1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim
;
68 /* This context's GFP mask */
71 /* Allocation order */
75 * Nodemask of nodes allowed by the caller. If NULL, all nodes
81 * The memory cgroup that hit its limit and as a result is the
82 * primary target of this reclaim invocation.
84 struct mem_cgroup
*target_mem_cgroup
;
86 /* Scan (total_size >> priority) pages at once */
89 /* The highest zone to isolate pages for reclaim from */
90 enum zone_type reclaim_idx
;
92 /* Writepage batching in laptop mode; RECLAIM_WRITE */
93 unsigned int may_writepage
:1;
95 /* Can mapped pages be reclaimed? */
96 unsigned int may_unmap
:1;
98 /* Can pages be swapped as part of reclaim? */
99 unsigned int may_swap
:1;
102 * Cgroups are not reclaimed below their configured memory.low,
103 * unless we threaten to OOM. If any cgroups are skipped due to
104 * memory.low and nothing was reclaimed, go back for memory.low.
106 unsigned int memcg_low_reclaim
:1;
107 unsigned int memcg_low_skipped
:1;
109 unsigned int hibernation_mode
:1;
111 /* One of the zones is ready for compaction */
112 unsigned int compaction_ready
:1;
114 /* Incremented by the number of inactive pages that were scanned */
115 unsigned long nr_scanned
;
117 /* Number of pages freed so far during a call to shrink_zones() */
118 unsigned long nr_reclaimed
;
121 #ifdef ARCH_HAS_PREFETCH
122 #define prefetch_prev_lru_page(_page, _base, _field) \
124 if ((_page)->lru.prev != _base) { \
127 prev = lru_to_page(&(_page->lru)); \
128 prefetch(&prev->_field); \
132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
135 #ifdef ARCH_HAS_PREFETCHW
136 #define prefetchw_prev_lru_page(_page, _base, _field) \
138 if ((_page)->lru.prev != _base) { \
141 prev = lru_to_page(&(_page->lru)); \
142 prefetchw(&prev->_field); \
146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
150 * From 0 .. 100. Higher means more swappy.
152 int vm_swappiness
= 60;
154 * The total number of pages which are beyond the high watermark within all
157 unsigned long vm_total_pages
;
159 static LIST_HEAD(shrinker_list
);
160 static DECLARE_RWSEM(shrinker_rwsem
);
163 static bool global_reclaim(struct scan_control
*sc
)
165 return !sc
->target_mem_cgroup
;
169 * sane_reclaim - is the usual dirty throttling mechanism operational?
170 * @sc: scan_control in question
172 * The normal page dirty throttling mechanism in balance_dirty_pages() is
173 * completely broken with the legacy memcg and direct stalling in
174 * shrink_page_list() is used for throttling instead, which lacks all the
175 * niceties such as fairness, adaptive pausing, bandwidth proportional
176 * allocation and configurability.
178 * This function tests whether the vmscan currently in progress can assume
179 * that the normal dirty throttling mechanism is operational.
181 static bool sane_reclaim(struct scan_control
*sc
)
183 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
187 #ifdef CONFIG_CGROUP_WRITEBACK
188 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
194 static bool global_reclaim(struct scan_control
*sc
)
199 static bool sane_reclaim(struct scan_control
*sc
)
206 * This misses isolated pages which are not accounted for to save counters.
207 * As the data only determines if reclaim or compaction continues, it is
208 * not expected that isolated pages will be a dominating factor.
210 unsigned long zone_reclaimable_pages(struct zone
*zone
)
214 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
215 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
216 if (get_nr_swap_pages() > 0)
217 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
218 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
223 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
227 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
228 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
229 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
231 if (get_nr_swap_pages() > 0)
232 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
233 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
234 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
240 * lruvec_lru_size - Returns the number of pages on the given LRU list.
241 * @lruvec: lru vector
243 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
245 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
247 unsigned long lru_size
;
250 if (!mem_cgroup_disabled())
251 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
253 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
255 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
256 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
259 if (!managed_zone(zone
))
262 if (!mem_cgroup_disabled())
263 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
265 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
266 NR_ZONE_LRU_BASE
+ lru
);
267 lru_size
-= min(size
, lru_size
);
275 * Add a shrinker callback to be called from the vm.
277 int prealloc_shrinker(struct shrinker
*shrinker
)
279 size_t size
= sizeof(*shrinker
->nr_deferred
);
281 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
284 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
285 if (!shrinker
->nr_deferred
)
290 void free_prealloced_shrinker(struct shrinker
*shrinker
)
292 kfree(shrinker
->nr_deferred
);
293 shrinker
->nr_deferred
= NULL
;
296 void register_shrinker_prepared(struct shrinker
*shrinker
)
298 down_write(&shrinker_rwsem
);
299 list_add_tail(&shrinker
->list
, &shrinker_list
);
300 up_write(&shrinker_rwsem
);
303 int register_shrinker(struct shrinker
*shrinker
)
305 int err
= prealloc_shrinker(shrinker
);
309 register_shrinker_prepared(shrinker
);
312 EXPORT_SYMBOL(register_shrinker
);
317 void unregister_shrinker(struct shrinker
*shrinker
)
319 if (!shrinker
->nr_deferred
)
321 down_write(&shrinker_rwsem
);
322 list_del(&shrinker
->list
);
323 up_write(&shrinker_rwsem
);
324 kfree(shrinker
->nr_deferred
);
325 shrinker
->nr_deferred
= NULL
;
327 EXPORT_SYMBOL(unregister_shrinker
);
329 #define SHRINK_BATCH 128
331 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
332 struct shrinker
*shrinker
,
333 unsigned long nr_scanned
,
334 unsigned long nr_eligible
)
336 unsigned long freed
= 0;
337 unsigned long long delta
;
342 int nid
= shrinkctl
->nid
;
343 long batch_size
= shrinker
->batch
? shrinker
->batch
345 long scanned
= 0, next_deferred
;
347 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
352 * copy the current shrinker scan count into a local variable
353 * and zero it so that other concurrent shrinker invocations
354 * don't also do this scanning work.
356 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
359 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
361 do_div(delta
, nr_eligible
+ 1);
363 if (total_scan
< 0) {
364 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
365 shrinker
->scan_objects
, total_scan
);
366 total_scan
= freeable
;
369 next_deferred
= total_scan
;
372 * We need to avoid excessive windup on filesystem shrinkers
373 * due to large numbers of GFP_NOFS allocations causing the
374 * shrinkers to return -1 all the time. This results in a large
375 * nr being built up so when a shrink that can do some work
376 * comes along it empties the entire cache due to nr >>>
377 * freeable. This is bad for sustaining a working set in
380 * Hence only allow the shrinker to scan the entire cache when
381 * a large delta change is calculated directly.
383 if (delta
< freeable
/ 4)
384 total_scan
= min(total_scan
, freeable
/ 2);
387 * Avoid risking looping forever due to too large nr value:
388 * never try to free more than twice the estimate number of
391 if (total_scan
> freeable
* 2)
392 total_scan
= freeable
* 2;
394 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
395 nr_scanned
, nr_eligible
,
396 freeable
, delta
, total_scan
);
399 * Normally, we should not scan less than batch_size objects in one
400 * pass to avoid too frequent shrinker calls, but if the slab has less
401 * than batch_size objects in total and we are really tight on memory,
402 * we will try to reclaim all available objects, otherwise we can end
403 * up failing allocations although there are plenty of reclaimable
404 * objects spread over several slabs with usage less than the
407 * We detect the "tight on memory" situations by looking at the total
408 * number of objects we want to scan (total_scan). If it is greater
409 * than the total number of objects on slab (freeable), we must be
410 * scanning at high prio and therefore should try to reclaim as much as
413 while (total_scan
>= batch_size
||
414 total_scan
>= freeable
) {
416 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
418 shrinkctl
->nr_to_scan
= nr_to_scan
;
419 shrinkctl
->nr_scanned
= nr_to_scan
;
420 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
421 if (ret
== SHRINK_STOP
)
425 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
426 total_scan
-= shrinkctl
->nr_scanned
;
427 scanned
+= shrinkctl
->nr_scanned
;
432 if (next_deferred
>= scanned
)
433 next_deferred
-= scanned
;
437 * move the unused scan count back into the shrinker in a
438 * manner that handles concurrent updates. If we exhausted the
439 * scan, there is no need to do an update.
441 if (next_deferred
> 0)
442 new_nr
= atomic_long_add_return(next_deferred
,
443 &shrinker
->nr_deferred
[nid
]);
445 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
447 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
452 * shrink_slab - shrink slab caches
453 * @gfp_mask: allocation context
454 * @nid: node whose slab caches to target
455 * @memcg: memory cgroup whose slab caches to target
456 * @nr_scanned: pressure numerator
457 * @nr_eligible: pressure denominator
459 * Call the shrink functions to age shrinkable caches.
461 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
462 * unaware shrinkers will receive a node id of 0 instead.
464 * @memcg specifies the memory cgroup to target. If it is not NULL,
465 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
466 * objects from the memory cgroup specified. Otherwise, only unaware
467 * shrinkers are called.
469 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
470 * the available objects should be scanned. Page reclaim for example
471 * passes the number of pages scanned and the number of pages on the
472 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
473 * when it encountered mapped pages. The ratio is further biased by
474 * the ->seeks setting of the shrink function, which indicates the
475 * cost to recreate an object relative to that of an LRU page.
477 * Returns the number of reclaimed slab objects.
479 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
480 struct mem_cgroup
*memcg
,
481 unsigned long nr_scanned
,
482 unsigned long nr_eligible
)
484 struct shrinker
*shrinker
;
485 unsigned long freed
= 0;
487 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
491 nr_scanned
= SWAP_CLUSTER_MAX
;
493 if (!down_read_trylock(&shrinker_rwsem
)) {
495 * If we would return 0, our callers would understand that we
496 * have nothing else to shrink and give up trying. By returning
497 * 1 we keep it going and assume we'll be able to shrink next
504 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
505 struct shrink_control sc
= {
506 .gfp_mask
= gfp_mask
,
512 * If kernel memory accounting is disabled, we ignore
513 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
514 * passing NULL for memcg.
516 if (memcg_kmem_enabled() &&
517 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
520 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
523 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
525 * Bail out if someone want to register a new shrinker to
526 * prevent the regsitration from being stalled for long periods
527 * by parallel ongoing shrinking.
529 if (rwsem_is_contended(&shrinker_rwsem
)) {
535 up_read(&shrinker_rwsem
);
541 void drop_slab_node(int nid
)
546 struct mem_cgroup
*memcg
= NULL
;
550 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
552 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
553 } while (freed
> 10);
560 for_each_online_node(nid
)
564 static inline int is_page_cache_freeable(struct page
*page
)
567 * A freeable page cache page is referenced only by the caller
568 * that isolated the page, the page cache radix tree and
569 * optional buffer heads at page->private.
571 int radix_pins
= PageTransHuge(page
) && PageSwapCache(page
) ?
573 return page_count(page
) - page_has_private(page
) == 1 + radix_pins
;
576 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
578 if (current
->flags
& PF_SWAPWRITE
)
580 if (!inode_write_congested(inode
))
582 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
588 * We detected a synchronous write error writing a page out. Probably
589 * -ENOSPC. We need to propagate that into the address_space for a subsequent
590 * fsync(), msync() or close().
592 * The tricky part is that after writepage we cannot touch the mapping: nothing
593 * prevents it from being freed up. But we have a ref on the page and once
594 * that page is locked, the mapping is pinned.
596 * We're allowed to run sleeping lock_page() here because we know the caller has
599 static void handle_write_error(struct address_space
*mapping
,
600 struct page
*page
, int error
)
603 if (page_mapping(page
) == mapping
)
604 mapping_set_error(mapping
, error
);
608 /* possible outcome of pageout() */
610 /* failed to write page out, page is locked */
612 /* move page to the active list, page is locked */
614 /* page has been sent to the disk successfully, page is unlocked */
616 /* page is clean and locked */
621 * pageout is called by shrink_page_list() for each dirty page.
622 * Calls ->writepage().
624 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
625 struct scan_control
*sc
)
628 * If the page is dirty, only perform writeback if that write
629 * will be non-blocking. To prevent this allocation from being
630 * stalled by pagecache activity. But note that there may be
631 * stalls if we need to run get_block(). We could test
632 * PagePrivate for that.
634 * If this process is currently in __generic_file_write_iter() against
635 * this page's queue, we can perform writeback even if that
638 * If the page is swapcache, write it back even if that would
639 * block, for some throttling. This happens by accident, because
640 * swap_backing_dev_info is bust: it doesn't reflect the
641 * congestion state of the swapdevs. Easy to fix, if needed.
643 if (!is_page_cache_freeable(page
))
647 * Some data journaling orphaned pages can have
648 * page->mapping == NULL while being dirty with clean buffers.
650 if (page_has_private(page
)) {
651 if (try_to_free_buffers(page
)) {
652 ClearPageDirty(page
);
653 pr_info("%s: orphaned page\n", __func__
);
659 if (mapping
->a_ops
->writepage
== NULL
)
660 return PAGE_ACTIVATE
;
661 if (!may_write_to_inode(mapping
->host
, sc
))
664 if (clear_page_dirty_for_io(page
)) {
666 struct writeback_control wbc
= {
667 .sync_mode
= WB_SYNC_NONE
,
668 .nr_to_write
= SWAP_CLUSTER_MAX
,
670 .range_end
= LLONG_MAX
,
674 SetPageReclaim(page
);
675 res
= mapping
->a_ops
->writepage(page
, &wbc
);
677 handle_write_error(mapping
, page
, res
);
678 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
679 ClearPageReclaim(page
);
680 return PAGE_ACTIVATE
;
683 if (!PageWriteback(page
)) {
684 /* synchronous write or broken a_ops? */
685 ClearPageReclaim(page
);
687 trace_mm_vmscan_writepage(page
);
688 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
696 * Same as remove_mapping, but if the page is removed from the mapping, it
697 * gets returned with a refcount of 0.
699 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
705 BUG_ON(!PageLocked(page
));
706 BUG_ON(mapping
!= page_mapping(page
));
708 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
710 * The non racy check for a busy page.
712 * Must be careful with the order of the tests. When someone has
713 * a ref to the page, it may be possible that they dirty it then
714 * drop the reference. So if PageDirty is tested before page_count
715 * here, then the following race may occur:
717 * get_user_pages(&page);
718 * [user mapping goes away]
720 * !PageDirty(page) [good]
721 * SetPageDirty(page);
723 * !page_count(page) [good, discard it]
725 * [oops, our write_to data is lost]
727 * Reversing the order of the tests ensures such a situation cannot
728 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
729 * load is not satisfied before that of page->_refcount.
731 * Note that if SetPageDirty is always performed via set_page_dirty,
732 * and thus under tree_lock, then this ordering is not required.
734 if (unlikely(PageTransHuge(page
)) && PageSwapCache(page
))
735 refcount
= 1 + HPAGE_PMD_NR
;
738 if (!page_ref_freeze(page
, refcount
))
740 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
741 if (unlikely(PageDirty(page
))) {
742 page_ref_unfreeze(page
, refcount
);
746 if (PageSwapCache(page
)) {
747 swp_entry_t swap
= { .val
= page_private(page
) };
748 mem_cgroup_swapout(page
, swap
);
749 __delete_from_swap_cache(page
);
750 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
751 put_swap_page(page
, swap
);
753 void (*freepage
)(struct page
*);
756 freepage
= mapping
->a_ops
->freepage
;
758 * Remember a shadow entry for reclaimed file cache in
759 * order to detect refaults, thus thrashing, later on.
761 * But don't store shadows in an address space that is
762 * already exiting. This is not just an optizimation,
763 * inode reclaim needs to empty out the radix tree or
764 * the nodes are lost. Don't plant shadows behind its
767 * We also don't store shadows for DAX mappings because the
768 * only page cache pages found in these are zero pages
769 * covering holes, and because we don't want to mix DAX
770 * exceptional entries and shadow exceptional entries in the
773 if (reclaimed
&& page_is_file_cache(page
) &&
774 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
775 shadow
= workingset_eviction(mapping
, page
);
776 __delete_from_page_cache(page
, shadow
);
777 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
779 if (freepage
!= NULL
)
786 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
791 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
792 * someone else has a ref on the page, abort and return 0. If it was
793 * successfully detached, return 1. Assumes the caller has a single ref on
796 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
798 if (__remove_mapping(mapping
, page
, false)) {
800 * Unfreezing the refcount with 1 rather than 2 effectively
801 * drops the pagecache ref for us without requiring another
804 page_ref_unfreeze(page
, 1);
811 * putback_lru_page - put previously isolated page onto appropriate LRU list
812 * @page: page to be put back to appropriate lru list
814 * Add previously isolated @page to appropriate LRU list.
815 * Page may still be unevictable for other reasons.
817 * lru_lock must not be held, interrupts must be enabled.
819 void putback_lru_page(struct page
*page
)
822 int was_unevictable
= PageUnevictable(page
);
824 VM_BUG_ON_PAGE(PageLRU(page
), page
);
827 ClearPageUnevictable(page
);
829 if (page_evictable(page
)) {
831 * For evictable pages, we can use the cache.
832 * In event of a race, worst case is we end up with an
833 * unevictable page on [in]active list.
834 * We know how to handle that.
836 is_unevictable
= false;
840 * Put unevictable pages directly on zone's unevictable
843 is_unevictable
= true;
844 add_page_to_unevictable_list(page
);
846 * When racing with an mlock or AS_UNEVICTABLE clearing
847 * (page is unlocked) make sure that if the other thread
848 * does not observe our setting of PG_lru and fails
849 * isolation/check_move_unevictable_pages,
850 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
851 * the page back to the evictable list.
853 * The other side is TestClearPageMlocked() or shmem_lock().
859 * page's status can change while we move it among lru. If an evictable
860 * page is on unevictable list, it never be freed. To avoid that,
861 * check after we added it to the list, again.
863 if (is_unevictable
&& page_evictable(page
)) {
864 if (!isolate_lru_page(page
)) {
868 /* This means someone else dropped this page from LRU
869 * So, it will be freed or putback to LRU again. There is
870 * nothing to do here.
874 if (was_unevictable
&& !is_unevictable
)
875 count_vm_event(UNEVICTABLE_PGRESCUED
);
876 else if (!was_unevictable
&& is_unevictable
)
877 count_vm_event(UNEVICTABLE_PGCULLED
);
879 put_page(page
); /* drop ref from isolate */
882 enum page_references
{
884 PAGEREF_RECLAIM_CLEAN
,
889 static enum page_references
page_check_references(struct page
*page
,
890 struct scan_control
*sc
)
892 int referenced_ptes
, referenced_page
;
893 unsigned long vm_flags
;
895 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
897 referenced_page
= TestClearPageReferenced(page
);
900 * Mlock lost the isolation race with us. Let try_to_unmap()
901 * move the page to the unevictable list.
903 if (vm_flags
& VM_LOCKED
)
904 return PAGEREF_RECLAIM
;
906 if (referenced_ptes
) {
907 if (PageSwapBacked(page
))
908 return PAGEREF_ACTIVATE
;
910 * All mapped pages start out with page table
911 * references from the instantiating fault, so we need
912 * to look twice if a mapped file page is used more
915 * Mark it and spare it for another trip around the
916 * inactive list. Another page table reference will
917 * lead to its activation.
919 * Note: the mark is set for activated pages as well
920 * so that recently deactivated but used pages are
923 SetPageReferenced(page
);
925 if (referenced_page
|| referenced_ptes
> 1)
926 return PAGEREF_ACTIVATE
;
929 * Activate file-backed executable pages after first usage.
931 if (vm_flags
& VM_EXEC
)
932 return PAGEREF_ACTIVATE
;
937 /* Reclaim if clean, defer dirty pages to writeback */
938 if (referenced_page
&& !PageSwapBacked(page
))
939 return PAGEREF_RECLAIM_CLEAN
;
941 return PAGEREF_RECLAIM
;
944 /* Check if a page is dirty or under writeback */
945 static void page_check_dirty_writeback(struct page
*page
,
946 bool *dirty
, bool *writeback
)
948 struct address_space
*mapping
;
951 * Anonymous pages are not handled by flushers and must be written
952 * from reclaim context. Do not stall reclaim based on them
954 if (!page_is_file_cache(page
) ||
955 (PageAnon(page
) && !PageSwapBacked(page
))) {
961 /* By default assume that the page flags are accurate */
962 *dirty
= PageDirty(page
);
963 *writeback
= PageWriteback(page
);
965 /* Verify dirty/writeback state if the filesystem supports it */
966 if (!page_has_private(page
))
969 mapping
= page_mapping(page
);
970 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
971 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
974 struct reclaim_stat
{
976 unsigned nr_unqueued_dirty
;
977 unsigned nr_congested
;
978 unsigned nr_writeback
;
979 unsigned nr_immediate
;
980 unsigned nr_activate
;
981 unsigned nr_ref_keep
;
982 unsigned nr_unmap_fail
;
986 * shrink_page_list() returns the number of reclaimed pages
988 static unsigned long shrink_page_list(struct list_head
*page_list
,
989 struct pglist_data
*pgdat
,
990 struct scan_control
*sc
,
991 enum ttu_flags ttu_flags
,
992 struct reclaim_stat
*stat
,
995 LIST_HEAD(ret_pages
);
996 LIST_HEAD(free_pages
);
998 unsigned nr_unqueued_dirty
= 0;
999 unsigned nr_dirty
= 0;
1000 unsigned nr_congested
= 0;
1001 unsigned nr_reclaimed
= 0;
1002 unsigned nr_writeback
= 0;
1003 unsigned nr_immediate
= 0;
1004 unsigned nr_ref_keep
= 0;
1005 unsigned nr_unmap_fail
= 0;
1009 while (!list_empty(page_list
)) {
1010 struct address_space
*mapping
;
1013 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
1014 bool dirty
, writeback
;
1018 page
= lru_to_page(page_list
);
1019 list_del(&page
->lru
);
1021 if (!trylock_page(page
))
1024 VM_BUG_ON_PAGE(PageActive(page
), page
);
1028 if (unlikely(!page_evictable(page
)))
1029 goto activate_locked
;
1031 if (!sc
->may_unmap
&& page_mapped(page
))
1034 /* Double the slab pressure for mapped and swapcache pages */
1035 if ((page_mapped(page
) || PageSwapCache(page
)) &&
1036 !(PageAnon(page
) && !PageSwapBacked(page
)))
1039 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1040 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1043 * The number of dirty pages determines if a zone is marked
1044 * reclaim_congested which affects wait_iff_congested. kswapd
1045 * will stall and start writing pages if the tail of the LRU
1046 * is all dirty unqueued pages.
1048 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1049 if (dirty
|| writeback
)
1052 if (dirty
&& !writeback
)
1053 nr_unqueued_dirty
++;
1056 * Treat this page as congested if the underlying BDI is or if
1057 * pages are cycling through the LRU so quickly that the
1058 * pages marked for immediate reclaim are making it to the
1059 * end of the LRU a second time.
1061 mapping
= page_mapping(page
);
1062 if (((dirty
|| writeback
) && mapping
&&
1063 inode_write_congested(mapping
->host
)) ||
1064 (writeback
&& PageReclaim(page
)))
1068 * If a page at the tail of the LRU is under writeback, there
1069 * are three cases to consider.
1071 * 1) If reclaim is encountering an excessive number of pages
1072 * under writeback and this page is both under writeback and
1073 * PageReclaim then it indicates that pages are being queued
1074 * for IO but are being recycled through the LRU before the
1075 * IO can complete. Waiting on the page itself risks an
1076 * indefinite stall if it is impossible to writeback the
1077 * page due to IO error or disconnected storage so instead
1078 * note that the LRU is being scanned too quickly and the
1079 * caller can stall after page list has been processed.
1081 * 2) Global or new memcg reclaim encounters a page that is
1082 * not marked for immediate reclaim, or the caller does not
1083 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1084 * not to fs). In this case mark the page for immediate
1085 * reclaim and continue scanning.
1087 * Require may_enter_fs because we would wait on fs, which
1088 * may not have submitted IO yet. And the loop driver might
1089 * enter reclaim, and deadlock if it waits on a page for
1090 * which it is needed to do the write (loop masks off
1091 * __GFP_IO|__GFP_FS for this reason); but more thought
1092 * would probably show more reasons.
1094 * 3) Legacy memcg encounters a page that is already marked
1095 * PageReclaim. memcg does not have any dirty pages
1096 * throttling so we could easily OOM just because too many
1097 * pages are in writeback and there is nothing else to
1098 * reclaim. Wait for the writeback to complete.
1100 * In cases 1) and 2) we activate the pages to get them out of
1101 * the way while we continue scanning for clean pages on the
1102 * inactive list and refilling from the active list. The
1103 * observation here is that waiting for disk writes is more
1104 * expensive than potentially causing reloads down the line.
1105 * Since they're marked for immediate reclaim, they won't put
1106 * memory pressure on the cache working set any longer than it
1107 * takes to write them to disk.
1109 if (PageWriteback(page
)) {
1111 if (current_is_kswapd() &&
1112 PageReclaim(page
) &&
1113 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1115 goto activate_locked
;
1118 } else if (sane_reclaim(sc
) ||
1119 !PageReclaim(page
) || !may_enter_fs
) {
1121 * This is slightly racy - end_page_writeback()
1122 * might have just cleared PageReclaim, then
1123 * setting PageReclaim here end up interpreted
1124 * as PageReadahead - but that does not matter
1125 * enough to care. What we do want is for this
1126 * page to have PageReclaim set next time memcg
1127 * reclaim reaches the tests above, so it will
1128 * then wait_on_page_writeback() to avoid OOM;
1129 * and it's also appropriate in global reclaim.
1131 SetPageReclaim(page
);
1133 goto activate_locked
;
1138 wait_on_page_writeback(page
);
1139 /* then go back and try same page again */
1140 list_add_tail(&page
->lru
, page_list
);
1146 references
= page_check_references(page
, sc
);
1148 switch (references
) {
1149 case PAGEREF_ACTIVATE
:
1150 goto activate_locked
;
1154 case PAGEREF_RECLAIM
:
1155 case PAGEREF_RECLAIM_CLEAN
:
1156 ; /* try to reclaim the page below */
1160 * Anonymous process memory has backing store?
1161 * Try to allocate it some swap space here.
1162 * Lazyfree page could be freed directly
1164 if (PageAnon(page
) && PageSwapBacked(page
)) {
1165 if (!PageSwapCache(page
)) {
1166 if (!(sc
->gfp_mask
& __GFP_IO
))
1168 if (PageTransHuge(page
)) {
1169 /* cannot split THP, skip it */
1170 if (!can_split_huge_page(page
, NULL
))
1171 goto activate_locked
;
1173 * Split pages without a PMD map right
1174 * away. Chances are some or all of the
1175 * tail pages can be freed without IO.
1177 if (!compound_mapcount(page
) &&
1178 split_huge_page_to_list(page
,
1180 goto activate_locked
;
1182 if (!add_to_swap(page
)) {
1183 if (!PageTransHuge(page
))
1184 goto activate_locked
;
1185 /* Fallback to swap normal pages */
1186 if (split_huge_page_to_list(page
,
1188 goto activate_locked
;
1189 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1190 count_vm_event(THP_SWPOUT_FALLBACK
);
1192 if (!add_to_swap(page
))
1193 goto activate_locked
;
1198 /* Adding to swap updated mapping */
1199 mapping
= page_mapping(page
);
1201 } else if (unlikely(PageTransHuge(page
))) {
1202 /* Split file THP */
1203 if (split_huge_page_to_list(page
, page_list
))
1208 * The page is mapped into the page tables of one or more
1209 * processes. Try to unmap it here.
1211 if (page_mapped(page
)) {
1212 enum ttu_flags flags
= ttu_flags
| TTU_BATCH_FLUSH
;
1214 if (unlikely(PageTransHuge(page
)))
1215 flags
|= TTU_SPLIT_HUGE_PMD
;
1216 if (!try_to_unmap(page
, flags
)) {
1218 goto activate_locked
;
1222 if (PageDirty(page
)) {
1224 * Only kswapd can writeback filesystem pages
1225 * to avoid risk of stack overflow. But avoid
1226 * injecting inefficient single-page IO into
1227 * flusher writeback as much as possible: only
1228 * write pages when we've encountered many
1229 * dirty pages, and when we've already scanned
1230 * the rest of the LRU for clean pages and see
1231 * the same dirty pages again (PageReclaim).
1233 if (page_is_file_cache(page
) &&
1234 (!current_is_kswapd() || !PageReclaim(page
) ||
1235 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1237 * Immediately reclaim when written back.
1238 * Similar in principal to deactivate_page()
1239 * except we already have the page isolated
1240 * and know it's dirty
1242 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1243 SetPageReclaim(page
);
1245 goto activate_locked
;
1248 if (references
== PAGEREF_RECLAIM_CLEAN
)
1252 if (!sc
->may_writepage
)
1256 * Page is dirty. Flush the TLB if a writable entry
1257 * potentially exists to avoid CPU writes after IO
1258 * starts and then write it out here.
1260 try_to_unmap_flush_dirty();
1261 switch (pageout(page
, mapping
, sc
)) {
1265 goto activate_locked
;
1267 if (PageWriteback(page
))
1269 if (PageDirty(page
))
1273 * A synchronous write - probably a ramdisk. Go
1274 * ahead and try to reclaim the page.
1276 if (!trylock_page(page
))
1278 if (PageDirty(page
) || PageWriteback(page
))
1280 mapping
= page_mapping(page
);
1282 ; /* try to free the page below */
1287 * If the page has buffers, try to free the buffer mappings
1288 * associated with this page. If we succeed we try to free
1291 * We do this even if the page is PageDirty().
1292 * try_to_release_page() does not perform I/O, but it is
1293 * possible for a page to have PageDirty set, but it is actually
1294 * clean (all its buffers are clean). This happens if the
1295 * buffers were written out directly, with submit_bh(). ext3
1296 * will do this, as well as the blockdev mapping.
1297 * try_to_release_page() will discover that cleanness and will
1298 * drop the buffers and mark the page clean - it can be freed.
1300 * Rarely, pages can have buffers and no ->mapping. These are
1301 * the pages which were not successfully invalidated in
1302 * truncate_complete_page(). We try to drop those buffers here
1303 * and if that worked, and the page is no longer mapped into
1304 * process address space (page_count == 1) it can be freed.
1305 * Otherwise, leave the page on the LRU so it is swappable.
1307 if (page_has_private(page
)) {
1308 if (!try_to_release_page(page
, sc
->gfp_mask
))
1309 goto activate_locked
;
1310 if (!mapping
&& page_count(page
) == 1) {
1312 if (put_page_testzero(page
))
1316 * rare race with speculative reference.
1317 * the speculative reference will free
1318 * this page shortly, so we may
1319 * increment nr_reclaimed here (and
1320 * leave it off the LRU).
1328 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1329 /* follow __remove_mapping for reference */
1330 if (!page_ref_freeze(page
, 1))
1332 if (PageDirty(page
)) {
1333 page_ref_unfreeze(page
, 1);
1337 count_vm_event(PGLAZYFREED
);
1338 count_memcg_page_event(page
, PGLAZYFREED
);
1339 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1342 * At this point, we have no other references and there is
1343 * no way to pick any more up (removed from LRU, removed
1344 * from pagecache). Can use non-atomic bitops now (and
1345 * we obviously don't have to worry about waking up a process
1346 * waiting on the page lock, because there are no references.
1348 __ClearPageLocked(page
);
1353 * Is there need to periodically free_page_list? It would
1354 * appear not as the counts should be low
1356 if (unlikely(PageTransHuge(page
))) {
1357 mem_cgroup_uncharge(page
);
1358 (*get_compound_page_dtor(page
))(page
);
1360 list_add(&page
->lru
, &free_pages
);
1364 /* Not a candidate for swapping, so reclaim swap space. */
1365 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1367 try_to_free_swap(page
);
1368 VM_BUG_ON_PAGE(PageActive(page
), page
);
1369 if (!PageMlocked(page
)) {
1370 SetPageActive(page
);
1372 count_memcg_page_event(page
, PGACTIVATE
);
1377 list_add(&page
->lru
, &ret_pages
);
1378 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1381 mem_cgroup_uncharge_list(&free_pages
);
1382 try_to_unmap_flush();
1383 free_unref_page_list(&free_pages
);
1385 list_splice(&ret_pages
, page_list
);
1386 count_vm_events(PGACTIVATE
, pgactivate
);
1389 stat
->nr_dirty
= nr_dirty
;
1390 stat
->nr_congested
= nr_congested
;
1391 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1392 stat
->nr_writeback
= nr_writeback
;
1393 stat
->nr_immediate
= nr_immediate
;
1394 stat
->nr_activate
= pgactivate
;
1395 stat
->nr_ref_keep
= nr_ref_keep
;
1396 stat
->nr_unmap_fail
= nr_unmap_fail
;
1398 return nr_reclaimed
;
1401 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1402 struct list_head
*page_list
)
1404 struct scan_control sc
= {
1405 .gfp_mask
= GFP_KERNEL
,
1406 .priority
= DEF_PRIORITY
,
1410 struct page
*page
, *next
;
1411 LIST_HEAD(clean_pages
);
1413 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1414 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1415 !__PageMovable(page
)) {
1416 ClearPageActive(page
);
1417 list_move(&page
->lru
, &clean_pages
);
1421 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1422 TTU_IGNORE_ACCESS
, NULL
, true);
1423 list_splice(&clean_pages
, page_list
);
1424 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1429 * Attempt to remove the specified page from its LRU. Only take this page
1430 * if it is of the appropriate PageActive status. Pages which are being
1431 * freed elsewhere are also ignored.
1433 * page: page to consider
1434 * mode: one of the LRU isolation modes defined above
1436 * returns 0 on success, -ve errno on failure.
1438 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1442 /* Only take pages on the LRU. */
1446 /* Compaction should not handle unevictable pages but CMA can do so */
1447 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1453 * To minimise LRU disruption, the caller can indicate that it only
1454 * wants to isolate pages it will be able to operate on without
1455 * blocking - clean pages for the most part.
1457 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1458 * that it is possible to migrate without blocking
1460 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1461 /* All the caller can do on PageWriteback is block */
1462 if (PageWriteback(page
))
1465 if (PageDirty(page
)) {
1466 struct address_space
*mapping
;
1470 * Only pages without mappings or that have a
1471 * ->migratepage callback are possible to migrate
1472 * without blocking. However, we can be racing with
1473 * truncation so it's necessary to lock the page
1474 * to stabilise the mapping as truncation holds
1475 * the page lock until after the page is removed
1476 * from the page cache.
1478 if (!trylock_page(page
))
1481 mapping
= page_mapping(page
);
1482 migrate_dirty
= !mapping
|| mapping
->a_ops
->migratepage
;
1489 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1492 if (likely(get_page_unless_zero(page
))) {
1494 * Be careful not to clear PageLRU until after we're
1495 * sure the page is not being freed elsewhere -- the
1496 * page release code relies on it.
1507 * Update LRU sizes after isolating pages. The LRU size updates must
1508 * be complete before mem_cgroup_update_lru_size due to a santity check.
1510 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1511 enum lru_list lru
, unsigned long *nr_zone_taken
)
1515 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1516 if (!nr_zone_taken
[zid
])
1519 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1521 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1528 * zone_lru_lock is heavily contended. Some of the functions that
1529 * shrink the lists perform better by taking out a batch of pages
1530 * and working on them outside the LRU lock.
1532 * For pagecache intensive workloads, this function is the hottest
1533 * spot in the kernel (apart from copy_*_user functions).
1535 * Appropriate locks must be held before calling this function.
1537 * @nr_to_scan: The number of eligible pages to look through on the list.
1538 * @lruvec: The LRU vector to pull pages from.
1539 * @dst: The temp list to put pages on to.
1540 * @nr_scanned: The number of pages that were scanned.
1541 * @sc: The scan_control struct for this reclaim session
1542 * @mode: One of the LRU isolation modes
1543 * @lru: LRU list id for isolating
1545 * returns how many pages were moved onto *@dst.
1547 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1548 struct lruvec
*lruvec
, struct list_head
*dst
,
1549 unsigned long *nr_scanned
, struct scan_control
*sc
,
1550 isolate_mode_t mode
, enum lru_list lru
)
1552 struct list_head
*src
= &lruvec
->lists
[lru
];
1553 unsigned long nr_taken
= 0;
1554 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1555 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1556 unsigned long skipped
= 0;
1557 unsigned long scan
, total_scan
, nr_pages
;
1558 LIST_HEAD(pages_skipped
);
1561 for (total_scan
= 0;
1562 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1566 page
= lru_to_page(src
);
1567 prefetchw_prev_lru_page(page
, src
, flags
);
1569 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1571 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1572 list_move(&page
->lru
, &pages_skipped
);
1573 nr_skipped
[page_zonenum(page
)]++;
1578 * Do not count skipped pages because that makes the function
1579 * return with no isolated pages if the LRU mostly contains
1580 * ineligible pages. This causes the VM to not reclaim any
1581 * pages, triggering a premature OOM.
1584 switch (__isolate_lru_page(page
, mode
)) {
1586 nr_pages
= hpage_nr_pages(page
);
1587 nr_taken
+= nr_pages
;
1588 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1589 list_move(&page
->lru
, dst
);
1593 /* else it is being freed elsewhere */
1594 list_move(&page
->lru
, src
);
1603 * Splice any skipped pages to the start of the LRU list. Note that
1604 * this disrupts the LRU order when reclaiming for lower zones but
1605 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1606 * scanning would soon rescan the same pages to skip and put the
1607 * system at risk of premature OOM.
1609 if (!list_empty(&pages_skipped
)) {
1612 list_splice(&pages_skipped
, src
);
1613 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1614 if (!nr_skipped
[zid
])
1617 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1618 skipped
+= nr_skipped
[zid
];
1621 *nr_scanned
= total_scan
;
1622 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1623 total_scan
, skipped
, nr_taken
, mode
, lru
);
1624 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1629 * isolate_lru_page - tries to isolate a page from its LRU list
1630 * @page: page to isolate from its LRU list
1632 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1633 * vmstat statistic corresponding to whatever LRU list the page was on.
1635 * Returns 0 if the page was removed from an LRU list.
1636 * Returns -EBUSY if the page was not on an LRU list.
1638 * The returned page will have PageLRU() cleared. If it was found on
1639 * the active list, it will have PageActive set. If it was found on
1640 * the unevictable list, it will have the PageUnevictable bit set. That flag
1641 * may need to be cleared by the caller before letting the page go.
1643 * The vmstat statistic corresponding to the list on which the page was
1644 * found will be decremented.
1647 * (1) Must be called with an elevated refcount on the page. This is a
1648 * fundamentnal difference from isolate_lru_pages (which is called
1649 * without a stable reference).
1650 * (2) the lru_lock must not be held.
1651 * (3) interrupts must be enabled.
1653 int isolate_lru_page(struct page
*page
)
1657 VM_BUG_ON_PAGE(!page_count(page
), page
);
1658 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1660 if (PageLRU(page
)) {
1661 struct zone
*zone
= page_zone(page
);
1662 struct lruvec
*lruvec
;
1664 spin_lock_irq(zone_lru_lock(zone
));
1665 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1666 if (PageLRU(page
)) {
1667 int lru
= page_lru(page
);
1670 del_page_from_lru_list(page
, lruvec
, lru
);
1673 spin_unlock_irq(zone_lru_lock(zone
));
1679 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1680 * then get resheduled. When there are massive number of tasks doing page
1681 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1682 * the LRU list will go small and be scanned faster than necessary, leading to
1683 * unnecessary swapping, thrashing and OOM.
1685 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1686 struct scan_control
*sc
)
1688 unsigned long inactive
, isolated
;
1690 if (current_is_kswapd())
1693 if (!sane_reclaim(sc
))
1697 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1698 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1700 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1701 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1705 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1706 * won't get blocked by normal direct-reclaimers, forming a circular
1709 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1712 return isolated
> inactive
;
1715 static noinline_for_stack
void
1716 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1718 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1719 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1720 LIST_HEAD(pages_to_free
);
1723 * Put back any unfreeable pages.
1725 while (!list_empty(page_list
)) {
1726 struct page
*page
= lru_to_page(page_list
);
1729 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1730 list_del(&page
->lru
);
1731 if (unlikely(!page_evictable(page
))) {
1732 spin_unlock_irq(&pgdat
->lru_lock
);
1733 putback_lru_page(page
);
1734 spin_lock_irq(&pgdat
->lru_lock
);
1738 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1741 lru
= page_lru(page
);
1742 add_page_to_lru_list(page
, lruvec
, lru
);
1744 if (is_active_lru(lru
)) {
1745 int file
= is_file_lru(lru
);
1746 int numpages
= hpage_nr_pages(page
);
1747 reclaim_stat
->recent_rotated
[file
] += numpages
;
1749 if (put_page_testzero(page
)) {
1750 __ClearPageLRU(page
);
1751 __ClearPageActive(page
);
1752 del_page_from_lru_list(page
, lruvec
, lru
);
1754 if (unlikely(PageCompound(page
))) {
1755 spin_unlock_irq(&pgdat
->lru_lock
);
1756 mem_cgroup_uncharge(page
);
1757 (*get_compound_page_dtor(page
))(page
);
1758 spin_lock_irq(&pgdat
->lru_lock
);
1760 list_add(&page
->lru
, &pages_to_free
);
1765 * To save our caller's stack, now use input list for pages to free.
1767 list_splice(&pages_to_free
, page_list
);
1771 * If a kernel thread (such as nfsd for loop-back mounts) services
1772 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1773 * In that case we should only throttle if the backing device it is
1774 * writing to is congested. In other cases it is safe to throttle.
1776 static int current_may_throttle(void)
1778 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1779 current
->backing_dev_info
== NULL
||
1780 bdi_write_congested(current
->backing_dev_info
);
1784 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1785 * of reclaimed pages
1787 static noinline_for_stack
unsigned long
1788 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1789 struct scan_control
*sc
, enum lru_list lru
)
1791 LIST_HEAD(page_list
);
1792 unsigned long nr_scanned
;
1793 unsigned long nr_reclaimed
= 0;
1794 unsigned long nr_taken
;
1795 struct reclaim_stat stat
= {};
1796 isolate_mode_t isolate_mode
= 0;
1797 int file
= is_file_lru(lru
);
1798 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1799 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1800 bool stalled
= false;
1802 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1806 /* wait a bit for the reclaimer. */
1810 /* We are about to die and free our memory. Return now. */
1811 if (fatal_signal_pending(current
))
1812 return SWAP_CLUSTER_MAX
;
1818 isolate_mode
|= ISOLATE_UNMAPPED
;
1820 spin_lock_irq(&pgdat
->lru_lock
);
1822 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1823 &nr_scanned
, sc
, isolate_mode
, lru
);
1825 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1826 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1828 if (current_is_kswapd()) {
1829 if (global_reclaim(sc
))
1830 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1831 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1834 if (global_reclaim(sc
))
1835 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1836 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1839 spin_unlock_irq(&pgdat
->lru_lock
);
1844 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1847 spin_lock_irq(&pgdat
->lru_lock
);
1849 if (current_is_kswapd()) {
1850 if (global_reclaim(sc
))
1851 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1852 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1855 if (global_reclaim(sc
))
1856 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1857 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1861 putback_inactive_pages(lruvec
, &page_list
);
1863 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1865 spin_unlock_irq(&pgdat
->lru_lock
);
1867 mem_cgroup_uncharge_list(&page_list
);
1868 free_unref_page_list(&page_list
);
1871 * If reclaim is isolating dirty pages under writeback, it implies
1872 * that the long-lived page allocation rate is exceeding the page
1873 * laundering rate. Either the global limits are not being effective
1874 * at throttling processes due to the page distribution throughout
1875 * zones or there is heavy usage of a slow backing device. The
1876 * only option is to throttle from reclaim context which is not ideal
1877 * as there is no guarantee the dirtying process is throttled in the
1878 * same way balance_dirty_pages() manages.
1880 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1881 * of pages under pages flagged for immediate reclaim and stall if any
1882 * are encountered in the nr_immediate check below.
1884 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1885 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1888 * If dirty pages are scanned that are not queued for IO, it
1889 * implies that flushers are not doing their job. This can
1890 * happen when memory pressure pushes dirty pages to the end of
1891 * the LRU before the dirty limits are breached and the dirty
1892 * data has expired. It can also happen when the proportion of
1893 * dirty pages grows not through writes but through memory
1894 * pressure reclaiming all the clean cache. And in some cases,
1895 * the flushers simply cannot keep up with the allocation
1896 * rate. Nudge the flusher threads in case they are asleep.
1898 if (stat
.nr_unqueued_dirty
== nr_taken
)
1899 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1902 * Legacy memcg will stall in page writeback so avoid forcibly
1905 if (sane_reclaim(sc
)) {
1907 * Tag a zone as congested if all the dirty pages scanned were
1908 * backed by a congested BDI and wait_iff_congested will stall.
1910 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1911 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1913 /* Allow kswapd to start writing pages during reclaim. */
1914 if (stat
.nr_unqueued_dirty
== nr_taken
)
1915 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1918 * If kswapd scans pages marked marked for immediate
1919 * reclaim and under writeback (nr_immediate), it implies
1920 * that pages are cycling through the LRU faster than
1921 * they are written so also forcibly stall.
1923 if (stat
.nr_immediate
&& current_may_throttle())
1924 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1928 * Stall direct reclaim for IO completions if underlying BDIs or zone
1929 * is congested. Allow kswapd to continue until it starts encountering
1930 * unqueued dirty pages or cycling through the LRU too quickly.
1932 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1933 current_may_throttle())
1934 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1936 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1937 nr_scanned
, nr_reclaimed
,
1938 stat
.nr_dirty
, stat
.nr_writeback
,
1939 stat
.nr_congested
, stat
.nr_immediate
,
1940 stat
.nr_activate
, stat
.nr_ref_keep
,
1942 sc
->priority
, file
);
1943 return nr_reclaimed
;
1947 * This moves pages from the active list to the inactive list.
1949 * We move them the other way if the page is referenced by one or more
1950 * processes, from rmap.
1952 * If the pages are mostly unmapped, the processing is fast and it is
1953 * appropriate to hold zone_lru_lock across the whole operation. But if
1954 * the pages are mapped, the processing is slow (page_referenced()) so we
1955 * should drop zone_lru_lock around each page. It's impossible to balance
1956 * this, so instead we remove the pages from the LRU while processing them.
1957 * It is safe to rely on PG_active against the non-LRU pages in here because
1958 * nobody will play with that bit on a non-LRU page.
1960 * The downside is that we have to touch page->_refcount against each page.
1961 * But we had to alter page->flags anyway.
1963 * Returns the number of pages moved to the given lru.
1966 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1967 struct list_head
*list
,
1968 struct list_head
*pages_to_free
,
1971 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1976 while (!list_empty(list
)) {
1977 page
= lru_to_page(list
);
1978 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1980 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1983 nr_pages
= hpage_nr_pages(page
);
1984 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1985 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1987 if (put_page_testzero(page
)) {
1988 __ClearPageLRU(page
);
1989 __ClearPageActive(page
);
1990 del_page_from_lru_list(page
, lruvec
, lru
);
1992 if (unlikely(PageCompound(page
))) {
1993 spin_unlock_irq(&pgdat
->lru_lock
);
1994 mem_cgroup_uncharge(page
);
1995 (*get_compound_page_dtor(page
))(page
);
1996 spin_lock_irq(&pgdat
->lru_lock
);
1998 list_add(&page
->lru
, pages_to_free
);
2000 nr_moved
+= nr_pages
;
2004 if (!is_active_lru(lru
)) {
2005 __count_vm_events(PGDEACTIVATE
, nr_moved
);
2006 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
2013 static void shrink_active_list(unsigned long nr_to_scan
,
2014 struct lruvec
*lruvec
,
2015 struct scan_control
*sc
,
2018 unsigned long nr_taken
;
2019 unsigned long nr_scanned
;
2020 unsigned long vm_flags
;
2021 LIST_HEAD(l_hold
); /* The pages which were snipped off */
2022 LIST_HEAD(l_active
);
2023 LIST_HEAD(l_inactive
);
2025 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2026 unsigned nr_deactivate
, nr_activate
;
2027 unsigned nr_rotated
= 0;
2028 isolate_mode_t isolate_mode
= 0;
2029 int file
= is_file_lru(lru
);
2030 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2035 isolate_mode
|= ISOLATE_UNMAPPED
;
2037 spin_lock_irq(&pgdat
->lru_lock
);
2039 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
2040 &nr_scanned
, sc
, isolate_mode
, lru
);
2042 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2043 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
2045 __count_vm_events(PGREFILL
, nr_scanned
);
2046 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2048 spin_unlock_irq(&pgdat
->lru_lock
);
2050 while (!list_empty(&l_hold
)) {
2052 page
= lru_to_page(&l_hold
);
2053 list_del(&page
->lru
);
2055 if (unlikely(!page_evictable(page
))) {
2056 putback_lru_page(page
);
2060 if (unlikely(buffer_heads_over_limit
)) {
2061 if (page_has_private(page
) && trylock_page(page
)) {
2062 if (page_has_private(page
))
2063 try_to_release_page(page
, 0);
2068 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2070 nr_rotated
+= hpage_nr_pages(page
);
2072 * Identify referenced, file-backed active pages and
2073 * give them one more trip around the active list. So
2074 * that executable code get better chances to stay in
2075 * memory under moderate memory pressure. Anon pages
2076 * are not likely to be evicted by use-once streaming
2077 * IO, plus JVM can create lots of anon VM_EXEC pages,
2078 * so we ignore them here.
2080 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2081 list_add(&page
->lru
, &l_active
);
2086 ClearPageActive(page
); /* we are de-activating */
2087 list_add(&page
->lru
, &l_inactive
);
2091 * Move pages back to the lru list.
2093 spin_lock_irq(&pgdat
->lru_lock
);
2095 * Count referenced pages from currently used mappings as rotated,
2096 * even though only some of them are actually re-activated. This
2097 * helps balance scan pressure between file and anonymous pages in
2100 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2102 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2103 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2104 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2105 spin_unlock_irq(&pgdat
->lru_lock
);
2107 mem_cgroup_uncharge_list(&l_hold
);
2108 free_unref_page_list(&l_hold
);
2109 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2110 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2114 * The inactive anon list should be small enough that the VM never has
2115 * to do too much work.
2117 * The inactive file list should be small enough to leave most memory
2118 * to the established workingset on the scan-resistant active list,
2119 * but large enough to avoid thrashing the aggregate readahead window.
2121 * Both inactive lists should also be large enough that each inactive
2122 * page has a chance to be referenced again before it is reclaimed.
2124 * If that fails and refaulting is observed, the inactive list grows.
2126 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2127 * on this LRU, maintained by the pageout code. An inactive_ratio
2128 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2131 * memory ratio inactive
2132 * -------------------------------------
2141 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2142 struct mem_cgroup
*memcg
,
2143 struct scan_control
*sc
, bool actual_reclaim
)
2145 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2146 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2147 enum lru_list inactive_lru
= file
* LRU_FILE
;
2148 unsigned long inactive
, active
;
2149 unsigned long inactive_ratio
;
2150 unsigned long refaults
;
2154 * If we don't have swap space, anonymous page deactivation
2157 if (!file
&& !total_swap_pages
)
2160 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2161 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2164 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2166 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2169 * When refaults are being observed, it means a new workingset
2170 * is being established. Disable active list protection to get
2171 * rid of the stale workingset quickly.
2173 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2176 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2178 inactive_ratio
= int_sqrt(10 * gb
);
2184 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2185 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2186 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2187 inactive_ratio
, file
);
2189 return inactive
* inactive_ratio
< active
;
2192 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2193 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2194 struct scan_control
*sc
)
2196 if (is_active_lru(lru
)) {
2197 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2199 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2203 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2214 * Determine how aggressively the anon and file LRU lists should be
2215 * scanned. The relative value of each set of LRU lists is determined
2216 * by looking at the fraction of the pages scanned we did rotate back
2217 * onto the active list instead of evict.
2219 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2220 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2222 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2223 struct scan_control
*sc
, unsigned long *nr
,
2224 unsigned long *lru_pages
)
2226 int swappiness
= mem_cgroup_swappiness(memcg
);
2227 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2229 u64 denominator
= 0; /* gcc */
2230 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2231 unsigned long anon_prio
, file_prio
;
2232 enum scan_balance scan_balance
;
2233 unsigned long anon
, file
;
2234 unsigned long ap
, fp
;
2237 /* If we have no swap space, do not bother scanning anon pages. */
2238 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2239 scan_balance
= SCAN_FILE
;
2244 * Global reclaim will swap to prevent OOM even with no
2245 * swappiness, but memcg users want to use this knob to
2246 * disable swapping for individual groups completely when
2247 * using the memory controller's swap limit feature would be
2250 if (!global_reclaim(sc
) && !swappiness
) {
2251 scan_balance
= SCAN_FILE
;
2256 * Do not apply any pressure balancing cleverness when the
2257 * system is close to OOM, scan both anon and file equally
2258 * (unless the swappiness setting disagrees with swapping).
2260 if (!sc
->priority
&& swappiness
) {
2261 scan_balance
= SCAN_EQUAL
;
2266 * Prevent the reclaimer from falling into the cache trap: as
2267 * cache pages start out inactive, every cache fault will tip
2268 * the scan balance towards the file LRU. And as the file LRU
2269 * shrinks, so does the window for rotation from references.
2270 * This means we have a runaway feedback loop where a tiny
2271 * thrashing file LRU becomes infinitely more attractive than
2272 * anon pages. Try to detect this based on file LRU size.
2274 if (global_reclaim(sc
)) {
2275 unsigned long pgdatfile
;
2276 unsigned long pgdatfree
;
2278 unsigned long total_high_wmark
= 0;
2280 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2281 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2282 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2284 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2285 struct zone
*zone
= &pgdat
->node_zones
[z
];
2286 if (!managed_zone(zone
))
2289 total_high_wmark
+= high_wmark_pages(zone
);
2292 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2294 * Force SCAN_ANON if there are enough inactive
2295 * anonymous pages on the LRU in eligible zones.
2296 * Otherwise, the small LRU gets thrashed.
2298 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2299 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2301 scan_balance
= SCAN_ANON
;
2308 * If there is enough inactive page cache, i.e. if the size of the
2309 * inactive list is greater than that of the active list *and* the
2310 * inactive list actually has some pages to scan on this priority, we
2311 * do not reclaim anything from the anonymous working set right now.
2312 * Without the second condition we could end up never scanning an
2313 * lruvec even if it has plenty of old anonymous pages unless the
2314 * system is under heavy pressure.
2316 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2317 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2318 scan_balance
= SCAN_FILE
;
2322 scan_balance
= SCAN_FRACT
;
2325 * With swappiness at 100, anonymous and file have the same priority.
2326 * This scanning priority is essentially the inverse of IO cost.
2328 anon_prio
= swappiness
;
2329 file_prio
= 200 - anon_prio
;
2332 * OK, so we have swap space and a fair amount of page cache
2333 * pages. We use the recently rotated / recently scanned
2334 * ratios to determine how valuable each cache is.
2336 * Because workloads change over time (and to avoid overflow)
2337 * we keep these statistics as a floating average, which ends
2338 * up weighing recent references more than old ones.
2340 * anon in [0], file in [1]
2343 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2344 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2345 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2346 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2348 spin_lock_irq(&pgdat
->lru_lock
);
2349 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2350 reclaim_stat
->recent_scanned
[0] /= 2;
2351 reclaim_stat
->recent_rotated
[0] /= 2;
2354 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2355 reclaim_stat
->recent_scanned
[1] /= 2;
2356 reclaim_stat
->recent_rotated
[1] /= 2;
2360 * The amount of pressure on anon vs file pages is inversely
2361 * proportional to the fraction of recently scanned pages on
2362 * each list that were recently referenced and in active use.
2364 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2365 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2367 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2368 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2369 spin_unlock_irq(&pgdat
->lru_lock
);
2373 denominator
= ap
+ fp
+ 1;
2376 for_each_evictable_lru(lru
) {
2377 int file
= is_file_lru(lru
);
2381 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2382 scan
= size
>> sc
->priority
;
2384 * If the cgroup's already been deleted, make sure to
2385 * scrape out the remaining cache.
2387 if (!scan
&& !mem_cgroup_online(memcg
))
2388 scan
= min(size
, SWAP_CLUSTER_MAX
);
2390 switch (scan_balance
) {
2392 /* Scan lists relative to size */
2396 * Scan types proportional to swappiness and
2397 * their relative recent reclaim efficiency.
2399 scan
= div64_u64(scan
* fraction
[file
],
2404 /* Scan one type exclusively */
2405 if ((scan_balance
== SCAN_FILE
) != file
) {
2411 /* Look ma, no brain */
2421 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2423 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2424 struct scan_control
*sc
, unsigned long *lru_pages
)
2426 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2427 unsigned long nr
[NR_LRU_LISTS
];
2428 unsigned long targets
[NR_LRU_LISTS
];
2429 unsigned long nr_to_scan
;
2431 unsigned long nr_reclaimed
= 0;
2432 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2433 struct blk_plug plug
;
2436 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2438 /* Record the original scan target for proportional adjustments later */
2439 memcpy(targets
, nr
, sizeof(nr
));
2442 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2443 * event that can occur when there is little memory pressure e.g.
2444 * multiple streaming readers/writers. Hence, we do not abort scanning
2445 * when the requested number of pages are reclaimed when scanning at
2446 * DEF_PRIORITY on the assumption that the fact we are direct
2447 * reclaiming implies that kswapd is not keeping up and it is best to
2448 * do a batch of work at once. For memcg reclaim one check is made to
2449 * abort proportional reclaim if either the file or anon lru has already
2450 * dropped to zero at the first pass.
2452 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2453 sc
->priority
== DEF_PRIORITY
);
2455 blk_start_plug(&plug
);
2456 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2457 nr
[LRU_INACTIVE_FILE
]) {
2458 unsigned long nr_anon
, nr_file
, percentage
;
2459 unsigned long nr_scanned
;
2461 for_each_evictable_lru(lru
) {
2463 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2464 nr
[lru
] -= nr_to_scan
;
2466 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2473 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2477 * For kswapd and memcg, reclaim at least the number of pages
2478 * requested. Ensure that the anon and file LRUs are scanned
2479 * proportionally what was requested by get_scan_count(). We
2480 * stop reclaiming one LRU and reduce the amount scanning
2481 * proportional to the original scan target.
2483 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2484 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2487 * It's just vindictive to attack the larger once the smaller
2488 * has gone to zero. And given the way we stop scanning the
2489 * smaller below, this makes sure that we only make one nudge
2490 * towards proportionality once we've got nr_to_reclaim.
2492 if (!nr_file
|| !nr_anon
)
2495 if (nr_file
> nr_anon
) {
2496 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2497 targets
[LRU_ACTIVE_ANON
] + 1;
2499 percentage
= nr_anon
* 100 / scan_target
;
2501 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2502 targets
[LRU_ACTIVE_FILE
] + 1;
2504 percentage
= nr_file
* 100 / scan_target
;
2507 /* Stop scanning the smaller of the LRU */
2509 nr
[lru
+ LRU_ACTIVE
] = 0;
2512 * Recalculate the other LRU scan count based on its original
2513 * scan target and the percentage scanning already complete
2515 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2516 nr_scanned
= targets
[lru
] - nr
[lru
];
2517 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2518 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2521 nr_scanned
= targets
[lru
] - nr
[lru
];
2522 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2523 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2525 scan_adjusted
= true;
2527 blk_finish_plug(&plug
);
2528 sc
->nr_reclaimed
+= nr_reclaimed
;
2531 * Even if we did not try to evict anon pages at all, we want to
2532 * rebalance the anon lru active/inactive ratio.
2534 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2535 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2536 sc
, LRU_ACTIVE_ANON
);
2539 /* Use reclaim/compaction for costly allocs or under memory pressure */
2540 static bool in_reclaim_compaction(struct scan_control
*sc
)
2542 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2543 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2544 sc
->priority
< DEF_PRIORITY
- 2))
2551 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2552 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2553 * true if more pages should be reclaimed such that when the page allocator
2554 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2555 * It will give up earlier than that if there is difficulty reclaiming pages.
2557 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2558 unsigned long nr_reclaimed
,
2559 unsigned long nr_scanned
,
2560 struct scan_control
*sc
)
2562 unsigned long pages_for_compaction
;
2563 unsigned long inactive_lru_pages
;
2566 /* If not in reclaim/compaction mode, stop */
2567 if (!in_reclaim_compaction(sc
))
2570 /* Consider stopping depending on scan and reclaim activity */
2571 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2573 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2574 * full LRU list has been scanned and we are still failing
2575 * to reclaim pages. This full LRU scan is potentially
2576 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2578 if (!nr_reclaimed
&& !nr_scanned
)
2582 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2583 * fail without consequence, stop if we failed to reclaim
2584 * any pages from the last SWAP_CLUSTER_MAX number of
2585 * pages that were scanned. This will return to the
2586 * caller faster at the risk reclaim/compaction and
2587 * the resulting allocation attempt fails
2594 * If we have not reclaimed enough pages for compaction and the
2595 * inactive lists are large enough, continue reclaiming
2597 pages_for_compaction
= compact_gap(sc
->order
);
2598 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2599 if (get_nr_swap_pages() > 0)
2600 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2601 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2602 inactive_lru_pages
> pages_for_compaction
)
2605 /* If compaction would go ahead or the allocation would succeed, stop */
2606 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2607 struct zone
*zone
= &pgdat
->node_zones
[z
];
2608 if (!managed_zone(zone
))
2611 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2612 case COMPACT_SUCCESS
:
2613 case COMPACT_CONTINUE
:
2616 /* check next zone */
2623 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2625 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2626 unsigned long nr_reclaimed
, nr_scanned
;
2627 bool reclaimable
= false;
2630 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2631 struct mem_cgroup_reclaim_cookie reclaim
= {
2633 .priority
= sc
->priority
,
2635 unsigned long node_lru_pages
= 0;
2636 struct mem_cgroup
*memcg
;
2638 nr_reclaimed
= sc
->nr_reclaimed
;
2639 nr_scanned
= sc
->nr_scanned
;
2641 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2643 unsigned long lru_pages
;
2644 unsigned long reclaimed
;
2645 unsigned long scanned
;
2647 if (mem_cgroup_low(root
, memcg
)) {
2648 if (!sc
->memcg_low_reclaim
) {
2649 sc
->memcg_low_skipped
= 1;
2652 mem_cgroup_event(memcg
, MEMCG_LOW
);
2655 reclaimed
= sc
->nr_reclaimed
;
2656 scanned
= sc
->nr_scanned
;
2658 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2659 node_lru_pages
+= lru_pages
;
2662 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2663 memcg
, sc
->nr_scanned
- scanned
,
2666 /* Record the group's reclaim efficiency */
2667 vmpressure(sc
->gfp_mask
, memcg
, false,
2668 sc
->nr_scanned
- scanned
,
2669 sc
->nr_reclaimed
- reclaimed
);
2672 * Direct reclaim and kswapd have to scan all memory
2673 * cgroups to fulfill the overall scan target for the
2676 * Limit reclaim, on the other hand, only cares about
2677 * nr_to_reclaim pages to be reclaimed and it will
2678 * retry with decreasing priority if one round over the
2679 * whole hierarchy is not sufficient.
2681 if (!global_reclaim(sc
) &&
2682 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2683 mem_cgroup_iter_break(root
, memcg
);
2686 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2689 * Shrink the slab caches in the same proportion that
2690 * the eligible LRU pages were scanned.
2692 if (global_reclaim(sc
))
2693 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2694 sc
->nr_scanned
- nr_scanned
,
2697 if (reclaim_state
) {
2698 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2699 reclaim_state
->reclaimed_slab
= 0;
2702 /* Record the subtree's reclaim efficiency */
2703 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2704 sc
->nr_scanned
- nr_scanned
,
2705 sc
->nr_reclaimed
- nr_reclaimed
);
2707 if (sc
->nr_reclaimed
- nr_reclaimed
)
2710 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2711 sc
->nr_scanned
- nr_scanned
, sc
));
2714 * Kswapd gives up on balancing particular nodes after too
2715 * many failures to reclaim anything from them and goes to
2716 * sleep. On reclaim progress, reset the failure counter. A
2717 * successful direct reclaim run will revive a dormant kswapd.
2720 pgdat
->kswapd_failures
= 0;
2726 * Returns true if compaction should go ahead for a costly-order request, or
2727 * the allocation would already succeed without compaction. Return false if we
2728 * should reclaim first.
2730 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2732 unsigned long watermark
;
2733 enum compact_result suitable
;
2735 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2736 if (suitable
== COMPACT_SUCCESS
)
2737 /* Allocation should succeed already. Don't reclaim. */
2739 if (suitable
== COMPACT_SKIPPED
)
2740 /* Compaction cannot yet proceed. Do reclaim. */
2744 * Compaction is already possible, but it takes time to run and there
2745 * are potentially other callers using the pages just freed. So proceed
2746 * with reclaim to make a buffer of free pages available to give
2747 * compaction a reasonable chance of completing and allocating the page.
2748 * Note that we won't actually reclaim the whole buffer in one attempt
2749 * as the target watermark in should_continue_reclaim() is lower. But if
2750 * we are already above the high+gap watermark, don't reclaim at all.
2752 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2754 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2758 * This is the direct reclaim path, for page-allocating processes. We only
2759 * try to reclaim pages from zones which will satisfy the caller's allocation
2762 * If a zone is deemed to be full of pinned pages then just give it a light
2763 * scan then give up on it.
2765 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2769 unsigned long nr_soft_reclaimed
;
2770 unsigned long nr_soft_scanned
;
2772 pg_data_t
*last_pgdat
= NULL
;
2775 * If the number of buffer_heads in the machine exceeds the maximum
2776 * allowed level, force direct reclaim to scan the highmem zone as
2777 * highmem pages could be pinning lowmem pages storing buffer_heads
2779 orig_mask
= sc
->gfp_mask
;
2780 if (buffer_heads_over_limit
) {
2781 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2782 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2785 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2786 sc
->reclaim_idx
, sc
->nodemask
) {
2788 * Take care memory controller reclaiming has small influence
2791 if (global_reclaim(sc
)) {
2792 if (!cpuset_zone_allowed(zone
,
2793 GFP_KERNEL
| __GFP_HARDWALL
))
2797 * If we already have plenty of memory free for
2798 * compaction in this zone, don't free any more.
2799 * Even though compaction is invoked for any
2800 * non-zero order, only frequent costly order
2801 * reclamation is disruptive enough to become a
2802 * noticeable problem, like transparent huge
2805 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2806 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2807 compaction_ready(zone
, sc
)) {
2808 sc
->compaction_ready
= true;
2813 * Shrink each node in the zonelist once. If the
2814 * zonelist is ordered by zone (not the default) then a
2815 * node may be shrunk multiple times but in that case
2816 * the user prefers lower zones being preserved.
2818 if (zone
->zone_pgdat
== last_pgdat
)
2822 * This steals pages from memory cgroups over softlimit
2823 * and returns the number of reclaimed pages and
2824 * scanned pages. This works for global memory pressure
2825 * and balancing, not for a memcg's limit.
2827 nr_soft_scanned
= 0;
2828 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2829 sc
->order
, sc
->gfp_mask
,
2831 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2832 sc
->nr_scanned
+= nr_soft_scanned
;
2833 /* need some check for avoid more shrink_zone() */
2836 /* See comment about same check for global reclaim above */
2837 if (zone
->zone_pgdat
== last_pgdat
)
2839 last_pgdat
= zone
->zone_pgdat
;
2840 shrink_node(zone
->zone_pgdat
, sc
);
2844 * Restore to original mask to avoid the impact on the caller if we
2845 * promoted it to __GFP_HIGHMEM.
2847 sc
->gfp_mask
= orig_mask
;
2850 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2852 struct mem_cgroup
*memcg
;
2854 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2856 unsigned long refaults
;
2857 struct lruvec
*lruvec
;
2860 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2862 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2864 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2865 lruvec
->refaults
= refaults
;
2866 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2870 * This is the main entry point to direct page reclaim.
2872 * If a full scan of the inactive list fails to free enough memory then we
2873 * are "out of memory" and something needs to be killed.
2875 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2876 * high - the zone may be full of dirty or under-writeback pages, which this
2877 * caller can't do much about. We kick the writeback threads and take explicit
2878 * naps in the hope that some of these pages can be written. But if the
2879 * allocating task holds filesystem locks which prevent writeout this might not
2880 * work, and the allocation attempt will fail.
2882 * returns: 0, if no pages reclaimed
2883 * else, the number of pages reclaimed
2885 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2886 struct scan_control
*sc
)
2888 int initial_priority
= sc
->priority
;
2889 pg_data_t
*last_pgdat
;
2893 delayacct_freepages_start();
2895 if (global_reclaim(sc
))
2896 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2899 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2902 shrink_zones(zonelist
, sc
);
2904 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2907 if (sc
->compaction_ready
)
2911 * If we're getting trouble reclaiming, start doing
2912 * writepage even in laptop mode.
2914 if (sc
->priority
< DEF_PRIORITY
- 2)
2915 sc
->may_writepage
= 1;
2916 } while (--sc
->priority
>= 0);
2919 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2921 if (zone
->zone_pgdat
== last_pgdat
)
2923 last_pgdat
= zone
->zone_pgdat
;
2924 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2927 delayacct_freepages_end();
2929 if (sc
->nr_reclaimed
)
2930 return sc
->nr_reclaimed
;
2932 /* Aborted reclaim to try compaction? don't OOM, then */
2933 if (sc
->compaction_ready
)
2936 /* Untapped cgroup reserves? Don't OOM, retry. */
2937 if (sc
->memcg_low_skipped
) {
2938 sc
->priority
= initial_priority
;
2939 sc
->memcg_low_reclaim
= 1;
2940 sc
->memcg_low_skipped
= 0;
2947 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2950 unsigned long pfmemalloc_reserve
= 0;
2951 unsigned long free_pages
= 0;
2955 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2958 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2959 zone
= &pgdat
->node_zones
[i
];
2960 if (!managed_zone(zone
))
2963 if (!zone_reclaimable_pages(zone
))
2966 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2967 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2970 /* If there are no reserves (unexpected config) then do not throttle */
2971 if (!pfmemalloc_reserve
)
2974 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2976 /* kswapd must be awake if processes are being throttled */
2977 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2978 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2979 (enum zone_type
)ZONE_NORMAL
);
2980 wake_up_interruptible(&pgdat
->kswapd_wait
);
2987 * Throttle direct reclaimers if backing storage is backed by the network
2988 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2989 * depleted. kswapd will continue to make progress and wake the processes
2990 * when the low watermark is reached.
2992 * Returns true if a fatal signal was delivered during throttling. If this
2993 * happens, the page allocator should not consider triggering the OOM killer.
2995 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2996 nodemask_t
*nodemask
)
3000 pg_data_t
*pgdat
= NULL
;
3003 * Kernel threads should not be throttled as they may be indirectly
3004 * responsible for cleaning pages necessary for reclaim to make forward
3005 * progress. kjournald for example may enter direct reclaim while
3006 * committing a transaction where throttling it could forcing other
3007 * processes to block on log_wait_commit().
3009 if (current
->flags
& PF_KTHREAD
)
3013 * If a fatal signal is pending, this process should not throttle.
3014 * It should return quickly so it can exit and free its memory
3016 if (fatal_signal_pending(current
))
3020 * Check if the pfmemalloc reserves are ok by finding the first node
3021 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3022 * GFP_KERNEL will be required for allocating network buffers when
3023 * swapping over the network so ZONE_HIGHMEM is unusable.
3025 * Throttling is based on the first usable node and throttled processes
3026 * wait on a queue until kswapd makes progress and wakes them. There
3027 * is an affinity then between processes waking up and where reclaim
3028 * progress has been made assuming the process wakes on the same node.
3029 * More importantly, processes running on remote nodes will not compete
3030 * for remote pfmemalloc reserves and processes on different nodes
3031 * should make reasonable progress.
3033 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
3034 gfp_zone(gfp_mask
), nodemask
) {
3035 if (zone_idx(zone
) > ZONE_NORMAL
)
3038 /* Throttle based on the first usable node */
3039 pgdat
= zone
->zone_pgdat
;
3040 if (allow_direct_reclaim(pgdat
))
3045 /* If no zone was usable by the allocation flags then do not throttle */
3049 /* Account for the throttling */
3050 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
3053 * If the caller cannot enter the filesystem, it's possible that it
3054 * is due to the caller holding an FS lock or performing a journal
3055 * transaction in the case of a filesystem like ext[3|4]. In this case,
3056 * it is not safe to block on pfmemalloc_wait as kswapd could be
3057 * blocked waiting on the same lock. Instead, throttle for up to a
3058 * second before continuing.
3060 if (!(gfp_mask
& __GFP_FS
)) {
3061 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3062 allow_direct_reclaim(pgdat
), HZ
);
3067 /* Throttle until kswapd wakes the process */
3068 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3069 allow_direct_reclaim(pgdat
));
3072 if (fatal_signal_pending(current
))
3079 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3080 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3082 unsigned long nr_reclaimed
;
3083 struct scan_control sc
= {
3084 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3085 .gfp_mask
= current_gfp_context(gfp_mask
),
3086 .reclaim_idx
= gfp_zone(gfp_mask
),
3088 .nodemask
= nodemask
,
3089 .priority
= DEF_PRIORITY
,
3090 .may_writepage
= !laptop_mode
,
3096 * Do not enter reclaim if fatal signal was delivered while throttled.
3097 * 1 is returned so that the page allocator does not OOM kill at this
3100 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3103 trace_mm_vmscan_direct_reclaim_begin(order
,
3108 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3110 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3112 return nr_reclaimed
;
3117 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3118 gfp_t gfp_mask
, bool noswap
,
3120 unsigned long *nr_scanned
)
3122 struct scan_control sc
= {
3123 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3124 .target_mem_cgroup
= memcg
,
3125 .may_writepage
= !laptop_mode
,
3127 .reclaim_idx
= MAX_NR_ZONES
- 1,
3128 .may_swap
= !noswap
,
3130 unsigned long lru_pages
;
3132 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3133 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3135 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3141 * NOTE: Although we can get the priority field, using it
3142 * here is not a good idea, since it limits the pages we can scan.
3143 * if we don't reclaim here, the shrink_node from balance_pgdat
3144 * will pick up pages from other mem cgroup's as well. We hack
3145 * the priority and make it zero.
3147 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3149 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3151 *nr_scanned
= sc
.nr_scanned
;
3152 return sc
.nr_reclaimed
;
3155 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3156 unsigned long nr_pages
,
3160 struct zonelist
*zonelist
;
3161 unsigned long nr_reclaimed
;
3163 unsigned int noreclaim_flag
;
3164 struct scan_control sc
= {
3165 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3166 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3167 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3168 .reclaim_idx
= MAX_NR_ZONES
- 1,
3169 .target_mem_cgroup
= memcg
,
3170 .priority
= DEF_PRIORITY
,
3171 .may_writepage
= !laptop_mode
,
3173 .may_swap
= may_swap
,
3177 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3178 * take care of from where we get pages. So the node where we start the
3179 * scan does not need to be the current node.
3181 nid
= mem_cgroup_select_victim_node(memcg
);
3183 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3185 trace_mm_vmscan_memcg_reclaim_begin(0,
3190 noreclaim_flag
= memalloc_noreclaim_save();
3191 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3192 memalloc_noreclaim_restore(noreclaim_flag
);
3194 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3196 return nr_reclaimed
;
3200 static void age_active_anon(struct pglist_data
*pgdat
,
3201 struct scan_control
*sc
)
3203 struct mem_cgroup
*memcg
;
3205 if (!total_swap_pages
)
3208 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3210 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3212 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3213 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3214 sc
, LRU_ACTIVE_ANON
);
3216 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3221 * Returns true if there is an eligible zone balanced for the request order
3224 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3227 unsigned long mark
= -1;
3230 for (i
= 0; i
<= classzone_idx
; i
++) {
3231 zone
= pgdat
->node_zones
+ i
;
3233 if (!managed_zone(zone
))
3236 mark
= high_wmark_pages(zone
);
3237 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3242 * If a node has no populated zone within classzone_idx, it does not
3243 * need balancing by definition. This can happen if a zone-restricted
3244 * allocation tries to wake a remote kswapd.
3252 /* Clear pgdat state for congested, dirty or under writeback. */
3253 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3255 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3256 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3257 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3261 * Prepare kswapd for sleeping. This verifies that there are no processes
3262 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3264 * Returns true if kswapd is ready to sleep
3266 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3269 * The throttled processes are normally woken up in balance_pgdat() as
3270 * soon as allow_direct_reclaim() is true. But there is a potential
3271 * race between when kswapd checks the watermarks and a process gets
3272 * throttled. There is also a potential race if processes get
3273 * throttled, kswapd wakes, a large process exits thereby balancing the
3274 * zones, which causes kswapd to exit balance_pgdat() before reaching
3275 * the wake up checks. If kswapd is going to sleep, no process should
3276 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3277 * the wake up is premature, processes will wake kswapd and get
3278 * throttled again. The difference from wake ups in balance_pgdat() is
3279 * that here we are under prepare_to_wait().
3281 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3282 wake_up_all(&pgdat
->pfmemalloc_wait
);
3284 /* Hopeless node, leave it to direct reclaim */
3285 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3288 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3289 clear_pgdat_congested(pgdat
);
3297 * kswapd shrinks a node of pages that are at or below the highest usable
3298 * zone that is currently unbalanced.
3300 * Returns true if kswapd scanned at least the requested number of pages to
3301 * reclaim or if the lack of progress was due to pages under writeback.
3302 * This is used to determine if the scanning priority needs to be raised.
3304 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3305 struct scan_control
*sc
)
3310 /* Reclaim a number of pages proportional to the number of zones */
3311 sc
->nr_to_reclaim
= 0;
3312 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3313 zone
= pgdat
->node_zones
+ z
;
3314 if (!managed_zone(zone
))
3317 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3321 * Historically care was taken to put equal pressure on all zones but
3322 * now pressure is applied based on node LRU order.
3324 shrink_node(pgdat
, sc
);
3327 * Fragmentation may mean that the system cannot be rebalanced for
3328 * high-order allocations. If twice the allocation size has been
3329 * reclaimed then recheck watermarks only at order-0 to prevent
3330 * excessive reclaim. Assume that a process requested a high-order
3331 * can direct reclaim/compact.
3333 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3336 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3340 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3341 * that are eligible for use by the caller until at least one zone is
3344 * Returns the order kswapd finished reclaiming at.
3346 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3347 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3348 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3349 * or lower is eligible for reclaim until at least one usable zone is
3352 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3355 unsigned long nr_soft_reclaimed
;
3356 unsigned long nr_soft_scanned
;
3358 struct scan_control sc
= {
3359 .gfp_mask
= GFP_KERNEL
,
3361 .priority
= DEF_PRIORITY
,
3362 .may_writepage
= !laptop_mode
,
3366 count_vm_event(PAGEOUTRUN
);
3369 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3370 bool raise_priority
= true;
3372 sc
.reclaim_idx
= classzone_idx
;
3375 * If the number of buffer_heads exceeds the maximum allowed
3376 * then consider reclaiming from all zones. This has a dual
3377 * purpose -- on 64-bit systems it is expected that
3378 * buffer_heads are stripped during active rotation. On 32-bit
3379 * systems, highmem pages can pin lowmem memory and shrinking
3380 * buffers can relieve lowmem pressure. Reclaim may still not
3381 * go ahead if all eligible zones for the original allocation
3382 * request are balanced to avoid excessive reclaim from kswapd.
3384 if (buffer_heads_over_limit
) {
3385 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3386 zone
= pgdat
->node_zones
+ i
;
3387 if (!managed_zone(zone
))
3396 * Only reclaim if there are no eligible zones. Note that
3397 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3400 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3404 * Do some background aging of the anon list, to give
3405 * pages a chance to be referenced before reclaiming. All
3406 * pages are rotated regardless of classzone as this is
3407 * about consistent aging.
3409 age_active_anon(pgdat
, &sc
);
3412 * If we're getting trouble reclaiming, start doing writepage
3413 * even in laptop mode.
3415 if (sc
.priority
< DEF_PRIORITY
- 2)
3416 sc
.may_writepage
= 1;
3418 /* Call soft limit reclaim before calling shrink_node. */
3420 nr_soft_scanned
= 0;
3421 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3422 sc
.gfp_mask
, &nr_soft_scanned
);
3423 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3426 * There should be no need to raise the scanning priority if
3427 * enough pages are already being scanned that that high
3428 * watermark would be met at 100% efficiency.
3430 if (kswapd_shrink_node(pgdat
, &sc
))
3431 raise_priority
= false;
3434 * If the low watermark is met there is no need for processes
3435 * to be throttled on pfmemalloc_wait as they should not be
3436 * able to safely make forward progress. Wake them
3438 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3439 allow_direct_reclaim(pgdat
))
3440 wake_up_all(&pgdat
->pfmemalloc_wait
);
3442 /* Check if kswapd should be suspending */
3443 if (try_to_freeze() || kthread_should_stop())
3447 * Raise priority if scanning rate is too low or there was no
3448 * progress in reclaiming pages
3450 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3451 if (raise_priority
|| !nr_reclaimed
)
3453 } while (sc
.priority
>= 1);
3455 if (!sc
.nr_reclaimed
)
3456 pgdat
->kswapd_failures
++;
3459 snapshot_refaults(NULL
, pgdat
);
3461 * Return the order kswapd stopped reclaiming at as
3462 * prepare_kswapd_sleep() takes it into account. If another caller
3463 * entered the allocator slow path while kswapd was awake, order will
3464 * remain at the higher level.
3470 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3471 * allocation request woke kswapd for. When kswapd has not woken recently,
3472 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3473 * given classzone and returns it or the highest classzone index kswapd
3474 * was recently woke for.
3476 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3477 enum zone_type classzone_idx
)
3479 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3480 return classzone_idx
;
3482 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3485 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3486 unsigned int classzone_idx
)
3491 if (freezing(current
) || kthread_should_stop())
3494 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3497 * Try to sleep for a short interval. Note that kcompactd will only be
3498 * woken if it is possible to sleep for a short interval. This is
3499 * deliberate on the assumption that if reclaim cannot keep an
3500 * eligible zone balanced that it's also unlikely that compaction will
3503 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3505 * Compaction records what page blocks it recently failed to
3506 * isolate pages from and skips them in the future scanning.
3507 * When kswapd is going to sleep, it is reasonable to assume
3508 * that pages and compaction may succeed so reset the cache.
3510 reset_isolation_suitable(pgdat
);
3513 * We have freed the memory, now we should compact it to make
3514 * allocation of the requested order possible.
3516 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3518 remaining
= schedule_timeout(HZ
/10);
3521 * If woken prematurely then reset kswapd_classzone_idx and
3522 * order. The values will either be from a wakeup request or
3523 * the previous request that slept prematurely.
3526 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3527 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3530 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3531 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3535 * After a short sleep, check if it was a premature sleep. If not, then
3536 * go fully to sleep until explicitly woken up.
3539 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3540 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3543 * vmstat counters are not perfectly accurate and the estimated
3544 * value for counters such as NR_FREE_PAGES can deviate from the
3545 * true value by nr_online_cpus * threshold. To avoid the zone
3546 * watermarks being breached while under pressure, we reduce the
3547 * per-cpu vmstat threshold while kswapd is awake and restore
3548 * them before going back to sleep.
3550 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3552 if (!kthread_should_stop())
3555 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3558 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3560 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3562 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3566 * The background pageout daemon, started as a kernel thread
3567 * from the init process.
3569 * This basically trickles out pages so that we have _some_
3570 * free memory available even if there is no other activity
3571 * that frees anything up. This is needed for things like routing
3572 * etc, where we otherwise might have all activity going on in
3573 * asynchronous contexts that cannot page things out.
3575 * If there are applications that are active memory-allocators
3576 * (most normal use), this basically shouldn't matter.
3578 static int kswapd(void *p
)
3580 unsigned int alloc_order
, reclaim_order
;
3581 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3582 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3583 struct task_struct
*tsk
= current
;
3585 struct reclaim_state reclaim_state
= {
3586 .reclaimed_slab
= 0,
3588 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3590 if (!cpumask_empty(cpumask
))
3591 set_cpus_allowed_ptr(tsk
, cpumask
);
3592 current
->reclaim_state
= &reclaim_state
;
3595 * Tell the memory management that we're a "memory allocator",
3596 * and that if we need more memory we should get access to it
3597 * regardless (see "__alloc_pages()"). "kswapd" should
3598 * never get caught in the normal page freeing logic.
3600 * (Kswapd normally doesn't need memory anyway, but sometimes
3601 * you need a small amount of memory in order to be able to
3602 * page out something else, and this flag essentially protects
3603 * us from recursively trying to free more memory as we're
3604 * trying to free the first piece of memory in the first place).
3606 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3609 pgdat
->kswapd_order
= 0;
3610 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3614 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3615 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3618 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3621 /* Read the new order and classzone_idx */
3622 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3623 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3624 pgdat
->kswapd_order
= 0;
3625 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3627 ret
= try_to_freeze();
3628 if (kthread_should_stop())
3632 * We can speed up thawing tasks if we don't call balance_pgdat
3633 * after returning from the refrigerator
3639 * Reclaim begins at the requested order but if a high-order
3640 * reclaim fails then kswapd falls back to reclaiming for
3641 * order-0. If that happens, kswapd will consider sleeping
3642 * for the order it finished reclaiming at (reclaim_order)
3643 * but kcompactd is woken to compact for the original
3644 * request (alloc_order).
3646 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3648 fs_reclaim_acquire(GFP_KERNEL
);
3649 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3650 fs_reclaim_release(GFP_KERNEL
);
3651 if (reclaim_order
< alloc_order
)
3652 goto kswapd_try_sleep
;
3655 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3656 current
->reclaim_state
= NULL
;
3662 * A zone is low on free memory, so wake its kswapd task to service it.
3664 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3668 if (!managed_zone(zone
))
3671 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3673 pgdat
= zone
->zone_pgdat
;
3674 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3676 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3677 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3680 /* Hopeless node, leave it to direct reclaim */
3681 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3684 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3687 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3688 wake_up_interruptible(&pgdat
->kswapd_wait
);
3691 #ifdef CONFIG_HIBERNATION
3693 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3696 * Rather than trying to age LRUs the aim is to preserve the overall
3697 * LRU order by reclaiming preferentially
3698 * inactive > active > active referenced > active mapped
3700 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3702 struct reclaim_state reclaim_state
;
3703 struct scan_control sc
= {
3704 .nr_to_reclaim
= nr_to_reclaim
,
3705 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3706 .reclaim_idx
= MAX_NR_ZONES
- 1,
3707 .priority
= DEF_PRIORITY
,
3711 .hibernation_mode
= 1,
3713 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3714 struct task_struct
*p
= current
;
3715 unsigned long nr_reclaimed
;
3716 unsigned int noreclaim_flag
;
3718 noreclaim_flag
= memalloc_noreclaim_save();
3719 fs_reclaim_acquire(sc
.gfp_mask
);
3720 reclaim_state
.reclaimed_slab
= 0;
3721 p
->reclaim_state
= &reclaim_state
;
3723 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3725 p
->reclaim_state
= NULL
;
3726 fs_reclaim_release(sc
.gfp_mask
);
3727 memalloc_noreclaim_restore(noreclaim_flag
);
3729 return nr_reclaimed
;
3731 #endif /* CONFIG_HIBERNATION */
3733 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3734 not required for correctness. So if the last cpu in a node goes
3735 away, we get changed to run anywhere: as the first one comes back,
3736 restore their cpu bindings. */
3737 static int kswapd_cpu_online(unsigned int cpu
)
3741 for_each_node_state(nid
, N_MEMORY
) {
3742 pg_data_t
*pgdat
= NODE_DATA(nid
);
3743 const struct cpumask
*mask
;
3745 mask
= cpumask_of_node(pgdat
->node_id
);
3747 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3748 /* One of our CPUs online: restore mask */
3749 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3755 * This kswapd start function will be called by init and node-hot-add.
3756 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3758 int kswapd_run(int nid
)
3760 pg_data_t
*pgdat
= NODE_DATA(nid
);
3766 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3767 if (IS_ERR(pgdat
->kswapd
)) {
3768 /* failure at boot is fatal */
3769 BUG_ON(system_state
< SYSTEM_RUNNING
);
3770 pr_err("Failed to start kswapd on node %d\n", nid
);
3771 ret
= PTR_ERR(pgdat
->kswapd
);
3772 pgdat
->kswapd
= NULL
;
3778 * Called by memory hotplug when all memory in a node is offlined. Caller must
3779 * hold mem_hotplug_begin/end().
3781 void kswapd_stop(int nid
)
3783 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3786 kthread_stop(kswapd
);
3787 NODE_DATA(nid
)->kswapd
= NULL
;
3791 static int __init
kswapd_init(void)
3796 for_each_node_state(nid
, N_MEMORY
)
3798 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3799 "mm/vmscan:online", kswapd_cpu_online
,
3805 module_init(kswapd_init
)
3811 * If non-zero call node_reclaim when the number of free pages falls below
3814 int node_reclaim_mode __read_mostly
;
3816 #define RECLAIM_OFF 0
3817 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3818 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3819 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3822 * Priority for NODE_RECLAIM. This determines the fraction of pages
3823 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3826 #define NODE_RECLAIM_PRIORITY 4
3829 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3832 int sysctl_min_unmapped_ratio
= 1;
3835 * If the number of slab pages in a zone grows beyond this percentage then
3836 * slab reclaim needs to occur.
3838 int sysctl_min_slab_ratio
= 5;
3840 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3842 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3843 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3844 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3847 * It's possible for there to be more file mapped pages than
3848 * accounted for by the pages on the file LRU lists because
3849 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3851 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3854 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3855 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3857 unsigned long nr_pagecache_reclaimable
;
3858 unsigned long delta
= 0;
3861 * If RECLAIM_UNMAP is set, then all file pages are considered
3862 * potentially reclaimable. Otherwise, we have to worry about
3863 * pages like swapcache and node_unmapped_file_pages() provides
3866 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3867 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3869 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3871 /* If we can't clean pages, remove dirty pages from consideration */
3872 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3873 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3875 /* Watch for any possible underflows due to delta */
3876 if (unlikely(delta
> nr_pagecache_reclaimable
))
3877 delta
= nr_pagecache_reclaimable
;
3879 return nr_pagecache_reclaimable
- delta
;
3883 * Try to free up some pages from this node through reclaim.
3885 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3887 /* Minimum pages needed in order to stay on node */
3888 const unsigned long nr_pages
= 1 << order
;
3889 struct task_struct
*p
= current
;
3890 struct reclaim_state reclaim_state
;
3891 unsigned int noreclaim_flag
;
3892 struct scan_control sc
= {
3893 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3894 .gfp_mask
= current_gfp_context(gfp_mask
),
3896 .priority
= NODE_RECLAIM_PRIORITY
,
3897 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3898 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3900 .reclaim_idx
= gfp_zone(gfp_mask
),
3905 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3906 * and we also need to be able to write out pages for RECLAIM_WRITE
3907 * and RECLAIM_UNMAP.
3909 noreclaim_flag
= memalloc_noreclaim_save();
3910 p
->flags
|= PF_SWAPWRITE
;
3911 fs_reclaim_acquire(sc
.gfp_mask
);
3912 reclaim_state
.reclaimed_slab
= 0;
3913 p
->reclaim_state
= &reclaim_state
;
3915 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3917 * Free memory by calling shrink zone with increasing
3918 * priorities until we have enough memory freed.
3921 shrink_node(pgdat
, &sc
);
3922 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3925 p
->reclaim_state
= NULL
;
3926 fs_reclaim_release(gfp_mask
);
3927 current
->flags
&= ~PF_SWAPWRITE
;
3928 memalloc_noreclaim_restore(noreclaim_flag
);
3929 return sc
.nr_reclaimed
>= nr_pages
;
3932 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3937 * Node reclaim reclaims unmapped file backed pages and
3938 * slab pages if we are over the defined limits.
3940 * A small portion of unmapped file backed pages is needed for
3941 * file I/O otherwise pages read by file I/O will be immediately
3942 * thrown out if the node is overallocated. So we do not reclaim
3943 * if less than a specified percentage of the node is used by
3944 * unmapped file backed pages.
3946 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3947 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3948 return NODE_RECLAIM_FULL
;
3951 * Do not scan if the allocation should not be delayed.
3953 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3954 return NODE_RECLAIM_NOSCAN
;
3957 * Only run node reclaim on the local node or on nodes that do not
3958 * have associated processors. This will favor the local processor
3959 * over remote processors and spread off node memory allocations
3960 * as wide as possible.
3962 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3963 return NODE_RECLAIM_NOSCAN
;
3965 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3966 return NODE_RECLAIM_NOSCAN
;
3968 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3969 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3972 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3979 * page_evictable - test whether a page is evictable
3980 * @page: the page to test
3982 * Test whether page is evictable--i.e., should be placed on active/inactive
3983 * lists vs unevictable list.
3985 * Reasons page might not be evictable:
3986 * (1) page's mapping marked unevictable
3987 * (2) page is part of an mlocked VMA
3990 int page_evictable(struct page
*page
)
3994 /* Prevent address_space of inode and swap cache from being freed */
3996 ret
= !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
4003 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
4004 * @pages: array of pages to check
4005 * @nr_pages: number of pages to check
4007 * Checks pages for evictability and moves them to the appropriate lru list.
4009 * This function is only used for SysV IPC SHM_UNLOCK.
4011 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
4013 struct lruvec
*lruvec
;
4014 struct pglist_data
*pgdat
= NULL
;
4019 for (i
= 0; i
< nr_pages
; i
++) {
4020 struct page
*page
= pages
[i
];
4021 struct pglist_data
*pagepgdat
= page_pgdat(page
);
4024 if (pagepgdat
!= pgdat
) {
4026 spin_unlock_irq(&pgdat
->lru_lock
);
4028 spin_lock_irq(&pgdat
->lru_lock
);
4030 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
4032 if (!PageLRU(page
) || !PageUnevictable(page
))
4035 if (page_evictable(page
)) {
4036 enum lru_list lru
= page_lru_base_type(page
);
4038 VM_BUG_ON_PAGE(PageActive(page
), page
);
4039 ClearPageUnevictable(page
);
4040 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
4041 add_page_to_lru_list(page
, lruvec
, lru
);
4047 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
4048 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
4049 spin_unlock_irq(&pgdat
->lru_lock
);
4052 #endif /* CONFIG_SHMEM */