1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim
;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup
*target_mem_cgroup
;
82 /* Can active pages be deactivated as part of reclaim? */
83 #define DEACTIVATE_ANON 1
84 #define DEACTIVATE_FILE 2
85 unsigned int may_deactivate
:2;
86 unsigned int force_deactivate
:1;
87 unsigned int skipped_deactivate
:1;
89 /* Writepage batching in laptop mode; RECLAIM_WRITE */
90 unsigned int may_writepage
:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap
:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap
:1;
99 * Cgroups are not reclaimed below their configured memory.low,
100 * unless we threaten to OOM. If any cgroups are skipped due to
101 * memory.low and nothing was reclaimed, go back for memory.low.
103 unsigned int memcg_low_reclaim
:1;
104 unsigned int memcg_low_skipped
:1;
106 unsigned int hibernation_mode
:1;
108 /* One of the zones is ready for compaction */
109 unsigned int compaction_ready
:1;
111 /* There is easily reclaimable cold cache in the current node */
112 unsigned int cache_trim_mode
:1;
114 /* The file pages on the current node are dangerously low */
115 unsigned int file_is_tiny
:1;
117 /* Allocation order */
120 /* Scan (total_size >> priority) pages at once */
123 /* The highest zone to isolate pages for reclaim from */
126 /* This context's GFP mask */
129 /* Incremented by the number of inactive pages that were scanned */
130 unsigned long nr_scanned
;
132 /* Number of pages freed so far during a call to shrink_zones() */
133 unsigned long nr_reclaimed
;
137 unsigned int unqueued_dirty
;
138 unsigned int congested
;
139 unsigned int writeback
;
140 unsigned int immediate
;
141 unsigned int file_taken
;
145 /* for recording the reclaimed slab by now */
146 struct reclaim_state reclaim_state
;
149 #ifdef ARCH_HAS_PREFETCHW
150 #define prefetchw_prev_lru_page(_page, _base, _field) \
152 if ((_page)->lru.prev != _base) { \
155 prev = lru_to_page(&(_page->lru)); \
156 prefetchw(&prev->_field); \
160 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
164 * From 0 .. 100. Higher means more swappy.
166 int vm_swappiness
= 60;
168 * The total number of pages which are beyond the high watermark within all
171 unsigned long vm_total_pages
;
173 static void set_task_reclaim_state(struct task_struct
*task
,
174 struct reclaim_state
*rs
)
176 /* Check for an overwrite */
177 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
179 /* Check for the nulling of an already-nulled member */
180 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
182 task
->reclaim_state
= rs
;
185 static LIST_HEAD(shrinker_list
);
186 static DECLARE_RWSEM(shrinker_rwsem
);
190 * We allow subsystems to populate their shrinker-related
191 * LRU lists before register_shrinker_prepared() is called
192 * for the shrinker, since we don't want to impose
193 * restrictions on their internal registration order.
194 * In this case shrink_slab_memcg() may find corresponding
195 * bit is set in the shrinkers map.
197 * This value is used by the function to detect registering
198 * shrinkers and to skip do_shrink_slab() calls for them.
200 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202 static DEFINE_IDR(shrinker_idr
);
203 static int shrinker_nr_max
;
205 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
207 int id
, ret
= -ENOMEM
;
209 down_write(&shrinker_rwsem
);
210 /* This may call shrinker, so it must use down_read_trylock() */
211 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
215 if (id
>= shrinker_nr_max
) {
216 if (memcg_expand_shrinker_maps(id
)) {
217 idr_remove(&shrinker_idr
, id
);
221 shrinker_nr_max
= id
+ 1;
226 up_write(&shrinker_rwsem
);
230 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
232 int id
= shrinker
->id
;
236 down_write(&shrinker_rwsem
);
237 idr_remove(&shrinker_idr
, id
);
238 up_write(&shrinker_rwsem
);
241 static bool cgroup_reclaim(struct scan_control
*sc
)
243 return sc
->target_mem_cgroup
;
247 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
248 * @sc: scan_control in question
250 * The normal page dirty throttling mechanism in balance_dirty_pages() is
251 * completely broken with the legacy memcg and direct stalling in
252 * shrink_page_list() is used for throttling instead, which lacks all the
253 * niceties such as fairness, adaptive pausing, bandwidth proportional
254 * allocation and configurability.
256 * This function tests whether the vmscan currently in progress can assume
257 * that the normal dirty throttling mechanism is operational.
259 static bool writeback_throttling_sane(struct scan_control
*sc
)
261 if (!cgroup_reclaim(sc
))
263 #ifdef CONFIG_CGROUP_WRITEBACK
264 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
270 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
275 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
279 static bool cgroup_reclaim(struct scan_control
*sc
)
284 static bool writeback_throttling_sane(struct scan_control
*sc
)
291 * This misses isolated pages which are not accounted for to save counters.
292 * As the data only determines if reclaim or compaction continues, it is
293 * not expected that isolated pages will be a dominating factor.
295 unsigned long zone_reclaimable_pages(struct zone
*zone
)
299 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
300 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
301 if (get_nr_swap_pages() > 0)
302 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
303 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
309 * lruvec_lru_size - Returns the number of pages on the given LRU list.
310 * @lruvec: lru vector
312 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
314 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
316 unsigned long size
= 0;
319 for (zid
= 0; zid
<= zone_idx
&& zid
< MAX_NR_ZONES
; zid
++) {
320 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
322 if (!managed_zone(zone
))
325 if (!mem_cgroup_disabled())
326 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
328 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
334 * Add a shrinker callback to be called from the vm.
336 int prealloc_shrinker(struct shrinker
*shrinker
)
338 unsigned int size
= sizeof(*shrinker
->nr_deferred
);
340 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
343 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
344 if (!shrinker
->nr_deferred
)
347 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
348 if (prealloc_memcg_shrinker(shrinker
))
355 kfree(shrinker
->nr_deferred
);
356 shrinker
->nr_deferred
= NULL
;
360 void free_prealloced_shrinker(struct shrinker
*shrinker
)
362 if (!shrinker
->nr_deferred
)
365 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
366 unregister_memcg_shrinker(shrinker
);
368 kfree(shrinker
->nr_deferred
);
369 shrinker
->nr_deferred
= NULL
;
372 void register_shrinker_prepared(struct shrinker
*shrinker
)
374 down_write(&shrinker_rwsem
);
375 list_add_tail(&shrinker
->list
, &shrinker_list
);
377 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
378 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
380 up_write(&shrinker_rwsem
);
383 int register_shrinker(struct shrinker
*shrinker
)
385 int err
= prealloc_shrinker(shrinker
);
389 register_shrinker_prepared(shrinker
);
392 EXPORT_SYMBOL(register_shrinker
);
397 void unregister_shrinker(struct shrinker
*shrinker
)
399 if (!shrinker
->nr_deferred
)
401 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
402 unregister_memcg_shrinker(shrinker
);
403 down_write(&shrinker_rwsem
);
404 list_del(&shrinker
->list
);
405 up_write(&shrinker_rwsem
);
406 kfree(shrinker
->nr_deferred
);
407 shrinker
->nr_deferred
= NULL
;
409 EXPORT_SYMBOL(unregister_shrinker
);
411 #define SHRINK_BATCH 128
413 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
414 struct shrinker
*shrinker
, int priority
)
416 unsigned long freed
= 0;
417 unsigned long long delta
;
422 int nid
= shrinkctl
->nid
;
423 long batch_size
= shrinker
->batch
? shrinker
->batch
425 long scanned
= 0, next_deferred
;
427 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
430 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
431 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
435 * copy the current shrinker scan count into a local variable
436 * and zero it so that other concurrent shrinker invocations
437 * don't also do this scanning work.
439 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
442 if (shrinker
->seeks
) {
443 delta
= freeable
>> priority
;
445 do_div(delta
, shrinker
->seeks
);
448 * These objects don't require any IO to create. Trim
449 * them aggressively under memory pressure to keep
450 * them from causing refetches in the IO caches.
452 delta
= freeable
/ 2;
456 if (total_scan
< 0) {
457 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
458 shrinker
->scan_objects
, total_scan
);
459 total_scan
= freeable
;
462 next_deferred
= total_scan
;
465 * We need to avoid excessive windup on filesystem shrinkers
466 * due to large numbers of GFP_NOFS allocations causing the
467 * shrinkers to return -1 all the time. This results in a large
468 * nr being built up so when a shrink that can do some work
469 * comes along it empties the entire cache due to nr >>>
470 * freeable. This is bad for sustaining a working set in
473 * Hence only allow the shrinker to scan the entire cache when
474 * a large delta change is calculated directly.
476 if (delta
< freeable
/ 4)
477 total_scan
= min(total_scan
, freeable
/ 2);
480 * Avoid risking looping forever due to too large nr value:
481 * never try to free more than twice the estimate number of
484 if (total_scan
> freeable
* 2)
485 total_scan
= freeable
* 2;
487 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
488 freeable
, delta
, total_scan
, priority
);
491 * Normally, we should not scan less than batch_size objects in one
492 * pass to avoid too frequent shrinker calls, but if the slab has less
493 * than batch_size objects in total and we are really tight on memory,
494 * we will try to reclaim all available objects, otherwise we can end
495 * up failing allocations although there are plenty of reclaimable
496 * objects spread over several slabs with usage less than the
499 * We detect the "tight on memory" situations by looking at the total
500 * number of objects we want to scan (total_scan). If it is greater
501 * than the total number of objects on slab (freeable), we must be
502 * scanning at high prio and therefore should try to reclaim as much as
505 while (total_scan
>= batch_size
||
506 total_scan
>= freeable
) {
508 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
510 shrinkctl
->nr_to_scan
= nr_to_scan
;
511 shrinkctl
->nr_scanned
= nr_to_scan
;
512 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
513 if (ret
== SHRINK_STOP
)
517 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
518 total_scan
-= shrinkctl
->nr_scanned
;
519 scanned
+= shrinkctl
->nr_scanned
;
524 if (next_deferred
>= scanned
)
525 next_deferred
-= scanned
;
529 * move the unused scan count back into the shrinker in a
530 * manner that handles concurrent updates. If we exhausted the
531 * scan, there is no need to do an update.
533 if (next_deferred
> 0)
534 new_nr
= atomic_long_add_return(next_deferred
,
535 &shrinker
->nr_deferred
[nid
]);
537 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
539 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
544 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
545 struct mem_cgroup
*memcg
, int priority
)
547 struct memcg_shrinker_map
*map
;
548 unsigned long ret
, freed
= 0;
551 if (!mem_cgroup_online(memcg
))
554 if (!down_read_trylock(&shrinker_rwsem
))
557 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
562 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
563 struct shrink_control sc
= {
564 .gfp_mask
= gfp_mask
,
568 struct shrinker
*shrinker
;
570 shrinker
= idr_find(&shrinker_idr
, i
);
571 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
573 clear_bit(i
, map
->map
);
577 /* Call non-slab shrinkers even though kmem is disabled */
578 if (!memcg_kmem_enabled() &&
579 !(shrinker
->flags
& SHRINKER_NONSLAB
))
582 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
583 if (ret
== SHRINK_EMPTY
) {
584 clear_bit(i
, map
->map
);
586 * After the shrinker reported that it had no objects to
587 * free, but before we cleared the corresponding bit in
588 * the memcg shrinker map, a new object might have been
589 * added. To make sure, we have the bit set in this
590 * case, we invoke the shrinker one more time and reset
591 * the bit if it reports that it is not empty anymore.
592 * The memory barrier here pairs with the barrier in
593 * memcg_set_shrinker_bit():
595 * list_lru_add() shrink_slab_memcg()
596 * list_add_tail() clear_bit()
598 * set_bit() do_shrink_slab()
600 smp_mb__after_atomic();
601 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
602 if (ret
== SHRINK_EMPTY
)
605 memcg_set_shrinker_bit(memcg
, nid
, i
);
609 if (rwsem_is_contended(&shrinker_rwsem
)) {
615 up_read(&shrinker_rwsem
);
618 #else /* CONFIG_MEMCG */
619 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
620 struct mem_cgroup
*memcg
, int priority
)
624 #endif /* CONFIG_MEMCG */
627 * shrink_slab - shrink slab caches
628 * @gfp_mask: allocation context
629 * @nid: node whose slab caches to target
630 * @memcg: memory cgroup whose slab caches to target
631 * @priority: the reclaim priority
633 * Call the shrink functions to age shrinkable caches.
635 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
636 * unaware shrinkers will receive a node id of 0 instead.
638 * @memcg specifies the memory cgroup to target. Unaware shrinkers
639 * are called only if it is the root cgroup.
641 * @priority is sc->priority, we take the number of objects and >> by priority
642 * in order to get the scan target.
644 * Returns the number of reclaimed slab objects.
646 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
647 struct mem_cgroup
*memcg
,
650 unsigned long ret
, freed
= 0;
651 struct shrinker
*shrinker
;
654 * The root memcg might be allocated even though memcg is disabled
655 * via "cgroup_disable=memory" boot parameter. This could make
656 * mem_cgroup_is_root() return false, then just run memcg slab
657 * shrink, but skip global shrink. This may result in premature
660 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
))
661 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
663 if (!down_read_trylock(&shrinker_rwsem
))
666 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
667 struct shrink_control sc
= {
668 .gfp_mask
= gfp_mask
,
673 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
674 if (ret
== SHRINK_EMPTY
)
678 * Bail out if someone want to register a new shrinker to
679 * prevent the regsitration from being stalled for long periods
680 * by parallel ongoing shrinking.
682 if (rwsem_is_contended(&shrinker_rwsem
)) {
688 up_read(&shrinker_rwsem
);
694 void drop_slab_node(int nid
)
699 struct mem_cgroup
*memcg
= NULL
;
702 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
704 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
705 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
706 } while (freed
> 10);
713 for_each_online_node(nid
)
717 static inline int is_page_cache_freeable(struct page
*page
)
720 * A freeable page cache page is referenced only by the caller
721 * that isolated the page, the page cache and optional buffer
722 * heads at page->private.
724 int page_cache_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
726 return page_count(page
) - page_has_private(page
) == 1 + page_cache_pins
;
729 static int may_write_to_inode(struct inode
*inode
)
731 if (current
->flags
& PF_SWAPWRITE
)
733 if (!inode_write_congested(inode
))
735 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
741 * We detected a synchronous write error writing a page out. Probably
742 * -ENOSPC. We need to propagate that into the address_space for a subsequent
743 * fsync(), msync() or close().
745 * The tricky part is that after writepage we cannot touch the mapping: nothing
746 * prevents it from being freed up. But we have a ref on the page and once
747 * that page is locked, the mapping is pinned.
749 * We're allowed to run sleeping lock_page() here because we know the caller has
752 static void handle_write_error(struct address_space
*mapping
,
753 struct page
*page
, int error
)
756 if (page_mapping(page
) == mapping
)
757 mapping_set_error(mapping
, error
);
761 /* possible outcome of pageout() */
763 /* failed to write page out, page is locked */
765 /* move page to the active list, page is locked */
767 /* page has been sent to the disk successfully, page is unlocked */
769 /* page is clean and locked */
774 * pageout is called by shrink_page_list() for each dirty page.
775 * Calls ->writepage().
777 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
780 * If the page is dirty, only perform writeback if that write
781 * will be non-blocking. To prevent this allocation from being
782 * stalled by pagecache activity. But note that there may be
783 * stalls if we need to run get_block(). We could test
784 * PagePrivate for that.
786 * If this process is currently in __generic_file_write_iter() against
787 * this page's queue, we can perform writeback even if that
790 * If the page is swapcache, write it back even if that would
791 * block, for some throttling. This happens by accident, because
792 * swap_backing_dev_info is bust: it doesn't reflect the
793 * congestion state of the swapdevs. Easy to fix, if needed.
795 if (!is_page_cache_freeable(page
))
799 * Some data journaling orphaned pages can have
800 * page->mapping == NULL while being dirty with clean buffers.
802 if (page_has_private(page
)) {
803 if (try_to_free_buffers(page
)) {
804 ClearPageDirty(page
);
805 pr_info("%s: orphaned page\n", __func__
);
811 if (mapping
->a_ops
->writepage
== NULL
)
812 return PAGE_ACTIVATE
;
813 if (!may_write_to_inode(mapping
->host
))
816 if (clear_page_dirty_for_io(page
)) {
818 struct writeback_control wbc
= {
819 .sync_mode
= WB_SYNC_NONE
,
820 .nr_to_write
= SWAP_CLUSTER_MAX
,
822 .range_end
= LLONG_MAX
,
826 SetPageReclaim(page
);
827 res
= mapping
->a_ops
->writepage(page
, &wbc
);
829 handle_write_error(mapping
, page
, res
);
830 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
831 ClearPageReclaim(page
);
832 return PAGE_ACTIVATE
;
835 if (!PageWriteback(page
)) {
836 /* synchronous write or broken a_ops? */
837 ClearPageReclaim(page
);
839 trace_mm_vmscan_writepage(page
);
840 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
848 * Same as remove_mapping, but if the page is removed from the mapping, it
849 * gets returned with a refcount of 0.
851 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
852 bool reclaimed
, struct mem_cgroup
*target_memcg
)
857 BUG_ON(!PageLocked(page
));
858 BUG_ON(mapping
!= page_mapping(page
));
860 xa_lock_irqsave(&mapping
->i_pages
, flags
);
862 * The non racy check for a busy page.
864 * Must be careful with the order of the tests. When someone has
865 * a ref to the page, it may be possible that they dirty it then
866 * drop the reference. So if PageDirty is tested before page_count
867 * here, then the following race may occur:
869 * get_user_pages(&page);
870 * [user mapping goes away]
872 * !PageDirty(page) [good]
873 * SetPageDirty(page);
875 * !page_count(page) [good, discard it]
877 * [oops, our write_to data is lost]
879 * Reversing the order of the tests ensures such a situation cannot
880 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
881 * load is not satisfied before that of page->_refcount.
883 * Note that if SetPageDirty is always performed via set_page_dirty,
884 * and thus under the i_pages lock, then this ordering is not required.
886 refcount
= 1 + compound_nr(page
);
887 if (!page_ref_freeze(page
, refcount
))
889 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
890 if (unlikely(PageDirty(page
))) {
891 page_ref_unfreeze(page
, refcount
);
895 if (PageSwapCache(page
)) {
896 swp_entry_t swap
= { .val
= page_private(page
) };
897 mem_cgroup_swapout(page
, swap
);
898 __delete_from_swap_cache(page
, swap
);
899 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
900 put_swap_page(page
, swap
);
902 void (*freepage
)(struct page
*);
905 freepage
= mapping
->a_ops
->freepage
;
907 * Remember a shadow entry for reclaimed file cache in
908 * order to detect refaults, thus thrashing, later on.
910 * But don't store shadows in an address space that is
911 * already exiting. This is not just an optizimation,
912 * inode reclaim needs to empty out the radix tree or
913 * the nodes are lost. Don't plant shadows behind its
916 * We also don't store shadows for DAX mappings because the
917 * only page cache pages found in these are zero pages
918 * covering holes, and because we don't want to mix DAX
919 * exceptional entries and shadow exceptional entries in the
920 * same address_space.
922 if (reclaimed
&& page_is_file_lru(page
) &&
923 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
924 shadow
= workingset_eviction(page
, target_memcg
);
925 __delete_from_page_cache(page
, shadow
);
926 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
928 if (freepage
!= NULL
)
935 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
940 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
941 * someone else has a ref on the page, abort and return 0. If it was
942 * successfully detached, return 1. Assumes the caller has a single ref on
945 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
947 if (__remove_mapping(mapping
, page
, false, NULL
)) {
949 * Unfreezing the refcount with 1 rather than 2 effectively
950 * drops the pagecache ref for us without requiring another
953 page_ref_unfreeze(page
, 1);
960 * putback_lru_page - put previously isolated page onto appropriate LRU list
961 * @page: page to be put back to appropriate lru list
963 * Add previously isolated @page to appropriate LRU list.
964 * Page may still be unevictable for other reasons.
966 * lru_lock must not be held, interrupts must be enabled.
968 void putback_lru_page(struct page
*page
)
971 put_page(page
); /* drop ref from isolate */
974 enum page_references
{
976 PAGEREF_RECLAIM_CLEAN
,
981 static enum page_references
page_check_references(struct page
*page
,
982 struct scan_control
*sc
)
984 int referenced_ptes
, referenced_page
;
985 unsigned long vm_flags
;
987 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
989 referenced_page
= TestClearPageReferenced(page
);
992 * Mlock lost the isolation race with us. Let try_to_unmap()
993 * move the page to the unevictable list.
995 if (vm_flags
& VM_LOCKED
)
996 return PAGEREF_RECLAIM
;
998 if (referenced_ptes
) {
999 if (PageSwapBacked(page
))
1000 return PAGEREF_ACTIVATE
;
1002 * All mapped pages start out with page table
1003 * references from the instantiating fault, so we need
1004 * to look twice if a mapped file page is used more
1007 * Mark it and spare it for another trip around the
1008 * inactive list. Another page table reference will
1009 * lead to its activation.
1011 * Note: the mark is set for activated pages as well
1012 * so that recently deactivated but used pages are
1013 * quickly recovered.
1015 SetPageReferenced(page
);
1017 if (referenced_page
|| referenced_ptes
> 1)
1018 return PAGEREF_ACTIVATE
;
1021 * Activate file-backed executable pages after first usage.
1023 if (vm_flags
& VM_EXEC
)
1024 return PAGEREF_ACTIVATE
;
1026 return PAGEREF_KEEP
;
1029 /* Reclaim if clean, defer dirty pages to writeback */
1030 if (referenced_page
&& !PageSwapBacked(page
))
1031 return PAGEREF_RECLAIM_CLEAN
;
1033 return PAGEREF_RECLAIM
;
1036 /* Check if a page is dirty or under writeback */
1037 static void page_check_dirty_writeback(struct page
*page
,
1038 bool *dirty
, bool *writeback
)
1040 struct address_space
*mapping
;
1043 * Anonymous pages are not handled by flushers and must be written
1044 * from reclaim context. Do not stall reclaim based on them
1046 if (!page_is_file_lru(page
) ||
1047 (PageAnon(page
) && !PageSwapBacked(page
))) {
1053 /* By default assume that the page flags are accurate */
1054 *dirty
= PageDirty(page
);
1055 *writeback
= PageWriteback(page
);
1057 /* Verify dirty/writeback state if the filesystem supports it */
1058 if (!page_has_private(page
))
1061 mapping
= page_mapping(page
);
1062 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1063 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1067 * shrink_page_list() returns the number of reclaimed pages
1069 static unsigned long shrink_page_list(struct list_head
*page_list
,
1070 struct pglist_data
*pgdat
,
1071 struct scan_control
*sc
,
1072 enum ttu_flags ttu_flags
,
1073 struct reclaim_stat
*stat
,
1074 bool ignore_references
)
1076 LIST_HEAD(ret_pages
);
1077 LIST_HEAD(free_pages
);
1078 unsigned nr_reclaimed
= 0;
1079 unsigned 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 (PageTransHuge(page
)) {
1242 /* cannot split THP, skip it */
1243 if (!can_split_huge_page(page
, NULL
))
1244 goto activate_locked
;
1246 * Split pages without a PMD map right
1247 * away. Chances are some or all of the
1248 * tail pages can be freed without IO.
1250 if (!compound_mapcount(page
) &&
1251 split_huge_page_to_list(page
,
1253 goto activate_locked
;
1255 if (!add_to_swap(page
)) {
1256 if (!PageTransHuge(page
))
1257 goto activate_locked_split
;
1258 /* Fallback to swap normal pages */
1259 if (split_huge_page_to_list(page
,
1261 goto activate_locked
;
1262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1263 count_vm_event(THP_SWPOUT_FALLBACK
);
1265 if (!add_to_swap(page
))
1266 goto activate_locked_split
;
1269 may_enter_fs
= true;
1271 /* Adding to swap updated mapping */
1272 mapping
= page_mapping(page
);
1274 } else if (unlikely(PageTransHuge(page
))) {
1275 /* Split file THP */
1276 if (split_huge_page_to_list(page
, page_list
))
1281 * THP may get split above, need minus tail pages and update
1282 * nr_pages to avoid accounting tail pages twice.
1284 * The tail pages that are added into swap cache successfully
1287 if ((nr_pages
> 1) && !PageTransHuge(page
)) {
1288 sc
->nr_scanned
-= (nr_pages
- 1);
1293 * The page is mapped into the page tables of one or more
1294 * processes. Try to unmap it here.
1296 if (page_mapped(page
)) {
1297 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1299 if (unlikely(PageTransHuge(page
)))
1300 flags
|= TTU_SPLIT_HUGE_PMD
;
1301 if (!try_to_unmap(page
, flags
)) {
1302 stat
->nr_unmap_fail
+= nr_pages
;
1303 goto activate_locked
;
1307 if (PageDirty(page
)) {
1309 * Only kswapd can writeback filesystem pages
1310 * to avoid risk of stack overflow. But avoid
1311 * injecting inefficient single-page IO into
1312 * flusher writeback as much as possible: only
1313 * write pages when we've encountered many
1314 * dirty pages, and when we've already scanned
1315 * the rest of the LRU for clean pages and see
1316 * the same dirty pages again (PageReclaim).
1318 if (page_is_file_lru(page
) &&
1319 (!current_is_kswapd() || !PageReclaim(page
) ||
1320 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1322 * Immediately reclaim when written back.
1323 * Similar in principal to deactivate_page()
1324 * except we already have the page isolated
1325 * and know it's dirty
1327 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1328 SetPageReclaim(page
);
1330 goto activate_locked
;
1333 if (references
== PAGEREF_RECLAIM_CLEAN
)
1337 if (!sc
->may_writepage
)
1341 * Page is dirty. Flush the TLB if a writable entry
1342 * potentially exists to avoid CPU writes after IO
1343 * starts and then write it out here.
1345 try_to_unmap_flush_dirty();
1346 switch (pageout(page
, mapping
)) {
1350 goto activate_locked
;
1352 if (PageWriteback(page
))
1354 if (PageDirty(page
))
1358 * A synchronous write - probably a ramdisk. Go
1359 * ahead and try to reclaim the page.
1361 if (!trylock_page(page
))
1363 if (PageDirty(page
) || PageWriteback(page
))
1365 mapping
= page_mapping(page
);
1367 ; /* try to free the page below */
1372 * If the page has buffers, try to free the buffer mappings
1373 * associated with this page. If we succeed we try to free
1376 * We do this even if the page is PageDirty().
1377 * try_to_release_page() does not perform I/O, but it is
1378 * possible for a page to have PageDirty set, but it is actually
1379 * clean (all its buffers are clean). This happens if the
1380 * buffers were written out directly, with submit_bh(). ext3
1381 * will do this, as well as the blockdev mapping.
1382 * try_to_release_page() will discover that cleanness and will
1383 * drop the buffers and mark the page clean - it can be freed.
1385 * Rarely, pages can have buffers and no ->mapping. These are
1386 * the pages which were not successfully invalidated in
1387 * truncate_complete_page(). We try to drop those buffers here
1388 * and if that worked, and the page is no longer mapped into
1389 * process address space (page_count == 1) it can be freed.
1390 * Otherwise, leave the page on the LRU so it is swappable.
1392 if (page_has_private(page
)) {
1393 if (!try_to_release_page(page
, sc
->gfp_mask
))
1394 goto activate_locked
;
1395 if (!mapping
&& page_count(page
) == 1) {
1397 if (put_page_testzero(page
))
1401 * rare race with speculative reference.
1402 * the speculative reference will free
1403 * this page shortly, so we may
1404 * increment nr_reclaimed here (and
1405 * leave it off the LRU).
1413 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1414 /* follow __remove_mapping for reference */
1415 if (!page_ref_freeze(page
, 1))
1417 if (PageDirty(page
)) {
1418 page_ref_unfreeze(page
, 1);
1422 count_vm_event(PGLAZYFREED
);
1423 count_memcg_page_event(page
, PGLAZYFREED
);
1424 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true,
1425 sc
->target_mem_cgroup
))
1431 * THP may get swapped out in a whole, need account
1434 nr_reclaimed
+= nr_pages
;
1437 * Is there need to periodically free_page_list? It would
1438 * appear not as the counts should be low
1440 if (unlikely(PageTransHuge(page
)))
1441 (*get_compound_page_dtor(page
))(page
);
1443 list_add(&page
->lru
, &free_pages
);
1446 activate_locked_split
:
1448 * The tail pages that are failed to add into swap cache
1449 * reach here. Fixup nr_scanned and nr_pages.
1452 sc
->nr_scanned
-= (nr_pages
- 1);
1456 /* Not a candidate for swapping, so reclaim swap space. */
1457 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1459 try_to_free_swap(page
);
1460 VM_BUG_ON_PAGE(PageActive(page
), page
);
1461 if (!PageMlocked(page
)) {
1462 int type
= page_is_file_lru(page
);
1463 SetPageActive(page
);
1464 stat
->nr_activate
[type
] += nr_pages
;
1465 count_memcg_page_event(page
, PGACTIVATE
);
1470 list_add(&page
->lru
, &ret_pages
);
1471 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1474 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1476 mem_cgroup_uncharge_list(&free_pages
);
1477 try_to_unmap_flush();
1478 free_unref_page_list(&free_pages
);
1480 list_splice(&ret_pages
, page_list
);
1481 count_vm_events(PGACTIVATE
, pgactivate
);
1483 return nr_reclaimed
;
1486 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1487 struct list_head
*page_list
)
1489 struct scan_control sc
= {
1490 .gfp_mask
= GFP_KERNEL
,
1491 .priority
= DEF_PRIORITY
,
1494 struct reclaim_stat dummy_stat
;
1496 struct page
*page
, *next
;
1497 LIST_HEAD(clean_pages
);
1499 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1500 if (page_is_file_lru(page
) && !PageDirty(page
) &&
1501 !__PageMovable(page
) && !PageUnevictable(page
)) {
1502 ClearPageActive(page
);
1503 list_move(&page
->lru
, &clean_pages
);
1507 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1508 TTU_IGNORE_ACCESS
, &dummy_stat
, true);
1509 list_splice(&clean_pages
, page_list
);
1510 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1515 * Attempt to remove the specified page from its LRU. Only take this page
1516 * if it is of the appropriate PageActive status. Pages which are being
1517 * freed elsewhere are also ignored.
1519 * page: page to consider
1520 * mode: one of the LRU isolation modes defined above
1522 * returns 0 on success, -ve errno on failure.
1524 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1528 /* Only take pages on the LRU. */
1532 /* Compaction should not handle unevictable pages but CMA can do so */
1533 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1539 * To minimise LRU disruption, the caller can indicate that it only
1540 * wants to isolate pages it will be able to operate on without
1541 * blocking - clean pages for the most part.
1543 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1544 * that it is possible to migrate without blocking
1546 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1547 /* All the caller can do on PageWriteback is block */
1548 if (PageWriteback(page
))
1551 if (PageDirty(page
)) {
1552 struct address_space
*mapping
;
1556 * Only pages without mappings or that have a
1557 * ->migratepage callback are possible to migrate
1558 * without blocking. However, we can be racing with
1559 * truncation so it's necessary to lock the page
1560 * to stabilise the mapping as truncation holds
1561 * the page lock until after the page is removed
1562 * from the page cache.
1564 if (!trylock_page(page
))
1567 mapping
= page_mapping(page
);
1568 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1575 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1578 if (likely(get_page_unless_zero(page
))) {
1580 * Be careful not to clear PageLRU until after we're
1581 * sure the page is not being freed elsewhere -- the
1582 * page release code relies on it.
1593 * Update LRU sizes after isolating pages. The LRU size updates must
1594 * be complete before mem_cgroup_update_lru_size due to a santity check.
1596 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1597 enum lru_list lru
, unsigned long *nr_zone_taken
)
1601 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1602 if (!nr_zone_taken
[zid
])
1605 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1607 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1614 * pgdat->lru_lock is heavily contended. Some of the functions that
1615 * shrink the lists perform better by taking out a batch of pages
1616 * and working on them outside the LRU lock.
1618 * For pagecache intensive workloads, this function is the hottest
1619 * spot in the kernel (apart from copy_*_user functions).
1621 * Appropriate locks must be held before calling this function.
1623 * @nr_to_scan: The number of eligible pages to look through on the list.
1624 * @lruvec: The LRU vector to pull pages from.
1625 * @dst: The temp list to put pages on to.
1626 * @nr_scanned: The number of pages that were scanned.
1627 * @sc: The scan_control struct for this reclaim session
1628 * @mode: One of the LRU isolation modes
1629 * @lru: LRU list id for isolating
1631 * returns how many pages were moved onto *@dst.
1633 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1634 struct lruvec
*lruvec
, struct list_head
*dst
,
1635 unsigned long *nr_scanned
, struct scan_control
*sc
,
1638 struct list_head
*src
= &lruvec
->lists
[lru
];
1639 unsigned long nr_taken
= 0;
1640 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1641 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1642 unsigned long skipped
= 0;
1643 unsigned long scan
, total_scan
, nr_pages
;
1644 LIST_HEAD(pages_skipped
);
1645 isolate_mode_t mode
= (sc
->may_unmap
? 0 : ISOLATE_UNMAPPED
);
1649 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1652 page
= lru_to_page(src
);
1653 prefetchw_prev_lru_page(page
, src
, flags
);
1655 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1657 nr_pages
= compound_nr(page
);
1658 total_scan
+= nr_pages
;
1660 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1661 list_move(&page
->lru
, &pages_skipped
);
1662 nr_skipped
[page_zonenum(page
)] += nr_pages
;
1667 * Do not count skipped pages because that makes the function
1668 * return with no isolated pages if the LRU mostly contains
1669 * ineligible pages. This causes the VM to not reclaim any
1670 * pages, triggering a premature OOM.
1672 * Account all tail pages of THP. This would not cause
1673 * premature OOM since __isolate_lru_page() returns -EBUSY
1674 * only when the page is being freed somewhere else.
1677 switch (__isolate_lru_page(page
, mode
)) {
1679 nr_taken
+= nr_pages
;
1680 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1681 list_move(&page
->lru
, dst
);
1685 /* else it is being freed elsewhere */
1686 list_move(&page
->lru
, src
);
1695 * Splice any skipped pages to the start of the LRU list. Note that
1696 * this disrupts the LRU order when reclaiming for lower zones but
1697 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1698 * scanning would soon rescan the same pages to skip and put the
1699 * system at risk of premature OOM.
1701 if (!list_empty(&pages_skipped
)) {
1704 list_splice(&pages_skipped
, src
);
1705 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1706 if (!nr_skipped
[zid
])
1709 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1710 skipped
+= nr_skipped
[zid
];
1713 *nr_scanned
= total_scan
;
1714 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1715 total_scan
, skipped
, nr_taken
, mode
, lru
);
1716 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1721 * isolate_lru_page - tries to isolate a page from its LRU list
1722 * @page: page to isolate from its LRU list
1724 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1725 * vmstat statistic corresponding to whatever LRU list the page was on.
1727 * Returns 0 if the page was removed from an LRU list.
1728 * Returns -EBUSY if the page was not on an LRU list.
1730 * The returned page will have PageLRU() cleared. If it was found on
1731 * the active list, it will have PageActive set. If it was found on
1732 * the unevictable list, it will have the PageUnevictable bit set. That flag
1733 * may need to be cleared by the caller before letting the page go.
1735 * The vmstat statistic corresponding to the list on which the page was
1736 * found will be decremented.
1740 * (1) Must be called with an elevated refcount on the page. This is a
1741 * fundamentnal difference from isolate_lru_pages (which is called
1742 * without a stable reference).
1743 * (2) the lru_lock must not be held.
1744 * (3) interrupts must be enabled.
1746 int isolate_lru_page(struct page
*page
)
1750 VM_BUG_ON_PAGE(!page_count(page
), page
);
1751 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1753 if (PageLRU(page
)) {
1754 pg_data_t
*pgdat
= page_pgdat(page
);
1755 struct lruvec
*lruvec
;
1757 spin_lock_irq(&pgdat
->lru_lock
);
1758 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1759 if (PageLRU(page
)) {
1760 int lru
= page_lru(page
);
1763 del_page_from_lru_list(page
, lruvec
, lru
);
1766 spin_unlock_irq(&pgdat
->lru_lock
);
1772 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1773 * then get rescheduled. When there are massive number of tasks doing page
1774 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1775 * the LRU list will go small and be scanned faster than necessary, leading to
1776 * unnecessary swapping, thrashing and OOM.
1778 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1779 struct scan_control
*sc
)
1781 unsigned long inactive
, isolated
;
1783 if (current_is_kswapd())
1786 if (!writeback_throttling_sane(sc
))
1790 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1791 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1793 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1794 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1798 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1799 * won't get blocked by normal direct-reclaimers, forming a circular
1802 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1805 return isolated
> inactive
;
1809 * This moves pages from @list to corresponding LRU list.
1811 * We move them the other way if the page is referenced by one or more
1812 * processes, from rmap.
1814 * If the pages are mostly unmapped, the processing is fast and it is
1815 * appropriate to hold zone_lru_lock across the whole operation. But if
1816 * the pages are mapped, the processing is slow (page_referenced()) so we
1817 * should drop zone_lru_lock around each page. It's impossible to balance
1818 * this, so instead we remove the pages from the LRU while processing them.
1819 * It is safe to rely on PG_active against the non-LRU pages in here because
1820 * nobody will play with that bit on a non-LRU page.
1822 * The downside is that we have to touch page->_refcount against each page.
1823 * But we had to alter page->flags anyway.
1825 * Returns the number of pages moved to the given lruvec.
1828 static unsigned noinline_for_stack
move_pages_to_lru(struct lruvec
*lruvec
,
1829 struct list_head
*list
)
1831 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1832 int nr_pages
, nr_moved
= 0;
1833 LIST_HEAD(pages_to_free
);
1837 while (!list_empty(list
)) {
1838 page
= lru_to_page(list
);
1839 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1840 if (unlikely(!page_evictable(page
))) {
1841 list_del(&page
->lru
);
1842 spin_unlock_irq(&pgdat
->lru_lock
);
1843 putback_lru_page(page
);
1844 spin_lock_irq(&pgdat
->lru_lock
);
1847 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1850 lru
= page_lru(page
);
1852 nr_pages
= hpage_nr_pages(page
);
1853 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1854 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1856 if (put_page_testzero(page
)) {
1857 __ClearPageLRU(page
);
1858 __ClearPageActive(page
);
1859 del_page_from_lru_list(page
, lruvec
, lru
);
1861 if (unlikely(PageCompound(page
))) {
1862 spin_unlock_irq(&pgdat
->lru_lock
);
1863 (*get_compound_page_dtor(page
))(page
);
1864 spin_lock_irq(&pgdat
->lru_lock
);
1866 list_add(&page
->lru
, &pages_to_free
);
1868 nr_moved
+= nr_pages
;
1873 * To save our caller's stack, now use input list for pages to free.
1875 list_splice(&pages_to_free
, list
);
1881 * If a kernel thread (such as nfsd for loop-back mounts) services
1882 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1883 * In that case we should only throttle if the backing device it is
1884 * writing to is congested. In other cases it is safe to throttle.
1886 static int current_may_throttle(void)
1888 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1889 current
->backing_dev_info
== NULL
||
1890 bdi_write_congested(current
->backing_dev_info
);
1894 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1895 * of reclaimed pages
1897 static noinline_for_stack
unsigned long
1898 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1899 struct scan_control
*sc
, enum lru_list lru
)
1901 LIST_HEAD(page_list
);
1902 unsigned long nr_scanned
;
1903 unsigned long nr_reclaimed
= 0;
1904 unsigned long nr_taken
;
1905 struct reclaim_stat stat
;
1906 int file
= is_file_lru(lru
);
1907 enum vm_event_item item
;
1908 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1909 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1910 bool stalled
= false;
1912 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1916 /* wait a bit for the reclaimer. */
1920 /* We are about to die and free our memory. Return now. */
1921 if (fatal_signal_pending(current
))
1922 return SWAP_CLUSTER_MAX
;
1927 spin_lock_irq(&pgdat
->lru_lock
);
1929 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1930 &nr_scanned
, sc
, lru
);
1932 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1933 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1935 item
= current_is_kswapd() ? PGSCAN_KSWAPD
: PGSCAN_DIRECT
;
1936 if (!cgroup_reclaim(sc
))
1937 __count_vm_events(item
, nr_scanned
);
1938 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1939 spin_unlock_irq(&pgdat
->lru_lock
);
1944 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1947 spin_lock_irq(&pgdat
->lru_lock
);
1949 item
= current_is_kswapd() ? PGSTEAL_KSWAPD
: PGSTEAL_DIRECT
;
1950 if (!cgroup_reclaim(sc
))
1951 __count_vm_events(item
, nr_reclaimed
);
1952 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
1953 reclaim_stat
->recent_rotated
[0] += stat
.nr_activate
[0];
1954 reclaim_stat
->recent_rotated
[1] += stat
.nr_activate
[1];
1956 move_pages_to_lru(lruvec
, &page_list
);
1958 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1960 spin_unlock_irq(&pgdat
->lru_lock
);
1962 mem_cgroup_uncharge_list(&page_list
);
1963 free_unref_page_list(&page_list
);
1966 * If dirty pages are scanned that are not queued for IO, it
1967 * implies that flushers are not doing their job. This can
1968 * happen when memory pressure pushes dirty pages to the end of
1969 * the LRU before the dirty limits are breached and the dirty
1970 * data has expired. It can also happen when the proportion of
1971 * dirty pages grows not through writes but through memory
1972 * pressure reclaiming all the clean cache. And in some cases,
1973 * the flushers simply cannot keep up with the allocation
1974 * rate. Nudge the flusher threads in case they are asleep.
1976 if (stat
.nr_unqueued_dirty
== nr_taken
)
1977 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1979 sc
->nr
.dirty
+= stat
.nr_dirty
;
1980 sc
->nr
.congested
+= stat
.nr_congested
;
1981 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1982 sc
->nr
.writeback
+= stat
.nr_writeback
;
1983 sc
->nr
.immediate
+= stat
.nr_immediate
;
1984 sc
->nr
.taken
+= nr_taken
;
1986 sc
->nr
.file_taken
+= nr_taken
;
1988 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1989 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
1990 return nr_reclaimed
;
1993 static void shrink_active_list(unsigned long nr_to_scan
,
1994 struct lruvec
*lruvec
,
1995 struct scan_control
*sc
,
1998 unsigned long nr_taken
;
1999 unsigned long nr_scanned
;
2000 unsigned long vm_flags
;
2001 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2002 LIST_HEAD(l_active
);
2003 LIST_HEAD(l_inactive
);
2005 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2006 unsigned nr_deactivate
, nr_activate
;
2007 unsigned nr_rotated
= 0;
2008 int file
= is_file_lru(lru
);
2009 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2013 spin_lock_irq(&pgdat
->lru_lock
);
2015 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2016 &nr_scanned
, sc
, lru
);
2018 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2019 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2021 __count_vm_events(PGREFILL
, nr_scanned
);
2022 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2024 spin_unlock_irq(&pgdat
->lru_lock
);
2026 while (!list_empty(&l_hold
)) {
2028 page
= lru_to_page(&l_hold
);
2029 list_del(&page
->lru
);
2031 if (unlikely(!page_evictable(page
))) {
2032 putback_lru_page(page
);
2036 if (unlikely(buffer_heads_over_limit
)) {
2037 if (page_has_private(page
) && trylock_page(page
)) {
2038 if (page_has_private(page
))
2039 try_to_release_page(page
, 0);
2044 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2046 nr_rotated
+= hpage_nr_pages(page
);
2048 * Identify referenced, file-backed active pages and
2049 * give them one more trip around the active list. So
2050 * that executable code get better chances to stay in
2051 * memory under moderate memory pressure. Anon pages
2052 * are not likely to be evicted by use-once streaming
2053 * IO, plus JVM can create lots of anon VM_EXEC pages,
2054 * so we ignore them here.
2056 if ((vm_flags
& VM_EXEC
) && page_is_file_lru(page
)) {
2057 list_add(&page
->lru
, &l_active
);
2062 ClearPageActive(page
); /* we are de-activating */
2063 SetPageWorkingset(page
);
2064 list_add(&page
->lru
, &l_inactive
);
2068 * Move pages back to the lru list.
2070 spin_lock_irq(&pgdat
->lru_lock
);
2072 * Count referenced pages from currently used mappings as rotated,
2073 * even though only some of them are actually re-activated. This
2074 * helps balance scan pressure between file and anonymous pages in
2077 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2079 nr_activate
= move_pages_to_lru(lruvec
, &l_active
);
2080 nr_deactivate
= move_pages_to_lru(lruvec
, &l_inactive
);
2081 /* Keep all free pages in l_active list */
2082 list_splice(&l_inactive
, &l_active
);
2084 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2085 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2087 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2088 spin_unlock_irq(&pgdat
->lru_lock
);
2090 mem_cgroup_uncharge_list(&l_active
);
2091 free_unref_page_list(&l_active
);
2092 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2093 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2096 unsigned long reclaim_pages(struct list_head
*page_list
)
2098 int nid
= NUMA_NO_NODE
;
2099 unsigned long nr_reclaimed
= 0;
2100 LIST_HEAD(node_page_list
);
2101 struct reclaim_stat dummy_stat
;
2103 struct scan_control sc
= {
2104 .gfp_mask
= GFP_KERNEL
,
2105 .priority
= DEF_PRIORITY
,
2111 while (!list_empty(page_list
)) {
2112 page
= lru_to_page(page_list
);
2113 if (nid
== NUMA_NO_NODE
) {
2114 nid
= page_to_nid(page
);
2115 INIT_LIST_HEAD(&node_page_list
);
2118 if (nid
== page_to_nid(page
)) {
2119 ClearPageActive(page
);
2120 list_move(&page
->lru
, &node_page_list
);
2124 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2127 &dummy_stat
, false);
2128 while (!list_empty(&node_page_list
)) {
2129 page
= lru_to_page(&node_page_list
);
2130 list_del(&page
->lru
);
2131 putback_lru_page(page
);
2137 if (!list_empty(&node_page_list
)) {
2138 nr_reclaimed
+= shrink_page_list(&node_page_list
,
2141 &dummy_stat
, false);
2142 while (!list_empty(&node_page_list
)) {
2143 page
= lru_to_page(&node_page_list
);
2144 list_del(&page
->lru
);
2145 putback_lru_page(page
);
2149 return nr_reclaimed
;
2152 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2153 struct lruvec
*lruvec
, struct scan_control
*sc
)
2155 if (is_active_lru(lru
)) {
2156 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2157 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2159 sc
->skipped_deactivate
= 1;
2163 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2167 * The inactive anon list should be small enough that the VM never has
2168 * to do too much work.
2170 * The inactive file list should be small enough to leave most memory
2171 * to the established workingset on the scan-resistant active list,
2172 * but large enough to avoid thrashing the aggregate readahead window.
2174 * Both inactive lists should also be large enough that each inactive
2175 * page has a chance to be referenced again before it is reclaimed.
2177 * If that fails and refaulting is observed, the inactive list grows.
2179 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2180 * on this LRU, maintained by the pageout code. An inactive_ratio
2181 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2184 * memory ratio inactive
2185 * -------------------------------------
2194 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2196 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2197 unsigned long inactive
, active
;
2198 unsigned long inactive_ratio
;
2201 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2202 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2204 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2206 inactive_ratio
= int_sqrt(10 * gb
);
2210 return inactive
* inactive_ratio
< active
;
2221 * Determine how aggressively the anon and file LRU lists should be
2222 * scanned. The relative value of each set of LRU lists is determined
2223 * by looking at the fraction of the pages scanned we did rotate back
2224 * onto the active list instead of evict.
2226 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2227 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2229 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2232 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2233 int swappiness
= mem_cgroup_swappiness(memcg
);
2234 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2236 u64 denominator
= 0; /* gcc */
2237 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2238 unsigned long anon_prio
, file_prio
;
2239 enum scan_balance scan_balance
;
2240 unsigned long anon
, file
;
2241 unsigned long ap
, fp
;
2244 /* If we have no swap space, do not bother scanning anon pages. */
2245 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2246 scan_balance
= SCAN_FILE
;
2251 * Global reclaim will swap to prevent OOM even with no
2252 * swappiness, but memcg users want to use this knob to
2253 * disable swapping for individual groups completely when
2254 * using the memory controller's swap limit feature would be
2257 if (cgroup_reclaim(sc
) && !swappiness
) {
2258 scan_balance
= SCAN_FILE
;
2263 * Do not apply any pressure balancing cleverness when the
2264 * system is close to OOM, scan both anon and file equally
2265 * (unless the swappiness setting disagrees with swapping).
2267 if (!sc
->priority
&& swappiness
) {
2268 scan_balance
= SCAN_EQUAL
;
2273 * If the system is almost out of file pages, force-scan anon.
2275 if (sc
->file_is_tiny
) {
2276 scan_balance
= SCAN_ANON
;
2281 * If there is enough inactive page cache, we do not reclaim
2282 * anything from the anonymous working right now.
2284 if (sc
->cache_trim_mode
) {
2285 scan_balance
= SCAN_FILE
;
2289 scan_balance
= SCAN_FRACT
;
2292 * With swappiness at 100, anonymous and file have the same priority.
2293 * This scanning priority is essentially the inverse of IO cost.
2295 anon_prio
= swappiness
;
2296 file_prio
= 200 - anon_prio
;
2299 * OK, so we have swap space and a fair amount of page cache
2300 * pages. We use the recently rotated / recently scanned
2301 * ratios to determine how valuable each cache is.
2303 * Because workloads change over time (and to avoid overflow)
2304 * we keep these statistics as a floating average, which ends
2305 * up weighing recent references more than old ones.
2307 * anon in [0], file in [1]
2310 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2311 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2312 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2313 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2315 spin_lock_irq(&pgdat
->lru_lock
);
2316 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2317 reclaim_stat
->recent_scanned
[0] /= 2;
2318 reclaim_stat
->recent_rotated
[0] /= 2;
2321 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2322 reclaim_stat
->recent_scanned
[1] /= 2;
2323 reclaim_stat
->recent_rotated
[1] /= 2;
2327 * The amount of pressure on anon vs file pages is inversely
2328 * proportional to the fraction of recently scanned pages on
2329 * each list that were recently referenced and in active use.
2331 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2332 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2334 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2335 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2336 spin_unlock_irq(&pgdat
->lru_lock
);
2340 denominator
= ap
+ fp
+ 1;
2342 for_each_evictable_lru(lru
) {
2343 int file
= is_file_lru(lru
);
2344 unsigned long lruvec_size
;
2346 unsigned long protection
;
2348 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2349 protection
= mem_cgroup_protection(memcg
,
2350 sc
->memcg_low_reclaim
);
2354 * Scale a cgroup's reclaim pressure by proportioning
2355 * its current usage to its memory.low or memory.min
2358 * This is important, as otherwise scanning aggression
2359 * becomes extremely binary -- from nothing as we
2360 * approach the memory protection threshold, to totally
2361 * nominal as we exceed it. This results in requiring
2362 * setting extremely liberal protection thresholds. It
2363 * also means we simply get no protection at all if we
2364 * set it too low, which is not ideal.
2366 * If there is any protection in place, we reduce scan
2367 * pressure by how much of the total memory used is
2368 * within protection thresholds.
2370 * There is one special case: in the first reclaim pass,
2371 * we skip over all groups that are within their low
2372 * protection. If that fails to reclaim enough pages to
2373 * satisfy the reclaim goal, we come back and override
2374 * the best-effort low protection. However, we still
2375 * ideally want to honor how well-behaved groups are in
2376 * that case instead of simply punishing them all
2377 * equally. As such, we reclaim them based on how much
2378 * memory they are using, reducing the scan pressure
2379 * again by how much of the total memory used is under
2382 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2384 /* Avoid TOCTOU with earlier protection check */
2385 cgroup_size
= max(cgroup_size
, protection
);
2387 scan
= lruvec_size
- lruvec_size
* protection
/
2391 * Minimally target SWAP_CLUSTER_MAX pages to keep
2392 * reclaim moving forwards, avoiding decremeting
2393 * sc->priority further than desirable.
2395 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2400 scan
>>= sc
->priority
;
2403 * If the cgroup's already been deleted, make sure to
2404 * scrape out the remaining cache.
2406 if (!scan
&& !mem_cgroup_online(memcg
))
2407 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2409 switch (scan_balance
) {
2411 /* Scan lists relative to size */
2415 * Scan types proportional to swappiness and
2416 * their relative recent reclaim efficiency.
2417 * Make sure we don't miss the last page on
2418 * the offlined memory cgroups because of a
2421 scan
= mem_cgroup_online(memcg
) ?
2422 div64_u64(scan
* fraction
[file
], denominator
) :
2423 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2428 /* Scan one type exclusively */
2429 if ((scan_balance
== SCAN_FILE
) != file
)
2433 /* Look ma, no brain */
2441 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
2443 unsigned long nr
[NR_LRU_LISTS
];
2444 unsigned long targets
[NR_LRU_LISTS
];
2445 unsigned long nr_to_scan
;
2447 unsigned long nr_reclaimed
= 0;
2448 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2449 struct blk_plug plug
;
2452 get_scan_count(lruvec
, sc
, nr
);
2454 /* Record the original scan target for proportional adjustments later */
2455 memcpy(targets
, nr
, sizeof(nr
));
2458 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2459 * event that can occur when there is little memory pressure e.g.
2460 * multiple streaming readers/writers. Hence, we do not abort scanning
2461 * when the requested number of pages are reclaimed when scanning at
2462 * DEF_PRIORITY on the assumption that the fact we are direct
2463 * reclaiming implies that kswapd is not keeping up and it is best to
2464 * do a batch of work at once. For memcg reclaim one check is made to
2465 * abort proportional reclaim if either the file or anon lru has already
2466 * dropped to zero at the first pass.
2468 scan_adjusted
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
2469 sc
->priority
== DEF_PRIORITY
);
2471 blk_start_plug(&plug
);
2472 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2473 nr
[LRU_INACTIVE_FILE
]) {
2474 unsigned long nr_anon
, nr_file
, percentage
;
2475 unsigned long nr_scanned
;
2477 for_each_evictable_lru(lru
) {
2479 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2480 nr
[lru
] -= nr_to_scan
;
2482 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2489 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2493 * For kswapd and memcg, reclaim at least the number of pages
2494 * requested. Ensure that the anon and file LRUs are scanned
2495 * proportionally what was requested by get_scan_count(). We
2496 * stop reclaiming one LRU and reduce the amount scanning
2497 * proportional to the original scan target.
2499 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2500 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2503 * It's just vindictive to attack the larger once the smaller
2504 * has gone to zero. And given the way we stop scanning the
2505 * smaller below, this makes sure that we only make one nudge
2506 * towards proportionality once we've got nr_to_reclaim.
2508 if (!nr_file
|| !nr_anon
)
2511 if (nr_file
> nr_anon
) {
2512 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2513 targets
[LRU_ACTIVE_ANON
] + 1;
2515 percentage
= nr_anon
* 100 / scan_target
;
2517 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2518 targets
[LRU_ACTIVE_FILE
] + 1;
2520 percentage
= nr_file
* 100 / scan_target
;
2523 /* Stop scanning the smaller of the LRU */
2525 nr
[lru
+ LRU_ACTIVE
] = 0;
2528 * Recalculate the other LRU scan count based on its original
2529 * scan target and the percentage scanning already complete
2531 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2532 nr_scanned
= targets
[lru
] - nr
[lru
];
2533 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2534 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2537 nr_scanned
= targets
[lru
] - nr
[lru
];
2538 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2539 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2541 scan_adjusted
= true;
2543 blk_finish_plug(&plug
);
2544 sc
->nr_reclaimed
+= nr_reclaimed
;
2547 * Even if we did not try to evict anon pages at all, we want to
2548 * rebalance the anon lru active/inactive ratio.
2550 if (total_swap_pages
&& inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
2551 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2552 sc
, LRU_ACTIVE_ANON
);
2555 /* Use reclaim/compaction for costly allocs or under memory pressure */
2556 static bool in_reclaim_compaction(struct scan_control
*sc
)
2558 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2559 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2560 sc
->priority
< DEF_PRIORITY
- 2))
2567 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2568 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2569 * true if more pages should be reclaimed such that when the page allocator
2570 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2571 * It will give up earlier than that if there is difficulty reclaiming pages.
2573 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2574 unsigned long nr_reclaimed
,
2575 struct scan_control
*sc
)
2577 unsigned long pages_for_compaction
;
2578 unsigned long inactive_lru_pages
;
2581 /* If not in reclaim/compaction mode, stop */
2582 if (!in_reclaim_compaction(sc
))
2586 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2587 * number of pages that were scanned. This will return to the caller
2588 * with the risk reclaim/compaction and the resulting allocation attempt
2589 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2590 * allocations through requiring that the full LRU list has been scanned
2591 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2592 * scan, but that approximation was wrong, and there were corner cases
2593 * where always a non-zero amount of pages were scanned.
2598 /* If compaction would go ahead or the allocation would succeed, stop */
2599 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2600 struct zone
*zone
= &pgdat
->node_zones
[z
];
2601 if (!managed_zone(zone
))
2604 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2605 case COMPACT_SUCCESS
:
2606 case COMPACT_CONTINUE
:
2609 /* check next zone */
2615 * If we have not reclaimed enough pages for compaction and the
2616 * inactive lists are large enough, continue reclaiming
2618 pages_for_compaction
= compact_gap(sc
->order
);
2619 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2620 if (get_nr_swap_pages() > 0)
2621 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2623 return inactive_lru_pages
> pages_for_compaction
;
2626 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
2628 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
2629 struct mem_cgroup
*memcg
;
2631 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
2633 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
2634 unsigned long reclaimed
;
2635 unsigned long scanned
;
2637 switch (mem_cgroup_protected(target_memcg
, memcg
)) {
2638 case MEMCG_PROT_MIN
:
2641 * If there is no reclaimable memory, OOM.
2644 case MEMCG_PROT_LOW
:
2647 * Respect the protection only as long as
2648 * there is an unprotected supply
2649 * of reclaimable memory from other cgroups.
2651 if (!sc
->memcg_low_reclaim
) {
2652 sc
->memcg_low_skipped
= 1;
2655 memcg_memory_event(memcg
, MEMCG_LOW
);
2657 case MEMCG_PROT_NONE
:
2659 * All protection thresholds breached. We may
2660 * still choose to vary the scan pressure
2661 * applied based on by how much the cgroup in
2662 * question has exceeded its protection
2663 * thresholds (see get_scan_count).
2668 reclaimed
= sc
->nr_reclaimed
;
2669 scanned
= sc
->nr_scanned
;
2671 shrink_lruvec(lruvec
, sc
);
2673 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
2676 /* Record the group's reclaim efficiency */
2677 vmpressure(sc
->gfp_mask
, memcg
, false,
2678 sc
->nr_scanned
- scanned
,
2679 sc
->nr_reclaimed
- reclaimed
);
2681 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
2684 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2686 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2687 unsigned long nr_reclaimed
, nr_scanned
;
2688 struct lruvec
*target_lruvec
;
2689 bool reclaimable
= false;
2692 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
2695 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2697 nr_reclaimed
= sc
->nr_reclaimed
;
2698 nr_scanned
= sc
->nr_scanned
;
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 if (inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
2708 sc
->may_deactivate
|= DEACTIVATE_ANON
;
2710 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
2713 * When refaults are being observed, it means a new
2714 * workingset is being established. Deactivate to get
2715 * rid of any stale active pages quickly.
2717 refaults
= lruvec_page_state(target_lruvec
,
2718 WORKINGSET_ACTIVATE
);
2719 if (refaults
!= target_lruvec
->refaults
||
2720 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
2721 sc
->may_deactivate
|= DEACTIVATE_FILE
;
2723 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
2725 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
2728 * If we have plenty of inactive file pages that aren't
2729 * thrashing, try to reclaim those first before touching
2732 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
2733 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
2734 sc
->cache_trim_mode
= 1;
2736 sc
->cache_trim_mode
= 0;
2739 * Prevent the reclaimer from falling into the cache trap: as
2740 * cache pages start out inactive, every cache fault will tip
2741 * the scan balance towards the file LRU. And as the file LRU
2742 * shrinks, so does the window for rotation from references.
2743 * This means we have a runaway feedback loop where a tiny
2744 * thrashing file LRU becomes infinitely more attractive than
2745 * anon pages. Try to detect this based on file LRU size.
2747 if (!cgroup_reclaim(sc
)) {
2748 unsigned long total_high_wmark
= 0;
2749 unsigned long free
, anon
;
2752 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2753 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2754 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2756 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2757 struct zone
*zone
= &pgdat
->node_zones
[z
];
2758 if (!managed_zone(zone
))
2761 total_high_wmark
+= high_wmark_pages(zone
);
2765 * Consider anon: if that's low too, this isn't a
2766 * runaway file reclaim problem, but rather just
2767 * extreme pressure. Reclaim as per usual then.
2769 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2772 file
+ free
<= total_high_wmark
&&
2773 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
2774 anon
>> sc
->priority
;
2777 shrink_node_memcgs(pgdat
, sc
);
2779 if (reclaim_state
) {
2780 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2781 reclaim_state
->reclaimed_slab
= 0;
2784 /* Record the subtree's reclaim efficiency */
2785 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2786 sc
->nr_scanned
- nr_scanned
,
2787 sc
->nr_reclaimed
- nr_reclaimed
);
2789 if (sc
->nr_reclaimed
- nr_reclaimed
)
2792 if (current_is_kswapd()) {
2794 * If reclaim is isolating dirty pages under writeback,
2795 * it implies that the long-lived page allocation rate
2796 * is exceeding the page laundering rate. Either the
2797 * global limits are not being effective at throttling
2798 * processes due to the page distribution throughout
2799 * zones or there is heavy usage of a slow backing
2800 * device. The only option is to throttle from reclaim
2801 * context which is not ideal as there is no guarantee
2802 * the dirtying process is throttled in the same way
2803 * balance_dirty_pages() manages.
2805 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2806 * count the number of pages under pages flagged for
2807 * immediate reclaim and stall if any are encountered
2808 * in the nr_immediate check below.
2810 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2811 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2813 /* Allow kswapd to start writing pages during reclaim.*/
2814 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2815 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2818 * If kswapd scans pages marked marked for immediate
2819 * reclaim and under writeback (nr_immediate), it
2820 * implies that pages are cycling through the LRU
2821 * faster than they are written so also forcibly stall.
2823 if (sc
->nr
.immediate
)
2824 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2828 * Tag a node/memcg as congested if all the dirty pages
2829 * scanned were backed by a congested BDI and
2830 * wait_iff_congested will stall.
2832 * Legacy memcg will stall in page writeback so avoid forcibly
2833 * stalling in wait_iff_congested().
2835 if ((current_is_kswapd() ||
2836 (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))) &&
2837 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2838 set_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
);
2841 * Stall direct reclaim for IO completions if underlying BDIs
2842 * and node is congested. Allow kswapd to continue until it
2843 * starts encountering unqueued dirty pages or cycling through
2844 * the LRU too quickly.
2846 if (!current_is_kswapd() && current_may_throttle() &&
2847 !sc
->hibernation_mode
&&
2848 test_bit(LRUVEC_CONGESTED
, &target_lruvec
->flags
))
2849 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2851 if (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2856 * Kswapd gives up on balancing particular nodes after too
2857 * many failures to reclaim anything from them and goes to
2858 * sleep. On reclaim progress, reset the failure counter. A
2859 * successful direct reclaim run will revive a dormant kswapd.
2862 pgdat
->kswapd_failures
= 0;
2866 * Returns true if compaction should go ahead for a costly-order request, or
2867 * the allocation would already succeed without compaction. Return false if we
2868 * should reclaim first.
2870 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2872 unsigned long watermark
;
2873 enum compact_result suitable
;
2875 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2876 if (suitable
== COMPACT_SUCCESS
)
2877 /* Allocation should succeed already. Don't reclaim. */
2879 if (suitable
== COMPACT_SKIPPED
)
2880 /* Compaction cannot yet proceed. Do reclaim. */
2884 * Compaction is already possible, but it takes time to run and there
2885 * are potentially other callers using the pages just freed. So proceed
2886 * with reclaim to make a buffer of free pages available to give
2887 * compaction a reasonable chance of completing and allocating the page.
2888 * Note that we won't actually reclaim the whole buffer in one attempt
2889 * as the target watermark in should_continue_reclaim() is lower. But if
2890 * we are already above the high+gap watermark, don't reclaim at all.
2892 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2894 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2898 * This is the direct reclaim path, for page-allocating processes. We only
2899 * try to reclaim pages from zones which will satisfy the caller's allocation
2902 * If a zone is deemed to be full of pinned pages then just give it a light
2903 * scan then give up on it.
2905 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2909 unsigned long nr_soft_reclaimed
;
2910 unsigned long nr_soft_scanned
;
2912 pg_data_t
*last_pgdat
= NULL
;
2915 * If the number of buffer_heads in the machine exceeds the maximum
2916 * allowed level, force direct reclaim to scan the highmem zone as
2917 * highmem pages could be pinning lowmem pages storing buffer_heads
2919 orig_mask
= sc
->gfp_mask
;
2920 if (buffer_heads_over_limit
) {
2921 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2922 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2925 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2926 sc
->reclaim_idx
, sc
->nodemask
) {
2928 * Take care memory controller reclaiming has small influence
2931 if (!cgroup_reclaim(sc
)) {
2932 if (!cpuset_zone_allowed(zone
,
2933 GFP_KERNEL
| __GFP_HARDWALL
))
2937 * If we already have plenty of memory free for
2938 * compaction in this zone, don't free any more.
2939 * Even though compaction is invoked for any
2940 * non-zero order, only frequent costly order
2941 * reclamation is disruptive enough to become a
2942 * noticeable problem, like transparent huge
2945 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2946 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2947 compaction_ready(zone
, sc
)) {
2948 sc
->compaction_ready
= true;
2953 * Shrink each node in the zonelist once. If the
2954 * zonelist is ordered by zone (not the default) then a
2955 * node may be shrunk multiple times but in that case
2956 * the user prefers lower zones being preserved.
2958 if (zone
->zone_pgdat
== last_pgdat
)
2962 * This steals pages from memory cgroups over softlimit
2963 * and returns the number of reclaimed pages and
2964 * scanned pages. This works for global memory pressure
2965 * and balancing, not for a memcg's limit.
2967 nr_soft_scanned
= 0;
2968 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2969 sc
->order
, sc
->gfp_mask
,
2971 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2972 sc
->nr_scanned
+= nr_soft_scanned
;
2973 /* need some check for avoid more shrink_zone() */
2976 /* See comment about same check for global reclaim above */
2977 if (zone
->zone_pgdat
== last_pgdat
)
2979 last_pgdat
= zone
->zone_pgdat
;
2980 shrink_node(zone
->zone_pgdat
, sc
);
2984 * Restore to original mask to avoid the impact on the caller if we
2985 * promoted it to __GFP_HIGHMEM.
2987 sc
->gfp_mask
= orig_mask
;
2990 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
2992 struct lruvec
*target_lruvec
;
2993 unsigned long refaults
;
2995 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
2996 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE
);
2997 target_lruvec
->refaults
= refaults
;
3001 * This is the main entry point to direct page reclaim.
3003 * If a full scan of the inactive list fails to free enough memory then we
3004 * are "out of memory" and something needs to be killed.
3006 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3007 * high - the zone may be full of dirty or under-writeback pages, which this
3008 * caller can't do much about. We kick the writeback threads and take explicit
3009 * naps in the hope that some of these pages can be written. But if the
3010 * allocating task holds filesystem locks which prevent writeout this might not
3011 * work, and the allocation attempt will fail.
3013 * returns: 0, if no pages reclaimed
3014 * else, the number of pages reclaimed
3016 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3017 struct scan_control
*sc
)
3019 int initial_priority
= sc
->priority
;
3020 pg_data_t
*last_pgdat
;
3024 delayacct_freepages_start();
3026 if (!cgroup_reclaim(sc
))
3027 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3030 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3033 shrink_zones(zonelist
, sc
);
3035 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3038 if (sc
->compaction_ready
)
3042 * If we're getting trouble reclaiming, start doing
3043 * writepage even in laptop mode.
3045 if (sc
->priority
< DEF_PRIORITY
- 2)
3046 sc
->may_writepage
= 1;
3047 } while (--sc
->priority
>= 0);
3050 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3052 if (zone
->zone_pgdat
== last_pgdat
)
3054 last_pgdat
= zone
->zone_pgdat
;
3056 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3058 if (cgroup_reclaim(sc
)) {
3059 struct lruvec
*lruvec
;
3061 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
3063 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3067 delayacct_freepages_end();
3069 if (sc
->nr_reclaimed
)
3070 return sc
->nr_reclaimed
;
3072 /* Aborted reclaim to try compaction? don't OOM, then */
3073 if (sc
->compaction_ready
)
3077 * We make inactive:active ratio decisions based on the node's
3078 * composition of memory, but a restrictive reclaim_idx or a
3079 * memory.low cgroup setting can exempt large amounts of
3080 * memory from reclaim. Neither of which are very common, so
3081 * instead of doing costly eligibility calculations of the
3082 * entire cgroup subtree up front, we assume the estimates are
3083 * good, and retry with forcible deactivation if that fails.
3085 if (sc
->skipped_deactivate
) {
3086 sc
->priority
= initial_priority
;
3087 sc
->force_deactivate
= 1;
3088 sc
->skipped_deactivate
= 0;
3092 /* Untapped cgroup reserves? Don't OOM, retry. */
3093 if (sc
->memcg_low_skipped
) {
3094 sc
->priority
= initial_priority
;
3095 sc
->force_deactivate
= 0;
3096 sc
->memcg_low_reclaim
= 1;
3097 sc
->memcg_low_skipped
= 0;
3104 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3107 unsigned long pfmemalloc_reserve
= 0;
3108 unsigned long free_pages
= 0;
3112 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3115 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3116 zone
= &pgdat
->node_zones
[i
];
3117 if (!managed_zone(zone
))
3120 if (!zone_reclaimable_pages(zone
))
3123 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3124 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3127 /* If there are no reserves (unexpected config) then do not throttle */
3128 if (!pfmemalloc_reserve
)
3131 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3133 /* kswapd must be awake if processes are being throttled */
3134 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3135 if (READ_ONCE(pgdat
->kswapd_classzone_idx
) > ZONE_NORMAL
)
3136 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, ZONE_NORMAL
);
3138 wake_up_interruptible(&pgdat
->kswapd_wait
);
3145 * Throttle direct reclaimers if backing storage is backed by the network
3146 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3147 * depleted. kswapd will continue to make progress and wake the processes
3148 * when the low watermark is reached.
3150 * Returns true if a fatal signal was delivered during throttling. If this
3151 * happens, the page allocator should not consider triggering the OOM killer.
3153 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3154 nodemask_t
*nodemask
)
3158 pg_data_t
*pgdat
= NULL
;
3161 * Kernel threads should not be throttled as they may be indirectly
3162 * responsible for cleaning pages necessary for reclaim to make forward
3163 * progress. kjournald for example may enter direct reclaim while
3164 * committing a transaction where throttling it could forcing other
3165 * processes to block on log_wait_commit().
3167 if (current
->flags
& PF_KTHREAD
)
3171 * If a fatal signal is pending, this process should not throttle.
3172 * It should return quickly so it can exit and free its memory
3174 if (fatal_signal_pending(current
))
3178 * Check if the pfmemalloc reserves are ok by finding the first node
3179 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3180 * GFP_KERNEL will be required for allocating network buffers when
3181 * swapping over the network so ZONE_HIGHMEM is unusable.
3183 * Throttling is based on the first usable node and throttled processes
3184 * wait on a queue until kswapd makes progress and wakes them. There
3185 * is an affinity then between processes waking up and where reclaim
3186 * progress has been made assuming the process wakes on the same node.
3187 * More importantly, processes running on remote nodes will not compete
3188 * for remote pfmemalloc reserves and processes on different nodes
3189 * should make reasonable progress.
3191 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3192 gfp_zone(gfp_mask
), nodemask
) {
3193 if (zone_idx(zone
) > ZONE_NORMAL
)
3196 /* Throttle based on the first usable node */
3197 pgdat
= zone
->zone_pgdat
;
3198 if (allow_direct_reclaim(pgdat
))
3203 /* If no zone was usable by the allocation flags then do not throttle */
3207 /* Account for the throttling */
3208 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3211 * If the caller cannot enter the filesystem, it's possible that it
3212 * is due to the caller holding an FS lock or performing a journal
3213 * transaction in the case of a filesystem like ext[3|4]. In this case,
3214 * it is not safe to block on pfmemalloc_wait as kswapd could be
3215 * blocked waiting on the same lock. Instead, throttle for up to a
3216 * second before continuing.
3218 if (!(gfp_mask
& __GFP_FS
)) {
3219 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3220 allow_direct_reclaim(pgdat
), HZ
);
3225 /* Throttle until kswapd wakes the process */
3226 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3227 allow_direct_reclaim(pgdat
));
3230 if (fatal_signal_pending(current
))
3237 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3238 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3240 unsigned long nr_reclaimed
;
3241 struct scan_control sc
= {
3242 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3243 .gfp_mask
= current_gfp_context(gfp_mask
),
3244 .reclaim_idx
= gfp_zone(gfp_mask
),
3246 .nodemask
= nodemask
,
3247 .priority
= DEF_PRIORITY
,
3248 .may_writepage
= !laptop_mode
,
3254 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3255 * Confirm they are large enough for max values.
3257 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3258 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3259 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3262 * Do not enter reclaim if fatal signal was delivered while throttled.
3263 * 1 is returned so that the page allocator does not OOM kill at this
3266 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3269 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3270 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
3272 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3274 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3275 set_task_reclaim_state(current
, NULL
);
3277 return nr_reclaimed
;
3282 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3283 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3284 gfp_t gfp_mask
, bool noswap
,
3286 unsigned long *nr_scanned
)
3288 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3289 struct scan_control sc
= {
3290 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3291 .target_mem_cgroup
= memcg
,
3292 .may_writepage
= !laptop_mode
,
3294 .reclaim_idx
= MAX_NR_ZONES
- 1,
3295 .may_swap
= !noswap
,
3298 WARN_ON_ONCE(!current
->reclaim_state
);
3300 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3301 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3303 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3307 * NOTE: Although we can get the priority field, using it
3308 * here is not a good idea, since it limits the pages we can scan.
3309 * if we don't reclaim here, the shrink_node from balance_pgdat
3310 * will pick up pages from other mem cgroup's as well. We hack
3311 * the priority and make it zero.
3313 shrink_lruvec(lruvec
, &sc
);
3315 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3317 *nr_scanned
= sc
.nr_scanned
;
3319 return sc
.nr_reclaimed
;
3322 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3323 unsigned long nr_pages
,
3327 unsigned long nr_reclaimed
;
3328 unsigned long pflags
;
3329 unsigned int noreclaim_flag
;
3330 struct scan_control sc
= {
3331 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3332 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3333 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3334 .reclaim_idx
= MAX_NR_ZONES
- 1,
3335 .target_mem_cgroup
= memcg
,
3336 .priority
= DEF_PRIORITY
,
3337 .may_writepage
= !laptop_mode
,
3339 .may_swap
= may_swap
,
3342 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3343 * equal pressure on all the nodes. This is based on the assumption that
3344 * the reclaim does not bail out early.
3346 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3348 set_task_reclaim_state(current
, &sc
.reclaim_state
);
3350 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
3352 psi_memstall_enter(&pflags
);
3353 noreclaim_flag
= memalloc_noreclaim_save();
3355 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3357 memalloc_noreclaim_restore(noreclaim_flag
);
3358 psi_memstall_leave(&pflags
);
3360 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3361 set_task_reclaim_state(current
, NULL
);
3363 return nr_reclaimed
;
3367 static void age_active_anon(struct pglist_data
*pgdat
,
3368 struct scan_control
*sc
)
3370 struct mem_cgroup
*memcg
;
3371 struct lruvec
*lruvec
;
3373 if (!total_swap_pages
)
3376 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3377 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
3380 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3382 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3383 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3384 sc
, LRU_ACTIVE_ANON
);
3385 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3389 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int classzone_idx
)
3395 * Check for watermark boosts top-down as the higher zones
3396 * are more likely to be boosted. Both watermarks and boosts
3397 * should not be checked at the time time as reclaim would
3398 * start prematurely when there is no boosting and a lower
3401 for (i
= classzone_idx
; i
>= 0; i
--) {
3402 zone
= pgdat
->node_zones
+ i
;
3403 if (!managed_zone(zone
))
3406 if (zone
->watermark_boost
)
3414 * Returns true if there is an eligible zone balanced for the request order
3417 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3420 unsigned long mark
= -1;
3424 * Check watermarks bottom-up as lower zones are more likely to
3427 for (i
= 0; i
<= classzone_idx
; i
++) {
3428 zone
= pgdat
->node_zones
+ i
;
3430 if (!managed_zone(zone
))
3433 mark
= high_wmark_pages(zone
);
3434 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3439 * If a node has no populated zone within classzone_idx, it does not
3440 * need balancing by definition. This can happen if a zone-restricted
3441 * allocation tries to wake a remote kswapd.
3449 /* Clear pgdat state for congested, dirty or under writeback. */
3450 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3452 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
3454 clear_bit(LRUVEC_CONGESTED
, &lruvec
->flags
);
3455 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3456 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3460 * Prepare kswapd for sleeping. This verifies that there are no processes
3461 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3463 * Returns true if kswapd is ready to sleep
3465 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
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
, classzone_idx
)) {
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 classzone_idx
)
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
<= classzone_idx
; 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
= classzone_idx
;
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
, classzone_idx
);
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
<= classzone_idx
; 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
, classzone_idx
);
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_classzone_idx is used to pass the highest zone index to be
3761 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3762 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3763 * after previous reclaim attempt (node is still unbalanced). In that case
3764 * return the zone index of the previous kswapd reclaim cycle.
3766 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3767 enum zone_type prev_classzone_idx
)
3769 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
3771 return curr_idx
== MAX_NR_ZONES
? prev_classzone_idx
: curr_idx
;
3774 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3775 unsigned int classzone_idx
)
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
, classzone_idx
)) {
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
, classzone_idx
);
3807 remaining
= schedule_timeout(HZ
/10);
3810 * If woken prematurely then reset kswapd_classzone_idx 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_classzone_idx
,
3816 kswapd_classzone_idx(pgdat
, classzone_idx
));
3818 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
3819 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
3822 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3823 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3827 * After a short sleep, check if it was a premature sleep. If not, then
3828 * go fully to sleep until explicitly woken up.
3831 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3832 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3835 * vmstat counters are not perfectly accurate and the estimated
3836 * value for counters such as NR_FREE_PAGES can deviate from the
3837 * true value by nr_online_cpus * threshold. To avoid the zone
3838 * watermarks being breached while under pressure, we reduce the
3839 * per-cpu vmstat threshold while kswapd is awake and restore
3840 * them before going back to sleep.
3842 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3844 if (!kthread_should_stop())
3847 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3850 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3852 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3854 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3858 * The background pageout daemon, started as a kernel thread
3859 * from the init process.
3861 * This basically trickles out pages so that we have _some_
3862 * free memory available even if there is no other activity
3863 * that frees anything up. This is needed for things like routing
3864 * etc, where we otherwise might have all activity going on in
3865 * asynchronous contexts that cannot page things out.
3867 * If there are applications that are active memory-allocators
3868 * (most normal use), this basically shouldn't matter.
3870 static int kswapd(void *p
)
3872 unsigned int alloc_order
, reclaim_order
;
3873 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3874 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3875 struct task_struct
*tsk
= current
;
3876 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3878 if (!cpumask_empty(cpumask
))
3879 set_cpus_allowed_ptr(tsk
, cpumask
);
3882 * Tell the memory management that we're a "memory allocator",
3883 * and that if we need more memory we should get access to it
3884 * regardless (see "__alloc_pages()"). "kswapd" should
3885 * never get caught in the normal page freeing logic.
3887 * (Kswapd normally doesn't need memory anyway, but sometimes
3888 * you need a small amount of memory in order to be able to
3889 * page out something else, and this flag essentially protects
3890 * us from recursively trying to free more memory as we're
3891 * trying to free the first piece of memory in the first place).
3893 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3896 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3897 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3901 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3902 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3905 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3908 /* Read the new order and classzone_idx */
3909 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
3910 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3911 WRITE_ONCE(pgdat
->kswapd_order
, 0);
3912 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, MAX_NR_ZONES
);
3914 ret
= try_to_freeze();
3915 if (kthread_should_stop())
3919 * We can speed up thawing tasks if we don't call balance_pgdat
3920 * after returning from the refrigerator
3926 * Reclaim begins at the requested order but if a high-order
3927 * reclaim fails then kswapd falls back to reclaiming for
3928 * order-0. If that happens, kswapd will consider sleeping
3929 * for the order it finished reclaiming at (reclaim_order)
3930 * but kcompactd is woken to compact for the original
3931 * request (alloc_order).
3933 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3935 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3936 if (reclaim_order
< alloc_order
)
3937 goto kswapd_try_sleep
;
3940 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3946 * A zone is low on free memory or too fragmented for high-order memory. If
3947 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3948 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3949 * has failed or is not needed, still wake up kcompactd if only compaction is
3952 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3953 enum zone_type classzone_idx
)
3956 enum zone_type curr_idx
;
3958 if (!managed_zone(zone
))
3961 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3964 pgdat
= zone
->zone_pgdat
;
3965 curr_idx
= READ_ONCE(pgdat
->kswapd_classzone_idx
);
3967 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< classzone_idx
)
3968 WRITE_ONCE(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3970 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
3971 WRITE_ONCE(pgdat
->kswapd_order
, order
);
3973 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3976 /* Hopeless node, leave it to direct reclaim if possible */
3977 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3978 (pgdat_balanced(pgdat
, order
, classzone_idx
) &&
3979 !pgdat_watermark_boosted(pgdat
, classzone_idx
))) {
3981 * There may be plenty of free memory available, but it's too
3982 * fragmented for high-order allocations. Wake up kcompactd
3983 * and rely on compaction_suitable() to determine if it's
3984 * needed. If it fails, it will defer subsequent attempts to
3985 * ratelimit its work.
3987 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3988 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3992 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
3994 wake_up_interruptible(&pgdat
->kswapd_wait
);
3997 #ifdef CONFIG_HIBERNATION
3999 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4002 * Rather than trying to age LRUs the aim is to preserve the overall
4003 * LRU order by reclaiming preferentially
4004 * inactive > active > active referenced > active mapped
4006 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
4008 struct scan_control sc
= {
4009 .nr_to_reclaim
= nr_to_reclaim
,
4010 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
4011 .reclaim_idx
= MAX_NR_ZONES
- 1,
4012 .priority
= DEF_PRIORITY
,
4016 .hibernation_mode
= 1,
4018 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
4019 unsigned long nr_reclaimed
;
4020 unsigned int noreclaim_flag
;
4022 fs_reclaim_acquire(sc
.gfp_mask
);
4023 noreclaim_flag
= memalloc_noreclaim_save();
4024 set_task_reclaim_state(current
, &sc
.reclaim_state
);
4026 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
4028 set_task_reclaim_state(current
, NULL
);
4029 memalloc_noreclaim_restore(noreclaim_flag
);
4030 fs_reclaim_release(sc
.gfp_mask
);
4032 return nr_reclaimed
;
4034 #endif /* CONFIG_HIBERNATION */
4037 * This kswapd start function will be called by init and node-hot-add.
4038 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4040 int kswapd_run(int nid
)
4042 pg_data_t
*pgdat
= NODE_DATA(nid
);
4048 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
4049 if (IS_ERR(pgdat
->kswapd
)) {
4050 /* failure at boot is fatal */
4051 BUG_ON(system_state
< SYSTEM_RUNNING
);
4052 pr_err("Failed to start kswapd on node %d\n", nid
);
4053 ret
= PTR_ERR(pgdat
->kswapd
);
4054 pgdat
->kswapd
= NULL
;
4060 * Called by memory hotplug when all memory in a node is offlined. Caller must
4061 * hold mem_hotplug_begin/end().
4063 void kswapd_stop(int nid
)
4065 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
4068 kthread_stop(kswapd
);
4069 NODE_DATA(nid
)->kswapd
= NULL
;
4073 static int __init
kswapd_init(void)
4078 for_each_node_state(nid
, N_MEMORY
)
4083 module_init(kswapd_init
)
4089 * If non-zero call node_reclaim when the number of free pages falls below
4092 int node_reclaim_mode __read_mostly
;
4094 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4095 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4098 * Priority for NODE_RECLAIM. This determines the fraction of pages
4099 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4102 #define NODE_RECLAIM_PRIORITY 4
4105 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4108 int sysctl_min_unmapped_ratio
= 1;
4111 * If the number of slab pages in a zone grows beyond this percentage then
4112 * slab reclaim needs to occur.
4114 int sysctl_min_slab_ratio
= 5;
4116 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4118 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4119 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4120 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4123 * It's possible for there to be more file mapped pages than
4124 * accounted for by the pages on the file LRU lists because
4125 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4127 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4130 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4131 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4133 unsigned long nr_pagecache_reclaimable
;
4134 unsigned long delta
= 0;
4137 * If RECLAIM_UNMAP is set, then all file pages are considered
4138 * potentially reclaimable. Otherwise, we have to worry about
4139 * pages like swapcache and node_unmapped_file_pages() provides
4142 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4143 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4145 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4147 /* If we can't clean pages, remove dirty pages from consideration */
4148 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4149 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4151 /* Watch for any possible underflows due to delta */
4152 if (unlikely(delta
> nr_pagecache_reclaimable
))
4153 delta
= nr_pagecache_reclaimable
;
4155 return nr_pagecache_reclaimable
- delta
;
4159 * Try to free up some pages from this node through reclaim.
4161 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4163 /* Minimum pages needed in order to stay on node */
4164 const unsigned long nr_pages
= 1 << order
;
4165 struct task_struct
*p
= current
;
4166 unsigned int noreclaim_flag
;
4167 struct scan_control sc
= {
4168 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4169 .gfp_mask
= current_gfp_context(gfp_mask
),
4171 .priority
= NODE_RECLAIM_PRIORITY
,
4172 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4173 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4175 .reclaim_idx
= gfp_zone(gfp_mask
),
4178 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
4182 fs_reclaim_acquire(sc
.gfp_mask
);
4184 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4185 * and we also need to be able to write out pages for RECLAIM_WRITE
4186 * and RECLAIM_UNMAP.
4188 noreclaim_flag
= memalloc_noreclaim_save();
4189 p
->flags
|= PF_SWAPWRITE
;
4190 set_task_reclaim_state(p
, &sc
.reclaim_state
);
4192 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4194 * Free memory by calling shrink node with increasing
4195 * priorities until we have enough memory freed.
4198 shrink_node(pgdat
, &sc
);
4199 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4202 set_task_reclaim_state(p
, NULL
);
4203 current
->flags
&= ~PF_SWAPWRITE
;
4204 memalloc_noreclaim_restore(noreclaim_flag
);
4205 fs_reclaim_release(sc
.gfp_mask
);
4207 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
4209 return sc
.nr_reclaimed
>= nr_pages
;
4212 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4217 * Node reclaim reclaims unmapped file backed pages and
4218 * slab pages if we are over the defined limits.
4220 * A small portion of unmapped file backed pages is needed for
4221 * file I/O otherwise pages read by file I/O will be immediately
4222 * thrown out if the node is overallocated. So we do not reclaim
4223 * if less than a specified percentage of the node is used by
4224 * unmapped file backed pages.
4226 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4227 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4228 return NODE_RECLAIM_FULL
;
4231 * Do not scan if the allocation should not be delayed.
4233 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4234 return NODE_RECLAIM_NOSCAN
;
4237 * Only run node reclaim on the local node or on nodes that do not
4238 * have associated processors. This will favor the local processor
4239 * over remote processors and spread off node memory allocations
4240 * as wide as possible.
4242 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4243 return NODE_RECLAIM_NOSCAN
;
4245 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4246 return NODE_RECLAIM_NOSCAN
;
4248 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4249 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4252 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4259 * check_move_unevictable_pages - check pages for evictability and move to
4260 * appropriate zone lru list
4261 * @pvec: pagevec with lru pages to check
4263 * Checks pages for evictability, if an evictable page is in the unevictable
4264 * lru list, moves it to the appropriate evictable lru list. This function
4265 * should be only used for lru pages.
4267 void check_move_unevictable_pages(struct pagevec
*pvec
)
4269 struct lruvec
*lruvec
;
4270 struct pglist_data
*pgdat
= NULL
;
4275 for (i
= 0; i
< pvec
->nr
; i
++) {
4276 struct page
*page
= pvec
->pages
[i
];
4277 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4280 if (pagepgdat
!= pgdat
) {
4282 spin_unlock_irq(&pgdat
->lru_lock
);
4284 spin_lock_irq(&pgdat
->lru_lock
);
4286 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4288 if (!PageLRU(page
) || !PageUnevictable(page
))
4291 if (page_evictable(page
)) {
4292 enum lru_list lru
= page_lru_base_type(page
);
4294 VM_BUG_ON_PAGE(PageActive(page
), page
);
4295 ClearPageUnevictable(page
);
4296 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4297 add_page_to_lru_list(page
, lruvec
, lru
);
4303 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4304 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4305 spin_unlock_irq(&pgdat
->lru_lock
);
4308 EXPORT_SYMBOL_GPL(check_move_unevictable_pages
);