If reclaim was boosted, account for the reclaim done in this pass */
3725 unsigned long flags
;
3727 for (i
= 0; i
<= highest_zoneidx
; i
++) {
3728 if (!zone_boosts
[i
])
3731 /* Increments are under the zone lock */
3732 zone
= pgdat
->node_zones
+ i
;
3733 spin_lock_irqsave(&zone
->lock
, flags
);
3734 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
3735 spin_unlock_irqrestore(&zone
->lock
, flags
);
3739 * As there is now likely space, wakeup kcompact to defragment
3742 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
3745 snapshot_refaults(NULL
, pgdat
);
3746 __fs_reclaim_release();
3747 psi_memstall_leave(&pflags
);
3748 set_task_reclaim_state(current
, NULL
);
3751 * Return the order kswapd stopped reclaiming at as
3752 * prepare_kswapd_sleep() takes it into account. If another caller
3753 * entered the allocator slow path while kswapd was awake, order will
3754 * remain at the higher level.
3760 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3761 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3762 * not a valid index then either kswapd runs for first time or kswapd couldn't
3763 * sleep after previous reclaim attempt (node is still unbalanced). In that
3764 * case return the zone index of the previous kswapd reclaim cycle.
3766 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
3767 enum zone_type prev_highest_zoneidx
)
3769 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
3771 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
3774 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3775 unsigned int highest_zoneidx
)
3780 if (freezing(current
) || kthread_should_stop())
3783 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3786 * Try to sleep for a short interval. Note that kcompactd will only be
3787 * woken if it is possible to sleep for a short interval. This is
3788 * deliberate on the assumption that if reclaim cannot keep an
3789 * eligible zone balanced that it's also unlikely that compaction will
3792 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
3794 * Compaction records what page blocks it recently failed to
3795 * isolate pages from and skips them in the future scanning.
3796 * When kswapd is going to sleep, it is reasonable to assume
3797 * that pages and compaction may succeed so reset the cache.
3799 reset_isolation_suitable(pgdat
);
3802 * We have freed the memory, now we should compact it to make
3803 * allocation of the requested order possible.
3805 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
3807 remaining
= schedule_timeout(HZ
/10);
3810 * If woken prematurely then reset kswapd_highest_zoneidx and
3811 * order. The values will either be from a wakeup request or
3812 * the previous request that slept prematurely.
3815 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
3816 kswapd_highest_zoneidx(pgdat
,
3819 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
3820 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
3823 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3824 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3828 * After a short sleep, check if it was a premature sleep. If not, then
3829 * go fully to sleep until explicitly woken up.
3832 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
3833 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3836 * vmstat counters are not perfectly accurate and the estimated
3837 * value for counters such as NR_FREE_PAGES can deviate from the
3838 * true value by nr_online_cpus * threshold. To avoid the zone
3839 * watermarks being breached while under pressure, we reduce the
3840 * per-cpu vmstat threshold while kswapd is awake and restore
3841 * them before going back to sleep.
3843 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3845 if (!kthread_should_stop())
3848 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3851 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3853 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3855 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3859 * The background pageout daemon, started as a kernel thread
3860 * from the init process.
3862 * This basically trickles out pages so that we have _some_
3863 * free memory available even if there is no other activity
3864 * that frees anything up. This is needed for things like routing
3865 * etc, where we otherwise might have all activity going on in
3866 * asynchronous contexts that cannot page things out.
3868 * If there are applications that are active memory-allocators
3869 * (most normal use), this basically shouldn't matter.
3871 static int kswapd(void *p
)
3873 unsigned int alloc_order
, reclaim_order
;
3874 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
3875 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3876 struct task_struct
*tsk
= current
;
3877 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3879 if (!cpumask_empty(cpumask
))
3880 set_cpus_allowed_ptr(tsk
, cpumask
);
3883 * Tell the memory management that we're a "memory allocator",
3884 * and that if we need more memory we should get access to it
3885 * regardless (see "__alloc_pages()"). "kswapd" should
3886 * never get caught in the normal page freeing logic.
3888 * (Kswapd normally doesn't need memory anyway, but sometimes
3889 * you need a small amount of memory in order to be able to
3890 * page out something else, and this flag essentially protects
3891 * us from recursively trying to free more memory as we're
3892 * trying to free the first piece of memory in the first place).
3894 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3897 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3898 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
3902 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3903 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
3907 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3910 /* Read the new order and highest_zoneidx */
3911 alloc_order
= READ_ONCE(pgdat
->kswapd_order
);
3912 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
3914 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3915 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
3917 ret
= try_to_freeze();
3918 if (kthread_should_stop())
3922 * We can speed up thawing tasks if we don't call balance_pgdat
3923 * after returning from the refrigerator
3929 * Reclaim begins at the requested order but if a high-order
3930 * reclaim fails then kswapd falls back to reclaiming for
3931 * order-0. If that happens, kswapd will consider sleeping
3932 * for the order it finished reclaiming at (reclaim_order)
3933 * but kcompactd is woken to compact for the original
3934 * request (alloc_order).
3936 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
3938 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
3940 if (reclaim_order
< alloc_order
)
3941 goto kswapd_try_sleep
;
3944 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3950 * A zone is low on free memory or too fragmented for high-order memory. If
3951 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3952 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3953 * has failed or is not needed, still wake up kcompactd if only compaction is
3956 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3957 enum zone_type highest_zoneidx
)
3960 enum zone_type curr_idx
;
3962 if (!managed_zone(zone
))
3965 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3968 pgdat
= zone
->zone_pgdat
;
3969 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
3971 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
3972 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
3974 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
3975 WRITE_ONCE(pgdat
->kswapd_order
, order
);
3977 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3980 /* Hopeless node, leave it to direct reclaim if possible */
3981 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3982 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
3983 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
3985 * There may be plenty of free memory available, but it's too
3986 * fragmented for high-order allocations. Wake up kcompactd
3987 * and rely on compaction_suitable() to determine if it's
3988 * needed. If it fails, it will defer subsequent attempts to
3989 * ratelimit its work.
3991 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3992 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
3996 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
3998 wake_up_interruptible(&pgdat
->kswapd_wait
);
4001 #ifdef CONFIG_HIBERNATION
4003 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4006 * Rather than trying to age LRUs the aim is to preserve the overall
4007 * LRU order by reclaiming preferentially
4008 * inactive > active > active referenced > active mapped
4010 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4012 struct scan_control sc
= {
4013 .nr_to_reclaim
= nr_to_reclaim
,
4014 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4015 .reclaim_idx
= MAX_NR_ZONES
- 1,
4016 .priority
= DEF_PRIORITY
,
4020 .hibernation_mode
= 1,
4022 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4023 unsigned long nr_reclaimed
;
4024 unsigned int noreclaim_flag
;
4026 fs_reclaim_acquire(sc
.gfp_mask
);
4027 noreclaim_flag
= memalloc_noreclaim_save();
4028 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4030 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4032 set_task_reclaim_state(current
, NULL
);
4033 memalloc_noreclaim_restore(noreclaim_flag
);
4034 fs_reclaim_release(sc
.gfp_mask
);
4036 return nr_reclaimed
;
4038 #endif /* CONFIG_HIBERNATION */
4041 * This kswapd start function will be called by init and node-hot-add.
4042 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4044 int kswapd_run(int nid
)
4046 pg_data_t
*pgdat
= NODE_DATA(nid
);
4052 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4053 if (IS_ERR(pgdat
->kswapd
)) {
4054 /* failure at boot is fatal */
4055 BUG_ON(system_state
< SYSTEM_RUNNING
);
4056 pr_err("Failed to start kswapd on node %d\n", nid
);
4057 ret
= PTR_ERR(pgdat
->kswapd
);
4058 pgdat
->kswapd
= NULL
;
4064 * Called by memory hotplug when all memory in a node is offlined. Caller must
4065 * hold mem_hotplug_begin/end().
4067 void kswapd_stop(int nid
)
4069 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4072 kthread_stop(kswapd
);
4073 NODE_DATA(nid
)->kswapd
= NULL
;
4077 static int __init
kswapd_init(void)
4082 for_each_node_state(nid
, N_MEMORY
)
4087 module_init(kswapd_init
)
4093 * If non-zero call node_reclaim when the number of free pages falls below
4096 int node_reclaim_mode __read_mostly
;
4099 * These bit locations are exposed in the vm.zone_reclaim_mode sysctl
4100 * ABI. New bits are OK, but existing bits can never change.
4102 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4103 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4104 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4107 * Priority for NODE_RECLAIM. This determines the fraction of pages
4108 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4111 #define NODE_RECLAIM_PRIORITY 4
4114 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4117 int sysctl_min_unmapped_ratio
= 1;
4120 * If the number of slab pages in a zone grows beyond this percentage then
4121 * slab reclaim needs to occur.
4123 int sysctl_min_slab_ratio
= 5;
4125 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4127 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4128 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4129 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4132 * It's possible for there to be more file mapped pages than
4133 * accounted for by the pages on the file LRU lists because
4134 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4136 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4139 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4140 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4142 unsigned long nr_pagecache_reclaimable
;
4143 unsigned long delta
= 0;
4146 * If RECLAIM_UNMAP is set, then all file pages are considered
4147 * potentially reclaimable. Otherwise, we have to worry about
4148 * pages like swapcache and node_unmapped_file_pages() provides
4151 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4152 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4154 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4156 /* If we can't clean pages, remove dirty pages from consideration */
4157 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4158 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4160 /* Watch for any possible underflows due to delta */
4161 if (unlikely(delta
> nr_pagecache_reclaimable
))
4162 delta
= nr_pagecache_reclaimable
;
4164 return nr_pagecache_reclaimable
- delta
;
4168 * Try to free up some pages from this node through reclaim.
4170 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4172 /* Minimum pages needed in order to stay on node */
4173 const unsigned long nr_pages
= 1 << order
;
4174 struct task_struct
*p
= current
;
4175 unsigned int noreclaim_flag
;
4176 struct scan_control sc
= {
4177 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4178 .gfp_mask
= current_gfp_context(gfp_mask
),
4180 .priority
= NODE_RECLAIM_PRIORITY
,
4181 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4182 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4184 .reclaim_idx
= gfp_zone(gfp_mask
),
4187 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4191 fs_reclaim_acquire(sc
.gfp_mask
);
4193 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4194 * and we also need to be able to write out pages for RECLAIM_WRITE
4195 * and RECLAIM_UNMAP.
4197 noreclaim_flag
= memalloc_noreclaim_save();
4198 p
->flags
|= PF_SWAPWRITE
;
4199 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4201 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4203 * Free memory by calling shrink node with increasing
4204 * priorities until we have enough memory freed.
4207 shrink_node(pgdat
, &sc
);
4208 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4211 set_task_reclaim_state(p
, NULL
);
4212 current
->flags
&= ~PF_SWAPWRITE
;
4213 memalloc_noreclaim_restore(noreclaim_flag
);
4214 fs_reclaim_release(sc
.gfp_mask
);
4216 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4218 return sc
.nr_reclaimed
>= nr_pages
;
4221 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4226 * Node reclaim reclaims unmapped file backed pages and
4227 * slab pages if we are over the defined limits.
4229 * A small portion of unmapped file backed pages is needed for
4230 * file I/O otherwise pages read by file I/O will be immediately
4231 * thrown out if the node is overallocated. So we do not reclaim
4232 * if less than a specified percentage of the node is used by
4233 * unmapped file backed pages.
4235 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4236 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
4237 pgdat
->min_slab_pages
)
4238 return NODE_RECLAIM_FULL
;
4241 * Do not scan if the allocation should not be delayed.
4243 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4244 return NODE_RECLAIM_NOSCAN
;
4247 * Only run node reclaim on the local node or on nodes that do not
4248 * have associated processors. This will favor the local processor
4249 * over remote processors and spread off node memory allocations
4250 * as wide as possible.
4252 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4253 return NODE_RECLAIM_NOSCAN
;
4255 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4256 return NODE_RECLAIM_NOSCAN
;
4258 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4259 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4262 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4269 * check_move_unevictable_pages - check pages for evictability and move to
4270 * appropriate zone lru list
4271 * @pvec: pagevec with lru pages to check
4273 * Checks pages for evictability, if an evictable page is in the unevictable
4274 * lru list, moves it to the appropriate evictable lru list. This function
4275 * should be only used for lru pages.
4277 void check_move_unevictable_pages(struct pagevec
*pvec
)
4279 struct lruvec
*lruvec
= NULL
;
4284 for (i
= 0; i
< pvec
->nr
; i
++) {
4285 struct page
*page
= pvec
->pages
[i
];
4288 if (PageTransTail(page
))
4291 nr_pages
= thp_nr_pages(page
);
4292 pgscanned
+= nr_pages
;
4294 /* block memcg migration during page moving between lru */
4295 if (!TestClearPageLRU(page
))
4298 lruvec
= relock_page_lruvec_irq(page
, lruvec
);
4299 if (page_evictable(page
) && PageUnevictable(page
)) {
4300 enum lru_list lru
= page_lru_base_type(page
);
4302 VM_BUG_ON_PAGE(PageActive(page
), page
);
4303 ClearPageUnevictable(page
);
4304 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4305 add_page_to_lru_list(page
, lruvec
, lru
);
4306 pgrescued
+= nr_pages
;
4312 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4313 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4314 unlock_page_lruvec_irq(lruvec
);
4315 } else if (pgscanned
) {
4316 count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4319 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);