4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/sched/mm.h>
18 #include <linux/module.h>
19 #include <linux/gfp.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/vmpressure.h>
26 #include <linux/vmstat.h>
27 #include <linux/file.h>
28 #include <linux/writeback.h>
29 #include <linux/blkdev.h>
30 #include <linux/buffer_head.h> /* for try_to_release_page(),
31 buffer_heads_over_limit */
32 #include <linux/mm_inline.h>
33 #include <linux/backing-dev.h>
34 #include <linux/rmap.h>
35 #include <linux/topology.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/compaction.h>
39 #include <linux/notifier.h>
40 #include <linux/rwsem.h>
41 #include <linux/delay.h>
42 #include <linux/kthread.h>
43 #include <linux/freezer.h>
44 #include <linux/memcontrol.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
55 #include <linux/swapops.h>
56 #include <linux/balloon_compaction.h>
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/vmscan.h>
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim
;
67 /* This context's GFP mask */
70 /* Allocation order */
74 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * The memory cgroup that hit its limit and as a result is the
81 * primary target of this reclaim invocation.
83 struct mem_cgroup
*target_mem_cgroup
;
85 /* Scan (total_size >> priority) pages at once */
88 /* The highest zone to isolate pages for reclaim from */
89 enum zone_type reclaim_idx
;
91 /* Writepage batching in laptop mode; RECLAIM_WRITE */
92 unsigned int may_writepage
:1;
94 /* Can mapped pages be reclaimed? */
95 unsigned int may_unmap
:1;
97 /* Can pages be swapped as part of reclaim? */
98 unsigned int may_swap
:1;
101 * Cgroups are not reclaimed below their configured memory.low,
102 * unless we threaten to OOM. If any cgroups are skipped due to
103 * memory.low and nothing was reclaimed, go back for memory.low.
105 unsigned int memcg_low_reclaim
:1;
106 unsigned int memcg_low_skipped
:1;
108 unsigned int hibernation_mode
:1;
110 /* One of the zones is ready for compaction */
111 unsigned int compaction_ready
:1;
113 /* Incremented by the number of inactive pages that were scanned */
114 unsigned long nr_scanned
;
116 /* Number of pages freed so far during a call to shrink_zones() */
117 unsigned long nr_reclaimed
;
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness
= 60;
153 * The total number of pages which are beyond the high watermark within all
156 unsigned long vm_total_pages
;
158 static LIST_HEAD(shrinker_list
);
159 static DECLARE_RWSEM(shrinker_rwsem
);
162 static bool global_reclaim(struct scan_control
*sc
)
164 return !sc
->target_mem_cgroup
;
168 * sane_reclaim - is the usual dirty throttling mechanism operational?
169 * @sc: scan_control in question
171 * The normal page dirty throttling mechanism in balance_dirty_pages() is
172 * completely broken with the legacy memcg and direct stalling in
173 * shrink_page_list() is used for throttling instead, which lacks all the
174 * niceties such as fairness, adaptive pausing, bandwidth proportional
175 * allocation and configurability.
177 * This function tests whether the vmscan currently in progress can assume
178 * that the normal dirty throttling mechanism is operational.
180 static bool sane_reclaim(struct scan_control
*sc
)
182 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
186 #ifdef CONFIG_CGROUP_WRITEBACK
187 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
193 static bool global_reclaim(struct scan_control
*sc
)
198 static bool sane_reclaim(struct scan_control
*sc
)
205 * This misses isolated pages which are not accounted for to save counters.
206 * As the data only determines if reclaim or compaction continues, it is
207 * not expected that isolated pages will be a dominating factor.
209 unsigned long zone_reclaimable_pages(struct zone
*zone
)
213 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
214 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
215 if (get_nr_swap_pages() > 0)
216 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
217 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
222 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
226 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
227 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
228 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
230 if (get_nr_swap_pages() > 0)
231 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
232 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
233 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
239 * lruvec_lru_size - Returns the number of pages on the given LRU list.
240 * @lruvec: lru vector
242 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
, int zone_idx
)
246 unsigned long lru_size
;
249 if (!mem_cgroup_disabled())
250 lru_size
= mem_cgroup_get_lru_size(lruvec
, lru
);
252 lru_size
= node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
254 for (zid
= zone_idx
+ 1; zid
< MAX_NR_ZONES
; zid
++) {
255 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
258 if (!managed_zone(zone
))
261 if (!mem_cgroup_disabled())
262 size
= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
264 size
= zone_page_state(&lruvec_pgdat(lruvec
)->node_zones
[zid
],
265 NR_ZONE_LRU_BASE
+ lru
);
266 lru_size
-= min(size
, lru_size
);
274 * Add a shrinker callback to be called from the vm.
276 int register_shrinker(struct shrinker
*shrinker
)
278 size_t size
= sizeof(*shrinker
->nr_deferred
);
280 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
283 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
284 if (!shrinker
->nr_deferred
)
287 down_write(&shrinker_rwsem
);
288 list_add_tail(&shrinker
->list
, &shrinker_list
);
289 up_write(&shrinker_rwsem
);
292 EXPORT_SYMBOL(register_shrinker
);
297 void unregister_shrinker(struct shrinker
*shrinker
)
299 down_write(&shrinker_rwsem
);
300 list_del(&shrinker
->list
);
301 up_write(&shrinker_rwsem
);
302 kfree(shrinker
->nr_deferred
);
304 EXPORT_SYMBOL(unregister_shrinker
);
306 #define SHRINK_BATCH 128
308 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
309 struct shrinker
*shrinker
,
310 unsigned long nr_scanned
,
311 unsigned long nr_eligible
)
313 unsigned long freed
= 0;
314 unsigned long long delta
;
319 int nid
= shrinkctl
->nid
;
320 long batch_size
= shrinker
->batch
? shrinker
->batch
322 long scanned
= 0, next_deferred
;
324 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
329 * copy the current shrinker scan count into a local variable
330 * and zero it so that other concurrent shrinker invocations
331 * don't also do this scanning work.
333 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
336 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
338 do_div(delta
, nr_eligible
+ 1);
340 if (total_scan
< 0) {
341 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
342 shrinker
->scan_objects
, total_scan
);
343 total_scan
= freeable
;
346 next_deferred
= total_scan
;
349 * We need to avoid excessive windup on filesystem shrinkers
350 * due to large numbers of GFP_NOFS allocations causing the
351 * shrinkers to return -1 all the time. This results in a large
352 * nr being built up so when a shrink that can do some work
353 * comes along it empties the entire cache due to nr >>>
354 * freeable. This is bad for sustaining a working set in
357 * Hence only allow the shrinker to scan the entire cache when
358 * a large delta change is calculated directly.
360 if (delta
< freeable
/ 4)
361 total_scan
= min(total_scan
, freeable
/ 2);
364 * Avoid risking looping forever due to too large nr value:
365 * never try to free more than twice the estimate number of
368 if (total_scan
> freeable
* 2)
369 total_scan
= freeable
* 2;
371 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
372 nr_scanned
, nr_eligible
,
373 freeable
, delta
, total_scan
);
376 * Normally, we should not scan less than batch_size objects in one
377 * pass to avoid too frequent shrinker calls, but if the slab has less
378 * than batch_size objects in total and we are really tight on memory,
379 * we will try to reclaim all available objects, otherwise we can end
380 * up failing allocations although there are plenty of reclaimable
381 * objects spread over several slabs with usage less than the
384 * We detect the "tight on memory" situations by looking at the total
385 * number of objects we want to scan (total_scan). If it is greater
386 * than the total number of objects on slab (freeable), we must be
387 * scanning at high prio and therefore should try to reclaim as much as
390 while (total_scan
>= batch_size
||
391 total_scan
>= freeable
) {
393 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
395 shrinkctl
->nr_to_scan
= nr_to_scan
;
396 shrinkctl
->nr_scanned
= nr_to_scan
;
397 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
398 if (ret
== SHRINK_STOP
)
402 count_vm_events(SLABS_SCANNED
, shrinkctl
->nr_scanned
);
403 total_scan
-= shrinkctl
->nr_scanned
;
404 scanned
+= shrinkctl
->nr_scanned
;
409 if (next_deferred
>= scanned
)
410 next_deferred
-= scanned
;
414 * move the unused scan count back into the shrinker in a
415 * manner that handles concurrent updates. If we exhausted the
416 * scan, there is no need to do an update.
418 if (next_deferred
> 0)
419 new_nr
= atomic_long_add_return(next_deferred
,
420 &shrinker
->nr_deferred
[nid
]);
422 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
424 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
429 * shrink_slab - shrink slab caches
430 * @gfp_mask: allocation context
431 * @nid: node whose slab caches to target
432 * @memcg: memory cgroup whose slab caches to target
433 * @nr_scanned: pressure numerator
434 * @nr_eligible: pressure denominator
436 * Call the shrink functions to age shrinkable caches.
438 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
439 * unaware shrinkers will receive a node id of 0 instead.
441 * @memcg specifies the memory cgroup to target. If it is not NULL,
442 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
443 * objects from the memory cgroup specified. Otherwise, only unaware
444 * shrinkers are called.
446 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
447 * the available objects should be scanned. Page reclaim for example
448 * passes the number of pages scanned and the number of pages on the
449 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
450 * when it encountered mapped pages. The ratio is further biased by
451 * the ->seeks setting of the shrink function, which indicates the
452 * cost to recreate an object relative to that of an LRU page.
454 * Returns the number of reclaimed slab objects.
456 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
457 struct mem_cgroup
*memcg
,
458 unsigned long nr_scanned
,
459 unsigned long nr_eligible
)
461 struct shrinker
*shrinker
;
462 unsigned long freed
= 0;
464 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
468 nr_scanned
= SWAP_CLUSTER_MAX
;
470 if (!down_read_trylock(&shrinker_rwsem
)) {
472 * If we would return 0, our callers would understand that we
473 * have nothing else to shrink and give up trying. By returning
474 * 1 we keep it going and assume we'll be able to shrink next
481 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
482 struct shrink_control sc
= {
483 .gfp_mask
= gfp_mask
,
489 * If kernel memory accounting is disabled, we ignore
490 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
491 * passing NULL for memcg.
493 if (memcg_kmem_enabled() &&
494 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
497 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
500 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
503 up_read(&shrinker_rwsem
);
509 void drop_slab_node(int nid
)
514 struct mem_cgroup
*memcg
= NULL
;
518 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
520 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
521 } while (freed
> 10);
528 for_each_online_node(nid
)
532 static inline int is_page_cache_freeable(struct page
*page
)
535 * A freeable page cache page is referenced only by the caller
536 * that isolated the page, the page cache radix tree and
537 * optional buffer heads at page->private.
539 return page_count(page
) - page_has_private(page
) == 2;
542 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
544 if (current
->flags
& PF_SWAPWRITE
)
546 if (!inode_write_congested(inode
))
548 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
554 * We detected a synchronous write error writing a page out. Probably
555 * -ENOSPC. We need to propagate that into the address_space for a subsequent
556 * fsync(), msync() or close().
558 * The tricky part is that after writepage we cannot touch the mapping: nothing
559 * prevents it from being freed up. But we have a ref on the page and once
560 * that page is locked, the mapping is pinned.
562 * We're allowed to run sleeping lock_page() here because we know the caller has
565 static void handle_write_error(struct address_space
*mapping
,
566 struct page
*page
, int error
)
569 if (page_mapping(page
) == mapping
)
570 mapping_set_error(mapping
, error
);
574 /* possible outcome of pageout() */
576 /* failed to write page out, page is locked */
578 /* move page to the active list, page is locked */
580 /* page has been sent to the disk successfully, page is unlocked */
582 /* page is clean and locked */
587 * pageout is called by shrink_page_list() for each dirty page.
588 * Calls ->writepage().
590 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
591 struct scan_control
*sc
)
594 * If the page is dirty, only perform writeback if that write
595 * will be non-blocking. To prevent this allocation from being
596 * stalled by pagecache activity. But note that there may be
597 * stalls if we need to run get_block(). We could test
598 * PagePrivate for that.
600 * If this process is currently in __generic_file_write_iter() against
601 * this page's queue, we can perform writeback even if that
604 * If the page is swapcache, write it back even if that would
605 * block, for some throttling. This happens by accident, because
606 * swap_backing_dev_info is bust: it doesn't reflect the
607 * congestion state of the swapdevs. Easy to fix, if needed.
609 if (!is_page_cache_freeable(page
))
613 * Some data journaling orphaned pages can have
614 * page->mapping == NULL while being dirty with clean buffers.
616 if (page_has_private(page
)) {
617 if (try_to_free_buffers(page
)) {
618 ClearPageDirty(page
);
619 pr_info("%s: orphaned page\n", __func__
);
625 if (mapping
->a_ops
->writepage
== NULL
)
626 return PAGE_ACTIVATE
;
627 if (!may_write_to_inode(mapping
->host
, sc
))
630 if (clear_page_dirty_for_io(page
)) {
632 struct writeback_control wbc
= {
633 .sync_mode
= WB_SYNC_NONE
,
634 .nr_to_write
= SWAP_CLUSTER_MAX
,
636 .range_end
= LLONG_MAX
,
640 SetPageReclaim(page
);
641 res
= mapping
->a_ops
->writepage(page
, &wbc
);
643 handle_write_error(mapping
, page
, res
);
644 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
645 ClearPageReclaim(page
);
646 return PAGE_ACTIVATE
;
649 if (!PageWriteback(page
)) {
650 /* synchronous write or broken a_ops? */
651 ClearPageReclaim(page
);
653 trace_mm_vmscan_writepage(page
);
654 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
662 * Same as remove_mapping, but if the page is removed from the mapping, it
663 * gets returned with a refcount of 0.
665 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
670 BUG_ON(!PageLocked(page
));
671 BUG_ON(mapping
!= page_mapping(page
));
673 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
675 * The non racy check for a busy page.
677 * Must be careful with the order of the tests. When someone has
678 * a ref to the page, it may be possible that they dirty it then
679 * drop the reference. So if PageDirty is tested before page_count
680 * here, then the following race may occur:
682 * get_user_pages(&page);
683 * [user mapping goes away]
685 * !PageDirty(page) [good]
686 * SetPageDirty(page);
688 * !page_count(page) [good, discard it]
690 * [oops, our write_to data is lost]
692 * Reversing the order of the tests ensures such a situation cannot
693 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
694 * load is not satisfied before that of page->_refcount.
696 * Note that if SetPageDirty is always performed via set_page_dirty,
697 * and thus under tree_lock, then this ordering is not required.
699 if (!page_ref_freeze(page
, 2))
701 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
702 if (unlikely(PageDirty(page
))) {
703 page_ref_unfreeze(page
, 2);
707 if (PageSwapCache(page
)) {
708 swp_entry_t swap
= { .val
= page_private(page
) };
709 mem_cgroup_swapout(page
, swap
);
710 __delete_from_swap_cache(page
);
711 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
712 put_swap_page(page
, swap
);
714 void (*freepage
)(struct page
*);
717 freepage
= mapping
->a_ops
->freepage
;
719 * Remember a shadow entry for reclaimed file cache in
720 * order to detect refaults, thus thrashing, later on.
722 * But don't store shadows in an address space that is
723 * already exiting. This is not just an optizimation,
724 * inode reclaim needs to empty out the radix tree or
725 * the nodes are lost. Don't plant shadows behind its
728 * We also don't store shadows for DAX mappings because the
729 * only page cache pages found in these are zero pages
730 * covering holes, and because we don't want to mix DAX
731 * exceptional entries and shadow exceptional entries in the
734 if (reclaimed
&& page_is_file_cache(page
) &&
735 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
736 shadow
= workingset_eviction(mapping
, page
);
737 __delete_from_page_cache(page
, shadow
);
738 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
740 if (freepage
!= NULL
)
747 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
752 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
753 * someone else has a ref on the page, abort and return 0. If it was
754 * successfully detached, return 1. Assumes the caller has a single ref on
757 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
759 if (__remove_mapping(mapping
, page
, false)) {
761 * Unfreezing the refcount with 1 rather than 2 effectively
762 * drops the pagecache ref for us without requiring another
765 page_ref_unfreeze(page
, 1);
772 * putback_lru_page - put previously isolated page onto appropriate LRU list
773 * @page: page to be put back to appropriate lru list
775 * Add previously isolated @page to appropriate LRU list.
776 * Page may still be unevictable for other reasons.
778 * lru_lock must not be held, interrupts must be enabled.
780 void putback_lru_page(struct page
*page
)
783 int was_unevictable
= PageUnevictable(page
);
785 VM_BUG_ON_PAGE(PageLRU(page
), page
);
788 ClearPageUnevictable(page
);
790 if (page_evictable(page
)) {
792 * For evictable pages, we can use the cache.
793 * In event of a race, worst case is we end up with an
794 * unevictable page on [in]active list.
795 * We know how to handle that.
797 is_unevictable
= false;
801 * Put unevictable pages directly on zone's unevictable
804 is_unevictable
= true;
805 add_page_to_unevictable_list(page
);
807 * When racing with an mlock or AS_UNEVICTABLE clearing
808 * (page is unlocked) make sure that if the other thread
809 * does not observe our setting of PG_lru and fails
810 * isolation/check_move_unevictable_pages,
811 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
812 * the page back to the evictable list.
814 * The other side is TestClearPageMlocked() or shmem_lock().
820 * page's status can change while we move it among lru. If an evictable
821 * page is on unevictable list, it never be freed. To avoid that,
822 * check after we added it to the list, again.
824 if (is_unevictable
&& page_evictable(page
)) {
825 if (!isolate_lru_page(page
)) {
829 /* This means someone else dropped this page from LRU
830 * So, it will be freed or putback to LRU again. There is
831 * nothing to do here.
835 if (was_unevictable
&& !is_unevictable
)
836 count_vm_event(UNEVICTABLE_PGRESCUED
);
837 else if (!was_unevictable
&& is_unevictable
)
838 count_vm_event(UNEVICTABLE_PGCULLED
);
840 put_page(page
); /* drop ref from isolate */
843 enum page_references
{
845 PAGEREF_RECLAIM_CLEAN
,
850 static enum page_references
page_check_references(struct page
*page
,
851 struct scan_control
*sc
)
853 int referenced_ptes
, referenced_page
;
854 unsigned long vm_flags
;
856 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
858 referenced_page
= TestClearPageReferenced(page
);
861 * Mlock lost the isolation race with us. Let try_to_unmap()
862 * move the page to the unevictable list.
864 if (vm_flags
& VM_LOCKED
)
865 return PAGEREF_RECLAIM
;
867 if (referenced_ptes
) {
868 if (PageSwapBacked(page
))
869 return PAGEREF_ACTIVATE
;
871 * All mapped pages start out with page table
872 * references from the instantiating fault, so we need
873 * to look twice if a mapped file page is used more
876 * Mark it and spare it for another trip around the
877 * inactive list. Another page table reference will
878 * lead to its activation.
880 * Note: the mark is set for activated pages as well
881 * so that recently deactivated but used pages are
884 SetPageReferenced(page
);
886 if (referenced_page
|| referenced_ptes
> 1)
887 return PAGEREF_ACTIVATE
;
890 * Activate file-backed executable pages after first usage.
892 if (vm_flags
& VM_EXEC
)
893 return PAGEREF_ACTIVATE
;
898 /* Reclaim if clean, defer dirty pages to writeback */
899 if (referenced_page
&& !PageSwapBacked(page
))
900 return PAGEREF_RECLAIM_CLEAN
;
902 return PAGEREF_RECLAIM
;
905 /* Check if a page is dirty or under writeback */
906 static void page_check_dirty_writeback(struct page
*page
,
907 bool *dirty
, bool *writeback
)
909 struct address_space
*mapping
;
912 * Anonymous pages are not handled by flushers and must be written
913 * from reclaim context. Do not stall reclaim based on them
915 if (!page_is_file_cache(page
) ||
916 (PageAnon(page
) && !PageSwapBacked(page
))) {
922 /* By default assume that the page flags are accurate */
923 *dirty
= PageDirty(page
);
924 *writeback
= PageWriteback(page
);
926 /* Verify dirty/writeback state if the filesystem supports it */
927 if (!page_has_private(page
))
930 mapping
= page_mapping(page
);
931 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
932 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
935 struct reclaim_stat
{
937 unsigned nr_unqueued_dirty
;
938 unsigned nr_congested
;
939 unsigned nr_writeback
;
940 unsigned nr_immediate
;
941 unsigned nr_activate
;
942 unsigned nr_ref_keep
;
943 unsigned nr_unmap_fail
;
947 * shrink_page_list() returns the number of reclaimed pages
949 static unsigned long shrink_page_list(struct list_head
*page_list
,
950 struct pglist_data
*pgdat
,
951 struct scan_control
*sc
,
952 enum ttu_flags ttu_flags
,
953 struct reclaim_stat
*stat
,
956 LIST_HEAD(ret_pages
);
957 LIST_HEAD(free_pages
);
959 unsigned nr_unqueued_dirty
= 0;
960 unsigned nr_dirty
= 0;
961 unsigned nr_congested
= 0;
962 unsigned nr_reclaimed
= 0;
963 unsigned nr_writeback
= 0;
964 unsigned nr_immediate
= 0;
965 unsigned nr_ref_keep
= 0;
966 unsigned nr_unmap_fail
= 0;
970 while (!list_empty(page_list
)) {
971 struct address_space
*mapping
;
974 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
975 bool dirty
, writeback
;
979 page
= lru_to_page(page_list
);
980 list_del(&page
->lru
);
982 if (!trylock_page(page
))
985 VM_BUG_ON_PAGE(PageActive(page
), page
);
989 if (unlikely(!page_evictable(page
)))
990 goto activate_locked
;
992 if (!sc
->may_unmap
&& page_mapped(page
))
995 /* Double the slab pressure for mapped and swapcache pages */
996 if ((page_mapped(page
) || PageSwapCache(page
)) &&
997 !(PageAnon(page
) && !PageSwapBacked(page
)))
1000 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
1001 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
1004 * The number of dirty pages determines if a zone is marked
1005 * reclaim_congested which affects wait_iff_congested. kswapd
1006 * will stall and start writing pages if the tail of the LRU
1007 * is all dirty unqueued pages.
1009 page_check_dirty_writeback(page
, &dirty
, &writeback
);
1010 if (dirty
|| writeback
)
1013 if (dirty
&& !writeback
)
1014 nr_unqueued_dirty
++;
1017 * Treat this page as congested if the underlying BDI is or if
1018 * pages are cycling through the LRU so quickly that the
1019 * pages marked for immediate reclaim are making it to the
1020 * end of the LRU a second time.
1022 mapping
= page_mapping(page
);
1023 if (((dirty
|| writeback
) && mapping
&&
1024 inode_write_congested(mapping
->host
)) ||
1025 (writeback
&& PageReclaim(page
)))
1029 * If a page at the tail of the LRU is under writeback, there
1030 * are three cases to consider.
1032 * 1) If reclaim is encountering an excessive number of pages
1033 * under writeback and this page is both under writeback and
1034 * PageReclaim then it indicates that pages are being queued
1035 * for IO but are being recycled through the LRU before the
1036 * IO can complete. Waiting on the page itself risks an
1037 * indefinite stall if it is impossible to writeback the
1038 * page due to IO error or disconnected storage so instead
1039 * note that the LRU is being scanned too quickly and the
1040 * caller can stall after page list has been processed.
1042 * 2) Global or new memcg reclaim encounters a page that is
1043 * not marked for immediate reclaim, or the caller does not
1044 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1045 * not to fs). In this case mark the page for immediate
1046 * reclaim and continue scanning.
1048 * Require may_enter_fs because we would wait on fs, which
1049 * may not have submitted IO yet. And the loop driver might
1050 * enter reclaim, and deadlock if it waits on a page for
1051 * which it is needed to do the write (loop masks off
1052 * __GFP_IO|__GFP_FS for this reason); but more thought
1053 * would probably show more reasons.
1055 * 3) Legacy memcg encounters a page that is already marked
1056 * PageReclaim. memcg does not have any dirty pages
1057 * throttling so we could easily OOM just because too many
1058 * pages are in writeback and there is nothing else to
1059 * reclaim. Wait for the writeback to complete.
1061 * In cases 1) and 2) we activate the pages to get them out of
1062 * the way while we continue scanning for clean pages on the
1063 * inactive list and refilling from the active list. The
1064 * observation here is that waiting for disk writes is more
1065 * expensive than potentially causing reloads down the line.
1066 * Since they're marked for immediate reclaim, they won't put
1067 * memory pressure on the cache working set any longer than it
1068 * takes to write them to disk.
1070 if (PageWriteback(page
)) {
1072 if (current_is_kswapd() &&
1073 PageReclaim(page
) &&
1074 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1076 goto activate_locked
;
1079 } else if (sane_reclaim(sc
) ||
1080 !PageReclaim(page
) || !may_enter_fs
) {
1082 * This is slightly racy - end_page_writeback()
1083 * might have just cleared PageReclaim, then
1084 * setting PageReclaim here end up interpreted
1085 * as PageReadahead - but that does not matter
1086 * enough to care. What we do want is for this
1087 * page to have PageReclaim set next time memcg
1088 * reclaim reaches the tests above, so it will
1089 * then wait_on_page_writeback() to avoid OOM;
1090 * and it's also appropriate in global reclaim.
1092 SetPageReclaim(page
);
1094 goto activate_locked
;
1099 wait_on_page_writeback(page
);
1100 /* then go back and try same page again */
1101 list_add_tail(&page
->lru
, page_list
);
1107 references
= page_check_references(page
, sc
);
1109 switch (references
) {
1110 case PAGEREF_ACTIVATE
:
1111 goto activate_locked
;
1115 case PAGEREF_RECLAIM
:
1116 case PAGEREF_RECLAIM_CLEAN
:
1117 ; /* try to reclaim the page below */
1121 * Anonymous process memory has backing store?
1122 * Try to allocate it some swap space here.
1123 * Lazyfree page could be freed directly
1125 if (PageAnon(page
) && PageSwapBacked(page
) &&
1126 !PageSwapCache(page
)) {
1127 if (!(sc
->gfp_mask
& __GFP_IO
))
1129 if (PageTransHuge(page
)) {
1130 /* cannot split THP, skip it */
1131 if (!can_split_huge_page(page
, NULL
))
1132 goto activate_locked
;
1134 * Split pages without a PMD map right
1135 * away. Chances are some or all of the
1136 * tail pages can be freed without IO.
1138 if (!compound_mapcount(page
) &&
1139 split_huge_page_to_list(page
, page_list
))
1140 goto activate_locked
;
1142 if (!add_to_swap(page
)) {
1143 if (!PageTransHuge(page
))
1144 goto activate_locked
;
1145 /* Split THP and swap individual base pages */
1146 if (split_huge_page_to_list(page
, page_list
))
1147 goto activate_locked
;
1148 if (!add_to_swap(page
))
1149 goto activate_locked
;
1152 /* XXX: We don't support THP writes */
1153 if (PageTransHuge(page
) &&
1154 split_huge_page_to_list(page
, page_list
)) {
1155 delete_from_swap_cache(page
);
1156 goto activate_locked
;
1161 /* Adding to swap updated mapping */
1162 mapping
= page_mapping(page
);
1163 } else if (unlikely(PageTransHuge(page
))) {
1164 /* Split file THP */
1165 if (split_huge_page_to_list(page
, page_list
))
1169 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1172 * The page is mapped into the page tables of one or more
1173 * processes. Try to unmap it here.
1175 if (page_mapped(page
)) {
1176 if (!try_to_unmap(page
, ttu_flags
| TTU_BATCH_FLUSH
)) {
1178 goto activate_locked
;
1182 if (PageDirty(page
)) {
1184 * Only kswapd can writeback filesystem pages
1185 * to avoid risk of stack overflow. But avoid
1186 * injecting inefficient single-page IO into
1187 * flusher writeback as much as possible: only
1188 * write pages when we've encountered many
1189 * dirty pages, and when we've already scanned
1190 * the rest of the LRU for clean pages and see
1191 * the same dirty pages again (PageReclaim).
1193 if (page_is_file_cache(page
) &&
1194 (!current_is_kswapd() || !PageReclaim(page
) ||
1195 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1197 * Immediately reclaim when written back.
1198 * Similar in principal to deactivate_page()
1199 * except we already have the page isolated
1200 * and know it's dirty
1202 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1203 SetPageReclaim(page
);
1205 goto activate_locked
;
1208 if (references
== PAGEREF_RECLAIM_CLEAN
)
1212 if (!sc
->may_writepage
)
1216 * Page is dirty. Flush the TLB if a writable entry
1217 * potentially exists to avoid CPU writes after IO
1218 * starts and then write it out here.
1220 try_to_unmap_flush_dirty();
1221 switch (pageout(page
, mapping
, sc
)) {
1225 goto activate_locked
;
1227 if (PageWriteback(page
))
1229 if (PageDirty(page
))
1233 * A synchronous write - probably a ramdisk. Go
1234 * ahead and try to reclaim the page.
1236 if (!trylock_page(page
))
1238 if (PageDirty(page
) || PageWriteback(page
))
1240 mapping
= page_mapping(page
);
1242 ; /* try to free the page below */
1247 * If the page has buffers, try to free the buffer mappings
1248 * associated with this page. If we succeed we try to free
1251 * We do this even if the page is PageDirty().
1252 * try_to_release_page() does not perform I/O, but it is
1253 * possible for a page to have PageDirty set, but it is actually
1254 * clean (all its buffers are clean). This happens if the
1255 * buffers were written out directly, with submit_bh(). ext3
1256 * will do this, as well as the blockdev mapping.
1257 * try_to_release_page() will discover that cleanness and will
1258 * drop the buffers and mark the page clean - it can be freed.
1260 * Rarely, pages can have buffers and no ->mapping. These are
1261 * the pages which were not successfully invalidated in
1262 * truncate_complete_page(). We try to drop those buffers here
1263 * and if that worked, and the page is no longer mapped into
1264 * process address space (page_count == 1) it can be freed.
1265 * Otherwise, leave the page on the LRU so it is swappable.
1267 if (page_has_private(page
)) {
1268 if (!try_to_release_page(page
, sc
->gfp_mask
))
1269 goto activate_locked
;
1270 if (!mapping
&& page_count(page
) == 1) {
1272 if (put_page_testzero(page
))
1276 * rare race with speculative reference.
1277 * the speculative reference will free
1278 * this page shortly, so we may
1279 * increment nr_reclaimed here (and
1280 * leave it off the LRU).
1288 if (PageAnon(page
) && !PageSwapBacked(page
)) {
1289 /* follow __remove_mapping for reference */
1290 if (!page_ref_freeze(page
, 1))
1292 if (PageDirty(page
)) {
1293 page_ref_unfreeze(page
, 1);
1297 count_vm_event(PGLAZYFREED
);
1298 count_memcg_page_event(page
, PGLAZYFREED
);
1299 } else if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1302 * At this point, we have no other references and there is
1303 * no way to pick any more up (removed from LRU, removed
1304 * from pagecache). Can use non-atomic bitops now (and
1305 * we obviously don't have to worry about waking up a process
1306 * waiting on the page lock, because there are no references.
1308 __ClearPageLocked(page
);
1313 * Is there need to periodically free_page_list? It would
1314 * appear not as the counts should be low
1316 list_add(&page
->lru
, &free_pages
);
1320 /* Not a candidate for swapping, so reclaim swap space. */
1321 if (PageSwapCache(page
) && (mem_cgroup_swap_full(page
) ||
1323 try_to_free_swap(page
);
1324 VM_BUG_ON_PAGE(PageActive(page
), page
);
1325 if (!PageMlocked(page
)) {
1326 SetPageActive(page
);
1328 count_memcg_page_event(page
, PGACTIVATE
);
1333 list_add(&page
->lru
, &ret_pages
);
1334 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1337 mem_cgroup_uncharge_list(&free_pages
);
1338 try_to_unmap_flush();
1339 free_hot_cold_page_list(&free_pages
, true);
1341 list_splice(&ret_pages
, page_list
);
1342 count_vm_events(PGACTIVATE
, pgactivate
);
1345 stat
->nr_dirty
= nr_dirty
;
1346 stat
->nr_congested
= nr_congested
;
1347 stat
->nr_unqueued_dirty
= nr_unqueued_dirty
;
1348 stat
->nr_writeback
= nr_writeback
;
1349 stat
->nr_immediate
= nr_immediate
;
1350 stat
->nr_activate
= pgactivate
;
1351 stat
->nr_ref_keep
= nr_ref_keep
;
1352 stat
->nr_unmap_fail
= nr_unmap_fail
;
1354 return nr_reclaimed
;
1357 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1358 struct list_head
*page_list
)
1360 struct scan_control sc
= {
1361 .gfp_mask
= GFP_KERNEL
,
1362 .priority
= DEF_PRIORITY
,
1366 struct page
*page
, *next
;
1367 LIST_HEAD(clean_pages
);
1369 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1370 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1371 !__PageMovable(page
)) {
1372 ClearPageActive(page
);
1373 list_move(&page
->lru
, &clean_pages
);
1377 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1378 TTU_IGNORE_ACCESS
, NULL
, true);
1379 list_splice(&clean_pages
, page_list
);
1380 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1385 * Attempt to remove the specified page from its LRU. Only take this page
1386 * if it is of the appropriate PageActive status. Pages which are being
1387 * freed elsewhere are also ignored.
1389 * page: page to consider
1390 * mode: one of the LRU isolation modes defined above
1392 * returns 0 on success, -ve errno on failure.
1394 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1398 /* Only take pages on the LRU. */
1402 /* Compaction should not handle unevictable pages but CMA can do so */
1403 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1409 * To minimise LRU disruption, the caller can indicate that it only
1410 * wants to isolate pages it will be able to operate on without
1411 * blocking - clean pages for the most part.
1413 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1414 * that it is possible to migrate without blocking
1416 if (mode
& ISOLATE_ASYNC_MIGRATE
) {
1417 /* All the caller can do on PageWriteback is block */
1418 if (PageWriteback(page
))
1421 if (PageDirty(page
)) {
1422 struct address_space
*mapping
;
1425 * Only pages without mappings or that have a
1426 * ->migratepage callback are possible to migrate
1429 mapping
= page_mapping(page
);
1430 if (mapping
&& !mapping
->a_ops
->migratepage
)
1435 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1438 if (likely(get_page_unless_zero(page
))) {
1440 * Be careful not to clear PageLRU until after we're
1441 * sure the page is not being freed elsewhere -- the
1442 * page release code relies on it.
1453 * Update LRU sizes after isolating pages. The LRU size updates must
1454 * be complete before mem_cgroup_update_lru_size due to a santity check.
1456 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1457 enum lru_list lru
, unsigned long *nr_zone_taken
)
1461 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1462 if (!nr_zone_taken
[zid
])
1465 __update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1467 mem_cgroup_update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1474 * zone_lru_lock is heavily contended. Some of the functions that
1475 * shrink the lists perform better by taking out a batch of pages
1476 * and working on them outside the LRU lock.
1478 * For pagecache intensive workloads, this function is the hottest
1479 * spot in the kernel (apart from copy_*_user functions).
1481 * Appropriate locks must be held before calling this function.
1483 * @nr_to_scan: The number of eligible pages to look through on the list.
1484 * @lruvec: The LRU vector to pull pages from.
1485 * @dst: The temp list to put pages on to.
1486 * @nr_scanned: The number of pages that were scanned.
1487 * @sc: The scan_control struct for this reclaim session
1488 * @mode: One of the LRU isolation modes
1489 * @lru: LRU list id for isolating
1491 * returns how many pages were moved onto *@dst.
1493 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1494 struct lruvec
*lruvec
, struct list_head
*dst
,
1495 unsigned long *nr_scanned
, struct scan_control
*sc
,
1496 isolate_mode_t mode
, enum lru_list lru
)
1498 struct list_head
*src
= &lruvec
->lists
[lru
];
1499 unsigned long nr_taken
= 0;
1500 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1501 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1502 unsigned long skipped
= 0;
1503 unsigned long scan
, total_scan
, nr_pages
;
1504 LIST_HEAD(pages_skipped
);
1507 for (total_scan
= 0;
1508 scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&& !list_empty(src
);
1512 page
= lru_to_page(src
);
1513 prefetchw_prev_lru_page(page
, src
, flags
);
1515 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1517 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1518 list_move(&page
->lru
, &pages_skipped
);
1519 nr_skipped
[page_zonenum(page
)]++;
1524 * Do not count skipped pages because that makes the function
1525 * return with no isolated pages if the LRU mostly contains
1526 * ineligible pages. This causes the VM to not reclaim any
1527 * pages, triggering a premature OOM.
1530 switch (__isolate_lru_page(page
, mode
)) {
1532 nr_pages
= hpage_nr_pages(page
);
1533 nr_taken
+= nr_pages
;
1534 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1535 list_move(&page
->lru
, dst
);
1539 /* else it is being freed elsewhere */
1540 list_move(&page
->lru
, src
);
1549 * Splice any skipped pages to the start of the LRU list. Note that
1550 * this disrupts the LRU order when reclaiming for lower zones but
1551 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1552 * scanning would soon rescan the same pages to skip and put the
1553 * system at risk of premature OOM.
1555 if (!list_empty(&pages_skipped
)) {
1558 list_splice(&pages_skipped
, src
);
1559 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1560 if (!nr_skipped
[zid
])
1563 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1564 skipped
+= nr_skipped
[zid
];
1567 *nr_scanned
= total_scan
;
1568 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1569 total_scan
, skipped
, nr_taken
, mode
, lru
);
1570 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1575 * isolate_lru_page - tries to isolate a page from its LRU list
1576 * @page: page to isolate from its LRU list
1578 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1579 * vmstat statistic corresponding to whatever LRU list the page was on.
1581 * Returns 0 if the page was removed from an LRU list.
1582 * Returns -EBUSY if the page was not on an LRU list.
1584 * The returned page will have PageLRU() cleared. If it was found on
1585 * the active list, it will have PageActive set. If it was found on
1586 * the unevictable list, it will have the PageUnevictable bit set. That flag
1587 * may need to be cleared by the caller before letting the page go.
1589 * The vmstat statistic corresponding to the list on which the page was
1590 * found will be decremented.
1593 * (1) Must be called with an elevated refcount on the page. This is a
1594 * fundamentnal difference from isolate_lru_pages (which is called
1595 * without a stable reference).
1596 * (2) the lru_lock must not be held.
1597 * (3) interrupts must be enabled.
1599 int isolate_lru_page(struct page
*page
)
1603 VM_BUG_ON_PAGE(!page_count(page
), page
);
1604 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1606 if (PageLRU(page
)) {
1607 struct zone
*zone
= page_zone(page
);
1608 struct lruvec
*lruvec
;
1610 spin_lock_irq(zone_lru_lock(zone
));
1611 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1612 if (PageLRU(page
)) {
1613 int lru
= page_lru(page
);
1616 del_page_from_lru_list(page
, lruvec
, lru
);
1619 spin_unlock_irq(zone_lru_lock(zone
));
1625 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1626 * then get resheduled. When there are massive number of tasks doing page
1627 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1628 * the LRU list will go small and be scanned faster than necessary, leading to
1629 * unnecessary swapping, thrashing and OOM.
1631 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1632 struct scan_control
*sc
)
1634 unsigned long inactive
, isolated
;
1636 if (current_is_kswapd())
1639 if (!sane_reclaim(sc
))
1643 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1644 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1646 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1647 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1651 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1652 * won't get blocked by normal direct-reclaimers, forming a circular
1655 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1658 return isolated
> inactive
;
1661 static noinline_for_stack
void
1662 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1664 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1665 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1666 LIST_HEAD(pages_to_free
);
1669 * Put back any unfreeable pages.
1671 while (!list_empty(page_list
)) {
1672 struct page
*page
= lru_to_page(page_list
);
1675 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1676 list_del(&page
->lru
);
1677 if (unlikely(!page_evictable(page
))) {
1678 spin_unlock_irq(&pgdat
->lru_lock
);
1679 putback_lru_page(page
);
1680 spin_lock_irq(&pgdat
->lru_lock
);
1684 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1687 lru
= page_lru(page
);
1688 add_page_to_lru_list(page
, lruvec
, lru
);
1690 if (is_active_lru(lru
)) {
1691 int file
= is_file_lru(lru
);
1692 int numpages
= hpage_nr_pages(page
);
1693 reclaim_stat
->recent_rotated
[file
] += numpages
;
1695 if (put_page_testzero(page
)) {
1696 __ClearPageLRU(page
);
1697 __ClearPageActive(page
);
1698 del_page_from_lru_list(page
, lruvec
, lru
);
1700 if (unlikely(PageCompound(page
))) {
1701 spin_unlock_irq(&pgdat
->lru_lock
);
1702 mem_cgroup_uncharge(page
);
1703 (*get_compound_page_dtor(page
))(page
);
1704 spin_lock_irq(&pgdat
->lru_lock
);
1706 list_add(&page
->lru
, &pages_to_free
);
1711 * To save our caller's stack, now use input list for pages to free.
1713 list_splice(&pages_to_free
, page_list
);
1717 * If a kernel thread (such as nfsd for loop-back mounts) services
1718 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1719 * In that case we should only throttle if the backing device it is
1720 * writing to is congested. In other cases it is safe to throttle.
1722 static int current_may_throttle(void)
1724 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1725 current
->backing_dev_info
== NULL
||
1726 bdi_write_congested(current
->backing_dev_info
);
1730 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1731 * of reclaimed pages
1733 static noinline_for_stack
unsigned long
1734 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1735 struct scan_control
*sc
, enum lru_list lru
)
1737 LIST_HEAD(page_list
);
1738 unsigned long nr_scanned
;
1739 unsigned long nr_reclaimed
= 0;
1740 unsigned long nr_taken
;
1741 struct reclaim_stat stat
= {};
1742 isolate_mode_t isolate_mode
= 0;
1743 int file
= is_file_lru(lru
);
1744 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1745 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1746 bool stalled
= false;
1748 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1752 /* wait a bit for the reclaimer. */
1756 /* We are about to die and free our memory. Return now. */
1757 if (fatal_signal_pending(current
))
1758 return SWAP_CLUSTER_MAX
;
1764 isolate_mode
|= ISOLATE_UNMAPPED
;
1766 spin_lock_irq(&pgdat
->lru_lock
);
1768 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1769 &nr_scanned
, sc
, isolate_mode
, lru
);
1771 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1772 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1774 if (current_is_kswapd()) {
1775 if (global_reclaim(sc
))
1776 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1777 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_KSWAPD
,
1780 if (global_reclaim(sc
))
1781 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1782 count_memcg_events(lruvec_memcg(lruvec
), PGSCAN_DIRECT
,
1785 spin_unlock_irq(&pgdat
->lru_lock
);
1790 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, 0,
1793 spin_lock_irq(&pgdat
->lru_lock
);
1795 if (current_is_kswapd()) {
1796 if (global_reclaim(sc
))
1797 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1798 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_KSWAPD
,
1801 if (global_reclaim(sc
))
1802 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1803 count_memcg_events(lruvec_memcg(lruvec
), PGSTEAL_DIRECT
,
1807 putback_inactive_pages(lruvec
, &page_list
);
1809 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1811 spin_unlock_irq(&pgdat
->lru_lock
);
1813 mem_cgroup_uncharge_list(&page_list
);
1814 free_hot_cold_page_list(&page_list
, true);
1817 * If reclaim is isolating dirty pages under writeback, it implies
1818 * that the long-lived page allocation rate is exceeding the page
1819 * laundering rate. Either the global limits are not being effective
1820 * at throttling processes due to the page distribution throughout
1821 * zones or there is heavy usage of a slow backing device. The
1822 * only option is to throttle from reclaim context which is not ideal
1823 * as there is no guarantee the dirtying process is throttled in the
1824 * same way balance_dirty_pages() manages.
1826 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1827 * of pages under pages flagged for immediate reclaim and stall if any
1828 * are encountered in the nr_immediate check below.
1830 if (stat
.nr_writeback
&& stat
.nr_writeback
== nr_taken
)
1831 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1834 * Legacy memcg will stall in page writeback so avoid forcibly
1837 if (sane_reclaim(sc
)) {
1839 * Tag a zone as congested if all the dirty pages scanned were
1840 * backed by a congested BDI and wait_iff_congested will stall.
1842 if (stat
.nr_dirty
&& stat
.nr_dirty
== stat
.nr_congested
)
1843 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1846 * If dirty pages are scanned that are not queued for IO, it
1847 * implies that flushers are not doing their job. This can
1848 * happen when memory pressure pushes dirty pages to the end of
1849 * the LRU before the dirty limits are breached and the dirty
1850 * data has expired. It can also happen when the proportion of
1851 * dirty pages grows not through writes but through memory
1852 * pressure reclaiming all the clean cache. And in some cases,
1853 * the flushers simply cannot keep up with the allocation
1854 * rate. Nudge the flusher threads in case they are asleep, but
1855 * also allow kswapd to start writing pages during reclaim.
1857 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1858 wakeup_flusher_threads(0, WB_REASON_VMSCAN
);
1859 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1863 * If kswapd scans pages marked marked for immediate
1864 * reclaim and under writeback (nr_immediate), it implies
1865 * that pages are cycling through the LRU faster than
1866 * they are written so also forcibly stall.
1868 if (stat
.nr_immediate
&& current_may_throttle())
1869 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1873 * Stall direct reclaim for IO completions if underlying BDIs or zone
1874 * is congested. Allow kswapd to continue until it starts encountering
1875 * unqueued dirty pages or cycling through the LRU too quickly.
1877 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1878 current_may_throttle())
1879 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1881 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1882 nr_scanned
, nr_reclaimed
,
1883 stat
.nr_dirty
, stat
.nr_writeback
,
1884 stat
.nr_congested
, stat
.nr_immediate
,
1885 stat
.nr_activate
, stat
.nr_ref_keep
,
1887 sc
->priority
, file
);
1888 return nr_reclaimed
;
1892 * This moves pages from the active list to the inactive list.
1894 * We move them the other way if the page is referenced by one or more
1895 * processes, from rmap.
1897 * If the pages are mostly unmapped, the processing is fast and it is
1898 * appropriate to hold zone_lru_lock across the whole operation. But if
1899 * the pages are mapped, the processing is slow (page_referenced()) so we
1900 * should drop zone_lru_lock around each page. It's impossible to balance
1901 * this, so instead we remove the pages from the LRU while processing them.
1902 * It is safe to rely on PG_active against the non-LRU pages in here because
1903 * nobody will play with that bit on a non-LRU page.
1905 * The downside is that we have to touch page->_refcount against each page.
1906 * But we had to alter page->flags anyway.
1908 * Returns the number of pages moved to the given lru.
1911 static unsigned move_active_pages_to_lru(struct lruvec
*lruvec
,
1912 struct list_head
*list
,
1913 struct list_head
*pages_to_free
,
1916 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1921 while (!list_empty(list
)) {
1922 page
= lru_to_page(list
);
1923 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1925 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1928 nr_pages
= hpage_nr_pages(page
);
1929 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1930 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1932 if (put_page_testzero(page
)) {
1933 __ClearPageLRU(page
);
1934 __ClearPageActive(page
);
1935 del_page_from_lru_list(page
, lruvec
, lru
);
1937 if (unlikely(PageCompound(page
))) {
1938 spin_unlock_irq(&pgdat
->lru_lock
);
1939 mem_cgroup_uncharge(page
);
1940 (*get_compound_page_dtor(page
))(page
);
1941 spin_lock_irq(&pgdat
->lru_lock
);
1943 list_add(&page
->lru
, pages_to_free
);
1945 nr_moved
+= nr_pages
;
1949 if (!is_active_lru(lru
)) {
1950 __count_vm_events(PGDEACTIVATE
, nr_moved
);
1951 count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
,
1958 static void shrink_active_list(unsigned long nr_to_scan
,
1959 struct lruvec
*lruvec
,
1960 struct scan_control
*sc
,
1963 unsigned long nr_taken
;
1964 unsigned long nr_scanned
;
1965 unsigned long vm_flags
;
1966 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1967 LIST_HEAD(l_active
);
1968 LIST_HEAD(l_inactive
);
1970 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1971 unsigned nr_deactivate
, nr_activate
;
1972 unsigned nr_rotated
= 0;
1973 isolate_mode_t isolate_mode
= 0;
1974 int file
= is_file_lru(lru
);
1975 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1980 isolate_mode
|= ISOLATE_UNMAPPED
;
1982 spin_lock_irq(&pgdat
->lru_lock
);
1984 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1985 &nr_scanned
, sc
, isolate_mode
, lru
);
1987 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1988 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1990 __count_vm_events(PGREFILL
, nr_scanned
);
1991 count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
1993 spin_unlock_irq(&pgdat
->lru_lock
);
1995 while (!list_empty(&l_hold
)) {
1997 page
= lru_to_page(&l_hold
);
1998 list_del(&page
->lru
);
2000 if (unlikely(!page_evictable(page
))) {
2001 putback_lru_page(page
);
2005 if (unlikely(buffer_heads_over_limit
)) {
2006 if (page_has_private(page
) && trylock_page(page
)) {
2007 if (page_has_private(page
))
2008 try_to_release_page(page
, 0);
2013 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
2015 nr_rotated
+= hpage_nr_pages(page
);
2017 * Identify referenced, file-backed active pages and
2018 * give them one more trip around the active list. So
2019 * that executable code get better chances to stay in
2020 * memory under moderate memory pressure. Anon pages
2021 * are not likely to be evicted by use-once streaming
2022 * IO, plus JVM can create lots of anon VM_EXEC pages,
2023 * so we ignore them here.
2025 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
2026 list_add(&page
->lru
, &l_active
);
2031 ClearPageActive(page
); /* we are de-activating */
2032 list_add(&page
->lru
, &l_inactive
);
2036 * Move pages back to the lru list.
2038 spin_lock_irq(&pgdat
->lru_lock
);
2040 * Count referenced pages from currently used mappings as rotated,
2041 * even though only some of them are actually re-activated. This
2042 * helps balance scan pressure between file and anonymous pages in
2045 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
2047 nr_activate
= move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
2048 nr_deactivate
= move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
2049 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2050 spin_unlock_irq(&pgdat
->lru_lock
);
2052 mem_cgroup_uncharge_list(&l_hold
);
2053 free_hot_cold_page_list(&l_hold
, true);
2054 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2055 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2059 * The inactive anon list should be small enough that the VM never has
2060 * to do too much work.
2062 * The inactive file list should be small enough to leave most memory
2063 * to the established workingset on the scan-resistant active list,
2064 * but large enough to avoid thrashing the aggregate readahead window.
2066 * Both inactive lists should also be large enough that each inactive
2067 * page has a chance to be referenced again before it is reclaimed.
2069 * If that fails and refaulting is observed, the inactive list grows.
2071 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2072 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2073 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2076 * memory ratio inactive
2077 * -------------------------------------
2086 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
,
2087 struct mem_cgroup
*memcg
,
2088 struct scan_control
*sc
, bool actual_reclaim
)
2090 enum lru_list active_lru
= file
* LRU_FILE
+ LRU_ACTIVE
;
2091 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2092 enum lru_list inactive_lru
= file
* LRU_FILE
;
2093 unsigned long inactive
, active
;
2094 unsigned long inactive_ratio
;
2095 unsigned long refaults
;
2099 * If we don't have swap space, anonymous page deactivation
2102 if (!file
&& !total_swap_pages
)
2105 inactive
= lruvec_lru_size(lruvec
, inactive_lru
, sc
->reclaim_idx
);
2106 active
= lruvec_lru_size(lruvec
, active_lru
, sc
->reclaim_idx
);
2109 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2111 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2114 * When refaults are being observed, it means a new workingset
2115 * is being established. Disable active list protection to get
2116 * rid of the stale workingset quickly.
2118 if (file
&& actual_reclaim
&& lruvec
->refaults
!= refaults
) {
2121 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2123 inactive_ratio
= int_sqrt(10 * gb
);
2129 trace_mm_vmscan_inactive_list_is_low(pgdat
->node_id
, sc
->reclaim_idx
,
2130 lruvec_lru_size(lruvec
, inactive_lru
, MAX_NR_ZONES
), inactive
,
2131 lruvec_lru_size(lruvec
, active_lru
, MAX_NR_ZONES
), active
,
2132 inactive_ratio
, file
);
2134 return inactive
* inactive_ratio
< active
;
2137 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2138 struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2139 struct scan_control
*sc
)
2141 if (is_active_lru(lru
)) {
2142 if (inactive_list_is_low(lruvec
, is_file_lru(lru
),
2144 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2148 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2159 * Determine how aggressively the anon and file LRU lists should be
2160 * scanned. The relative value of each set of LRU lists is determined
2161 * by looking at the fraction of the pages scanned we did rotate back
2162 * onto the active list instead of evict.
2164 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2165 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2167 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
2168 struct scan_control
*sc
, unsigned long *nr
,
2169 unsigned long *lru_pages
)
2171 int swappiness
= mem_cgroup_swappiness(memcg
);
2172 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
2174 u64 denominator
= 0; /* gcc */
2175 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2176 unsigned long anon_prio
, file_prio
;
2177 enum scan_balance scan_balance
;
2178 unsigned long anon
, file
;
2179 unsigned long ap
, fp
;
2182 /* If we have no swap space, do not bother scanning anon pages. */
2183 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2184 scan_balance
= SCAN_FILE
;
2189 * Global reclaim will swap to prevent OOM even with no
2190 * swappiness, but memcg users want to use this knob to
2191 * disable swapping for individual groups completely when
2192 * using the memory controller's swap limit feature would be
2195 if (!global_reclaim(sc
) && !swappiness
) {
2196 scan_balance
= SCAN_FILE
;
2201 * Do not apply any pressure balancing cleverness when the
2202 * system is close to OOM, scan both anon and file equally
2203 * (unless the swappiness setting disagrees with swapping).
2205 if (!sc
->priority
&& swappiness
) {
2206 scan_balance
= SCAN_EQUAL
;
2211 * Prevent the reclaimer from falling into the cache trap: as
2212 * cache pages start out inactive, every cache fault will tip
2213 * the scan balance towards the file LRU. And as the file LRU
2214 * shrinks, so does the window for rotation from references.
2215 * This means we have a runaway feedback loop where a tiny
2216 * thrashing file LRU becomes infinitely more attractive than
2217 * anon pages. Try to detect this based on file LRU size.
2219 if (global_reclaim(sc
)) {
2220 unsigned long pgdatfile
;
2221 unsigned long pgdatfree
;
2223 unsigned long total_high_wmark
= 0;
2225 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2226 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2227 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2229 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2230 struct zone
*zone
= &pgdat
->node_zones
[z
];
2231 if (!managed_zone(zone
))
2234 total_high_wmark
+= high_wmark_pages(zone
);
2237 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2239 * Force SCAN_ANON if there are enough inactive
2240 * anonymous pages on the LRU in eligible zones.
2241 * Otherwise, the small LRU gets thrashed.
2243 if (!inactive_list_is_low(lruvec
, false, memcg
, sc
, false) &&
2244 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, sc
->reclaim_idx
)
2246 scan_balance
= SCAN_ANON
;
2253 * If there is enough inactive page cache, i.e. if the size of the
2254 * inactive list is greater than that of the active list *and* the
2255 * inactive list actually has some pages to scan on this priority, we
2256 * do not reclaim anything from the anonymous working set right now.
2257 * Without the second condition we could end up never scanning an
2258 * lruvec even if it has plenty of old anonymous pages unless the
2259 * system is under heavy pressure.
2261 if (!inactive_list_is_low(lruvec
, true, memcg
, sc
, false) &&
2262 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, sc
->reclaim_idx
) >> sc
->priority
) {
2263 scan_balance
= SCAN_FILE
;
2267 scan_balance
= SCAN_FRACT
;
2270 * With swappiness at 100, anonymous and file have the same priority.
2271 * This scanning priority is essentially the inverse of IO cost.
2273 anon_prio
= swappiness
;
2274 file_prio
= 200 - anon_prio
;
2277 * OK, so we have swap space and a fair amount of page cache
2278 * pages. We use the recently rotated / recently scanned
2279 * ratios to determine how valuable each cache is.
2281 * Because workloads change over time (and to avoid overflow)
2282 * we keep these statistics as a floating average, which ends
2283 * up weighing recent references more than old ones.
2285 * anon in [0], file in [1]
2288 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
, MAX_NR_ZONES
) +
2289 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
, MAX_NR_ZONES
);
2290 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
, MAX_NR_ZONES
) +
2291 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
, MAX_NR_ZONES
);
2293 spin_lock_irq(&pgdat
->lru_lock
);
2294 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2295 reclaim_stat
->recent_scanned
[0] /= 2;
2296 reclaim_stat
->recent_rotated
[0] /= 2;
2299 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2300 reclaim_stat
->recent_scanned
[1] /= 2;
2301 reclaim_stat
->recent_rotated
[1] /= 2;
2305 * The amount of pressure on anon vs file pages is inversely
2306 * proportional to the fraction of recently scanned pages on
2307 * each list that were recently referenced and in active use.
2309 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2310 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2312 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2313 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2314 spin_unlock_irq(&pgdat
->lru_lock
);
2318 denominator
= ap
+ fp
+ 1;
2321 for_each_evictable_lru(lru
) {
2322 int file
= is_file_lru(lru
);
2326 size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2327 scan
= size
>> sc
->priority
;
2329 * If the cgroup's already been deleted, make sure to
2330 * scrape out the remaining cache.
2332 if (!scan
&& !mem_cgroup_online(memcg
))
2333 scan
= min(size
, SWAP_CLUSTER_MAX
);
2335 switch (scan_balance
) {
2337 /* Scan lists relative to size */
2341 * Scan types proportional to swappiness and
2342 * their relative recent reclaim efficiency.
2344 scan
= div64_u64(scan
* fraction
[file
],
2349 /* Scan one type exclusively */
2350 if ((scan_balance
== SCAN_FILE
) != file
) {
2356 /* Look ma, no brain */
2366 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2368 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2369 struct scan_control
*sc
, unsigned long *lru_pages
)
2371 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2372 unsigned long nr
[NR_LRU_LISTS
];
2373 unsigned long targets
[NR_LRU_LISTS
];
2374 unsigned long nr_to_scan
;
2376 unsigned long nr_reclaimed
= 0;
2377 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2378 struct blk_plug plug
;
2381 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2383 /* Record the original scan target for proportional adjustments later */
2384 memcpy(targets
, nr
, sizeof(nr
));
2387 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2388 * event that can occur when there is little memory pressure e.g.
2389 * multiple streaming readers/writers. Hence, we do not abort scanning
2390 * when the requested number of pages are reclaimed when scanning at
2391 * DEF_PRIORITY on the assumption that the fact we are direct
2392 * reclaiming implies that kswapd is not keeping up and it is best to
2393 * do a batch of work at once. For memcg reclaim one check is made to
2394 * abort proportional reclaim if either the file or anon lru has already
2395 * dropped to zero at the first pass.
2397 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2398 sc
->priority
== DEF_PRIORITY
);
2400 blk_start_plug(&plug
);
2401 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2402 nr
[LRU_INACTIVE_FILE
]) {
2403 unsigned long nr_anon
, nr_file
, percentage
;
2404 unsigned long nr_scanned
;
2406 for_each_evictable_lru(lru
) {
2408 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2409 nr
[lru
] -= nr_to_scan
;
2411 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2418 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2422 * For kswapd and memcg, reclaim at least the number of pages
2423 * requested. Ensure that the anon and file LRUs are scanned
2424 * proportionally what was requested by get_scan_count(). We
2425 * stop reclaiming one LRU and reduce the amount scanning
2426 * proportional to the original scan target.
2428 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2429 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2432 * It's just vindictive to attack the larger once the smaller
2433 * has gone to zero. And given the way we stop scanning the
2434 * smaller below, this makes sure that we only make one nudge
2435 * towards proportionality once we've got nr_to_reclaim.
2437 if (!nr_file
|| !nr_anon
)
2440 if (nr_file
> nr_anon
) {
2441 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2442 targets
[LRU_ACTIVE_ANON
] + 1;
2444 percentage
= nr_anon
* 100 / scan_target
;
2446 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2447 targets
[LRU_ACTIVE_FILE
] + 1;
2449 percentage
= nr_file
* 100 / scan_target
;
2452 /* Stop scanning the smaller of the LRU */
2454 nr
[lru
+ LRU_ACTIVE
] = 0;
2457 * Recalculate the other LRU scan count based on its original
2458 * scan target and the percentage scanning already complete
2460 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2461 nr_scanned
= targets
[lru
] - nr
[lru
];
2462 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2463 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2466 nr_scanned
= targets
[lru
] - nr
[lru
];
2467 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2468 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2470 scan_adjusted
= true;
2472 blk_finish_plug(&plug
);
2473 sc
->nr_reclaimed
+= nr_reclaimed
;
2476 * Even if we did not try to evict anon pages at all, we want to
2477 * rebalance the anon lru active/inactive ratio.
2479 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
2480 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2481 sc
, LRU_ACTIVE_ANON
);
2484 /* Use reclaim/compaction for costly allocs or under memory pressure */
2485 static bool in_reclaim_compaction(struct scan_control
*sc
)
2487 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2488 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2489 sc
->priority
< DEF_PRIORITY
- 2))
2496 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2497 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2498 * true if more pages should be reclaimed such that when the page allocator
2499 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2500 * It will give up earlier than that if there is difficulty reclaiming pages.
2502 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2503 unsigned long nr_reclaimed
,
2504 unsigned long nr_scanned
,
2505 struct scan_control
*sc
)
2507 unsigned long pages_for_compaction
;
2508 unsigned long inactive_lru_pages
;
2511 /* If not in reclaim/compaction mode, stop */
2512 if (!in_reclaim_compaction(sc
))
2515 /* Consider stopping depending on scan and reclaim activity */
2516 if (sc
->gfp_mask
& __GFP_RETRY_MAYFAIL
) {
2518 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2519 * full LRU list has been scanned and we are still failing
2520 * to reclaim pages. This full LRU scan is potentially
2521 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2523 if (!nr_reclaimed
&& !nr_scanned
)
2527 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2528 * fail without consequence, stop if we failed to reclaim
2529 * any pages from the last SWAP_CLUSTER_MAX number of
2530 * pages that were scanned. This will return to the
2531 * caller faster at the risk reclaim/compaction and
2532 * the resulting allocation attempt fails
2539 * If we have not reclaimed enough pages for compaction and the
2540 * inactive lists are large enough, continue reclaiming
2542 pages_for_compaction
= compact_gap(sc
->order
);
2543 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2544 if (get_nr_swap_pages() > 0)
2545 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2546 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2547 inactive_lru_pages
> pages_for_compaction
)
2550 /* If compaction would go ahead or the allocation would succeed, stop */
2551 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2552 struct zone
*zone
= &pgdat
->node_zones
[z
];
2553 if (!managed_zone(zone
))
2556 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2557 case COMPACT_SUCCESS
:
2558 case COMPACT_CONTINUE
:
2561 /* check next zone */
2568 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2570 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2571 unsigned long nr_reclaimed
, nr_scanned
;
2572 bool reclaimable
= false;
2575 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2576 struct mem_cgroup_reclaim_cookie reclaim
= {
2578 .priority
= sc
->priority
,
2580 unsigned long node_lru_pages
= 0;
2581 struct mem_cgroup
*memcg
;
2583 nr_reclaimed
= sc
->nr_reclaimed
;
2584 nr_scanned
= sc
->nr_scanned
;
2586 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2588 unsigned long lru_pages
;
2589 unsigned long reclaimed
;
2590 unsigned long scanned
;
2592 if (mem_cgroup_low(root
, memcg
)) {
2593 if (!sc
->memcg_low_reclaim
) {
2594 sc
->memcg_low_skipped
= 1;
2597 mem_cgroup_event(memcg
, MEMCG_LOW
);
2600 reclaimed
= sc
->nr_reclaimed
;
2601 scanned
= sc
->nr_scanned
;
2603 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2604 node_lru_pages
+= lru_pages
;
2607 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2608 memcg
, sc
->nr_scanned
- scanned
,
2611 /* Record the group's reclaim efficiency */
2612 vmpressure(sc
->gfp_mask
, memcg
, false,
2613 sc
->nr_scanned
- scanned
,
2614 sc
->nr_reclaimed
- reclaimed
);
2617 * Direct reclaim and kswapd have to scan all memory
2618 * cgroups to fulfill the overall scan target for the
2621 * Limit reclaim, on the other hand, only cares about
2622 * nr_to_reclaim pages to be reclaimed and it will
2623 * retry with decreasing priority if one round over the
2624 * whole hierarchy is not sufficient.
2626 if (!global_reclaim(sc
) &&
2627 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2628 mem_cgroup_iter_break(root
, memcg
);
2631 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2634 * Shrink the slab caches in the same proportion that
2635 * the eligible LRU pages were scanned.
2637 if (global_reclaim(sc
))
2638 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2639 sc
->nr_scanned
- nr_scanned
,
2642 if (reclaim_state
) {
2643 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2644 reclaim_state
->reclaimed_slab
= 0;
2647 /* Record the subtree's reclaim efficiency */
2648 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2649 sc
->nr_scanned
- nr_scanned
,
2650 sc
->nr_reclaimed
- nr_reclaimed
);
2652 if (sc
->nr_reclaimed
- nr_reclaimed
)
2655 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2656 sc
->nr_scanned
- nr_scanned
, sc
));
2659 * Kswapd gives up on balancing particular nodes after too
2660 * many failures to reclaim anything from them and goes to
2661 * sleep. On reclaim progress, reset the failure counter. A
2662 * successful direct reclaim run will revive a dormant kswapd.
2665 pgdat
->kswapd_failures
= 0;
2671 * Returns true if compaction should go ahead for a costly-order request, or
2672 * the allocation would already succeed without compaction. Return false if we
2673 * should reclaim first.
2675 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2677 unsigned long watermark
;
2678 enum compact_result suitable
;
2680 suitable
= compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
);
2681 if (suitable
== COMPACT_SUCCESS
)
2682 /* Allocation should succeed already. Don't reclaim. */
2684 if (suitable
== COMPACT_SKIPPED
)
2685 /* Compaction cannot yet proceed. Do reclaim. */
2689 * Compaction is already possible, but it takes time to run and there
2690 * are potentially other callers using the pages just freed. So proceed
2691 * with reclaim to make a buffer of free pages available to give
2692 * compaction a reasonable chance of completing and allocating the page.
2693 * Note that we won't actually reclaim the whole buffer in one attempt
2694 * as the target watermark in should_continue_reclaim() is lower. But if
2695 * we are already above the high+gap watermark, don't reclaim at all.
2697 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
2699 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2703 * This is the direct reclaim path, for page-allocating processes. We only
2704 * try to reclaim pages from zones which will satisfy the caller's allocation
2707 * If a zone is deemed to be full of pinned pages then just give it a light
2708 * scan then give up on it.
2710 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2714 unsigned long nr_soft_reclaimed
;
2715 unsigned long nr_soft_scanned
;
2717 pg_data_t
*last_pgdat
= NULL
;
2720 * If the number of buffer_heads in the machine exceeds the maximum
2721 * allowed level, force direct reclaim to scan the highmem zone as
2722 * highmem pages could be pinning lowmem pages storing buffer_heads
2724 orig_mask
= sc
->gfp_mask
;
2725 if (buffer_heads_over_limit
) {
2726 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2727 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2730 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2731 sc
->reclaim_idx
, sc
->nodemask
) {
2733 * Take care memory controller reclaiming has small influence
2736 if (global_reclaim(sc
)) {
2737 if (!cpuset_zone_allowed(zone
,
2738 GFP_KERNEL
| __GFP_HARDWALL
))
2742 * If we already have plenty of memory free for
2743 * compaction in this zone, don't free any more.
2744 * Even though compaction is invoked for any
2745 * non-zero order, only frequent costly order
2746 * reclamation is disruptive enough to become a
2747 * noticeable problem, like transparent huge
2750 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2751 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2752 compaction_ready(zone
, sc
)) {
2753 sc
->compaction_ready
= true;
2758 * Shrink each node in the zonelist once. If the
2759 * zonelist is ordered by zone (not the default) then a
2760 * node may be shrunk multiple times but in that case
2761 * the user prefers lower zones being preserved.
2763 if (zone
->zone_pgdat
== last_pgdat
)
2767 * This steals pages from memory cgroups over softlimit
2768 * and returns the number of reclaimed pages and
2769 * scanned pages. This works for global memory pressure
2770 * and balancing, not for a memcg's limit.
2772 nr_soft_scanned
= 0;
2773 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2774 sc
->order
, sc
->gfp_mask
,
2776 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2777 sc
->nr_scanned
+= nr_soft_scanned
;
2778 /* need some check for avoid more shrink_zone() */
2781 /* See comment about same check for global reclaim above */
2782 if (zone
->zone_pgdat
== last_pgdat
)
2784 last_pgdat
= zone
->zone_pgdat
;
2785 shrink_node(zone
->zone_pgdat
, sc
);
2789 * Restore to original mask to avoid the impact on the caller if we
2790 * promoted it to __GFP_HIGHMEM.
2792 sc
->gfp_mask
= orig_mask
;
2795 static void snapshot_refaults(struct mem_cgroup
*root_memcg
, pg_data_t
*pgdat
)
2797 struct mem_cgroup
*memcg
;
2799 memcg
= mem_cgroup_iter(root_memcg
, NULL
, NULL
);
2801 unsigned long refaults
;
2802 struct lruvec
*lruvec
;
2805 refaults
= memcg_page_state(memcg
, WORKINGSET_ACTIVATE
);
2807 refaults
= node_page_state(pgdat
, WORKINGSET_ACTIVATE
);
2809 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2810 lruvec
->refaults
= refaults
;
2811 } while ((memcg
= mem_cgroup_iter(root_memcg
, memcg
, NULL
)));
2815 * This is the main entry point to direct page reclaim.
2817 * If a full scan of the inactive list fails to free enough memory then we
2818 * are "out of memory" and something needs to be killed.
2820 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2821 * high - the zone may be full of dirty or under-writeback pages, which this
2822 * caller can't do much about. We kick the writeback threads and take explicit
2823 * naps in the hope that some of these pages can be written. But if the
2824 * allocating task holds filesystem locks which prevent writeout this might not
2825 * work, and the allocation attempt will fail.
2827 * returns: 0, if no pages reclaimed
2828 * else, the number of pages reclaimed
2830 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2831 struct scan_control
*sc
)
2833 int initial_priority
= sc
->priority
;
2834 pg_data_t
*last_pgdat
;
2838 delayacct_freepages_start();
2840 if (global_reclaim(sc
))
2841 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2844 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2847 shrink_zones(zonelist
, sc
);
2849 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2852 if (sc
->compaction_ready
)
2856 * If we're getting trouble reclaiming, start doing
2857 * writepage even in laptop mode.
2859 if (sc
->priority
< DEF_PRIORITY
- 2)
2860 sc
->may_writepage
= 1;
2861 } while (--sc
->priority
>= 0);
2864 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
2866 if (zone
->zone_pgdat
== last_pgdat
)
2868 last_pgdat
= zone
->zone_pgdat
;
2869 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
2872 delayacct_freepages_end();
2874 if (sc
->nr_reclaimed
)
2875 return sc
->nr_reclaimed
;
2877 /* Aborted reclaim to try compaction? don't OOM, then */
2878 if (sc
->compaction_ready
)
2881 /* Untapped cgroup reserves? Don't OOM, retry. */
2882 if (sc
->memcg_low_skipped
) {
2883 sc
->priority
= initial_priority
;
2884 sc
->memcg_low_reclaim
= 1;
2885 sc
->memcg_low_skipped
= 0;
2892 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
2895 unsigned long pfmemalloc_reserve
= 0;
2896 unsigned long free_pages
= 0;
2900 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
2903 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2904 zone
= &pgdat
->node_zones
[i
];
2905 if (!managed_zone(zone
))
2908 if (!zone_reclaimable_pages(zone
))
2911 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2912 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2915 /* If there are no reserves (unexpected config) then do not throttle */
2916 if (!pfmemalloc_reserve
)
2919 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2921 /* kswapd must be awake if processes are being throttled */
2922 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2923 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2924 (enum zone_type
)ZONE_NORMAL
);
2925 wake_up_interruptible(&pgdat
->kswapd_wait
);
2932 * Throttle direct reclaimers if backing storage is backed by the network
2933 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2934 * depleted. kswapd will continue to make progress and wake the processes
2935 * when the low watermark is reached.
2937 * Returns true if a fatal signal was delivered during throttling. If this
2938 * happens, the page allocator should not consider triggering the OOM killer.
2940 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2941 nodemask_t
*nodemask
)
2945 pg_data_t
*pgdat
= NULL
;
2948 * Kernel threads should not be throttled as they may be indirectly
2949 * responsible for cleaning pages necessary for reclaim to make forward
2950 * progress. kjournald for example may enter direct reclaim while
2951 * committing a transaction where throttling it could forcing other
2952 * processes to block on log_wait_commit().
2954 if (current
->flags
& PF_KTHREAD
)
2958 * If a fatal signal is pending, this process should not throttle.
2959 * It should return quickly so it can exit and free its memory
2961 if (fatal_signal_pending(current
))
2965 * Check if the pfmemalloc reserves are ok by finding the first node
2966 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2967 * GFP_KERNEL will be required for allocating network buffers when
2968 * swapping over the network so ZONE_HIGHMEM is unusable.
2970 * Throttling is based on the first usable node and throttled processes
2971 * wait on a queue until kswapd makes progress and wakes them. There
2972 * is an affinity then between processes waking up and where reclaim
2973 * progress has been made assuming the process wakes on the same node.
2974 * More importantly, processes running on remote nodes will not compete
2975 * for remote pfmemalloc reserves and processes on different nodes
2976 * should make reasonable progress.
2978 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2979 gfp_zone(gfp_mask
), nodemask
) {
2980 if (zone_idx(zone
) > ZONE_NORMAL
)
2983 /* Throttle based on the first usable node */
2984 pgdat
= zone
->zone_pgdat
;
2985 if (allow_direct_reclaim(pgdat
))
2990 /* If no zone was usable by the allocation flags then do not throttle */
2994 /* Account for the throttling */
2995 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2998 * If the caller cannot enter the filesystem, it's possible that it
2999 * is due to the caller holding an FS lock or performing a journal
3000 * transaction in the case of a filesystem like ext[3|4]. In this case,
3001 * it is not safe to block on pfmemalloc_wait as kswapd could be
3002 * blocked waiting on the same lock. Instead, throttle for up to a
3003 * second before continuing.
3005 if (!(gfp_mask
& __GFP_FS
)) {
3006 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
3007 allow_direct_reclaim(pgdat
), HZ
);
3012 /* Throttle until kswapd wakes the process */
3013 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
3014 allow_direct_reclaim(pgdat
));
3017 if (fatal_signal_pending(current
))
3024 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
3025 gfp_t gfp_mask
, nodemask_t
*nodemask
)
3027 unsigned long nr_reclaimed
;
3028 struct scan_control sc
= {
3029 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3030 .gfp_mask
= current_gfp_context(gfp_mask
),
3031 .reclaim_idx
= gfp_zone(gfp_mask
),
3033 .nodemask
= nodemask
,
3034 .priority
= DEF_PRIORITY
,
3035 .may_writepage
= !laptop_mode
,
3041 * Do not enter reclaim if fatal signal was delivered while throttled.
3042 * 1 is returned so that the page allocator does not OOM kill at this
3045 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
3048 trace_mm_vmscan_direct_reclaim_begin(order
,
3053 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3055 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
3057 return nr_reclaimed
;
3062 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
3063 gfp_t gfp_mask
, bool noswap
,
3065 unsigned long *nr_scanned
)
3067 struct scan_control sc
= {
3068 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
3069 .target_mem_cgroup
= memcg
,
3070 .may_writepage
= !laptop_mode
,
3072 .reclaim_idx
= MAX_NR_ZONES
- 1,
3073 .may_swap
= !noswap
,
3075 unsigned long lru_pages
;
3077 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
3078 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
3080 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
3086 * NOTE: Although we can get the priority field, using it
3087 * here is not a good idea, since it limits the pages we can scan.
3088 * if we don't reclaim here, the shrink_node from balance_pgdat
3089 * will pick up pages from other mem cgroup's as well. We hack
3090 * the priority and make it zero.
3092 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
3094 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
3096 *nr_scanned
= sc
.nr_scanned
;
3097 return sc
.nr_reclaimed
;
3100 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
3101 unsigned long nr_pages
,
3105 struct zonelist
*zonelist
;
3106 unsigned long nr_reclaimed
;
3108 unsigned int noreclaim_flag
;
3109 struct scan_control sc
= {
3110 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3111 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
3112 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
3113 .reclaim_idx
= MAX_NR_ZONES
- 1,
3114 .target_mem_cgroup
= memcg
,
3115 .priority
= DEF_PRIORITY
,
3116 .may_writepage
= !laptop_mode
,
3118 .may_swap
= may_swap
,
3122 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3123 * take care of from where we get pages. So the node where we start the
3124 * scan does not need to be the current node.
3126 nid
= mem_cgroup_select_victim_node(memcg
);
3128 zonelist
= &NODE_DATA(nid
)->node_zonelists
[ZONELIST_FALLBACK
];
3130 trace_mm_vmscan_memcg_reclaim_begin(0,
3135 noreclaim_flag
= memalloc_noreclaim_save();
3136 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3137 memalloc_noreclaim_restore(noreclaim_flag
);
3139 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
3141 return nr_reclaimed
;
3145 static void age_active_anon(struct pglist_data
*pgdat
,
3146 struct scan_control
*sc
)
3148 struct mem_cgroup
*memcg
;
3150 if (!total_swap_pages
)
3153 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3155 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
3157 if (inactive_list_is_low(lruvec
, false, memcg
, sc
, true))
3158 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
3159 sc
, LRU_ACTIVE_ANON
);
3161 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
3166 * Returns true if there is an eligible zone balanced for the request order
3169 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3172 unsigned long mark
= -1;
3175 for (i
= 0; i
<= classzone_idx
; i
++) {
3176 zone
= pgdat
->node_zones
+ i
;
3178 if (!managed_zone(zone
))
3181 mark
= high_wmark_pages(zone
);
3182 if (zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3187 * If a node has no populated zone within classzone_idx, it does not
3188 * need balancing by definition. This can happen if a zone-restricted
3189 * allocation tries to wake a remote kswapd.
3197 /* Clear pgdat state for congested, dirty or under writeback. */
3198 static void clear_pgdat_congested(pg_data_t
*pgdat
)
3200 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3201 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3202 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3206 * Prepare kswapd for sleeping. This verifies that there are no processes
3207 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3209 * Returns true if kswapd is ready to sleep
3211 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3214 * The throttled processes are normally woken up in balance_pgdat() as
3215 * soon as allow_direct_reclaim() is true. But there is a potential
3216 * race between when kswapd checks the watermarks and a process gets
3217 * throttled. There is also a potential race if processes get
3218 * throttled, kswapd wakes, a large process exits thereby balancing the
3219 * zones, which causes kswapd to exit balance_pgdat() before reaching
3220 * the wake up checks. If kswapd is going to sleep, no process should
3221 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3222 * the wake up is premature, processes will wake kswapd and get
3223 * throttled again. The difference from wake ups in balance_pgdat() is
3224 * that here we are under prepare_to_wait().
3226 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3227 wake_up_all(&pgdat
->pfmemalloc_wait
);
3229 /* Hopeless node, leave it to direct reclaim */
3230 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3233 if (pgdat_balanced(pgdat
, order
, classzone_idx
)) {
3234 clear_pgdat_congested(pgdat
);
3242 * kswapd shrinks a node of pages that are at or below the highest usable
3243 * zone that is currently unbalanced.
3245 * Returns true if kswapd scanned at least the requested number of pages to
3246 * reclaim or if the lack of progress was due to pages under writeback.
3247 * This is used to determine if the scanning priority needs to be raised.
3249 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3250 struct scan_control
*sc
)
3255 /* Reclaim a number of pages proportional to the number of zones */
3256 sc
->nr_to_reclaim
= 0;
3257 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3258 zone
= pgdat
->node_zones
+ z
;
3259 if (!managed_zone(zone
))
3262 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3266 * Historically care was taken to put equal pressure on all zones but
3267 * now pressure is applied based on node LRU order.
3269 shrink_node(pgdat
, sc
);
3272 * Fragmentation may mean that the system cannot be rebalanced for
3273 * high-order allocations. If twice the allocation size has been
3274 * reclaimed then recheck watermarks only at order-0 to prevent
3275 * excessive reclaim. Assume that a process requested a high-order
3276 * can direct reclaim/compact.
3278 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
3281 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3285 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3286 * that are eligible for use by the caller until at least one zone is
3289 * Returns the order kswapd finished reclaiming at.
3291 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3292 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3293 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3294 * or lower is eligible for reclaim until at least one usable zone is
3297 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3300 unsigned long nr_soft_reclaimed
;
3301 unsigned long nr_soft_scanned
;
3303 struct scan_control sc
= {
3304 .gfp_mask
= GFP_KERNEL
,
3306 .priority
= DEF_PRIORITY
,
3307 .may_writepage
= !laptop_mode
,
3311 count_vm_event(PAGEOUTRUN
);
3314 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
3315 bool raise_priority
= true;
3317 sc
.reclaim_idx
= classzone_idx
;
3320 * If the number of buffer_heads exceeds the maximum allowed
3321 * then consider reclaiming from all zones. This has a dual
3322 * purpose -- on 64-bit systems it is expected that
3323 * buffer_heads are stripped during active rotation. On 32-bit
3324 * systems, highmem pages can pin lowmem memory and shrinking
3325 * buffers can relieve lowmem pressure. Reclaim may still not
3326 * go ahead if all eligible zones for the original allocation
3327 * request are balanced to avoid excessive reclaim from kswapd.
3329 if (buffer_heads_over_limit
) {
3330 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3331 zone
= pgdat
->node_zones
+ i
;
3332 if (!managed_zone(zone
))
3341 * Only reclaim if there are no eligible zones. Note that
3342 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3345 if (pgdat_balanced(pgdat
, sc
.order
, classzone_idx
))
3349 * Do some background aging of the anon list, to give
3350 * pages a chance to be referenced before reclaiming. All
3351 * pages are rotated regardless of classzone as this is
3352 * about consistent aging.
3354 age_active_anon(pgdat
, &sc
);
3357 * If we're getting trouble reclaiming, start doing writepage
3358 * even in laptop mode.
3360 if (sc
.priority
< DEF_PRIORITY
- 2)
3361 sc
.may_writepage
= 1;
3363 /* Call soft limit reclaim before calling shrink_node. */
3365 nr_soft_scanned
= 0;
3366 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3367 sc
.gfp_mask
, &nr_soft_scanned
);
3368 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3371 * There should be no need to raise the scanning priority if
3372 * enough pages are already being scanned that that high
3373 * watermark would be met at 100% efficiency.
3375 if (kswapd_shrink_node(pgdat
, &sc
))
3376 raise_priority
= false;
3379 * If the low watermark is met there is no need for processes
3380 * to be throttled on pfmemalloc_wait as they should not be
3381 * able to safely make forward progress. Wake them
3383 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3384 allow_direct_reclaim(pgdat
))
3385 wake_up_all(&pgdat
->pfmemalloc_wait
);
3387 /* Check if kswapd should be suspending */
3388 if (try_to_freeze() || kthread_should_stop())
3392 * Raise priority if scanning rate is too low or there was no
3393 * progress in reclaiming pages
3395 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
3396 if (raise_priority
|| !nr_reclaimed
)
3398 } while (sc
.priority
>= 1);
3400 if (!sc
.nr_reclaimed
)
3401 pgdat
->kswapd_failures
++;
3404 snapshot_refaults(NULL
, pgdat
);
3406 * Return the order kswapd stopped reclaiming at as
3407 * prepare_kswapd_sleep() takes it into account. If another caller
3408 * entered the allocator slow path while kswapd was awake, order will
3409 * remain at the higher level.
3415 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3416 * allocation request woke kswapd for. When kswapd has not woken recently,
3417 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3418 * given classzone and returns it or the highest classzone index kswapd
3419 * was recently woke for.
3421 static enum zone_type
kswapd_classzone_idx(pg_data_t
*pgdat
,
3422 enum zone_type classzone_idx
)
3424 if (pgdat
->kswapd_classzone_idx
== MAX_NR_ZONES
)
3425 return classzone_idx
;
3427 return max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3430 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3431 unsigned int classzone_idx
)
3436 if (freezing(current
) || kthread_should_stop())
3439 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3442 * Try to sleep for a short interval. Note that kcompactd will only be
3443 * woken if it is possible to sleep for a short interval. This is
3444 * deliberate on the assumption that if reclaim cannot keep an
3445 * eligible zone balanced that it's also unlikely that compaction will
3448 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3450 * Compaction records what page blocks it recently failed to
3451 * isolate pages from and skips them in the future scanning.
3452 * When kswapd is going to sleep, it is reasonable to assume
3453 * that pages and compaction may succeed so reset the cache.
3455 reset_isolation_suitable(pgdat
);
3458 * We have freed the memory, now we should compact it to make
3459 * allocation of the requested order possible.
3461 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3463 remaining
= schedule_timeout(HZ
/10);
3466 * If woken prematurely then reset kswapd_classzone_idx and
3467 * order. The values will either be from a wakeup request or
3468 * the previous request that slept prematurely.
3471 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3472 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3475 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3476 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3480 * After a short sleep, check if it was a premature sleep. If not, then
3481 * go fully to sleep until explicitly woken up.
3484 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3485 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3488 * vmstat counters are not perfectly accurate and the estimated
3489 * value for counters such as NR_FREE_PAGES can deviate from the
3490 * true value by nr_online_cpus * threshold. To avoid the zone
3491 * watermarks being breached while under pressure, we reduce the
3492 * per-cpu vmstat threshold while kswapd is awake and restore
3493 * them before going back to sleep.
3495 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3497 if (!kthread_should_stop())
3500 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3503 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3505 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3507 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3511 * The background pageout daemon, started as a kernel thread
3512 * from the init process.
3514 * This basically trickles out pages so that we have _some_
3515 * free memory available even if there is no other activity
3516 * that frees anything up. This is needed for things like routing
3517 * etc, where we otherwise might have all activity going on in
3518 * asynchronous contexts that cannot page things out.
3520 * If there are applications that are active memory-allocators
3521 * (most normal use), this basically shouldn't matter.
3523 static int kswapd(void *p
)
3525 unsigned int alloc_order
, reclaim_order
;
3526 unsigned int classzone_idx
= MAX_NR_ZONES
- 1;
3527 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3528 struct task_struct
*tsk
= current
;
3530 struct reclaim_state reclaim_state
= {
3531 .reclaimed_slab
= 0,
3533 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3535 if (!cpumask_empty(cpumask
))
3536 set_cpus_allowed_ptr(tsk
, cpumask
);
3537 current
->reclaim_state
= &reclaim_state
;
3540 * Tell the memory management that we're a "memory allocator",
3541 * and that if we need more memory we should get access to it
3542 * regardless (see "__alloc_pages()"). "kswapd" should
3543 * never get caught in the normal page freeing logic.
3545 * (Kswapd normally doesn't need memory anyway, but sometimes
3546 * you need a small amount of memory in order to be able to
3547 * page out something else, and this flag essentially protects
3548 * us from recursively trying to free more memory as we're
3549 * trying to free the first piece of memory in the first place).
3551 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3554 pgdat
->kswapd_order
= 0;
3555 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3559 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3560 classzone_idx
= kswapd_classzone_idx(pgdat
, classzone_idx
);
3563 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3566 /* Read the new order and classzone_idx */
3567 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3568 classzone_idx
= kswapd_classzone_idx(pgdat
, 0);
3569 pgdat
->kswapd_order
= 0;
3570 pgdat
->kswapd_classzone_idx
= MAX_NR_ZONES
;
3572 ret
= try_to_freeze();
3573 if (kthread_should_stop())
3577 * We can speed up thawing tasks if we don't call balance_pgdat
3578 * after returning from the refrigerator
3584 * Reclaim begins at the requested order but if a high-order
3585 * reclaim fails then kswapd falls back to reclaiming for
3586 * order-0. If that happens, kswapd will consider sleeping
3587 * for the order it finished reclaiming at (reclaim_order)
3588 * but kcompactd is woken to compact for the original
3589 * request (alloc_order).
3591 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3593 fs_reclaim_acquire(GFP_KERNEL
);
3594 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3595 fs_reclaim_release(GFP_KERNEL
);
3596 if (reclaim_order
< alloc_order
)
3597 goto kswapd_try_sleep
;
3600 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3601 current
->reclaim_state
= NULL
;
3607 * A zone is low on free memory, so wake its kswapd task to service it.
3609 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3613 if (!managed_zone(zone
))
3616 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3618 pgdat
= zone
->zone_pgdat
;
3619 pgdat
->kswapd_classzone_idx
= kswapd_classzone_idx(pgdat
,
3621 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3622 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3625 /* Hopeless node, leave it to direct reclaim */
3626 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
3629 if (pgdat_balanced(pgdat
, order
, classzone_idx
))
3632 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, classzone_idx
, order
);
3633 wake_up_interruptible(&pgdat
->kswapd_wait
);
3636 #ifdef CONFIG_HIBERNATION
3638 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3641 * Rather than trying to age LRUs the aim is to preserve the overall
3642 * LRU order by reclaiming preferentially
3643 * inactive > active > active referenced > active mapped
3645 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3647 struct reclaim_state reclaim_state
;
3648 struct scan_control sc
= {
3649 .nr_to_reclaim
= nr_to_reclaim
,
3650 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3651 .reclaim_idx
= MAX_NR_ZONES
- 1,
3652 .priority
= DEF_PRIORITY
,
3656 .hibernation_mode
= 1,
3658 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3659 struct task_struct
*p
= current
;
3660 unsigned long nr_reclaimed
;
3661 unsigned int noreclaim_flag
;
3663 noreclaim_flag
= memalloc_noreclaim_save();
3664 fs_reclaim_acquire(sc
.gfp_mask
);
3665 reclaim_state
.reclaimed_slab
= 0;
3666 p
->reclaim_state
= &reclaim_state
;
3668 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3670 p
->reclaim_state
= NULL
;
3671 fs_reclaim_release(sc
.gfp_mask
);
3672 memalloc_noreclaim_restore(noreclaim_flag
);
3674 return nr_reclaimed
;
3676 #endif /* CONFIG_HIBERNATION */
3678 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3679 not required for correctness. So if the last cpu in a node goes
3680 away, we get changed to run anywhere: as the first one comes back,
3681 restore their cpu bindings. */
3682 static int kswapd_cpu_online(unsigned int cpu
)
3686 for_each_node_state(nid
, N_MEMORY
) {
3687 pg_data_t
*pgdat
= NODE_DATA(nid
);
3688 const struct cpumask
*mask
;
3690 mask
= cpumask_of_node(pgdat
->node_id
);
3692 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3693 /* One of our CPUs online: restore mask */
3694 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3700 * This kswapd start function will be called by init and node-hot-add.
3701 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3703 int kswapd_run(int nid
)
3705 pg_data_t
*pgdat
= NODE_DATA(nid
);
3711 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3712 if (IS_ERR(pgdat
->kswapd
)) {
3713 /* failure at boot is fatal */
3714 BUG_ON(system_state
< SYSTEM_RUNNING
);
3715 pr_err("Failed to start kswapd on node %d\n", nid
);
3716 ret
= PTR_ERR(pgdat
->kswapd
);
3717 pgdat
->kswapd
= NULL
;
3723 * Called by memory hotplug when all memory in a node is offlined. Caller must
3724 * hold mem_hotplug_begin/end().
3726 void kswapd_stop(int nid
)
3728 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3731 kthread_stop(kswapd
);
3732 NODE_DATA(nid
)->kswapd
= NULL
;
3736 static int __init
kswapd_init(void)
3741 for_each_node_state(nid
, N_MEMORY
)
3743 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3744 "mm/vmscan:online", kswapd_cpu_online
,
3750 module_init(kswapd_init
)
3756 * If non-zero call node_reclaim when the number of free pages falls below
3759 int node_reclaim_mode __read_mostly
;
3761 #define RECLAIM_OFF 0
3762 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3763 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3764 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3767 * Priority for NODE_RECLAIM. This determines the fraction of pages
3768 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3771 #define NODE_RECLAIM_PRIORITY 4
3774 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3777 int sysctl_min_unmapped_ratio
= 1;
3780 * If the number of slab pages in a zone grows beyond this percentage then
3781 * slab reclaim needs to occur.
3783 int sysctl_min_slab_ratio
= 5;
3785 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3787 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3788 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3789 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3792 * It's possible for there to be more file mapped pages than
3793 * accounted for by the pages on the file LRU lists because
3794 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3796 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3799 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3800 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3802 unsigned long nr_pagecache_reclaimable
;
3803 unsigned long delta
= 0;
3806 * If RECLAIM_UNMAP is set, then all file pages are considered
3807 * potentially reclaimable. Otherwise, we have to worry about
3808 * pages like swapcache and node_unmapped_file_pages() provides
3811 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3812 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3814 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3816 /* If we can't clean pages, remove dirty pages from consideration */
3817 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3818 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3820 /* Watch for any possible underflows due to delta */
3821 if (unlikely(delta
> nr_pagecache_reclaimable
))
3822 delta
= nr_pagecache_reclaimable
;
3824 return nr_pagecache_reclaimable
- delta
;
3828 * Try to free up some pages from this node through reclaim.
3830 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3832 /* Minimum pages needed in order to stay on node */
3833 const unsigned long nr_pages
= 1 << order
;
3834 struct task_struct
*p
= current
;
3835 struct reclaim_state reclaim_state
;
3836 unsigned int noreclaim_flag
;
3837 struct scan_control sc
= {
3838 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3839 .gfp_mask
= current_gfp_context(gfp_mask
),
3841 .priority
= NODE_RECLAIM_PRIORITY
,
3842 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3843 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3845 .reclaim_idx
= gfp_zone(gfp_mask
),
3850 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3851 * and we also need to be able to write out pages for RECLAIM_WRITE
3852 * and RECLAIM_UNMAP.
3854 noreclaim_flag
= memalloc_noreclaim_save();
3855 p
->flags
|= PF_SWAPWRITE
;
3856 fs_reclaim_acquire(sc
.gfp_mask
);
3857 reclaim_state
.reclaimed_slab
= 0;
3858 p
->reclaim_state
= &reclaim_state
;
3860 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3862 * Free memory by calling shrink zone with increasing
3863 * priorities until we have enough memory freed.
3866 shrink_node(pgdat
, &sc
);
3867 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3870 p
->reclaim_state
= NULL
;
3871 fs_reclaim_release(gfp_mask
);
3872 current
->flags
&= ~PF_SWAPWRITE
;
3873 memalloc_noreclaim_restore(noreclaim_flag
);
3874 return sc
.nr_reclaimed
>= nr_pages
;
3877 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3882 * Node reclaim reclaims unmapped file backed pages and
3883 * slab pages if we are over the defined limits.
3885 * A small portion of unmapped file backed pages is needed for
3886 * file I/O otherwise pages read by file I/O will be immediately
3887 * thrown out if the node is overallocated. So we do not reclaim
3888 * if less than a specified percentage of the node is used by
3889 * unmapped file backed pages.
3891 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3892 node_page_state(pgdat
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3893 return NODE_RECLAIM_FULL
;
3896 * Do not scan if the allocation should not be delayed.
3898 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3899 return NODE_RECLAIM_NOSCAN
;
3902 * Only run node reclaim on the local node or on nodes that do not
3903 * have associated processors. This will favor the local processor
3904 * over remote processors and spread off node memory allocations
3905 * as wide as possible.
3907 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3908 return NODE_RECLAIM_NOSCAN
;
3910 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3911 return NODE_RECLAIM_NOSCAN
;
3913 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3914 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3917 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3924 * page_evictable - test whether a page is evictable
3925 * @page: the page to test
3927 * Test whether page is evictable--i.e., should be placed on active/inactive
3928 * lists vs unevictable list.
3930 * Reasons page might not be evictable:
3931 * (1) page's mapping marked unevictable
3932 * (2) page is part of an mlocked VMA
3935 int page_evictable(struct page
*page
)
3937 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3942 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3943 * @pages: array of pages to check
3944 * @nr_pages: number of pages to check
3946 * Checks pages for evictability and moves them to the appropriate lru list.
3948 * This function is only used for SysV IPC SHM_UNLOCK.
3950 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3952 struct lruvec
*lruvec
;
3953 struct pglist_data
*pgdat
= NULL
;
3958 for (i
= 0; i
< nr_pages
; i
++) {
3959 struct page
*page
= pages
[i
];
3960 struct pglist_data
*pagepgdat
= page_pgdat(page
);
3963 if (pagepgdat
!= pgdat
) {
3965 spin_unlock_irq(&pgdat
->lru_lock
);
3967 spin_lock_irq(&pgdat
->lru_lock
);
3969 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
3971 if (!PageLRU(page
) || !PageUnevictable(page
))
3974 if (page_evictable(page
)) {
3975 enum lru_list lru
= page_lru_base_type(page
);
3977 VM_BUG_ON_PAGE(PageActive(page
), page
);
3978 ClearPageUnevictable(page
);
3979 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3980 add_page_to_lru_list(page
, lruvec
, lru
);
3986 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3987 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3988 spin_unlock_irq(&pgdat
->lru_lock
);
3991 #endif /* CONFIG_SHMEM */