4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/sched/mm.h>
18 #include <linux/module.h>
19 #include <linux/gfp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/vmpressure.h>
26 #include <linux/vmstat.h>
27 #include <linux/file.h>
28 #include <linux/writeback.h>
29 #include <linux/blkdev.h>
30 #include <linux/buffer_head.h> /* for try_to_release_page(),
31 buffer_heads_over_limit */
32 #include <linux/mm_inline.h>
33 #include <linux/backing-dev.h>
34 #include <linux/rmap.h>
35 #include <linux/topology.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/compaction.h>
39 #include <linux/notifier.h>
40 #include <linux/rwsem.h>
41 #include <linux/delay.h>
42 #include <linux/kthread.h>
43 #include <linux/freezer.h>
44 #include <linux/memcontrol.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
55 #include <linux/swapops.h>
56 #include <linux/balloon_compaction.h>
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/vmscan.h>
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim
;
67 /* This context's GFP mask */
70 /* Allocation order */
74 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * The memory cgroup that hit its limit and as a result is the
81 * primary target of this reclaim invocation.
83 struct mem_cgroup
*target_mem_cgroup
;
85 /* Scan (total_size >> priority) pages at once */
88 /* The highest zone to isolate pages for reclaim from */
89 enum zone_type reclaim_idx
;
91 /* Writepage batching in laptop mode; RECLAIM_WRITE */
92 unsigned int may_writepage
:1;
94 /* Can mapped pages be reclaimed? */
95 unsigned int may_unmap
:1;
97 /* Can pages be swapped as part of reclaim? */
98 unsigned int may_swap
:1;
101 * Cgroups are not reclaimed below their configured memory.low,
102 * unless we threaten to OOM. If any cgroups are skipped due to
103 * memory.low and nothing was reclaimed, go back for memory.low.
105 unsigned int memcg_low_reclaim
:1;
106 unsigned int memcg_low_skipped
:1;
108 unsigned int hibernation_mode
:1;
110 /* One of the zones is ready for compaction */
111 unsigned int compaction_ready
:1;
113 /* Incremented by the number of inactive pages that were scanned */
114 unsigned long nr_scanned
;
116 /* Number of pages freed so far during a call to shrink_zones() */
117 unsigned long nr_reclaimed
;
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness
= 60;
153 * The total number of pages which are beyond the high watermark within all
156 unsigned long vm_total_pages
;
158 static LIST_HEAD(shrinker_list
);
159 static DECLARE_RWSEM(shrinker_rwsem
);
162 static bool global_reclaim(struct scan_control
*sc
)
164 return !sc
->target_mem_cgroup
;
168 * sane_reclaim - is the usual dirty throttling mechanism operational?
169 * @sc: scan_control in question
171 * The normal page dirty throttling mechanism in balance_dirty_pages() is
172 * completely broken with the legacy memcg and direct stalling in
173 * shrink_page_list() is used for throttling instead, which lacks all the
174 * niceties such as fairness, adaptive pausing, bandwidth proportional
175 * allocation and configurability.
177 * This function tests whether the vmscan currently in progress can assume
178 * that the normal dirty throttling mechanism is operational.
180 static bool sane_reclaim(struct scan_control
*sc
)
182 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
186 #ifdef CONFIG_CGROUP_WRITEBACK
187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
193 static bool global_reclaim(struct scan_control
*sc
)
198 static bool sane_reclaim(struct scan_control
*sc
)
205 * This misses isolated pages which are not accounted for to save counters.
206 * As the data only determines if reclaim or compaction continues, it is
207 * not expected that isolated pages will be a dominating factor.
209 unsigned long zone_reclaimable_pages(struct zone
*zone
)
213 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
214 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
215 if (get_nr_swap_pages() > 0)
216 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
217 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
222 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
226 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
227 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
228 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
230 if (get_nr_swap_pages() > 0)
231 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
232 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
233 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
239 * lruvec_lru_size - Returns the number of pages on the given LRU list.
240 * @lruvec: lru vector
242 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
246 unsigned long lru_size
;
249 if (!mem_cgroup_disabled())
250 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
252 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
254 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
255 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
258 if (!managed_zone(zone
))
261 if (!mem_cgroup_disabled())
262 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
264 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
265 NR_ZONE_LRU_BASE
+ lru
);
266 lru_size
-= min(size
, lru_size
);
274 * Add a shrinker callback to be called from the vm.
276 int register_shrinker(struct shrinker
*shrinker
)
278 size_t size
= sizeof(*shrinker
->nr_deferred
);
280 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
283 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
284 if (!shrinker
->nr_deferred
)
287 down_write(&shrinker_rwsem
);
288 list_add_tail(&shrinker
->list
, &shrinker_list
);
289 up_write(&shrinker_rwsem
);
292 EXPORT_SYMBOL(register_shrinker
);
297 void unregister_shrinker(struct shrinker
*shrinker
)
299 down_write(&shrinker_rwsem
);
300 list_del(&shrinker
->list
);
301 up_write(&shrinker_rwsem
);
302 kfree(shrinker
->nr_deferred
);
304 EXPORT_SYMBOL(unregister_shrinker
);
306 #define SHRINK_BATCH 128
308 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
309 struct shrinker
*shrinker
,
310 unsigned long nr_scanned
,
311 unsigned long nr_eligible
)
313 unsigned long freed
= 0;
314 unsigned long long delta
;
319 int nid
= shrinkctl
->nid
;
320 long batch_size
= shrinker
->batch
? shrinker
->batch
322 long scanned
= 0, next_deferred
;
324 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
333 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
336 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
338 do_div(delta
, nr_eligible
+ 1);
340 if (total_scan
< 0) {
341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342 shrinker
->scan_objects
, total_scan
);
343 total_scan
= freeable
;
346 next_deferred
= total_scan
;
349 * We need to avoid excessive windup on filesystem shrinkers
350 * due to large numbers of GFP_NOFS allocations causing the
351 * shrinkers to return -1 all the time. This results in a large
352 * nr being built up so when a shrink that can do some work
353 * comes along it empties the entire cache due to nr >>>
354 * freeable. This is bad for sustaining a working set in
357 * Hence only allow the shrinker to scan the entire cache when
358 * a large delta change is calculated directly.
360 if (delta
< freeable
/ 4)
361 total_scan
= min(total_scan
, freeable
/ 2);
364 * Avoid risking looping forever due to too large nr value:
365 * never try to free more than twice the estimate number of
368 if (total_scan
> freeable
* 2)
369 total_scan
= freeable
* 2;
371 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
372 nr_scanned
, nr_eligible
,
373 freeable
, delta
, total_scan
);
376 * Normally, we should not scan less than batch_size objects in one
377 * pass to avoid too frequent shrinker calls, but if the slab has less
378 * than batch_size objects in total and we are really tight on memory,
379 * we will try to reclaim all available objects, otherwise we can end
380 * up failing allocations although there are plenty of reclaimable
381 * objects spread over several slabs with usage less than the
384 * We detect the "tight on memory" situations by looking at the total
385 * number of objects we want to scan (total_scan). If it is greater
386 * than the total number of objects on slab (freeable), we must be
387 * scanning at high prio and therefore should try to reclaim as much as
390 while (total_scan
>= batch_size
||
391 total_scan
>= freeable
) {
393 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
395 shrinkctl
->nr_to_scan
= nr_to_scan
;
396 shrinkctl
->nr_scanned
= nr_to_scan
;
397 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
398 if (ret
== SHRINK_STOP
)
402 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
403 total_scan
-= shrinkctl
->nr_scanned
;
404 scanned
+= shrinkctl
->nr_scanned
;
409 if (next_deferred
>= scanned
)
410 next_deferred
-= scanned
;
414 * move the unused scan count back into the shrinker in a
415 * manner that handles concurrent updates. If we exhausted the
416 * scan, there is no need to do an update.
418 if (next_deferred
> 0)
419 new_nr
= atomic_long_add_return(next_deferred
,
420 &shrinker
->nr_deferred
[nid
]);
422 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
424 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
429 * shrink_slab - shrink slab caches
430 * @gfp_mask: allocation context
431 * @nid: node whose slab caches to target
432 * @memcg: memory cgroup whose slab caches to target
433 * @nr_scanned: pressure numerator
434 * @nr_eligible: pressure denominator
436 * Call the shrink functions to age shrinkable caches.
438 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
439 * unaware shrinkers will receive a node id of 0 instead.
441 * @memcg specifies the memory cgroup to target. If it is not NULL,
442 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
443 * objects from the memory cgroup specified. Otherwise, only unaware
444 * shrinkers are called.
446 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
447 * the available objects should be scanned. Page reclaim for example
448 * passes the number of pages scanned and the number of pages on the
449 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
450 * when it encountered mapped pages. The ratio is further biased by
451 * the ->seeks setting of the shrink function, which indicates the
452 * cost to recreate an object relative to that of an LRU page.
454 * Returns the number of reclaimed slab objects.
456 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
457 struct mem_cgroup
*memcg
,
458 unsigned long nr_scanned
,
459 unsigned long nr_eligible
)
461 struct shrinker
*shrinker
;
462 unsigned long freed
= 0;
464 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
468 nr_scanned
= SWAP_CLUSTER_MAX
;
470 if (!down_read_trylock(&shrinker_rwsem
)) {
472 * If we would return 0, our callers would understand that we
473 * have nothing else to shrink and give up trying. By returning
474 * 1 we keep it going and assume we'll be able to shrink next
481 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
482 struct shrink_control sc
= {
483 .gfp_mask
= gfp_mask
,
489 * If kernel memory accounting is disabled, we ignore
490 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
491 * passing NULL for memcg.
493 if (memcg_kmem_enabled() &&
494 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
497 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
500 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
503 up_read(&shrinker_rwsem
);
509 void drop_slab_node(int nid
)
514 struct mem_cgroup
*memcg
= NULL
;
518 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
520 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
521 } while (freed
> 10);
528 for_each_online_node(nid
)
532 static inline int is_page_cache_freeable(struct page
*page
)
535 * A freeable page cache page is referenced only by the caller
536 * that isolated the page, the page cache radix tree and
537 * optional buffer heads at page->private.
539 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
541 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
544 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
546 if (current
->flags
& PF_SWAPWRITE
)
548 if (!inode_write_congested(inode
))
550 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
556 * We detected a synchronous write error writing a page out. Probably
557 * -ENOSPC. We need to propagate that into the address_space for a subsequent
558 * fsync(), msync() or close().
560 * The tricky part is that after writepage we cannot touch the mapping: nothing
561 * prevents it from being freed up. But we have a ref on the page and once
562 * that page is locked, the mapping is pinned.
564 * We're allowed to run sleeping lock_page() here because we know the caller has
567 static void handle_write_error(struct address_space
*mapping
,
568 struct page
*page
, int error
)
571 if (page_mapping(page
) == mapping
)
572 mapping_set_error(mapping
, error
);
576 /* possible outcome of pageout() */
578 /* failed to write page out, page is locked */
580 /* move page to the active list, page is locked */
582 /* page has been sent to the disk successfully, page is unlocked */
584 /* page is clean and locked */
589 * pageout is called by shrink_page_list() for each dirty page.
590 * Calls ->writepage().
592 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
593 struct scan_control
*sc
)
596 * If the page is dirty, only perform writeback if that write
597 * will be non-blocking. To prevent this allocation from being
598 * stalled by pagecache activity. But note that there may be
599 * stalls if we need to run get_block(). We could test
600 * PagePrivate for that.
602 * If this process is currently in __generic_file_write_iter() against
603 * this page's queue, we can perform writeback even if that
606 * If the page is swapcache, write it back even if that would
607 * block, for some throttling. This happens by accident, because
608 * swap_backing_dev_info is bust: it doesn't reflect the
609 * congestion state of the swapdevs. Easy to fix, if needed.
611 if (!is_page_cache_freeable(page
))
615 * Some data journaling orphaned pages can have
616 * page->mapping == NULL while being dirty with clean buffers.
618 if (page_has_private(page
)) {
619 if (try_to_free_buffers(page
)) {
620 ClearPageDirty(page
);
621 pr_info("%s: orphaned page\n", __func__
);
627 if (mapping
->a_ops
->writepage
== NULL
)
628 return PAGE_ACTIVATE
;
629 if (!may_write_to_inode(mapping
->host
, sc
))
632 if (clear_page_dirty_for_io(page
)) {
634 struct writeback_control wbc
= {
635 .sync_mode
= WB_SYNC_NONE
,
636 .nr_to_write
= SWAP_CLUSTER_MAX
,
638 .range_end
= LLONG_MAX
,
642 SetPageReclaim(page
);
643 res
= mapping
->a_ops
->writepage(page
, &wbc
);
645 handle_write_error(mapping
, page
, res
);
646 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
647 ClearPageReclaim(page
);
648 return PAGE_ACTIVATE
;
651 if (!PageWriteback(page
)) {
652 /* synchronous write or broken a_ops? */
653 ClearPageReclaim(page
);
655 trace_mm_vmscan_writepage(page
);
656 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
664 * Same as remove_mapping, but if the page is removed from the mapping, it
665 * gets returned with a refcount of 0.
667 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
673 BUG_ON(!PageLocked(page
));
674 BUG_ON(mapping
!= page_mapping(page
));
676 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
678 * The non racy check for a busy page.
680 * Must be careful with the order of the tests. When someone has
681 * a ref to the page, it may be possible that they dirty it then
682 * drop the reference. So if PageDirty is tested before page_count
683 * here, then the following race may occur:
685 * get_user_pages(&page);
686 * [user mapping goes away]
688 * !PageDirty(page) [good]
689 * SetPageDirty(page);
691 * !page_count(page) [good, discard it]
693 * [oops, our write_to data is lost]
695 * Reversing the order of the tests ensures such a situation cannot
696 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
697 * load is not satisfied before that of page->_refcount.
699 * Note that if SetPageDirty is always performed via set_page_dirty,
700 * and thus under tree_lock, then this ordering is not required.
702 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
703 refcount
= 1 + HPAGE_PMD_NR
;
706 if (!page_ref_freeze(page
, refcount
))
708 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
709 if (unlikely(PageDirty(page
))) {
710 page_ref_unfreeze(page
, refcount
);
714 if (PageSwapCache(page
)) {
715 swp_entry_t swap
= { .val
= page_private(page
) };
716 mem_cgroup_swapout(page
, swap
);
717 __delete_from_swap_cache(page
);
718 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
719 put_swap_page(page
, swap
);
721 void (*freepage
)(struct page
*);
724 freepage
= mapping
->a_ops
->freepage
;
726 * Remember a shadow entry for reclaimed file cache in
727 * order to detect refaults, thus thrashing, later on.
729 * But don't store shadows in an address space that is
730 * already exiting. This is not just an optizimation,
731 * inode reclaim needs to empty out the radix tree or
732 * the nodes are lost. Don't plant shadows behind its
735 * We also don't store shadows for DAX mappings because the
736 * only page cache pages found in these are zero pages
737 * covering holes, and because we don't want to mix DAX
738 * exceptional entries and shadow exceptional entries in the
741 if (reclaimed
&& page_is_file_cache(page
) &&
742 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
743 shadow
= workingset_eviction(mapping
, page
);
744 __delete_from_page_cache(page
, shadow
);
745 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
747 if (freepage
!= NULL
)
754 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
759 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
760 * someone else has a ref on the page, abort and return 0. If it was
761 * successfully detached, return 1. Assumes the caller has a single ref on
764 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
766 if (__remove_mapping(mapping
, page
, false)) {
768 * Unfreezing the refcount with 1 rather than 2 effectively
769 * drops the pagecache ref for us without requiring another
772 page_ref_unfreeze(page
, 1);
779 * putback_lru_page - put previously isolated page onto appropriate LRU list
780 * @page: page to be put back to appropriate lru list
782 * Add previously isolated @page to appropriate LRU list.
783 * Page may still be unevictable for other reasons.
785 * lru_lock must not be held, interrupts must be enabled.
787 void putback_lru_page(struct page
*page
)
790 int was_unevictable
= PageUnevictable(page
);
792 VM_BUG_ON_PAGE(PageLRU(page
), page
);
795 ClearPageUnevictable(page
);
797 if (page_evictable(page
)) {
799 * For evictable pages, we can use the cache.
800 * In event of a race, worst case is we end up with an
801 * unevictable page on [in]active list.
802 * We know how to handle that.
804 is_unevictable
= false;
808 * Put unevictable pages directly on zone's unevictable
811 is_unevictable
= true;
812 add_page_to_unevictable_list(page
);
814 * When racing with an mlock or AS_UNEVICTABLE clearing
815 * (page is unlocked) make sure that if the other thread
816 * does not observe our setting of PG_lru and fails
817 * isolation/check_move_unevictable_pages,
818 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
819 * the page back to the evictable list.
821 * The other side is TestClearPageMlocked() or shmem_lock().
827 * page's status can change while we move it among lru. If an evictable
828 * page is on unevictable list, it never be freed. To avoid that,
829 * check after we added it to the list, again.
831 if (is_unevictable
&& page_evictable(page
)) {
832 if (!isolate_lru_page(page
)) {
836 /* This means someone else dropped this page from LRU
837 * So, it will be freed or putback to LRU again. There is
838 * nothing to do here.
842 if (was_unevictable
&& !is_unevictable
)
843 count_vm_event(UNEVICTABLE_PGRESCUED
);
844 else if (!was_unevictable
&& is_unevictable
)
845 count_vm_event(UNEVICTABLE_PGCULLED
);
847 put_page(page
); /* drop ref from isolate */
850 enum page_references
{
852 PAGEREF_RECLAIM_CLEAN
,
857 static enum page_references
page_check_references(struct page
*page
,
858 struct scan_control
*sc
)
860 int referenced_ptes
, referenced_page
;
861 unsigned long vm_flags
;
863 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
865 referenced_page
= TestClearPageReferenced(page
);
868 * Mlock lost the isolation race with us. Let try_to_unmap()
869 * move the page to the unevictable list.
871 if (vm_flags
& VM_LOCKED
)
872 return PAGEREF_RECLAIM
;
874 if (referenced_ptes
) {
875 if (PageSwapBacked(page
))
876 return PAGEREF_ACTIVATE
;
878 * All mapped pages start out with page table
879 * references from the instantiating fault, so we need
880 * to look twice if a mapped file page is used more
883 * Mark it and spare it for another trip around the
884 * inactive list. Another page table reference will
885 * lead to its activation.
887 * Note: the mark is set for activated pages as well
888 * so that recently deactivated but used pages are
891 SetPageReferenced(page
);
893 if (referenced_page
|| referenced_ptes
> 1)
894 return PAGEREF_ACTIVATE
;
897 * Activate file-backed executable pages after first usage.
899 if (vm_flags
& VM_EXEC
)
900 return PAGEREF_ACTIVATE
;
905 /* Reclaim if clean, defer dirty pages to writeback */
906 if (referenced_page
&& !PageSwapBacked(page
))
907 return PAGEREF_RECLAIM_CLEAN
;
909 return PAGEREF_RECLAIM
;
912 /* Check if a page is dirty or under writeback */
913 static void page_check_dirty_writeback(struct page
*page
,
914 bool *dirty
, bool *writeback
)
916 struct address_space
*mapping
;
919 * Anonymous pages are not handled by flushers and must be written
920 * from reclaim context. Do not stall reclaim based on them
922 if (!page_is_file_cache(page
) ||
923 (PageAnon(page
) && !PageSwapBacked(page
))) {
929 /* By default assume that the page flags are accurate */
930 *dirty
= PageDirty(page
);
931 *writeback
= PageWriteback(page
);
933 /* Verify dirty/writeback state if the filesystem supports it */
934 if (!page_has_private(page
))
937 mapping
= page_mapping(page
);
938 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
939 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
942 struct reclaim_stat
{
944 unsigned nr_unqueued_dirty
;
945 unsigned nr_congested
;
946 unsigned nr_writeback
;
947 unsigned nr_immediate
;
948 unsigned nr_activate
;
949 unsigned nr_ref_keep
;
950 unsigned nr_unmap_fail
;
954 * shrink_page_list() returns the number of reclaimed pages
956 static unsigned long shrink_page_list(struct list_head
*page_list
,
957 struct pglist_data
*pgdat
,
958 struct scan_control
*sc
,
959 enum ttu_flags ttu_flags
,
960 struct reclaim_stat
*stat
,
963 LIST_HEAD(ret_pages
);
964 LIST_HEAD(free_pages
);
966 unsigned nr_unqueued_dirty
= 0;
967 unsigned nr_dirty
= 0;
968 unsigned nr_congested
= 0;
969 unsigned nr_reclaimed
= 0;
970 unsigned nr_writeback
= 0;
971 unsigned nr_immediate
= 0;
972 unsigned nr_ref_keep
= 0;
973 unsigned nr_unmap_fail
= 0;
977 while (!list_empty(page_list
)) {
978 struct address_space
*mapping
;
981 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
982 bool dirty
, writeback
;
986 page
= lru_to_page(page_list
);
987 list_del(&page
->lru
);
989 if (!trylock_page(page
))
992 VM_BUG_ON_PAGE(PageActive(page
), page
);
996 if (unlikely(!page_evictable(page
)))
997 goto activate_locked
;
999 if (!sc
->may_unmap
&& page_mapped(page
))
1002 /* Double the slab pressure for mapped and swapcache pages */
1003 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1004 !(PageAnon(page
) && !PageSwapBacked(page
)))
1007 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1008 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1011 * The number of dirty pages determines if a zone is marked
1012 * reclaim_congested which affects wait_iff_congested. kswapd
1013 * will stall and start writing pages if the tail of the LRU
1014 * is all dirty unqueued pages.
1016 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1017 if (dirty
|| writeback
)
1020 if (dirty
&& !writeback
)
1021 nr_unqueued_dirty
++;
1024 * Treat this page as congested if the underlying BDI is or if
1025 * pages are cycling through the LRU so quickly that the
1026 * pages marked for immediate reclaim are making it to the
1027 * end of the LRU a second time.
1029 mapping
= page_mapping(page
);
1030 if (((dirty
|| writeback
) && mapping
&&
1031 inode_write_congested(mapping
->host
)) ||
1032 (writeback
&& PageReclaim(page
)))
1036 * If a page at the tail of the LRU is under writeback, there
1037 * are three cases to consider.
1039 * 1) If reclaim is encountering an excessive number of pages
1040 * under writeback and this page is both under writeback and
1041 * PageReclaim then it indicates that pages are being queued
1042 * for IO but are being recycled through the LRU before the
1043 * IO can complete. Waiting on the page itself risks an
1044 * indefinite stall if it is impossible to writeback the
1045 * page due to IO error or disconnected storage so instead
1046 * note that the LRU is being scanned too quickly and the
1047 * caller can stall after page list has been processed.
1049 * 2) Global or new memcg reclaim encounters a page that is
1050 * not marked for immediate reclaim, or the caller does not
1051 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1052 * not to fs). In this case mark the page for immediate
1053 * reclaim and continue scanning.
1055 * Require may_enter_fs because we would wait on fs, which
1056 * may not have submitted IO yet. And the loop driver might
1057 * enter reclaim, and deadlock if it waits on a page for
1058 * which it is needed to do the write (loop masks off
1059 * __GFP_IO|__GFP_FS for this reason); but more thought
1060 * would probably show more reasons.
1062 * 3) Legacy memcg encounters a page that is already marked
1063 * PageReclaim. memcg does not have any dirty pages
1064 * throttling so we could easily OOM just because too many
1065 * pages are in writeback and there is nothing else to
1066 * reclaim. Wait for the writeback to complete.
1068 * In cases 1) and 2) we activate the pages to get them out of
1069 * the way while we continue scanning for clean pages on the
1070 * inactive list and refilling from the active list. The
1071 * observation here is that waiting for disk writes is more
1072 * expensive than potentially causing reloads down the line.
1073 * Since they're marked for immediate reclaim, they won't put
1074 * memory pressure on the cache working set any longer than it
1075 * takes to write them to disk.
1077 if (PageWriteback(page
)) {
1079 if (current_is_kswapd() &&
1080 PageReclaim(page
) &&
1081 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1083 goto activate_locked
;
1086 } else if (sane_reclaim(sc
) ||
1087 !PageReclaim(page
) || !may_enter_fs
) {
1089 * This is slightly racy - end_page_writeback()
1090 * might have just cleared PageReclaim, then
1091 * setting PageReclaim here end up interpreted
1092 * as PageReadahead - but that does not matter
1093 * enough to care. What we do want is for this
1094 * page to have PageReclaim set next time memcg
1095 * reclaim reaches the tests above, so it will
1096 * then wait_on_page_writeback() to avoid OOM;
1097 * and it's also appropriate in global reclaim.
1099 SetPageReclaim(page
);
1101 goto activate_locked
;
1106 wait_on_page_writeback(page
);
1107 /* then go back and try same page again */
1108 list_add_tail(&page
->lru
, page_list
);
1114 references
= page_check_references(page
, sc
);
1116 switch (references
) {
1117 case PAGEREF_ACTIVATE
:
1118 goto activate_locked
;
1122 case PAGEREF_RECLAIM
:
1123 case PAGEREF_RECLAIM_CLEAN
:
1124 ; /* try to reclaim the page below */
1128 * Anonymous process memory has backing store?
1129 * Try to allocate it some swap space here.
1130 * Lazyfree page could be freed directly
1132 if (PageAnon(page
) && PageSwapBacked(page
)) {
1133 if (!PageSwapCache(page
)) {
1134 if (!(sc
->gfp_mask
& __GFP_IO
))
1136 if (PageTransHuge(page
)) {
1137 /* cannot split THP, skip it */
1138 if (!can_split_huge_page(page
, NULL
))
1139 goto activate_locked
;
1141 * Split pages without a PMD map right
1142 * away. Chances are some or all of the
1143 * tail pages can be freed without IO.
1145 if (!compound_mapcount(page
) &&
1146 split_huge_page_to_list(page
,
1148 goto activate_locked
;
1150 if (!add_to_swap(page
)) {
1151 if (!PageTransHuge(page
))
1152 goto activate_locked
;
1153 /* Fallback to swap normal pages */
1154 if (split_huge_page_to_list(page
,
1156 goto activate_locked
;
1157 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1158 count_vm_event(THP_SWPOUT_FALLBACK
);
1160 if (!add_to_swap(page
))
1161 goto activate_locked
;
1166 /* Adding to swap updated mapping */
1167 mapping
= page_mapping(page
);
1169 } else if (unlikely(PageTransHuge(page
))) {
1170 /* Split file THP */
1171 if (split_huge_page_to_list(page
, page_list
))
1176 * The page is mapped into the page tables of one or more
1177 * processes. Try to unmap it here.
1179 if (page_mapped(page
)) {
1180 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1182 if (unlikely(PageTransHuge(page
)))
1183 flags
|= TTU_SPLIT_HUGE_PMD
;
1184 if (!try_to_unmap(page
, flags
)) {
1186 goto activate_locked
;
1190 if (PageDirty(page
)) {
1192 * Only kswapd can writeback filesystem pages
1193 * to avoid risk of stack overflow. But avoid
1194 * injecting inefficient single-page IO into
1195 * flusher writeback as much as possible: only
1196 * write pages when we've encountered many
1197 * dirty pages, and when we've already scanned
1198 * the rest of the LRU for clean pages and see
1199 * the same dirty pages again (PageReclaim).
1201 if (page_is_file_cache(page
) &&
1202 (!current_is_kswapd() || !PageReclaim(page
) ||
1203 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1205 * Immediately reclaim when written back.
1206 * Similar in principal to deactivate_page()
1207 * except we already have the page isolated
1208 * and know it's dirty
1210 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1211 SetPageReclaim(page
);
1213 goto activate_locked
;
1216 if (references
== PAGEREF_RECLAIM_CLEAN
)
1220 if (!sc
->may_writepage
)
1224 * Page is dirty. Flush the TLB if a writable entry
1225 * potentially exists to avoid CPU writes after IO
1226 * starts and then write it out here.
1228 try_to_unmap_flush_dirty();
1229 switch (pageout(page
, mapping
, sc
)) {
1233 goto activate_locked
;
1235 if (PageWriteback(page
))
1237 if (PageDirty(page
))
1241 * A synchronous write - probably a ramdisk. Go
1242 * ahead and try to reclaim the page.
1244 if (!trylock_page(page
))
1246 if (PageDirty(page
) || PageWriteback(page
))
1248 mapping
= page_mapping(page
);
1250 ; /* try to free the page below */
1255 * If the page has buffers, try to free the buffer mappings
1256 * associated with this page. If we succeed we try to free
1259 * We do this even if the page is PageDirty().
1260 * try_to_release_page() does not perform I/O, but it is
1261 * possible for a page to have PageDirty set, but it is actually
1262 * clean (all its buffers are clean). This happens if the
1263 * buffers were written out directly, with submit_bh(). ext3
1264 * will do this, as well as the blockdev mapping.
1265 * try_to_release_page() will discover that cleanness and will
1266 * drop the buffers and mark the page clean - it can be freed.
1268 * Rarely, pages can have buffers and no ->mapping. These are
1269 * the pages which were not successfully invalidated in
1270 * truncate_complete_page(). We try to drop those buffers here
1271 * and if that worked, and the page is no longer mapped into
1272 * process address space (page_count == 1) it can be freed.
1273 * Otherwise, leave the page on the LRU so it is swappable.
1275 if (page_has_private(page
)) {
1276 if (!try_to_release_page(page
, sc
->gfp_mask
))
1277 goto activate_locked
;
1278 if (!mapping
&& page_count(page
) == 1) {
1280 if (put_page_testzero(page
))
1284 * rare race with speculative reference.
1285 * the speculative reference will free
1286 * this page shortly, so we may
1287 * increment nr_reclaimed here (and
1288 * leave it off the LRU).
1296 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1297 /* follow __remove_mapping for reference */
1298 if (!page_ref_freeze(page
, 1))
1300 if (PageDirty(page
)) {
1301 page_ref_unfreeze(page
, 1);
1305 count_vm_event(PGLAZYFREED
);
1306 count_memcg_page_event(page
, PGLAZYFREED
);
1307 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1310 * At this point, we have no other references and there is
1311 * no way to pick any more up (removed from LRU, removed
1312 * from pagecache). Can use non-atomic bitops now (and
1313 * we obviously don't have to worry about waking up a process
1314 * waiting on the page lock, because there are no references.
1316 __ClearPageLocked(page
);
1321 * Is there need to periodically free_page_list? It would
1322 * appear not as the counts should be low
1324 if (unlikely(PageTransHuge(page
))) {
1325 mem_cgroup_uncharge(page
);
1326 (*get_compound_page_dtor(page
))(page
);
1328 list_add(&page
->lru
, &free_pages
);
1332 /* Not a candidate for swapping, so reclaim swap space. */
1333 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1335 try_to_free_swap(page
);
1336 VM_BUG_ON_PAGE(PageActive(page
), page
);
1337 if (!PageMlocked(page
)) {
1338 SetPageActive(page
);
1340 count_memcg_page_event(page
, PGACTIVATE
);
1345 list_add(&page
->lru
, &ret_pages
);
1346 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1349 mem_cgroup_uncharge_list(&free_pages
);
1350 try_to_unmap_flush();
1351 free_hot_cold_page_list(&free_pages
, true);
1353 list_splice(&ret_pages
, page_list
);
1354 count_vm_events(PGACTIVATE
, pgactivate
);
1357 stat
->nr_dirty
= nr_dirty
;
1358 stat
->nr_congested
= nr_congested
;
1359 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1360 stat
->nr_writeback
= nr_writeback
;
1361 stat
->nr_immediate
= nr_immediate
;
1362 stat
->nr_activate
= pgactivate
;
1363 stat
->nr_ref_keep
= nr_ref_keep
;
1364 stat
->nr_unmap_fail
= nr_unmap_fail
;
1366 return nr_reclaimed
;
1369 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1370 struct list_head
*page_list
)
1372 struct scan_control sc
= {
1373 .gfp_mask
= GFP_KERNEL
,
1374 .priority
= DEF_PRIORITY
,
1378 struct page
*page
, *next
;
1379 LIST_HEAD(clean_pages
);
1381 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1382 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1383 !__PageMovable(page
)) {
1384 ClearPageActive(page
);
1385 list_move(&page
->lru
, &clean_pages
);
1389 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1390 TTU_IGNORE_ACCESS
, NULL
, true);
1391 list_splice(&clean_pages
, page_list
);
1392 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1397 * Attempt to remove the specified page from its LRU. Only take this page
1398 * if it is of the appropriate PageActive status. Pages which are being
1399 * freed elsewhere are also ignored.
1401 * page: page to consider
1402 * mode: one of the LRU isolation modes defined above
1404 * returns 0 on success, -ve errno on failure.
1406 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1410 /* Only take pages on the LRU. */
1414 /* Compaction should not handle unevictable pages but CMA can do so */
1415 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1421 * To minimise LRU disruption, the caller can indicate that it only
1422 * wants to isolate pages it will be able to operate on without
1423 * blocking - clean pages for the most part.
1425 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1426 * that it is possible to migrate without blocking
1428 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1429 /* All the caller can do on PageWriteback is block */
1430 if (PageWriteback(page
))
1433 if (PageDirty(page
)) {
1434 struct address_space
*mapping
;
1437 * Only pages without mappings or that have a
1438 * ->migratepage callback are possible to migrate
1441 mapping
= page_mapping(page
);
1442 if (mapping
&& !mapping
->a_ops
->migratepage
)
1447 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1450 if (likely(get_page_unless_zero(page
))) {
1452 * Be careful not to clear PageLRU until after we're
1453 * sure the page is not being freed elsewhere -- the
1454 * page release code relies on it.
1465 * Update LRU sizes after isolating pages. The LRU size updates must
1466 * be complete before mem_cgroup_update_lru_size due to a santity check.
1468 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1469 enum lru_list lru
, unsigned long *nr_zone_taken
)
1473 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1474 if (!nr_zone_taken
[zid
])
1477 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1479 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1486 * zone_lru_lock is heavily contended. Some of the functions that
1487 * shrink the lists perform better by taking out a batch of pages
1488 * and working on them outside the LRU lock.
1490 * For pagecache intensive workloads, this function is the hottest
1491 * spot in the kernel (apart from copy_*_user functions).
1493 * Appropriate locks must be held before calling this function.
1495 * @nr_to_scan: The number of eligible pages to look through on the list.
1496 * @lruvec: The LRU vector to pull pages from.
1497 * @dst: The temp list to put pages on to.
1498 * @nr_scanned: The number of pages that were scanned.
1499 * @sc: The scan_control struct for this reclaim session
1500 * @mode: One of the LRU isolation modes
1501 * @lru: LRU list id for isolating
1503 * returns how many pages were moved onto *@dst.
1505 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1506 struct lruvec
*lruvec
, struct list_head
*dst
,
1507 unsigned long *nr_scanned
, struct scan_control
*sc
,
1508 isolate_mode_t mode
, enum lru_list lru
)
1510 struct list_head
*src
= &lruvec
->lists
[lru
];
1511 unsigned long nr_taken
= 0;
1512 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1513 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1514 unsigned long skipped
= 0;
1515 unsigned long scan
, total_scan
, nr_pages
;
1516 LIST_HEAD(pages_skipped
);
1519 for (total_scan
= 0;
1520 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1524 page
= lru_to_page(src
);
1525 prefetchw_prev_lru_page(page
, src
, flags
);
1527 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1529 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1530 list_move(&page
->lru
, &pages_skipped
);
1531 nr_skipped
[page_zonenum(page
)]++;
1536 * Do not count skipped pages because that makes the function
1537 * return with no isolated pages if the LRU mostly contains
1538 * ineligible pages. This causes the VM to not reclaim any
1539 * pages, triggering a premature OOM.
1542 switch (__isolate_lru_page(page
, mode
)) {
1544 nr_pages
= hpage_nr_pages(page
);
1545 nr_taken
+= nr_pages
;
1546 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1547 list_move(&page
->lru
, dst
);
1551 /* else it is being freed elsewhere */
1552 list_move(&page
->lru
, src
);
1561 * Splice any skipped pages to the start of the LRU list. Note that
1562 * this disrupts the LRU order when reclaiming for lower zones but
1563 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1564 * scanning would soon rescan the same pages to skip and put the
1565 * system at risk of premature OOM.
1567 if (!list_empty(&pages_skipped
)) {
1570 list_splice(&pages_skipped
, src
);
1571 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1572 if (!nr_skipped
[zid
])
1575 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1576 skipped
+= nr_skipped
[zid
];
1579 *nr_scanned
= total_scan
;
1580 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1581 total_scan
, skipped
, nr_taken
, mode
, lru
);
1582 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1587 * isolate_lru_page - tries to isolate a page from its LRU list
1588 * @page: page to isolate from its LRU list
1590 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1591 * vmstat statistic corresponding to whatever LRU list the page was on.
1593 * Returns 0 if the page was removed from an LRU list.
1594 * Returns -EBUSY if the page was not on an LRU list.
1596 * The returned page will have PageLRU() cleared. If it was found on
1597 * the active list, it will have PageActive set. If it was found on
1598 * the unevictable list, it will have the PageUnevictable bit set. That flag
1599 * may need to be cleared by the caller before letting the page go.
1601 * The vmstat statistic corresponding to the list on which the page was
1602 * found will be decremented.
1605 * (1) Must be called with an elevated refcount on the page. This is a
1606 * fundamentnal difference from isolate_lru_pages (which is called
1607 * without a stable reference).
1608 * (2) the lru_lock must not be held.
1609 * (3) interrupts must be enabled.
1611 int isolate_lru_page(struct page
*page
)
1615 VM_BUG_ON_PAGE(!page_count(page
), page
);
1616 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1618 if (PageLRU(page
)) {
1619 struct zone
*zone
= page_zone(page
);
1620 struct lruvec
*lruvec
;
1622 spin_lock_irq(zone_lru_lock(zone
));
1623 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1624 if (PageLRU(page
)) {
1625 int lru
= page_lru(page
);
1628 del_page_from_lru_list(page
, lruvec
, lru
);
1631 spin_unlock_irq(zone_lru_lock(zone
));
1637 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1638 * then get resheduled. When there are massive number of tasks doing page
1639 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1640 * the LRU list will go small and be scanned faster than necessary, leading to
1641 * unnecessary swapping, thrashing and OOM.
1643 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1644 struct scan_control
*sc
)
1646 unsigned long inactive
, isolated
;
1648 if (current_is_kswapd())
1651 if (!sane_reclaim(sc
))
1655 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1656 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1658 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1659 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1663 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1664 * won't get blocked by normal direct-reclaimers, forming a circular
1667 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1670 return isolated
> inactive
;
1673 static noinline_for_stack
void
1674 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1676 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1677 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1678 LIST_HEAD(pages_to_free
);
1681 * Put back any unfreeable pages.
1683 while (!list_empty(page_list
)) {
1684 struct page
*page
= lru_to_page(page_list
);
1687 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1688 list_del(&page
->lru
);
1689 if (unlikely(!page_evictable(page
))) {
1690 spin_unlock_irq(&pgdat
->lru_lock
);
1691 putback_lru_page(page
);
1692 spin_lock_irq(&pgdat
->lru_lock
);
1696 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1699 lru
= page_lru(page
);
1700 add_page_to_lru_list(page
, lruvec
, lru
);
1702 if (is_active_lru(lru
)) {
1703 int file
= is_file_lru(lru
);
1704 int numpages
= hpage_nr_pages(page
);
1705 reclaim_stat
->recent_rotated
[file
] += numpages
;
1707 if (put_page_testzero(page
)) {
1708 __ClearPageLRU(page
);
1709 __ClearPageActive(page
);
1710 del_page_from_lru_list(page
, lruvec
, lru
);
1712 if (unlikely(PageCompound(page
))) {
1713 spin_unlock_irq(&pgdat
->lru_lock
);
1714 mem_cgroup_uncharge(page
);
1715 (*get_compound_page_dtor(page
))(page
);
1716 spin_lock_irq(&pgdat
->lru_lock
);
1718 list_add(&page
->lru
, &pages_to_free
);
1723 * To save our caller's stack, now use input list for pages to free.
1725 list_splice(&pages_to_free
, page_list
);
1729 * If a kernel thread (such as nfsd for loop-back mounts) services
1730 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1731 * In that case we should only throttle if the backing device it is
1732 * writing to is congested. In other cases it is safe to throttle.
1734 static int current_may_throttle(void)
1736 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1737 current
->backing_dev_info
== NULL
||
1738 bdi_write_congested(current
->backing_dev_info
);
1742 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1743 * of reclaimed pages
1745 static noinline_for_stack
unsigned long
1746 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1747 struct scan_control
*sc
, enum lru_list lru
)
1749 LIST_HEAD(page_list
);
1750 unsigned long nr_scanned
;
1751 unsigned long nr_reclaimed
= 0;
1752 unsigned long nr_taken
;
1753 struct reclaim_stat stat
= {};
1754 isolate_mode_t isolate_mode
= 0;
1755 int file
= is_file_lru(lru
);
1756 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1757 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1758 bool stalled
= false;
1760 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1764 /* wait a bit for the reclaimer. */
1768 /* We are about to die and free our memory. Return now. */
1769 if (fatal_signal_pending(current
))
1770 return SWAP_CLUSTER_MAX
;
1776 isolate_mode
|= ISOLATE_UNMAPPED
;
1778 spin_lock_irq(&pgdat
->lru_lock
);
1780 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1781 &nr_scanned
, sc
, isolate_mode
, lru
);
1783 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1784 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1786 if (current_is_kswapd()) {
1787 if (global_reclaim(sc
))
1788 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1789 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1792 if (global_reclaim(sc
))
1793 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1794 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1797 spin_unlock_irq(&pgdat
->lru_lock
);
1802 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1805 spin_lock_irq(&pgdat
->lru_lock
);
1807 if (current_is_kswapd()) {
1808 if (global_reclaim(sc
))
1809 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1810 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1813 if (global_reclaim(sc
))
1814 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1815 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1819 putback_inactive_pages(lruvec
, &page_list
);
1821 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1823 spin_unlock_irq(&pgdat
->lru_lock
);
1825 mem_cgroup_uncharge_list(&page_list
);
1826 free_hot_cold_page_list(&page_list
, true);
1829 * If reclaim is isolating dirty pages under writeback, it implies
1830 * that the long-lived page allocation rate is exceeding the page
1831 * laundering rate. Either the global limits are not being effective
1832 * at throttling processes due to the page distribution throughout
1833 * zones or there is heavy usage of a slow backing device. The
1834 * only option is to throttle from reclaim context which is not ideal
1835 * as there is no guarantee the dirtying process is throttled in the
1836 * same way balance_dirty_pages() manages.
1838 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1839 * of pages under pages flagged for immediate reclaim and stall if any
1840 * are encountered in the nr_immediate check below.
1842 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1843 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1846 * Legacy memcg will stall in page writeback so avoid forcibly
1849 if (sane_reclaim(sc
)) {
1851 * Tag a zone as congested if all the dirty pages scanned were
1852 * backed by a congested BDI and wait_iff_congested will stall.
1854 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1855 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1858 * If dirty pages are scanned that are not queued for IO, it
1859 * implies that flushers are not doing their job. This can
1860 * happen when memory pressure pushes dirty pages to the end of
1861 * the LRU before the dirty limits are breached and the dirty
1862 * data has expired. It can also happen when the proportion of
1863 * dirty pages grows not through writes but through memory
1864 * pressure reclaiming all the clean cache. And in some cases,
1865 * the flushers simply cannot keep up with the allocation
1866 * rate. Nudge the flusher threads in case they are asleep, but
1867 * also allow kswapd to start writing pages during reclaim.
1869 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1870 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1871 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1875 * If kswapd scans pages marked marked for immediate
1876 * reclaim and under writeback (nr_immediate), it implies
1877 * that pages are cycling through the LRU faster than
1878 * they are written so also forcibly stall.
1880 if (stat
.nr_immediate
&& current_may_throttle())
1881 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1885 * Stall direct reclaim for IO completions if underlying BDIs or zone
1886 * is congested. Allow kswapd to continue until it starts encountering
1887 * unqueued dirty pages or cycling through the LRU too quickly.
1889 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1890 current_may_throttle())
1891 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1893 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1894 nr_scanned
, nr_reclaimed
,
1895 stat
.nr_dirty
, stat
.nr_writeback
,
1896 stat
.nr_congested
, stat
.nr_immediate
,
1897 stat
.nr_activate
, stat
.nr_ref_keep
,
1899 sc
->priority
, file
);
1900 return nr_reclaimed
;
1904 * This moves pages from the active list to the inactive list.
1906 * We move them the other way if the page is referenced by one or more
1907 * processes, from rmap.
1909 * If the pages are mostly unmapped, the processing is fast and it is
1910 * appropriate to hold zone_lru_lock across the whole operation. But if
1911 * the pages are mapped, the processing is slow (page_referenced()) so we
1912 * should drop zone_lru_lock around each page. It's impossible to balance
1913 * this, so instead we remove the pages from the LRU while processing them.
1914 * It is safe to rely on PG_active against the non-LRU pages in here because
1915 * nobody will play with that bit on a non-LRU page.
1917 * The downside is that we have to touch page->_refcount against each page.
1918 * But we had to alter page->flags anyway.
1920 * Returns the number of pages moved to the given lru.
1923 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1924 struct list_head
*list
,
1925 struct list_head
*pages_to_free
,
1928 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1933 while (!list_empty(list
)) {
1934 page
= lru_to_page(list
);
1935 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1937 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1940 nr_pages
= hpage_nr_pages(page
);
1941 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1942 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1944 if (put_page_testzero(page
)) {
1945 __ClearPageLRU(page
);
1946 __ClearPageActive(page
);
1947 del_page_from_lru_list(page
, lruvec
, lru
);
1949 if (unlikely(PageCompound(page
))) {
1950 spin_unlock_irq(&pgdat
->lru_lock
);
1951 mem_cgroup_uncharge(page
);
1952 (*get_compound_page_dtor(page
))(page
);
1953 spin_lock_irq(&pgdat
->lru_lock
);
1955 list_add(&page
->lru
, pages_to_free
);
1957 nr_moved
+= nr_pages
;
1961 if (!is_active_lru(lru
)) {
1962 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1963 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
1970 static void shrink_active_list(unsigned long nr_to_scan
,
1971 struct lruvec
*lruvec
,
1972 struct scan_control
*sc
,
1975 unsigned long nr_taken
;
1976 unsigned long nr_scanned
;
1977 unsigned long vm_flags
;
1978 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1979 LIST_HEAD(l_active
);
1980 LIST_HEAD(l_inactive
);
1982 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1983 unsigned nr_deactivate
, nr_activate
;
1984 unsigned nr_rotated
= 0;
1985 isolate_mode_t isolate_mode
= 0;
1986 int file
= is_file_lru(lru
);
1987 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1992 isolate_mode
|= ISOLATE_UNMAPPED
;
1994 spin_lock_irq(&pgdat
->lru_lock
);
1996 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1997 &nr_scanned
, sc
, isolate_mode
, lru
);
1999 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2000 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2002 __count_vm_events(PGREFILL
, nr_scanned
);
2003 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2005 spin_unlock_irq(&pgdat
->lru_lock
);
2007 while (!list_empty(&l_hold
)) {
2009 page
= lru_to_page(&l_hold
);
2010 list_del(&page
->lru
);
2012 if (unlikely(!page_evictable(page
))) {
2013 putback_lru_page(page
);
2017 if (unlikely(buffer_heads_over_limit
)) {
2018 if (page_has_private(page
) && trylock_page(page
)) {
2019 if (page_has_private(page
))
2020 try_to_release_page(page
, 0);
2025 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2027 nr_rotated
+= hpage_nr_pages(page
);
2029 * Identify referenced, file-backed active pages and
2030 * give them one more trip around the active list. So
2031 * that executable code get better chances to stay in
2032 * memory under moderate memory pressure. Anon pages
2033 * are not likely to be evicted by use-once streaming
2034 * IO, plus JVM can create lots of anon VM_EXEC pages,
2035 * so we ignore them here.
2037 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2038 list_add(&page
->lru
, &l_active
);
2043 ClearPageActive(page
); /* we are de-activating */
2044 list_add(&page
->lru
, &l_inactive
);
2048 * Move pages back to the lru list.
2050 spin_lock_irq(&pgdat
->lru_lock
);
2052 * Count referenced pages from currently used mappings as rotated,
2053 * even though only some of them are actually re-activated. This
2054 * helps balance scan pressure between file and anonymous pages in
2057 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2059 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2060 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2061 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2062 spin_unlock_irq(&pgdat
->lru_lock
);
2064 mem_cgroup_uncharge_list(&l_hold
);
2065 free_hot_cold_page_list(&l_hold
, true);
2066 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2067 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2071 * The inactive anon list should be small enough that the VM never has
2072 * to do too much work.
2074 * The inactive file list should be small enough to leave most memory
2075 * to the established workingset on the scan-resistant active list,
2076 * but large enough to avoid thrashing the aggregate readahead window.
2078 * Both inactive lists should also be large enough that each inactive
2079 * page has a chance to be referenced again before it is reclaimed.
2081 * If that fails and refaulting is observed, the inactive list grows.
2083 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2084 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2085 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2088 * memory ratio inactive
2089 * -------------------------------------
2098 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2099 struct mem_cgroup
*memcg
,
2100 struct scan_control
*sc
, bool actual_reclaim
)
2102 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2103 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2104 enum lru_list inactive_lru
= file
* LRU_FILE
;
2105 unsigned long inactive
, active
;
2106 unsigned long inactive_ratio
;
2107 unsigned long refaults
;
2111 * If we don't have swap space, anonymous page deactivation
2114 if (!file
&& !total_swap_pages
)
2117 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2118 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2121 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2123 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2126 * When refaults are being observed, it means a new workingset
2127 * is being established. Disable active list protection to get
2128 * rid of the stale workingset quickly.
2130 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2133 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2135 inactive_ratio
= int_sqrt(10 * gb
);
2141 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2142 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2143 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2144 inactive_ratio
, file
);
2146 return inactive
* inactive_ratio
< active
;
2149 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2150 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2151 struct scan_control
*sc
)
2153 if (is_active_lru(lru
)) {
2154 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2156 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2160 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2171 * Determine how aggressively the anon and file LRU lists should be
2172 * scanned. The relative value of each set of LRU lists is determined
2173 * by looking at the fraction of the pages scanned we did rotate back
2174 * onto the active list instead of evict.
2176 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2177 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2179 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2180 struct scan_control
*sc
, unsigned long *nr
,
2181 unsigned long *lru_pages
)
2183 int swappiness
= mem_cgroup_swappiness(memcg
);
2184 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2186 u64 denominator
= 0; /* gcc */
2187 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2188 unsigned long anon_prio
, file_prio
;
2189 enum scan_balance scan_balance
;
2190 unsigned long anon
, file
;
2191 unsigned long ap
, fp
;
2194 /* If we have no swap space, do not bother scanning anon pages. */
2195 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2196 scan_balance
= SCAN_FILE
;
2201 * Global reclaim will swap to prevent OOM even with no
2202 * swappiness, but memcg users want to use this knob to
2203 * disable swapping for individual groups completely when
2204 * using the memory controller's swap limit feature would be
2207 if (!global_reclaim(sc
) && !swappiness
) {
2208 scan_balance
= SCAN_FILE
;
2213 * Do not apply any pressure balancing cleverness when the
2214 * system is close to OOM, scan both anon and file equally
2215 * (unless the swappiness setting disagrees with swapping).
2217 if (!sc
->priority
&& swappiness
) {
2218 scan_balance
= SCAN_EQUAL
;
2223 * Prevent the reclaimer from falling into the cache trap: as
2224 * cache pages start out inactive, every cache fault will tip
2225 * the scan balance towards the file LRU. And as the file LRU
2226 * shrinks, so does the window for rotation from references.
2227 * This means we have a runaway feedback loop where a tiny
2228 * thrashing file LRU becomes infinitely more attractive than
2229 * anon pages. Try to detect this based on file LRU size.
2231 if (global_reclaim(sc
)) {
2232 unsigned long pgdatfile
;
2233 unsigned long pgdatfree
;
2235 unsigned long total_high_wmark
= 0;
2237 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2238 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2239 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2241 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2242 struct zone
*zone
= &pgdat
->node_zones
[z
];
2243 if (!managed_zone(zone
))
2246 total_high_wmark
+= high_wmark_pages(zone
);
2249 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2251 * Force SCAN_ANON if there are enough inactive
2252 * anonymous pages on the LRU in eligible zones.
2253 * Otherwise, the small LRU gets thrashed.
2255 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2256 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2258 scan_balance
= SCAN_ANON
;
2265 * If there is enough inactive page cache, i.e. if the size of the
2266 * inactive list is greater than that of the active list *and* the
2267 * inactive list actually has some pages to scan on this priority, we
2268 * do not reclaim anything from the anonymous working set right now.
2269 * Without the second condition we could end up never scanning an
2270 * lruvec even if it has plenty of old anonymous pages unless the
2271 * system is under heavy pressure.
2273 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2274 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2275 scan_balance
= SCAN_FILE
;
2279 scan_balance
= SCAN_FRACT
;
2282 * With swappiness at 100, anonymous and file have the same priority.
2283 * This scanning priority is essentially the inverse of IO cost.
2285 anon_prio
= swappiness
;
2286 file_prio
= 200 - anon_prio
;
2289 * OK, so we have swap space and a fair amount of page cache
2290 * pages. We use the recently rotated / recently scanned
2291 * ratios to determine how valuable each cache is.
2293 * Because workloads change over time (and to avoid overflow)
2294 * we keep these statistics as a floating average, which ends
2295 * up weighing recent references more than old ones.
2297 * anon in [0], file in [1]
2300 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2301 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2302 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2303 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2305 spin_lock_irq(&pgdat
->lru_lock
);
2306 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2307 reclaim_stat
->recent_scanned
[0] /= 2;
2308 reclaim_stat
->recent_rotated
[0] /= 2;
2311 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2312 reclaim_stat
->recent_scanned
[1] /= 2;
2313 reclaim_stat
->recent_rotated
[1] /= 2;
2317 * The amount of pressure on anon vs file pages is inversely
2318 * proportional to the fraction of recently scanned pages on
2319 * each list that were recently referenced and in active use.
2321 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2322 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2324 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2325 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2326 spin_unlock_irq(&pgdat
->lru_lock
);
2330 denominator
= ap
+ fp
+ 1;
2333 for_each_evictable_lru(lru
) {
2334 int file
= is_file_lru(lru
);
2338 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2339 scan
= size
>> sc
->priority
;
2341 * If the cgroup's already been deleted, make sure to
2342 * scrape out the remaining cache.
2344 if (!scan
&& !mem_cgroup_online(memcg
))
2345 scan
= min(size
, SWAP_CLUSTER_MAX
);
2347 switch (scan_balance
) {
2349 /* Scan lists relative to size */
2353 * Scan types proportional to swappiness and
2354 * their relative recent reclaim efficiency.
2356 scan
= div64_u64(scan
* fraction
[file
],
2361 /* Scan one type exclusively */
2362 if ((scan_balance
== SCAN_FILE
) != file
) {
2368 /* Look ma, no brain */
2378 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2380 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2381 struct scan_control
*sc
, unsigned long *lru_pages
)
2383 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2384 unsigned long nr
[NR_LRU_LISTS
];
2385 unsigned long targets
[NR_LRU_LISTS
];
2386 unsigned long nr_to_scan
;
2388 unsigned long nr_reclaimed
= 0;
2389 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2390 struct blk_plug plug
;
2393 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2395 /* Record the original scan target for proportional adjustments later */
2396 memcpy(targets
, nr
, sizeof(nr
));
2399 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2400 * event that can occur when there is little memory pressure e.g.
2401 * multiple streaming readers/writers. Hence, we do not abort scanning
2402 * when the requested number of pages are reclaimed when scanning at
2403 * DEF_PRIORITY on the assumption that the fact we are direct
2404 * reclaiming implies that kswapd is not keeping up and it is best to
2405 * do a batch of work at once. For memcg reclaim one check is made to
2406 * abort proportional reclaim if either the file or anon lru has already
2407 * dropped to zero at the first pass.
2409 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2410 sc
->priority
== DEF_PRIORITY
);
2412 blk_start_plug(&plug
);
2413 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2414 nr
[LRU_INACTIVE_FILE
]) {
2415 unsigned long nr_anon
, nr_file
, percentage
;
2416 unsigned long nr_scanned
;
2418 for_each_evictable_lru(lru
) {
2420 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2421 nr
[lru
] -= nr_to_scan
;
2423 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2430 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2434 * For kswapd and memcg, reclaim at least the number of pages
2435 * requested. Ensure that the anon and file LRUs are scanned
2436 * proportionally what was requested by get_scan_count(). We
2437 * stop reclaiming one LRU and reduce the amount scanning
2438 * proportional to the original scan target.
2440 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2441 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2444 * It's just vindictive to attack the larger once the smaller
2445 * has gone to zero. And given the way we stop scanning the
2446 * smaller below, this makes sure that we only make one nudge
2447 * towards proportionality once we've got nr_to_reclaim.
2449 if (!nr_file
|| !nr_anon
)
2452 if (nr_file
> nr_anon
) {
2453 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2454 targets
[LRU_ACTIVE_ANON
] + 1;
2456 percentage
= nr_anon
* 100 / scan_target
;
2458 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2459 targets
[LRU_ACTIVE_FILE
] + 1;
2461 percentage
= nr_file
* 100 / scan_target
;
2464 /* Stop scanning the smaller of the LRU */
2466 nr
[lru
+ LRU_ACTIVE
] = 0;
2469 * Recalculate the other LRU scan count based on its original
2470 * scan target and the percentage scanning already complete
2472 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2473 nr_scanned
= targets
[lru
] - nr
[lru
];
2474 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2475 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2478 nr_scanned
= targets
[lru
] - nr
[lru
];
2479 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2480 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2482 scan_adjusted
= true;
2484 blk_finish_plug(&plug
);
2485 sc
->nr_reclaimed
+= nr_reclaimed
;
2488 * Even if we did not try to evict anon pages at all, we want to
2489 * rebalance the anon lru active/inactive ratio.
2491 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2492 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2493 sc
, LRU_ACTIVE_ANON
);
2496 /* Use reclaim/compaction for costly allocs or under memory pressure */
2497 static bool in_reclaim_compaction(struct scan_control
*sc
)
2499 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2500 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2501 sc
->priority
< DEF_PRIORITY
- 2))
2508 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2509 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2510 * true if more pages should be reclaimed such that when the page allocator
2511 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2512 * It will give up earlier than that if there is difficulty reclaiming pages.
2514 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2515 unsigned long nr_reclaimed
,
2516 unsigned long nr_scanned
,
2517 struct scan_control
*sc
)
2519 unsigned long pages_for_compaction
;
2520 unsigned long inactive_lru_pages
;
2523 /* If not in reclaim/compaction mode, stop */
2524 if (!in_reclaim_compaction(sc
))
2527 /* Consider stopping depending on scan and reclaim activity */
2528 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2530 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2531 * full LRU list has been scanned and we are still failing
2532 * to reclaim pages. This full LRU scan is potentially
2533 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2535 if (!nr_reclaimed
&& !nr_scanned
)
2539 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2540 * fail without consequence, stop if we failed to reclaim
2541 * any pages from the last SWAP_CLUSTER_MAX number of
2542 * pages that were scanned. This will return to the
2543 * caller faster at the risk reclaim/compaction and
2544 * the resulting allocation attempt fails
2551 * If we have not reclaimed enough pages for compaction and the
2552 * inactive lists are large enough, continue reclaiming
2554 pages_for_compaction
= compact_gap(sc
->order
);
2555 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2556 if (get_nr_swap_pages() > 0)
2557 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2558 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2559 inactive_lru_pages
> pages_for_compaction
)
2562 /* If compaction would go ahead or the allocation would succeed, stop */
2563 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2564 struct zone
*zone
= &pgdat
->node_zones
[z
];
2565 if (!managed_zone(zone
))
2568 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2569 case COMPACT_SUCCESS
:
2570 case COMPACT_CONTINUE
:
2573 /* check next zone */
2580 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2582 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2583 unsigned long nr_reclaimed
, nr_scanned
;
2584 bool reclaimable
= false;
2587 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2588 struct mem_cgroup_reclaim_cookie reclaim
= {
2590 .priority
= sc
->priority
,
2592 unsigned long node_lru_pages
= 0;
2593 struct mem_cgroup
*memcg
;
2595 nr_reclaimed
= sc
->nr_reclaimed
;
2596 nr_scanned
= sc
->nr_scanned
;
2598 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2600 unsigned long lru_pages
;
2601 unsigned long reclaimed
;
2602 unsigned long scanned
;
2604 if (mem_cgroup_low(root
, memcg
)) {
2605 if (!sc
->memcg_low_reclaim
) {
2606 sc
->memcg_low_skipped
= 1;
2609 mem_cgroup_event(memcg
, MEMCG_LOW
);
2612 reclaimed
= sc
->nr_reclaimed
;
2613 scanned
= sc
->nr_scanned
;
2615 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2616 node_lru_pages
+= lru_pages
;
2619 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2620 memcg
, sc
->nr_scanned
- scanned
,
2623 /* Record the group's reclaim efficiency */
2624 vmpressure(sc
->gfp_mask
, memcg
, false,
2625 sc
->nr_scanned
- scanned
,
2626 sc
->nr_reclaimed
- reclaimed
);
2629 * Direct reclaim and kswapd have to scan all memory
2630 * cgroups to fulfill the overall scan target for the
2633 * Limit reclaim, on the other hand, only cares about
2634 * nr_to_reclaim pages to be reclaimed and it will
2635 * retry with decreasing priority if one round over the
2636 * whole hierarchy is not sufficient.
2638 if (!global_reclaim(sc
) &&
2639 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2640 mem_cgroup_iter_break(root
, memcg
);
2643 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2646 * Shrink the slab caches in the same proportion that
2647 * the eligible LRU pages were scanned.
2649 if (global_reclaim(sc
))
2650 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2651 sc
->nr_scanned
- nr_scanned
,
2654 if (reclaim_state
) {
2655 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2656 reclaim_state
->reclaimed_slab
= 0;
2659 /* Record the subtree's reclaim efficiency */
2660 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2661 sc
->nr_scanned
- nr_scanned
,
2662 sc
->nr_reclaimed
- nr_reclaimed
);
2664 if (sc
->nr_reclaimed
- nr_reclaimed
)
2667 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2668 sc
->nr_scanned
- nr_scanned
, sc
));
2671 * Kswapd gives up on balancing particular nodes after too
2672 * many failures to reclaim anything from them and goes to
2673 * sleep. On reclaim progress, reset the failure counter. A
2674 * successful direct reclaim run will revive a dormant kswapd.
2677 pgdat
->kswapd_failures
= 0;
2683 * Returns true if compaction should go ahead for a costly-order request, or
2684 * the allocation would already succeed without compaction. Return false if we
2685 * should reclaim first.
2687 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2689 unsigned long watermark
;
2690 enum compact_result suitable
;
2692 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2693 if (suitable
== COMPACT_SUCCESS
)
2694 /* Allocation should succeed already. Don't reclaim. */
2696 if (suitable
== COMPACT_SKIPPED
)
2697 /* Compaction cannot yet proceed. Do reclaim. */
2701 * Compaction is already possible, but it takes time to run and there
2702 * are potentially other callers using the pages just freed. So proceed
2703 * with reclaim to make a buffer of free pages available to give
2704 * compaction a reasonable chance of completing and allocating the page.
2705 * Note that we won't actually reclaim the whole buffer in one attempt
2706 * as the target watermark in should_continue_reclaim() is lower. But if
2707 * we are already above the high+gap watermark, don't reclaim at all.
2709 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2711 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2715 * This is the direct reclaim path, for page-allocating processes. We only
2716 * try to reclaim pages from zones which will satisfy the caller's allocation
2719 * If a zone is deemed to be full of pinned pages then just give it a light
2720 * scan then give up on it.
2722 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2726 unsigned long nr_soft_reclaimed
;
2727 unsigned long nr_soft_scanned
;
2729 pg_data_t
*last_pgdat
= NULL
;
2732 * If the number of buffer_heads in the machine exceeds the maximum
2733 * allowed level, force direct reclaim to scan the highmem zone as
2734 * highmem pages could be pinning lowmem pages storing buffer_heads
2736 orig_mask
= sc
->gfp_mask
;
2737 if (buffer_heads_over_limit
) {
2738 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2739 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2742 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2743 sc
->reclaim_idx
, sc
->nodemask
) {
2745 * Take care memory controller reclaiming has small influence
2748 if (global_reclaim(sc
)) {
2749 if (!cpuset_zone_allowed(zone
,
2750 GFP_KERNEL
| __GFP_HARDWALL
))
2754 * If we already have plenty of memory free for
2755 * compaction in this zone, don't free any more.
2756 * Even though compaction is invoked for any
2757 * non-zero order, only frequent costly order
2758 * reclamation is disruptive enough to become a
2759 * noticeable problem, like transparent huge
2762 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2763 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2764 compaction_ready(zone
, sc
)) {
2765 sc
->compaction_ready
= true;
2770 * Shrink each node in the zonelist once. If the
2771 * zonelist is ordered by zone (not the default) then a
2772 * node may be shrunk multiple times but in that case
2773 * the user prefers lower zones being preserved.
2775 if (zone
->zone_pgdat
== last_pgdat
)
2779 * This steals pages from memory cgroups over softlimit
2780 * and returns the number of reclaimed pages and
2781 * scanned pages. This works for global memory pressure
2782 * and balancing, not for a memcg's limit.
2784 nr_soft_scanned
= 0;
2785 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2786 sc
->order
, sc
->gfp_mask
,
2788 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2789 sc
->nr_scanned
+= nr_soft_scanned
;
2790 /* need some check for avoid more shrink_zone() */
2793 /* See comment about same check for global reclaim above */
2794 if (zone
->zone_pgdat
== last_pgdat
)
2796 last_pgdat
= zone
->zone_pgdat
;
2797 shrink_node(zone
->zone_pgdat
, sc
);
2801 * Restore to original mask to avoid the impact on the caller if we
2802 * promoted it to __GFP_HIGHMEM.
2804 sc
->gfp_mask
= orig_mask
;
2807 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2809 struct mem_cgroup
*memcg
;
2811 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2813 unsigned long refaults
;
2814 struct lruvec
*lruvec
;
2817 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2819 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2821 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2822 lruvec
->refaults
= refaults
;
2823 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2827 * This is the main entry point to direct page reclaim.
2829 * If a full scan of the inactive list fails to free enough memory then we
2830 * are "out of memory" and something needs to be killed.
2832 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2833 * high - the zone may be full of dirty or under-writeback pages, which this
2834 * caller can't do much about. We kick the writeback threads and take explicit
2835 * naps in the hope that some of these pages can be written. But if the
2836 * allocating task holds filesystem locks which prevent writeout this might not
2837 * work, and the allocation attempt will fail.
2839 * returns: 0, if no pages reclaimed
2840 * else, the number of pages reclaimed
2842 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2843 struct scan_control
*sc
)
2845 int initial_priority
= sc
->priority
;
2846 pg_data_t
*last_pgdat
;
2850 delayacct_freepages_start();
2852 if (global_reclaim(sc
))
2853 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2856 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2859 shrink_zones(zonelist
, sc
);
2861 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2864 if (sc
->compaction_ready
)
2868 * If we're getting trouble reclaiming, start doing
2869 * writepage even in laptop mode.
2871 if (sc
->priority
< DEF_PRIORITY
- 2)
2872 sc
->may_writepage
= 1;
2873 } while (--sc
->priority
>= 0);
2876 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2878 if (zone
->zone_pgdat
== last_pgdat
)
2880 last_pgdat
= zone
->zone_pgdat
;
2881 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2884 delayacct_freepages_end();
2886 if (sc
->nr_reclaimed
)
2887 return sc
->nr_reclaimed
;
2889 /* Aborted reclaim to try compaction? don't OOM, then */
2890 if (sc
->compaction_ready
)
2893 /* Untapped cgroup reserves? Don't OOM, retry. */
2894 if (sc
->memcg_low_skipped
) {
2895 sc
->priority
= initial_priority
;
2896 sc
->memcg_low_reclaim
= 1;
2897 sc
->memcg_low_skipped
= 0;
2904 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2907 unsigned long pfmemalloc_reserve
= 0;
2908 unsigned long free_pages
= 0;
2912 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2915 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2916 zone
= &pgdat
->node_zones
[i
];
2917 if (!managed_zone(zone
))
2920 if (!zone_reclaimable_pages(zone
))
2923 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2924 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2927 /* If there are no reserves (unexpected config) then do not throttle */
2928 if (!pfmemalloc_reserve
)
2931 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2933 /* kswapd must be awake if processes are being throttled */
2934 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2935 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2936 (enum zone_type
)ZONE_NORMAL
);
2937 wake_up_interruptible(&pgdat
->kswapd_wait
);
2944 * Throttle direct reclaimers if backing storage is backed by the network
2945 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2946 * depleted. kswapd will continue to make progress and wake the processes
2947 * when the low watermark is reached.
2949 * Returns true if a fatal signal was delivered during throttling. If this
2950 * happens, the page allocator should not consider triggering the OOM killer.
2952 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2953 nodemask_t
*nodemask
)
2957 pg_data_t
*pgdat
= NULL
;
2960 * Kernel threads should not be throttled as they may be indirectly
2961 * responsible for cleaning pages necessary for reclaim to make forward
2962 * progress. kjournald for example may enter direct reclaim while
2963 * committing a transaction where throttling it could forcing other
2964 * processes to block on log_wait_commit().
2966 if (current
->flags
& PF_KTHREAD
)
2970 * If a fatal signal is pending, this process should not throttle.
2971 * It should return quickly so it can exit and free its memory
2973 if (fatal_signal_pending(current
))
2977 * Check if the pfmemalloc reserves are ok by finding the first node
2978 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2979 * GFP_KERNEL will be required for allocating network buffers when
2980 * swapping over the network so ZONE_HIGHMEM is unusable.
2982 * Throttling is based on the first usable node and throttled processes
2983 * wait on a queue until kswapd makes progress and wakes them. There
2984 * is an affinity then between processes waking up and where reclaim
2985 * progress has been made assuming the process wakes on the same node.
2986 * More importantly, processes running on remote nodes will not compete
2987 * for remote pfmemalloc reserves and processes on different nodes
2988 * should make reasonable progress.
2990 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2991 gfp_zone(gfp_mask
), nodemask
) {
2992 if (zone_idx(zone
) > ZONE_NORMAL
)
2995 /* Throttle based on the first usable node */
2996 pgdat
= zone
->zone_pgdat
;
2997 if (allow_direct_reclaim(pgdat
))
3002 /* If no zone was usable by the allocation flags then do not throttle */
3006 /* Account for the throttling */
3007 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3010 * If the caller cannot enter the filesystem, it's possible that it
3011 * is due to the caller holding an FS lock or performing a journal
3012 * transaction in the case of a filesystem like ext[3|4]. In this case,
3013 * it is not safe to block on pfmemalloc_wait as kswapd could be
3014 * blocked waiting on the same lock. Instead, throttle for up to a
3015 * second before continuing.
3017 if (!(gfp_mask
& __GFP_FS
)) {
3018 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3019 allow_direct_reclaim(pgdat
), HZ
);
3024 /* Throttle until kswapd wakes the process */
3025 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3026 allow_direct_reclaim(pgdat
));
3029 if (fatal_signal_pending(current
))
3036 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3037 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3039 unsigned long nr_reclaimed
;
3040 struct scan_control sc
= {
3041 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3042 .gfp_mask
= current_gfp_context(gfp_mask
),
3043 .reclaim_idx
= gfp_zone(gfp_mask
),
3045 .nodemask
= nodemask
,
3046 .priority
= DEF_PRIORITY
,
3047 .may_writepage
= !laptop_mode
,
3053 * Do not enter reclaim if fatal signal was delivered while throttled.
3054 * 1 is returned so that the page allocator does not OOM kill at this
3057 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3060 trace_mm_vmscan_direct_reclaim_begin(order
,
3065 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3067 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3069 return nr_reclaimed
;
3074 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3075 gfp_t gfp_mask
, bool noswap
,
3077 unsigned long *nr_scanned
)
3079 struct scan_control sc
= {
3080 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3081 .target_mem_cgroup
= memcg
,
3082 .may_writepage
= !laptop_mode
,
3084 .reclaim_idx
= MAX_NR_ZONES
- 1,
3085 .may_swap
= !noswap
,
3087 unsigned long lru_pages
;
3089 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3090 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3092 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3098 * NOTE: Although we can get the priority field, using it
3099 * here is not a good idea, since it limits the pages we can scan.
3100 * if we don't reclaim here, the shrink_node from balance_pgdat
3101 * will pick up pages from other mem cgroup's as well. We hack
3102 * the priority and make it zero.
3104 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3106 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3108 *nr_scanned
= sc
.nr_scanned
;
3109 return sc
.nr_reclaimed
;
3112 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3113 unsigned long nr_pages
,
3117 struct zonelist
*zonelist
;
3118 unsigned long nr_reclaimed
;
3120 unsigned int noreclaim_flag
;
3121 struct scan_control sc
= {
3122 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3123 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3124 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3125 .reclaim_idx
= MAX_NR_ZONES
- 1,
3126 .target_mem_cgroup
= memcg
,
3127 .priority
= DEF_PRIORITY
,
3128 .may_writepage
= !laptop_mode
,
3130 .may_swap
= may_swap
,
3134 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3135 * take care of from where we get pages. So the node where we start the
3136 * scan does not need to be the current node.
3138 nid
= mem_cgroup_select_victim_node(memcg
);
3140 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3142 trace_mm_vmscan_memcg_reclaim_begin(0,
3147 noreclaim_flag
= memalloc_noreclaim_save();
3148 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3149 memalloc_noreclaim_restore(noreclaim_flag
);
3151 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3153 return nr_reclaimed
;
3157 static void age_active_anon(struct pglist_data
*pgdat
,
3158 struct scan_control
*sc
)
3160 struct mem_cgroup
*memcg
;
3162 if (!total_swap_pages
)
3165 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3167 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3169 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3170 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3171 sc
, LRU_ACTIVE_ANON
);
3173 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3178 * Returns true if there is an eligible zone balanced for the request order
3181 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3184 unsigned long mark
= -1;
3187 for (i
= 0; i
<= classzone_idx
; i
++) {
3188 zone
= pgdat
->node_zones
+ i
;
3190 if (!managed_zone(zone
))
3193 mark
= high_wmark_pages(zone
);
3194 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3199 * If a node has no populated zone within classzone_idx, it does not
3200 * need balancing by definition. This can happen if a zone-restricted
3201 * allocation tries to wake a remote kswapd.
3209 /* Clear pgdat state for congested, dirty or under writeback. */
3210 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3212 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3213 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3214 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3218 * Prepare kswapd for sleeping. This verifies that there are no processes
3219 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3221 * Returns true if kswapd is ready to sleep
3223 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3226 * The throttled processes are normally woken up in balance_pgdat() as
3227 * soon as allow_direct_reclaim() is true. But there is a potential
3228 * race between when kswapd checks the watermarks and a process gets
3229 * throttled. There is also a potential race if processes get
3230 * throttled, kswapd wakes, a large process exits thereby balancing the
3231 * zones, which causes kswapd to exit balance_pgdat() before reaching
3232 * the wake up checks. If kswapd is going to sleep, no process should
3233 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3234 * the wake up is premature, processes will wake kswapd and get
3235 * throttled again. The difference from wake ups in balance_pgdat() is
3236 * that here we are under prepare_to_wait().
3238 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3239 wake_up_all(&pgdat
->pfmemalloc_wait
);
3241 /* Hopeless node, leave it to direct reclaim */
3242 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3245 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3246 clear_pgdat_congested(pgdat
);
3254 * kswapd shrinks a node of pages that are at or below the highest usable
3255 * zone that is currently unbalanced.
3257 * Returns true if kswapd scanned at least the requested number of pages to
3258 * reclaim or if the lack of progress was due to pages under writeback.
3259 * This is used to determine if the scanning priority needs to be raised.
3261 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3262 struct scan_control
*sc
)
3267 /* Reclaim a number of pages proportional to the number of zones */
3268 sc
->nr_to_reclaim
= 0;
3269 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3270 zone
= pgdat
->node_zones
+ z
;
3271 if (!managed_zone(zone
))
3274 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3278 * Historically care was taken to put equal pressure on all zones but
3279 * now pressure is applied based on node LRU order.
3281 shrink_node(pgdat
, sc
);
3284 * Fragmentation may mean that the system cannot be rebalanced for
3285 * high-order allocations. If twice the allocation size has been
3286 * reclaimed then recheck watermarks only at order-0 to prevent
3287 * excessive reclaim. Assume that a process requested a high-order
3288 * can direct reclaim/compact.
3290 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3293 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3297 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3298 * that are eligible for use by the caller until at least one zone is
3301 * Returns the order kswapd finished reclaiming at.
3303 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3304 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3305 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3306 * or lower is eligible for reclaim until at least one usable zone is
3309 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3312 unsigned long nr_soft_reclaimed
;
3313 unsigned long nr_soft_scanned
;
3315 struct scan_control sc
= {
3316 .gfp_mask
= GFP_KERNEL
,
3318 .priority
= DEF_PRIORITY
,
3319 .may_writepage
= !laptop_mode
,
3323 count_vm_event(PAGEOUTRUN
);
3326 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3327 bool raise_priority
= true;
3329 sc
.reclaim_idx
= classzone_idx
;
3332 * If the number of buffer_heads exceeds the maximum allowed
3333 * then consider reclaiming from all zones. This has a dual
3334 * purpose -- on 64-bit systems it is expected that
3335 * buffer_heads are stripped during active rotation. On 32-bit
3336 * systems, highmem pages can pin lowmem memory and shrinking
3337 * buffers can relieve lowmem pressure. Reclaim may still not
3338 * go ahead if all eligible zones for the original allocation
3339 * request are balanced to avoid excessive reclaim from kswapd.
3341 if (buffer_heads_over_limit
) {
3342 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3343 zone
= pgdat
->node_zones
+ i
;
3344 if (!managed_zone(zone
))
3353 * Only reclaim if there are no eligible zones. Note that
3354 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3357 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3361 * Do some background aging of the anon list, to give
3362 * pages a chance to be referenced before reclaiming. All
3363 * pages are rotated regardless of classzone as this is
3364 * about consistent aging.
3366 age_active_anon(pgdat
, &sc
);
3369 * If we're getting trouble reclaiming, start doing writepage
3370 * even in laptop mode.
3372 if (sc
.priority
< DEF_PRIORITY
- 2)
3373 sc
.may_writepage
= 1;
3375 /* Call soft limit reclaim before calling shrink_node. */
3377 nr_soft_scanned
= 0;
3378 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3379 sc
.gfp_mask
, &nr_soft_scanned
);
3380 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3383 * There should be no need to raise the scanning priority if
3384 * enough pages are already being scanned that that high
3385 * watermark would be met at 100% efficiency.
3387 if (kswapd_shrink_node(pgdat
, &sc
))
3388 raise_priority
= false;
3391 * If the low watermark is met there is no need for processes
3392 * to be throttled on pfmemalloc_wait as they should not be
3393 * able to safely make forward progress. Wake them
3395 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3396 allow_direct_reclaim(pgdat
))
3397 wake_up_all(&pgdat
->pfmemalloc_wait
);
3399 /* Check if kswapd should be suspending */
3400 if (try_to_freeze() || kthread_should_stop())
3404 * Raise priority if scanning rate is too low or there was no
3405 * progress in reclaiming pages
3407 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3408 if (raise_priority
|| !nr_reclaimed
)
3410 } while (sc
.priority
>= 1);
3412 if (!sc
.nr_reclaimed
)
3413 pgdat
->kswapd_failures
++;
3416 snapshot_refaults(NULL
, pgdat
);
3418 * Return the order kswapd stopped reclaiming at as
3419 * prepare_kswapd_sleep() takes it into account. If another caller
3420 * entered the allocator slow path while kswapd was awake, order will
3421 * remain at the higher level.
3427 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3428 * allocation request woke kswapd for. When kswapd has not woken recently,
3429 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3430 * given classzone and returns it or the highest classzone index kswapd
3431 * was recently woke for.
3433 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3434 enum zone_type classzone_idx
)
3436 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3437 return classzone_idx
;
3439 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3442 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3443 unsigned int classzone_idx
)
3448 if (freezing(current
) || kthread_should_stop())
3451 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3454 * Try to sleep for a short interval. Note that kcompactd will only be
3455 * woken if it is possible to sleep for a short interval. This is
3456 * deliberate on the assumption that if reclaim cannot keep an
3457 * eligible zone balanced that it's also unlikely that compaction will
3460 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3462 * Compaction records what page blocks it recently failed to
3463 * isolate pages from and skips them in the future scanning.
3464 * When kswapd is going to sleep, it is reasonable to assume
3465 * that pages and compaction may succeed so reset the cache.
3467 reset_isolation_suitable(pgdat
);
3470 * We have freed the memory, now we should compact it to make
3471 * allocation of the requested order possible.
3473 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3475 remaining
= schedule_timeout(HZ
/10);
3478 * If woken prematurely then reset kswapd_classzone_idx and
3479 * order. The values will either be from a wakeup request or
3480 * the previous request that slept prematurely.
3483 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3484 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3487 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3488 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3492 * After a short sleep, check if it was a premature sleep. If not, then
3493 * go fully to sleep until explicitly woken up.
3496 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3497 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3500 * vmstat counters are not perfectly accurate and the estimated
3501 * value for counters such as NR_FREE_PAGES can deviate from the
3502 * true value by nr_online_cpus * threshold. To avoid the zone
3503 * watermarks being breached while under pressure, we reduce the
3504 * per-cpu vmstat threshold while kswapd is awake and restore
3505 * them before going back to sleep.
3507 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3509 if (!kthread_should_stop())
3512 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3515 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3517 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3519 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3523 * The background pageout daemon, started as a kernel thread
3524 * from the init process.
3526 * This basically trickles out pages so that we have _some_
3527 * free memory available even if there is no other activity
3528 * that frees anything up. This is needed for things like routing
3529 * etc, where we otherwise might have all activity going on in
3530 * asynchronous contexts that cannot page things out.
3532 * If there are applications that are active memory-allocators
3533 * (most normal use), this basically shouldn't matter.
3535 static int kswapd(void *p
)
3537 unsigned int alloc_order
, reclaim_order
;
3538 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3539 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3540 struct task_struct
*tsk
= current
;
3542 struct reclaim_state reclaim_state
= {
3543 .reclaimed_slab
= 0,
3545 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3547 if (!cpumask_empty(cpumask
))
3548 set_cpus_allowed_ptr(tsk
, cpumask
);
3549 current
->reclaim_state
= &reclaim_state
;
3552 * Tell the memory management that we're a "memory allocator",
3553 * and that if we need more memory we should get access to it
3554 * regardless (see "__alloc_pages()"). "kswapd" should
3555 * never get caught in the normal page freeing logic.
3557 * (Kswapd normally doesn't need memory anyway, but sometimes
3558 * you need a small amount of memory in order to be able to
3559 * page out something else, and this flag essentially protects
3560 * us from recursively trying to free more memory as we're
3561 * trying to free the first piece of memory in the first place).
3563 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3566 pgdat
->kswapd_order
= 0;
3567 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3571 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3572 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3575 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3578 /* Read the new order and classzone_idx */
3579 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3580 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3581 pgdat
->kswapd_order
= 0;
3582 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3584 ret
= try_to_freeze();
3585 if (kthread_should_stop())
3589 * We can speed up thawing tasks if we don't call balance_pgdat
3590 * after returning from the refrigerator
3596 * Reclaim begins at the requested order but if a high-order
3597 * reclaim fails then kswapd falls back to reclaiming for
3598 * order-0. If that happens, kswapd will consider sleeping
3599 * for the order it finished reclaiming at (reclaim_order)
3600 * but kcompactd is woken to compact for the original
3601 * request (alloc_order).
3603 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3605 fs_reclaim_acquire(GFP_KERNEL
);
3606 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3607 fs_reclaim_release(GFP_KERNEL
);
3608 if (reclaim_order
< alloc_order
)
3609 goto kswapd_try_sleep
;
3612 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3613 current
->reclaim_state
= NULL
;
3619 * A zone is low on free memory, so wake its kswapd task to service it.
3621 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3625 if (!managed_zone(zone
))
3628 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3630 pgdat
= zone
->zone_pgdat
;
3631 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3633 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3634 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3637 /* Hopeless node, leave it to direct reclaim */
3638 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3641 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3644 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3645 wake_up_interruptible(&pgdat
->kswapd_wait
);
3648 #ifdef CONFIG_HIBERNATION
3650 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3653 * Rather than trying to age LRUs the aim is to preserve the overall
3654 * LRU order by reclaiming preferentially
3655 * inactive > active > active referenced > active mapped
3657 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3659 struct reclaim_state reclaim_state
;
3660 struct scan_control sc
= {
3661 .nr_to_reclaim
= nr_to_reclaim
,
3662 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3663 .reclaim_idx
= MAX_NR_ZONES
- 1,
3664 .priority
= DEF_PRIORITY
,
3668 .hibernation_mode
= 1,
3670 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3671 struct task_struct
*p
= current
;
3672 unsigned long nr_reclaimed
;
3673 unsigned int noreclaim_flag
;
3675 noreclaim_flag
= memalloc_noreclaim_save();
3676 fs_reclaim_acquire(sc
.gfp_mask
);
3677 reclaim_state
.reclaimed_slab
= 0;
3678 p
->reclaim_state
= &reclaim_state
;
3680 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3682 p
->reclaim_state
= NULL
;
3683 fs_reclaim_release(sc
.gfp_mask
);
3684 memalloc_noreclaim_restore(noreclaim_flag
);
3686 return nr_reclaimed
;
3688 #endif /* CONFIG_HIBERNATION */
3690 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3691 not required for correctness. So if the last cpu in a node goes
3692 away, we get changed to run anywhere: as the first one comes back,
3693 restore their cpu bindings. */
3694 static int kswapd_cpu_online(unsigned int cpu
)
3698 for_each_node_state(nid
, N_MEMORY
) {
3699 pg_data_t
*pgdat
= NODE_DATA(nid
);
3700 const struct cpumask
*mask
;
3702 mask
= cpumask_of_node(pgdat
->node_id
);
3704 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3705 /* One of our CPUs online: restore mask */
3706 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3712 * This kswapd start function will be called by init and node-hot-add.
3713 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3715 int kswapd_run(int nid
)
3717 pg_data_t
*pgdat
= NODE_DATA(nid
);
3723 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3724 if (IS_ERR(pgdat
->kswapd
)) {
3725 /* failure at boot is fatal */
3726 BUG_ON(system_state
< SYSTEM_RUNNING
);
3727 pr_err("Failed to start kswapd on node %d\n", nid
);
3728 ret
= PTR_ERR(pgdat
->kswapd
);
3729 pgdat
->kswapd
= NULL
;
3735 * Called by memory hotplug when all memory in a node is offlined. Caller must
3736 * hold mem_hotplug_begin/end().
3738 void kswapd_stop(int nid
)
3740 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3743 kthread_stop(kswapd
);
3744 NODE_DATA(nid
)->kswapd
= NULL
;
3748 static int __init
kswapd_init(void)
3753 for_each_node_state(nid
, N_MEMORY
)
3755 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3756 "mm/vmscan:online", kswapd_cpu_online
,
3762 module_init(kswapd_init
)
3768 * If non-zero call node_reclaim when the number of free pages falls below
3771 int node_reclaim_mode __read_mostly
;
3773 #define RECLAIM_OFF 0
3774 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3775 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3776 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3779 * Priority for NODE_RECLAIM. This determines the fraction of pages
3780 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3783 #define NODE_RECLAIM_PRIORITY 4
3786 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3789 int sysctl_min_unmapped_ratio
= 1;
3792 * If the number of slab pages in a zone grows beyond this percentage then
3793 * slab reclaim needs to occur.
3795 int sysctl_min_slab_ratio
= 5;
3797 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3799 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3800 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3801 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3804 * It's possible for there to be more file mapped pages than
3805 * accounted for by the pages on the file LRU lists because
3806 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3808 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3811 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3812 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3814 unsigned long nr_pagecache_reclaimable
;
3815 unsigned long delta
= 0;
3818 * If RECLAIM_UNMAP is set, then all file pages are considered
3819 * potentially reclaimable. Otherwise, we have to worry about
3820 * pages like swapcache and node_unmapped_file_pages() provides
3823 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3824 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3826 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3828 /* If we can't clean pages, remove dirty pages from consideration */
3829 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3830 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3832 /* Watch for any possible underflows due to delta */
3833 if (unlikely(delta
> nr_pagecache_reclaimable
))
3834 delta
= nr_pagecache_reclaimable
;
3836 return nr_pagecache_reclaimable
- delta
;
3840 * Try to free up some pages from this node through reclaim.
3842 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3844 /* Minimum pages needed in order to stay on node */
3845 const unsigned long nr_pages
= 1 << order
;
3846 struct task_struct
*p
= current
;
3847 struct reclaim_state reclaim_state
;
3848 unsigned int noreclaim_flag
;
3849 struct scan_control sc
= {
3850 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3851 .gfp_mask
= current_gfp_context(gfp_mask
),
3853 .priority
= NODE_RECLAIM_PRIORITY
,
3854 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3855 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3857 .reclaim_idx
= gfp_zone(gfp_mask
),
3862 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3863 * and we also need to be able to write out pages for RECLAIM_WRITE
3864 * and RECLAIM_UNMAP.
3866 noreclaim_flag
= memalloc_noreclaim_save();
3867 p
->flags
|= PF_SWAPWRITE
;
3868 fs_reclaim_acquire(sc
.gfp_mask
);
3869 reclaim_state
.reclaimed_slab
= 0;
3870 p
->reclaim_state
= &reclaim_state
;
3872 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3874 * Free memory by calling shrink zone with increasing
3875 * priorities until we have enough memory freed.
3878 shrink_node(pgdat
, &sc
);
3879 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3882 p
->reclaim_state
= NULL
;
3883 fs_reclaim_release(gfp_mask
);
3884 current
->flags
&= ~PF_SWAPWRITE
;
3885 memalloc_noreclaim_restore(noreclaim_flag
);
3886 return sc
.nr_reclaimed
>= nr_pages
;
3889 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3894 * Node reclaim reclaims unmapped file backed pages and
3895 * slab pages if we are over the defined limits.
3897 * A small portion of unmapped file backed pages is needed for
3898 * file I/O otherwise pages read by file I/O will be immediately
3899 * thrown out if the node is overallocated. So we do not reclaim
3900 * if less than a specified percentage of the node is used by
3901 * unmapped file backed pages.
3903 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3904 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3905 return NODE_RECLAIM_FULL
;
3908 * Do not scan if the allocation should not be delayed.
3910 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3911 return NODE_RECLAIM_NOSCAN
;
3914 * Only run node reclaim on the local node or on nodes that do not
3915 * have associated processors. This will favor the local processor
3916 * over remote processors and spread off node memory allocations
3917 * as wide as possible.
3919 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3920 return NODE_RECLAIM_NOSCAN
;
3922 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3923 return NODE_RECLAIM_NOSCAN
;
3925 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3926 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3929 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3936 * page_evictable - test whether a page is evictable
3937 * @page: the page to test
3939 * Test whether page is evictable--i.e., should be placed on active/inactive
3940 * lists vs unevictable list.
3942 * Reasons page might not be evictable:
3943 * (1) page's mapping marked unevictable
3944 * (2) page is part of an mlocked VMA
3947 int page_evictable(struct page
*page
)
3949 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3954 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3955 * @pages: array of pages to check
3956 * @nr_pages: number of pages to check
3958 * Checks pages for evictability and moves them to the appropriate lru list.
3960 * This function is only used for SysV IPC SHM_UNLOCK.
3962 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3964 struct lruvec
*lruvec
;
3965 struct pglist_data
*pgdat
= NULL
;
3970 for (i
= 0; i
< nr_pages
; i
++) {
3971 struct page
*page
= pages
[i
];
3972 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3975 if (pagepgdat
!= pgdat
) {
3977 spin_unlock_irq(&pgdat
->lru_lock
);
3979 spin_lock_irq(&pgdat
->lru_lock
);
3981 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
3983 if (!PageLRU(page
) || !PageUnevictable(page
))
3986 if (page_evictable(page
)) {
3987 enum lru_list lru
= page_lru_base_type(page
);
3989 VM_BUG_ON_PAGE(PageActive(page
), page
);
3990 ClearPageUnevictable(page
);
3991 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3992 add_page_to_lru_list(page
, lruvec
, lru
);
3998 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3999 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4000 spin_unlock_irq(&pgdat
->lru_lock
);
4003 #endif /* CONFIG_SHMEM */