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/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim
;
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
78 struct mem_cgroup
*target_mem_cgroup
;
80 /* Writepage batching in laptop mode; RECLAIM_WRITE */
81 unsigned int may_writepage
:1;
83 /* Can mapped pages be reclaimed? */
84 unsigned int may_unmap
:1;
86 /* Can pages be swapped as part of reclaim? */
87 unsigned int may_swap
:1;
90 * Cgroups are not reclaimed below their configured memory.low,
91 * unless we threaten to OOM. If any cgroups are skipped due to
92 * memory.low and nothing was reclaimed, go back for memory.low.
94 unsigned int memcg_low_reclaim
:1;
95 unsigned int memcg_low_skipped
:1;
97 unsigned int hibernation_mode
:1;
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready
:1;
102 /* Allocation order */
105 /* Scan (total_size >> priority) pages at once */
108 /* The highest zone to isolate pages for reclaim from */
111 /* This context's GFP mask */
114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned
;
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed
;
122 unsigned int unqueued_dirty
;
123 unsigned int congested
;
124 unsigned int writeback
;
125 unsigned int immediate
;
126 unsigned int file_taken
;
131 #ifdef ARCH_HAS_PREFETCH
132 #define prefetch_prev_lru_page(_page, _base, _field) \
134 if ((_page)->lru.prev != _base) { \
137 prev = lru_to_page(&(_page->lru)); \
138 prefetch(&prev->_field); \
142 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
145 #ifdef ARCH_HAS_PREFETCHW
146 #define prefetchw_prev_lru_page(_page, _base, _field) \
148 if ((_page)->lru.prev != _base) { \
151 prev = lru_to_page(&(_page->lru)); \
152 prefetchw(&prev->_field); \
156 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
160 * From 0 .. 100. Higher means more swappy.
162 int vm_swappiness
= 60;
164 * The total number of pages which are beyond the high watermark within all
167 unsigned long vm_total_pages
;
169 static LIST_HEAD(shrinker_list
);
170 static DECLARE_RWSEM(shrinker_rwsem
);
172 #ifdef CONFIG_MEMCG_KMEM
175 * We allow subsystems to populate their shrinker-related
176 * LRU lists before register_shrinker_prepared() is called
177 * for the shrinker, since we don't want to impose
178 * restrictions on their internal registration order.
179 * In this case shrink_slab_memcg() may find corresponding
180 * bit is set in the shrinkers map.
182 * This value is used by the function to detect registering
183 * shrinkers and to skip do_shrink_slab() calls for them.
185 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
187 static DEFINE_IDR(shrinker_idr
);
188 static int shrinker_nr_max
;
190 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
192 int id
, ret
= -ENOMEM
;
194 down_write(&shrinker_rwsem
);
195 /* This may call shrinker, so it must use down_read_trylock() */
196 id
= idr_alloc(&shrinker_idr
, SHRINKER_REGISTERING
, 0, 0, GFP_KERNEL
);
200 if (id
>= shrinker_nr_max
) {
201 if (memcg_expand_shrinker_maps(id
)) {
202 idr_remove(&shrinker_idr
, id
);
206 shrinker_nr_max
= id
+ 1;
211 up_write(&shrinker_rwsem
);
215 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
217 int id
= shrinker
->id
;
221 down_write(&shrinker_rwsem
);
222 idr_remove(&shrinker_idr
, id
);
223 up_write(&shrinker_rwsem
);
225 #else /* CONFIG_MEMCG_KMEM */
226 static int prealloc_memcg_shrinker(struct shrinker
*shrinker
)
231 static void unregister_memcg_shrinker(struct shrinker
*shrinker
)
234 #endif /* CONFIG_MEMCG_KMEM */
237 static bool global_reclaim(struct scan_control
*sc
)
239 return !sc
->target_mem_cgroup
;
243 * sane_reclaim - is the usual dirty throttling mechanism operational?
244 * @sc: scan_control in question
246 * The normal page dirty throttling mechanism in balance_dirty_pages() is
247 * completely broken with the legacy memcg and direct stalling in
248 * shrink_page_list() is used for throttling instead, which lacks all the
249 * niceties such as fairness, adaptive pausing, bandwidth proportional
250 * allocation and configurability.
252 * This function tests whether the vmscan currently in progress can assume
253 * that the normal dirty throttling mechanism is operational.
255 static bool sane_reclaim(struct scan_control
*sc
)
257 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
261 #ifdef CONFIG_CGROUP_WRITEBACK
262 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
268 static void set_memcg_congestion(pg_data_t
*pgdat
,
269 struct mem_cgroup
*memcg
,
272 struct mem_cgroup_per_node
*mn
;
277 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
278 WRITE_ONCE(mn
->congested
, congested
);
281 static bool memcg_congested(pg_data_t
*pgdat
,
282 struct mem_cgroup
*memcg
)
284 struct mem_cgroup_per_node
*mn
;
286 mn
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
287 return READ_ONCE(mn
->congested
);
291 static bool global_reclaim(struct scan_control
*sc
)
296 static bool sane_reclaim(struct scan_control
*sc
)
301 static inline void set_memcg_congestion(struct pglist_data
*pgdat
,
302 struct mem_cgroup
*memcg
, bool congested
)
306 static inline bool memcg_congested(struct pglist_data
*pgdat
,
307 struct mem_cgroup
*memcg
)
315 * This misses isolated pages which are not accounted for to save counters.
316 * As the data only determines if reclaim or compaction continues, it is
317 * not expected that isolated pages will be a dominating factor.
319 unsigned long zone_reclaimable_pages(struct zone
*zone
)
323 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
324 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
325 if (get_nr_swap_pages() > 0)
326 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
327 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
333 * lruvec_lru_size - Returns the number of pages on the given LRU list.
334 * @lruvec: lru vector
336 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
338 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
340 unsigned long lru_size
;
343 if (!mem_cgroup_disabled())
344 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
346 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
348 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
349 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
352 if (!managed_zone(zone
))
355 if (!mem_cgroup_disabled())
356 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
358 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
359 NR_ZONE_LRU_BASE
+ lru
);
360 lru_size
-= min(size
, lru_size
);
368 * Add a shrinker callback to be called from the vm.
370 int prealloc_shrinker(struct shrinker
*shrinker
)
372 size_t size
= sizeof(*shrinker
->nr_deferred
);
374 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
377 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
378 if (!shrinker
->nr_deferred
)
381 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
) {
382 if (prealloc_memcg_shrinker(shrinker
))
389 kfree(shrinker
->nr_deferred
);
390 shrinker
->nr_deferred
= NULL
;
394 void free_prealloced_shrinker(struct shrinker
*shrinker
)
396 if (!shrinker
->nr_deferred
)
399 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
400 unregister_memcg_shrinker(shrinker
);
402 kfree(shrinker
->nr_deferred
);
403 shrinker
->nr_deferred
= NULL
;
406 void register_shrinker_prepared(struct shrinker
*shrinker
)
408 down_write(&shrinker_rwsem
);
409 list_add_tail(&shrinker
->list
, &shrinker_list
);
410 #ifdef CONFIG_MEMCG_KMEM
411 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
412 idr_replace(&shrinker_idr
, shrinker
, shrinker
->id
);
414 up_write(&shrinker_rwsem
);
417 int register_shrinker(struct shrinker
*shrinker
)
419 int err
= prealloc_shrinker(shrinker
);
423 register_shrinker_prepared(shrinker
);
426 EXPORT_SYMBOL(register_shrinker
);
431 void unregister_shrinker(struct shrinker
*shrinker
)
433 if (!shrinker
->nr_deferred
)
435 if (shrinker
->flags
& SHRINKER_MEMCG_AWARE
)
436 unregister_memcg_shrinker(shrinker
);
437 down_write(&shrinker_rwsem
);
438 list_del(&shrinker
->list
);
439 up_write(&shrinker_rwsem
);
440 kfree(shrinker
->nr_deferred
);
441 shrinker
->nr_deferred
= NULL
;
443 EXPORT_SYMBOL(unregister_shrinker
);
445 #define SHRINK_BATCH 128
447 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
448 struct shrinker
*shrinker
, int priority
)
450 unsigned long freed
= 0;
451 unsigned long long delta
;
456 int nid
= shrinkctl
->nid
;
457 long batch_size
= shrinker
->batch
? shrinker
->batch
459 long scanned
= 0, next_deferred
;
461 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
464 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
465 if (freeable
== 0 || freeable
== SHRINK_EMPTY
)
469 * copy the current shrinker scan count into a local variable
470 * and zero it so that other concurrent shrinker invocations
471 * don't also do this scanning work.
473 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
476 delta
= freeable
>> priority
;
478 do_div(delta
, shrinker
->seeks
);
481 * Make sure we apply some minimal pressure on default priority
482 * even on small cgroups. Stale objects are not only consuming memory
483 * by themselves, but can also hold a reference to a dying cgroup,
484 * preventing it from being reclaimed. A dying cgroup with all
485 * corresponding structures like per-cpu stats and kmem caches
486 * can be really big, so it may lead to a significant waste of memory.
488 delta
= max_t(unsigned long long, delta
, min(freeable
, batch_size
));
491 if (total_scan
< 0) {
492 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
493 shrinker
->scan_objects
, total_scan
);
494 total_scan
= freeable
;
497 next_deferred
= total_scan
;
500 * We need to avoid excessive windup on filesystem shrinkers
501 * due to large numbers of GFP_NOFS allocations causing the
502 * shrinkers to return -1 all the time. This results in a large
503 * nr being built up so when a shrink that can do some work
504 * comes along it empties the entire cache due to nr >>>
505 * freeable. This is bad for sustaining a working set in
508 * Hence only allow the shrinker to scan the entire cache when
509 * a large delta change is calculated directly.
511 if (delta
< freeable
/ 4)
512 total_scan
= min(total_scan
, freeable
/ 2);
515 * Avoid risking looping forever due to too large nr value:
516 * never try to free more than twice the estimate number of
519 if (total_scan
> freeable
* 2)
520 total_scan
= freeable
* 2;
522 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
523 freeable
, delta
, total_scan
, priority
);
526 * Normally, we should not scan less than batch_size objects in one
527 * pass to avoid too frequent shrinker calls, but if the slab has less
528 * than batch_size objects in total and we are really tight on memory,
529 * we will try to reclaim all available objects, otherwise we can end
530 * up failing allocations although there are plenty of reclaimable
531 * objects spread over several slabs with usage less than the
534 * We detect the "tight on memory" situations by looking at the total
535 * number of objects we want to scan (total_scan). If it is greater
536 * than the total number of objects on slab (freeable), we must be
537 * scanning at high prio and therefore should try to reclaim as much as
540 while (total_scan
>= batch_size
||
541 total_scan
>= freeable
) {
543 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
545 shrinkctl
->nr_to_scan
= nr_to_scan
;
546 shrinkctl
->nr_scanned
= nr_to_scan
;
547 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
548 if (ret
== SHRINK_STOP
)
552 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
553 total_scan
-= shrinkctl
->nr_scanned
;
554 scanned
+= shrinkctl
->nr_scanned
;
559 if (next_deferred
>= scanned
)
560 next_deferred
-= scanned
;
564 * move the unused scan count back into the shrinker in a
565 * manner that handles concurrent updates. If we exhausted the
566 * scan, there is no need to do an update.
568 if (next_deferred
> 0)
569 new_nr
= atomic_long_add_return(next_deferred
,
570 &shrinker
->nr_deferred
[nid
]);
572 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
574 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
578 #ifdef CONFIG_MEMCG_KMEM
579 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
580 struct mem_cgroup
*memcg
, int priority
)
582 struct memcg_shrinker_map
*map
;
583 unsigned long ret
, freed
= 0;
586 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
))
589 if (!down_read_trylock(&shrinker_rwsem
))
592 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
597 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
598 struct shrink_control sc
= {
599 .gfp_mask
= gfp_mask
,
603 struct shrinker
*shrinker
;
605 shrinker
= idr_find(&shrinker_idr
, i
);
606 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
608 clear_bit(i
, map
->map
);
612 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
613 if (ret
== SHRINK_EMPTY
) {
614 clear_bit(i
, map
->map
);
616 * After the shrinker reported that it had no objects to
617 * free, but before we cleared the corresponding bit in
618 * the memcg shrinker map, a new object might have been
619 * added. To make sure, we have the bit set in this
620 * case, we invoke the shrinker one more time and reset
621 * the bit if it reports that it is not empty anymore.
622 * The memory barrier here pairs with the barrier in
623 * memcg_set_shrinker_bit():
625 * list_lru_add() shrink_slab_memcg()
626 * list_add_tail() clear_bit()
628 * set_bit() do_shrink_slab()
630 smp_mb__after_atomic();
631 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
632 if (ret
== SHRINK_EMPTY
)
635 memcg_set_shrinker_bit(memcg
, nid
, i
);
639 if (rwsem_is_contended(&shrinker_rwsem
)) {
645 up_read(&shrinker_rwsem
);
648 #else /* CONFIG_MEMCG_KMEM */
649 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
650 struct mem_cgroup
*memcg
, int priority
)
654 #endif /* CONFIG_MEMCG_KMEM */
657 * shrink_slab - shrink slab caches
658 * @gfp_mask: allocation context
659 * @nid: node whose slab caches to target
660 * @memcg: memory cgroup whose slab caches to target
661 * @priority: the reclaim priority
663 * Call the shrink functions to age shrinkable caches.
665 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
666 * unaware shrinkers will receive a node id of 0 instead.
668 * @memcg specifies the memory cgroup to target. Unaware shrinkers
669 * are called only if it is the root cgroup.
671 * @priority is sc->priority, we take the number of objects and >> by priority
672 * in order to get the scan target.
674 * Returns the number of reclaimed slab objects.
676 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
677 struct mem_cgroup
*memcg
,
680 unsigned long ret
, freed
= 0;
681 struct shrinker
*shrinker
;
683 if (!mem_cgroup_is_root(memcg
))
684 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
686 if (!down_read_trylock(&shrinker_rwsem
))
689 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
690 struct shrink_control sc
= {
691 .gfp_mask
= gfp_mask
,
696 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
697 if (ret
== SHRINK_EMPTY
)
701 * Bail out if someone want to register a new shrinker to
702 * prevent the regsitration from being stalled for long periods
703 * by parallel ongoing shrinking.
705 if (rwsem_is_contended(&shrinker_rwsem
)) {
711 up_read(&shrinker_rwsem
);
717 void drop_slab_node(int nid
)
722 struct mem_cgroup
*memcg
= NULL
;
725 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
727 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
728 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
729 } while (freed
> 10);
736 for_each_online_node(nid
)
740 static inline int is_page_cache_freeable(struct page
*page
)
743 * A freeable page cache page is referenced only by the caller
744 * that isolated the page, the page cache radix tree and
745 * optional buffer heads at page->private.
747 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
749 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
752 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
754 if (current
->flags
& PF_SWAPWRITE
)
756 if (!inode_write_congested(inode
))
758 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
764 * We detected a synchronous write error writing a page out. Probably
765 * -ENOSPC. We need to propagate that into the address_space for a subsequent
766 * fsync(), msync() or close().
768 * The tricky part is that after writepage we cannot touch the mapping: nothing
769 * prevents it from being freed up. But we have a ref on the page and once
770 * that page is locked, the mapping is pinned.
772 * We're allowed to run sleeping lock_page() here because we know the caller has
775 static void handle_write_error(struct address_space
*mapping
,
776 struct page
*page
, int error
)
779 if (page_mapping(page
) == mapping
)
780 mapping_set_error(mapping
, error
);
784 /* possible outcome of pageout() */
786 /* failed to write page out, page is locked */
788 /* move page to the active list, page is locked */
790 /* page has been sent to the disk successfully, page is unlocked */
792 /* page is clean and locked */
797 * pageout is called by shrink_page_list() for each dirty page.
798 * Calls ->writepage().
800 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
801 struct scan_control
*sc
)
804 * If the page is dirty, only perform writeback if that write
805 * will be non-blocking. To prevent this allocation from being
806 * stalled by pagecache activity. But note that there may be
807 * stalls if we need to run get_block(). We could test
808 * PagePrivate for that.
810 * If this process is currently in __generic_file_write_iter() against
811 * this page's queue, we can perform writeback even if that
814 * If the page is swapcache, write it back even if that would
815 * block, for some throttling. This happens by accident, because
816 * swap_backing_dev_info is bust: it doesn't reflect the
817 * congestion state of the swapdevs. Easy to fix, if needed.
819 if (!is_page_cache_freeable(page
))
823 * Some data journaling orphaned pages can have
824 * page->mapping == NULL while being dirty with clean buffers.
826 if (page_has_private(page
)) {
827 if (try_to_free_buffers(page
)) {
828 ClearPageDirty(page
);
829 pr_info("%s: orphaned page\n", __func__
);
835 if (mapping
->a_ops
->writepage
== NULL
)
836 return PAGE_ACTIVATE
;
837 if (!may_write_to_inode(mapping
->host
, sc
))
840 if (clear_page_dirty_for_io(page
)) {
842 struct writeback_control wbc
= {
843 .sync_mode
= WB_SYNC_NONE
,
844 .nr_to_write
= SWAP_CLUSTER_MAX
,
846 .range_end
= LLONG_MAX
,
850 SetPageReclaim(page
);
851 res
= mapping
->a_ops
->writepage(page
, &wbc
);
853 handle_write_error(mapping
, page
, res
);
854 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
855 ClearPageReclaim(page
);
856 return PAGE_ACTIVATE
;
859 if (!PageWriteback(page
)) {
860 /* synchronous write or broken a_ops? */
861 ClearPageReclaim(page
);
863 trace_mm_vmscan_writepage(page
);
864 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
872 * Same as remove_mapping, but if the page is removed from the mapping, it
873 * gets returned with a refcount of 0.
875 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
881 BUG_ON(!PageLocked(page
));
882 BUG_ON(mapping
!= page_mapping(page
));
884 xa_lock_irqsave(&mapping
->i_pages
, flags
);
886 * The non racy check for a busy page.
888 * Must be careful with the order of the tests. When someone has
889 * a ref to the page, it may be possible that they dirty it then
890 * drop the reference. So if PageDirty is tested before page_count
891 * here, then the following race may occur:
893 * get_user_pages(&page);
894 * [user mapping goes away]
896 * !PageDirty(page) [good]
897 * SetPageDirty(page);
899 * !page_count(page) [good, discard it]
901 * [oops, our write_to data is lost]
903 * Reversing the order of the tests ensures such a situation cannot
904 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
905 * load is not satisfied before that of page->_refcount.
907 * Note that if SetPageDirty is always performed via set_page_dirty,
908 * and thus under the i_pages lock, then this ordering is not required.
910 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
911 refcount
= 1 + HPAGE_PMD_NR
;
914 if (!page_ref_freeze(page
, refcount
))
916 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
917 if (unlikely(PageDirty(page
))) {
918 page_ref_unfreeze(page
, refcount
);
922 if (PageSwapCache(page
)) {
923 swp_entry_t swap
= { .val
= page_private(page
) };
924 mem_cgroup_swapout(page
, swap
);
925 __delete_from_swap_cache(page
);
926 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
927 put_swap_page(page
, swap
);
929 void (*freepage
)(struct page
*);
932 freepage
= mapping
->a_ops
->freepage
;
934 * Remember a shadow entry for reclaimed file cache in
935 * order to detect refaults, thus thrashing, later on.
937 * But don't store shadows in an address space that is
938 * already exiting. This is not just an optizimation,
939 * inode reclaim needs to empty out the radix tree or
940 * the nodes are lost. Don't plant shadows behind its
943 * We also don't store shadows for DAX mappings because the
944 * only page cache pages found in these are zero pages
945 * covering holes, and because we don't want to mix DAX
946 * exceptional entries and shadow exceptional entries in the
947 * same address_space.
949 if (reclaimed
&& page_is_file_cache(page
) &&
950 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
951 shadow
= workingset_eviction(mapping
, page
);
952 __delete_from_page_cache(page
, shadow
);
953 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
955 if (freepage
!= NULL
)
962 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
967 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
968 * someone else has a ref on the page, abort and return 0. If it was
969 * successfully detached, return 1. Assumes the caller has a single ref on
972 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
974 if (__remove_mapping(mapping
, page
, false)) {
976 * Unfreezing the refcount with 1 rather than 2 effectively
977 * drops the pagecache ref for us without requiring another
980 page_ref_unfreeze(page
, 1);
987 * putback_lru_page - put previously isolated page onto appropriate LRU list
988 * @page: page to be put back to appropriate lru list
990 * Add previously isolated @page to appropriate LRU list.
991 * Page may still be unevictable for other reasons.
993 * lru_lock must not be held, interrupts must be enabled.
995 void putback_lru_page(struct page
*page
)
998 put_page(page
); /* drop ref from isolate */
1001 enum page_references
{
1003 PAGEREF_RECLAIM_CLEAN
,
1008 static enum page_references
page_check_references(struct page
*page
,
1009 struct scan_control
*sc
)
1011 int referenced_ptes
, referenced_page
;
1012 unsigned long vm_flags
;
1014 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1016 referenced_page
= TestClearPageReferenced(page
);
1019 * Mlock lost the isolation race with us. Let try_to_unmap()
1020 * move the page to the unevictable list.
1022 if (vm_flags
& VM_LOCKED
)
1023 return PAGEREF_RECLAIM
;
1025 if (referenced_ptes
) {
1026 if (PageSwapBacked(page
))
1027 return PAGEREF_ACTIVATE
;
1029 * All mapped pages start out with page table
1030 * references from the instantiating fault, so we need
1031 * to look twice if a mapped file page is used more
1034 * Mark it and spare it for another trip around the
1035 * inactive list. Another page table reference will
1036 * lead to its activation.
1038 * Note: the mark is set for activated pages as well
1039 * so that recently deactivated but used pages are
1040 * quickly recovered.
1042 SetPageReferenced(page
);
1044 if (referenced_page
|| referenced_ptes
> 1)
1045 return PAGEREF_ACTIVATE
;
1048 * Activate file-backed executable pages after first usage.
1050 if (vm_flags
& VM_EXEC
)
1051 return PAGEREF_ACTIVATE
;
1053 return PAGEREF_KEEP
;
1056 /* Reclaim if clean, defer dirty pages to writeback */
1057 if (referenced_page
&& !PageSwapBacked(page
))
1058 return PAGEREF_RECLAIM_CLEAN
;
1060 return PAGEREF_RECLAIM
;
1063 /* Check if a page is dirty or under writeback */
1064 static void page_check_dirty_writeback(struct page
*page
,
1065 bool *dirty
, bool *writeback
)
1067 struct address_space
*mapping
;
1070 * Anonymous pages are not handled by flushers and must be written
1071 * from reclaim context. Do not stall reclaim based on them
1073 if (!page_is_file_cache(page
) ||
1074 (PageAnon(page
) && !PageSwapBacked(page
))) {
1080 /* By default assume that the page flags are accurate */
1081 *dirty
= PageDirty(page
);
1082 *writeback
= PageWriteback(page
);
1084 /* Verify dirty/writeback state if the filesystem supports it */
1085 if (!page_has_private(page
))
1088 mapping
= page_mapping(page
);
1089 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1090 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1094 * shrink_page_list() returns the number of reclaimed pages
1096 static unsigned long shrink_page_list(struct list_head
*page_list
,
1097 struct pglist_data
*pgdat
,
1098 struct scan_control
*sc
,
1099 enum ttu_flags ttu_flags
,
1100 struct reclaim_stat
*stat
,
1103 LIST_HEAD(ret_pages
);
1104 LIST_HEAD(free_pages
);
1106 unsigned nr_unqueued_dirty
= 0;
1107 unsigned nr_dirty
= 0;
1108 unsigned nr_congested
= 0;
1109 unsigned nr_reclaimed
= 0;
1110 unsigned nr_writeback
= 0;
1111 unsigned nr_immediate
= 0;
1112 unsigned nr_ref_keep
= 0;
1113 unsigned nr_unmap_fail
= 0;
1117 while (!list_empty(page_list
)) {
1118 struct address_space
*mapping
;
1121 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
1122 bool dirty
, writeback
;
1126 page
= lru_to_page(page_list
);
1127 list_del(&page
->lru
);
1129 if (!trylock_page(page
))
1132 VM_BUG_ON_PAGE(PageActive(page
), page
);
1136 if (unlikely(!page_evictable(page
)))
1137 goto activate_locked
;
1139 if (!sc
->may_unmap
&& page_mapped(page
))
1142 /* Double the slab pressure for mapped and swapcache pages */
1143 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1144 !(PageAnon(page
) && !PageSwapBacked(page
)))
1147 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1148 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1151 * The number of dirty pages determines if a node is marked
1152 * reclaim_congested which affects wait_iff_congested. kswapd
1153 * will stall and start writing pages if the tail of the LRU
1154 * is all dirty unqueued pages.
1156 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1157 if (dirty
|| writeback
)
1160 if (dirty
&& !writeback
)
1161 nr_unqueued_dirty
++;
1164 * Treat this page as congested if the underlying BDI is or if
1165 * pages are cycling through the LRU so quickly that the
1166 * pages marked for immediate reclaim are making it to the
1167 * end of the LRU a second time.
1169 mapping
= page_mapping(page
);
1170 if (((dirty
|| writeback
) && mapping
&&
1171 inode_write_congested(mapping
->host
)) ||
1172 (writeback
&& PageReclaim(page
)))
1176 * If a page at the tail of the LRU is under writeback, there
1177 * are three cases to consider.
1179 * 1) If reclaim is encountering an excessive number of pages
1180 * under writeback and this page is both under writeback and
1181 * PageReclaim then it indicates that pages are being queued
1182 * for IO but are being recycled through the LRU before the
1183 * IO can complete. Waiting on the page itself risks an
1184 * indefinite stall if it is impossible to writeback the
1185 * page due to IO error or disconnected storage so instead
1186 * note that the LRU is being scanned too quickly and the
1187 * caller can stall after page list has been processed.
1189 * 2) Global or new memcg reclaim encounters a page that is
1190 * not marked for immediate reclaim, or the caller does not
1191 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1192 * not to fs). In this case mark the page for immediate
1193 * reclaim and continue scanning.
1195 * Require may_enter_fs because we would wait on fs, which
1196 * may not have submitted IO yet. And the loop driver might
1197 * enter reclaim, and deadlock if it waits on a page for
1198 * which it is needed to do the write (loop masks off
1199 * __GFP_IO|__GFP_FS for this reason); but more thought
1200 * would probably show more reasons.
1202 * 3) Legacy memcg encounters a page that is already marked
1203 * PageReclaim. memcg does not have any dirty pages
1204 * throttling so we could easily OOM just because too many
1205 * pages are in writeback and there is nothing else to
1206 * reclaim. Wait for the writeback to complete.
1208 * In cases 1) and 2) we activate the pages to get them out of
1209 * the way while we continue scanning for clean pages on the
1210 * inactive list and refilling from the active list. The
1211 * observation here is that waiting for disk writes is more
1212 * expensive than potentially causing reloads down the line.
1213 * Since they're marked for immediate reclaim, they won't put
1214 * memory pressure on the cache working set any longer than it
1215 * takes to write them to disk.
1217 if (PageWriteback(page
)) {
1219 if (current_is_kswapd() &&
1220 PageReclaim(page
) &&
1221 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1223 goto activate_locked
;
1226 } else if (sane_reclaim(sc
) ||
1227 !PageReclaim(page
) || !may_enter_fs
) {
1229 * This is slightly racy - end_page_writeback()
1230 * might have just cleared PageReclaim, then
1231 * setting PageReclaim here end up interpreted
1232 * as PageReadahead - but that does not matter
1233 * enough to care. What we do want is for this
1234 * page to have PageReclaim set next time memcg
1235 * reclaim reaches the tests above, so it will
1236 * then wait_on_page_writeback() to avoid OOM;
1237 * and it's also appropriate in global reclaim.
1239 SetPageReclaim(page
);
1241 goto activate_locked
;
1246 wait_on_page_writeback(page
);
1247 /* then go back and try same page again */
1248 list_add_tail(&page
->lru
, page_list
);
1254 references
= page_check_references(page
, sc
);
1256 switch (references
) {
1257 case PAGEREF_ACTIVATE
:
1258 goto activate_locked
;
1262 case PAGEREF_RECLAIM
:
1263 case PAGEREF_RECLAIM_CLEAN
:
1264 ; /* try to reclaim the page below */
1268 * Anonymous process memory has backing store?
1269 * Try to allocate it some swap space here.
1270 * Lazyfree page could be freed directly
1272 if (PageAnon(page
) && PageSwapBacked(page
)) {
1273 if (!PageSwapCache(page
)) {
1274 if (!(sc
->gfp_mask
& __GFP_IO
))
1276 if (PageTransHuge(page
)) {
1277 /* cannot split THP, skip it */
1278 if (!can_split_huge_page(page
, NULL
))
1279 goto activate_locked
;
1281 * Split pages without a PMD map right
1282 * away. Chances are some or all of the
1283 * tail pages can be freed without IO.
1285 if (!compound_mapcount(page
) &&
1286 split_huge_page_to_list(page
,
1288 goto activate_locked
;
1290 if (!add_to_swap(page
)) {
1291 if (!PageTransHuge(page
))
1292 goto activate_locked
;
1293 /* Fallback to swap normal pages */
1294 if (split_huge_page_to_list(page
,
1296 goto activate_locked
;
1297 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1298 count_vm_event(THP_SWPOUT_FALLBACK
);
1300 if (!add_to_swap(page
))
1301 goto activate_locked
;
1306 /* Adding to swap updated mapping */
1307 mapping
= page_mapping(page
);
1309 } else if (unlikely(PageTransHuge(page
))) {
1310 /* Split file THP */
1311 if (split_huge_page_to_list(page
, page_list
))
1316 * The page is mapped into the page tables of one or more
1317 * processes. Try to unmap it here.
1319 if (page_mapped(page
)) {
1320 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1322 if (unlikely(PageTransHuge(page
)))
1323 flags
|= TTU_SPLIT_HUGE_PMD
;
1324 if (!try_to_unmap(page
, flags
)) {
1326 goto activate_locked
;
1330 if (PageDirty(page
)) {
1332 * Only kswapd can writeback filesystem pages
1333 * to avoid risk of stack overflow. But avoid
1334 * injecting inefficient single-page IO into
1335 * flusher writeback as much as possible: only
1336 * write pages when we've encountered many
1337 * dirty pages, and when we've already scanned
1338 * the rest of the LRU for clean pages and see
1339 * the same dirty pages again (PageReclaim).
1341 if (page_is_file_cache(page
) &&
1342 (!current_is_kswapd() || !PageReclaim(page
) ||
1343 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1345 * Immediately reclaim when written back.
1346 * Similar in principal to deactivate_page()
1347 * except we already have the page isolated
1348 * and know it's dirty
1350 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1351 SetPageReclaim(page
);
1353 goto activate_locked
;
1356 if (references
== PAGEREF_RECLAIM_CLEAN
)
1360 if (!sc
->may_writepage
)
1364 * Page is dirty. Flush the TLB if a writable entry
1365 * potentially exists to avoid CPU writes after IO
1366 * starts and then write it out here.
1368 try_to_unmap_flush_dirty();
1369 switch (pageout(page
, mapping
, sc
)) {
1373 goto activate_locked
;
1375 if (PageWriteback(page
))
1377 if (PageDirty(page
))
1381 * A synchronous write - probably a ramdisk. Go
1382 * ahead and try to reclaim the page.
1384 if (!trylock_page(page
))
1386 if (PageDirty(page
) || PageWriteback(page
))
1388 mapping
= page_mapping(page
);
1390 ; /* try to free the page below */
1395 * If the page has buffers, try to free the buffer mappings
1396 * associated with this page. If we succeed we try to free
1399 * We do this even if the page is PageDirty().
1400 * try_to_release_page() does not perform I/O, but it is
1401 * possible for a page to have PageDirty set, but it is actually
1402 * clean (all its buffers are clean). This happens if the
1403 * buffers were written out directly, with submit_bh(). ext3
1404 * will do this, as well as the blockdev mapping.
1405 * try_to_release_page() will discover that cleanness and will
1406 * drop the buffers and mark the page clean - it can be freed.
1408 * Rarely, pages can have buffers and no ->mapping. These are
1409 * the pages which were not successfully invalidated in
1410 * truncate_complete_page(). We try to drop those buffers here
1411 * and if that worked, and the page is no longer mapped into
1412 * process address space (page_count == 1) it can be freed.
1413 * Otherwise, leave the page on the LRU so it is swappable.
1415 if (page_has_private(page
)) {
1416 if (!try_to_release_page(page
, sc
->gfp_mask
))
1417 goto activate_locked
;
1418 if (!mapping
&& page_count(page
) == 1) {
1420 if (put_page_testzero(page
))
1424 * rare race with speculative reference.
1425 * the speculative reference will free
1426 * this page shortly, so we may
1427 * increment nr_reclaimed here (and
1428 * leave it off the LRU).
1436 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1437 /* follow __remove_mapping for reference */
1438 if (!page_ref_freeze(page
, 1))
1440 if (PageDirty(page
)) {
1441 page_ref_unfreeze(page
, 1);
1445 count_vm_event(PGLAZYFREED
);
1446 count_memcg_page_event(page
, PGLAZYFREED
);
1447 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1450 * At this point, we have no other references and there is
1451 * no way to pick any more up (removed from LRU, removed
1452 * from pagecache). Can use non-atomic bitops now (and
1453 * we obviously don't have to worry about waking up a process
1454 * waiting on the page lock, because there are no references.
1456 __ClearPageLocked(page
);
1461 * Is there need to periodically free_page_list? It would
1462 * appear not as the counts should be low
1464 if (unlikely(PageTransHuge(page
))) {
1465 mem_cgroup_uncharge(page
);
1466 (*get_compound_page_dtor(page
))(page
);
1468 list_add(&page
->lru
, &free_pages
);
1472 /* Not a candidate for swapping, so reclaim swap space. */
1473 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1475 try_to_free_swap(page
);
1476 VM_BUG_ON_PAGE(PageActive(page
), page
);
1477 if (!PageMlocked(page
)) {
1478 SetPageActive(page
);
1480 count_memcg_page_event(page
, PGACTIVATE
);
1485 list_add(&page
->lru
, &ret_pages
);
1486 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1489 mem_cgroup_uncharge_list(&free_pages
);
1490 try_to_unmap_flush();
1491 free_unref_page_list(&free_pages
);
1493 list_splice(&ret_pages
, page_list
);
1494 count_vm_events(PGACTIVATE
, pgactivate
);
1497 stat
->nr_dirty
= nr_dirty
;
1498 stat
->nr_congested
= nr_congested
;
1499 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1500 stat
->nr_writeback
= nr_writeback
;
1501 stat
->nr_immediate
= nr_immediate
;
1502 stat
->nr_activate
= pgactivate
;
1503 stat
->nr_ref_keep
= nr_ref_keep
;
1504 stat
->nr_unmap_fail
= nr_unmap_fail
;
1506 return nr_reclaimed
;
1509 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1510 struct list_head
*page_list
)
1512 struct scan_control sc
= {
1513 .gfp_mask
= GFP_KERNEL
,
1514 .priority
= DEF_PRIORITY
,
1518 struct page
*page
, *next
;
1519 LIST_HEAD(clean_pages
);
1521 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1522 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1523 !__PageMovable(page
)) {
1524 ClearPageActive(page
);
1525 list_move(&page
->lru
, &clean_pages
);
1529 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1530 TTU_IGNORE_ACCESS
, NULL
, true);
1531 list_splice(&clean_pages
, page_list
);
1532 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1537 * Attempt to remove the specified page from its LRU. Only take this page
1538 * if it is of the appropriate PageActive status. Pages which are being
1539 * freed elsewhere are also ignored.
1541 * page: page to consider
1542 * mode: one of the LRU isolation modes defined above
1544 * returns 0 on success, -ve errno on failure.
1546 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1550 /* Only take pages on the LRU. */
1554 /* Compaction should not handle unevictable pages but CMA can do so */
1555 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1561 * To minimise LRU disruption, the caller can indicate that it only
1562 * wants to isolate pages it will be able to operate on without
1563 * blocking - clean pages for the most part.
1565 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1566 * that it is possible to migrate without blocking
1568 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1569 /* All the caller can do on PageWriteback is block */
1570 if (PageWriteback(page
))
1573 if (PageDirty(page
)) {
1574 struct address_space
*mapping
;
1578 * Only pages without mappings or that have a
1579 * ->migratepage callback are possible to migrate
1580 * without blocking. However, we can be racing with
1581 * truncation so it's necessary to lock the page
1582 * to stabilise the mapping as truncation holds
1583 * the page lock until after the page is removed
1584 * from the page cache.
1586 if (!trylock_page(page
))
1589 mapping
= page_mapping(page
);
1590 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1597 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1600 if (likely(get_page_unless_zero(page
))) {
1602 * Be careful not to clear PageLRU until after we're
1603 * sure the page is not being freed elsewhere -- the
1604 * page release code relies on it.
1615 * Update LRU sizes after isolating pages. The LRU size updates must
1616 * be complete before mem_cgroup_update_lru_size due to a santity check.
1618 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1619 enum lru_list lru
, unsigned long *nr_zone_taken
)
1623 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1624 if (!nr_zone_taken
[zid
])
1627 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1629 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1636 * zone_lru_lock is heavily contended. Some of the functions that
1637 * shrink the lists perform better by taking out a batch of pages
1638 * and working on them outside the LRU lock.
1640 * For pagecache intensive workloads, this function is the hottest
1641 * spot in the kernel (apart from copy_*_user functions).
1643 * Appropriate locks must be held before calling this function.
1645 * @nr_to_scan: The number of eligible pages to look through on the list.
1646 * @lruvec: The LRU vector to pull pages from.
1647 * @dst: The temp list to put pages on to.
1648 * @nr_scanned: The number of pages that were scanned.
1649 * @sc: The scan_control struct for this reclaim session
1650 * @mode: One of the LRU isolation modes
1651 * @lru: LRU list id for isolating
1653 * returns how many pages were moved onto *@dst.
1655 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1656 struct lruvec
*lruvec
, struct list_head
*dst
,
1657 unsigned long *nr_scanned
, struct scan_control
*sc
,
1658 isolate_mode_t mode
, enum lru_list lru
)
1660 struct list_head
*src
= &lruvec
->lists
[lru
];
1661 unsigned long nr_taken
= 0;
1662 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1663 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1664 unsigned long skipped
= 0;
1665 unsigned long scan
, total_scan
, nr_pages
;
1666 LIST_HEAD(pages_skipped
);
1669 for (total_scan
= 0;
1670 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1674 page
= lru_to_page(src
);
1675 prefetchw_prev_lru_page(page
, src
, flags
);
1677 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1679 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1680 list_move(&page
->lru
, &pages_skipped
);
1681 nr_skipped
[page_zonenum(page
)]++;
1686 * Do not count skipped pages because that makes the function
1687 * return with no isolated pages if the LRU mostly contains
1688 * ineligible pages. This causes the VM to not reclaim any
1689 * pages, triggering a premature OOM.
1692 switch (__isolate_lru_page(page
, mode
)) {
1694 nr_pages
= hpage_nr_pages(page
);
1695 nr_taken
+= nr_pages
;
1696 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1697 list_move(&page
->lru
, dst
);
1701 /* else it is being freed elsewhere */
1702 list_move(&page
->lru
, src
);
1711 * Splice any skipped pages to the start of the LRU list. Note that
1712 * this disrupts the LRU order when reclaiming for lower zones but
1713 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1714 * scanning would soon rescan the same pages to skip and put the
1715 * system at risk of premature OOM.
1717 if (!list_empty(&pages_skipped
)) {
1720 list_splice(&pages_skipped
, src
);
1721 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1722 if (!nr_skipped
[zid
])
1725 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1726 skipped
+= nr_skipped
[zid
];
1729 *nr_scanned
= total_scan
;
1730 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1731 total_scan
, skipped
, nr_taken
, mode
, lru
);
1732 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1737 * isolate_lru_page - tries to isolate a page from its LRU list
1738 * @page: page to isolate from its LRU list
1740 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1741 * vmstat statistic corresponding to whatever LRU list the page was on.
1743 * Returns 0 if the page was removed from an LRU list.
1744 * Returns -EBUSY if the page was not on an LRU list.
1746 * The returned page will have PageLRU() cleared. If it was found on
1747 * the active list, it will have PageActive set. If it was found on
1748 * the unevictable list, it will have the PageUnevictable bit set. That flag
1749 * may need to be cleared by the caller before letting the page go.
1751 * The vmstat statistic corresponding to the list on which the page was
1752 * found will be decremented.
1756 * (1) Must be called with an elevated refcount on the page. This is a
1757 * fundamentnal difference from isolate_lru_pages (which is called
1758 * without a stable reference).
1759 * (2) the lru_lock must not be held.
1760 * (3) interrupts must be enabled.
1762 int isolate_lru_page(struct page
*page
)
1766 VM_BUG_ON_PAGE(!page_count(page
), page
);
1767 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1769 if (PageLRU(page
)) {
1770 struct zone
*zone
= page_zone(page
);
1771 struct lruvec
*lruvec
;
1773 spin_lock_irq(zone_lru_lock(zone
));
1774 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1775 if (PageLRU(page
)) {
1776 int lru
= page_lru(page
);
1779 del_page_from_lru_list(page
, lruvec
, lru
);
1782 spin_unlock_irq(zone_lru_lock(zone
));
1788 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1789 * then get resheduled. When there are massive number of tasks doing page
1790 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1791 * the LRU list will go small and be scanned faster than necessary, leading to
1792 * unnecessary swapping, thrashing and OOM.
1794 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1795 struct scan_control
*sc
)
1797 unsigned long inactive
, isolated
;
1799 if (current_is_kswapd())
1802 if (!sane_reclaim(sc
))
1806 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1807 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1809 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1810 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1814 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1815 * won't get blocked by normal direct-reclaimers, forming a circular
1818 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1821 return isolated
> inactive
;
1824 static noinline_for_stack
void
1825 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1827 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1828 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1829 LIST_HEAD(pages_to_free
);
1832 * Put back any unfreeable pages.
1834 while (!list_empty(page_list
)) {
1835 struct page
*page
= lru_to_page(page_list
);
1838 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1839 list_del(&page
->lru
);
1840 if (unlikely(!page_evictable(page
))) {
1841 spin_unlock_irq(&pgdat
->lru_lock
);
1842 putback_lru_page(page
);
1843 spin_lock_irq(&pgdat
->lru_lock
);
1847 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1850 lru
= page_lru(page
);
1851 add_page_to_lru_list(page
, lruvec
, lru
);
1853 if (is_active_lru(lru
)) {
1854 int file
= is_file_lru(lru
);
1855 int numpages
= hpage_nr_pages(page
);
1856 reclaim_stat
->recent_rotated
[file
] += numpages
;
1858 if (put_page_testzero(page
)) {
1859 __ClearPageLRU(page
);
1860 __ClearPageActive(page
);
1861 del_page_from_lru_list(page
, lruvec
, lru
);
1863 if (unlikely(PageCompound(page
))) {
1864 spin_unlock_irq(&pgdat
->lru_lock
);
1865 mem_cgroup_uncharge(page
);
1866 (*get_compound_page_dtor(page
))(page
);
1867 spin_lock_irq(&pgdat
->lru_lock
);
1869 list_add(&page
->lru
, &pages_to_free
);
1874 * To save our caller's stack, now use input list for pages to free.
1876 list_splice(&pages_to_free
, page_list
);
1880 * If a kernel thread (such as nfsd for loop-back mounts) services
1881 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1882 * In that case we should only throttle if the backing device it is
1883 * writing to is congested. In other cases it is safe to throttle.
1885 static int current_may_throttle(void)
1887 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1888 current
->backing_dev_info
== NULL
||
1889 bdi_write_congested(current
->backing_dev_info
);
1893 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1894 * of reclaimed pages
1896 static noinline_for_stack
unsigned long
1897 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1898 struct scan_control
*sc
, enum lru_list lru
)
1900 LIST_HEAD(page_list
);
1901 unsigned long nr_scanned
;
1902 unsigned long nr_reclaimed
= 0;
1903 unsigned long nr_taken
;
1904 struct reclaim_stat stat
= {};
1905 isolate_mode_t isolate_mode
= 0;
1906 int file
= is_file_lru(lru
);
1907 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1908 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1909 bool stalled
= false;
1911 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1915 /* wait a bit for the reclaimer. */
1919 /* We are about to die and free our memory. Return now. */
1920 if (fatal_signal_pending(current
))
1921 return SWAP_CLUSTER_MAX
;
1927 isolate_mode
|= ISOLATE_UNMAPPED
;
1929 spin_lock_irq(&pgdat
->lru_lock
);
1931 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1932 &nr_scanned
, sc
, isolate_mode
, lru
);
1934 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1935 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1937 if (current_is_kswapd()) {
1938 if (global_reclaim(sc
))
1939 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1940 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1943 if (global_reclaim(sc
))
1944 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1945 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1948 spin_unlock_irq(&pgdat
->lru_lock
);
1953 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1956 spin_lock_irq(&pgdat
->lru_lock
);
1958 if (current_is_kswapd()) {
1959 if (global_reclaim(sc
))
1960 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1961 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1964 if (global_reclaim(sc
))
1965 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1966 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1970 putback_inactive_pages(lruvec
, &page_list
);
1972 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1974 spin_unlock_irq(&pgdat
->lru_lock
);
1976 mem_cgroup_uncharge_list(&page_list
);
1977 free_unref_page_list(&page_list
);
1980 * If dirty pages are scanned that are not queued for IO, it
1981 * implies that flushers are not doing their job. This can
1982 * happen when memory pressure pushes dirty pages to the end of
1983 * the LRU before the dirty limits are breached and the dirty
1984 * data has expired. It can also happen when the proportion of
1985 * dirty pages grows not through writes but through memory
1986 * pressure reclaiming all the clean cache. And in some cases,
1987 * the flushers simply cannot keep up with the allocation
1988 * rate. Nudge the flusher threads in case they are asleep.
1990 if (stat
.nr_unqueued_dirty
== nr_taken
)
1991 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1993 sc
->nr
.dirty
+= stat
.nr_dirty
;
1994 sc
->nr
.congested
+= stat
.nr_congested
;
1995 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1996 sc
->nr
.writeback
+= stat
.nr_writeback
;
1997 sc
->nr
.immediate
+= stat
.nr_immediate
;
1998 sc
->nr
.taken
+= nr_taken
;
2000 sc
->nr
.file_taken
+= nr_taken
;
2002 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
2003 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
2004 return nr_reclaimed
;
2008 * This moves pages from the active list to the inactive list.
2010 * We move them the other way if the page is referenced by one or more
2011 * processes, from rmap.
2013 * If the pages are mostly unmapped, the processing is fast and it is
2014 * appropriate to hold zone_lru_lock across the whole operation. But if
2015 * the pages are mapped, the processing is slow (page_referenced()) so we
2016 * should drop zone_lru_lock around each page. It's impossible to balance
2017 * this, so instead we remove the pages from the LRU while processing them.
2018 * It is safe to rely on PG_active against the non-LRU pages in here because
2019 * nobody will play with that bit on a non-LRU page.
2021 * The downside is that we have to touch page->_refcount against each page.
2022 * But we had to alter page->flags anyway.
2024 * Returns the number of pages moved to the given lru.
2027 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
2028 struct list_head
*list
,
2029 struct list_head
*pages_to_free
,
2032 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2037 while (!list_empty(list
)) {
2038 page
= lru_to_page(list
);
2039 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2041 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2044 nr_pages
= hpage_nr_pages(page
);
2045 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
2046 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
2048 if (put_page_testzero(page
)) {
2049 __ClearPageLRU(page
);
2050 __ClearPageActive(page
);
2051 del_page_from_lru_list(page
, lruvec
, lru
);
2053 if (unlikely(PageCompound(page
))) {
2054 spin_unlock_irq(&pgdat
->lru_lock
);
2055 mem_cgroup_uncharge(page
);
2056 (*get_compound_page_dtor(page
))(page
);
2057 spin_lock_irq(&pgdat
->lru_lock
);
2059 list_add(&page
->lru
, pages_to_free
);
2061 nr_moved
+= nr_pages
;
2065 if (!is_active_lru(lru
)) {
2066 __count_vm_events(PGDEACTIVATE
, nr_moved
);
2067 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
2074 static void shrink_active_list(unsigned long nr_to_scan
,
2075 struct lruvec
*lruvec
,
2076 struct scan_control
*sc
,
2079 unsigned long nr_taken
;
2080 unsigned long nr_scanned
;
2081 unsigned long vm_flags
;
2082 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2083 LIST_HEAD(l_active
);
2084 LIST_HEAD(l_inactive
);
2086 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2087 unsigned nr_deactivate
, nr_activate
;
2088 unsigned nr_rotated
= 0;
2089 isolate_mode_t isolate_mode
= 0;
2090 int file
= is_file_lru(lru
);
2091 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2096 isolate_mode
|= ISOLATE_UNMAPPED
;
2098 spin_lock_irq(&pgdat
->lru_lock
);
2100 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2101 &nr_scanned
, sc
, isolate_mode
, lru
);
2103 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2104 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2106 __count_vm_events(PGREFILL
, nr_scanned
);
2107 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2109 spin_unlock_irq(&pgdat
->lru_lock
);
2111 while (!list_empty(&l_hold
)) {
2113 page
= lru_to_page(&l_hold
);
2114 list_del(&page
->lru
);
2116 if (unlikely(!page_evictable(page
))) {
2117 putback_lru_page(page
);
2121 if (unlikely(buffer_heads_over_limit
)) {
2122 if (page_has_private(page
) && trylock_page(page
)) {
2123 if (page_has_private(page
))
2124 try_to_release_page(page
, 0);
2129 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2131 nr_rotated
+= hpage_nr_pages(page
);
2133 * Identify referenced, file-backed active pages and
2134 * give them one more trip around the active list. So
2135 * that executable code get better chances to stay in
2136 * memory under moderate memory pressure. Anon pages
2137 * are not likely to be evicted by use-once streaming
2138 * IO, plus JVM can create lots of anon VM_EXEC pages,
2139 * so we ignore them here.
2141 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2142 list_add(&page
->lru
, &l_active
);
2147 ClearPageActive(page
); /* we are de-activating */
2148 list_add(&page
->lru
, &l_inactive
);
2152 * Move pages back to the lru list.
2154 spin_lock_irq(&pgdat
->lru_lock
);
2156 * Count referenced pages from currently used mappings as rotated,
2157 * even though only some of them are actually re-activated. This
2158 * helps balance scan pressure between file and anonymous pages in
2161 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2163 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2164 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2165 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2166 spin_unlock_irq(&pgdat
->lru_lock
);
2168 mem_cgroup_uncharge_list(&l_hold
);
2169 free_unref_page_list(&l_hold
);
2170 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2171 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2175 * The inactive anon list should be small enough that the VM never has
2176 * to do too much work.
2178 * The inactive file list should be small enough to leave most memory
2179 * to the established workingset on the scan-resistant active list,
2180 * but large enough to avoid thrashing the aggregate readahead window.
2182 * Both inactive lists should also be large enough that each inactive
2183 * page has a chance to be referenced again before it is reclaimed.
2185 * If that fails and refaulting is observed, the inactive list grows.
2187 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2188 * on this LRU, maintained by the pageout code. An inactive_ratio
2189 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2192 * memory ratio inactive
2193 * -------------------------------------
2202 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2203 struct mem_cgroup
*memcg
,
2204 struct scan_control
*sc
, bool actual_reclaim
)
2206 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2207 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2208 enum lru_list inactive_lru
= file
* LRU_FILE
;
2209 unsigned long inactive
, active
;
2210 unsigned long inactive_ratio
;
2211 unsigned long refaults
;
2215 * If we don't have swap space, anonymous page deactivation
2218 if (!file
&& !total_swap_pages
)
2221 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2222 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2225 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2227 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2230 * When refaults are being observed, it means a new workingset
2231 * is being established. Disable active list protection to get
2232 * rid of the stale workingset quickly.
2234 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2237 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2239 inactive_ratio
= int_sqrt(10 * gb
);
2245 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2246 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2247 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2248 inactive_ratio
, file
);
2250 return inactive
* inactive_ratio
< active
;
2253 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2254 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2255 struct scan_control
*sc
)
2257 if (is_active_lru(lru
)) {
2258 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2260 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2264 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2275 * Determine how aggressively the anon and file LRU lists should be
2276 * scanned. The relative value of each set of LRU lists is determined
2277 * by looking at the fraction of the pages scanned we did rotate back
2278 * onto the active list instead of evict.
2280 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2281 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2283 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2284 struct scan_control
*sc
, unsigned long *nr
,
2285 unsigned long *lru_pages
)
2287 int swappiness
= mem_cgroup_swappiness(memcg
);
2288 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2290 u64 denominator
= 0; /* gcc */
2291 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2292 unsigned long anon_prio
, file_prio
;
2293 enum scan_balance scan_balance
;
2294 unsigned long anon
, file
;
2295 unsigned long ap
, fp
;
2298 /* If we have no swap space, do not bother scanning anon pages. */
2299 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2300 scan_balance
= SCAN_FILE
;
2305 * Global reclaim will swap to prevent OOM even with no
2306 * swappiness, but memcg users want to use this knob to
2307 * disable swapping for individual groups completely when
2308 * using the memory controller's swap limit feature would be
2311 if (!global_reclaim(sc
) && !swappiness
) {
2312 scan_balance
= SCAN_FILE
;
2317 * Do not apply any pressure balancing cleverness when the
2318 * system is close to OOM, scan both anon and file equally
2319 * (unless the swappiness setting disagrees with swapping).
2321 if (!sc
->priority
&& swappiness
) {
2322 scan_balance
= SCAN_EQUAL
;
2327 * Prevent the reclaimer from falling into the cache trap: as
2328 * cache pages start out inactive, every cache fault will tip
2329 * the scan balance towards the file LRU. And as the file LRU
2330 * shrinks, so does the window for rotation from references.
2331 * This means we have a runaway feedback loop where a tiny
2332 * thrashing file LRU becomes infinitely more attractive than
2333 * anon pages. Try to detect this based on file LRU size.
2335 if (global_reclaim(sc
)) {
2336 unsigned long pgdatfile
;
2337 unsigned long pgdatfree
;
2339 unsigned long total_high_wmark
= 0;
2341 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2342 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2343 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2345 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2346 struct zone
*zone
= &pgdat
->node_zones
[z
];
2347 if (!managed_zone(zone
))
2350 total_high_wmark
+= high_wmark_pages(zone
);
2353 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2355 * Force SCAN_ANON if there are enough inactive
2356 * anonymous pages on the LRU in eligible zones.
2357 * Otherwise, the small LRU gets thrashed.
2359 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2360 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2362 scan_balance
= SCAN_ANON
;
2369 * If there is enough inactive page cache, i.e. if the size of the
2370 * inactive list is greater than that of the active list *and* the
2371 * inactive list actually has some pages to scan on this priority, we
2372 * do not reclaim anything from the anonymous working set right now.
2373 * Without the second condition we could end up never scanning an
2374 * lruvec even if it has plenty of old anonymous pages unless the
2375 * system is under heavy pressure.
2377 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2378 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2379 scan_balance
= SCAN_FILE
;
2383 scan_balance
= SCAN_FRACT
;
2386 * With swappiness at 100, anonymous and file have the same priority.
2387 * This scanning priority is essentially the inverse of IO cost.
2389 anon_prio
= swappiness
;
2390 file_prio
= 200 - anon_prio
;
2393 * OK, so we have swap space and a fair amount of page cache
2394 * pages. We use the recently rotated / recently scanned
2395 * ratios to determine how valuable each cache is.
2397 * Because workloads change over time (and to avoid overflow)
2398 * we keep these statistics as a floating average, which ends
2399 * up weighing recent references more than old ones.
2401 * anon in [0], file in [1]
2404 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2405 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2406 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2407 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2409 spin_lock_irq(&pgdat
->lru_lock
);
2410 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2411 reclaim_stat
->recent_scanned
[0] /= 2;
2412 reclaim_stat
->recent_rotated
[0] /= 2;
2415 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2416 reclaim_stat
->recent_scanned
[1] /= 2;
2417 reclaim_stat
->recent_rotated
[1] /= 2;
2421 * The amount of pressure on anon vs file pages is inversely
2422 * proportional to the fraction of recently scanned pages on
2423 * each list that were recently referenced and in active use.
2425 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2426 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2428 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2429 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2430 spin_unlock_irq(&pgdat
->lru_lock
);
2434 denominator
= ap
+ fp
+ 1;
2437 for_each_evictable_lru(lru
) {
2438 int file
= is_file_lru(lru
);
2442 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2443 scan
= size
>> sc
->priority
;
2445 * If the cgroup's already been deleted, make sure to
2446 * scrape out the remaining cache.
2448 if (!scan
&& !mem_cgroup_online(memcg
))
2449 scan
= min(size
, SWAP_CLUSTER_MAX
);
2451 switch (scan_balance
) {
2453 /* Scan lists relative to size */
2457 * Scan types proportional to swappiness and
2458 * their relative recent reclaim efficiency.
2460 scan
= div64_u64(scan
* fraction
[file
],
2465 /* Scan one type exclusively */
2466 if ((scan_balance
== SCAN_FILE
) != file
) {
2472 /* Look ma, no brain */
2482 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2484 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2485 struct scan_control
*sc
, unsigned long *lru_pages
)
2487 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2488 unsigned long nr
[NR_LRU_LISTS
];
2489 unsigned long targets
[NR_LRU_LISTS
];
2490 unsigned long nr_to_scan
;
2492 unsigned long nr_reclaimed
= 0;
2493 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2494 struct blk_plug plug
;
2497 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2499 /* Record the original scan target for proportional adjustments later */
2500 memcpy(targets
, nr
, sizeof(nr
));
2503 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2504 * event that can occur when there is little memory pressure e.g.
2505 * multiple streaming readers/writers. Hence, we do not abort scanning
2506 * when the requested number of pages are reclaimed when scanning at
2507 * DEF_PRIORITY on the assumption that the fact we are direct
2508 * reclaiming implies that kswapd is not keeping up and it is best to
2509 * do a batch of work at once. For memcg reclaim one check is made to
2510 * abort proportional reclaim if either the file or anon lru has already
2511 * dropped to zero at the first pass.
2513 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2514 sc
->priority
== DEF_PRIORITY
);
2516 blk_start_plug(&plug
);
2517 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2518 nr
[LRU_INACTIVE_FILE
]) {
2519 unsigned long nr_anon
, nr_file
, percentage
;
2520 unsigned long nr_scanned
;
2522 for_each_evictable_lru(lru
) {
2524 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2525 nr
[lru
] -= nr_to_scan
;
2527 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2534 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2538 * For kswapd and memcg, reclaim at least the number of pages
2539 * requested. Ensure that the anon and file LRUs are scanned
2540 * proportionally what was requested by get_scan_count(). We
2541 * stop reclaiming one LRU and reduce the amount scanning
2542 * proportional to the original scan target.
2544 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2545 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2548 * It's just vindictive to attack the larger once the smaller
2549 * has gone to zero. And given the way we stop scanning the
2550 * smaller below, this makes sure that we only make one nudge
2551 * towards proportionality once we've got nr_to_reclaim.
2553 if (!nr_file
|| !nr_anon
)
2556 if (nr_file
> nr_anon
) {
2557 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2558 targets
[LRU_ACTIVE_ANON
] + 1;
2560 percentage
= nr_anon
* 100 / scan_target
;
2562 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2563 targets
[LRU_ACTIVE_FILE
] + 1;
2565 percentage
= nr_file
* 100 / scan_target
;
2568 /* Stop scanning the smaller of the LRU */
2570 nr
[lru
+ LRU_ACTIVE
] = 0;
2573 * Recalculate the other LRU scan count based on its original
2574 * scan target and the percentage scanning already complete
2576 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2577 nr_scanned
= targets
[lru
] - nr
[lru
];
2578 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2579 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2582 nr_scanned
= targets
[lru
] - nr
[lru
];
2583 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2584 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2586 scan_adjusted
= true;
2588 blk_finish_plug(&plug
);
2589 sc
->nr_reclaimed
+= nr_reclaimed
;
2592 * Even if we did not try to evict anon pages at all, we want to
2593 * rebalance the anon lru active/inactive ratio.
2595 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2596 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2597 sc
, LRU_ACTIVE_ANON
);
2600 /* Use reclaim/compaction for costly allocs or under memory pressure */
2601 static bool in_reclaim_compaction(struct scan_control
*sc
)
2603 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2604 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2605 sc
->priority
< DEF_PRIORITY
- 2))
2612 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2613 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2614 * true if more pages should be reclaimed such that when the page allocator
2615 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2616 * It will give up earlier than that if there is difficulty reclaiming pages.
2618 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2619 unsigned long nr_reclaimed
,
2620 unsigned long nr_scanned
,
2621 struct scan_control
*sc
)
2623 unsigned long pages_for_compaction
;
2624 unsigned long inactive_lru_pages
;
2627 /* If not in reclaim/compaction mode, stop */
2628 if (!in_reclaim_compaction(sc
))
2631 /* Consider stopping depending on scan and reclaim activity */
2632 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2634 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2635 * full LRU list has been scanned and we are still failing
2636 * to reclaim pages. This full LRU scan is potentially
2637 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2639 if (!nr_reclaimed
&& !nr_scanned
)
2643 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2644 * fail without consequence, stop if we failed to reclaim
2645 * any pages from the last SWAP_CLUSTER_MAX number of
2646 * pages that were scanned. This will return to the
2647 * caller faster at the risk reclaim/compaction and
2648 * the resulting allocation attempt fails
2655 * If we have not reclaimed enough pages for compaction and the
2656 * inactive lists are large enough, continue reclaiming
2658 pages_for_compaction
= compact_gap(sc
->order
);
2659 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2660 if (get_nr_swap_pages() > 0)
2661 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2662 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2663 inactive_lru_pages
> pages_for_compaction
)
2666 /* If compaction would go ahead or the allocation would succeed, stop */
2667 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2668 struct zone
*zone
= &pgdat
->node_zones
[z
];
2669 if (!managed_zone(zone
))
2672 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2673 case COMPACT_SUCCESS
:
2674 case COMPACT_CONTINUE
:
2677 /* check next zone */
2684 static bool pgdat_memcg_congested(pg_data_t
*pgdat
, struct mem_cgroup
*memcg
)
2686 return test_bit(PGDAT_CONGESTED
, &pgdat
->flags
) ||
2687 (memcg
&& memcg_congested(pgdat
, memcg
));
2690 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2692 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2693 unsigned long nr_reclaimed
, nr_scanned
;
2694 bool reclaimable
= false;
2697 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2698 struct mem_cgroup_reclaim_cookie reclaim
= {
2700 .priority
= sc
->priority
,
2702 unsigned long node_lru_pages
= 0;
2703 struct mem_cgroup
*memcg
;
2705 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2707 nr_reclaimed
= sc
->nr_reclaimed
;
2708 nr_scanned
= sc
->nr_scanned
;
2710 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2712 unsigned long lru_pages
;
2713 unsigned long reclaimed
;
2714 unsigned long scanned
;
2716 switch (mem_cgroup_protected(root
, memcg
)) {
2717 case MEMCG_PROT_MIN
:
2720 * If there is no reclaimable memory, OOM.
2723 case MEMCG_PROT_LOW
:
2726 * Respect the protection only as long as
2727 * there is an unprotected supply
2728 * of reclaimable memory from other cgroups.
2730 if (!sc
->memcg_low_reclaim
) {
2731 sc
->memcg_low_skipped
= 1;
2734 memcg_memory_event(memcg
, MEMCG_LOW
);
2736 case MEMCG_PROT_NONE
:
2740 reclaimed
= sc
->nr_reclaimed
;
2741 scanned
= sc
->nr_scanned
;
2742 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2743 node_lru_pages
+= lru_pages
;
2745 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2746 memcg
, sc
->priority
);
2748 /* Record the group's reclaim efficiency */
2749 vmpressure(sc
->gfp_mask
, memcg
, false,
2750 sc
->nr_scanned
- scanned
,
2751 sc
->nr_reclaimed
- reclaimed
);
2754 * Direct reclaim and kswapd have to scan all memory
2755 * cgroups to fulfill the overall scan target for the
2758 * Limit reclaim, on the other hand, only cares about
2759 * nr_to_reclaim pages to be reclaimed and it will
2760 * retry with decreasing priority if one round over the
2761 * whole hierarchy is not sufficient.
2763 if (!global_reclaim(sc
) &&
2764 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2765 mem_cgroup_iter_break(root
, memcg
);
2768 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2770 if (reclaim_state
) {
2771 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2772 reclaim_state
->reclaimed_slab
= 0;
2775 /* Record the subtree's reclaim efficiency */
2776 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2777 sc
->nr_scanned
- nr_scanned
,
2778 sc
->nr_reclaimed
- nr_reclaimed
);
2780 if (sc
->nr_reclaimed
- nr_reclaimed
)
2783 if (current_is_kswapd()) {
2785 * If reclaim is isolating dirty pages under writeback,
2786 * it implies that the long-lived page allocation rate
2787 * is exceeding the page laundering rate. Either the
2788 * global limits are not being effective at throttling
2789 * processes due to the page distribution throughout
2790 * zones or there is heavy usage of a slow backing
2791 * device. The only option is to throttle from reclaim
2792 * context which is not ideal as there is no guarantee
2793 * the dirtying process is throttled in the same way
2794 * balance_dirty_pages() manages.
2796 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2797 * count the number of pages under pages flagged for
2798 * immediate reclaim and stall if any are encountered
2799 * in the nr_immediate check below.
2801 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2802 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2805 * Tag a node as congested if all the dirty pages
2806 * scanned were backed by a congested BDI and
2807 * wait_iff_congested will stall.
2809 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2810 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
2812 /* Allow kswapd to start writing pages during reclaim.*/
2813 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2814 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2817 * If kswapd scans pages marked marked for immediate
2818 * reclaim and under writeback (nr_immediate), it
2819 * implies that pages are cycling through the LRU
2820 * faster than they are written so also forcibly stall.
2822 if (sc
->nr
.immediate
)
2823 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2827 * Legacy memcg will stall in page writeback so avoid forcibly
2828 * stalling in wait_iff_congested().
2830 if (!global_reclaim(sc
) && sane_reclaim(sc
) &&
2831 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2832 set_memcg_congestion(pgdat
, root
, true);
2835 * Stall direct reclaim for IO completions if underlying BDIs
2836 * and node is congested. Allow kswapd to continue until it
2837 * starts encountering unqueued dirty pages or cycling through
2838 * the LRU too quickly.
2840 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
2841 current_may_throttle() && pgdat_memcg_congested(pgdat
, root
))
2842 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2844 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2845 sc
->nr_scanned
- nr_scanned
, sc
));
2848 * Kswapd gives up on balancing particular nodes after too
2849 * many failures to reclaim anything from them and goes to
2850 * sleep. On reclaim progress, reset the failure counter. A
2851 * successful direct reclaim run will revive a dormant kswapd.
2854 pgdat
->kswapd_failures
= 0;
2860 * Returns true if compaction should go ahead for a costly-order request, or
2861 * the allocation would already succeed without compaction. Return false if we
2862 * should reclaim first.
2864 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2866 unsigned long watermark
;
2867 enum compact_result suitable
;
2869 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2870 if (suitable
== COMPACT_SUCCESS
)
2871 /* Allocation should succeed already. Don't reclaim. */
2873 if (suitable
== COMPACT_SKIPPED
)
2874 /* Compaction cannot yet proceed. Do reclaim. */
2878 * Compaction is already possible, but it takes time to run and there
2879 * are potentially other callers using the pages just freed. So proceed
2880 * with reclaim to make a buffer of free pages available to give
2881 * compaction a reasonable chance of completing and allocating the page.
2882 * Note that we won't actually reclaim the whole buffer in one attempt
2883 * as the target watermark in should_continue_reclaim() is lower. But if
2884 * we are already above the high+gap watermark, don't reclaim at all.
2886 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2888 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2892 * This is the direct reclaim path, for page-allocating processes. We only
2893 * try to reclaim pages from zones which will satisfy the caller's allocation
2896 * If a zone is deemed to be full of pinned pages then just give it a light
2897 * scan then give up on it.
2899 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2903 unsigned long nr_soft_reclaimed
;
2904 unsigned long nr_soft_scanned
;
2906 pg_data_t
*last_pgdat
= NULL
;
2909 * If the number of buffer_heads in the machine exceeds the maximum
2910 * allowed level, force direct reclaim to scan the highmem zone as
2911 * highmem pages could be pinning lowmem pages storing buffer_heads
2913 orig_mask
= sc
->gfp_mask
;
2914 if (buffer_heads_over_limit
) {
2915 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2916 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2919 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2920 sc
->reclaim_idx
, sc
->nodemask
) {
2922 * Take care memory controller reclaiming has small influence
2925 if (global_reclaim(sc
)) {
2926 if (!cpuset_zone_allowed(zone
,
2927 GFP_KERNEL
| __GFP_HARDWALL
))
2931 * If we already have plenty of memory free for
2932 * compaction in this zone, don't free any more.
2933 * Even though compaction is invoked for any
2934 * non-zero order, only frequent costly order
2935 * reclamation is disruptive enough to become a
2936 * noticeable problem, like transparent huge
2939 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2940 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2941 compaction_ready(zone
, sc
)) {
2942 sc
->compaction_ready
= true;
2947 * Shrink each node in the zonelist once. If the
2948 * zonelist is ordered by zone (not the default) then a
2949 * node may be shrunk multiple times but in that case
2950 * the user prefers lower zones being preserved.
2952 if (zone
->zone_pgdat
== last_pgdat
)
2956 * This steals pages from memory cgroups over softlimit
2957 * and returns the number of reclaimed pages and
2958 * scanned pages. This works for global memory pressure
2959 * and balancing, not for a memcg's limit.
2961 nr_soft_scanned
= 0;
2962 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2963 sc
->order
, sc
->gfp_mask
,
2965 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2966 sc
->nr_scanned
+= nr_soft_scanned
;
2967 /* need some check for avoid more shrink_zone() */
2970 /* See comment about same check for global reclaim above */
2971 if (zone
->zone_pgdat
== last_pgdat
)
2973 last_pgdat
= zone
->zone_pgdat
;
2974 shrink_node(zone
->zone_pgdat
, sc
);
2978 * Restore to original mask to avoid the impact on the caller if we
2979 * promoted it to __GFP_HIGHMEM.
2981 sc
->gfp_mask
= orig_mask
;
2984 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2986 struct mem_cgroup
*memcg
;
2988 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2990 unsigned long refaults
;
2991 struct lruvec
*lruvec
;
2994 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2996 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2998 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2999 lruvec
->refaults
= refaults
;
3000 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
3004 * This is the main entry point to direct page reclaim.
3006 * If a full scan of the inactive list fails to free enough memory then we
3007 * are "out of memory" and something needs to be killed.
3009 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3010 * high - the zone may be full of dirty or under-writeback pages, which this
3011 * caller can't do much about. We kick the writeback threads and take explicit
3012 * naps in the hope that some of these pages can be written. But if the
3013 * allocating task holds filesystem locks which prevent writeout this might not
3014 * work, and the allocation attempt will fail.
3016 * returns: 0, if no pages reclaimed
3017 * else, the number of pages reclaimed
3019 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3020 struct scan_control
*sc
)
3022 int initial_priority
= sc
->priority
;
3023 pg_data_t
*last_pgdat
;
3027 delayacct_freepages_start();
3029 if (global_reclaim(sc
))
3030 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3033 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3036 shrink_zones(zonelist
, sc
);
3038 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3041 if (sc
->compaction_ready
)
3045 * If we're getting trouble reclaiming, start doing
3046 * writepage even in laptop mode.
3048 if (sc
->priority
< DEF_PRIORITY
- 2)
3049 sc
->may_writepage
= 1;
3050 } while (--sc
->priority
>= 0);
3053 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3055 if (zone
->zone_pgdat
== last_pgdat
)
3057 last_pgdat
= zone
->zone_pgdat
;
3058 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3059 set_memcg_congestion(last_pgdat
, sc
->target_mem_cgroup
, false);
3062 delayacct_freepages_end();
3064 if (sc
->nr_reclaimed
)
3065 return sc
->nr_reclaimed
;
3067 /* Aborted reclaim to try compaction? don't OOM, then */
3068 if (sc
->compaction_ready
)
3071 /* Untapped cgroup reserves? Don't OOM, retry. */
3072 if (sc
->memcg_low_skipped
) {
3073 sc
->priority
= initial_priority
;
3074 sc
->memcg_low_reclaim
= 1;
3075 sc
->memcg_low_skipped
= 0;
3082 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3085 unsigned long pfmemalloc_reserve
= 0;
3086 unsigned long free_pages
= 0;
3090 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3093 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3094 zone
= &pgdat
->node_zones
[i
];
3095 if (!managed_zone(zone
))
3098 if (!zone_reclaimable_pages(zone
))
3101 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3102 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3105 /* If there are no reserves (unexpected config) then do not throttle */
3106 if (!pfmemalloc_reserve
)
3109 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3111 /* kswapd must be awake if processes are being throttled */
3112 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3113 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
3114 (enum zone_type
)ZONE_NORMAL
);
3115 wake_up_interruptible(&pgdat
->kswapd_wait
);
3122 * Throttle direct reclaimers if backing storage is backed by the network
3123 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3124 * depleted. kswapd will continue to make progress and wake the processes
3125 * when the low watermark is reached.
3127 * Returns true if a fatal signal was delivered during throttling. If this
3128 * happens, the page allocator should not consider triggering the OOM killer.
3130 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3131 nodemask_t
*nodemask
)
3135 pg_data_t
*pgdat
= NULL
;
3138 * Kernel threads should not be throttled as they may be indirectly
3139 * responsible for cleaning pages necessary for reclaim to make forward
3140 * progress. kjournald for example may enter direct reclaim while
3141 * committing a transaction where throttling it could forcing other
3142 * processes to block on log_wait_commit().
3144 if (current
->flags
& PF_KTHREAD
)
3148 * If a fatal signal is pending, this process should not throttle.
3149 * It should return quickly so it can exit and free its memory
3151 if (fatal_signal_pending(current
))
3155 * Check if the pfmemalloc reserves are ok by finding the first node
3156 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3157 * GFP_KERNEL will be required for allocating network buffers when
3158 * swapping over the network so ZONE_HIGHMEM is unusable.
3160 * Throttling is based on the first usable node and throttled processes
3161 * wait on a queue until kswapd makes progress and wakes them. There
3162 * is an affinity then between processes waking up and where reclaim
3163 * progress has been made assuming the process wakes on the same node.
3164 * More importantly, processes running on remote nodes will not compete
3165 * for remote pfmemalloc reserves and processes on different nodes
3166 * should make reasonable progress.
3168 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3169 gfp_zone(gfp_mask
), nodemask
) {
3170 if (zone_idx(zone
) > ZONE_NORMAL
)
3173 /* Throttle based on the first usable node */
3174 pgdat
= zone
->zone_pgdat
;
3175 if (allow_direct_reclaim(pgdat
))
3180 /* If no zone was usable by the allocation flags then do not throttle */
3184 /* Account for the throttling */
3185 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3188 * If the caller cannot enter the filesystem, it's possible that it
3189 * is due to the caller holding an FS lock or performing a journal
3190 * transaction in the case of a filesystem like ext[3|4]. In this case,
3191 * it is not safe to block on pfmemalloc_wait as kswapd could be
3192 * blocked waiting on the same lock. Instead, throttle for up to a
3193 * second before continuing.
3195 if (!(gfp_mask
& __GFP_FS
)) {
3196 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3197 allow_direct_reclaim(pgdat
), HZ
);
3202 /* Throttle until kswapd wakes the process */
3203 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3204 allow_direct_reclaim(pgdat
));
3207 if (fatal_signal_pending(current
))
3214 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3215 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3217 unsigned long nr_reclaimed
;
3218 struct scan_control sc
= {
3219 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3220 .gfp_mask
= current_gfp_context(gfp_mask
),
3221 .reclaim_idx
= gfp_zone(gfp_mask
),
3223 .nodemask
= nodemask
,
3224 .priority
= DEF_PRIORITY
,
3225 .may_writepage
= !laptop_mode
,
3231 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3232 * Confirm they are large enough for max values.
3234 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3235 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3236 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3239 * Do not enter reclaim if fatal signal was delivered while throttled.
3240 * 1 is returned so that the page allocator does not OOM kill at this
3243 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3246 trace_mm_vmscan_direct_reclaim_begin(order
,
3251 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3253 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3255 return nr_reclaimed
;
3260 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3261 gfp_t gfp_mask
, bool noswap
,
3263 unsigned long *nr_scanned
)
3265 struct scan_control sc
= {
3266 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3267 .target_mem_cgroup
= memcg
,
3268 .may_writepage
= !laptop_mode
,
3270 .reclaim_idx
= MAX_NR_ZONES
- 1,
3271 .may_swap
= !noswap
,
3273 unsigned long lru_pages
;
3275 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3276 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3278 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3284 * NOTE: Although we can get the priority field, using it
3285 * here is not a good idea, since it limits the pages we can scan.
3286 * if we don't reclaim here, the shrink_node from balance_pgdat
3287 * will pick up pages from other mem cgroup's as well. We hack
3288 * the priority and make it zero.
3290 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3292 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3294 *nr_scanned
= sc
.nr_scanned
;
3295 return sc
.nr_reclaimed
;
3298 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3299 unsigned long nr_pages
,
3303 struct zonelist
*zonelist
;
3304 unsigned long nr_reclaimed
;
3306 unsigned int noreclaim_flag
;
3307 struct scan_control sc
= {
3308 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3309 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3310 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3311 .reclaim_idx
= MAX_NR_ZONES
- 1,
3312 .target_mem_cgroup
= memcg
,
3313 .priority
= DEF_PRIORITY
,
3314 .may_writepage
= !laptop_mode
,
3316 .may_swap
= may_swap
,
3320 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3321 * take care of from where we get pages. So the node where we start the
3322 * scan does not need to be the current node.
3324 nid
= mem_cgroup_select_victim_node(memcg
);
3326 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3328 trace_mm_vmscan_memcg_reclaim_begin(0,
3333 noreclaim_flag
= memalloc_noreclaim_save();
3334 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3335 memalloc_noreclaim_restore(noreclaim_flag
);
3337 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3339 return nr_reclaimed
;
3343 static void age_active_anon(struct pglist_data
*pgdat
,
3344 struct scan_control
*sc
)
3346 struct mem_cgroup
*memcg
;
3348 if (!total_swap_pages
)
3351 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3353 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3355 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3356 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3357 sc
, LRU_ACTIVE_ANON
);
3359 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3364 * Returns true if there is an eligible zone balanced for the request order
3367 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3370 unsigned long mark
= -1;
3373 for (i
= 0; i
<= classzone_idx
; i
++) {
3374 zone
= pgdat
->node_zones
+ i
;
3376 if (!managed_zone(zone
))
3379 mark
= high_wmark_pages(zone
);
3380 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3385 * If a node has no populated zone within classzone_idx, it does not
3386 * need balancing by definition. This can happen if a zone-restricted
3387 * allocation tries to wake a remote kswapd.
3395 /* Clear pgdat state for congested, dirty or under writeback. */
3396 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3398 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3399 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3400 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3404 * Prepare kswapd for sleeping. This verifies that there are no processes
3405 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3407 * Returns true if kswapd is ready to sleep
3409 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3412 * The throttled processes are normally woken up in balance_pgdat() as
3413 * soon as allow_direct_reclaim() is true. But there is a potential
3414 * race between when kswapd checks the watermarks and a process gets
3415 * throttled. There is also a potential race if processes get
3416 * throttled, kswapd wakes, a large process exits thereby balancing the
3417 * zones, which causes kswapd to exit balance_pgdat() before reaching
3418 * the wake up checks. If kswapd is going to sleep, no process should
3419 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3420 * the wake up is premature, processes will wake kswapd and get
3421 * throttled again. The difference from wake ups in balance_pgdat() is
3422 * that here we are under prepare_to_wait().
3424 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3425 wake_up_all(&pgdat
->pfmemalloc_wait
);
3427 /* Hopeless node, leave it to direct reclaim */
3428 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3431 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3432 clear_pgdat_congested(pgdat
);
3440 * kswapd shrinks a node of pages that are at or below the highest usable
3441 * zone that is currently unbalanced.
3443 * Returns true if kswapd scanned at least the requested number of pages to
3444 * reclaim or if the lack of progress was due to pages under writeback.
3445 * This is used to determine if the scanning priority needs to be raised.
3447 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3448 struct scan_control
*sc
)
3453 /* Reclaim a number of pages proportional to the number of zones */
3454 sc
->nr_to_reclaim
= 0;
3455 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3456 zone
= pgdat
->node_zones
+ z
;
3457 if (!managed_zone(zone
))
3460 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3464 * Historically care was taken to put equal pressure on all zones but
3465 * now pressure is applied based on node LRU order.
3467 shrink_node(pgdat
, sc
);
3470 * Fragmentation may mean that the system cannot be rebalanced for
3471 * high-order allocations. If twice the allocation size has been
3472 * reclaimed then recheck watermarks only at order-0 to prevent
3473 * excessive reclaim. Assume that a process requested a high-order
3474 * can direct reclaim/compact.
3476 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3479 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3483 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3484 * that are eligible for use by the caller until at least one zone is
3487 * Returns the order kswapd finished reclaiming at.
3489 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3490 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3491 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3492 * or lower is eligible for reclaim until at least one usable zone is
3495 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3498 unsigned long nr_soft_reclaimed
;
3499 unsigned long nr_soft_scanned
;
3501 struct scan_control sc
= {
3502 .gfp_mask
= GFP_KERNEL
,
3504 .priority
= DEF_PRIORITY
,
3505 .may_writepage
= !laptop_mode
,
3510 __fs_reclaim_acquire();
3512 count_vm_event(PAGEOUTRUN
);
3515 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3516 bool raise_priority
= true;
3519 sc
.reclaim_idx
= classzone_idx
;
3522 * If the number of buffer_heads exceeds the maximum allowed
3523 * then consider reclaiming from all zones. This has a dual
3524 * purpose -- on 64-bit systems it is expected that
3525 * buffer_heads are stripped during active rotation. On 32-bit
3526 * systems, highmem pages can pin lowmem memory and shrinking
3527 * buffers can relieve lowmem pressure. Reclaim may still not
3528 * go ahead if all eligible zones for the original allocation
3529 * request are balanced to avoid excessive reclaim from kswapd.
3531 if (buffer_heads_over_limit
) {
3532 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3533 zone
= pgdat
->node_zones
+ i
;
3534 if (!managed_zone(zone
))
3543 * Only reclaim if there are no eligible zones. Note that
3544 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3547 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3551 * Do some background aging of the anon list, to give
3552 * pages a chance to be referenced before reclaiming. All
3553 * pages are rotated regardless of classzone as this is
3554 * about consistent aging.
3556 age_active_anon(pgdat
, &sc
);
3559 * If we're getting trouble reclaiming, start doing writepage
3560 * even in laptop mode.
3562 if (sc
.priority
< DEF_PRIORITY
- 2)
3563 sc
.may_writepage
= 1;
3565 /* Call soft limit reclaim before calling shrink_node. */
3567 nr_soft_scanned
= 0;
3568 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3569 sc
.gfp_mask
, &nr_soft_scanned
);
3570 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3573 * There should be no need to raise the scanning priority if
3574 * enough pages are already being scanned that that high
3575 * watermark would be met at 100% efficiency.
3577 if (kswapd_shrink_node(pgdat
, &sc
))
3578 raise_priority
= false;
3581 * If the low watermark is met there is no need for processes
3582 * to be throttled on pfmemalloc_wait as they should not be
3583 * able to safely make forward progress. Wake them
3585 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3586 allow_direct_reclaim(pgdat
))
3587 wake_up_all(&pgdat
->pfmemalloc_wait
);
3589 /* Check if kswapd should be suspending */
3590 __fs_reclaim_release();
3591 ret
= try_to_freeze();
3592 __fs_reclaim_acquire();
3593 if (ret
|| kthread_should_stop())
3597 * Raise priority if scanning rate is too low or there was no
3598 * progress in reclaiming pages
3600 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3601 if (raise_priority
|| !nr_reclaimed
)
3603 } while (sc
.priority
>= 1);
3605 if (!sc
.nr_reclaimed
)
3606 pgdat
->kswapd_failures
++;
3609 snapshot_refaults(NULL
, pgdat
);
3610 __fs_reclaim_release();
3612 * Return the order kswapd stopped reclaiming at as
3613 * prepare_kswapd_sleep() takes it into account. If another caller
3614 * entered the allocator slow path while kswapd was awake, order will
3615 * remain at the higher level.
3621 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3622 * allocation request woke kswapd for. When kswapd has not woken recently,
3623 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3624 * given classzone and returns it or the highest classzone index kswapd
3625 * was recently woke for.
3627 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3628 enum zone_type classzone_idx
)
3630 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3631 return classzone_idx
;
3633 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3636 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3637 unsigned int classzone_idx
)
3642 if (freezing(current
) || kthread_should_stop())
3645 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3648 * Try to sleep for a short interval. Note that kcompactd will only be
3649 * woken if it is possible to sleep for a short interval. This is
3650 * deliberate on the assumption that if reclaim cannot keep an
3651 * eligible zone balanced that it's also unlikely that compaction will
3654 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3656 * Compaction records what page blocks it recently failed to
3657 * isolate pages from and skips them in the future scanning.
3658 * When kswapd is going to sleep, it is reasonable to assume
3659 * that pages and compaction may succeed so reset the cache.
3661 reset_isolation_suitable(pgdat
);
3664 * We have freed the memory, now we should compact it to make
3665 * allocation of the requested order possible.
3667 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3669 remaining
= schedule_timeout(HZ
/10);
3672 * If woken prematurely then reset kswapd_classzone_idx and
3673 * order. The values will either be from a wakeup request or
3674 * the previous request that slept prematurely.
3677 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3678 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3681 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3682 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3686 * After a short sleep, check if it was a premature sleep. If not, then
3687 * go fully to sleep until explicitly woken up.
3690 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3691 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3694 * vmstat counters are not perfectly accurate and the estimated
3695 * value for counters such as NR_FREE_PAGES can deviate from the
3696 * true value by nr_online_cpus * threshold. To avoid the zone
3697 * watermarks being breached while under pressure, we reduce the
3698 * per-cpu vmstat threshold while kswapd is awake and restore
3699 * them before going back to sleep.
3701 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3703 if (!kthread_should_stop())
3706 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3709 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3711 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3713 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3717 * The background pageout daemon, started as a kernel thread
3718 * from the init process.
3720 * This basically trickles out pages so that we have _some_
3721 * free memory available even if there is no other activity
3722 * that frees anything up. This is needed for things like routing
3723 * etc, where we otherwise might have all activity going on in
3724 * asynchronous contexts that cannot page things out.
3726 * If there are applications that are active memory-allocators
3727 * (most normal use), this basically shouldn't matter.
3729 static int kswapd(void *p
)
3731 unsigned int alloc_order
, reclaim_order
;
3732 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3733 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3734 struct task_struct
*tsk
= current
;
3736 struct reclaim_state reclaim_state
= {
3737 .reclaimed_slab
= 0,
3739 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3741 if (!cpumask_empty(cpumask
))
3742 set_cpus_allowed_ptr(tsk
, cpumask
);
3743 current
->reclaim_state
= &reclaim_state
;
3746 * Tell the memory management that we're a "memory allocator",
3747 * and that if we need more memory we should get access to it
3748 * regardless (see "__alloc_pages()"). "kswapd" should
3749 * never get caught in the normal page freeing logic.
3751 * (Kswapd normally doesn't need memory anyway, but sometimes
3752 * you need a small amount of memory in order to be able to
3753 * page out something else, and this flag essentially protects
3754 * us from recursively trying to free more memory as we're
3755 * trying to free the first piece of memory in the first place).
3757 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3760 pgdat
->kswapd_order
= 0;
3761 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3765 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3766 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3769 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3772 /* Read the new order and classzone_idx */
3773 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3774 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3775 pgdat
->kswapd_order
= 0;
3776 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3778 ret
= try_to_freeze();
3779 if (kthread_should_stop())
3783 * We can speed up thawing tasks if we don't call balance_pgdat
3784 * after returning from the refrigerator
3790 * Reclaim begins at the requested order but if a high-order
3791 * reclaim fails then kswapd falls back to reclaiming for
3792 * order-0. If that happens, kswapd will consider sleeping
3793 * for the order it finished reclaiming at (reclaim_order)
3794 * but kcompactd is woken to compact for the original
3795 * request (alloc_order).
3797 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3799 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3800 if (reclaim_order
< alloc_order
)
3801 goto kswapd_try_sleep
;
3804 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3805 current
->reclaim_state
= NULL
;
3811 * A zone is low on free memory or too fragmented for high-order memory. If
3812 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3813 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3814 * has failed or is not needed, still wake up kcompactd if only compaction is
3817 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3818 enum zone_type classzone_idx
)
3822 if (!managed_zone(zone
))
3825 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3827 pgdat
= zone
->zone_pgdat
;
3828 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3830 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3831 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3834 /* Hopeless node, leave it to direct reclaim if possible */
3835 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3836 pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3838 * There may be plenty of free memory available, but it's too
3839 * fragmented for high-order allocations. Wake up kcompactd
3840 * and rely on compaction_suitable() to determine if it's
3841 * needed. If it fails, it will defer subsequent attempts to
3842 * ratelimit its work.
3844 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3845 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3849 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
3851 wake_up_interruptible(&pgdat
->kswapd_wait
);
3854 #ifdef CONFIG_HIBERNATION
3856 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3859 * Rather than trying to age LRUs the aim is to preserve the overall
3860 * LRU order by reclaiming preferentially
3861 * inactive > active > active referenced > active mapped
3863 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3865 struct reclaim_state reclaim_state
;
3866 struct scan_control sc
= {
3867 .nr_to_reclaim
= nr_to_reclaim
,
3868 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3869 .reclaim_idx
= MAX_NR_ZONES
- 1,
3870 .priority
= DEF_PRIORITY
,
3874 .hibernation_mode
= 1,
3876 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3877 struct task_struct
*p
= current
;
3878 unsigned long nr_reclaimed
;
3879 unsigned int noreclaim_flag
;
3881 fs_reclaim_acquire(sc
.gfp_mask
);
3882 noreclaim_flag
= memalloc_noreclaim_save();
3883 reclaim_state
.reclaimed_slab
= 0;
3884 p
->reclaim_state
= &reclaim_state
;
3886 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3888 p
->reclaim_state
= NULL
;
3889 memalloc_noreclaim_restore(noreclaim_flag
);
3890 fs_reclaim_release(sc
.gfp_mask
);
3892 return nr_reclaimed
;
3894 #endif /* CONFIG_HIBERNATION */
3896 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3897 not required for correctness. So if the last cpu in a node goes
3898 away, we get changed to run anywhere: as the first one comes back,
3899 restore their cpu bindings. */
3900 static int kswapd_cpu_online(unsigned int cpu
)
3904 for_each_node_state(nid
, N_MEMORY
) {
3905 pg_data_t
*pgdat
= NODE_DATA(nid
);
3906 const struct cpumask
*mask
;
3908 mask
= cpumask_of_node(pgdat
->node_id
);
3910 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3911 /* One of our CPUs online: restore mask */
3912 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3918 * This kswapd start function will be called by init and node-hot-add.
3919 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3921 int kswapd_run(int nid
)
3923 pg_data_t
*pgdat
= NODE_DATA(nid
);
3929 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3930 if (IS_ERR(pgdat
->kswapd
)) {
3931 /* failure at boot is fatal */
3932 BUG_ON(system_state
< SYSTEM_RUNNING
);
3933 pr_err("Failed to start kswapd on node %d\n", nid
);
3934 ret
= PTR_ERR(pgdat
->kswapd
);
3935 pgdat
->kswapd
= NULL
;
3941 * Called by memory hotplug when all memory in a node is offlined. Caller must
3942 * hold mem_hotplug_begin/end().
3944 void kswapd_stop(int nid
)
3946 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3949 kthread_stop(kswapd
);
3950 NODE_DATA(nid
)->kswapd
= NULL
;
3954 static int __init
kswapd_init(void)
3959 for_each_node_state(nid
, N_MEMORY
)
3961 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3962 "mm/vmscan:online", kswapd_cpu_online
,
3968 module_init(kswapd_init
)
3974 * If non-zero call node_reclaim when the number of free pages falls below
3977 int node_reclaim_mode __read_mostly
;
3979 #define RECLAIM_OFF 0
3980 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3981 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3982 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3985 * Priority for NODE_RECLAIM. This determines the fraction of pages
3986 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3989 #define NODE_RECLAIM_PRIORITY 4
3992 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3995 int sysctl_min_unmapped_ratio
= 1;
3998 * If the number of slab pages in a zone grows beyond this percentage then
3999 * slab reclaim needs to occur.
4001 int sysctl_min_slab_ratio
= 5;
4003 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
4005 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
4006 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
4007 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4010 * It's possible for there to be more file mapped pages than
4011 * accounted for by the pages on the file LRU lists because
4012 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4014 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4017 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4018 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4020 unsigned long nr_pagecache_reclaimable
;
4021 unsigned long delta
= 0;
4024 * If RECLAIM_UNMAP is set, then all file pages are considered
4025 * potentially reclaimable. Otherwise, we have to worry about
4026 * pages like swapcache and node_unmapped_file_pages() provides
4029 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4030 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4032 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4034 /* If we can't clean pages, remove dirty pages from consideration */
4035 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4036 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4038 /* Watch for any possible underflows due to delta */
4039 if (unlikely(delta
> nr_pagecache_reclaimable
))
4040 delta
= nr_pagecache_reclaimable
;
4042 return nr_pagecache_reclaimable
- delta
;
4046 * Try to free up some pages from this node through reclaim.
4048 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4050 /* Minimum pages needed in order to stay on node */
4051 const unsigned long nr_pages
= 1 << order
;
4052 struct task_struct
*p
= current
;
4053 struct reclaim_state reclaim_state
;
4054 unsigned int noreclaim_flag
;
4055 struct scan_control sc
= {
4056 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4057 .gfp_mask
= current_gfp_context(gfp_mask
),
4059 .priority
= NODE_RECLAIM_PRIORITY
,
4060 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4061 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4063 .reclaim_idx
= gfp_zone(gfp_mask
),
4067 fs_reclaim_acquire(sc
.gfp_mask
);
4069 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4070 * and we also need to be able to write out pages for RECLAIM_WRITE
4071 * and RECLAIM_UNMAP.
4073 noreclaim_flag
= memalloc_noreclaim_save();
4074 p
->flags
|= PF_SWAPWRITE
;
4075 reclaim_state
.reclaimed_slab
= 0;
4076 p
->reclaim_state
= &reclaim_state
;
4078 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4080 * Free memory by calling shrink node with increasing
4081 * priorities until we have enough memory freed.
4084 shrink_node(pgdat
, &sc
);
4085 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4088 p
->reclaim_state
= NULL
;
4089 current
->flags
&= ~PF_SWAPWRITE
;
4090 memalloc_noreclaim_restore(noreclaim_flag
);
4091 fs_reclaim_release(sc
.gfp_mask
);
4092 return sc
.nr_reclaimed
>= nr_pages
;
4095 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4100 * Node reclaim reclaims unmapped file backed pages and
4101 * slab pages if we are over the defined limits.
4103 * A small portion of unmapped file backed pages is needed for
4104 * file I/O otherwise pages read by file I/O will be immediately
4105 * thrown out if the node is overallocated. So we do not reclaim
4106 * if less than a specified percentage of the node is used by
4107 * unmapped file backed pages.
4109 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4110 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4111 return NODE_RECLAIM_FULL
;
4114 * Do not scan if the allocation should not be delayed.
4116 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4117 return NODE_RECLAIM_NOSCAN
;
4120 * Only run node reclaim on the local node or on nodes that do not
4121 * have associated processors. This will favor the local processor
4122 * over remote processors and spread off node memory allocations
4123 * as wide as possible.
4125 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4126 return NODE_RECLAIM_NOSCAN
;
4128 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4129 return NODE_RECLAIM_NOSCAN
;
4131 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4132 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4135 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4142 * page_evictable - test whether a page is evictable
4143 * @page: the page to test
4145 * Test whether page is evictable--i.e., should be placed on active/inactive
4146 * lists vs unevictable list.
4148 * Reasons page might not be evictable:
4149 * (1) page's mapping marked unevictable
4150 * (2) page is part of an mlocked VMA
4153 int page_evictable(struct page
*page
)
4157 /* Prevent address_space of inode and swap cache from being freed */
4159 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4166 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
4167 * @pages: array of pages to check
4168 * @nr_pages: number of pages to check
4170 * Checks pages for evictability and moves them to the appropriate lru list.
4172 * This function is only used for SysV IPC SHM_UNLOCK.
4174 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
4176 struct lruvec
*lruvec
;
4177 struct pglist_data
*pgdat
= NULL
;
4182 for (i
= 0; i
< nr_pages
; i
++) {
4183 struct page
*page
= pages
[i
];
4184 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4187 if (pagepgdat
!= pgdat
) {
4189 spin_unlock_irq(&pgdat
->lru_lock
);
4191 spin_lock_irq(&pgdat
->lru_lock
);
4193 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4195 if (!PageLRU(page
) || !PageUnevictable(page
))
4198 if (page_evictable(page
)) {
4199 enum lru_list lru
= page_lru_base_type(page
);
4201 VM_BUG_ON_PAGE(PageActive(page
), page
);
4202 ClearPageUnevictable(page
);
4203 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4204 add_page_to_lru_list(page
, lruvec
, lru
);
4210 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4211 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4212 spin_unlock_irq(&pgdat
->lru_lock
);
4215 #endif /* CONFIG_SHMEM */