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>
52 #include <linux/psi.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
57 #include <linux/swapops.h>
58 #include <linux/balloon_compaction.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/vmscan.h>
66 /* How many pages shrink_list() should reclaim */
67 unsigned long nr_to_reclaim
;
69 /* This context's GFP mask */
72 /* Allocation order */
76 * Nodemask of nodes allowed by the caller. If NULL, all nodes
82 * The memory cgroup that hit its limit and as a result is the
83 * primary target of this reclaim invocation.
85 struct mem_cgroup
*target_mem_cgroup
;
87 /* Scan (total_size >> priority) pages at once */
90 /* The highest zone to isolate pages for reclaim from */
91 enum zone_type reclaim_idx
;
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage
:1;
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap
:1;
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap
:1;
103 * Cgroups are not reclaimed below their configured memory.low,
104 * unless we threaten to OOM. If any cgroups are skipped due to
105 * memory.low and nothing was reclaimed, go back for memory.low.
107 unsigned int memcg_low_reclaim
:1;
108 unsigned int memcg_low_skipped
:1;
110 unsigned int hibernation_mode
:1;
112 /* One of the zones is ready for compaction */
113 unsigned int compaction_ready
:1;
115 /* Incremented by the number of inactive pages that were scanned */
116 unsigned long nr_scanned
;
118 /* Number of pages freed so far during a call to shrink_zones() */
119 unsigned long nr_reclaimed
;
122 #ifdef ARCH_HAS_PREFETCH
123 #define prefetch_prev_lru_page(_page, _base, _field) \
125 if ((_page)->lru.prev != _base) { \
128 prev = lru_to_page(&(_page->lru)); \
129 prefetch(&prev->_field); \
133 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
136 #ifdef ARCH_HAS_PREFETCHW
137 #define prefetchw_prev_lru_page(_page, _base, _field) \
139 if ((_page)->lru.prev != _base) { \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetchw(&prev->_field); \
147 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
151 * From 0 .. 100. Higher means more swappy.
153 int vm_swappiness
= 60;
155 * The total number of pages which are beyond the high watermark within all
158 unsigned long vm_total_pages
;
160 static LIST_HEAD(shrinker_list
);
161 static DECLARE_RWSEM(shrinker_rwsem
);
164 static bool global_reclaim(struct scan_control
*sc
)
166 return !sc
->target_mem_cgroup
;
170 * sane_reclaim - is the usual dirty throttling mechanism operational?
171 * @sc: scan_control in question
173 * The normal page dirty throttling mechanism in balance_dirty_pages() is
174 * completely broken with the legacy memcg and direct stalling in
175 * shrink_page_list() is used for throttling instead, which lacks all the
176 * niceties such as fairness, adaptive pausing, bandwidth proportional
177 * allocation and configurability.
179 * This function tests whether the vmscan currently in progress can assume
180 * that the normal dirty throttling mechanism is operational.
182 static bool sane_reclaim(struct scan_control
*sc
)
184 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
188 #ifdef CONFIG_CGROUP_WRITEBACK
189 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
195 static bool global_reclaim(struct scan_control
*sc
)
200 static bool sane_reclaim(struct scan_control
*sc
)
207 * This misses isolated pages which are not accounted for to save counters.
208 * As the data only determines if reclaim or compaction continues, it is
209 * not expected that isolated pages will be a dominating factor.
211 unsigned long zone_reclaimable_pages(struct zone
*zone
)
215 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
216 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
217 if (get_nr_swap_pages() > 0)
218 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
219 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
224 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
228 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
229 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
230 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
232 if (get_nr_swap_pages() > 0)
233 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
234 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
235 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
241 * lruvec_lru_size - Returns the number of pages on the given LRU list.
242 * @lruvec: lru vector
244 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
246 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
248 unsigned long lru_size
;
251 if (!mem_cgroup_disabled())
252 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
254 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
256 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
257 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
260 if (!managed_zone(zone
))
263 if (!mem_cgroup_disabled())
264 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
266 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
267 NR_ZONE_LRU_BASE
+ lru
);
268 lru_size
-= min(size
, lru_size
);
276 * Add a shrinker callback to be called from the vm.
278 int register_shrinker(struct shrinker
*shrinker
)
280 size_t size
= sizeof(*shrinker
->nr_deferred
);
282 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
285 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
286 if (!shrinker
->nr_deferred
)
289 down_write(&shrinker_rwsem
);
290 list_add_tail(&shrinker
->list
, &shrinker_list
);
291 up_write(&shrinker_rwsem
);
294 EXPORT_SYMBOL(register_shrinker
);
299 void unregister_shrinker(struct shrinker
*shrinker
)
301 if (!shrinker
->nr_deferred
)
303 down_write(&shrinker_rwsem
);
304 list_del(&shrinker
->list
);
305 up_write(&shrinker_rwsem
);
306 kfree(shrinker
->nr_deferred
);
307 shrinker
->nr_deferred
= NULL
;
309 EXPORT_SYMBOL(unregister_shrinker
);
311 #define SHRINK_BATCH 128
313 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
314 struct shrinker
*shrinker
,
315 unsigned long nr_scanned
,
316 unsigned long nr_eligible
)
318 unsigned long freed
= 0;
319 unsigned long long delta
;
324 int nid
= shrinkctl
->nid
;
325 long batch_size
= shrinker
->batch
? shrinker
->batch
327 long scanned
= 0, next_deferred
;
329 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
334 * copy the current shrinker scan count into a local variable
335 * and zero it so that other concurrent shrinker invocations
336 * don't also do this scanning work.
338 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
341 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
343 do_div(delta
, nr_eligible
+ 1);
345 if (total_scan
< 0) {
346 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
347 shrinker
->scan_objects
, total_scan
);
348 total_scan
= freeable
;
351 next_deferred
= total_scan
;
354 * We need to avoid excessive windup on filesystem shrinkers
355 * due to large numbers of GFP_NOFS allocations causing the
356 * shrinkers to return -1 all the time. This results in a large
357 * nr being built up so when a shrink that can do some work
358 * comes along it empties the entire cache due to nr >>>
359 * freeable. This is bad for sustaining a working set in
362 * Hence only allow the shrinker to scan the entire cache when
363 * a large delta change is calculated directly.
365 if (delta
< freeable
/ 4)
366 total_scan
= min(total_scan
, freeable
/ 2);
369 * Avoid risking looping forever due to too large nr value:
370 * never try to free more than twice the estimate number of
373 if (total_scan
> freeable
* 2)
374 total_scan
= freeable
* 2;
376 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
377 nr_scanned
, nr_eligible
,
378 freeable
, delta
, total_scan
);
381 * Normally, we should not scan less than batch_size objects in one
382 * pass to avoid too frequent shrinker calls, but if the slab has less
383 * than batch_size objects in total and we are really tight on memory,
384 * we will try to reclaim all available objects, otherwise we can end
385 * up failing allocations although there are plenty of reclaimable
386 * objects spread over several slabs with usage less than the
389 * We detect the "tight on memory" situations by looking at the total
390 * number of objects we want to scan (total_scan). If it is greater
391 * than the total number of objects on slab (freeable), we must be
392 * scanning at high prio and therefore should try to reclaim as much as
395 while (total_scan
>= batch_size
||
396 total_scan
>= freeable
) {
398 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
400 shrinkctl
->nr_to_scan
= nr_to_scan
;
401 shrinkctl
->nr_scanned
= nr_to_scan
;
402 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
403 if (ret
== SHRINK_STOP
)
407 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
408 total_scan
-= shrinkctl
->nr_scanned
;
409 scanned
+= shrinkctl
->nr_scanned
;
414 if (next_deferred
>= scanned
)
415 next_deferred
-= scanned
;
419 * move the unused scan count back into the shrinker in a
420 * manner that handles concurrent updates. If we exhausted the
421 * scan, there is no need to do an update.
423 if (next_deferred
> 0)
424 new_nr
= atomic_long_add_return(next_deferred
,
425 &shrinker
->nr_deferred
[nid
]);
427 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
429 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
434 * shrink_slab - shrink slab caches
435 * @gfp_mask: allocation context
436 * @nid: node whose slab caches to target
437 * @memcg: memory cgroup whose slab caches to target
438 * @nr_scanned: pressure numerator
439 * @nr_eligible: pressure denominator
441 * Call the shrink functions to age shrinkable caches.
443 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
444 * unaware shrinkers will receive a node id of 0 instead.
446 * @memcg specifies the memory cgroup to target. If it is not NULL,
447 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
448 * objects from the memory cgroup specified. Otherwise, only unaware
449 * shrinkers are called.
451 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
452 * the available objects should be scanned. Page reclaim for example
453 * passes the number of pages scanned and the number of pages on the
454 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
455 * when it encountered mapped pages. The ratio is further biased by
456 * the ->seeks setting of the shrink function, which indicates the
457 * cost to recreate an object relative to that of an LRU page.
459 * Returns the number of reclaimed slab objects.
461 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
462 struct mem_cgroup
*memcg
,
463 unsigned long nr_scanned
,
464 unsigned long nr_eligible
)
466 struct shrinker
*shrinker
;
467 unsigned long freed
= 0;
469 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
473 nr_scanned
= SWAP_CLUSTER_MAX
;
475 if (!down_read_trylock(&shrinker_rwsem
)) {
477 * If we would return 0, our callers would understand that we
478 * have nothing else to shrink and give up trying. By returning
479 * 1 we keep it going and assume we'll be able to shrink next
486 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
487 struct shrink_control sc
= {
488 .gfp_mask
= gfp_mask
,
494 * If kernel memory accounting is disabled, we ignore
495 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
496 * passing NULL for memcg.
498 if (memcg_kmem_enabled() &&
499 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
502 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
505 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
508 up_read(&shrinker_rwsem
);
514 void drop_slab_node(int nid
)
519 struct mem_cgroup
*memcg
= NULL
;
523 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
525 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
526 } while (freed
> 10);
533 for_each_online_node(nid
)
537 static inline int is_page_cache_freeable(struct page
*page
)
540 * A freeable page cache page is referenced only by the caller
541 * that isolated the page, the page cache radix tree and
542 * optional buffer heads at page->private.
544 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
546 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
549 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
551 if (current
->flags
& PF_SWAPWRITE
)
553 if (!inode_write_congested(inode
))
555 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
561 * We detected a synchronous write error writing a page out. Probably
562 * -ENOSPC. We need to propagate that into the address_space for a subsequent
563 * fsync(), msync() or close().
565 * The tricky part is that after writepage we cannot touch the mapping: nothing
566 * prevents it from being freed up. But we have a ref on the page and once
567 * that page is locked, the mapping is pinned.
569 * We're allowed to run sleeping lock_page() here because we know the caller has
572 static void handle_write_error(struct address_space
*mapping
,
573 struct page
*page
, int error
)
576 if (page_mapping(page
) == mapping
)
577 mapping_set_error(mapping
, error
);
581 /* possible outcome of pageout() */
583 /* failed to write page out, page is locked */
585 /* move page to the active list, page is locked */
587 /* page has been sent to the disk successfully, page is unlocked */
589 /* page is clean and locked */
594 * pageout is called by shrink_page_list() for each dirty page.
595 * Calls ->writepage().
597 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
598 struct scan_control
*sc
)
601 * If the page is dirty, only perform writeback if that write
602 * will be non-blocking. To prevent this allocation from being
603 * stalled by pagecache activity. But note that there may be
604 * stalls if we need to run get_block(). We could test
605 * PagePrivate for that.
607 * If this process is currently in __generic_file_write_iter() against
608 * this page's queue, we can perform writeback even if that
611 * If the page is swapcache, write it back even if that would
612 * block, for some throttling. This happens by accident, because
613 * swap_backing_dev_info is bust: it doesn't reflect the
614 * congestion state of the swapdevs. Easy to fix, if needed.
616 if (!is_page_cache_freeable(page
))
620 * Some data journaling orphaned pages can have
621 * page->mapping == NULL while being dirty with clean buffers.
623 if (page_has_private(page
)) {
624 if (try_to_free_buffers(page
)) {
625 ClearPageDirty(page
);
626 pr_info("%s: orphaned page\n", __func__
);
632 if (mapping
->a_ops
->writepage
== NULL
)
633 return PAGE_ACTIVATE
;
634 if (!may_write_to_inode(mapping
->host
, sc
))
637 if (clear_page_dirty_for_io(page
)) {
639 struct writeback_control wbc
= {
640 .sync_mode
= WB_SYNC_NONE
,
641 .nr_to_write
= SWAP_CLUSTER_MAX
,
643 .range_end
= LLONG_MAX
,
647 SetPageReclaim(page
);
648 res
= mapping
->a_ops
->writepage(page
, &wbc
);
650 handle_write_error(mapping
, page
, res
);
651 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
652 ClearPageReclaim(page
);
653 return PAGE_ACTIVATE
;
656 if (!PageWriteback(page
)) {
657 /* synchronous write or broken a_ops? */
658 ClearPageReclaim(page
);
660 trace_mm_vmscan_writepage(page
);
661 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
669 * Same as remove_mapping, but if the page is removed from the mapping, it
670 * gets returned with a refcount of 0.
672 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
678 BUG_ON(!PageLocked(page
));
679 BUG_ON(mapping
!= page_mapping(page
));
681 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
683 * The non racy check for a busy page.
685 * Must be careful with the order of the tests. When someone has
686 * a ref to the page, it may be possible that they dirty it then
687 * drop the reference. So if PageDirty is tested before page_count
688 * here, then the following race may occur:
690 * get_user_pages(&page);
691 * [user mapping goes away]
693 * !PageDirty(page) [good]
694 * SetPageDirty(page);
696 * !page_count(page) [good, discard it]
698 * [oops, our write_to data is lost]
700 * Reversing the order of the tests ensures such a situation cannot
701 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
702 * load is not satisfied before that of page->_refcount.
704 * Note that if SetPageDirty is always performed via set_page_dirty,
705 * and thus under tree_lock, then this ordering is not required.
707 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
708 refcount
= 1 + HPAGE_PMD_NR
;
711 if (!page_ref_freeze(page
, refcount
))
713 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
714 if (unlikely(PageDirty(page
))) {
715 page_ref_unfreeze(page
, refcount
);
719 if (PageSwapCache(page
)) {
720 swp_entry_t swap
= { .val
= page_private(page
) };
721 mem_cgroup_swapout(page
, swap
);
722 __delete_from_swap_cache(page
);
723 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
724 put_swap_page(page
, swap
);
726 void (*freepage
)(struct page
*);
729 freepage
= mapping
->a_ops
->freepage
;
731 * Remember a shadow entry for reclaimed file cache in
732 * order to detect refaults, thus thrashing, later on.
734 * But don't store shadows in an address space that is
735 * already exiting. This is not just an optizimation,
736 * inode reclaim needs to empty out the radix tree or
737 * the nodes are lost. Don't plant shadows behind its
740 * We also don't store shadows for DAX mappings because the
741 * only page cache pages found in these are zero pages
742 * covering holes, and because we don't want to mix DAX
743 * exceptional entries and shadow exceptional entries in the
746 if (reclaimed
&& page_is_file_cache(page
) &&
747 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
748 shadow
= workingset_eviction(mapping
, page
);
749 __delete_from_page_cache(page
, shadow
);
750 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
752 if (freepage
!= NULL
)
759 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
764 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
765 * someone else has a ref on the page, abort and return 0. If it was
766 * successfully detached, return 1. Assumes the caller has a single ref on
769 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
771 if (__remove_mapping(mapping
, page
, false)) {
773 * Unfreezing the refcount with 1 rather than 2 effectively
774 * drops the pagecache ref for us without requiring another
777 page_ref_unfreeze(page
, 1);
784 * putback_lru_page - put previously isolated page onto appropriate LRU list
785 * @page: page to be put back to appropriate lru list
787 * Add previously isolated @page to appropriate LRU list.
788 * Page may still be unevictable for other reasons.
790 * lru_lock must not be held, interrupts must be enabled.
792 void putback_lru_page(struct page
*page
)
795 int was_unevictable
= PageUnevictable(page
);
797 VM_BUG_ON_PAGE(PageLRU(page
), page
);
800 ClearPageUnevictable(page
);
802 if (page_evictable(page
)) {
804 * For evictable pages, we can use the cache.
805 * In event of a race, worst case is we end up with an
806 * unevictable page on [in]active list.
807 * We know how to handle that.
809 is_unevictable
= false;
813 * Put unevictable pages directly on zone's unevictable
816 is_unevictable
= true;
817 add_page_to_unevictable_list(page
);
819 * When racing with an mlock or AS_UNEVICTABLE clearing
820 * (page is unlocked) make sure that if the other thread
821 * does not observe our setting of PG_lru and fails
822 * isolation/check_move_unevictable_pages,
823 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
824 * the page back to the evictable list.
826 * The other side is TestClearPageMlocked() or shmem_lock().
832 * page's status can change while we move it among lru. If an evictable
833 * page is on unevictable list, it never be freed. To avoid that,
834 * check after we added it to the list, again.
836 if (is_unevictable
&& page_evictable(page
)) {
837 if (!isolate_lru_page(page
)) {
841 /* This means someone else dropped this page from LRU
842 * So, it will be freed or putback to LRU again. There is
843 * nothing to do here.
847 if (was_unevictable
&& !is_unevictable
)
848 count_vm_event(UNEVICTABLE_PGRESCUED
);
849 else if (!was_unevictable
&& is_unevictable
)
850 count_vm_event(UNEVICTABLE_PGCULLED
);
852 put_page(page
); /* drop ref from isolate */
855 enum page_references
{
857 PAGEREF_RECLAIM_CLEAN
,
862 static enum page_references
page_check_references(struct page
*page
,
863 struct scan_control
*sc
)
865 int referenced_ptes
, referenced_page
;
866 unsigned long vm_flags
;
868 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
870 referenced_page
= TestClearPageReferenced(page
);
873 * Mlock lost the isolation race with us. Let try_to_unmap()
874 * move the page to the unevictable list.
876 if (vm_flags
& VM_LOCKED
)
877 return PAGEREF_RECLAIM
;
879 if (referenced_ptes
) {
880 if (PageSwapBacked(page
))
881 return PAGEREF_ACTIVATE
;
883 * All mapped pages start out with page table
884 * references from the instantiating fault, so we need
885 * to look twice if a mapped file page is used more
888 * Mark it and spare it for another trip around the
889 * inactive list. Another page table reference will
890 * lead to its activation.
892 * Note: the mark is set for activated pages as well
893 * so that recently deactivated but used pages are
896 SetPageReferenced(page
);
898 if (referenced_page
|| referenced_ptes
> 1)
899 return PAGEREF_ACTIVATE
;
902 * Activate file-backed executable pages after first usage.
904 if (vm_flags
& VM_EXEC
)
905 return PAGEREF_ACTIVATE
;
910 /* Reclaim if clean, defer dirty pages to writeback */
911 if (referenced_page
&& !PageSwapBacked(page
))
912 return PAGEREF_RECLAIM_CLEAN
;
914 return PAGEREF_RECLAIM
;
917 /* Check if a page is dirty or under writeback */
918 static void page_check_dirty_writeback(struct page
*page
,
919 bool *dirty
, bool *writeback
)
921 struct address_space
*mapping
;
924 * Anonymous pages are not handled by flushers and must be written
925 * from reclaim context. Do not stall reclaim based on them
927 if (!page_is_file_cache(page
) ||
928 (PageAnon(page
) && !PageSwapBacked(page
))) {
934 /* By default assume that the page flags are accurate */
935 *dirty
= PageDirty(page
);
936 *writeback
= PageWriteback(page
);
938 /* Verify dirty/writeback state if the filesystem supports it */
939 if (!page_has_private(page
))
942 mapping
= page_mapping(page
);
943 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
944 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
947 struct reclaim_stat
{
949 unsigned nr_unqueued_dirty
;
950 unsigned nr_congested
;
951 unsigned nr_writeback
;
952 unsigned nr_immediate
;
953 unsigned nr_activate
;
954 unsigned nr_ref_keep
;
955 unsigned nr_unmap_fail
;
959 * shrink_page_list() returns the number of reclaimed pages
961 static unsigned long shrink_page_list(struct list_head
*page_list
,
962 struct pglist_data
*pgdat
,
963 struct scan_control
*sc
,
964 enum ttu_flags ttu_flags
,
965 struct reclaim_stat
*stat
,
968 LIST_HEAD(ret_pages
);
969 LIST_HEAD(free_pages
);
971 unsigned nr_unqueued_dirty
= 0;
972 unsigned nr_dirty
= 0;
973 unsigned nr_congested
= 0;
974 unsigned nr_reclaimed
= 0;
975 unsigned nr_writeback
= 0;
976 unsigned nr_immediate
= 0;
977 unsigned nr_ref_keep
= 0;
978 unsigned nr_unmap_fail
= 0;
982 while (!list_empty(page_list
)) {
983 struct address_space
*mapping
;
986 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
987 bool dirty
, writeback
;
991 page
= lru_to_page(page_list
);
992 list_del(&page
->lru
);
994 if (!trylock_page(page
))
997 VM_BUG_ON_PAGE(PageActive(page
), page
);
1001 if (unlikely(!page_evictable(page
)))
1002 goto activate_locked
;
1004 if (!sc
->may_unmap
&& page_mapped(page
))
1007 /* Double the slab pressure for mapped and swapcache pages */
1008 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1009 !(PageAnon(page
) && !PageSwapBacked(page
)))
1012 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1013 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1016 * The number of dirty pages determines if a zone is marked
1017 * reclaim_congested which affects wait_iff_congested. kswapd
1018 * will stall and start writing pages if the tail of the LRU
1019 * is all dirty unqueued pages.
1021 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1022 if (dirty
|| writeback
)
1025 if (dirty
&& !writeback
)
1026 nr_unqueued_dirty
++;
1029 * Treat this page as congested if the underlying BDI is or if
1030 * pages are cycling through the LRU so quickly that the
1031 * pages marked for immediate reclaim are making it to the
1032 * end of the LRU a second time.
1034 mapping
= page_mapping(page
);
1035 if (((dirty
|| writeback
) && mapping
&&
1036 inode_write_congested(mapping
->host
)) ||
1037 (writeback
&& PageReclaim(page
)))
1041 * If a page at the tail of the LRU is under writeback, there
1042 * are three cases to consider.
1044 * 1) If reclaim is encountering an excessive number of pages
1045 * under writeback and this page is both under writeback and
1046 * PageReclaim then it indicates that pages are being queued
1047 * for IO but are being recycled through the LRU before the
1048 * IO can complete. Waiting on the page itself risks an
1049 * indefinite stall if it is impossible to writeback the
1050 * page due to IO error or disconnected storage so instead
1051 * note that the LRU is being scanned too quickly and the
1052 * caller can stall after page list has been processed.
1054 * 2) Global or new memcg reclaim encounters a page that is
1055 * not marked for immediate reclaim, or the caller does not
1056 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1057 * not to fs). In this case mark the page for immediate
1058 * reclaim and continue scanning.
1060 * Require may_enter_fs because we would wait on fs, which
1061 * may not have submitted IO yet. And the loop driver might
1062 * enter reclaim, and deadlock if it waits on a page for
1063 * which it is needed to do the write (loop masks off
1064 * __GFP_IO|__GFP_FS for this reason); but more thought
1065 * would probably show more reasons.
1067 * 3) Legacy memcg encounters a page that is already marked
1068 * PageReclaim. memcg does not have any dirty pages
1069 * throttling so we could easily OOM just because too many
1070 * pages are in writeback and there is nothing else to
1071 * reclaim. Wait for the writeback to complete.
1073 * In cases 1) and 2) we activate the pages to get them out of
1074 * the way while we continue scanning for clean pages on the
1075 * inactive list and refilling from the active list. The
1076 * observation here is that waiting for disk writes is more
1077 * expensive than potentially causing reloads down the line.
1078 * Since they're marked for immediate reclaim, they won't put
1079 * memory pressure on the cache working set any longer than it
1080 * takes to write them to disk.
1082 if (PageWriteback(page
)) {
1084 if (current_is_kswapd() &&
1085 PageReclaim(page
) &&
1086 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1088 goto activate_locked
;
1091 } else if (sane_reclaim(sc
) ||
1092 !PageReclaim(page
) || !may_enter_fs
) {
1094 * This is slightly racy - end_page_writeback()
1095 * might have just cleared PageReclaim, then
1096 * setting PageReclaim here end up interpreted
1097 * as PageReadahead - but that does not matter
1098 * enough to care. What we do want is for this
1099 * page to have PageReclaim set next time memcg
1100 * reclaim reaches the tests above, so it will
1101 * then wait_on_page_writeback() to avoid OOM;
1102 * and it's also appropriate in global reclaim.
1104 SetPageReclaim(page
);
1106 goto activate_locked
;
1111 wait_on_page_writeback(page
);
1112 /* then go back and try same page again */
1113 list_add_tail(&page
->lru
, page_list
);
1119 references
= page_check_references(page
, sc
);
1121 switch (references
) {
1122 case PAGEREF_ACTIVATE
:
1123 goto activate_locked
;
1127 case PAGEREF_RECLAIM
:
1128 case PAGEREF_RECLAIM_CLEAN
:
1129 ; /* try to reclaim the page below */
1133 * Anonymous process memory has backing store?
1134 * Try to allocate it some swap space here.
1135 * Lazyfree page could be freed directly
1137 if (PageAnon(page
) && PageSwapBacked(page
)) {
1138 if (!PageSwapCache(page
)) {
1139 if (!(sc
->gfp_mask
& __GFP_IO
))
1141 if (PageTransHuge(page
)) {
1142 /* cannot split THP, skip it */
1143 if (!can_split_huge_page(page
, NULL
))
1144 goto activate_locked
;
1146 * Split pages without a PMD map right
1147 * away. Chances are some or all of the
1148 * tail pages can be freed without IO.
1150 if (!compound_mapcount(page
) &&
1151 split_huge_page_to_list(page
,
1153 goto activate_locked
;
1155 if (!add_to_swap(page
)) {
1156 if (!PageTransHuge(page
))
1157 goto activate_locked
;
1158 /* Fallback to swap normal pages */
1159 if (split_huge_page_to_list(page
,
1161 goto activate_locked
;
1162 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1163 count_vm_event(THP_SWPOUT_FALLBACK
);
1165 if (!add_to_swap(page
))
1166 goto activate_locked
;
1171 /* Adding to swap updated mapping */
1172 mapping
= page_mapping(page
);
1174 } else if (unlikely(PageTransHuge(page
))) {
1175 /* Split file THP */
1176 if (split_huge_page_to_list(page
, page_list
))
1181 * The page is mapped into the page tables of one or more
1182 * processes. Try to unmap it here.
1184 if (page_mapped(page
)) {
1185 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1187 if (unlikely(PageTransHuge(page
)))
1188 flags
|= TTU_SPLIT_HUGE_PMD
;
1189 if (!try_to_unmap(page
, flags
)) {
1191 goto activate_locked
;
1195 if (PageDirty(page
)) {
1197 * Only kswapd can writeback filesystem pages
1198 * to avoid risk of stack overflow. But avoid
1199 * injecting inefficient single-page IO into
1200 * flusher writeback as much as possible: only
1201 * write pages when we've encountered many
1202 * dirty pages, and when we've already scanned
1203 * the rest of the LRU for clean pages and see
1204 * the same dirty pages again (PageReclaim).
1206 if (page_is_file_cache(page
) &&
1207 (!current_is_kswapd() || !PageReclaim(page
) ||
1208 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1210 * Immediately reclaim when written back.
1211 * Similar in principal to deactivate_page()
1212 * except we already have the page isolated
1213 * and know it's dirty
1215 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1216 SetPageReclaim(page
);
1218 goto activate_locked
;
1221 if (references
== PAGEREF_RECLAIM_CLEAN
)
1225 if (!sc
->may_writepage
)
1229 * Page is dirty. Flush the TLB if a writable entry
1230 * potentially exists to avoid CPU writes after IO
1231 * starts and then write it out here.
1233 try_to_unmap_flush_dirty();
1234 switch (pageout(page
, mapping
, sc
)) {
1238 goto activate_locked
;
1240 if (PageWriteback(page
))
1242 if (PageDirty(page
))
1246 * A synchronous write - probably a ramdisk. Go
1247 * ahead and try to reclaim the page.
1249 if (!trylock_page(page
))
1251 if (PageDirty(page
) || PageWriteback(page
))
1253 mapping
= page_mapping(page
);
1255 ; /* try to free the page below */
1260 * If the page has buffers, try to free the buffer mappings
1261 * associated with this page. If we succeed we try to free
1264 * We do this even if the page is PageDirty().
1265 * try_to_release_page() does not perform I/O, but it is
1266 * possible for a page to have PageDirty set, but it is actually
1267 * clean (all its buffers are clean). This happens if the
1268 * buffers were written out directly, with submit_bh(). ext3
1269 * will do this, as well as the blockdev mapping.
1270 * try_to_release_page() will discover that cleanness and will
1271 * drop the buffers and mark the page clean - it can be freed.
1273 * Rarely, pages can have buffers and no ->mapping. These are
1274 * the pages which were not successfully invalidated in
1275 * truncate_complete_page(). We try to drop those buffers here
1276 * and if that worked, and the page is no longer mapped into
1277 * process address space (page_count == 1) it can be freed.
1278 * Otherwise, leave the page on the LRU so it is swappable.
1280 if (page_has_private(page
)) {
1281 if (!try_to_release_page(page
, sc
->gfp_mask
))
1282 goto activate_locked
;
1283 if (!mapping
&& page_count(page
) == 1) {
1285 if (put_page_testzero(page
))
1289 * rare race with speculative reference.
1290 * the speculative reference will free
1291 * this page shortly, so we may
1292 * increment nr_reclaimed here (and
1293 * leave it off the LRU).
1301 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1302 /* follow __remove_mapping for reference */
1303 if (!page_ref_freeze(page
, 1))
1305 if (PageDirty(page
)) {
1306 page_ref_unfreeze(page
, 1);
1310 count_vm_event(PGLAZYFREED
);
1311 count_memcg_page_event(page
, PGLAZYFREED
);
1312 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1315 * At this point, we have no other references and there is
1316 * no way to pick any more up (removed from LRU, removed
1317 * from pagecache). Can use non-atomic bitops now (and
1318 * we obviously don't have to worry about waking up a process
1319 * waiting on the page lock, because there are no references.
1321 __ClearPageLocked(page
);
1326 * Is there need to periodically free_page_list? It would
1327 * appear not as the counts should be low
1329 if (unlikely(PageTransHuge(page
))) {
1330 mem_cgroup_uncharge(page
);
1331 (*get_compound_page_dtor(page
))(page
);
1333 list_add(&page
->lru
, &free_pages
);
1337 /* Not a candidate for swapping, so reclaim swap space. */
1338 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1340 try_to_free_swap(page
);
1341 VM_BUG_ON_PAGE(PageActive(page
), page
);
1342 if (!PageMlocked(page
)) {
1343 SetPageActive(page
);
1345 count_memcg_page_event(page
, PGACTIVATE
);
1350 list_add(&page
->lru
, &ret_pages
);
1351 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1354 mem_cgroup_uncharge_list(&free_pages
);
1355 try_to_unmap_flush();
1356 free_hot_cold_page_list(&free_pages
, true);
1358 list_splice(&ret_pages
, page_list
);
1359 count_vm_events(PGACTIVATE
, pgactivate
);
1362 stat
->nr_dirty
= nr_dirty
;
1363 stat
->nr_congested
= nr_congested
;
1364 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1365 stat
->nr_writeback
= nr_writeback
;
1366 stat
->nr_immediate
= nr_immediate
;
1367 stat
->nr_activate
= pgactivate
;
1368 stat
->nr_ref_keep
= nr_ref_keep
;
1369 stat
->nr_unmap_fail
= nr_unmap_fail
;
1371 return nr_reclaimed
;
1374 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1375 struct list_head
*page_list
)
1377 struct scan_control sc
= {
1378 .gfp_mask
= GFP_KERNEL
,
1379 .priority
= DEF_PRIORITY
,
1383 struct page
*page
, *next
;
1384 LIST_HEAD(clean_pages
);
1386 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1387 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1388 !__PageMovable(page
)) {
1389 ClearPageActive(page
);
1390 list_move(&page
->lru
, &clean_pages
);
1394 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1395 TTU_IGNORE_ACCESS
, NULL
, true);
1396 list_splice(&clean_pages
, page_list
);
1397 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1402 * Attempt to remove the specified page from its LRU. Only take this page
1403 * if it is of the appropriate PageActive status. Pages which are being
1404 * freed elsewhere are also ignored.
1406 * page: page to consider
1407 * mode: one of the LRU isolation modes defined above
1409 * returns 0 on success, -ve errno on failure.
1411 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1415 /* Only take pages on the LRU. */
1419 /* Compaction should not handle unevictable pages but CMA can do so */
1420 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1426 * To minimise LRU disruption, the caller can indicate that it only
1427 * wants to isolate pages it will be able to operate on without
1428 * blocking - clean pages for the most part.
1430 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1431 * that it is possible to migrate without blocking
1433 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1434 /* All the caller can do on PageWriteback is block */
1435 if (PageWriteback(page
))
1438 if (PageDirty(page
)) {
1439 struct address_space
*mapping
;
1443 * Only pages without mappings or that have a
1444 * ->migratepage callback are possible to migrate
1445 * without blocking. However, we can be racing with
1446 * truncation so it's necessary to lock the page
1447 * to stabilise the mapping as truncation holds
1448 * the page lock until after the page is removed
1449 * from the page cache.
1451 if (!trylock_page(page
))
1454 mapping
= page_mapping(page
);
1455 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1462 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1465 if (likely(get_page_unless_zero(page
))) {
1467 * Be careful not to clear PageLRU until after we're
1468 * sure the page is not being freed elsewhere -- the
1469 * page release code relies on it.
1480 * Update LRU sizes after isolating pages. The LRU size updates must
1481 * be complete before mem_cgroup_update_lru_size due to a santity check.
1483 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1484 enum lru_list lru
, unsigned long *nr_zone_taken
)
1488 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1489 if (!nr_zone_taken
[zid
])
1492 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1494 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1501 * zone_lru_lock is heavily contended. Some of the functions that
1502 * shrink the lists perform better by taking out a batch of pages
1503 * and working on them outside the LRU lock.
1505 * For pagecache intensive workloads, this function is the hottest
1506 * spot in the kernel (apart from copy_*_user functions).
1508 * Appropriate locks must be held before calling this function.
1510 * @nr_to_scan: The number of eligible pages to look through on the list.
1511 * @lruvec: The LRU vector to pull pages from.
1512 * @dst: The temp list to put pages on to.
1513 * @nr_scanned: The number of pages that were scanned.
1514 * @sc: The scan_control struct for this reclaim session
1515 * @mode: One of the LRU isolation modes
1516 * @lru: LRU list id for isolating
1518 * returns how many pages were moved onto *@dst.
1520 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1521 struct lruvec
*lruvec
, struct list_head
*dst
,
1522 unsigned long *nr_scanned
, struct scan_control
*sc
,
1523 isolate_mode_t mode
, enum lru_list lru
)
1525 struct list_head
*src
= &lruvec
->lists
[lru
];
1526 unsigned long nr_taken
= 0;
1527 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1528 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1529 unsigned long skipped
= 0;
1530 unsigned long scan
, total_scan
, nr_pages
;
1531 LIST_HEAD(pages_skipped
);
1534 for (total_scan
= 0;
1535 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1539 page
= lru_to_page(src
);
1540 prefetchw_prev_lru_page(page
, src
, flags
);
1542 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1544 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1545 list_move(&page
->lru
, &pages_skipped
);
1546 nr_skipped
[page_zonenum(page
)]++;
1551 * Do not count skipped pages because that makes the function
1552 * return with no isolated pages if the LRU mostly contains
1553 * ineligible pages. This causes the VM to not reclaim any
1554 * pages, triggering a premature OOM.
1557 switch (__isolate_lru_page(page
, mode
)) {
1559 nr_pages
= hpage_nr_pages(page
);
1560 nr_taken
+= nr_pages
;
1561 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1562 list_move(&page
->lru
, dst
);
1566 /* else it is being freed elsewhere */
1567 list_move(&page
->lru
, src
);
1576 * Splice any skipped pages to the start of the LRU list. Note that
1577 * this disrupts the LRU order when reclaiming for lower zones but
1578 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1579 * scanning would soon rescan the same pages to skip and put the
1580 * system at risk of premature OOM.
1582 if (!list_empty(&pages_skipped
)) {
1585 list_splice(&pages_skipped
, src
);
1586 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1587 if (!nr_skipped
[zid
])
1590 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1591 skipped
+= nr_skipped
[zid
];
1594 *nr_scanned
= total_scan
;
1595 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1596 total_scan
, skipped
, nr_taken
, mode
, lru
);
1597 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1602 * isolate_lru_page - tries to isolate a page from its LRU list
1603 * @page: page to isolate from its LRU list
1605 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1606 * vmstat statistic corresponding to whatever LRU list the page was on.
1608 * Returns 0 if the page was removed from an LRU list.
1609 * Returns -EBUSY if the page was not on an LRU list.
1611 * The returned page will have PageLRU() cleared. If it was found on
1612 * the active list, it will have PageActive set. If it was found on
1613 * the unevictable list, it will have the PageUnevictable bit set. That flag
1614 * may need to be cleared by the caller before letting the page go.
1616 * The vmstat statistic corresponding to the list on which the page was
1617 * found will be decremented.
1620 * (1) Must be called with an elevated refcount on the page. This is a
1621 * fundamentnal difference from isolate_lru_pages (which is called
1622 * without a stable reference).
1623 * (2) the lru_lock must not be held.
1624 * (3) interrupts must be enabled.
1626 int isolate_lru_page(struct page
*page
)
1630 VM_BUG_ON_PAGE(!page_count(page
), page
);
1631 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1633 if (PageLRU(page
)) {
1634 struct zone
*zone
= page_zone(page
);
1635 struct lruvec
*lruvec
;
1637 spin_lock_irq(zone_lru_lock(zone
));
1638 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1639 if (PageLRU(page
)) {
1640 int lru
= page_lru(page
);
1643 del_page_from_lru_list(page
, lruvec
, lru
);
1646 spin_unlock_irq(zone_lru_lock(zone
));
1652 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1653 * then get resheduled. When there are massive number of tasks doing page
1654 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1655 * the LRU list will go small and be scanned faster than necessary, leading to
1656 * unnecessary swapping, thrashing and OOM.
1658 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1659 struct scan_control
*sc
)
1661 unsigned long inactive
, isolated
;
1663 if (current_is_kswapd())
1666 if (!sane_reclaim(sc
))
1670 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1671 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1673 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1674 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1678 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1679 * won't get blocked by normal direct-reclaimers, forming a circular
1682 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1685 return isolated
> inactive
;
1688 static noinline_for_stack
void
1689 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1691 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1692 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1693 LIST_HEAD(pages_to_free
);
1696 * Put back any unfreeable pages.
1698 while (!list_empty(page_list
)) {
1699 struct page
*page
= lru_to_page(page_list
);
1702 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1703 list_del(&page
->lru
);
1704 if (unlikely(!page_evictable(page
))) {
1705 spin_unlock_irq(&pgdat
->lru_lock
);
1706 putback_lru_page(page
);
1707 spin_lock_irq(&pgdat
->lru_lock
);
1711 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1714 lru
= page_lru(page
);
1715 add_page_to_lru_list(page
, lruvec
, lru
);
1717 if (is_active_lru(lru
)) {
1718 int file
= is_file_lru(lru
);
1719 int numpages
= hpage_nr_pages(page
);
1720 reclaim_stat
->recent_rotated
[file
] += numpages
;
1722 if (put_page_testzero(page
)) {
1723 __ClearPageLRU(page
);
1724 __ClearPageActive(page
);
1725 del_page_from_lru_list(page
, lruvec
, lru
);
1727 if (unlikely(PageCompound(page
))) {
1728 spin_unlock_irq(&pgdat
->lru_lock
);
1729 mem_cgroup_uncharge(page
);
1730 (*get_compound_page_dtor(page
))(page
);
1731 spin_lock_irq(&pgdat
->lru_lock
);
1733 list_add(&page
->lru
, &pages_to_free
);
1738 * To save our caller's stack, now use input list for pages to free.
1740 list_splice(&pages_to_free
, page_list
);
1744 * If a kernel thread (such as nfsd for loop-back mounts) services
1745 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1746 * In that case we should only throttle if the backing device it is
1747 * writing to is congested. In other cases it is safe to throttle.
1749 static int current_may_throttle(void)
1751 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1752 current
->backing_dev_info
== NULL
||
1753 bdi_write_congested(current
->backing_dev_info
);
1757 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1758 * of reclaimed pages
1760 static noinline_for_stack
unsigned long
1761 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1762 struct scan_control
*sc
, enum lru_list lru
)
1764 LIST_HEAD(page_list
);
1765 unsigned long nr_scanned
;
1766 unsigned long nr_reclaimed
= 0;
1767 unsigned long nr_taken
;
1768 struct reclaim_stat stat
= {};
1769 isolate_mode_t isolate_mode
= 0;
1770 int file
= is_file_lru(lru
);
1771 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1772 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1773 bool stalled
= false;
1775 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1779 /* wait a bit for the reclaimer. */
1783 /* We are about to die and free our memory. Return now. */
1784 if (fatal_signal_pending(current
))
1785 return SWAP_CLUSTER_MAX
;
1791 isolate_mode
|= ISOLATE_UNMAPPED
;
1793 spin_lock_irq(&pgdat
->lru_lock
);
1795 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1796 &nr_scanned
, sc
, isolate_mode
, lru
);
1798 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1799 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1801 if (current_is_kswapd()) {
1802 if (global_reclaim(sc
))
1803 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1804 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1807 if (global_reclaim(sc
))
1808 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1809 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1812 spin_unlock_irq(&pgdat
->lru_lock
);
1817 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1820 spin_lock_irq(&pgdat
->lru_lock
);
1822 if (current_is_kswapd()) {
1823 if (global_reclaim(sc
))
1824 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1825 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1828 if (global_reclaim(sc
))
1829 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1830 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1834 putback_inactive_pages(lruvec
, &page_list
);
1836 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1838 spin_unlock_irq(&pgdat
->lru_lock
);
1840 mem_cgroup_uncharge_list(&page_list
);
1841 free_hot_cold_page_list(&page_list
, true);
1844 * If reclaim is isolating dirty pages under writeback, it implies
1845 * that the long-lived page allocation rate is exceeding the page
1846 * laundering rate. Either the global limits are not being effective
1847 * at throttling processes due to the page distribution throughout
1848 * zones or there is heavy usage of a slow backing device. The
1849 * only option is to throttle from reclaim context which is not ideal
1850 * as there is no guarantee the dirtying process is throttled in the
1851 * same way balance_dirty_pages() manages.
1853 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1854 * of pages under pages flagged for immediate reclaim and stall if any
1855 * are encountered in the nr_immediate check below.
1857 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1858 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1861 * If dirty pages are scanned that are not queued for IO, it
1862 * implies that flushers are not doing their job. This can
1863 * happen when memory pressure pushes dirty pages to the end of
1864 * the LRU before the dirty limits are breached and the dirty
1865 * data has expired. It can also happen when the proportion of
1866 * dirty pages grows not through writes but through memory
1867 * pressure reclaiming all the clean cache. And in some cases,
1868 * the flushers simply cannot keep up with the allocation
1869 * rate. Nudge the flusher threads in case they are asleep.
1871 if (stat
.nr_unqueued_dirty
== nr_taken
)
1872 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1875 * Legacy memcg will stall in page writeback so avoid forcibly
1878 if (sane_reclaim(sc
)) {
1880 * Tag a zone as congested if all the dirty pages scanned were
1881 * backed by a congested BDI and wait_iff_congested will stall.
1883 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1884 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1886 /* Allow kswapd to start writing pages during reclaim. */
1887 if (stat
.nr_unqueued_dirty
== nr_taken
)
1888 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1891 * If kswapd scans pages marked marked for immediate
1892 * reclaim and under writeback (nr_immediate), it implies
1893 * that pages are cycling through the LRU faster than
1894 * they are written so also forcibly stall.
1896 if (stat
.nr_immediate
&& current_may_throttle())
1897 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1901 * Stall direct reclaim for IO completions if underlying BDIs or zone
1902 * is congested. Allow kswapd to continue until it starts encountering
1903 * unqueued dirty pages or cycling through the LRU too quickly.
1905 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1906 current_may_throttle())
1907 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1909 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1910 nr_scanned
, nr_reclaimed
,
1911 stat
.nr_dirty
, stat
.nr_writeback
,
1912 stat
.nr_congested
, stat
.nr_immediate
,
1913 stat
.nr_activate
, stat
.nr_ref_keep
,
1915 sc
->priority
, file
);
1916 return nr_reclaimed
;
1920 * This moves pages from the active list to the inactive list.
1922 * We move them the other way if the page is referenced by one or more
1923 * processes, from rmap.
1925 * If the pages are mostly unmapped, the processing is fast and it is
1926 * appropriate to hold zone_lru_lock across the whole operation. But if
1927 * the pages are mapped, the processing is slow (page_referenced()) so we
1928 * should drop zone_lru_lock around each page. It's impossible to balance
1929 * this, so instead we remove the pages from the LRU while processing them.
1930 * It is safe to rely on PG_active against the non-LRU pages in here because
1931 * nobody will play with that bit on a non-LRU page.
1933 * The downside is that we have to touch page->_refcount against each page.
1934 * But we had to alter page->flags anyway.
1936 * Returns the number of pages moved to the given lru.
1939 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1940 struct list_head
*list
,
1941 struct list_head
*pages_to_free
,
1944 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1949 while (!list_empty(list
)) {
1950 page
= lru_to_page(list
);
1951 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1953 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1956 nr_pages
= hpage_nr_pages(page
);
1957 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1958 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1960 if (put_page_testzero(page
)) {
1961 __ClearPageLRU(page
);
1962 __ClearPageActive(page
);
1963 del_page_from_lru_list(page
, lruvec
, lru
);
1965 if (unlikely(PageCompound(page
))) {
1966 spin_unlock_irq(&pgdat
->lru_lock
);
1967 mem_cgroup_uncharge(page
);
1968 (*get_compound_page_dtor(page
))(page
);
1969 spin_lock_irq(&pgdat
->lru_lock
);
1971 list_add(&page
->lru
, pages_to_free
);
1973 nr_moved
+= nr_pages
;
1977 if (!is_active_lru(lru
)) {
1978 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1979 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
1986 static void shrink_active_list(unsigned long nr_to_scan
,
1987 struct lruvec
*lruvec
,
1988 struct scan_control
*sc
,
1991 unsigned long nr_taken
;
1992 unsigned long nr_scanned
;
1993 unsigned long vm_flags
;
1994 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1995 LIST_HEAD(l_active
);
1996 LIST_HEAD(l_inactive
);
1998 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1999 unsigned nr_deactivate
, nr_activate
;
2000 unsigned nr_rotated
= 0;
2001 isolate_mode_t isolate_mode
= 0;
2002 int file
= is_file_lru(lru
);
2003 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2008 isolate_mode
|= ISOLATE_UNMAPPED
;
2010 spin_lock_irq(&pgdat
->lru_lock
);
2012 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2013 &nr_scanned
, sc
, isolate_mode
, lru
);
2015 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2016 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2018 __count_vm_events(PGREFILL
, nr_scanned
);
2019 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2021 spin_unlock_irq(&pgdat
->lru_lock
);
2023 while (!list_empty(&l_hold
)) {
2025 page
= lru_to_page(&l_hold
);
2026 list_del(&page
->lru
);
2028 if (unlikely(!page_evictable(page
))) {
2029 putback_lru_page(page
);
2033 if (unlikely(buffer_heads_over_limit
)) {
2034 if (page_has_private(page
) && trylock_page(page
)) {
2035 if (page_has_private(page
))
2036 try_to_release_page(page
, 0);
2041 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2043 nr_rotated
+= hpage_nr_pages(page
);
2045 * Identify referenced, file-backed active pages and
2046 * give them one more trip around the active list. So
2047 * that executable code get better chances to stay in
2048 * memory under moderate memory pressure. Anon pages
2049 * are not likely to be evicted by use-once streaming
2050 * IO, plus JVM can create lots of anon VM_EXEC pages,
2051 * so we ignore them here.
2053 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2054 list_add(&page
->lru
, &l_active
);
2059 ClearPageActive(page
); /* we are de-activating */
2060 SetPageWorkingset(page
);
2061 list_add(&page
->lru
, &l_inactive
);
2065 * Move pages back to the lru list.
2067 spin_lock_irq(&pgdat
->lru_lock
);
2069 * Count referenced pages from currently used mappings as rotated,
2070 * even though only some of them are actually re-activated. This
2071 * helps balance scan pressure between file and anonymous pages in
2074 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2076 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2077 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2078 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2079 spin_unlock_irq(&pgdat
->lru_lock
);
2081 mem_cgroup_uncharge_list(&l_hold
);
2082 free_hot_cold_page_list(&l_hold
, true);
2083 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2084 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2088 * The inactive anon list should be small enough that the VM never has
2089 * to do too much work.
2091 * The inactive file list should be small enough to leave most memory
2092 * to the established workingset on the scan-resistant active list,
2093 * but large enough to avoid thrashing the aggregate readahead window.
2095 * Both inactive lists should also be large enough that each inactive
2096 * page has a chance to be referenced again before it is reclaimed.
2098 * If that fails and refaulting is observed, the inactive list grows.
2100 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2101 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2102 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2105 * memory ratio inactive
2106 * -------------------------------------
2115 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2116 struct mem_cgroup
*memcg
,
2117 struct scan_control
*sc
, bool actual_reclaim
)
2119 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2120 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2121 enum lru_list inactive_lru
= file
* LRU_FILE
;
2122 unsigned long inactive
, active
;
2123 unsigned long inactive_ratio
;
2124 unsigned long refaults
;
2128 * If we don't have swap space, anonymous page deactivation
2131 if (!file
&& !total_swap_pages
)
2134 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2135 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2138 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2140 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2143 * When refaults are being observed, it means a new workingset
2144 * is being established. Disable active list protection to get
2145 * rid of the stale workingset quickly.
2147 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2150 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2152 inactive_ratio
= int_sqrt(10 * gb
);
2158 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2159 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2160 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2161 inactive_ratio
, file
);
2163 return inactive
* inactive_ratio
< active
;
2166 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2167 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2168 struct scan_control
*sc
)
2170 if (is_active_lru(lru
)) {
2171 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2173 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2177 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2188 * Determine how aggressively the anon and file LRU lists should be
2189 * scanned. The relative value of each set of LRU lists is determined
2190 * by looking at the fraction of the pages scanned we did rotate back
2191 * onto the active list instead of evict.
2193 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2194 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2196 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2197 struct scan_control
*sc
, unsigned long *nr
,
2198 unsigned long *lru_pages
)
2200 int swappiness
= mem_cgroup_swappiness(memcg
);
2201 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2203 u64 denominator
= 0; /* gcc */
2204 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2205 unsigned long anon_prio
, file_prio
;
2206 enum scan_balance scan_balance
;
2207 unsigned long anon
, file
;
2208 unsigned long ap
, fp
;
2211 /* If we have no swap space, do not bother scanning anon pages. */
2212 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2213 scan_balance
= SCAN_FILE
;
2218 * Global reclaim will swap to prevent OOM even with no
2219 * swappiness, but memcg users want to use this knob to
2220 * disable swapping for individual groups completely when
2221 * using the memory controller's swap limit feature would be
2224 if (!global_reclaim(sc
) && !swappiness
) {
2225 scan_balance
= SCAN_FILE
;
2230 * Do not apply any pressure balancing cleverness when the
2231 * system is close to OOM, scan both anon and file equally
2232 * (unless the swappiness setting disagrees with swapping).
2234 if (!sc
->priority
&& swappiness
) {
2235 scan_balance
= SCAN_EQUAL
;
2240 * Prevent the reclaimer from falling into the cache trap: as
2241 * cache pages start out inactive, every cache fault will tip
2242 * the scan balance towards the file LRU. And as the file LRU
2243 * shrinks, so does the window for rotation from references.
2244 * This means we have a runaway feedback loop where a tiny
2245 * thrashing file LRU becomes infinitely more attractive than
2246 * anon pages. Try to detect this based on file LRU size.
2248 if (global_reclaim(sc
)) {
2249 unsigned long pgdatfile
;
2250 unsigned long pgdatfree
;
2252 unsigned long total_high_wmark
= 0;
2254 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2255 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2256 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2258 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2259 struct zone
*zone
= &pgdat
->node_zones
[z
];
2260 if (!managed_zone(zone
))
2263 total_high_wmark
+= high_wmark_pages(zone
);
2266 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2268 * Force SCAN_ANON if there are enough inactive
2269 * anonymous pages on the LRU in eligible zones.
2270 * Otherwise, the small LRU gets thrashed.
2272 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2273 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2275 scan_balance
= SCAN_ANON
;
2282 * If there is enough inactive page cache, i.e. if the size of the
2283 * inactive list is greater than that of the active list *and* the
2284 * inactive list actually has some pages to scan on this priority, we
2285 * do not reclaim anything from the anonymous working set right now.
2286 * Without the second condition we could end up never scanning an
2287 * lruvec even if it has plenty of old anonymous pages unless the
2288 * system is under heavy pressure.
2290 if (!IS_ENABLED(CONFIG_BALANCE_ANON_FILE_RECLAIM
) &&
2291 !inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2292 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2293 scan_balance
= SCAN_FILE
;
2297 scan_balance
= SCAN_FRACT
;
2300 * With swappiness at 100, anonymous and file have the same priority.
2301 * This scanning priority is essentially the inverse of IO cost.
2303 anon_prio
= swappiness
;
2304 file_prio
= 200 - anon_prio
;
2307 * OK, so we have swap space and a fair amount of page cache
2308 * pages. We use the recently rotated / recently scanned
2309 * ratios to determine how valuable each cache is.
2311 * Because workloads change over time (and to avoid overflow)
2312 * we keep these statistics as a floating average, which ends
2313 * up weighing recent references more than old ones.
2315 * anon in [0], file in [1]
2318 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2319 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2320 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2321 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2323 spin_lock_irq(&pgdat
->lru_lock
);
2324 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2325 reclaim_stat
->recent_scanned
[0] /= 2;
2326 reclaim_stat
->recent_rotated
[0] /= 2;
2329 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2330 reclaim_stat
->recent_scanned
[1] /= 2;
2331 reclaim_stat
->recent_rotated
[1] /= 2;
2335 * The amount of pressure on anon vs file pages is inversely
2336 * proportional to the fraction of recently scanned pages on
2337 * each list that were recently referenced and in active use.
2339 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2340 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2342 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2343 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2344 spin_unlock_irq(&pgdat
->lru_lock
);
2348 denominator
= ap
+ fp
+ 1;
2351 for_each_evictable_lru(lru
) {
2352 int file
= is_file_lru(lru
);
2356 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2357 scan
= size
>> sc
->priority
;
2359 * If the cgroup's already been deleted, make sure to
2360 * scrape out the remaining cache.
2362 if (!scan
&& !mem_cgroup_online(memcg
))
2363 scan
= min(size
, SWAP_CLUSTER_MAX
);
2365 switch (scan_balance
) {
2367 /* Scan lists relative to size */
2371 * Scan types proportional to swappiness and
2372 * their relative recent reclaim efficiency.
2373 * Make sure we don't miss the last page
2374 * because of a round-off error.
2376 scan
= DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2381 /* Scan one type exclusively */
2382 if ((scan_balance
== SCAN_FILE
) != file
) {
2388 /* Look ma, no brain */
2398 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2400 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2401 struct scan_control
*sc
, unsigned long *lru_pages
)
2403 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2404 unsigned long nr
[NR_LRU_LISTS
];
2405 unsigned long targets
[NR_LRU_LISTS
];
2406 unsigned long nr_to_scan
;
2408 unsigned long nr_reclaimed
= 0;
2409 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2410 struct blk_plug plug
;
2413 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2415 /* Record the original scan target for proportional adjustments later */
2416 memcpy(targets
, nr
, sizeof(nr
));
2419 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2420 * event that can occur when there is little memory pressure e.g.
2421 * multiple streaming readers/writers. Hence, we do not abort scanning
2422 * when the requested number of pages are reclaimed when scanning at
2423 * DEF_PRIORITY on the assumption that the fact we are direct
2424 * reclaiming implies that kswapd is not keeping up and it is best to
2425 * do a batch of work at once. For memcg reclaim one check is made to
2426 * abort proportional reclaim if either the file or anon lru has already
2427 * dropped to zero at the first pass.
2429 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2430 sc
->priority
== DEF_PRIORITY
);
2432 blk_start_plug(&plug
);
2433 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2434 nr
[LRU_INACTIVE_FILE
]) {
2435 unsigned long nr_anon
, nr_file
, percentage
;
2436 unsigned long nr_scanned
;
2438 for_each_evictable_lru(lru
) {
2440 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2441 nr
[lru
] -= nr_to_scan
;
2443 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2450 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2454 * For kswapd and memcg, reclaim at least the number of pages
2455 * requested. Ensure that the anon and file LRUs are scanned
2456 * proportionally what was requested by get_scan_count(). We
2457 * stop reclaiming one LRU and reduce the amount scanning
2458 * proportional to the original scan target.
2460 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2461 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2464 * It's just vindictive to attack the larger once the smaller
2465 * has gone to zero. And given the way we stop scanning the
2466 * smaller below, this makes sure that we only make one nudge
2467 * towards proportionality once we've got nr_to_reclaim.
2469 if (!nr_file
|| !nr_anon
)
2472 if (nr_file
> nr_anon
) {
2473 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2474 targets
[LRU_ACTIVE_ANON
] + 1;
2476 percentage
= nr_anon
* 100 / scan_target
;
2478 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2479 targets
[LRU_ACTIVE_FILE
] + 1;
2481 percentage
= nr_file
* 100 / scan_target
;
2484 /* Stop scanning the smaller of the LRU */
2486 nr
[lru
+ LRU_ACTIVE
] = 0;
2489 * Recalculate the other LRU scan count based on its original
2490 * scan target and the percentage scanning already complete
2492 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2493 nr_scanned
= targets
[lru
] - nr
[lru
];
2494 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2495 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2498 nr_scanned
= targets
[lru
] - nr
[lru
];
2499 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2500 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2502 scan_adjusted
= true;
2504 blk_finish_plug(&plug
);
2505 sc
->nr_reclaimed
+= nr_reclaimed
;
2508 * Even if we did not try to evict anon pages at all, we want to
2509 * rebalance the anon lru active/inactive ratio.
2511 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2512 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2513 sc
, LRU_ACTIVE_ANON
);
2516 /* Use reclaim/compaction for costly allocs or under memory pressure */
2517 static bool in_reclaim_compaction(struct scan_control
*sc
)
2519 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2520 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2521 sc
->priority
< DEF_PRIORITY
- 2))
2528 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2529 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2530 * true if more pages should be reclaimed such that when the page allocator
2531 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2532 * It will give up earlier than that if there is difficulty reclaiming pages.
2534 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2535 unsigned long nr_reclaimed
,
2536 unsigned long nr_scanned
,
2537 struct scan_control
*sc
)
2539 unsigned long pages_for_compaction
;
2540 unsigned long inactive_lru_pages
;
2543 /* If not in reclaim/compaction mode, stop */
2544 if (!in_reclaim_compaction(sc
))
2547 /* Consider stopping depending on scan and reclaim activity */
2548 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2550 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2551 * full LRU list has been scanned and we are still failing
2552 * to reclaim pages. This full LRU scan is potentially
2553 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2555 if (!nr_reclaimed
&& !nr_scanned
)
2559 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2560 * fail without consequence, stop if we failed to reclaim
2561 * any pages from the last SWAP_CLUSTER_MAX number of
2562 * pages that were scanned. This will return to the
2563 * caller faster at the risk reclaim/compaction and
2564 * the resulting allocation attempt fails
2571 * If we have not reclaimed enough pages for compaction and the
2572 * inactive lists are large enough, continue reclaiming
2574 pages_for_compaction
= compact_gap(sc
->order
);
2575 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2576 if (get_nr_swap_pages() > 0)
2577 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2578 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2579 inactive_lru_pages
> pages_for_compaction
)
2582 /* If compaction would go ahead or the allocation would succeed, stop */
2583 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2584 struct zone
*zone
= &pgdat
->node_zones
[z
];
2585 if (!managed_zone(zone
))
2588 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2589 case COMPACT_SUCCESS
:
2590 case COMPACT_CONTINUE
:
2593 /* check next zone */
2600 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2602 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2603 unsigned long nr_reclaimed
, nr_scanned
;
2604 bool reclaimable
= false;
2607 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2608 struct mem_cgroup_reclaim_cookie reclaim
= {
2610 .priority
= sc
->priority
,
2612 unsigned long node_lru_pages
= 0;
2613 struct mem_cgroup
*memcg
;
2615 nr_reclaimed
= sc
->nr_reclaimed
;
2616 nr_scanned
= sc
->nr_scanned
;
2618 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2620 unsigned long lru_pages
;
2621 unsigned long reclaimed
;
2622 unsigned long scanned
;
2624 if (mem_cgroup_low(root
, memcg
)) {
2625 if (!sc
->memcg_low_reclaim
) {
2626 sc
->memcg_low_skipped
= 1;
2629 mem_cgroup_event(memcg
, MEMCG_LOW
);
2632 reclaimed
= sc
->nr_reclaimed
;
2633 scanned
= sc
->nr_scanned
;
2635 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2636 node_lru_pages
+= lru_pages
;
2639 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2640 memcg
, sc
->nr_scanned
- scanned
,
2643 /* Record the group's reclaim efficiency */
2644 vmpressure(sc
->gfp_mask
, memcg
, false,
2645 sc
->nr_scanned
- scanned
,
2646 sc
->nr_reclaimed
- reclaimed
);
2649 * Direct reclaim and kswapd have to scan all memory
2650 * cgroups to fulfill the overall scan target for the
2653 * Limit reclaim, on the other hand, only cares about
2654 * nr_to_reclaim pages to be reclaimed and it will
2655 * retry with decreasing priority if one round over the
2656 * whole hierarchy is not sufficient.
2658 if (!global_reclaim(sc
) &&
2659 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2660 mem_cgroup_iter_break(root
, memcg
);
2663 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2666 * Shrink the slab caches in the same proportion that
2667 * the eligible LRU pages were scanned.
2669 if (global_reclaim(sc
))
2670 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2671 sc
->nr_scanned
- nr_scanned
,
2674 if (reclaim_state
) {
2675 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2676 reclaim_state
->reclaimed_slab
= 0;
2679 /* Record the subtree's reclaim efficiency */
2680 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2681 sc
->nr_scanned
- nr_scanned
,
2682 sc
->nr_reclaimed
- nr_reclaimed
);
2684 if (sc
->nr_reclaimed
- nr_reclaimed
)
2687 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2688 sc
->nr_scanned
- nr_scanned
, sc
));
2691 * Kswapd gives up on balancing particular nodes after too
2692 * many failures to reclaim anything from them and goes to
2693 * sleep. On reclaim progress, reset the failure counter. A
2694 * successful direct reclaim run will revive a dormant kswapd.
2697 pgdat
->kswapd_failures
= 0;
2703 * Returns true if compaction should go ahead for a costly-order request, or
2704 * the allocation would already succeed without compaction. Return false if we
2705 * should reclaim first.
2707 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2709 unsigned long watermark
;
2710 enum compact_result suitable
;
2712 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2713 if (suitable
== COMPACT_SUCCESS
)
2714 /* Allocation should succeed already. Don't reclaim. */
2716 if (suitable
== COMPACT_SKIPPED
)
2717 /* Compaction cannot yet proceed. Do reclaim. */
2721 * Compaction is already possible, but it takes time to run and there
2722 * are potentially other callers using the pages just freed. So proceed
2723 * with reclaim to make a buffer of free pages available to give
2724 * compaction a reasonable chance of completing and allocating the page.
2725 * Note that we won't actually reclaim the whole buffer in one attempt
2726 * as the target watermark in should_continue_reclaim() is lower. But if
2727 * we are already above the high+gap watermark, don't reclaim at all.
2729 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2731 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2735 * This is the direct reclaim path, for page-allocating processes. We only
2736 * try to reclaim pages from zones which will satisfy the caller's allocation
2739 * If a zone is deemed to be full of pinned pages then just give it a light
2740 * scan then give up on it.
2742 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2746 unsigned long nr_soft_reclaimed
;
2747 unsigned long nr_soft_scanned
;
2749 pg_data_t
*last_pgdat
= NULL
;
2752 * If the number of buffer_heads in the machine exceeds the maximum
2753 * allowed level, force direct reclaim to scan the highmem zone as
2754 * highmem pages could be pinning lowmem pages storing buffer_heads
2756 orig_mask
= sc
->gfp_mask
;
2757 if (buffer_heads_over_limit
) {
2758 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2759 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2762 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2763 sc
->reclaim_idx
, sc
->nodemask
) {
2765 * Take care memory controller reclaiming has small influence
2768 if (global_reclaim(sc
)) {
2769 if (!cpuset_zone_allowed(zone
,
2770 GFP_KERNEL
| __GFP_HARDWALL
))
2774 * If we already have plenty of memory free for
2775 * compaction in this zone, don't free any more.
2776 * Even though compaction is invoked for any
2777 * non-zero order, only frequent costly order
2778 * reclamation is disruptive enough to become a
2779 * noticeable problem, like transparent huge
2782 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2783 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2784 compaction_ready(zone
, sc
)) {
2785 sc
->compaction_ready
= true;
2790 * Shrink each node in the zonelist once. If the
2791 * zonelist is ordered by zone (not the default) then a
2792 * node may be shrunk multiple times but in that case
2793 * the user prefers lower zones being preserved.
2795 if (zone
->zone_pgdat
== last_pgdat
)
2799 * This steals pages from memory cgroups over softlimit
2800 * and returns the number of reclaimed pages and
2801 * scanned pages. This works for global memory pressure
2802 * and balancing, not for a memcg's limit.
2804 nr_soft_scanned
= 0;
2805 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2806 sc
->order
, sc
->gfp_mask
,
2808 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2809 sc
->nr_scanned
+= nr_soft_scanned
;
2810 /* need some check for avoid more shrink_zone() */
2813 /* See comment about same check for global reclaim above */
2814 if (zone
->zone_pgdat
== last_pgdat
)
2816 last_pgdat
= zone
->zone_pgdat
;
2817 shrink_node(zone
->zone_pgdat
, sc
);
2821 * Restore to original mask to avoid the impact on the caller if we
2822 * promoted it to __GFP_HIGHMEM.
2824 sc
->gfp_mask
= orig_mask
;
2827 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2829 struct mem_cgroup
*memcg
;
2831 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2833 unsigned long refaults
;
2834 struct lruvec
*lruvec
;
2837 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2839 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2841 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2842 lruvec
->refaults
= refaults
;
2843 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2847 * This is the main entry point to direct page reclaim.
2849 * If a full scan of the inactive list fails to free enough memory then we
2850 * are "out of memory" and something needs to be killed.
2852 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2853 * high - the zone may be full of dirty or under-writeback pages, which this
2854 * caller can't do much about. We kick the writeback threads and take explicit
2855 * naps in the hope that some of these pages can be written. But if the
2856 * allocating task holds filesystem locks which prevent writeout this might not
2857 * work, and the allocation attempt will fail.
2859 * returns: 0, if no pages reclaimed
2860 * else, the number of pages reclaimed
2862 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2863 struct scan_control
*sc
)
2865 int initial_priority
= sc
->priority
;
2866 pg_data_t
*last_pgdat
;
2870 delayacct_freepages_start();
2872 if (global_reclaim(sc
))
2873 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2876 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2879 shrink_zones(zonelist
, sc
);
2881 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2884 if (sc
->compaction_ready
)
2888 * If we're getting trouble reclaiming, start doing
2889 * writepage even in laptop mode.
2891 if (sc
->priority
< DEF_PRIORITY
- 2)
2892 sc
->may_writepage
= 1;
2893 } while (--sc
->priority
>= 0);
2896 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2898 if (zone
->zone_pgdat
== last_pgdat
)
2900 last_pgdat
= zone
->zone_pgdat
;
2901 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2904 delayacct_freepages_end();
2906 if (sc
->nr_reclaimed
)
2907 return sc
->nr_reclaimed
;
2909 /* Aborted reclaim to try compaction? don't OOM, then */
2910 if (sc
->compaction_ready
)
2913 /* Untapped cgroup reserves? Don't OOM, retry. */
2914 if (sc
->memcg_low_skipped
) {
2915 sc
->priority
= initial_priority
;
2916 sc
->memcg_low_reclaim
= 1;
2917 sc
->memcg_low_skipped
= 0;
2924 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2927 unsigned long pfmemalloc_reserve
= 0;
2928 unsigned long free_pages
= 0;
2932 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2935 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2936 zone
= &pgdat
->node_zones
[i
];
2937 if (!managed_zone(zone
))
2940 if (!zone_reclaimable_pages(zone
))
2943 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2944 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2947 /* If there are no reserves (unexpected config) then do not throttle */
2948 if (!pfmemalloc_reserve
)
2951 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2953 /* kswapd must be awake if processes are being throttled */
2954 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2955 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2956 (enum zone_type
)ZONE_NORMAL
);
2957 wake_up_interruptible(&pgdat
->kswapd_wait
);
2964 * Throttle direct reclaimers if backing storage is backed by the network
2965 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2966 * depleted. kswapd will continue to make progress and wake the processes
2967 * when the low watermark is reached.
2969 * Returns true if a fatal signal was delivered during throttling. If this
2970 * happens, the page allocator should not consider triggering the OOM killer.
2972 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2973 nodemask_t
*nodemask
)
2977 pg_data_t
*pgdat
= NULL
;
2980 * Kernel threads should not be throttled as they may be indirectly
2981 * responsible for cleaning pages necessary for reclaim to make forward
2982 * progress. kjournald for example may enter direct reclaim while
2983 * committing a transaction where throttling it could forcing other
2984 * processes to block on log_wait_commit().
2986 if (current
->flags
& PF_KTHREAD
)
2990 * If a fatal signal is pending, this process should not throttle.
2991 * It should return quickly so it can exit and free its memory
2993 if (fatal_signal_pending(current
))
2997 * Check if the pfmemalloc reserves are ok by finding the first node
2998 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2999 * GFP_KERNEL will be required for allocating network buffers when
3000 * swapping over the network so ZONE_HIGHMEM is unusable.
3002 * Throttling is based on the first usable node and throttled processes
3003 * wait on a queue until kswapd makes progress and wakes them. There
3004 * is an affinity then between processes waking up and where reclaim
3005 * progress has been made assuming the process wakes on the same node.
3006 * More importantly, processes running on remote nodes will not compete
3007 * for remote pfmemalloc reserves and processes on different nodes
3008 * should make reasonable progress.
3010 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3011 gfp_zone(gfp_mask
), nodemask
) {
3012 if (zone_idx(zone
) > ZONE_NORMAL
)
3015 /* Throttle based on the first usable node */
3016 pgdat
= zone
->zone_pgdat
;
3017 if (allow_direct_reclaim(pgdat
))
3022 /* If no zone was usable by the allocation flags then do not throttle */
3026 /* Account for the throttling */
3027 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3030 * If the caller cannot enter the filesystem, it's possible that it
3031 * is due to the caller holding an FS lock or performing a journal
3032 * transaction in the case of a filesystem like ext[3|4]. In this case,
3033 * it is not safe to block on pfmemalloc_wait as kswapd could be
3034 * blocked waiting on the same lock. Instead, throttle for up to a
3035 * second before continuing.
3037 if (!(gfp_mask
& __GFP_FS
)) {
3038 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3039 allow_direct_reclaim(pgdat
), HZ
);
3044 /* Throttle until kswapd wakes the process */
3045 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3046 allow_direct_reclaim(pgdat
));
3049 if (fatal_signal_pending(current
))
3056 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3057 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3059 unsigned long nr_reclaimed
;
3060 struct scan_control sc
= {
3061 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3062 .gfp_mask
= current_gfp_context(gfp_mask
),
3063 .reclaim_idx
= gfp_zone(gfp_mask
),
3065 .nodemask
= nodemask
,
3066 .priority
= DEF_PRIORITY
,
3067 .may_writepage
= !laptop_mode
,
3073 * Do not enter reclaim if fatal signal was delivered while throttled.
3074 * 1 is returned so that the page allocator does not OOM kill at this
3077 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3080 trace_mm_vmscan_direct_reclaim_begin(order
,
3085 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3087 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3089 return nr_reclaimed
;
3094 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3095 gfp_t gfp_mask
, bool noswap
,
3097 unsigned long *nr_scanned
)
3099 struct scan_control sc
= {
3100 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3101 .target_mem_cgroup
= memcg
,
3102 .may_writepage
= !laptop_mode
,
3104 .reclaim_idx
= MAX_NR_ZONES
- 1,
3105 .may_swap
= !noswap
,
3107 unsigned long lru_pages
;
3109 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3110 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3112 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3118 * NOTE: Although we can get the priority field, using it
3119 * here is not a good idea, since it limits the pages we can scan.
3120 * if we don't reclaim here, the shrink_node from balance_pgdat
3121 * will pick up pages from other mem cgroup's as well. We hack
3122 * the priority and make it zero.
3124 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3126 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3128 *nr_scanned
= sc
.nr_scanned
;
3129 return sc
.nr_reclaimed
;
3132 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3133 unsigned long nr_pages
,
3137 struct zonelist
*zonelist
;
3138 unsigned long nr_reclaimed
;
3139 unsigned long pflags
;
3141 unsigned int noreclaim_flag
;
3142 struct scan_control sc
= {
3143 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3144 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3145 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3146 .reclaim_idx
= MAX_NR_ZONES
- 1,
3147 .target_mem_cgroup
= memcg
,
3148 .priority
= DEF_PRIORITY
,
3149 .may_writepage
= !laptop_mode
,
3151 .may_swap
= may_swap
,
3155 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3156 * take care of from where we get pages. So the node where we start the
3157 * scan does not need to be the current node.
3159 nid
= mem_cgroup_select_victim_node(memcg
);
3161 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3163 trace_mm_vmscan_memcg_reclaim_begin(0,
3168 psi_memstall_enter(&pflags
);
3169 noreclaim_flag
= memalloc_noreclaim_save();
3171 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3173 memalloc_noreclaim_restore(noreclaim_flag
);
3174 psi_memstall_leave(&pflags
);
3176 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3178 return nr_reclaimed
;
3182 static void age_active_anon(struct pglist_data
*pgdat
,
3183 struct scan_control
*sc
)
3185 struct mem_cgroup
*memcg
;
3187 if (!total_swap_pages
)
3190 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3192 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3194 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3195 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3196 sc
, LRU_ACTIVE_ANON
);
3198 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3203 * Returns true if there is an eligible zone balanced for the request order
3206 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3209 unsigned long mark
= -1;
3212 for (i
= 0; i
<= classzone_idx
; i
++) {
3213 zone
= pgdat
->node_zones
+ i
;
3215 if (!managed_zone(zone
))
3218 mark
= high_wmark_pages(zone
);
3219 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3224 * If a node has no populated zone within classzone_idx, it does not
3225 * need balancing by definition. This can happen if a zone-restricted
3226 * allocation tries to wake a remote kswapd.
3234 /* Clear pgdat state for congested, dirty or under writeback. */
3235 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3237 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3238 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3239 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3243 * Prepare kswapd for sleeping. This verifies that there are no processes
3244 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3246 * Returns true if kswapd is ready to sleep
3248 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3251 * The throttled processes are normally woken up in balance_pgdat() as
3252 * soon as allow_direct_reclaim() is true. But there is a potential
3253 * race between when kswapd checks the watermarks and a process gets
3254 * throttled. There is also a potential race if processes get
3255 * throttled, kswapd wakes, a large process exits thereby balancing the
3256 * zones, which causes kswapd to exit balance_pgdat() before reaching
3257 * the wake up checks. If kswapd is going to sleep, no process should
3258 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3259 * the wake up is premature, processes will wake kswapd and get
3260 * throttled again. The difference from wake ups in balance_pgdat() is
3261 * that here we are under prepare_to_wait().
3263 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3264 wake_up_all(&pgdat
->pfmemalloc_wait
);
3266 /* Hopeless node, leave it to direct reclaim */
3267 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3270 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3271 clear_pgdat_congested(pgdat
);
3279 * kswapd shrinks a node of pages that are at or below the highest usable
3280 * zone that is currently unbalanced.
3282 * Returns true if kswapd scanned at least the requested number of pages to
3283 * reclaim or if the lack of progress was due to pages under writeback.
3284 * This is used to determine if the scanning priority needs to be raised.
3286 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3287 struct scan_control
*sc
)
3292 /* Reclaim a number of pages proportional to the number of zones */
3293 sc
->nr_to_reclaim
= 0;
3294 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3295 zone
= pgdat
->node_zones
+ z
;
3296 if (!managed_zone(zone
))
3299 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3303 * Historically care was taken to put equal pressure on all zones but
3304 * now pressure is applied based on node LRU order.
3306 shrink_node(pgdat
, sc
);
3309 * Fragmentation may mean that the system cannot be rebalanced for
3310 * high-order allocations. If twice the allocation size has been
3311 * reclaimed then recheck watermarks only at order-0 to prevent
3312 * excessive reclaim. Assume that a process requested a high-order
3313 * can direct reclaim/compact.
3315 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3318 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3322 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3323 * that are eligible for use by the caller until at least one zone is
3326 * Returns the order kswapd finished reclaiming at.
3328 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3329 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3330 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3331 * or lower is eligible for reclaim until at least one usable zone is
3334 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3337 unsigned long nr_soft_reclaimed
;
3338 unsigned long nr_soft_scanned
;
3339 unsigned long pflags
;
3341 struct scan_control sc
= {
3342 .gfp_mask
= GFP_KERNEL
,
3344 .priority
= DEF_PRIORITY
,
3345 .may_writepage
= !laptop_mode
,
3349 psi_memstall_enter(&pflags
);
3350 count_vm_event(PAGEOUTRUN
);
3353 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3354 bool raise_priority
= true;
3356 sc
.reclaim_idx
= classzone_idx
;
3359 * If the number of buffer_heads exceeds the maximum allowed
3360 * then consider reclaiming from all zones. This has a dual
3361 * purpose -- on 64-bit systems it is expected that
3362 * buffer_heads are stripped during active rotation. On 32-bit
3363 * systems, highmem pages can pin lowmem memory and shrinking
3364 * buffers can relieve lowmem pressure. Reclaim may still not
3365 * go ahead if all eligible zones for the original allocation
3366 * request are balanced to avoid excessive reclaim from kswapd.
3368 if (buffer_heads_over_limit
) {
3369 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3370 zone
= pgdat
->node_zones
+ i
;
3371 if (!managed_zone(zone
))
3380 * Only reclaim if there are no eligible zones. Note that
3381 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3384 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3388 * Do some background aging of the anon list, to give
3389 * pages a chance to be referenced before reclaiming. All
3390 * pages are rotated regardless of classzone as this is
3391 * about consistent aging.
3393 age_active_anon(pgdat
, &sc
);
3396 * If we're getting trouble reclaiming, start doing writepage
3397 * even in laptop mode.
3399 if (sc
.priority
< DEF_PRIORITY
- 2)
3400 sc
.may_writepage
= 1;
3402 /* Call soft limit reclaim before calling shrink_node. */
3404 nr_soft_scanned
= 0;
3405 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3406 sc
.gfp_mask
, &nr_soft_scanned
);
3407 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3410 * There should be no need to raise the scanning priority if
3411 * enough pages are already being scanned that that high
3412 * watermark would be met at 100% efficiency.
3414 if (kswapd_shrink_node(pgdat
, &sc
))
3415 raise_priority
= false;
3418 * If the low watermark is met there is no need for processes
3419 * to be throttled on pfmemalloc_wait as they should not be
3420 * able to safely make forward progress. Wake them
3422 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3423 allow_direct_reclaim(pgdat
))
3424 wake_up_all(&pgdat
->pfmemalloc_wait
);
3426 /* Check if kswapd should be suspending */
3427 if (try_to_freeze() || kthread_should_stop())
3431 * Raise priority if scanning rate is too low or there was no
3432 * progress in reclaiming pages
3434 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3435 if (raise_priority
|| !nr_reclaimed
)
3437 } while (sc
.priority
>= 1);
3439 if (!sc
.nr_reclaimed
)
3440 pgdat
->kswapd_failures
++;
3443 snapshot_refaults(NULL
, pgdat
);
3444 psi_memstall_leave(&pflags
);
3446 * Return the order kswapd stopped reclaiming at as
3447 * prepare_kswapd_sleep() takes it into account. If another caller
3448 * entered the allocator slow path while kswapd was awake, order will
3449 * remain at the higher level.
3455 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3456 * allocation request woke kswapd for. When kswapd has not woken recently,
3457 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3458 * given classzone and returns it or the highest classzone index kswapd
3459 * was recently woke for.
3461 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3462 enum zone_type classzone_idx
)
3464 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3465 return classzone_idx
;
3467 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3470 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3471 unsigned int classzone_idx
)
3476 if (freezing(current
) || kthread_should_stop())
3479 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3482 * Try to sleep for a short interval. Note that kcompactd will only be
3483 * woken if it is possible to sleep for a short interval. This is
3484 * deliberate on the assumption that if reclaim cannot keep an
3485 * eligible zone balanced that it's also unlikely that compaction will
3488 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3490 * Compaction records what page blocks it recently failed to
3491 * isolate pages from and skips them in the future scanning.
3492 * When kswapd is going to sleep, it is reasonable to assume
3493 * that pages and compaction may succeed so reset the cache.
3495 reset_isolation_suitable(pgdat
);
3498 * We have freed the memory, now we should compact it to make
3499 * allocation of the requested order possible.
3501 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3503 remaining
= schedule_timeout(HZ
/10);
3506 * If woken prematurely then reset kswapd_classzone_idx and
3507 * order. The values will either be from a wakeup request or
3508 * the previous request that slept prematurely.
3511 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3512 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3515 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3516 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3520 * After a short sleep, check if it was a premature sleep. If not, then
3521 * go fully to sleep until explicitly woken up.
3524 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3525 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3528 * vmstat counters are not perfectly accurate and the estimated
3529 * value for counters such as NR_FREE_PAGES can deviate from the
3530 * true value by nr_online_cpus * threshold. To avoid the zone
3531 * watermarks being breached while under pressure, we reduce the
3532 * per-cpu vmstat threshold while kswapd is awake and restore
3533 * them before going back to sleep.
3535 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3537 if (!kthread_should_stop())
3540 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3543 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3545 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3547 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3551 * The background pageout daemon, started as a kernel thread
3552 * from the init process.
3554 * This basically trickles out pages so that we have _some_
3555 * free memory available even if there is no other activity
3556 * that frees anything up. This is needed for things like routing
3557 * etc, where we otherwise might have all activity going on in
3558 * asynchronous contexts that cannot page things out.
3560 * If there are applications that are active memory-allocators
3561 * (most normal use), this basically shouldn't matter.
3563 static int kswapd(void *p
)
3565 unsigned int alloc_order
, reclaim_order
;
3566 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3567 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3568 struct task_struct
*tsk
= current
;
3570 struct reclaim_state reclaim_state
= {
3571 .reclaimed_slab
= 0,
3573 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3575 if (!cpumask_empty(cpumask
))
3576 set_cpus_allowed_ptr(tsk
, cpumask
);
3577 current
->reclaim_state
= &reclaim_state
;
3580 * Tell the memory management that we're a "memory allocator",
3581 * and that if we need more memory we should get access to it
3582 * regardless (see "__alloc_pages()"). "kswapd" should
3583 * never get caught in the normal page freeing logic.
3585 * (Kswapd normally doesn't need memory anyway, but sometimes
3586 * you need a small amount of memory in order to be able to
3587 * page out something else, and this flag essentially protects
3588 * us from recursively trying to free more memory as we're
3589 * trying to free the first piece of memory in the first place).
3591 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3594 pgdat
->kswapd_order
= 0;
3595 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3599 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3600 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3603 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3606 /* Read the new order and classzone_idx */
3607 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3608 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3609 pgdat
->kswapd_order
= 0;
3610 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3612 ret
= try_to_freeze();
3613 if (kthread_should_stop())
3617 * We can speed up thawing tasks if we don't call balance_pgdat
3618 * after returning from the refrigerator
3624 * Reclaim begins at the requested order but if a high-order
3625 * reclaim fails then kswapd falls back to reclaiming for
3626 * order-0. If that happens, kswapd will consider sleeping
3627 * for the order it finished reclaiming at (reclaim_order)
3628 * but kcompactd is woken to compact for the original
3629 * request (alloc_order).
3631 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3633 fs_reclaim_acquire(GFP_KERNEL
);
3634 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3635 fs_reclaim_release(GFP_KERNEL
);
3636 if (reclaim_order
< alloc_order
)
3637 goto kswapd_try_sleep
;
3640 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3641 current
->reclaim_state
= NULL
;
3647 * A zone is low on free memory, so wake its kswapd task to service it.
3649 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3653 if (!managed_zone(zone
))
3656 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3658 pgdat
= zone
->zone_pgdat
;
3659 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3661 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3662 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3665 /* Hopeless node, leave it to direct reclaim */
3666 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3669 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3672 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3673 wake_up_interruptible(&pgdat
->kswapd_wait
);
3676 #ifdef CONFIG_HIBERNATION
3678 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3681 * Rather than trying to age LRUs the aim is to preserve the overall
3682 * LRU order by reclaiming preferentially
3683 * inactive > active > active referenced > active mapped
3685 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3687 struct reclaim_state reclaim_state
;
3688 struct scan_control sc
= {
3689 .nr_to_reclaim
= nr_to_reclaim
,
3690 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3691 .reclaim_idx
= MAX_NR_ZONES
- 1,
3692 .priority
= DEF_PRIORITY
,
3696 .hibernation_mode
= 1,
3698 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3699 struct task_struct
*p
= current
;
3700 unsigned long nr_reclaimed
;
3701 unsigned int noreclaim_flag
;
3703 noreclaim_flag
= memalloc_noreclaim_save();
3704 fs_reclaim_acquire(sc
.gfp_mask
);
3705 reclaim_state
.reclaimed_slab
= 0;
3706 p
->reclaim_state
= &reclaim_state
;
3708 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3710 p
->reclaim_state
= NULL
;
3711 fs_reclaim_release(sc
.gfp_mask
);
3712 memalloc_noreclaim_restore(noreclaim_flag
);
3714 return nr_reclaimed
;
3716 #endif /* CONFIG_HIBERNATION */
3718 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3719 not required for correctness. So if the last cpu in a node goes
3720 away, we get changed to run anywhere: as the first one comes back,
3721 restore their cpu bindings. */
3722 static int kswapd_cpu_online(unsigned int cpu
)
3726 for_each_node_state(nid
, N_MEMORY
) {
3727 pg_data_t
*pgdat
= NODE_DATA(nid
);
3728 const struct cpumask
*mask
;
3730 mask
= cpumask_of_node(pgdat
->node_id
);
3732 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3733 /* One of our CPUs online: restore mask */
3734 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3740 * This kswapd start function will be called by init and node-hot-add.
3741 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3743 int kswapd_run(int nid
)
3745 pg_data_t
*pgdat
= NODE_DATA(nid
);
3751 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3752 if (IS_ERR(pgdat
->kswapd
)) {
3753 /* failure at boot is fatal */
3754 BUG_ON(system_state
< SYSTEM_RUNNING
);
3755 pr_err("Failed to start kswapd on node %d\n", nid
);
3756 ret
= PTR_ERR(pgdat
->kswapd
);
3757 pgdat
->kswapd
= NULL
;
3763 * Called by memory hotplug when all memory in a node is offlined. Caller must
3764 * hold mem_hotplug_begin/end().
3766 void kswapd_stop(int nid
)
3768 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3771 kthread_stop(kswapd
);
3772 NODE_DATA(nid
)->kswapd
= NULL
;
3776 static int __init
kswapd_init(void)
3781 for_each_node_state(nid
, N_MEMORY
)
3783 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3784 "mm/vmscan:online", kswapd_cpu_online
,
3790 module_init(kswapd_init
)
3796 * If non-zero call node_reclaim when the number of free pages falls below
3799 int node_reclaim_mode __read_mostly
;
3801 #define RECLAIM_OFF 0
3802 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3803 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3804 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3807 * Priority for NODE_RECLAIM. This determines the fraction of pages
3808 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3811 #define NODE_RECLAIM_PRIORITY 4
3814 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3817 int sysctl_min_unmapped_ratio
= 1;
3820 * If the number of slab pages in a zone grows beyond this percentage then
3821 * slab reclaim needs to occur.
3823 int sysctl_min_slab_ratio
= 5;
3825 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3827 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3828 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3829 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3832 * It's possible for there to be more file mapped pages than
3833 * accounted for by the pages on the file LRU lists because
3834 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3836 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3839 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3840 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3842 unsigned long nr_pagecache_reclaimable
;
3843 unsigned long delta
= 0;
3846 * If RECLAIM_UNMAP is set, then all file pages are considered
3847 * potentially reclaimable. Otherwise, we have to worry about
3848 * pages like swapcache and node_unmapped_file_pages() provides
3851 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3852 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3854 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3856 /* If we can't clean pages, remove dirty pages from consideration */
3857 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3858 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3860 /* Watch for any possible underflows due to delta */
3861 if (unlikely(delta
> nr_pagecache_reclaimable
))
3862 delta
= nr_pagecache_reclaimable
;
3864 return nr_pagecache_reclaimable
- delta
;
3868 * Try to free up some pages from this node through reclaim.
3870 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3872 /* Minimum pages needed in order to stay on node */
3873 const unsigned long nr_pages
= 1 << order
;
3874 struct task_struct
*p
= current
;
3875 struct reclaim_state reclaim_state
;
3876 unsigned int noreclaim_flag
;
3877 struct scan_control sc
= {
3878 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3879 .gfp_mask
= current_gfp_context(gfp_mask
),
3881 .priority
= NODE_RECLAIM_PRIORITY
,
3882 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3883 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3885 .reclaim_idx
= gfp_zone(gfp_mask
),
3890 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3891 * and we also need to be able to write out pages for RECLAIM_WRITE
3892 * and RECLAIM_UNMAP.
3894 noreclaim_flag
= memalloc_noreclaim_save();
3895 p
->flags
|= PF_SWAPWRITE
;
3896 fs_reclaim_acquire(sc
.gfp_mask
);
3897 reclaim_state
.reclaimed_slab
= 0;
3898 p
->reclaim_state
= &reclaim_state
;
3900 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3902 * Free memory by calling shrink zone with increasing
3903 * priorities until we have enough memory freed.
3906 shrink_node(pgdat
, &sc
);
3907 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3910 p
->reclaim_state
= NULL
;
3911 fs_reclaim_release(gfp_mask
);
3912 current
->flags
&= ~PF_SWAPWRITE
;
3913 memalloc_noreclaim_restore(noreclaim_flag
);
3914 return sc
.nr_reclaimed
>= nr_pages
;
3917 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3922 * Node reclaim reclaims unmapped file backed pages and
3923 * slab pages if we are over the defined limits.
3925 * A small portion of unmapped file backed pages is needed for
3926 * file I/O otherwise pages read by file I/O will be immediately
3927 * thrown out if the node is overallocated. So we do not reclaim
3928 * if less than a specified percentage of the node is used by
3929 * unmapped file backed pages.
3931 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3932 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3933 return NODE_RECLAIM_FULL
;
3936 * Do not scan if the allocation should not be delayed.
3938 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3939 return NODE_RECLAIM_NOSCAN
;
3942 * Only run node reclaim on the local node or on nodes that do not
3943 * have associated processors. This will favor the local processor
3944 * over remote processors and spread off node memory allocations
3945 * as wide as possible.
3947 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3948 return NODE_RECLAIM_NOSCAN
;
3950 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3951 return NODE_RECLAIM_NOSCAN
;
3953 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3954 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3957 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3964 * page_evictable - test whether a page is evictable
3965 * @page: the page to test
3967 * Test whether page is evictable--i.e., should be placed on active/inactive
3968 * lists vs unevictable list.
3970 * Reasons page might not be evictable:
3971 * (1) page's mapping marked unevictable
3972 * (2) page is part of an mlocked VMA
3975 int page_evictable(struct page
*page
)
3979 /* Prevent address_space of inode and swap cache from being freed */
3981 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3988 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3989 * @pages: array of pages to check
3990 * @nr_pages: number of pages to check
3992 * Checks pages for evictability and moves them to the appropriate lru list.
3994 * This function is only used for SysV IPC SHM_UNLOCK.
3996 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3998 struct lruvec
*lruvec
;
3999 struct pglist_data
*pgdat
= NULL
;
4004 for (i
= 0; i
< nr_pages
; i
++) {
4005 struct page
*page
= pages
[i
];
4006 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4009 if (pagepgdat
!= pgdat
) {
4011 spin_unlock_irq(&pgdat
->lru_lock
);
4013 spin_lock_irq(&pgdat
->lru_lock
);
4015 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4017 if (!PageLRU(page
) || !PageUnevictable(page
))
4020 if (page_evictable(page
)) {
4021 enum lru_list lru
= page_lru_base_type(page
);
4023 VM_BUG_ON_PAGE(PageActive(page
), page
);
4024 ClearPageUnevictable(page
);
4025 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4026 add_page_to_lru_list(page
, lruvec
, lru
);
4032 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4033 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4034 spin_unlock_irq(&pgdat
->lru_lock
);
4037 #endif /* CONFIG_SHMEM */