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
);
480 if (total_scan
< 0) {
481 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
482 shrinker
->scan_objects
, total_scan
);
483 total_scan
= freeable
;
486 next_deferred
= total_scan
;
489 * We need to avoid excessive windup on filesystem shrinkers
490 * due to large numbers of GFP_NOFS allocations causing the
491 * shrinkers to return -1 all the time. This results in a large
492 * nr being built up so when a shrink that can do some work
493 * comes along it empties the entire cache due to nr >>>
494 * freeable. This is bad for sustaining a working set in
497 * Hence only allow the shrinker to scan the entire cache when
498 * a large delta change is calculated directly.
500 if (delta
< freeable
/ 4)
501 total_scan
= min(total_scan
, freeable
/ 2);
504 * Avoid risking looping forever due to too large nr value:
505 * never try to free more than twice the estimate number of
508 if (total_scan
> freeable
* 2)
509 total_scan
= freeable
* 2;
511 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
512 freeable
, delta
, total_scan
, priority
);
515 * Normally, we should not scan less than batch_size objects in one
516 * pass to avoid too frequent shrinker calls, but if the slab has less
517 * than batch_size objects in total and we are really tight on memory,
518 * we will try to reclaim all available objects, otherwise we can end
519 * up failing allocations although there are plenty of reclaimable
520 * objects spread over several slabs with usage less than the
523 * We detect the "tight on memory" situations by looking at the total
524 * number of objects we want to scan (total_scan). If it is greater
525 * than the total number of objects on slab (freeable), we must be
526 * scanning at high prio and therefore should try to reclaim as much as
529 while (total_scan
>= batch_size
||
530 total_scan
>= freeable
) {
532 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
534 shrinkctl
->nr_to_scan
= nr_to_scan
;
535 shrinkctl
->nr_scanned
= nr_to_scan
;
536 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
537 if (ret
== SHRINK_STOP
)
541 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
542 total_scan
-= shrinkctl
->nr_scanned
;
543 scanned
+= shrinkctl
->nr_scanned
;
548 if (next_deferred
>= scanned
)
549 next_deferred
-= scanned
;
553 * move the unused scan count back into the shrinker in a
554 * manner that handles concurrent updates. If we exhausted the
555 * scan, there is no need to do an update.
557 if (next_deferred
> 0)
558 new_nr
= atomic_long_add_return(next_deferred
,
559 &shrinker
->nr_deferred
[nid
]);
561 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
563 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
567 #ifdef CONFIG_MEMCG_KMEM
568 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
569 struct mem_cgroup
*memcg
, int priority
)
571 struct memcg_shrinker_map
*map
;
572 unsigned long freed
= 0;
575 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
))
578 if (!down_read_trylock(&shrinker_rwsem
))
581 map
= rcu_dereference_protected(memcg
->nodeinfo
[nid
]->shrinker_map
,
586 for_each_set_bit(i
, map
->map
, shrinker_nr_max
) {
587 struct shrink_control sc
= {
588 .gfp_mask
= gfp_mask
,
592 struct shrinker
*shrinker
;
594 shrinker
= idr_find(&shrinker_idr
, i
);
595 if (unlikely(!shrinker
|| shrinker
== SHRINKER_REGISTERING
)) {
597 clear_bit(i
, map
->map
);
601 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
602 if (ret
== SHRINK_EMPTY
) {
603 clear_bit(i
, map
->map
);
605 * After the shrinker reported that it had no objects to
606 * free, but before we cleared the corresponding bit in
607 * the memcg shrinker map, a new object might have been
608 * added. To make sure, we have the bit set in this
609 * case, we invoke the shrinker one more time and reset
610 * the bit if it reports that it is not empty anymore.
611 * The memory barrier here pairs with the barrier in
612 * memcg_set_shrinker_bit():
614 * list_lru_add() shrink_slab_memcg()
615 * list_add_tail() clear_bit()
617 * set_bit() do_shrink_slab()
619 smp_mb__after_atomic();
620 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
621 if (ret
== SHRINK_EMPTY
)
624 memcg_set_shrinker_bit(memcg
, nid
, i
);
628 if (rwsem_is_contended(&shrinker_rwsem
)) {
634 up_read(&shrinker_rwsem
);
637 #else /* CONFIG_MEMCG_KMEM */
638 static unsigned long shrink_slab_memcg(gfp_t gfp_mask
, int nid
,
639 struct mem_cgroup
*memcg
, int priority
)
643 #endif /* CONFIG_MEMCG_KMEM */
646 * shrink_slab - shrink slab caches
647 * @gfp_mask: allocation context
648 * @nid: node whose slab caches to target
649 * @memcg: memory cgroup whose slab caches to target
650 * @priority: the reclaim priority
652 * Call the shrink functions to age shrinkable caches.
654 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
655 * unaware shrinkers will receive a node id of 0 instead.
657 * @memcg specifies the memory cgroup to target. Unaware shrinkers
658 * are called only if it is the root cgroup.
660 * @priority is sc->priority, we take the number of objects and >> by priority
661 * in order to get the scan target.
663 * Returns the number of reclaimed slab objects.
665 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
666 struct mem_cgroup
*memcg
,
669 struct shrinker
*shrinker
;
670 unsigned long freed
= 0;
673 if (!mem_cgroup_is_root(memcg
))
674 return shrink_slab_memcg(gfp_mask
, nid
, memcg
, priority
);
676 if (!down_read_trylock(&shrinker_rwsem
))
679 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
680 struct shrink_control sc
= {
681 .gfp_mask
= gfp_mask
,
686 ret
= do_shrink_slab(&sc
, shrinker
, priority
);
687 if (ret
== SHRINK_EMPTY
)
691 * Bail out if someone want to register a new shrinker to
692 * prevent the regsitration from being stalled for long periods
693 * by parallel ongoing shrinking.
695 if (rwsem_is_contended(&shrinker_rwsem
)) {
701 up_read(&shrinker_rwsem
);
707 void drop_slab_node(int nid
)
712 struct mem_cgroup
*memcg
= NULL
;
715 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
717 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
718 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
719 } while (freed
> 10);
726 for_each_online_node(nid
)
730 static inline int is_page_cache_freeable(struct page
*page
)
733 * A freeable page cache page is referenced only by the caller
734 * that isolated the page, the page cache radix tree and
735 * optional buffer heads at page->private.
737 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
739 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
742 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
744 if (current
->flags
& PF_SWAPWRITE
)
746 if (!inode_write_congested(inode
))
748 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
754 * We detected a synchronous write error writing a page out. Probably
755 * -ENOSPC. We need to propagate that into the address_space for a subsequent
756 * fsync(), msync() or close().
758 * The tricky part is that after writepage we cannot touch the mapping: nothing
759 * prevents it from being freed up. But we have a ref on the page and once
760 * that page is locked, the mapping is pinned.
762 * We're allowed to run sleeping lock_page() here because we know the caller has
765 static void handle_write_error(struct address_space
*mapping
,
766 struct page
*page
, int error
)
769 if (page_mapping(page
) == mapping
)
770 mapping_set_error(mapping
, error
);
774 /* possible outcome of pageout() */
776 /* failed to write page out, page is locked */
778 /* move page to the active list, page is locked */
780 /* page has been sent to the disk successfully, page is unlocked */
782 /* page is clean and locked */
787 * pageout is called by shrink_page_list() for each dirty page.
788 * Calls ->writepage().
790 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
791 struct scan_control
*sc
)
794 * If the page is dirty, only perform writeback if that write
795 * will be non-blocking. To prevent this allocation from being
796 * stalled by pagecache activity. But note that there may be
797 * stalls if we need to run get_block(). We could test
798 * PagePrivate for that.
800 * If this process is currently in __generic_file_write_iter() against
801 * this page's queue, we can perform writeback even if that
804 * If the page is swapcache, write it back even if that would
805 * block, for some throttling. This happens by accident, because
806 * swap_backing_dev_info is bust: it doesn't reflect the
807 * congestion state of the swapdevs. Easy to fix, if needed.
809 if (!is_page_cache_freeable(page
))
813 * Some data journaling orphaned pages can have
814 * page->mapping == NULL while being dirty with clean buffers.
816 if (page_has_private(page
)) {
817 if (try_to_free_buffers(page
)) {
818 ClearPageDirty(page
);
819 pr_info("%s: orphaned page\n", __func__
);
825 if (mapping
->a_ops
->writepage
== NULL
)
826 return PAGE_ACTIVATE
;
827 if (!may_write_to_inode(mapping
->host
, sc
))
830 if (clear_page_dirty_for_io(page
)) {
832 struct writeback_control wbc
= {
833 .sync_mode
= WB_SYNC_NONE
,
834 .nr_to_write
= SWAP_CLUSTER_MAX
,
836 .range_end
= LLONG_MAX
,
840 SetPageReclaim(page
);
841 res
= mapping
->a_ops
->writepage(page
, &wbc
);
843 handle_write_error(mapping
, page
, res
);
844 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
845 ClearPageReclaim(page
);
846 return PAGE_ACTIVATE
;
849 if (!PageWriteback(page
)) {
850 /* synchronous write or broken a_ops? */
851 ClearPageReclaim(page
);
853 trace_mm_vmscan_writepage(page
);
854 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
862 * Same as remove_mapping, but if the page is removed from the mapping, it
863 * gets returned with a refcount of 0.
865 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
871 BUG_ON(!PageLocked(page
));
872 BUG_ON(mapping
!= page_mapping(page
));
874 xa_lock_irqsave(&mapping
->i_pages
, flags
);
876 * The non racy check for a busy page.
878 * Must be careful with the order of the tests. When someone has
879 * a ref to the page, it may be possible that they dirty it then
880 * drop the reference. So if PageDirty is tested before page_count
881 * here, then the following race may occur:
883 * get_user_pages(&page);
884 * [user mapping goes away]
886 * !PageDirty(page) [good]
887 * SetPageDirty(page);
889 * !page_count(page) [good, discard it]
891 * [oops, our write_to data is lost]
893 * Reversing the order of the tests ensures such a situation cannot
894 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
895 * load is not satisfied before that of page->_refcount.
897 * Note that if SetPageDirty is always performed via set_page_dirty,
898 * and thus under the i_pages lock, then this ordering is not required.
900 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
901 refcount
= 1 + HPAGE_PMD_NR
;
904 if (!page_ref_freeze(page
, refcount
))
906 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
907 if (unlikely(PageDirty(page
))) {
908 page_ref_unfreeze(page
, refcount
);
912 if (PageSwapCache(page
)) {
913 swp_entry_t swap
= { .val
= page_private(page
) };
914 mem_cgroup_swapout(page
, swap
);
915 __delete_from_swap_cache(page
);
916 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
917 put_swap_page(page
, swap
);
919 void (*freepage
)(struct page
*);
922 freepage
= mapping
->a_ops
->freepage
;
924 * Remember a shadow entry for reclaimed file cache in
925 * order to detect refaults, thus thrashing, later on.
927 * But don't store shadows in an address space that is
928 * already exiting. This is not just an optizimation,
929 * inode reclaim needs to empty out the radix tree or
930 * the nodes are lost. Don't plant shadows behind its
933 * We also don't store shadows for DAX mappings because the
934 * only page cache pages found in these are zero pages
935 * covering holes, and because we don't want to mix DAX
936 * exceptional entries and shadow exceptional entries in the
937 * same address_space.
939 if (reclaimed
&& page_is_file_cache(page
) &&
940 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
941 shadow
= workingset_eviction(mapping
, page
);
942 __delete_from_page_cache(page
, shadow
);
943 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
945 if (freepage
!= NULL
)
952 xa_unlock_irqrestore(&mapping
->i_pages
, flags
);
957 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
958 * someone else has a ref on the page, abort and return 0. If it was
959 * successfully detached, return 1. Assumes the caller has a single ref on
962 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
964 if (__remove_mapping(mapping
, page
, false)) {
966 * Unfreezing the refcount with 1 rather than 2 effectively
967 * drops the pagecache ref for us without requiring another
970 page_ref_unfreeze(page
, 1);
977 * putback_lru_page - put previously isolated page onto appropriate LRU list
978 * @page: page to be put back to appropriate lru list
980 * Add previously isolated @page to appropriate LRU list.
981 * Page may still be unevictable for other reasons.
983 * lru_lock must not be held, interrupts must be enabled.
985 void putback_lru_page(struct page
*page
)
988 put_page(page
); /* drop ref from isolate */
991 enum page_references
{
993 PAGEREF_RECLAIM_CLEAN
,
998 static enum page_references
page_check_references(struct page
*page
,
999 struct scan_control
*sc
)
1001 int referenced_ptes
, referenced_page
;
1002 unsigned long vm_flags
;
1004 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
1006 referenced_page
= TestClearPageReferenced(page
);
1009 * Mlock lost the isolation race with us. Let try_to_unmap()
1010 * move the page to the unevictable list.
1012 if (vm_flags
& VM_LOCKED
)
1013 return PAGEREF_RECLAIM
;
1015 if (referenced_ptes
) {
1016 if (PageSwapBacked(page
))
1017 return PAGEREF_ACTIVATE
;
1019 * All mapped pages start out with page table
1020 * references from the instantiating fault, so we need
1021 * to look twice if a mapped file page is used more
1024 * Mark it and spare it for another trip around the
1025 * inactive list. Another page table reference will
1026 * lead to its activation.
1028 * Note: the mark is set for activated pages as well
1029 * so that recently deactivated but used pages are
1030 * quickly recovered.
1032 SetPageReferenced(page
);
1034 if (referenced_page
|| referenced_ptes
> 1)
1035 return PAGEREF_ACTIVATE
;
1038 * Activate file-backed executable pages after first usage.
1040 if (vm_flags
& VM_EXEC
)
1041 return PAGEREF_ACTIVATE
;
1043 return PAGEREF_KEEP
;
1046 /* Reclaim if clean, defer dirty pages to writeback */
1047 if (referenced_page
&& !PageSwapBacked(page
))
1048 return PAGEREF_RECLAIM_CLEAN
;
1050 return PAGEREF_RECLAIM
;
1053 /* Check if a page is dirty or under writeback */
1054 static void page_check_dirty_writeback(struct page
*page
,
1055 bool *dirty
, bool *writeback
)
1057 struct address_space
*mapping
;
1060 * Anonymous pages are not handled by flushers and must be written
1061 * from reclaim context. Do not stall reclaim based on them
1063 if (!page_is_file_cache(page
) ||
1064 (PageAnon(page
) && !PageSwapBacked(page
))) {
1070 /* By default assume that the page flags are accurate */
1071 *dirty
= PageDirty(page
);
1072 *writeback
= PageWriteback(page
);
1074 /* Verify dirty/writeback state if the filesystem supports it */
1075 if (!page_has_private(page
))
1078 mapping
= page_mapping(page
);
1079 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
1080 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
1084 * shrink_page_list() returns the number of reclaimed pages
1086 static unsigned long shrink_page_list(struct list_head
*page_list
,
1087 struct pglist_data
*pgdat
,
1088 struct scan_control
*sc
,
1089 enum ttu_flags ttu_flags
,
1090 struct reclaim_stat
*stat
,
1093 LIST_HEAD(ret_pages
);
1094 LIST_HEAD(free_pages
);
1096 unsigned nr_unqueued_dirty
= 0;
1097 unsigned nr_dirty
= 0;
1098 unsigned nr_congested
= 0;
1099 unsigned nr_reclaimed
= 0;
1100 unsigned nr_writeback
= 0;
1101 unsigned nr_immediate
= 0;
1102 unsigned nr_ref_keep
= 0;
1103 unsigned nr_unmap_fail
= 0;
1107 while (!list_empty(page_list
)) {
1108 struct address_space
*mapping
;
1111 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
1112 bool dirty
, writeback
;
1116 page
= lru_to_page(page_list
);
1117 list_del(&page
->lru
);
1119 if (!trylock_page(page
))
1122 VM_BUG_ON_PAGE(PageActive(page
), page
);
1126 if (unlikely(!page_evictable(page
)))
1127 goto activate_locked
;
1129 if (!sc
->may_unmap
&& page_mapped(page
))
1132 /* Double the slab pressure for mapped and swapcache pages */
1133 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1134 !(PageAnon(page
) && !PageSwapBacked(page
)))
1137 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1138 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1141 * The number of dirty pages determines if a node is marked
1142 * reclaim_congested which affects wait_iff_congested. kswapd
1143 * will stall and start writing pages if the tail of the LRU
1144 * is all dirty unqueued pages.
1146 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1147 if (dirty
|| writeback
)
1150 if (dirty
&& !writeback
)
1151 nr_unqueued_dirty
++;
1154 * Treat this page as congested if the underlying BDI is or if
1155 * pages are cycling through the LRU so quickly that the
1156 * pages marked for immediate reclaim are making it to the
1157 * end of the LRU a second time.
1159 mapping
= page_mapping(page
);
1160 if (((dirty
|| writeback
) && mapping
&&
1161 inode_write_congested(mapping
->host
)) ||
1162 (writeback
&& PageReclaim(page
)))
1166 * If a page at the tail of the LRU is under writeback, there
1167 * are three cases to consider.
1169 * 1) If reclaim is encountering an excessive number of pages
1170 * under writeback and this page is both under writeback and
1171 * PageReclaim then it indicates that pages are being queued
1172 * for IO but are being recycled through the LRU before the
1173 * IO can complete. Waiting on the page itself risks an
1174 * indefinite stall if it is impossible to writeback the
1175 * page due to IO error or disconnected storage so instead
1176 * note that the LRU is being scanned too quickly and the
1177 * caller can stall after page list has been processed.
1179 * 2) Global or new memcg reclaim encounters a page that is
1180 * not marked for immediate reclaim, or the caller does not
1181 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1182 * not to fs). In this case mark the page for immediate
1183 * reclaim and continue scanning.
1185 * Require may_enter_fs because we would wait on fs, which
1186 * may not have submitted IO yet. And the loop driver might
1187 * enter reclaim, and deadlock if it waits on a page for
1188 * which it is needed to do the write (loop masks off
1189 * __GFP_IO|__GFP_FS for this reason); but more thought
1190 * would probably show more reasons.
1192 * 3) Legacy memcg encounters a page that is already marked
1193 * PageReclaim. memcg does not have any dirty pages
1194 * throttling so we could easily OOM just because too many
1195 * pages are in writeback and there is nothing else to
1196 * reclaim. Wait for the writeback to complete.
1198 * In cases 1) and 2) we activate the pages to get them out of
1199 * the way while we continue scanning for clean pages on the
1200 * inactive list and refilling from the active list. The
1201 * observation here is that waiting for disk writes is more
1202 * expensive than potentially causing reloads down the line.
1203 * Since they're marked for immediate reclaim, they won't put
1204 * memory pressure on the cache working set any longer than it
1205 * takes to write them to disk.
1207 if (PageWriteback(page
)) {
1209 if (current_is_kswapd() &&
1210 PageReclaim(page
) &&
1211 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1213 goto activate_locked
;
1216 } else if (sane_reclaim(sc
) ||
1217 !PageReclaim(page
) || !may_enter_fs
) {
1219 * This is slightly racy - end_page_writeback()
1220 * might have just cleared PageReclaim, then
1221 * setting PageReclaim here end up interpreted
1222 * as PageReadahead - but that does not matter
1223 * enough to care. What we do want is for this
1224 * page to have PageReclaim set next time memcg
1225 * reclaim reaches the tests above, so it will
1226 * then wait_on_page_writeback() to avoid OOM;
1227 * and it's also appropriate in global reclaim.
1229 SetPageReclaim(page
);
1231 goto activate_locked
;
1236 wait_on_page_writeback(page
);
1237 /* then go back and try same page again */
1238 list_add_tail(&page
->lru
, page_list
);
1244 references
= page_check_references(page
, sc
);
1246 switch (references
) {
1247 case PAGEREF_ACTIVATE
:
1248 goto activate_locked
;
1252 case PAGEREF_RECLAIM
:
1253 case PAGEREF_RECLAIM_CLEAN
:
1254 ; /* try to reclaim the page below */
1258 * Anonymous process memory has backing store?
1259 * Try to allocate it some swap space here.
1260 * Lazyfree page could be freed directly
1262 if (PageAnon(page
) && PageSwapBacked(page
)) {
1263 if (!PageSwapCache(page
)) {
1264 if (!(sc
->gfp_mask
& __GFP_IO
))
1266 if (PageTransHuge(page
)) {
1267 /* cannot split THP, skip it */
1268 if (!can_split_huge_page(page
, NULL
))
1269 goto activate_locked
;
1271 * Split pages without a PMD map right
1272 * away. Chances are some or all of the
1273 * tail pages can be freed without IO.
1275 if (!compound_mapcount(page
) &&
1276 split_huge_page_to_list(page
,
1278 goto activate_locked
;
1280 if (!add_to_swap(page
)) {
1281 if (!PageTransHuge(page
))
1282 goto activate_locked
;
1283 /* Fallback to swap normal pages */
1284 if (split_huge_page_to_list(page
,
1286 goto activate_locked
;
1287 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1288 count_vm_event(THP_SWPOUT_FALLBACK
);
1290 if (!add_to_swap(page
))
1291 goto activate_locked
;
1296 /* Adding to swap updated mapping */
1297 mapping
= page_mapping(page
);
1299 } else if (unlikely(PageTransHuge(page
))) {
1300 /* Split file THP */
1301 if (split_huge_page_to_list(page
, page_list
))
1306 * The page is mapped into the page tables of one or more
1307 * processes. Try to unmap it here.
1309 if (page_mapped(page
)) {
1310 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1312 if (unlikely(PageTransHuge(page
)))
1313 flags
|= TTU_SPLIT_HUGE_PMD
;
1314 if (!try_to_unmap(page
, flags
)) {
1316 goto activate_locked
;
1320 if (PageDirty(page
)) {
1322 * Only kswapd can writeback filesystem pages
1323 * to avoid risk of stack overflow. But avoid
1324 * injecting inefficient single-page IO into
1325 * flusher writeback as much as possible: only
1326 * write pages when we've encountered many
1327 * dirty pages, and when we've already scanned
1328 * the rest of the LRU for clean pages and see
1329 * the same dirty pages again (PageReclaim).
1331 if (page_is_file_cache(page
) &&
1332 (!current_is_kswapd() || !PageReclaim(page
) ||
1333 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1335 * Immediately reclaim when written back.
1336 * Similar in principal to deactivate_page()
1337 * except we already have the page isolated
1338 * and know it's dirty
1340 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1341 SetPageReclaim(page
);
1343 goto activate_locked
;
1346 if (references
== PAGEREF_RECLAIM_CLEAN
)
1350 if (!sc
->may_writepage
)
1354 * Page is dirty. Flush the TLB if a writable entry
1355 * potentially exists to avoid CPU writes after IO
1356 * starts and then write it out here.
1358 try_to_unmap_flush_dirty();
1359 switch (pageout(page
, mapping
, sc
)) {
1363 goto activate_locked
;
1365 if (PageWriteback(page
))
1367 if (PageDirty(page
))
1371 * A synchronous write - probably a ramdisk. Go
1372 * ahead and try to reclaim the page.
1374 if (!trylock_page(page
))
1376 if (PageDirty(page
) || PageWriteback(page
))
1378 mapping
= page_mapping(page
);
1380 ; /* try to free the page below */
1385 * If the page has buffers, try to free the buffer mappings
1386 * associated with this page. If we succeed we try to free
1389 * We do this even if the page is PageDirty().
1390 * try_to_release_page() does not perform I/O, but it is
1391 * possible for a page to have PageDirty set, but it is actually
1392 * clean (all its buffers are clean). This happens if the
1393 * buffers were written out directly, with submit_bh(). ext3
1394 * will do this, as well as the blockdev mapping.
1395 * try_to_release_page() will discover that cleanness and will
1396 * drop the buffers and mark the page clean - it can be freed.
1398 * Rarely, pages can have buffers and no ->mapping. These are
1399 * the pages which were not successfully invalidated in
1400 * truncate_complete_page(). We try to drop those buffers here
1401 * and if that worked, and the page is no longer mapped into
1402 * process address space (page_count == 1) it can be freed.
1403 * Otherwise, leave the page on the LRU so it is swappable.
1405 if (page_has_private(page
)) {
1406 if (!try_to_release_page(page
, sc
->gfp_mask
))
1407 goto activate_locked
;
1408 if (!mapping
&& page_count(page
) == 1) {
1410 if (put_page_testzero(page
))
1414 * rare race with speculative reference.
1415 * the speculative reference will free
1416 * this page shortly, so we may
1417 * increment nr_reclaimed here (and
1418 * leave it off the LRU).
1426 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1427 /* follow __remove_mapping for reference */
1428 if (!page_ref_freeze(page
, 1))
1430 if (PageDirty(page
)) {
1431 page_ref_unfreeze(page
, 1);
1435 count_vm_event(PGLAZYFREED
);
1436 count_memcg_page_event(page
, PGLAZYFREED
);
1437 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1440 * At this point, we have no other references and there is
1441 * no way to pick any more up (removed from LRU, removed
1442 * from pagecache). Can use non-atomic bitops now (and
1443 * we obviously don't have to worry about waking up a process
1444 * waiting on the page lock, because there are no references.
1446 __ClearPageLocked(page
);
1451 * Is there need to periodically free_page_list? It would
1452 * appear not as the counts should be low
1454 if (unlikely(PageTransHuge(page
))) {
1455 mem_cgroup_uncharge(page
);
1456 (*get_compound_page_dtor(page
))(page
);
1458 list_add(&page
->lru
, &free_pages
);
1462 /* Not a candidate for swapping, so reclaim swap space. */
1463 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1465 try_to_free_swap(page
);
1466 VM_BUG_ON_PAGE(PageActive(page
), page
);
1467 if (!PageMlocked(page
)) {
1468 SetPageActive(page
);
1470 count_memcg_page_event(page
, PGACTIVATE
);
1475 list_add(&page
->lru
, &ret_pages
);
1476 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1479 mem_cgroup_uncharge_list(&free_pages
);
1480 try_to_unmap_flush();
1481 free_unref_page_list(&free_pages
);
1483 list_splice(&ret_pages
, page_list
);
1484 count_vm_events(PGACTIVATE
, pgactivate
);
1487 stat
->nr_dirty
= nr_dirty
;
1488 stat
->nr_congested
= nr_congested
;
1489 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1490 stat
->nr_writeback
= nr_writeback
;
1491 stat
->nr_immediate
= nr_immediate
;
1492 stat
->nr_activate
= pgactivate
;
1493 stat
->nr_ref_keep
= nr_ref_keep
;
1494 stat
->nr_unmap_fail
= nr_unmap_fail
;
1496 return nr_reclaimed
;
1499 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1500 struct list_head
*page_list
)
1502 struct scan_control sc
= {
1503 .gfp_mask
= GFP_KERNEL
,
1504 .priority
= DEF_PRIORITY
,
1508 struct page
*page
, *next
;
1509 LIST_HEAD(clean_pages
);
1511 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1512 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1513 !__PageMovable(page
)) {
1514 ClearPageActive(page
);
1515 list_move(&page
->lru
, &clean_pages
);
1519 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1520 TTU_IGNORE_ACCESS
, NULL
, true);
1521 list_splice(&clean_pages
, page_list
);
1522 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1527 * Attempt to remove the specified page from its LRU. Only take this page
1528 * if it is of the appropriate PageActive status. Pages which are being
1529 * freed elsewhere are also ignored.
1531 * page: page to consider
1532 * mode: one of the LRU isolation modes defined above
1534 * returns 0 on success, -ve errno on failure.
1536 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1540 /* Only take pages on the LRU. */
1544 /* Compaction should not handle unevictable pages but CMA can do so */
1545 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1551 * To minimise LRU disruption, the caller can indicate that it only
1552 * wants to isolate pages it will be able to operate on without
1553 * blocking - clean pages for the most part.
1555 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1556 * that it is possible to migrate without blocking
1558 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1559 /* All the caller can do on PageWriteback is block */
1560 if (PageWriteback(page
))
1563 if (PageDirty(page
)) {
1564 struct address_space
*mapping
;
1568 * Only pages without mappings or that have a
1569 * ->migratepage callback are possible to migrate
1570 * without blocking. However, we can be racing with
1571 * truncation so it's necessary to lock the page
1572 * to stabilise the mapping as truncation holds
1573 * the page lock until after the page is removed
1574 * from the page cache.
1576 if (!trylock_page(page
))
1579 mapping
= page_mapping(page
);
1580 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1587 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1590 if (likely(get_page_unless_zero(page
))) {
1592 * Be careful not to clear PageLRU until after we're
1593 * sure the page is not being freed elsewhere -- the
1594 * page release code relies on it.
1605 * Update LRU sizes after isolating pages. The LRU size updates must
1606 * be complete before mem_cgroup_update_lru_size due to a santity check.
1608 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1609 enum lru_list lru
, unsigned long *nr_zone_taken
)
1613 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1614 if (!nr_zone_taken
[zid
])
1617 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1619 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1626 * zone_lru_lock is heavily contended. Some of the functions that
1627 * shrink the lists perform better by taking out a batch of pages
1628 * and working on them outside the LRU lock.
1630 * For pagecache intensive workloads, this function is the hottest
1631 * spot in the kernel (apart from copy_*_user functions).
1633 * Appropriate locks must be held before calling this function.
1635 * @nr_to_scan: The number of eligible pages to look through on the list.
1636 * @lruvec: The LRU vector to pull pages from.
1637 * @dst: The temp list to put pages on to.
1638 * @nr_scanned: The number of pages that were scanned.
1639 * @sc: The scan_control struct for this reclaim session
1640 * @mode: One of the LRU isolation modes
1641 * @lru: LRU list id for isolating
1643 * returns how many pages were moved onto *@dst.
1645 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1646 struct lruvec
*lruvec
, struct list_head
*dst
,
1647 unsigned long *nr_scanned
, struct scan_control
*sc
,
1648 isolate_mode_t mode
, enum lru_list lru
)
1650 struct list_head
*src
= &lruvec
->lists
[lru
];
1651 unsigned long nr_taken
= 0;
1652 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1653 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1654 unsigned long skipped
= 0;
1655 unsigned long scan
, total_scan
, nr_pages
;
1656 LIST_HEAD(pages_skipped
);
1659 for (total_scan
= 0;
1660 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1664 page
= lru_to_page(src
);
1665 prefetchw_prev_lru_page(page
, src
, flags
);
1667 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1669 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1670 list_move(&page
->lru
, &pages_skipped
);
1671 nr_skipped
[page_zonenum(page
)]++;
1676 * Do not count skipped pages because that makes the function
1677 * return with no isolated pages if the LRU mostly contains
1678 * ineligible pages. This causes the VM to not reclaim any
1679 * pages, triggering a premature OOM.
1682 switch (__isolate_lru_page(page
, mode
)) {
1684 nr_pages
= hpage_nr_pages(page
);
1685 nr_taken
+= nr_pages
;
1686 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1687 list_move(&page
->lru
, dst
);
1691 /* else it is being freed elsewhere */
1692 list_move(&page
->lru
, src
);
1701 * Splice any skipped pages to the start of the LRU list. Note that
1702 * this disrupts the LRU order when reclaiming for lower zones but
1703 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1704 * scanning would soon rescan the same pages to skip and put the
1705 * system at risk of premature OOM.
1707 if (!list_empty(&pages_skipped
)) {
1710 list_splice(&pages_skipped
, src
);
1711 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1712 if (!nr_skipped
[zid
])
1715 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1716 skipped
+= nr_skipped
[zid
];
1719 *nr_scanned
= total_scan
;
1720 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1721 total_scan
, skipped
, nr_taken
, mode
, lru
);
1722 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1727 * isolate_lru_page - tries to isolate a page from its LRU list
1728 * @page: page to isolate from its LRU list
1730 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1731 * vmstat statistic corresponding to whatever LRU list the page was on.
1733 * Returns 0 if the page was removed from an LRU list.
1734 * Returns -EBUSY if the page was not on an LRU list.
1736 * The returned page will have PageLRU() cleared. If it was found on
1737 * the active list, it will have PageActive set. If it was found on
1738 * the unevictable list, it will have the PageUnevictable bit set. That flag
1739 * may need to be cleared by the caller before letting the page go.
1741 * The vmstat statistic corresponding to the list on which the page was
1742 * found will be decremented.
1746 * (1) Must be called with an elevated refcount on the page. This is a
1747 * fundamentnal difference from isolate_lru_pages (which is called
1748 * without a stable reference).
1749 * (2) the lru_lock must not be held.
1750 * (3) interrupts must be enabled.
1752 int isolate_lru_page(struct page
*page
)
1756 VM_BUG_ON_PAGE(!page_count(page
), page
);
1757 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1759 if (PageLRU(page
)) {
1760 struct zone
*zone
= page_zone(page
);
1761 struct lruvec
*lruvec
;
1763 spin_lock_irq(zone_lru_lock(zone
));
1764 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1765 if (PageLRU(page
)) {
1766 int lru
= page_lru(page
);
1769 del_page_from_lru_list(page
, lruvec
, lru
);
1772 spin_unlock_irq(zone_lru_lock(zone
));
1778 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1779 * then get resheduled. When there are massive number of tasks doing page
1780 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1781 * the LRU list will go small and be scanned faster than necessary, leading to
1782 * unnecessary swapping, thrashing and OOM.
1784 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1785 struct scan_control
*sc
)
1787 unsigned long inactive
, isolated
;
1789 if (current_is_kswapd())
1792 if (!sane_reclaim(sc
))
1796 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1797 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1799 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1800 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1804 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1805 * won't get blocked by normal direct-reclaimers, forming a circular
1808 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1811 return isolated
> inactive
;
1814 static noinline_for_stack
void
1815 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1817 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1818 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1819 LIST_HEAD(pages_to_free
);
1822 * Put back any unfreeable pages.
1824 while (!list_empty(page_list
)) {
1825 struct page
*page
= lru_to_page(page_list
);
1828 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1829 list_del(&page
->lru
);
1830 if (unlikely(!page_evictable(page
))) {
1831 spin_unlock_irq(&pgdat
->lru_lock
);
1832 putback_lru_page(page
);
1833 spin_lock_irq(&pgdat
->lru_lock
);
1837 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1840 lru
= page_lru(page
);
1841 add_page_to_lru_list(page
, lruvec
, lru
);
1843 if (is_active_lru(lru
)) {
1844 int file
= is_file_lru(lru
);
1845 int numpages
= hpage_nr_pages(page
);
1846 reclaim_stat
->recent_rotated
[file
] += numpages
;
1848 if (put_page_testzero(page
)) {
1849 __ClearPageLRU(page
);
1850 __ClearPageActive(page
);
1851 del_page_from_lru_list(page
, lruvec
, lru
);
1853 if (unlikely(PageCompound(page
))) {
1854 spin_unlock_irq(&pgdat
->lru_lock
);
1855 mem_cgroup_uncharge(page
);
1856 (*get_compound_page_dtor(page
))(page
);
1857 spin_lock_irq(&pgdat
->lru_lock
);
1859 list_add(&page
->lru
, &pages_to_free
);
1864 * To save our caller's stack, now use input list for pages to free.
1866 list_splice(&pages_to_free
, page_list
);
1870 * If a kernel thread (such as nfsd for loop-back mounts) services
1871 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1872 * In that case we should only throttle if the backing device it is
1873 * writing to is congested. In other cases it is safe to throttle.
1875 static int current_may_throttle(void)
1877 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1878 current
->backing_dev_info
== NULL
||
1879 bdi_write_congested(current
->backing_dev_info
);
1883 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1884 * of reclaimed pages
1886 static noinline_for_stack
unsigned long
1887 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1888 struct scan_control
*sc
, enum lru_list lru
)
1890 LIST_HEAD(page_list
);
1891 unsigned long nr_scanned
;
1892 unsigned long nr_reclaimed
= 0;
1893 unsigned long nr_taken
;
1894 struct reclaim_stat stat
= {};
1895 isolate_mode_t isolate_mode
= 0;
1896 int file
= is_file_lru(lru
);
1897 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1898 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1899 bool stalled
= false;
1901 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1905 /* wait a bit for the reclaimer. */
1909 /* We are about to die and free our memory. Return now. */
1910 if (fatal_signal_pending(current
))
1911 return SWAP_CLUSTER_MAX
;
1917 isolate_mode
|= ISOLATE_UNMAPPED
;
1919 spin_lock_irq(&pgdat
->lru_lock
);
1921 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1922 &nr_scanned
, sc
, isolate_mode
, lru
);
1924 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1925 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1927 if (current_is_kswapd()) {
1928 if (global_reclaim(sc
))
1929 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1930 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1933 if (global_reclaim(sc
))
1934 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1935 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1938 spin_unlock_irq(&pgdat
->lru_lock
);
1943 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1946 spin_lock_irq(&pgdat
->lru_lock
);
1948 if (current_is_kswapd()) {
1949 if (global_reclaim(sc
))
1950 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1951 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1954 if (global_reclaim(sc
))
1955 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1956 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1960 putback_inactive_pages(lruvec
, &page_list
);
1962 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1964 spin_unlock_irq(&pgdat
->lru_lock
);
1966 mem_cgroup_uncharge_list(&page_list
);
1967 free_unref_page_list(&page_list
);
1970 * If dirty pages are scanned that are not queued for IO, it
1971 * implies that flushers are not doing their job. This can
1972 * happen when memory pressure pushes dirty pages to the end of
1973 * the LRU before the dirty limits are breached and the dirty
1974 * data has expired. It can also happen when the proportion of
1975 * dirty pages grows not through writes but through memory
1976 * pressure reclaiming all the clean cache. And in some cases,
1977 * the flushers simply cannot keep up with the allocation
1978 * rate. Nudge the flusher threads in case they are asleep.
1980 if (stat
.nr_unqueued_dirty
== nr_taken
)
1981 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1983 sc
->nr
.dirty
+= stat
.nr_dirty
;
1984 sc
->nr
.congested
+= stat
.nr_congested
;
1985 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1986 sc
->nr
.writeback
+= stat
.nr_writeback
;
1987 sc
->nr
.immediate
+= stat
.nr_immediate
;
1988 sc
->nr
.taken
+= nr_taken
;
1990 sc
->nr
.file_taken
+= nr_taken
;
1992 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1993 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
1994 return nr_reclaimed
;
1998 * This moves pages from the active list to the inactive list.
2000 * We move them the other way if the page is referenced by one or more
2001 * processes, from rmap.
2003 * If the pages are mostly unmapped, the processing is fast and it is
2004 * appropriate to hold zone_lru_lock across the whole operation. But if
2005 * the pages are mapped, the processing is slow (page_referenced()) so we
2006 * should drop zone_lru_lock around each page. It's impossible to balance
2007 * this, so instead we remove the pages from the LRU while processing them.
2008 * It is safe to rely on PG_active against the non-LRU pages in here because
2009 * nobody will play with that bit on a non-LRU page.
2011 * The downside is that we have to touch page->_refcount against each page.
2012 * But we had to alter page->flags anyway.
2014 * Returns the number of pages moved to the given lru.
2017 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
2018 struct list_head
*list
,
2019 struct list_head
*pages_to_free
,
2022 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2027 while (!list_empty(list
)) {
2028 page
= lru_to_page(list
);
2029 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2031 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2034 nr_pages
= hpage_nr_pages(page
);
2035 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
2036 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
2038 if (put_page_testzero(page
)) {
2039 __ClearPageLRU(page
);
2040 __ClearPageActive(page
);
2041 del_page_from_lru_list(page
, lruvec
, lru
);
2043 if (unlikely(PageCompound(page
))) {
2044 spin_unlock_irq(&pgdat
->lru_lock
);
2045 mem_cgroup_uncharge(page
);
2046 (*get_compound_page_dtor(page
))(page
);
2047 spin_lock_irq(&pgdat
->lru_lock
);
2049 list_add(&page
->lru
, pages_to_free
);
2051 nr_moved
+= nr_pages
;
2055 if (!is_active_lru(lru
)) {
2056 __count_vm_events(PGDEACTIVATE
, nr_moved
);
2057 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
2064 static void shrink_active_list(unsigned long nr_to_scan
,
2065 struct lruvec
*lruvec
,
2066 struct scan_control
*sc
,
2069 unsigned long nr_taken
;
2070 unsigned long nr_scanned
;
2071 unsigned long vm_flags
;
2072 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2073 LIST_HEAD(l_active
);
2074 LIST_HEAD(l_inactive
);
2076 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2077 unsigned nr_deactivate
, nr_activate
;
2078 unsigned nr_rotated
= 0;
2079 isolate_mode_t isolate_mode
= 0;
2080 int file
= is_file_lru(lru
);
2081 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2086 isolate_mode
|= ISOLATE_UNMAPPED
;
2088 spin_lock_irq(&pgdat
->lru_lock
);
2090 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2091 &nr_scanned
, sc
, isolate_mode
, lru
);
2093 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2094 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2096 __count_vm_events(PGREFILL
, nr_scanned
);
2097 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2099 spin_unlock_irq(&pgdat
->lru_lock
);
2101 while (!list_empty(&l_hold
)) {
2103 page
= lru_to_page(&l_hold
);
2104 list_del(&page
->lru
);
2106 if (unlikely(!page_evictable(page
))) {
2107 putback_lru_page(page
);
2111 if (unlikely(buffer_heads_over_limit
)) {
2112 if (page_has_private(page
) && trylock_page(page
)) {
2113 if (page_has_private(page
))
2114 try_to_release_page(page
, 0);
2119 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2121 nr_rotated
+= hpage_nr_pages(page
);
2123 * Identify referenced, file-backed active pages and
2124 * give them one more trip around the active list. So
2125 * that executable code get better chances to stay in
2126 * memory under moderate memory pressure. Anon pages
2127 * are not likely to be evicted by use-once streaming
2128 * IO, plus JVM can create lots of anon VM_EXEC pages,
2129 * so we ignore them here.
2131 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2132 list_add(&page
->lru
, &l_active
);
2137 ClearPageActive(page
); /* we are de-activating */
2138 list_add(&page
->lru
, &l_inactive
);
2142 * Move pages back to the lru list.
2144 spin_lock_irq(&pgdat
->lru_lock
);
2146 * Count referenced pages from currently used mappings as rotated,
2147 * even though only some of them are actually re-activated. This
2148 * helps balance scan pressure between file and anonymous pages in
2151 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2153 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2154 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2155 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2156 spin_unlock_irq(&pgdat
->lru_lock
);
2158 mem_cgroup_uncharge_list(&l_hold
);
2159 free_unref_page_list(&l_hold
);
2160 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2161 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2165 * The inactive anon list should be small enough that the VM never has
2166 * to do too much work.
2168 * The inactive file list should be small enough to leave most memory
2169 * to the established workingset on the scan-resistant active list,
2170 * but large enough to avoid thrashing the aggregate readahead window.
2172 * Both inactive lists should also be large enough that each inactive
2173 * page has a chance to be referenced again before it is reclaimed.
2175 * If that fails and refaulting is observed, the inactive list grows.
2177 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2178 * on this LRU, maintained by the pageout code. An inactive_ratio
2179 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2182 * memory ratio inactive
2183 * -------------------------------------
2192 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2193 struct mem_cgroup
*memcg
,
2194 struct scan_control
*sc
, bool actual_reclaim
)
2196 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2197 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2198 enum lru_list inactive_lru
= file
* LRU_FILE
;
2199 unsigned long inactive
, active
;
2200 unsigned long inactive_ratio
;
2201 unsigned long refaults
;
2205 * If we don't have swap space, anonymous page deactivation
2208 if (!file
&& !total_swap_pages
)
2211 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2212 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2215 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2217 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2220 * When refaults are being observed, it means a new workingset
2221 * is being established. Disable active list protection to get
2222 * rid of the stale workingset quickly.
2224 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2227 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2229 inactive_ratio
= int_sqrt(10 * gb
);
2235 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2236 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2237 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2238 inactive_ratio
, file
);
2240 return inactive
* inactive_ratio
< active
;
2243 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2244 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2245 struct scan_control
*sc
)
2247 if (is_active_lru(lru
)) {
2248 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2250 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2254 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2265 * Determine how aggressively the anon and file LRU lists should be
2266 * scanned. The relative value of each set of LRU lists is determined
2267 * by looking at the fraction of the pages scanned we did rotate back
2268 * onto the active list instead of evict.
2270 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2271 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2273 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2274 struct scan_control
*sc
, unsigned long *nr
,
2275 unsigned long *lru_pages
)
2277 int swappiness
= mem_cgroup_swappiness(memcg
);
2278 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2280 u64 denominator
= 0; /* gcc */
2281 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2282 unsigned long anon_prio
, file_prio
;
2283 enum scan_balance scan_balance
;
2284 unsigned long anon
, file
;
2285 unsigned long ap
, fp
;
2288 /* If we have no swap space, do not bother scanning anon pages. */
2289 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2290 scan_balance
= SCAN_FILE
;
2295 * Global reclaim will swap to prevent OOM even with no
2296 * swappiness, but memcg users want to use this knob to
2297 * disable swapping for individual groups completely when
2298 * using the memory controller's swap limit feature would be
2301 if (!global_reclaim(sc
) && !swappiness
) {
2302 scan_balance
= SCAN_FILE
;
2307 * Do not apply any pressure balancing cleverness when the
2308 * system is close to OOM, scan both anon and file equally
2309 * (unless the swappiness setting disagrees with swapping).
2311 if (!sc
->priority
&& swappiness
) {
2312 scan_balance
= SCAN_EQUAL
;
2317 * Prevent the reclaimer from falling into the cache trap: as
2318 * cache pages start out inactive, every cache fault will tip
2319 * the scan balance towards the file LRU. And as the file LRU
2320 * shrinks, so does the window for rotation from references.
2321 * This means we have a runaway feedback loop where a tiny
2322 * thrashing file LRU becomes infinitely more attractive than
2323 * anon pages. Try to detect this based on file LRU size.
2325 if (global_reclaim(sc
)) {
2326 unsigned long pgdatfile
;
2327 unsigned long pgdatfree
;
2329 unsigned long total_high_wmark
= 0;
2331 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2332 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2333 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2335 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2336 struct zone
*zone
= &pgdat
->node_zones
[z
];
2337 if (!managed_zone(zone
))
2340 total_high_wmark
+= high_wmark_pages(zone
);
2343 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2345 * Force SCAN_ANON if there are enough inactive
2346 * anonymous pages on the LRU in eligible zones.
2347 * Otherwise, the small LRU gets thrashed.
2349 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2350 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2352 scan_balance
= SCAN_ANON
;
2359 * If there is enough inactive page cache, i.e. if the size of the
2360 * inactive list is greater than that of the active list *and* the
2361 * inactive list actually has some pages to scan on this priority, we
2362 * do not reclaim anything from the anonymous working set right now.
2363 * Without the second condition we could end up never scanning an
2364 * lruvec even if it has plenty of old anonymous pages unless the
2365 * system is under heavy pressure.
2367 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2368 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2369 scan_balance
= SCAN_FILE
;
2373 scan_balance
= SCAN_FRACT
;
2376 * With swappiness at 100, anonymous and file have the same priority.
2377 * This scanning priority is essentially the inverse of IO cost.
2379 anon_prio
= swappiness
;
2380 file_prio
= 200 - anon_prio
;
2383 * OK, so we have swap space and a fair amount of page cache
2384 * pages. We use the recently rotated / recently scanned
2385 * ratios to determine how valuable each cache is.
2387 * Because workloads change over time (and to avoid overflow)
2388 * we keep these statistics as a floating average, which ends
2389 * up weighing recent references more than old ones.
2391 * anon in [0], file in [1]
2394 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2395 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2396 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2397 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2399 spin_lock_irq(&pgdat
->lru_lock
);
2400 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2401 reclaim_stat
->recent_scanned
[0] /= 2;
2402 reclaim_stat
->recent_rotated
[0] /= 2;
2405 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2406 reclaim_stat
->recent_scanned
[1] /= 2;
2407 reclaim_stat
->recent_rotated
[1] /= 2;
2411 * The amount of pressure on anon vs file pages is inversely
2412 * proportional to the fraction of recently scanned pages on
2413 * each list that were recently referenced and in active use.
2415 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2416 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2418 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2419 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2420 spin_unlock_irq(&pgdat
->lru_lock
);
2424 denominator
= ap
+ fp
+ 1;
2427 for_each_evictable_lru(lru
) {
2428 int file
= is_file_lru(lru
);
2432 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2433 scan
= size
>> sc
->priority
;
2435 * If the cgroup's already been deleted, make sure to
2436 * scrape out the remaining cache.
2438 if (!scan
&& !mem_cgroup_online(memcg
))
2439 scan
= min(size
, SWAP_CLUSTER_MAX
);
2441 switch (scan_balance
) {
2443 /* Scan lists relative to size */
2447 * Scan types proportional to swappiness and
2448 * their relative recent reclaim efficiency.
2450 scan
= div64_u64(scan
* fraction
[file
],
2455 /* Scan one type exclusively */
2456 if ((scan_balance
== SCAN_FILE
) != file
) {
2462 /* Look ma, no brain */
2472 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2474 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2475 struct scan_control
*sc
, unsigned long *lru_pages
)
2477 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2478 unsigned long nr
[NR_LRU_LISTS
];
2479 unsigned long targets
[NR_LRU_LISTS
];
2480 unsigned long nr_to_scan
;
2482 unsigned long nr_reclaimed
= 0;
2483 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2484 struct blk_plug plug
;
2487 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2489 /* Record the original scan target for proportional adjustments later */
2490 memcpy(targets
, nr
, sizeof(nr
));
2493 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2494 * event that can occur when there is little memory pressure e.g.
2495 * multiple streaming readers/writers. Hence, we do not abort scanning
2496 * when the requested number of pages are reclaimed when scanning at
2497 * DEF_PRIORITY on the assumption that the fact we are direct
2498 * reclaiming implies that kswapd is not keeping up and it is best to
2499 * do a batch of work at once. For memcg reclaim one check is made to
2500 * abort proportional reclaim if either the file or anon lru has already
2501 * dropped to zero at the first pass.
2503 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2504 sc
->priority
== DEF_PRIORITY
);
2506 blk_start_plug(&plug
);
2507 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2508 nr
[LRU_INACTIVE_FILE
]) {
2509 unsigned long nr_anon
, nr_file
, percentage
;
2510 unsigned long nr_scanned
;
2512 for_each_evictable_lru(lru
) {
2514 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2515 nr
[lru
] -= nr_to_scan
;
2517 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2524 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2528 * For kswapd and memcg, reclaim at least the number of pages
2529 * requested. Ensure that the anon and file LRUs are scanned
2530 * proportionally what was requested by get_scan_count(). We
2531 * stop reclaiming one LRU and reduce the amount scanning
2532 * proportional to the original scan target.
2534 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2535 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2538 * It's just vindictive to attack the larger once the smaller
2539 * has gone to zero. And given the way we stop scanning the
2540 * smaller below, this makes sure that we only make one nudge
2541 * towards proportionality once we've got nr_to_reclaim.
2543 if (!nr_file
|| !nr_anon
)
2546 if (nr_file
> nr_anon
) {
2547 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2548 targets
[LRU_ACTIVE_ANON
] + 1;
2550 percentage
= nr_anon
* 100 / scan_target
;
2552 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2553 targets
[LRU_ACTIVE_FILE
] + 1;
2555 percentage
= nr_file
* 100 / scan_target
;
2558 /* Stop scanning the smaller of the LRU */
2560 nr
[lru
+ LRU_ACTIVE
] = 0;
2563 * Recalculate the other LRU scan count based on its original
2564 * scan target and the percentage scanning already complete
2566 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2567 nr_scanned
= targets
[lru
] - nr
[lru
];
2568 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2569 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2572 nr_scanned
= targets
[lru
] - nr
[lru
];
2573 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2574 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2576 scan_adjusted
= true;
2578 blk_finish_plug(&plug
);
2579 sc
->nr_reclaimed
+= nr_reclaimed
;
2582 * Even if we did not try to evict anon pages at all, we want to
2583 * rebalance the anon lru active/inactive ratio.
2585 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2586 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2587 sc
, LRU_ACTIVE_ANON
);
2590 /* Use reclaim/compaction for costly allocs or under memory pressure */
2591 static bool in_reclaim_compaction(struct scan_control
*sc
)
2593 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2594 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2595 sc
->priority
< DEF_PRIORITY
- 2))
2602 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2603 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2604 * true if more pages should be reclaimed such that when the page allocator
2605 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2606 * It will give up earlier than that if there is difficulty reclaiming pages.
2608 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2609 unsigned long nr_reclaimed
,
2610 unsigned long nr_scanned
,
2611 struct scan_control
*sc
)
2613 unsigned long pages_for_compaction
;
2614 unsigned long inactive_lru_pages
;
2617 /* If not in reclaim/compaction mode, stop */
2618 if (!in_reclaim_compaction(sc
))
2621 /* Consider stopping depending on scan and reclaim activity */
2622 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2624 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2625 * full LRU list has been scanned and we are still failing
2626 * to reclaim pages. This full LRU scan is potentially
2627 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2629 if (!nr_reclaimed
&& !nr_scanned
)
2633 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2634 * fail without consequence, stop if we failed to reclaim
2635 * any pages from the last SWAP_CLUSTER_MAX number of
2636 * pages that were scanned. This will return to the
2637 * caller faster at the risk reclaim/compaction and
2638 * the resulting allocation attempt fails
2645 * If we have not reclaimed enough pages for compaction and the
2646 * inactive lists are large enough, continue reclaiming
2648 pages_for_compaction
= compact_gap(sc
->order
);
2649 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2650 if (get_nr_swap_pages() > 0)
2651 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2652 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2653 inactive_lru_pages
> pages_for_compaction
)
2656 /* If compaction would go ahead or the allocation would succeed, stop */
2657 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2658 struct zone
*zone
= &pgdat
->node_zones
[z
];
2659 if (!managed_zone(zone
))
2662 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2663 case COMPACT_SUCCESS
:
2664 case COMPACT_CONTINUE
:
2667 /* check next zone */
2674 static bool pgdat_memcg_congested(pg_data_t
*pgdat
, struct mem_cgroup
*memcg
)
2676 return test_bit(PGDAT_CONGESTED
, &pgdat
->flags
) ||
2677 (memcg
&& memcg_congested(pgdat
, memcg
));
2680 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2682 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2683 unsigned long nr_reclaimed
, nr_scanned
;
2684 bool reclaimable
= false;
2687 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2688 struct mem_cgroup_reclaim_cookie reclaim
= {
2690 .priority
= sc
->priority
,
2692 unsigned long node_lru_pages
= 0;
2693 struct mem_cgroup
*memcg
;
2695 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
2697 nr_reclaimed
= sc
->nr_reclaimed
;
2698 nr_scanned
= sc
->nr_scanned
;
2700 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2702 unsigned long lru_pages
;
2703 unsigned long reclaimed
;
2704 unsigned long scanned
;
2706 switch (mem_cgroup_protected(root
, memcg
)) {
2707 case MEMCG_PROT_MIN
:
2710 * If there is no reclaimable memory, OOM.
2713 case MEMCG_PROT_LOW
:
2716 * Respect the protection only as long as
2717 * there is an unprotected supply
2718 * of reclaimable memory from other cgroups.
2720 if (!sc
->memcg_low_reclaim
) {
2721 sc
->memcg_low_skipped
= 1;
2724 memcg_memory_event(memcg
, MEMCG_LOW
);
2726 case MEMCG_PROT_NONE
:
2730 reclaimed
= sc
->nr_reclaimed
;
2731 scanned
= sc
->nr_scanned
;
2732 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2733 node_lru_pages
+= lru_pages
;
2735 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2736 memcg
, sc
->priority
);
2738 /* Record the group's reclaim efficiency */
2739 vmpressure(sc
->gfp_mask
, memcg
, false,
2740 sc
->nr_scanned
- scanned
,
2741 sc
->nr_reclaimed
- reclaimed
);
2744 * Direct reclaim and kswapd have to scan all memory
2745 * cgroups to fulfill the overall scan target for the
2748 * Limit reclaim, on the other hand, only cares about
2749 * nr_to_reclaim pages to be reclaimed and it will
2750 * retry with decreasing priority if one round over the
2751 * whole hierarchy is not sufficient.
2753 if (!global_reclaim(sc
) &&
2754 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2755 mem_cgroup_iter_break(root
, memcg
);
2758 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2760 if (reclaim_state
) {
2761 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2762 reclaim_state
->reclaimed_slab
= 0;
2765 /* Record the subtree's reclaim efficiency */
2766 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2767 sc
->nr_scanned
- nr_scanned
,
2768 sc
->nr_reclaimed
- nr_reclaimed
);
2770 if (sc
->nr_reclaimed
- nr_reclaimed
)
2773 if (current_is_kswapd()) {
2775 * If reclaim is isolating dirty pages under writeback,
2776 * it implies that the long-lived page allocation rate
2777 * is exceeding the page laundering rate. Either the
2778 * global limits are not being effective at throttling
2779 * processes due to the page distribution throughout
2780 * zones or there is heavy usage of a slow backing
2781 * device. The only option is to throttle from reclaim
2782 * context which is not ideal as there is no guarantee
2783 * the dirtying process is throttled in the same way
2784 * balance_dirty_pages() manages.
2786 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2787 * count the number of pages under pages flagged for
2788 * immediate reclaim and stall if any are encountered
2789 * in the nr_immediate check below.
2791 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
2792 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
2795 * Tag a node as congested if all the dirty pages
2796 * scanned were backed by a congested BDI and
2797 * wait_iff_congested will stall.
2799 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2800 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
2802 /* Allow kswapd to start writing pages during reclaim.*/
2803 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
2804 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
2807 * If kswapd scans pages marked marked for immediate
2808 * reclaim and under writeback (nr_immediate), it
2809 * implies that pages are cycling through the LRU
2810 * faster than they are written so also forcibly stall.
2812 if (sc
->nr
.immediate
)
2813 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2817 * Legacy memcg will stall in page writeback so avoid forcibly
2818 * stalling in wait_iff_congested().
2820 if (!global_reclaim(sc
) && sane_reclaim(sc
) &&
2821 sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
)
2822 set_memcg_congestion(pgdat
, root
, true);
2825 * Stall direct reclaim for IO completions if underlying BDIs
2826 * and node is congested. Allow kswapd to continue until it
2827 * starts encountering unqueued dirty pages or cycling through
2828 * the LRU too quickly.
2830 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
2831 current_may_throttle() && pgdat_memcg_congested(pgdat
, root
))
2832 wait_iff_congested(BLK_RW_ASYNC
, HZ
/10);
2834 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2835 sc
->nr_scanned
- nr_scanned
, sc
));
2838 * Kswapd gives up on balancing particular nodes after too
2839 * many failures to reclaim anything from them and goes to
2840 * sleep. On reclaim progress, reset the failure counter. A
2841 * successful direct reclaim run will revive a dormant kswapd.
2844 pgdat
->kswapd_failures
= 0;
2850 * Returns true if compaction should go ahead for a costly-order request, or
2851 * the allocation would already succeed without compaction. Return false if we
2852 * should reclaim first.
2854 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2856 unsigned long watermark
;
2857 enum compact_result suitable
;
2859 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2860 if (suitable
== COMPACT_SUCCESS
)
2861 /* Allocation should succeed already. Don't reclaim. */
2863 if (suitable
== COMPACT_SKIPPED
)
2864 /* Compaction cannot yet proceed. Do reclaim. */
2868 * Compaction is already possible, but it takes time to run and there
2869 * are potentially other callers using the pages just freed. So proceed
2870 * with reclaim to make a buffer of free pages available to give
2871 * compaction a reasonable chance of completing and allocating the page.
2872 * Note that we won't actually reclaim the whole buffer in one attempt
2873 * as the target watermark in should_continue_reclaim() is lower. But if
2874 * we are already above the high+gap watermark, don't reclaim at all.
2876 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2878 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2882 * This is the direct reclaim path, for page-allocating processes. We only
2883 * try to reclaim pages from zones which will satisfy the caller's allocation
2886 * If a zone is deemed to be full of pinned pages then just give it a light
2887 * scan then give up on it.
2889 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2893 unsigned long nr_soft_reclaimed
;
2894 unsigned long nr_soft_scanned
;
2896 pg_data_t
*last_pgdat
= NULL
;
2899 * If the number of buffer_heads in the machine exceeds the maximum
2900 * allowed level, force direct reclaim to scan the highmem zone as
2901 * highmem pages could be pinning lowmem pages storing buffer_heads
2903 orig_mask
= sc
->gfp_mask
;
2904 if (buffer_heads_over_limit
) {
2905 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2906 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2909 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2910 sc
->reclaim_idx
, sc
->nodemask
) {
2912 * Take care memory controller reclaiming has small influence
2915 if (global_reclaim(sc
)) {
2916 if (!cpuset_zone_allowed(zone
,
2917 GFP_KERNEL
| __GFP_HARDWALL
))
2921 * If we already have plenty of memory free for
2922 * compaction in this zone, don't free any more.
2923 * Even though compaction is invoked for any
2924 * non-zero order, only frequent costly order
2925 * reclamation is disruptive enough to become a
2926 * noticeable problem, like transparent huge
2929 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2930 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2931 compaction_ready(zone
, sc
)) {
2932 sc
->compaction_ready
= true;
2937 * Shrink each node in the zonelist once. If the
2938 * zonelist is ordered by zone (not the default) then a
2939 * node may be shrunk multiple times but in that case
2940 * the user prefers lower zones being preserved.
2942 if (zone
->zone_pgdat
== last_pgdat
)
2946 * This steals pages from memory cgroups over softlimit
2947 * and returns the number of reclaimed pages and
2948 * scanned pages. This works for global memory pressure
2949 * and balancing, not for a memcg's limit.
2951 nr_soft_scanned
= 0;
2952 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2953 sc
->order
, sc
->gfp_mask
,
2955 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2956 sc
->nr_scanned
+= nr_soft_scanned
;
2957 /* need some check for avoid more shrink_zone() */
2960 /* See comment about same check for global reclaim above */
2961 if (zone
->zone_pgdat
== last_pgdat
)
2963 last_pgdat
= zone
->zone_pgdat
;
2964 shrink_node(zone
->zone_pgdat
, sc
);
2968 * Restore to original mask to avoid the impact on the caller if we
2969 * promoted it to __GFP_HIGHMEM.
2971 sc
->gfp_mask
= orig_mask
;
2974 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2976 struct mem_cgroup
*memcg
;
2978 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2980 unsigned long refaults
;
2981 struct lruvec
*lruvec
;
2984 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2986 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2988 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2989 lruvec
->refaults
= refaults
;
2990 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2994 * This is the main entry point to direct page reclaim.
2996 * If a full scan of the inactive list fails to free enough memory then we
2997 * are "out of memory" and something needs to be killed.
2999 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3000 * high - the zone may be full of dirty or under-writeback pages, which this
3001 * caller can't do much about. We kick the writeback threads and take explicit
3002 * naps in the hope that some of these pages can be written. But if the
3003 * allocating task holds filesystem locks which prevent writeout this might not
3004 * work, and the allocation attempt will fail.
3006 * returns: 0, if no pages reclaimed
3007 * else, the number of pages reclaimed
3009 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
3010 struct scan_control
*sc
)
3012 int initial_priority
= sc
->priority
;
3013 pg_data_t
*last_pgdat
;
3017 delayacct_freepages_start();
3019 if (global_reclaim(sc
))
3020 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
3023 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
3026 shrink_zones(zonelist
, sc
);
3028 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
3031 if (sc
->compaction_ready
)
3035 * If we're getting trouble reclaiming, start doing
3036 * writepage even in laptop mode.
3038 if (sc
->priority
< DEF_PRIORITY
- 2)
3039 sc
->may_writepage
= 1;
3040 } while (--sc
->priority
>= 0);
3043 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
3045 if (zone
->zone_pgdat
== last_pgdat
)
3047 last_pgdat
= zone
->zone_pgdat
;
3048 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
3049 set_memcg_congestion(last_pgdat
, sc
->target_mem_cgroup
, false);
3052 delayacct_freepages_end();
3054 if (sc
->nr_reclaimed
)
3055 return sc
->nr_reclaimed
;
3057 /* Aborted reclaim to try compaction? don't OOM, then */
3058 if (sc
->compaction_ready
)
3061 /* Untapped cgroup reserves? Don't OOM, retry. */
3062 if (sc
->memcg_low_skipped
) {
3063 sc
->priority
= initial_priority
;
3064 sc
->memcg_low_reclaim
= 1;
3065 sc
->memcg_low_skipped
= 0;
3072 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
3075 unsigned long pfmemalloc_reserve
= 0;
3076 unsigned long free_pages
= 0;
3080 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3083 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
3084 zone
= &pgdat
->node_zones
[i
];
3085 if (!managed_zone(zone
))
3088 if (!zone_reclaimable_pages(zone
))
3091 pfmemalloc_reserve
+= min_wmark_pages(zone
);
3092 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
3095 /* If there are no reserves (unexpected config) then do not throttle */
3096 if (!pfmemalloc_reserve
)
3099 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
3101 /* kswapd must be awake if processes are being throttled */
3102 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
3103 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
3104 (enum zone_type
)ZONE_NORMAL
);
3105 wake_up_interruptible(&pgdat
->kswapd_wait
);
3112 * Throttle direct reclaimers if backing storage is backed by the network
3113 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3114 * depleted. kswapd will continue to make progress and wake the processes
3115 * when the low watermark is reached.
3117 * Returns true if a fatal signal was delivered during throttling. If this
3118 * happens, the page allocator should not consider triggering the OOM killer.
3120 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
3121 nodemask_t
*nodemask
)
3125 pg_data_t
*pgdat
= NULL
;
3128 * Kernel threads should not be throttled as they may be indirectly
3129 * responsible for cleaning pages necessary for reclaim to make forward
3130 * progress. kjournald for example may enter direct reclaim while
3131 * committing a transaction where throttling it could forcing other
3132 * processes to block on log_wait_commit().
3134 if (current
->flags
& PF_KTHREAD
)
3138 * If a fatal signal is pending, this process should not throttle.
3139 * It should return quickly so it can exit and free its memory
3141 if (fatal_signal_pending(current
))
3145 * Check if the pfmemalloc reserves are ok by finding the first node
3146 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3147 * GFP_KERNEL will be required for allocating network buffers when
3148 * swapping over the network so ZONE_HIGHMEM is unusable.
3150 * Throttling is based on the first usable node and throttled processes
3151 * wait on a queue until kswapd makes progress and wakes them. There
3152 * is an affinity then between processes waking up and where reclaim
3153 * progress has been made assuming the process wakes on the same node.
3154 * More importantly, processes running on remote nodes will not compete
3155 * for remote pfmemalloc reserves and processes on different nodes
3156 * should make reasonable progress.
3158 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3159 gfp_zone(gfp_mask
), nodemask
) {
3160 if (zone_idx(zone
) > ZONE_NORMAL
)
3163 /* Throttle based on the first usable node */
3164 pgdat
= zone
->zone_pgdat
;
3165 if (allow_direct_reclaim(pgdat
))
3170 /* If no zone was usable by the allocation flags then do not throttle */
3174 /* Account for the throttling */
3175 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3178 * If the caller cannot enter the filesystem, it's possible that it
3179 * is due to the caller holding an FS lock or performing a journal
3180 * transaction in the case of a filesystem like ext[3|4]. In this case,
3181 * it is not safe to block on pfmemalloc_wait as kswapd could be
3182 * blocked waiting on the same lock. Instead, throttle for up to a
3183 * second before continuing.
3185 if (!(gfp_mask
& __GFP_FS
)) {
3186 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3187 allow_direct_reclaim(pgdat
), HZ
);
3192 /* Throttle until kswapd wakes the process */
3193 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3194 allow_direct_reclaim(pgdat
));
3197 if (fatal_signal_pending(current
))
3204 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3205 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3207 unsigned long nr_reclaimed
;
3208 struct scan_control sc
= {
3209 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3210 .gfp_mask
= current_gfp_context(gfp_mask
),
3211 .reclaim_idx
= gfp_zone(gfp_mask
),
3213 .nodemask
= nodemask
,
3214 .priority
= DEF_PRIORITY
,
3215 .may_writepage
= !laptop_mode
,
3221 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3222 * Confirm they are large enough for max values.
3224 BUILD_BUG_ON(MAX_ORDER
> S8_MAX
);
3225 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
3226 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
3229 * Do not enter reclaim if fatal signal was delivered while throttled.
3230 * 1 is returned so that the page allocator does not OOM kill at this
3233 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3236 trace_mm_vmscan_direct_reclaim_begin(order
,
3241 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3243 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3245 return nr_reclaimed
;
3250 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3251 gfp_t gfp_mask
, bool noswap
,
3253 unsigned long *nr_scanned
)
3255 struct scan_control sc
= {
3256 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3257 .target_mem_cgroup
= memcg
,
3258 .may_writepage
= !laptop_mode
,
3260 .reclaim_idx
= MAX_NR_ZONES
- 1,
3261 .may_swap
= !noswap
,
3263 unsigned long lru_pages
;
3265 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3266 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3268 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3274 * NOTE: Although we can get the priority field, using it
3275 * here is not a good idea, since it limits the pages we can scan.
3276 * if we don't reclaim here, the shrink_node from balance_pgdat
3277 * will pick up pages from other mem cgroup's as well. We hack
3278 * the priority and make it zero.
3280 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3282 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3284 *nr_scanned
= sc
.nr_scanned
;
3285 return sc
.nr_reclaimed
;
3288 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3289 unsigned long nr_pages
,
3293 struct zonelist
*zonelist
;
3294 unsigned long nr_reclaimed
;
3296 unsigned int noreclaim_flag
;
3297 struct scan_control sc
= {
3298 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3299 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3300 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3301 .reclaim_idx
= MAX_NR_ZONES
- 1,
3302 .target_mem_cgroup
= memcg
,
3303 .priority
= DEF_PRIORITY
,
3304 .may_writepage
= !laptop_mode
,
3306 .may_swap
= may_swap
,
3310 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3311 * take care of from where we get pages. So the node where we start the
3312 * scan does not need to be the current node.
3314 nid
= mem_cgroup_select_victim_node(memcg
);
3316 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3318 trace_mm_vmscan_memcg_reclaim_begin(0,
3323 noreclaim_flag
= memalloc_noreclaim_save();
3324 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3325 memalloc_noreclaim_restore(noreclaim_flag
);
3327 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3329 return nr_reclaimed
;
3333 static void age_active_anon(struct pglist_data
*pgdat
,
3334 struct scan_control
*sc
)
3336 struct mem_cgroup
*memcg
;
3338 if (!total_swap_pages
)
3341 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3343 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3345 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3346 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3347 sc
, LRU_ACTIVE_ANON
);
3349 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3354 * Returns true if there is an eligible zone balanced for the request order
3357 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3360 unsigned long mark
= -1;
3363 for (i
= 0; i
<= classzone_idx
; i
++) {
3364 zone
= pgdat
->node_zones
+ i
;
3366 if (!managed_zone(zone
))
3369 mark
= high_wmark_pages(zone
);
3370 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3375 * If a node has no populated zone within classzone_idx, it does not
3376 * need balancing by definition. This can happen if a zone-restricted
3377 * allocation tries to wake a remote kswapd.
3385 /* Clear pgdat state for congested, dirty or under writeback. */
3386 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3388 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3389 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3390 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3394 * Prepare kswapd for sleeping. This verifies that there are no processes
3395 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3397 * Returns true if kswapd is ready to sleep
3399 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3402 * The throttled processes are normally woken up in balance_pgdat() as
3403 * soon as allow_direct_reclaim() is true. But there is a potential
3404 * race between when kswapd checks the watermarks and a process gets
3405 * throttled. There is also a potential race if processes get
3406 * throttled, kswapd wakes, a large process exits thereby balancing the
3407 * zones, which causes kswapd to exit balance_pgdat() before reaching
3408 * the wake up checks. If kswapd is going to sleep, no process should
3409 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3410 * the wake up is premature, processes will wake kswapd and get
3411 * throttled again. The difference from wake ups in balance_pgdat() is
3412 * that here we are under prepare_to_wait().
3414 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3415 wake_up_all(&pgdat
->pfmemalloc_wait
);
3417 /* Hopeless node, leave it to direct reclaim */
3418 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3421 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3422 clear_pgdat_congested(pgdat
);
3430 * kswapd shrinks a node of pages that are at or below the highest usable
3431 * zone that is currently unbalanced.
3433 * Returns true if kswapd scanned at least the requested number of pages to
3434 * reclaim or if the lack of progress was due to pages under writeback.
3435 * This is used to determine if the scanning priority needs to be raised.
3437 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3438 struct scan_control
*sc
)
3443 /* Reclaim a number of pages proportional to the number of zones */
3444 sc
->nr_to_reclaim
= 0;
3445 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3446 zone
= pgdat
->node_zones
+ z
;
3447 if (!managed_zone(zone
))
3450 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3454 * Historically care was taken to put equal pressure on all zones but
3455 * now pressure is applied based on node LRU order.
3457 shrink_node(pgdat
, sc
);
3460 * Fragmentation may mean that the system cannot be rebalanced for
3461 * high-order allocations. If twice the allocation size has been
3462 * reclaimed then recheck watermarks only at order-0 to prevent
3463 * excessive reclaim. Assume that a process requested a high-order
3464 * can direct reclaim/compact.
3466 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3469 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3473 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3474 * that are eligible for use by the caller until at least one zone is
3477 * Returns the order kswapd finished reclaiming at.
3479 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3480 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3481 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3482 * or lower is eligible for reclaim until at least one usable zone is
3485 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3488 unsigned long nr_soft_reclaimed
;
3489 unsigned long nr_soft_scanned
;
3491 struct scan_control sc
= {
3492 .gfp_mask
= GFP_KERNEL
,
3494 .priority
= DEF_PRIORITY
,
3495 .may_writepage
= !laptop_mode
,
3500 __fs_reclaim_acquire();
3502 count_vm_event(PAGEOUTRUN
);
3505 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3506 bool raise_priority
= true;
3509 sc
.reclaim_idx
= classzone_idx
;
3512 * If the number of buffer_heads exceeds the maximum allowed
3513 * then consider reclaiming from all zones. This has a dual
3514 * purpose -- on 64-bit systems it is expected that
3515 * buffer_heads are stripped during active rotation. On 32-bit
3516 * systems, highmem pages can pin lowmem memory and shrinking
3517 * buffers can relieve lowmem pressure. Reclaim may still not
3518 * go ahead if all eligible zones for the original allocation
3519 * request are balanced to avoid excessive reclaim from kswapd.
3521 if (buffer_heads_over_limit
) {
3522 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3523 zone
= pgdat
->node_zones
+ i
;
3524 if (!managed_zone(zone
))
3533 * Only reclaim if there are no eligible zones. Note that
3534 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3537 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3541 * Do some background aging of the anon list, to give
3542 * pages a chance to be referenced before reclaiming. All
3543 * pages are rotated regardless of classzone as this is
3544 * about consistent aging.
3546 age_active_anon(pgdat
, &sc
);
3549 * If we're getting trouble reclaiming, start doing writepage
3550 * even in laptop mode.
3552 if (sc
.priority
< DEF_PRIORITY
- 2)
3553 sc
.may_writepage
= 1;
3555 /* Call soft limit reclaim before calling shrink_node. */
3557 nr_soft_scanned
= 0;
3558 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3559 sc
.gfp_mask
, &nr_soft_scanned
);
3560 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3563 * There should be no need to raise the scanning priority if
3564 * enough pages are already being scanned that that high
3565 * watermark would be met at 100% efficiency.
3567 if (kswapd_shrink_node(pgdat
, &sc
))
3568 raise_priority
= false;
3571 * If the low watermark is met there is no need for processes
3572 * to be throttled on pfmemalloc_wait as they should not be
3573 * able to safely make forward progress. Wake them
3575 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3576 allow_direct_reclaim(pgdat
))
3577 wake_up_all(&pgdat
->pfmemalloc_wait
);
3579 /* Check if kswapd should be suspending */
3580 __fs_reclaim_release();
3581 ret
= try_to_freeze();
3582 __fs_reclaim_acquire();
3583 if (ret
|| kthread_should_stop())
3587 * Raise priority if scanning rate is too low or there was no
3588 * progress in reclaiming pages
3590 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3591 if (raise_priority
|| !nr_reclaimed
)
3593 } while (sc
.priority
>= 1);
3595 if (!sc
.nr_reclaimed
)
3596 pgdat
->kswapd_failures
++;
3599 snapshot_refaults(NULL
, pgdat
);
3600 __fs_reclaim_release();
3602 * Return the order kswapd stopped reclaiming at as
3603 * prepare_kswapd_sleep() takes it into account. If another caller
3604 * entered the allocator slow path while kswapd was awake, order will
3605 * remain at the higher level.
3611 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3612 * allocation request woke kswapd for. When kswapd has not woken recently,
3613 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3614 * given classzone and returns it or the highest classzone index kswapd
3615 * was recently woke for.
3617 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3618 enum zone_type classzone_idx
)
3620 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3621 return classzone_idx
;
3623 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3626 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3627 unsigned int classzone_idx
)
3632 if (freezing(current
) || kthread_should_stop())
3635 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3638 * Try to sleep for a short interval. Note that kcompactd will only be
3639 * woken if it is possible to sleep for a short interval. This is
3640 * deliberate on the assumption that if reclaim cannot keep an
3641 * eligible zone balanced that it's also unlikely that compaction will
3644 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3646 * Compaction records what page blocks it recently failed to
3647 * isolate pages from and skips them in the future scanning.
3648 * When kswapd is going to sleep, it is reasonable to assume
3649 * that pages and compaction may succeed so reset the cache.
3651 reset_isolation_suitable(pgdat
);
3654 * We have freed the memory, now we should compact it to make
3655 * allocation of the requested order possible.
3657 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3659 remaining
= schedule_timeout(HZ
/10);
3662 * If woken prematurely then reset kswapd_classzone_idx and
3663 * order. The values will either be from a wakeup request or
3664 * the previous request that slept prematurely.
3667 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3668 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3671 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3672 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3676 * After a short sleep, check if it was a premature sleep. If not, then
3677 * go fully to sleep until explicitly woken up.
3680 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3681 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3684 * vmstat counters are not perfectly accurate and the estimated
3685 * value for counters such as NR_FREE_PAGES can deviate from the
3686 * true value by nr_online_cpus * threshold. To avoid the zone
3687 * watermarks being breached while under pressure, we reduce the
3688 * per-cpu vmstat threshold while kswapd is awake and restore
3689 * them before going back to sleep.
3691 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3693 if (!kthread_should_stop())
3696 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3699 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3701 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3703 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3707 * The background pageout daemon, started as a kernel thread
3708 * from the init process.
3710 * This basically trickles out pages so that we have _some_
3711 * free memory available even if there is no other activity
3712 * that frees anything up. This is needed for things like routing
3713 * etc, where we otherwise might have all activity going on in
3714 * asynchronous contexts that cannot page things out.
3716 * If there are applications that are active memory-allocators
3717 * (most normal use), this basically shouldn't matter.
3719 static int kswapd(void *p
)
3721 unsigned int alloc_order
, reclaim_order
;
3722 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3723 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3724 struct task_struct
*tsk
= current
;
3726 struct reclaim_state reclaim_state
= {
3727 .reclaimed_slab
= 0,
3729 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3731 if (!cpumask_empty(cpumask
))
3732 set_cpus_allowed_ptr(tsk
, cpumask
);
3733 current
->reclaim_state
= &reclaim_state
;
3736 * Tell the memory management that we're a "memory allocator",
3737 * and that if we need more memory we should get access to it
3738 * regardless (see "__alloc_pages()"). "kswapd" should
3739 * never get caught in the normal page freeing logic.
3741 * (Kswapd normally doesn't need memory anyway, but sometimes
3742 * you need a small amount of memory in order to be able to
3743 * page out something else, and this flag essentially protects
3744 * us from recursively trying to free more memory as we're
3745 * trying to free the first piece of memory in the first place).
3747 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3750 pgdat
->kswapd_order
= 0;
3751 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3755 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3756 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3759 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3762 /* Read the new order and classzone_idx */
3763 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3764 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3765 pgdat
->kswapd_order
= 0;
3766 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3768 ret
= try_to_freeze();
3769 if (kthread_should_stop())
3773 * We can speed up thawing tasks if we don't call balance_pgdat
3774 * after returning from the refrigerator
3780 * Reclaim begins at the requested order but if a high-order
3781 * reclaim fails then kswapd falls back to reclaiming for
3782 * order-0. If that happens, kswapd will consider sleeping
3783 * for the order it finished reclaiming at (reclaim_order)
3784 * but kcompactd is woken to compact for the original
3785 * request (alloc_order).
3787 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3789 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3790 if (reclaim_order
< alloc_order
)
3791 goto kswapd_try_sleep
;
3794 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3795 current
->reclaim_state
= NULL
;
3801 * A zone is low on free memory or too fragmented for high-order memory. If
3802 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3803 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3804 * has failed or is not needed, still wake up kcompactd if only compaction is
3807 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
3808 enum zone_type classzone_idx
)
3812 if (!managed_zone(zone
))
3815 if (!cpuset_zone_allowed(zone
, gfp_flags
))
3817 pgdat
= zone
->zone_pgdat
;
3818 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3820 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3821 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3824 /* Hopeless node, leave it to direct reclaim if possible */
3825 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
3826 pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3828 * There may be plenty of free memory available, but it's too
3829 * fragmented for high-order allocations. Wake up kcompactd
3830 * and rely on compaction_suitable() to determine if it's
3831 * needed. If it fails, it will defer subsequent attempts to
3832 * ratelimit its work.
3834 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
3835 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3839 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
,
3841 wake_up_interruptible(&pgdat
->kswapd_wait
);
3844 #ifdef CONFIG_HIBERNATION
3846 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3849 * Rather than trying to age LRUs the aim is to preserve the overall
3850 * LRU order by reclaiming preferentially
3851 * inactive > active > active referenced > active mapped
3853 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3855 struct reclaim_state reclaim_state
;
3856 struct scan_control sc
= {
3857 .nr_to_reclaim
= nr_to_reclaim
,
3858 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3859 .reclaim_idx
= MAX_NR_ZONES
- 1,
3860 .priority
= DEF_PRIORITY
,
3864 .hibernation_mode
= 1,
3866 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3867 struct task_struct
*p
= current
;
3868 unsigned long nr_reclaimed
;
3869 unsigned int noreclaim_flag
;
3871 fs_reclaim_acquire(sc
.gfp_mask
);
3872 noreclaim_flag
= memalloc_noreclaim_save();
3873 reclaim_state
.reclaimed_slab
= 0;
3874 p
->reclaim_state
= &reclaim_state
;
3876 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3878 p
->reclaim_state
= NULL
;
3879 memalloc_noreclaim_restore(noreclaim_flag
);
3880 fs_reclaim_release(sc
.gfp_mask
);
3882 return nr_reclaimed
;
3884 #endif /* CONFIG_HIBERNATION */
3886 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3887 not required for correctness. So if the last cpu in a node goes
3888 away, we get changed to run anywhere: as the first one comes back,
3889 restore their cpu bindings. */
3890 static int kswapd_cpu_online(unsigned int cpu
)
3894 for_each_node_state(nid
, N_MEMORY
) {
3895 pg_data_t
*pgdat
= NODE_DATA(nid
);
3896 const struct cpumask
*mask
;
3898 mask
= cpumask_of_node(pgdat
->node_id
);
3900 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3901 /* One of our CPUs online: restore mask */
3902 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3908 * This kswapd start function will be called by init and node-hot-add.
3909 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3911 int kswapd_run(int nid
)
3913 pg_data_t
*pgdat
= NODE_DATA(nid
);
3919 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3920 if (IS_ERR(pgdat
->kswapd
)) {
3921 /* failure at boot is fatal */
3922 BUG_ON(system_state
< SYSTEM_RUNNING
);
3923 pr_err("Failed to start kswapd on node %d\n", nid
);
3924 ret
= PTR_ERR(pgdat
->kswapd
);
3925 pgdat
->kswapd
= NULL
;
3931 * Called by memory hotplug when all memory in a node is offlined. Caller must
3932 * hold mem_hotplug_begin/end().
3934 void kswapd_stop(int nid
)
3936 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3939 kthread_stop(kswapd
);
3940 NODE_DATA(nid
)->kswapd
= NULL
;
3944 static int __init
kswapd_init(void)
3949 for_each_node_state(nid
, N_MEMORY
)
3951 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3952 "mm/vmscan:online", kswapd_cpu_online
,
3958 module_init(kswapd_init
)
3964 * If non-zero call node_reclaim when the number of free pages falls below
3967 int node_reclaim_mode __read_mostly
;
3969 #define RECLAIM_OFF 0
3970 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3971 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3972 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3975 * Priority for NODE_RECLAIM. This determines the fraction of pages
3976 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3979 #define NODE_RECLAIM_PRIORITY 4
3982 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3985 int sysctl_min_unmapped_ratio
= 1;
3988 * If the number of slab pages in a zone grows beyond this percentage then
3989 * slab reclaim needs to occur.
3991 int sysctl_min_slab_ratio
= 5;
3993 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3995 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3996 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3997 node_page_state(pgdat
, NR_ACTIVE_FILE
);
4000 * It's possible for there to be more file mapped pages than
4001 * accounted for by the pages on the file LRU lists because
4002 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4004 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
4007 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4008 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
4010 unsigned long nr_pagecache_reclaimable
;
4011 unsigned long delta
= 0;
4014 * If RECLAIM_UNMAP is set, then all file pages are considered
4015 * potentially reclaimable. Otherwise, we have to worry about
4016 * pages like swapcache and node_unmapped_file_pages() provides
4019 if (node_reclaim_mode
& RECLAIM_UNMAP
)
4020 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
4022 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
4024 /* If we can't clean pages, remove dirty pages from consideration */
4025 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
4026 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
4028 /* Watch for any possible underflows due to delta */
4029 if (unlikely(delta
> nr_pagecache_reclaimable
))
4030 delta
= nr_pagecache_reclaimable
;
4032 return nr_pagecache_reclaimable
- delta
;
4036 * Try to free up some pages from this node through reclaim.
4038 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4040 /* Minimum pages needed in order to stay on node */
4041 const unsigned long nr_pages
= 1 << order
;
4042 struct task_struct
*p
= current
;
4043 struct reclaim_state reclaim_state
;
4044 unsigned int noreclaim_flag
;
4045 struct scan_control sc
= {
4046 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
4047 .gfp_mask
= current_gfp_context(gfp_mask
),
4049 .priority
= NODE_RECLAIM_PRIORITY
,
4050 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
4051 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
4053 .reclaim_idx
= gfp_zone(gfp_mask
),
4057 fs_reclaim_acquire(sc
.gfp_mask
);
4059 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4060 * and we also need to be able to write out pages for RECLAIM_WRITE
4061 * and RECLAIM_UNMAP.
4063 noreclaim_flag
= memalloc_noreclaim_save();
4064 p
->flags
|= PF_SWAPWRITE
;
4065 reclaim_state
.reclaimed_slab
= 0;
4066 p
->reclaim_state
= &reclaim_state
;
4068 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
4070 * Free memory by calling shrink node with increasing
4071 * priorities until we have enough memory freed.
4074 shrink_node(pgdat
, &sc
);
4075 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
4078 p
->reclaim_state
= NULL
;
4079 current
->flags
&= ~PF_SWAPWRITE
;
4080 memalloc_noreclaim_restore(noreclaim_flag
);
4081 fs_reclaim_release(sc
.gfp_mask
);
4082 return sc
.nr_reclaimed
>= nr_pages
;
4085 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
4090 * Node reclaim reclaims unmapped file backed pages and
4091 * slab pages if we are over the defined limits.
4093 * A small portion of unmapped file backed pages is needed for
4094 * file I/O otherwise pages read by file I/O will be immediately
4095 * thrown out if the node is overallocated. So we do not reclaim
4096 * if less than a specified percentage of the node is used by
4097 * unmapped file backed pages.
4099 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
4100 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
4101 return NODE_RECLAIM_FULL
;
4104 * Do not scan if the allocation should not be delayed.
4106 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
4107 return NODE_RECLAIM_NOSCAN
;
4110 * Only run node reclaim on the local node or on nodes that do not
4111 * have associated processors. This will favor the local processor
4112 * over remote processors and spread off node memory allocations
4113 * as wide as possible.
4115 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
4116 return NODE_RECLAIM_NOSCAN
;
4118 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
4119 return NODE_RECLAIM_NOSCAN
;
4121 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
4122 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
4125 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
4132 * page_evictable - test whether a page is evictable
4133 * @page: the page to test
4135 * Test whether page is evictable--i.e., should be placed on active/inactive
4136 * lists vs unevictable list.
4138 * Reasons page might not be evictable:
4139 * (1) page's mapping marked unevictable
4140 * (2) page is part of an mlocked VMA
4143 int page_evictable(struct page
*page
)
4147 /* Prevent address_space of inode and swap cache from being freed */
4149 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4156 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
4157 * @pages: array of pages to check
4158 * @nr_pages: number of pages to check
4160 * Checks pages for evictability and moves them to the appropriate lru list.
4162 * This function is only used for SysV IPC SHM_UNLOCK.
4164 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
4166 struct lruvec
*lruvec
;
4167 struct pglist_data
*pgdat
= NULL
;
4172 for (i
= 0; i
< nr_pages
; i
++) {
4173 struct page
*page
= pages
[i
];
4174 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4177 if (pagepgdat
!= pgdat
) {
4179 spin_unlock_irq(&pgdat
->lru_lock
);
4181 spin_lock_irq(&pgdat
->lru_lock
);
4183 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4185 if (!PageLRU(page
) || !PageUnevictable(page
))
4188 if (page_evictable(page
)) {
4189 enum lru_list lru
= page_lru_base_type(page
);
4191 VM_BUG_ON_PAGE(PageActive(page
), page
);
4192 ClearPageUnevictable(page
);
4193 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4194 add_page_to_lru_list(page
, lruvec
, lru
);
4200 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4201 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4202 spin_unlock_irq(&pgdat
->lru_lock
);
4205 #endif /* CONFIG_SHMEM */