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/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup
*target_mem_cgroup
;
84 /* Scan (total_size >> priority) pages at once */
87 unsigned int may_writepage
:1;
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap
:1;
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap
:1;
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash
:1;
98 unsigned int hibernation_mode
:1;
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready
:1;
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned
;
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed
;
110 #ifdef ARCH_HAS_PREFETCH
111 #define prefetch_prev_lru_page(_page, _base, _field) \
113 if ((_page)->lru.prev != _base) { \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
121 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
124 #ifdef ARCH_HAS_PREFETCHW
125 #define prefetchw_prev_lru_page(_page, _base, _field) \
127 if ((_page)->lru.prev != _base) { \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
135 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139 * From 0 .. 100. Higher means more swappy.
141 int vm_swappiness
= 60;
143 * The total number of pages which are beyond the high watermark within all
146 unsigned long vm_total_pages
;
148 static LIST_HEAD(shrinker_list
);
149 static DECLARE_RWSEM(shrinker_rwsem
);
152 static bool global_reclaim(struct scan_control
*sc
)
154 return !sc
->target_mem_cgroup
;
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
170 static bool sane_reclaim(struct scan_control
*sc
)
172 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
176 #ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
183 static bool global_reclaim(struct scan_control
*sc
)
188 static bool sane_reclaim(struct scan_control
*sc
)
195 * This misses isolated pages which are not accounted for to save counters.
196 * As the data only determines if reclaim or compaction continues, it is
197 * not expected that isolated pages will be a dominating factor.
199 unsigned long zone_reclaimable_pages(struct zone
*zone
)
203 nr
= zone_page_state_snapshot(zone
, NR_ZONE_LRU_FILE
);
204 if (get_nr_swap_pages() > 0)
205 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_LRU_ANON
);
210 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
214 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
215 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
216 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
218 if (get_nr_swap_pages() > 0)
219 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
220 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
221 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
226 bool pgdat_reclaimable(struct pglist_data
*pgdat
)
228 return node_page_state_snapshot(pgdat
, NR_PAGES_SCANNED
) <
229 pgdat_reclaimable_pages(pgdat
) * 6;
232 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
234 if (!mem_cgroup_disabled())
235 return mem_cgroup_get_lru_size(lruvec
, lru
);
237 return node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
241 * Add a shrinker callback to be called from the vm.
243 int register_shrinker(struct shrinker
*shrinker
)
245 size_t size
= sizeof(*shrinker
->nr_deferred
);
247 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
250 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
251 if (!shrinker
->nr_deferred
)
254 down_write(&shrinker_rwsem
);
255 list_add_tail(&shrinker
->list
, &shrinker_list
);
256 up_write(&shrinker_rwsem
);
259 EXPORT_SYMBOL(register_shrinker
);
264 void unregister_shrinker(struct shrinker
*shrinker
)
266 down_write(&shrinker_rwsem
);
267 list_del(&shrinker
->list
);
268 up_write(&shrinker_rwsem
);
269 kfree(shrinker
->nr_deferred
);
271 EXPORT_SYMBOL(unregister_shrinker
);
273 #define SHRINK_BATCH 128
275 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
276 struct shrinker
*shrinker
,
277 unsigned long nr_scanned
,
278 unsigned long nr_eligible
)
280 unsigned long freed
= 0;
281 unsigned long long delta
;
286 int nid
= shrinkctl
->nid
;
287 long batch_size
= shrinker
->batch
? shrinker
->batch
290 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
295 * copy the current shrinker scan count into a local variable
296 * and zero it so that other concurrent shrinker invocations
297 * don't also do this scanning work.
299 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
302 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
304 do_div(delta
, nr_eligible
+ 1);
306 if (total_scan
< 0) {
307 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
308 shrinker
->scan_objects
, total_scan
);
309 total_scan
= freeable
;
313 * We need to avoid excessive windup on filesystem shrinkers
314 * due to large numbers of GFP_NOFS allocations causing the
315 * shrinkers to return -1 all the time. This results in a large
316 * nr being built up so when a shrink that can do some work
317 * comes along it empties the entire cache due to nr >>>
318 * freeable. This is bad for sustaining a working set in
321 * Hence only allow the shrinker to scan the entire cache when
322 * a large delta change is calculated directly.
324 if (delta
< freeable
/ 4)
325 total_scan
= min(total_scan
, freeable
/ 2);
328 * Avoid risking looping forever due to too large nr value:
329 * never try to free more than twice the estimate number of
332 if (total_scan
> freeable
* 2)
333 total_scan
= freeable
* 2;
335 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
336 nr_scanned
, nr_eligible
,
337 freeable
, delta
, total_scan
);
340 * Normally, we should not scan less than batch_size objects in one
341 * pass to avoid too frequent shrinker calls, but if the slab has less
342 * than batch_size objects in total and we are really tight on memory,
343 * we will try to reclaim all available objects, otherwise we can end
344 * up failing allocations although there are plenty of reclaimable
345 * objects spread over several slabs with usage less than the
348 * We detect the "tight on memory" situations by looking at the total
349 * number of objects we want to scan (total_scan). If it is greater
350 * than the total number of objects on slab (freeable), we must be
351 * scanning at high prio and therefore should try to reclaim as much as
354 while (total_scan
>= batch_size
||
355 total_scan
>= freeable
) {
357 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
359 shrinkctl
->nr_to_scan
= nr_to_scan
;
360 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
361 if (ret
== SHRINK_STOP
)
365 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
366 total_scan
-= nr_to_scan
;
372 * move the unused scan count back into the shrinker in a
373 * manner that handles concurrent updates. If we exhausted the
374 * scan, there is no need to do an update.
377 new_nr
= atomic_long_add_return(total_scan
,
378 &shrinker
->nr_deferred
[nid
]);
380 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
382 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
387 * shrink_slab - shrink slab caches
388 * @gfp_mask: allocation context
389 * @nid: node whose slab caches to target
390 * @memcg: memory cgroup whose slab caches to target
391 * @nr_scanned: pressure numerator
392 * @nr_eligible: pressure denominator
394 * Call the shrink functions to age shrinkable caches.
396 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
397 * unaware shrinkers will receive a node id of 0 instead.
399 * @memcg specifies the memory cgroup to target. If it is not NULL,
400 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
401 * objects from the memory cgroup specified. Otherwise, only unaware
402 * shrinkers are called.
404 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
405 * the available objects should be scanned. Page reclaim for example
406 * passes the number of pages scanned and the number of pages on the
407 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
408 * when it encountered mapped pages. The ratio is further biased by
409 * the ->seeks setting of the shrink function, which indicates the
410 * cost to recreate an object relative to that of an LRU page.
412 * Returns the number of reclaimed slab objects.
414 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
415 struct mem_cgroup
*memcg
,
416 unsigned long nr_scanned
,
417 unsigned long nr_eligible
)
419 struct shrinker
*shrinker
;
420 unsigned long freed
= 0;
422 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
426 nr_scanned
= SWAP_CLUSTER_MAX
;
428 if (!down_read_trylock(&shrinker_rwsem
)) {
430 * If we would return 0, our callers would understand that we
431 * have nothing else to shrink and give up trying. By returning
432 * 1 we keep it going and assume we'll be able to shrink next
439 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
440 struct shrink_control sc
= {
441 .gfp_mask
= gfp_mask
,
447 * If kernel memory accounting is disabled, we ignore
448 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
449 * passing NULL for memcg.
451 if (memcg_kmem_enabled() &&
452 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
455 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
458 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
461 up_read(&shrinker_rwsem
);
467 void drop_slab_node(int nid
)
472 struct mem_cgroup
*memcg
= NULL
;
476 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
478 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
479 } while (freed
> 10);
486 for_each_online_node(nid
)
490 static inline int is_page_cache_freeable(struct page
*page
)
493 * A freeable page cache page is referenced only by the caller
494 * that isolated the page, the page cache radix tree and
495 * optional buffer heads at page->private.
497 return page_count(page
) - page_has_private(page
) == 2;
500 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
502 if (current
->flags
& PF_SWAPWRITE
)
504 if (!inode_write_congested(inode
))
506 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
512 * We detected a synchronous write error writing a page out. Probably
513 * -ENOSPC. We need to propagate that into the address_space for a subsequent
514 * fsync(), msync() or close().
516 * The tricky part is that after writepage we cannot touch the mapping: nothing
517 * prevents it from being freed up. But we have a ref on the page and once
518 * that page is locked, the mapping is pinned.
520 * We're allowed to run sleeping lock_page() here because we know the caller has
523 static void handle_write_error(struct address_space
*mapping
,
524 struct page
*page
, int error
)
527 if (page_mapping(page
) == mapping
)
528 mapping_set_error(mapping
, error
);
532 /* possible outcome of pageout() */
534 /* failed to write page out, page is locked */
536 /* move page to the active list, page is locked */
538 /* page has been sent to the disk successfully, page is unlocked */
540 /* page is clean and locked */
545 * pageout is called by shrink_page_list() for each dirty page.
546 * Calls ->writepage().
548 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
549 struct scan_control
*sc
)
552 * If the page is dirty, only perform writeback if that write
553 * will be non-blocking. To prevent this allocation from being
554 * stalled by pagecache activity. But note that there may be
555 * stalls if we need to run get_block(). We could test
556 * PagePrivate for that.
558 * If this process is currently in __generic_file_write_iter() against
559 * this page's queue, we can perform writeback even if that
562 * If the page is swapcache, write it back even if that would
563 * block, for some throttling. This happens by accident, because
564 * swap_backing_dev_info is bust: it doesn't reflect the
565 * congestion state of the swapdevs. Easy to fix, if needed.
567 if (!is_page_cache_freeable(page
))
571 * Some data journaling orphaned pages can have
572 * page->mapping == NULL while being dirty with clean buffers.
574 if (page_has_private(page
)) {
575 if (try_to_free_buffers(page
)) {
576 ClearPageDirty(page
);
577 pr_info("%s: orphaned page\n", __func__
);
583 if (mapping
->a_ops
->writepage
== NULL
)
584 return PAGE_ACTIVATE
;
585 if (!may_write_to_inode(mapping
->host
, sc
))
588 if (clear_page_dirty_for_io(page
)) {
590 struct writeback_control wbc
= {
591 .sync_mode
= WB_SYNC_NONE
,
592 .nr_to_write
= SWAP_CLUSTER_MAX
,
594 .range_end
= LLONG_MAX
,
598 SetPageReclaim(page
);
599 res
= mapping
->a_ops
->writepage(page
, &wbc
);
601 handle_write_error(mapping
, page
, res
);
602 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
603 ClearPageReclaim(page
);
604 return PAGE_ACTIVATE
;
607 if (!PageWriteback(page
)) {
608 /* synchronous write or broken a_ops? */
609 ClearPageReclaim(page
);
611 trace_mm_vmscan_writepage(page
);
612 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
620 * Same as remove_mapping, but if the page is removed from the mapping, it
621 * gets returned with a refcount of 0.
623 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
628 BUG_ON(!PageLocked(page
));
629 BUG_ON(mapping
!= page_mapping(page
));
631 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
633 * The non racy check for a busy page.
635 * Must be careful with the order of the tests. When someone has
636 * a ref to the page, it may be possible that they dirty it then
637 * drop the reference. So if PageDirty is tested before page_count
638 * here, then the following race may occur:
640 * get_user_pages(&page);
641 * [user mapping goes away]
643 * !PageDirty(page) [good]
644 * SetPageDirty(page);
646 * !page_count(page) [good, discard it]
648 * [oops, our write_to data is lost]
650 * Reversing the order of the tests ensures such a situation cannot
651 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
652 * load is not satisfied before that of page->_refcount.
654 * Note that if SetPageDirty is always performed via set_page_dirty,
655 * and thus under tree_lock, then this ordering is not required.
657 if (!page_ref_freeze(page
, 2))
659 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
660 if (unlikely(PageDirty(page
))) {
661 page_ref_unfreeze(page
, 2);
665 if (PageSwapCache(page
)) {
666 swp_entry_t swap
= { .val
= page_private(page
) };
667 mem_cgroup_swapout(page
, swap
);
668 __delete_from_swap_cache(page
);
669 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
670 swapcache_free(swap
);
672 void (*freepage
)(struct page
*);
675 freepage
= mapping
->a_ops
->freepage
;
677 * Remember a shadow entry for reclaimed file cache in
678 * order to detect refaults, thus thrashing, later on.
680 * But don't store shadows in an address space that is
681 * already exiting. This is not just an optizimation,
682 * inode reclaim needs to empty out the radix tree or
683 * the nodes are lost. Don't plant shadows behind its
686 * We also don't store shadows for DAX mappings because the
687 * only page cache pages found in these are zero pages
688 * covering holes, and because we don't want to mix DAX
689 * exceptional entries and shadow exceptional entries in the
692 if (reclaimed
&& page_is_file_cache(page
) &&
693 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
694 shadow
= workingset_eviction(mapping
, page
);
695 __delete_from_page_cache(page
, shadow
);
696 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
698 if (freepage
!= NULL
)
705 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
710 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
711 * someone else has a ref on the page, abort and return 0. If it was
712 * successfully detached, return 1. Assumes the caller has a single ref on
715 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
717 if (__remove_mapping(mapping
, page
, false)) {
719 * Unfreezing the refcount with 1 rather than 2 effectively
720 * drops the pagecache ref for us without requiring another
723 page_ref_unfreeze(page
, 1);
730 * putback_lru_page - put previously isolated page onto appropriate LRU list
731 * @page: page to be put back to appropriate lru list
733 * Add previously isolated @page to appropriate LRU list.
734 * Page may still be unevictable for other reasons.
736 * lru_lock must not be held, interrupts must be enabled.
738 void putback_lru_page(struct page
*page
)
741 int was_unevictable
= PageUnevictable(page
);
743 VM_BUG_ON_PAGE(PageLRU(page
), page
);
746 ClearPageUnevictable(page
);
748 if (page_evictable(page
)) {
750 * For evictable pages, we can use the cache.
751 * In event of a race, worst case is we end up with an
752 * unevictable page on [in]active list.
753 * We know how to handle that.
755 is_unevictable
= false;
759 * Put unevictable pages directly on zone's unevictable
762 is_unevictable
= true;
763 add_page_to_unevictable_list(page
);
765 * When racing with an mlock or AS_UNEVICTABLE clearing
766 * (page is unlocked) make sure that if the other thread
767 * does not observe our setting of PG_lru and fails
768 * isolation/check_move_unevictable_pages,
769 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
770 * the page back to the evictable list.
772 * The other side is TestClearPageMlocked() or shmem_lock().
778 * page's status can change while we move it among lru. If an evictable
779 * page is on unevictable list, it never be freed. To avoid that,
780 * check after we added it to the list, again.
782 if (is_unevictable
&& page_evictable(page
)) {
783 if (!isolate_lru_page(page
)) {
787 /* This means someone else dropped this page from LRU
788 * So, it will be freed or putback to LRU again. There is
789 * nothing to do here.
793 if (was_unevictable
&& !is_unevictable
)
794 count_vm_event(UNEVICTABLE_PGRESCUED
);
795 else if (!was_unevictable
&& is_unevictable
)
796 count_vm_event(UNEVICTABLE_PGCULLED
);
798 put_page(page
); /* drop ref from isolate */
801 enum page_references
{
803 PAGEREF_RECLAIM_CLEAN
,
808 static enum page_references
page_check_references(struct page
*page
,
809 struct scan_control
*sc
)
811 int referenced_ptes
, referenced_page
;
812 unsigned long vm_flags
;
814 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
816 referenced_page
= TestClearPageReferenced(page
);
819 * Mlock lost the isolation race with us. Let try_to_unmap()
820 * move the page to the unevictable list.
822 if (vm_flags
& VM_LOCKED
)
823 return PAGEREF_RECLAIM
;
825 if (referenced_ptes
) {
826 if (PageSwapBacked(page
))
827 return PAGEREF_ACTIVATE
;
829 * All mapped pages start out with page table
830 * references from the instantiating fault, so we need
831 * to look twice if a mapped file page is used more
834 * Mark it and spare it for another trip around the
835 * inactive list. Another page table reference will
836 * lead to its activation.
838 * Note: the mark is set for activated pages as well
839 * so that recently deactivated but used pages are
842 SetPageReferenced(page
);
844 if (referenced_page
|| referenced_ptes
> 1)
845 return PAGEREF_ACTIVATE
;
848 * Activate file-backed executable pages after first usage.
850 if (vm_flags
& VM_EXEC
)
851 return PAGEREF_ACTIVATE
;
856 /* Reclaim if clean, defer dirty pages to writeback */
857 if (referenced_page
&& !PageSwapBacked(page
))
858 return PAGEREF_RECLAIM_CLEAN
;
860 return PAGEREF_RECLAIM
;
863 /* Check if a page is dirty or under writeback */
864 static void page_check_dirty_writeback(struct page
*page
,
865 bool *dirty
, bool *writeback
)
867 struct address_space
*mapping
;
870 * Anonymous pages are not handled by flushers and must be written
871 * from reclaim context. Do not stall reclaim based on them
873 if (!page_is_file_cache(page
)) {
879 /* By default assume that the page flags are accurate */
880 *dirty
= PageDirty(page
);
881 *writeback
= PageWriteback(page
);
883 /* Verify dirty/writeback state if the filesystem supports it */
884 if (!page_has_private(page
))
887 mapping
= page_mapping(page
);
888 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
889 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
893 * shrink_page_list() returns the number of reclaimed pages
895 static unsigned long shrink_page_list(struct list_head
*page_list
,
896 struct pglist_data
*pgdat
,
897 struct scan_control
*sc
,
898 enum ttu_flags ttu_flags
,
899 unsigned long *ret_nr_dirty
,
900 unsigned long *ret_nr_unqueued_dirty
,
901 unsigned long *ret_nr_congested
,
902 unsigned long *ret_nr_writeback
,
903 unsigned long *ret_nr_immediate
,
906 LIST_HEAD(ret_pages
);
907 LIST_HEAD(free_pages
);
909 unsigned long nr_unqueued_dirty
= 0;
910 unsigned long nr_dirty
= 0;
911 unsigned long nr_congested
= 0;
912 unsigned long nr_reclaimed
= 0;
913 unsigned long nr_writeback
= 0;
914 unsigned long nr_immediate
= 0;
918 while (!list_empty(page_list
)) {
919 struct address_space
*mapping
;
922 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
923 bool dirty
, writeback
;
924 bool lazyfree
= false;
925 int ret
= SWAP_SUCCESS
;
929 page
= lru_to_page(page_list
);
930 list_del(&page
->lru
);
932 if (!trylock_page(page
))
935 VM_BUG_ON_PAGE(PageActive(page
), page
);
939 if (unlikely(!page_evictable(page
)))
942 if (!sc
->may_unmap
&& page_mapped(page
))
945 /* Double the slab pressure for mapped and swapcache pages */
946 if (page_mapped(page
) || PageSwapCache(page
))
949 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
950 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
953 * The number of dirty pages determines if a zone is marked
954 * reclaim_congested which affects wait_iff_congested. kswapd
955 * will stall and start writing pages if the tail of the LRU
956 * is all dirty unqueued pages.
958 page_check_dirty_writeback(page
, &dirty
, &writeback
);
959 if (dirty
|| writeback
)
962 if (dirty
&& !writeback
)
966 * Treat this page as congested if the underlying BDI is or if
967 * pages are cycling through the LRU so quickly that the
968 * pages marked for immediate reclaim are making it to the
969 * end of the LRU a second time.
971 mapping
= page_mapping(page
);
972 if (((dirty
|| writeback
) && mapping
&&
973 inode_write_congested(mapping
->host
)) ||
974 (writeback
&& PageReclaim(page
)))
978 * If a page at the tail of the LRU is under writeback, there
979 * are three cases to consider.
981 * 1) If reclaim is encountering an excessive number of pages
982 * under writeback and this page is both under writeback and
983 * PageReclaim then it indicates that pages are being queued
984 * for IO but are being recycled through the LRU before the
985 * IO can complete. Waiting on the page itself risks an
986 * indefinite stall if it is impossible to writeback the
987 * page due to IO error or disconnected storage so instead
988 * note that the LRU is being scanned too quickly and the
989 * caller can stall after page list has been processed.
991 * 2) Global or new memcg reclaim encounters a page that is
992 * not marked for immediate reclaim, or the caller does not
993 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
994 * not to fs). In this case mark the page for immediate
995 * reclaim and continue scanning.
997 * Require may_enter_fs because we would wait on fs, which
998 * may not have submitted IO yet. And the loop driver might
999 * enter reclaim, and deadlock if it waits on a page for
1000 * which it is needed to do the write (loop masks off
1001 * __GFP_IO|__GFP_FS for this reason); but more thought
1002 * would probably show more reasons.
1004 * 3) Legacy memcg encounters a page that is already marked
1005 * PageReclaim. memcg does not have any dirty pages
1006 * throttling so we could easily OOM just because too many
1007 * pages are in writeback and there is nothing else to
1008 * reclaim. Wait for the writeback to complete.
1010 if (PageWriteback(page
)) {
1012 if (current_is_kswapd() &&
1013 PageReclaim(page
) &&
1014 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1019 } else if (sane_reclaim(sc
) ||
1020 !PageReclaim(page
) || !may_enter_fs
) {
1022 * This is slightly racy - end_page_writeback()
1023 * might have just cleared PageReclaim, then
1024 * setting PageReclaim here end up interpreted
1025 * as PageReadahead - but that does not matter
1026 * enough to care. What we do want is for this
1027 * page to have PageReclaim set next time memcg
1028 * reclaim reaches the tests above, so it will
1029 * then wait_on_page_writeback() to avoid OOM;
1030 * and it's also appropriate in global reclaim.
1032 SetPageReclaim(page
);
1039 wait_on_page_writeback(page
);
1040 /* then go back and try same page again */
1041 list_add_tail(&page
->lru
, page_list
);
1047 references
= page_check_references(page
, sc
);
1049 switch (references
) {
1050 case PAGEREF_ACTIVATE
:
1051 goto activate_locked
;
1054 case PAGEREF_RECLAIM
:
1055 case PAGEREF_RECLAIM_CLEAN
:
1056 ; /* try to reclaim the page below */
1060 * Anonymous process memory has backing store?
1061 * Try to allocate it some swap space here.
1063 if (PageAnon(page
) && !PageSwapCache(page
)) {
1064 if (!(sc
->gfp_mask
& __GFP_IO
))
1066 if (!add_to_swap(page
, page_list
))
1067 goto activate_locked
;
1071 /* Adding to swap updated mapping */
1072 mapping
= page_mapping(page
);
1073 } else if (unlikely(PageTransHuge(page
))) {
1074 /* Split file THP */
1075 if (split_huge_page_to_list(page
, page_list
))
1079 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1082 * The page is mapped into the page tables of one or more
1083 * processes. Try to unmap it here.
1085 if (page_mapped(page
) && mapping
) {
1086 switch (ret
= try_to_unmap(page
, lazyfree
?
1087 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1088 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1090 goto activate_locked
;
1098 ; /* try to free the page below */
1102 if (PageDirty(page
)) {
1104 * Only kswapd can writeback filesystem pages to
1105 * avoid risk of stack overflow but only writeback
1106 * if many dirty pages have been encountered.
1108 if (page_is_file_cache(page
) &&
1109 (!current_is_kswapd() ||
1110 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1112 * Immediately reclaim when written back.
1113 * Similar in principal to deactivate_page()
1114 * except we already have the page isolated
1115 * and know it's dirty
1117 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1118 SetPageReclaim(page
);
1123 if (references
== PAGEREF_RECLAIM_CLEAN
)
1127 if (!sc
->may_writepage
)
1131 * Page is dirty. Flush the TLB if a writable entry
1132 * potentially exists to avoid CPU writes after IO
1133 * starts and then write it out here.
1135 try_to_unmap_flush_dirty();
1136 switch (pageout(page
, mapping
, sc
)) {
1140 goto activate_locked
;
1142 if (PageWriteback(page
))
1144 if (PageDirty(page
))
1148 * A synchronous write - probably a ramdisk. Go
1149 * ahead and try to reclaim the page.
1151 if (!trylock_page(page
))
1153 if (PageDirty(page
) || PageWriteback(page
))
1155 mapping
= page_mapping(page
);
1157 ; /* try to free the page below */
1162 * If the page has buffers, try to free the buffer mappings
1163 * associated with this page. If we succeed we try to free
1166 * We do this even if the page is PageDirty().
1167 * try_to_release_page() does not perform I/O, but it is
1168 * possible for a page to have PageDirty set, but it is actually
1169 * clean (all its buffers are clean). This happens if the
1170 * buffers were written out directly, with submit_bh(). ext3
1171 * will do this, as well as the blockdev mapping.
1172 * try_to_release_page() will discover that cleanness and will
1173 * drop the buffers and mark the page clean - it can be freed.
1175 * Rarely, pages can have buffers and no ->mapping. These are
1176 * the pages which were not successfully invalidated in
1177 * truncate_complete_page(). We try to drop those buffers here
1178 * and if that worked, and the page is no longer mapped into
1179 * process address space (page_count == 1) it can be freed.
1180 * Otherwise, leave the page on the LRU so it is swappable.
1182 if (page_has_private(page
)) {
1183 if (!try_to_release_page(page
, sc
->gfp_mask
))
1184 goto activate_locked
;
1185 if (!mapping
&& page_count(page
) == 1) {
1187 if (put_page_testzero(page
))
1191 * rare race with speculative reference.
1192 * the speculative reference will free
1193 * this page shortly, so we may
1194 * increment nr_reclaimed here (and
1195 * leave it off the LRU).
1204 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1208 * At this point, we have no other references and there is
1209 * no way to pick any more up (removed from LRU, removed
1210 * from pagecache). Can use non-atomic bitops now (and
1211 * we obviously don't have to worry about waking up a process
1212 * waiting on the page lock, because there are no references.
1214 __ClearPageLocked(page
);
1216 if (ret
== SWAP_LZFREE
)
1217 count_vm_event(PGLAZYFREED
);
1222 * Is there need to periodically free_page_list? It would
1223 * appear not as the counts should be low
1225 list_add(&page
->lru
, &free_pages
);
1229 if (PageSwapCache(page
))
1230 try_to_free_swap(page
);
1232 list_add(&page
->lru
, &ret_pages
);
1236 /* Not a candidate for swapping, so reclaim swap space. */
1237 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1238 try_to_free_swap(page
);
1239 VM_BUG_ON_PAGE(PageActive(page
), page
);
1240 SetPageActive(page
);
1245 list_add(&page
->lru
, &ret_pages
);
1246 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1249 mem_cgroup_uncharge_list(&free_pages
);
1250 try_to_unmap_flush();
1251 free_hot_cold_page_list(&free_pages
, true);
1253 list_splice(&ret_pages
, page_list
);
1254 count_vm_events(PGACTIVATE
, pgactivate
);
1256 *ret_nr_dirty
+= nr_dirty
;
1257 *ret_nr_congested
+= nr_congested
;
1258 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1259 *ret_nr_writeback
+= nr_writeback
;
1260 *ret_nr_immediate
+= nr_immediate
;
1261 return nr_reclaimed
;
1264 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1265 struct list_head
*page_list
)
1267 struct scan_control sc
= {
1268 .gfp_mask
= GFP_KERNEL
,
1269 .priority
= DEF_PRIORITY
,
1272 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1273 struct page
*page
, *next
;
1274 LIST_HEAD(clean_pages
);
1276 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1277 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1278 !__PageMovable(page
)) {
1279 ClearPageActive(page
);
1280 list_move(&page
->lru
, &clean_pages
);
1284 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1285 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1286 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1287 list_splice(&clean_pages
, page_list
);
1288 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1293 * Attempt to remove the specified page from its LRU. Only take this page
1294 * if it is of the appropriate PageActive status. Pages which are being
1295 * freed elsewhere are also ignored.
1297 * page: page to consider
1298 * mode: one of the LRU isolation modes defined above
1300 * returns 0 on success, -ve errno on failure.
1302 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1306 /* Only take pages on the LRU. */
1310 /* Compaction should not handle unevictable pages but CMA can do so */
1311 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1317 * To minimise LRU disruption, the caller can indicate that it only
1318 * wants to isolate pages it will be able to operate on without
1319 * blocking - clean pages for the most part.
1321 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1322 * is used by reclaim when it is cannot write to backing storage
1324 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1325 * that it is possible to migrate without blocking
1327 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1328 /* All the caller can do on PageWriteback is block */
1329 if (PageWriteback(page
))
1332 if (PageDirty(page
)) {
1333 struct address_space
*mapping
;
1335 /* ISOLATE_CLEAN means only clean pages */
1336 if (mode
& ISOLATE_CLEAN
)
1340 * Only pages without mappings or that have a
1341 * ->migratepage callback are possible to migrate
1344 mapping
= page_mapping(page
);
1345 if (mapping
&& !mapping
->a_ops
->migratepage
)
1350 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1353 if (likely(get_page_unless_zero(page
))) {
1355 * Be careful not to clear PageLRU until after we're
1356 * sure the page is not being freed elsewhere -- the
1357 * page release code relies on it.
1367 * zone_lru_lock is heavily contended. Some of the functions that
1368 * shrink the lists perform better by taking out a batch of pages
1369 * and working on them outside the LRU lock.
1371 * For pagecache intensive workloads, this function is the hottest
1372 * spot in the kernel (apart from copy_*_user functions).
1374 * Appropriate locks must be held before calling this function.
1376 * @nr_to_scan: The number of pages to look through on the list.
1377 * @lruvec: The LRU vector to pull pages from.
1378 * @dst: The temp list to put pages on to.
1379 * @nr_scanned: The number of pages that were scanned.
1380 * @sc: The scan_control struct for this reclaim session
1381 * @mode: One of the LRU isolation modes
1382 * @lru: LRU list id for isolating
1384 * returns how many pages were moved onto *@dst.
1386 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1387 struct lruvec
*lruvec
, struct list_head
*dst
,
1388 unsigned long *nr_scanned
, struct scan_control
*sc
,
1389 isolate_mode_t mode
, enum lru_list lru
)
1391 struct list_head
*src
= &lruvec
->lists
[lru
];
1392 unsigned long nr_taken
= 0;
1393 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1394 unsigned long scan
, nr_pages
;
1396 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1397 !list_empty(src
); scan
++) {
1400 page
= lru_to_page(src
);
1401 prefetchw_prev_lru_page(page
, src
, flags
);
1403 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1405 switch (__isolate_lru_page(page
, mode
)) {
1407 nr_pages
= hpage_nr_pages(page
);
1408 nr_taken
+= nr_pages
;
1409 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1410 list_move(&page
->lru
, dst
);
1414 /* else it is being freed elsewhere */
1415 list_move(&page
->lru
, src
);
1424 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1425 nr_taken
, mode
, is_file_lru(lru
));
1426 for (scan
= 0; scan
< MAX_NR_ZONES
; scan
++) {
1427 nr_pages
= nr_zone_taken
[scan
];
1431 update_lru_size(lruvec
, lru
, scan
, -nr_pages
);
1437 * isolate_lru_page - tries to isolate a page from its LRU list
1438 * @page: page to isolate from its LRU list
1440 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1441 * vmstat statistic corresponding to whatever LRU list the page was on.
1443 * Returns 0 if the page was removed from an LRU list.
1444 * Returns -EBUSY if the page was not on an LRU list.
1446 * The returned page will have PageLRU() cleared. If it was found on
1447 * the active list, it will have PageActive set. If it was found on
1448 * the unevictable list, it will have the PageUnevictable bit set. That flag
1449 * may need to be cleared by the caller before letting the page go.
1451 * The vmstat statistic corresponding to the list on which the page was
1452 * found will be decremented.
1455 * (1) Must be called with an elevated refcount on the page. This is a
1456 * fundamentnal difference from isolate_lru_pages (which is called
1457 * without a stable reference).
1458 * (2) the lru_lock must not be held.
1459 * (3) interrupts must be enabled.
1461 int isolate_lru_page(struct page
*page
)
1465 VM_BUG_ON_PAGE(!page_count(page
), page
);
1466 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1468 if (PageLRU(page
)) {
1469 struct zone
*zone
= page_zone(page
);
1470 struct lruvec
*lruvec
;
1472 spin_lock_irq(zone_lru_lock(zone
));
1473 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1474 if (PageLRU(page
)) {
1475 int lru
= page_lru(page
);
1478 del_page_from_lru_list(page
, lruvec
, lru
);
1481 spin_unlock_irq(zone_lru_lock(zone
));
1487 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1488 * then get resheduled. When there are massive number of tasks doing page
1489 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1490 * the LRU list will go small and be scanned faster than necessary, leading to
1491 * unnecessary swapping, thrashing and OOM.
1493 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1494 struct scan_control
*sc
)
1496 unsigned long inactive
, isolated
;
1498 if (current_is_kswapd())
1501 if (!sane_reclaim(sc
))
1505 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1506 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1508 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1509 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1513 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1514 * won't get blocked by normal direct-reclaimers, forming a circular
1517 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1520 return isolated
> inactive
;
1523 static noinline_for_stack
void
1524 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1526 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1527 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1528 LIST_HEAD(pages_to_free
);
1531 * Put back any unfreeable pages.
1533 while (!list_empty(page_list
)) {
1534 struct page
*page
= lru_to_page(page_list
);
1537 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1538 list_del(&page
->lru
);
1539 if (unlikely(!page_evictable(page
))) {
1540 spin_unlock_irq(&pgdat
->lru_lock
);
1541 putback_lru_page(page
);
1542 spin_lock_irq(&pgdat
->lru_lock
);
1546 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1549 lru
= page_lru(page
);
1550 add_page_to_lru_list(page
, lruvec
, lru
);
1552 if (is_active_lru(lru
)) {
1553 int file
= is_file_lru(lru
);
1554 int numpages
= hpage_nr_pages(page
);
1555 reclaim_stat
->recent_rotated
[file
] += numpages
;
1557 if (put_page_testzero(page
)) {
1558 __ClearPageLRU(page
);
1559 __ClearPageActive(page
);
1560 del_page_from_lru_list(page
, lruvec
, lru
);
1562 if (unlikely(PageCompound(page
))) {
1563 spin_unlock_irq(&pgdat
->lru_lock
);
1564 mem_cgroup_uncharge(page
);
1565 (*get_compound_page_dtor(page
))(page
);
1566 spin_lock_irq(&pgdat
->lru_lock
);
1568 list_add(&page
->lru
, &pages_to_free
);
1573 * To save our caller's stack, now use input list for pages to free.
1575 list_splice(&pages_to_free
, page_list
);
1579 * If a kernel thread (such as nfsd for loop-back mounts) services
1580 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1581 * In that case we should only throttle if the backing device it is
1582 * writing to is congested. In other cases it is safe to throttle.
1584 static int current_may_throttle(void)
1586 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1587 current
->backing_dev_info
== NULL
||
1588 bdi_write_congested(current
->backing_dev_info
);
1592 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1593 * of reclaimed pages
1595 static noinline_for_stack
unsigned long
1596 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1597 struct scan_control
*sc
, enum lru_list lru
)
1599 LIST_HEAD(page_list
);
1600 unsigned long nr_scanned
;
1601 unsigned long nr_reclaimed
= 0;
1602 unsigned long nr_taken
;
1603 unsigned long nr_dirty
= 0;
1604 unsigned long nr_congested
= 0;
1605 unsigned long nr_unqueued_dirty
= 0;
1606 unsigned long nr_writeback
= 0;
1607 unsigned long nr_immediate
= 0;
1608 isolate_mode_t isolate_mode
= 0;
1609 int file
= is_file_lru(lru
);
1610 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1611 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1613 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1614 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1616 /* We are about to die and free our memory. Return now. */
1617 if (fatal_signal_pending(current
))
1618 return SWAP_CLUSTER_MAX
;
1624 isolate_mode
|= ISOLATE_UNMAPPED
;
1625 if (!sc
->may_writepage
)
1626 isolate_mode
|= ISOLATE_CLEAN
;
1628 spin_lock_irq(&pgdat
->lru_lock
);
1630 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1631 &nr_scanned
, sc
, isolate_mode
, lru
);
1633 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1634 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1636 if (global_reclaim(sc
)) {
1637 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1638 if (current_is_kswapd())
1639 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1641 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1643 spin_unlock_irq(&pgdat
->lru_lock
);
1648 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, TTU_UNMAP
,
1649 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1650 &nr_writeback
, &nr_immediate
,
1653 spin_lock_irq(&pgdat
->lru_lock
);
1655 if (global_reclaim(sc
)) {
1656 if (current_is_kswapd())
1657 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1659 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1662 putback_inactive_pages(lruvec
, &page_list
);
1664 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1666 spin_unlock_irq(&pgdat
->lru_lock
);
1668 mem_cgroup_uncharge_list(&page_list
);
1669 free_hot_cold_page_list(&page_list
, true);
1672 * If reclaim is isolating dirty pages under writeback, it implies
1673 * that the long-lived page allocation rate is exceeding the page
1674 * laundering rate. Either the global limits are not being effective
1675 * at throttling processes due to the page distribution throughout
1676 * zones or there is heavy usage of a slow backing device. The
1677 * only option is to throttle from reclaim context which is not ideal
1678 * as there is no guarantee the dirtying process is throttled in the
1679 * same way balance_dirty_pages() manages.
1681 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1682 * of pages under pages flagged for immediate reclaim and stall if any
1683 * are encountered in the nr_immediate check below.
1685 if (nr_writeback
&& nr_writeback
== nr_taken
)
1686 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1689 * Legacy memcg will stall in page writeback so avoid forcibly
1692 if (sane_reclaim(sc
)) {
1694 * Tag a zone as congested if all the dirty pages scanned were
1695 * backed by a congested BDI and wait_iff_congested will stall.
1697 if (nr_dirty
&& nr_dirty
== nr_congested
)
1698 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1701 * If dirty pages are scanned that are not queued for IO, it
1702 * implies that flushers are not keeping up. In this case, flag
1703 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1706 if (nr_unqueued_dirty
== nr_taken
)
1707 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1710 * If kswapd scans pages marked marked for immediate
1711 * reclaim and under writeback (nr_immediate), it implies
1712 * that pages are cycling through the LRU faster than
1713 * they are written so also forcibly stall.
1715 if (nr_immediate
&& current_may_throttle())
1716 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1720 * Stall direct reclaim for IO completions if underlying BDIs or zone
1721 * is congested. Allow kswapd to continue until it starts encountering
1722 * unqueued dirty pages or cycling through the LRU too quickly.
1724 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1725 current_may_throttle())
1726 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1728 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1729 nr_scanned
, nr_reclaimed
,
1730 sc
->priority
, file
);
1731 return nr_reclaimed
;
1735 * This moves pages from the active list to the inactive list.
1737 * We move them the other way if the page is referenced by one or more
1738 * processes, from rmap.
1740 * If the pages are mostly unmapped, the processing is fast and it is
1741 * appropriate to hold zone_lru_lock across the whole operation. But if
1742 * the pages are mapped, the processing is slow (page_referenced()) so we
1743 * should drop zone_lru_lock around each page. It's impossible to balance
1744 * this, so instead we remove the pages from the LRU while processing them.
1745 * It is safe to rely on PG_active against the non-LRU pages in here because
1746 * nobody will play with that bit on a non-LRU page.
1748 * The downside is that we have to touch page->_refcount against each page.
1749 * But we had to alter page->flags anyway.
1752 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1753 struct list_head
*list
,
1754 struct list_head
*pages_to_free
,
1757 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1758 unsigned long pgmoved
= 0;
1762 while (!list_empty(list
)) {
1763 page
= lru_to_page(list
);
1764 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1766 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1769 nr_pages
= hpage_nr_pages(page
);
1770 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1771 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1772 pgmoved
+= nr_pages
;
1774 if (put_page_testzero(page
)) {
1775 __ClearPageLRU(page
);
1776 __ClearPageActive(page
);
1777 del_page_from_lru_list(page
, lruvec
, lru
);
1779 if (unlikely(PageCompound(page
))) {
1780 spin_unlock_irq(&pgdat
->lru_lock
);
1781 mem_cgroup_uncharge(page
);
1782 (*get_compound_page_dtor(page
))(page
);
1783 spin_lock_irq(&pgdat
->lru_lock
);
1785 list_add(&page
->lru
, pages_to_free
);
1789 if (!is_active_lru(lru
))
1790 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1793 static void shrink_active_list(unsigned long nr_to_scan
,
1794 struct lruvec
*lruvec
,
1795 struct scan_control
*sc
,
1798 unsigned long nr_taken
;
1799 unsigned long nr_scanned
;
1800 unsigned long vm_flags
;
1801 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1802 LIST_HEAD(l_active
);
1803 LIST_HEAD(l_inactive
);
1805 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1806 unsigned long nr_rotated
= 0;
1807 isolate_mode_t isolate_mode
= 0;
1808 int file
= is_file_lru(lru
);
1809 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1814 isolate_mode
|= ISOLATE_UNMAPPED
;
1815 if (!sc
->may_writepage
)
1816 isolate_mode
|= ISOLATE_CLEAN
;
1818 spin_lock_irq(&pgdat
->lru_lock
);
1820 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1821 &nr_scanned
, sc
, isolate_mode
, lru
);
1823 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1824 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1826 if (global_reclaim(sc
))
1827 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1828 __count_vm_events(PGREFILL
, nr_scanned
);
1830 spin_unlock_irq(&pgdat
->lru_lock
);
1832 while (!list_empty(&l_hold
)) {
1834 page
= lru_to_page(&l_hold
);
1835 list_del(&page
->lru
);
1837 if (unlikely(!page_evictable(page
))) {
1838 putback_lru_page(page
);
1842 if (unlikely(buffer_heads_over_limit
)) {
1843 if (page_has_private(page
) && trylock_page(page
)) {
1844 if (page_has_private(page
))
1845 try_to_release_page(page
, 0);
1850 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1852 nr_rotated
+= hpage_nr_pages(page
);
1854 * Identify referenced, file-backed active pages and
1855 * give them one more trip around the active list. So
1856 * that executable code get better chances to stay in
1857 * memory under moderate memory pressure. Anon pages
1858 * are not likely to be evicted by use-once streaming
1859 * IO, plus JVM can create lots of anon VM_EXEC pages,
1860 * so we ignore them here.
1862 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1863 list_add(&page
->lru
, &l_active
);
1868 ClearPageActive(page
); /* we are de-activating */
1869 list_add(&page
->lru
, &l_inactive
);
1873 * Move pages back to the lru list.
1875 spin_lock_irq(&pgdat
->lru_lock
);
1877 * Count referenced pages from currently used mappings as rotated,
1878 * even though only some of them are actually re-activated. This
1879 * helps balance scan pressure between file and anonymous pages in
1882 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1884 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1885 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1886 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1887 spin_unlock_irq(&pgdat
->lru_lock
);
1889 mem_cgroup_uncharge_list(&l_hold
);
1890 free_hot_cold_page_list(&l_hold
, true);
1894 * The inactive anon list should be small enough that the VM never has
1895 * to do too much work.
1897 * The inactive file list should be small enough to leave most memory
1898 * to the established workingset on the scan-resistant active list,
1899 * but large enough to avoid thrashing the aggregate readahead window.
1901 * Both inactive lists should also be large enough that each inactive
1902 * page has a chance to be referenced again before it is reclaimed.
1904 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1905 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1906 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1909 * memory ratio inactive
1910 * -------------------------------------
1919 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
)
1921 unsigned long inactive_ratio
;
1922 unsigned long inactive
;
1923 unsigned long active
;
1927 * If we don't have swap space, anonymous page deactivation
1930 if (!file
&& !total_swap_pages
)
1933 inactive
= lruvec_lru_size(lruvec
, file
* LRU_FILE
);
1934 active
= lruvec_lru_size(lruvec
, file
* LRU_FILE
+ LRU_ACTIVE
);
1936 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1938 inactive_ratio
= int_sqrt(10 * gb
);
1942 return inactive
* inactive_ratio
< active
;
1945 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1946 struct lruvec
*lruvec
, struct scan_control
*sc
)
1948 if (is_active_lru(lru
)) {
1949 if (inactive_list_is_low(lruvec
, is_file_lru(lru
)))
1950 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1954 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1965 * Determine how aggressively the anon and file LRU lists should be
1966 * scanned. The relative value of each set of LRU lists is determined
1967 * by looking at the fraction of the pages scanned we did rotate back
1968 * onto the active list instead of evict.
1970 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1971 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1973 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
1974 struct scan_control
*sc
, unsigned long *nr
,
1975 unsigned long *lru_pages
)
1977 int swappiness
= mem_cgroup_swappiness(memcg
);
1978 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1980 u64 denominator
= 0; /* gcc */
1981 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1982 unsigned long anon_prio
, file_prio
;
1983 enum scan_balance scan_balance
;
1984 unsigned long anon
, file
;
1985 bool force_scan
= false;
1986 unsigned long ap
, fp
;
1992 * If the zone or memcg is small, nr[l] can be 0. This
1993 * results in no scanning on this priority and a potential
1994 * priority drop. Global direct reclaim can go to the next
1995 * zone and tends to have no problems. Global kswapd is for
1996 * zone balancing and it needs to scan a minimum amount. When
1997 * reclaiming for a memcg, a priority drop can cause high
1998 * latencies, so it's better to scan a minimum amount there as
2001 if (current_is_kswapd()) {
2002 if (!pgdat_reclaimable(pgdat
))
2004 if (!mem_cgroup_online(memcg
))
2007 if (!global_reclaim(sc
))
2010 /* If we have no swap space, do not bother scanning anon pages. */
2011 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2012 scan_balance
= SCAN_FILE
;
2017 * Global reclaim will swap to prevent OOM even with no
2018 * swappiness, but memcg users want to use this knob to
2019 * disable swapping for individual groups completely when
2020 * using the memory controller's swap limit feature would be
2023 if (!global_reclaim(sc
) && !swappiness
) {
2024 scan_balance
= SCAN_FILE
;
2029 * Do not apply any pressure balancing cleverness when the
2030 * system is close to OOM, scan both anon and file equally
2031 * (unless the swappiness setting disagrees with swapping).
2033 if (!sc
->priority
&& swappiness
) {
2034 scan_balance
= SCAN_EQUAL
;
2039 * Prevent the reclaimer from falling into the cache trap: as
2040 * cache pages start out inactive, every cache fault will tip
2041 * the scan balance towards the file LRU. And as the file LRU
2042 * shrinks, so does the window for rotation from references.
2043 * This means we have a runaway feedback loop where a tiny
2044 * thrashing file LRU becomes infinitely more attractive than
2045 * anon pages. Try to detect this based on file LRU size.
2047 if (global_reclaim(sc
)) {
2048 unsigned long pgdatfile
;
2049 unsigned long pgdatfree
;
2051 unsigned long total_high_wmark
= 0;
2053 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2054 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2055 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2057 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2058 struct zone
*zone
= &pgdat
->node_zones
[z
];
2059 if (!populated_zone(zone
))
2062 total_high_wmark
+= high_wmark_pages(zone
);
2065 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2066 scan_balance
= SCAN_ANON
;
2072 * If there is enough inactive page cache, i.e. if the size of the
2073 * inactive list is greater than that of the active list *and* the
2074 * inactive list actually has some pages to scan on this priority, we
2075 * do not reclaim anything from the anonymous working set right now.
2076 * Without the second condition we could end up never scanning an
2077 * lruvec even if it has plenty of old anonymous pages unless the
2078 * system is under heavy pressure.
2080 if (!inactive_list_is_low(lruvec
, true) &&
2081 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2082 scan_balance
= SCAN_FILE
;
2086 scan_balance
= SCAN_FRACT
;
2089 * With swappiness at 100, anonymous and file have the same priority.
2090 * This scanning priority is essentially the inverse of IO cost.
2092 anon_prio
= swappiness
;
2093 file_prio
= 200 - anon_prio
;
2096 * OK, so we have swap space and a fair amount of page cache
2097 * pages. We use the recently rotated / recently scanned
2098 * ratios to determine how valuable each cache is.
2100 * Because workloads change over time (and to avoid overflow)
2101 * we keep these statistics as a floating average, which ends
2102 * up weighing recent references more than old ones.
2104 * anon in [0], file in [1]
2107 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2108 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2109 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2110 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2112 spin_lock_irq(&pgdat
->lru_lock
);
2113 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2114 reclaim_stat
->recent_scanned
[0] /= 2;
2115 reclaim_stat
->recent_rotated
[0] /= 2;
2118 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2119 reclaim_stat
->recent_scanned
[1] /= 2;
2120 reclaim_stat
->recent_rotated
[1] /= 2;
2124 * The amount of pressure on anon vs file pages is inversely
2125 * proportional to the fraction of recently scanned pages on
2126 * each list that were recently referenced and in active use.
2128 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2129 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2131 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2132 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2133 spin_unlock_irq(&pgdat
->lru_lock
);
2137 denominator
= ap
+ fp
+ 1;
2139 some_scanned
= false;
2140 /* Only use force_scan on second pass. */
2141 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2143 for_each_evictable_lru(lru
) {
2144 int file
= is_file_lru(lru
);
2148 size
= lruvec_lru_size(lruvec
, lru
);
2149 scan
= size
>> sc
->priority
;
2151 if (!scan
&& pass
&& force_scan
)
2152 scan
= min(size
, SWAP_CLUSTER_MAX
);
2154 switch (scan_balance
) {
2156 /* Scan lists relative to size */
2160 * Scan types proportional to swappiness and
2161 * their relative recent reclaim efficiency.
2163 scan
= div64_u64(scan
* fraction
[file
],
2168 /* Scan one type exclusively */
2169 if ((scan_balance
== SCAN_FILE
) != file
) {
2175 /* Look ma, no brain */
2183 * Skip the second pass and don't force_scan,
2184 * if we found something to scan.
2186 some_scanned
|= !!scan
;
2191 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2192 static void init_tlb_ubc(void)
2195 * This deliberately does not clear the cpumask as it's expensive
2196 * and unnecessary. If there happens to be data in there then the
2197 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2198 * then will be cleared.
2200 current
->tlb_ubc
.flush_required
= false;
2203 static inline void init_tlb_ubc(void)
2206 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2209 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2211 static void shrink_zone_memcg(struct zone
*zone
, struct mem_cgroup
*memcg
,
2212 struct scan_control
*sc
, unsigned long *lru_pages
)
2214 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2215 unsigned long nr
[NR_LRU_LISTS
];
2216 unsigned long targets
[NR_LRU_LISTS
];
2217 unsigned long nr_to_scan
;
2219 unsigned long nr_reclaimed
= 0;
2220 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2221 struct blk_plug plug
;
2224 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2226 /* Record the original scan target for proportional adjustments later */
2227 memcpy(targets
, nr
, sizeof(nr
));
2230 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2231 * event that can occur when there is little memory pressure e.g.
2232 * multiple streaming readers/writers. Hence, we do not abort scanning
2233 * when the requested number of pages are reclaimed when scanning at
2234 * DEF_PRIORITY on the assumption that the fact we are direct
2235 * reclaiming implies that kswapd is not keeping up and it is best to
2236 * do a batch of work at once. For memcg reclaim one check is made to
2237 * abort proportional reclaim if either the file or anon lru has already
2238 * dropped to zero at the first pass.
2240 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2241 sc
->priority
== DEF_PRIORITY
);
2245 blk_start_plug(&plug
);
2246 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2247 nr
[LRU_INACTIVE_FILE
]) {
2248 unsigned long nr_anon
, nr_file
, percentage
;
2249 unsigned long nr_scanned
;
2251 for_each_evictable_lru(lru
) {
2253 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2254 nr
[lru
] -= nr_to_scan
;
2256 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2261 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2265 * For kswapd and memcg, reclaim at least the number of pages
2266 * requested. Ensure that the anon and file LRUs are scanned
2267 * proportionally what was requested by get_scan_count(). We
2268 * stop reclaiming one LRU and reduce the amount scanning
2269 * proportional to the original scan target.
2271 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2272 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2275 * It's just vindictive to attack the larger once the smaller
2276 * has gone to zero. And given the way we stop scanning the
2277 * smaller below, this makes sure that we only make one nudge
2278 * towards proportionality once we've got nr_to_reclaim.
2280 if (!nr_file
|| !nr_anon
)
2283 if (nr_file
> nr_anon
) {
2284 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2285 targets
[LRU_ACTIVE_ANON
] + 1;
2287 percentage
= nr_anon
* 100 / scan_target
;
2289 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2290 targets
[LRU_ACTIVE_FILE
] + 1;
2292 percentage
= nr_file
* 100 / scan_target
;
2295 /* Stop scanning the smaller of the LRU */
2297 nr
[lru
+ LRU_ACTIVE
] = 0;
2300 * Recalculate the other LRU scan count based on its original
2301 * scan target and the percentage scanning already complete
2303 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2304 nr_scanned
= targets
[lru
] - nr
[lru
];
2305 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2306 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2309 nr_scanned
= targets
[lru
] - nr
[lru
];
2310 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2311 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2313 scan_adjusted
= true;
2315 blk_finish_plug(&plug
);
2316 sc
->nr_reclaimed
+= nr_reclaimed
;
2319 * Even if we did not try to evict anon pages at all, we want to
2320 * rebalance the anon lru active/inactive ratio.
2322 if (inactive_list_is_low(lruvec
, false))
2323 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2324 sc
, LRU_ACTIVE_ANON
);
2326 throttle_vm_writeout(sc
->gfp_mask
);
2329 /* Use reclaim/compaction for costly allocs or under memory pressure */
2330 static bool in_reclaim_compaction(struct scan_control
*sc
)
2332 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2333 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2334 sc
->priority
< DEF_PRIORITY
- 2))
2341 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2342 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2343 * true if more pages should be reclaimed such that when the page allocator
2344 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2345 * It will give up earlier than that if there is difficulty reclaiming pages.
2347 static inline bool should_continue_reclaim(struct zone
*zone
,
2348 unsigned long nr_reclaimed
,
2349 unsigned long nr_scanned
,
2350 struct scan_control
*sc
)
2352 unsigned long pages_for_compaction
;
2353 unsigned long inactive_lru_pages
;
2355 /* If not in reclaim/compaction mode, stop */
2356 if (!in_reclaim_compaction(sc
))
2359 /* Consider stopping depending on scan and reclaim activity */
2360 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2362 * For __GFP_REPEAT allocations, stop reclaiming if the
2363 * full LRU list has been scanned and we are still failing
2364 * to reclaim pages. This full LRU scan is potentially
2365 * expensive but a __GFP_REPEAT caller really wants to succeed
2367 if (!nr_reclaimed
&& !nr_scanned
)
2371 * For non-__GFP_REPEAT allocations which can presumably
2372 * fail without consequence, stop if we failed to reclaim
2373 * any pages from the last SWAP_CLUSTER_MAX number of
2374 * pages that were scanned. This will return to the
2375 * caller faster at the risk reclaim/compaction and
2376 * the resulting allocation attempt fails
2383 * If we have not reclaimed enough pages for compaction and the
2384 * inactive lists are large enough, continue reclaiming
2386 pages_for_compaction
= (2UL << sc
->order
);
2387 inactive_lru_pages
= node_page_state(zone
->zone_pgdat
, NR_INACTIVE_FILE
);
2388 if (get_nr_swap_pages() > 0)
2389 inactive_lru_pages
+= node_page_state(zone
->zone_pgdat
, NR_INACTIVE_ANON
);
2390 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2391 inactive_lru_pages
> pages_for_compaction
)
2394 /* If compaction would go ahead or the allocation would succeed, stop */
2395 switch (compaction_suitable(zone
, sc
->order
, 0, 0)) {
2396 case COMPACT_PARTIAL
:
2397 case COMPACT_CONTINUE
:
2404 static bool shrink_zone(struct zone
*zone
, struct scan_control
*sc
,
2407 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2408 unsigned long nr_reclaimed
, nr_scanned
;
2409 bool reclaimable
= false;
2412 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2413 struct mem_cgroup_reclaim_cookie reclaim
= {
2415 .priority
= sc
->priority
,
2417 unsigned long zone_lru_pages
= 0;
2418 struct mem_cgroup
*memcg
;
2420 nr_reclaimed
= sc
->nr_reclaimed
;
2421 nr_scanned
= sc
->nr_scanned
;
2423 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2425 unsigned long lru_pages
;
2426 unsigned long reclaimed
;
2427 unsigned long scanned
;
2429 if (mem_cgroup_low(root
, memcg
)) {
2430 if (!sc
->may_thrash
)
2432 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2435 reclaimed
= sc
->nr_reclaimed
;
2436 scanned
= sc
->nr_scanned
;
2438 shrink_zone_memcg(zone
, memcg
, sc
, &lru_pages
);
2439 zone_lru_pages
+= lru_pages
;
2441 if (memcg
&& is_classzone
)
2442 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
),
2443 memcg
, sc
->nr_scanned
- scanned
,
2446 /* Record the group's reclaim efficiency */
2447 vmpressure(sc
->gfp_mask
, memcg
, false,
2448 sc
->nr_scanned
- scanned
,
2449 sc
->nr_reclaimed
- reclaimed
);
2452 * Direct reclaim and kswapd have to scan all memory
2453 * cgroups to fulfill the overall scan target for the
2456 * Limit reclaim, on the other hand, only cares about
2457 * nr_to_reclaim pages to be reclaimed and it will
2458 * retry with decreasing priority if one round over the
2459 * whole hierarchy is not sufficient.
2461 if (!global_reclaim(sc
) &&
2462 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2463 mem_cgroup_iter_break(root
, memcg
);
2466 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2469 * Shrink the slab caches in the same proportion that
2470 * the eligible LRU pages were scanned.
2472 if (global_reclaim(sc
) && is_classzone
)
2473 shrink_slab(sc
->gfp_mask
, zone_to_nid(zone
), NULL
,
2474 sc
->nr_scanned
- nr_scanned
,
2477 if (reclaim_state
) {
2478 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2479 reclaim_state
->reclaimed_slab
= 0;
2482 /* Record the subtree's reclaim efficiency */
2483 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2484 sc
->nr_scanned
- nr_scanned
,
2485 sc
->nr_reclaimed
- nr_reclaimed
);
2487 if (sc
->nr_reclaimed
- nr_reclaimed
)
2490 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2491 sc
->nr_scanned
- nr_scanned
, sc
));
2497 * Returns true if compaction should go ahead for a high-order request, or
2498 * the high-order allocation would succeed without compaction.
2500 static inline bool compaction_ready(struct zone
*zone
, int order
, int classzone_idx
)
2502 unsigned long balance_gap
, watermark
;
2506 * Compaction takes time to run and there are potentially other
2507 * callers using the pages just freed. Continue reclaiming until
2508 * there is a buffer of free pages available to give compaction
2509 * a reasonable chance of completing and allocating the page
2511 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
2512 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
2513 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << order
);
2514 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, classzone_idx
);
2517 * If compaction is deferred, reclaim up to a point where
2518 * compaction will have a chance of success when re-enabled
2520 if (compaction_deferred(zone
, order
))
2521 return watermark_ok
;
2524 * If compaction is not ready to start and allocation is not likely
2525 * to succeed without it, then keep reclaiming.
2527 if (compaction_suitable(zone
, order
, 0, classzone_idx
) == COMPACT_SKIPPED
)
2530 return watermark_ok
;
2534 * This is the direct reclaim path, for page-allocating processes. We only
2535 * try to reclaim pages from zones which will satisfy the caller's allocation
2538 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2540 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2542 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2543 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2544 * zone defense algorithm.
2546 * If a zone is deemed to be full of pinned pages then just give it a light
2547 * scan then give up on it.
2549 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2553 unsigned long nr_soft_reclaimed
;
2554 unsigned long nr_soft_scanned
;
2556 enum zone_type requested_highidx
= gfp_zone(sc
->gfp_mask
);
2559 * If the number of buffer_heads in the machine exceeds the maximum
2560 * allowed level, force direct reclaim to scan the highmem zone as
2561 * highmem pages could be pinning lowmem pages storing buffer_heads
2563 orig_mask
= sc
->gfp_mask
;
2564 if (buffer_heads_over_limit
)
2565 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2567 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2568 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2569 enum zone_type classzone_idx
;
2571 if (!populated_zone(zone
))
2574 classzone_idx
= requested_highidx
;
2575 while (!populated_zone(zone
->zone_pgdat
->node_zones
+
2580 * Take care memory controller reclaiming has small influence
2583 if (global_reclaim(sc
)) {
2584 if (!cpuset_zone_allowed(zone
,
2585 GFP_KERNEL
| __GFP_HARDWALL
))
2588 if (sc
->priority
!= DEF_PRIORITY
&&
2589 !pgdat_reclaimable(zone
->zone_pgdat
))
2590 continue; /* Let kswapd poll it */
2593 * If we already have plenty of memory free for
2594 * compaction in this zone, don't free any more.
2595 * Even though compaction is invoked for any
2596 * non-zero order, only frequent costly order
2597 * reclamation is disruptive enough to become a
2598 * noticeable problem, like transparent huge
2601 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2602 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2603 zonelist_zone_idx(z
) <= requested_highidx
&&
2604 compaction_ready(zone
, sc
->order
, requested_highidx
)) {
2605 sc
->compaction_ready
= true;
2610 * This steals pages from memory cgroups over softlimit
2611 * and returns the number of reclaimed pages and
2612 * scanned pages. This works for global memory pressure
2613 * and balancing, not for a memcg's limit.
2615 nr_soft_scanned
= 0;
2616 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2617 sc
->order
, sc
->gfp_mask
,
2619 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2620 sc
->nr_scanned
+= nr_soft_scanned
;
2621 /* need some check for avoid more shrink_zone() */
2624 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
2628 * Restore to original mask to avoid the impact on the caller if we
2629 * promoted it to __GFP_HIGHMEM.
2631 sc
->gfp_mask
= orig_mask
;
2635 * This is the main entry point to direct page reclaim.
2637 * If a full scan of the inactive list fails to free enough memory then we
2638 * are "out of memory" and something needs to be killed.
2640 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2641 * high - the zone may be full of dirty or under-writeback pages, which this
2642 * caller can't do much about. We kick the writeback threads and take explicit
2643 * naps in the hope that some of these pages can be written. But if the
2644 * allocating task holds filesystem locks which prevent writeout this might not
2645 * work, and the allocation attempt will fail.
2647 * returns: 0, if no pages reclaimed
2648 * else, the number of pages reclaimed
2650 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2651 struct scan_control
*sc
)
2653 int initial_priority
= sc
->priority
;
2654 unsigned long total_scanned
= 0;
2655 unsigned long writeback_threshold
;
2657 delayacct_freepages_start();
2659 if (global_reclaim(sc
))
2660 count_vm_event(ALLOCSTALL
);
2663 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2666 shrink_zones(zonelist
, sc
);
2668 total_scanned
+= sc
->nr_scanned
;
2669 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2672 if (sc
->compaction_ready
)
2676 * If we're getting trouble reclaiming, start doing
2677 * writepage even in laptop mode.
2679 if (sc
->priority
< DEF_PRIORITY
- 2)
2680 sc
->may_writepage
= 1;
2683 * Try to write back as many pages as we just scanned. This
2684 * tends to cause slow streaming writers to write data to the
2685 * disk smoothly, at the dirtying rate, which is nice. But
2686 * that's undesirable in laptop mode, where we *want* lumpy
2687 * writeout. So in laptop mode, write out the whole world.
2689 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2690 if (total_scanned
> writeback_threshold
) {
2691 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2692 WB_REASON_TRY_TO_FREE_PAGES
);
2693 sc
->may_writepage
= 1;
2695 } while (--sc
->priority
>= 0);
2697 delayacct_freepages_end();
2699 if (sc
->nr_reclaimed
)
2700 return sc
->nr_reclaimed
;
2702 /* Aborted reclaim to try compaction? don't OOM, then */
2703 if (sc
->compaction_ready
)
2706 /* Untapped cgroup reserves? Don't OOM, retry. */
2707 if (!sc
->may_thrash
) {
2708 sc
->priority
= initial_priority
;
2716 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2719 unsigned long pfmemalloc_reserve
= 0;
2720 unsigned long free_pages
= 0;
2724 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2725 zone
= &pgdat
->node_zones
[i
];
2726 if (!populated_zone(zone
) ||
2727 pgdat_reclaimable_pages(pgdat
) == 0)
2730 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2731 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2734 /* If there are no reserves (unexpected config) then do not throttle */
2735 if (!pfmemalloc_reserve
)
2738 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2740 /* kswapd must be awake if processes are being throttled */
2741 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2742 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2743 (enum zone_type
)ZONE_NORMAL
);
2744 wake_up_interruptible(&pgdat
->kswapd_wait
);
2751 * Throttle direct reclaimers if backing storage is backed by the network
2752 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2753 * depleted. kswapd will continue to make progress and wake the processes
2754 * when the low watermark is reached.
2756 * Returns true if a fatal signal was delivered during throttling. If this
2757 * happens, the page allocator should not consider triggering the OOM killer.
2759 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2760 nodemask_t
*nodemask
)
2764 pg_data_t
*pgdat
= NULL
;
2767 * Kernel threads should not be throttled as they may be indirectly
2768 * responsible for cleaning pages necessary for reclaim to make forward
2769 * progress. kjournald for example may enter direct reclaim while
2770 * committing a transaction where throttling it could forcing other
2771 * processes to block on log_wait_commit().
2773 if (current
->flags
& PF_KTHREAD
)
2777 * If a fatal signal is pending, this process should not throttle.
2778 * It should return quickly so it can exit and free its memory
2780 if (fatal_signal_pending(current
))
2784 * Check if the pfmemalloc reserves are ok by finding the first node
2785 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2786 * GFP_KERNEL will be required for allocating network buffers when
2787 * swapping over the network so ZONE_HIGHMEM is unusable.
2789 * Throttling is based on the first usable node and throttled processes
2790 * wait on a queue until kswapd makes progress and wakes them. There
2791 * is an affinity then between processes waking up and where reclaim
2792 * progress has been made assuming the process wakes on the same node.
2793 * More importantly, processes running on remote nodes will not compete
2794 * for remote pfmemalloc reserves and processes on different nodes
2795 * should make reasonable progress.
2797 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2798 gfp_zone(gfp_mask
), nodemask
) {
2799 if (zone_idx(zone
) > ZONE_NORMAL
)
2802 /* Throttle based on the first usable node */
2803 pgdat
= zone
->zone_pgdat
;
2804 if (pfmemalloc_watermark_ok(pgdat
))
2809 /* If no zone was usable by the allocation flags then do not throttle */
2813 /* Account for the throttling */
2814 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2817 * If the caller cannot enter the filesystem, it's possible that it
2818 * is due to the caller holding an FS lock or performing a journal
2819 * transaction in the case of a filesystem like ext[3|4]. In this case,
2820 * it is not safe to block on pfmemalloc_wait as kswapd could be
2821 * blocked waiting on the same lock. Instead, throttle for up to a
2822 * second before continuing.
2824 if (!(gfp_mask
& __GFP_FS
)) {
2825 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2826 pfmemalloc_watermark_ok(pgdat
), HZ
);
2831 /* Throttle until kswapd wakes the process */
2832 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2833 pfmemalloc_watermark_ok(pgdat
));
2836 if (fatal_signal_pending(current
))
2843 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2844 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2846 unsigned long nr_reclaimed
;
2847 struct scan_control sc
= {
2848 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2849 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2851 .nodemask
= nodemask
,
2852 .priority
= DEF_PRIORITY
,
2853 .may_writepage
= !laptop_mode
,
2859 * Do not enter reclaim if fatal signal was delivered while throttled.
2860 * 1 is returned so that the page allocator does not OOM kill at this
2863 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2866 trace_mm_vmscan_direct_reclaim_begin(order
,
2870 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2872 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2874 return nr_reclaimed
;
2879 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2880 gfp_t gfp_mask
, bool noswap
,
2882 unsigned long *nr_scanned
)
2884 struct scan_control sc
= {
2885 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2886 .target_mem_cgroup
= memcg
,
2887 .may_writepage
= !laptop_mode
,
2889 .may_swap
= !noswap
,
2891 unsigned long lru_pages
;
2893 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2894 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2896 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2901 * NOTE: Although we can get the priority field, using it
2902 * here is not a good idea, since it limits the pages we can scan.
2903 * if we don't reclaim here, the shrink_zone from balance_pgdat
2904 * will pick up pages from other mem cgroup's as well. We hack
2905 * the priority and make it zero.
2907 shrink_zone_memcg(zone
, memcg
, &sc
, &lru_pages
);
2909 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2911 *nr_scanned
= sc
.nr_scanned
;
2912 return sc
.nr_reclaimed
;
2915 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2916 unsigned long nr_pages
,
2920 struct zonelist
*zonelist
;
2921 unsigned long nr_reclaimed
;
2923 struct scan_control sc
= {
2924 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2925 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2926 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2927 .target_mem_cgroup
= memcg
,
2928 .priority
= DEF_PRIORITY
,
2929 .may_writepage
= !laptop_mode
,
2931 .may_swap
= may_swap
,
2935 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2936 * take care of from where we get pages. So the node where we start the
2937 * scan does not need to be the current node.
2939 nid
= mem_cgroup_select_victim_node(memcg
);
2941 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2943 trace_mm_vmscan_memcg_reclaim_begin(0,
2947 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2949 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2951 return nr_reclaimed
;
2955 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2957 struct mem_cgroup
*memcg
;
2959 if (!total_swap_pages
)
2962 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2964 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2966 if (inactive_list_is_low(lruvec
, false))
2967 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2968 sc
, LRU_ACTIVE_ANON
);
2970 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2974 static bool zone_balanced(struct zone
*zone
, int order
, bool highorder
,
2975 unsigned long balance_gap
, int classzone_idx
)
2977 unsigned long mark
= high_wmark_pages(zone
) + balance_gap
;
2980 * When checking from pgdat_balanced(), kswapd should stop and sleep
2981 * when it reaches the high order-0 watermark and let kcompactd take
2982 * over. Other callers such as wakeup_kswapd() want to determine the
2983 * true high-order watermark.
2985 if (IS_ENABLED(CONFIG_COMPACTION
) && !highorder
) {
2986 mark
+= (1UL << order
);
2990 return zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
);
2994 * pgdat_balanced() is used when checking if a node is balanced.
2996 * For order-0, all zones must be balanced!
2998 * For high-order allocations only zones that meet watermarks and are in a
2999 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3000 * total of balanced pages must be at least 25% of the zones allowed by
3001 * classzone_idx for the node to be considered balanced. Forcing all zones to
3002 * be balanced for high orders can cause excessive reclaim when there are
3004 * The choice of 25% is due to
3005 * o a 16M DMA zone that is balanced will not balance a zone on any
3006 * reasonable sized machine
3007 * o On all other machines, the top zone must be at least a reasonable
3008 * percentage of the middle zones. For example, on 32-bit x86, highmem
3009 * would need to be at least 256M for it to be balance a whole node.
3010 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3011 * to balance a node on its own. These seemed like reasonable ratios.
3013 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3015 unsigned long managed_pages
= 0;
3016 unsigned long balanced_pages
= 0;
3019 /* Check the watermark levels */
3020 for (i
= 0; i
<= classzone_idx
; i
++) {
3021 struct zone
*zone
= pgdat
->node_zones
+ i
;
3023 if (!populated_zone(zone
))
3026 managed_pages
+= zone
->managed_pages
;
3029 * A special case here:
3031 * balance_pgdat() skips over all_unreclaimable after
3032 * DEF_PRIORITY. Effectively, it considers them balanced so
3033 * they must be considered balanced here as well!
3035 if (!pgdat_reclaimable(zone
->zone_pgdat
)) {
3036 balanced_pages
+= zone
->managed_pages
;
3040 if (zone_balanced(zone
, order
, false, 0, i
))
3041 balanced_pages
+= zone
->managed_pages
;
3047 return balanced_pages
>= (managed_pages
>> 2);
3053 * Prepare kswapd for sleeping. This verifies that there are no processes
3054 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3056 * Returns true if kswapd is ready to sleep
3058 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
3061 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3066 * The throttled processes are normally woken up in balance_pgdat() as
3067 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3068 * race between when kswapd checks the watermarks and a process gets
3069 * throttled. There is also a potential race if processes get
3070 * throttled, kswapd wakes, a large process exits thereby balancing the
3071 * zones, which causes kswapd to exit balance_pgdat() before reaching
3072 * the wake up checks. If kswapd is going to sleep, no process should
3073 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3074 * the wake up is premature, processes will wake kswapd and get
3075 * throttled again. The difference from wake ups in balance_pgdat() is
3076 * that here we are under prepare_to_wait().
3078 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3079 wake_up_all(&pgdat
->pfmemalloc_wait
);
3081 return pgdat_balanced(pgdat
, order
, classzone_idx
);
3085 * kswapd shrinks the zone by the number of pages required to reach
3086 * the high watermark.
3088 * Returns true if kswapd scanned at least the requested number of pages to
3089 * reclaim or if the lack of progress was due to pages under writeback.
3090 * This is used to determine if the scanning priority needs to be raised.
3092 static bool kswapd_shrink_zone(struct zone
*zone
,
3094 struct scan_control
*sc
)
3096 unsigned long balance_gap
;
3097 bool lowmem_pressure
;
3098 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3100 /* Reclaim above the high watermark. */
3101 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
3104 * We put equal pressure on every zone, unless one zone has way too
3105 * many pages free already. The "too many pages" is defined as the
3106 * high wmark plus a "gap" where the gap is either the low
3107 * watermark or 1% of the zone, whichever is smaller.
3109 balance_gap
= min(low_wmark_pages(zone
), DIV_ROUND_UP(
3110 zone
->managed_pages
, KSWAPD_ZONE_BALANCE_GAP_RATIO
));
3113 * If there is no low memory pressure or the zone is balanced then no
3114 * reclaim is necessary
3116 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
3117 if (!lowmem_pressure
&& zone_balanced(zone
, sc
->order
, false,
3118 balance_gap
, classzone_idx
))
3121 shrink_zone(zone
, sc
, zone_idx(zone
) == classzone_idx
);
3124 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
3127 * If a zone reaches its high watermark, consider it to be no longer
3128 * congested. It's possible there are dirty pages backed by congested
3129 * BDIs but as pressure is relieved, speculatively avoid congestion
3132 if (pgdat_reclaimable(zone
->zone_pgdat
) &&
3133 zone_balanced(zone
, sc
->order
, false, 0, classzone_idx
)) {
3134 clear_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
3135 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
3138 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3142 * For kswapd, balance_pgdat() will work across all this node's zones until
3143 * they are all at high_wmark_pages(zone).
3145 * Returns the highest zone idx kswapd was reclaiming at
3147 * There is special handling here for zones which are full of pinned pages.
3148 * This can happen if the pages are all mlocked, or if they are all used by
3149 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3150 * What we do is to detect the case where all pages in the zone have been
3151 * scanned twice and there has been zero successful reclaim. Mark the zone as
3152 * dead and from now on, only perform a short scan. Basically we're polling
3153 * the zone for when the problem goes away.
3155 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3156 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3157 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3158 * lower zones regardless of the number of free pages in the lower zones. This
3159 * interoperates with the page allocator fallback scheme to ensure that aging
3160 * of pages is balanced across the zones.
3162 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3165 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
3166 unsigned long nr_soft_reclaimed
;
3167 unsigned long nr_soft_scanned
;
3168 struct scan_control sc
= {
3169 .gfp_mask
= GFP_KERNEL
,
3171 .priority
= DEF_PRIORITY
,
3172 .may_writepage
= !laptop_mode
,
3176 count_vm_event(PAGEOUTRUN
);
3179 bool raise_priority
= true;
3181 sc
.nr_reclaimed
= 0;
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3187 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
3188 struct zone
*zone
= pgdat
->node_zones
+ i
;
3190 if (!populated_zone(zone
))
3193 if (sc
.priority
!= DEF_PRIORITY
&&
3194 !pgdat_reclaimable(zone
->zone_pgdat
))
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3201 age_active_anon(zone
, &sc
);
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3209 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
3214 if (!zone_balanced(zone
, order
, false, 0, 0)) {
3219 * If balanced, clear the dirty and congested
3224 clear_bit(PGDAT_CONGESTED
, &zone
->zone_pgdat
->flags
);
3225 clear_bit(PGDAT_DIRTY
, &zone
->zone_pgdat
->flags
);
3233 * If we're getting trouble reclaiming, start doing writepage
3234 * even in laptop mode.
3236 if (sc
.priority
< DEF_PRIORITY
- 2)
3237 sc
.may_writepage
= 1;
3240 * Now scan the zone in the dma->highmem direction, stopping
3241 * at the last zone which needs scanning.
3243 * We do this because the page allocator works in the opposite
3244 * direction. This prevents the page allocator from allocating
3245 * pages behind kswapd's direction of progress, which would
3246 * cause too much scanning of the lower zones.
3248 for (i
= 0; i
<= end_zone
; i
++) {
3249 struct zone
*zone
= pgdat
->node_zones
+ i
;
3251 if (!populated_zone(zone
))
3254 if (sc
.priority
!= DEF_PRIORITY
&&
3255 !pgdat_reclaimable(zone
->zone_pgdat
))
3260 nr_soft_scanned
= 0;
3262 * Call soft limit reclaim before calling shrink_zone.
3264 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3267 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3270 * There should be no need to raise the scanning
3271 * priority if enough pages are already being scanned
3272 * that that high watermark would be met at 100%
3275 if (kswapd_shrink_zone(zone
, end_zone
, &sc
))
3276 raise_priority
= false;
3280 * If the low watermark is met there is no need for processes
3281 * to be throttled on pfmemalloc_wait as they should not be
3282 * able to safely make forward progress. Wake them
3284 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3285 pfmemalloc_watermark_ok(pgdat
))
3286 wake_up_all(&pgdat
->pfmemalloc_wait
);
3288 /* Check if kswapd should be suspending */
3289 if (try_to_freeze() || kthread_should_stop())
3293 * Raise priority if scanning rate is too low or there was no
3294 * progress in reclaiming pages
3296 if (raise_priority
|| !sc
.nr_reclaimed
)
3298 } while (sc
.priority
>= 1 &&
3299 !pgdat_balanced(pgdat
, order
, classzone_idx
));
3303 * Return the highest zone idx we were reclaiming at so
3304 * prepare_kswapd_sleep() makes the same decisions as here.
3309 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
,
3310 int classzone_idx
, int balanced_classzone_idx
)
3315 if (freezing(current
) || kthread_should_stop())
3318 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3320 /* Try to sleep for a short interval */
3321 if (prepare_kswapd_sleep(pgdat
, order
, remaining
,
3322 balanced_classzone_idx
)) {
3324 * Compaction records what page blocks it recently failed to
3325 * isolate pages from and skips them in the future scanning.
3326 * When kswapd is going to sleep, it is reasonable to assume
3327 * that pages and compaction may succeed so reset the cache.
3329 reset_isolation_suitable(pgdat
);
3332 * We have freed the memory, now we should compact it to make
3333 * allocation of the requested order possible.
3335 wakeup_kcompactd(pgdat
, order
, classzone_idx
);
3337 remaining
= schedule_timeout(HZ
/10);
3338 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3339 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3343 * After a short sleep, check if it was a premature sleep. If not, then
3344 * go fully to sleep until explicitly woken up.
3346 if (prepare_kswapd_sleep(pgdat
, order
, remaining
,
3347 balanced_classzone_idx
)) {
3348 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3351 * vmstat counters are not perfectly accurate and the estimated
3352 * value for counters such as NR_FREE_PAGES can deviate from the
3353 * true value by nr_online_cpus * threshold. To avoid the zone
3354 * watermarks being breached while under pressure, we reduce the
3355 * per-cpu vmstat threshold while kswapd is awake and restore
3356 * them before going back to sleep.
3358 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3360 if (!kthread_should_stop())
3363 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3366 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3368 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3370 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3374 * The background pageout daemon, started as a kernel thread
3375 * from the init process.
3377 * This basically trickles out pages so that we have _some_
3378 * free memory available even if there is no other activity
3379 * that frees anything up. This is needed for things like routing
3380 * etc, where we otherwise might have all activity going on in
3381 * asynchronous contexts that cannot page things out.
3383 * If there are applications that are active memory-allocators
3384 * (most normal use), this basically shouldn't matter.
3386 static int kswapd(void *p
)
3388 unsigned long order
, new_order
;
3389 int classzone_idx
, new_classzone_idx
;
3390 int balanced_classzone_idx
;
3391 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3392 struct task_struct
*tsk
= current
;
3394 struct reclaim_state reclaim_state
= {
3395 .reclaimed_slab
= 0,
3397 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3399 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3401 if (!cpumask_empty(cpumask
))
3402 set_cpus_allowed_ptr(tsk
, cpumask
);
3403 current
->reclaim_state
= &reclaim_state
;
3406 * Tell the memory management that we're a "memory allocator",
3407 * and that if we need more memory we should get access to it
3408 * regardless (see "__alloc_pages()"). "kswapd" should
3409 * never get caught in the normal page freeing logic.
3411 * (Kswapd normally doesn't need memory anyway, but sometimes
3412 * you need a small amount of memory in order to be able to
3413 * page out something else, and this flag essentially protects
3414 * us from recursively trying to free more memory as we're
3415 * trying to free the first piece of memory in the first place).
3417 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3420 order
= new_order
= 0;
3421 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3422 balanced_classzone_idx
= classzone_idx
;
3427 * While we were reclaiming, there might have been another
3428 * wakeup, so check the values.
3430 new_order
= pgdat
->kswapd_max_order
;
3431 new_classzone_idx
= pgdat
->classzone_idx
;
3432 pgdat
->kswapd_max_order
= 0;
3433 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3435 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3437 * Don't sleep if someone wants a larger 'order'
3438 * allocation or has tigher zone constraints
3441 classzone_idx
= new_classzone_idx
;
3443 kswapd_try_to_sleep(pgdat
, order
, classzone_idx
,
3444 balanced_classzone_idx
);
3445 order
= pgdat
->kswapd_max_order
;
3446 classzone_idx
= pgdat
->classzone_idx
;
3448 new_classzone_idx
= classzone_idx
;
3449 pgdat
->kswapd_max_order
= 0;
3450 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3453 ret
= try_to_freeze();
3454 if (kthread_should_stop())
3458 * We can speed up thawing tasks if we don't call balance_pgdat
3459 * after returning from the refrigerator
3462 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3463 balanced_classzone_idx
= balance_pgdat(pgdat
, order
,
3468 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3469 current
->reclaim_state
= NULL
;
3470 lockdep_clear_current_reclaim_state();
3476 * A zone is low on free memory, so wake its kswapd task to service it.
3478 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3482 if (!populated_zone(zone
))
3485 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3487 pgdat
= zone
->zone_pgdat
;
3488 if (pgdat
->kswapd_max_order
< order
) {
3489 pgdat
->kswapd_max_order
= order
;
3490 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3492 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3494 if (zone_balanced(zone
, order
, true, 0, 0))
3497 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3498 wake_up_interruptible(&pgdat
->kswapd_wait
);
3501 #ifdef CONFIG_HIBERNATION
3503 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3506 * Rather than trying to age LRUs the aim is to preserve the overall
3507 * LRU order by reclaiming preferentially
3508 * inactive > active > active referenced > active mapped
3510 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3512 struct reclaim_state reclaim_state
;
3513 struct scan_control sc
= {
3514 .nr_to_reclaim
= nr_to_reclaim
,
3515 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3516 .priority
= DEF_PRIORITY
,
3520 .hibernation_mode
= 1,
3522 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3523 struct task_struct
*p
= current
;
3524 unsigned long nr_reclaimed
;
3526 p
->flags
|= PF_MEMALLOC
;
3527 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3528 reclaim_state
.reclaimed_slab
= 0;
3529 p
->reclaim_state
= &reclaim_state
;
3531 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3533 p
->reclaim_state
= NULL
;
3534 lockdep_clear_current_reclaim_state();
3535 p
->flags
&= ~PF_MEMALLOC
;
3537 return nr_reclaimed
;
3539 #endif /* CONFIG_HIBERNATION */
3541 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3542 not required for correctness. So if the last cpu in a node goes
3543 away, we get changed to run anywhere: as the first one comes back,
3544 restore their cpu bindings. */
3545 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3550 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3551 for_each_node_state(nid
, N_MEMORY
) {
3552 pg_data_t
*pgdat
= NODE_DATA(nid
);
3553 const struct cpumask
*mask
;
3555 mask
= cpumask_of_node(pgdat
->node_id
);
3557 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3558 /* One of our CPUs online: restore mask */
3559 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3566 * This kswapd start function will be called by init and node-hot-add.
3567 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3569 int kswapd_run(int nid
)
3571 pg_data_t
*pgdat
= NODE_DATA(nid
);
3577 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3578 if (IS_ERR(pgdat
->kswapd
)) {
3579 /* failure at boot is fatal */
3580 BUG_ON(system_state
== SYSTEM_BOOTING
);
3581 pr_err("Failed to start kswapd on node %d\n", nid
);
3582 ret
= PTR_ERR(pgdat
->kswapd
);
3583 pgdat
->kswapd
= NULL
;
3589 * Called by memory hotplug when all memory in a node is offlined. Caller must
3590 * hold mem_hotplug_begin/end().
3592 void kswapd_stop(int nid
)
3594 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3597 kthread_stop(kswapd
);
3598 NODE_DATA(nid
)->kswapd
= NULL
;
3602 static int __init
kswapd_init(void)
3607 for_each_node_state(nid
, N_MEMORY
)
3609 hotcpu_notifier(cpu_callback
, 0);
3613 module_init(kswapd_init
)
3619 * If non-zero call zone_reclaim when the number of free pages falls below
3622 int zone_reclaim_mode __read_mostly
;
3624 #define RECLAIM_OFF 0
3625 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3626 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3627 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3630 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3631 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3634 #define ZONE_RECLAIM_PRIORITY 4
3637 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3640 int sysctl_min_unmapped_ratio
= 1;
3643 * If the number of slab pages in a zone grows beyond this percentage then
3644 * slab reclaim needs to occur.
3646 int sysctl_min_slab_ratio
= 5;
3648 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3650 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3651 unsigned long file_lru
= node_page_state(zone
->zone_pgdat
, NR_INACTIVE_FILE
) +
3652 node_page_state(zone
->zone_pgdat
, NR_ACTIVE_FILE
);
3655 * It's possible for there to be more file mapped pages than
3656 * accounted for by the pages on the file LRU lists because
3657 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3659 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3662 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3663 static unsigned long zone_pagecache_reclaimable(struct zone
*zone
)
3665 unsigned long nr_pagecache_reclaimable
;
3666 unsigned long delta
= 0;
3669 * If RECLAIM_UNMAP is set, then all file pages are considered
3670 * potentially reclaimable. Otherwise, we have to worry about
3671 * pages like swapcache and zone_unmapped_file_pages() provides
3674 if (zone_reclaim_mode
& RECLAIM_UNMAP
)
3675 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3677 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3679 /* If we can't clean pages, remove dirty pages from consideration */
3680 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3681 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3683 /* Watch for any possible underflows due to delta */
3684 if (unlikely(delta
> nr_pagecache_reclaimable
))
3685 delta
= nr_pagecache_reclaimable
;
3687 return nr_pagecache_reclaimable
- delta
;
3691 * Try to free up some pages from this zone through reclaim.
3693 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3695 /* Minimum pages needed in order to stay on node */
3696 const unsigned long nr_pages
= 1 << order
;
3697 struct task_struct
*p
= current
;
3698 struct reclaim_state reclaim_state
;
3699 struct scan_control sc
= {
3700 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3701 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3703 .priority
= ZONE_RECLAIM_PRIORITY
,
3704 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3705 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_UNMAP
),
3711 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3712 * and we also need to be able to write out pages for RECLAIM_WRITE
3713 * and RECLAIM_UNMAP.
3715 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3716 lockdep_set_current_reclaim_state(gfp_mask
);
3717 reclaim_state
.reclaimed_slab
= 0;
3718 p
->reclaim_state
= &reclaim_state
;
3720 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3722 * Free memory by calling shrink zone with increasing
3723 * priorities until we have enough memory freed.
3726 shrink_zone(zone
, &sc
, true);
3727 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3730 p
->reclaim_state
= NULL
;
3731 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3732 lockdep_clear_current_reclaim_state();
3733 return sc
.nr_reclaimed
>= nr_pages
;
3736 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3742 * Zone reclaim reclaims unmapped file backed pages and
3743 * slab pages if we are over the defined limits.
3745 * A small portion of unmapped file backed pages is needed for
3746 * file I/O otherwise pages read by file I/O will be immediately
3747 * thrown out if the zone is overallocated. So we do not reclaim
3748 * if less than a specified percentage of the zone is used by
3749 * unmapped file backed pages.
3751 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3752 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3753 return ZONE_RECLAIM_FULL
;
3755 if (!pgdat_reclaimable(zone
->zone_pgdat
))
3756 return ZONE_RECLAIM_FULL
;
3759 * Do not scan if the allocation should not be delayed.
3761 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3762 return ZONE_RECLAIM_NOSCAN
;
3765 * Only run zone reclaim on the local zone or on zones that do not
3766 * have associated processors. This will favor the local processor
3767 * over remote processors and spread off node memory allocations
3768 * as wide as possible.
3770 node_id
= zone_to_nid(zone
);
3771 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3772 return ZONE_RECLAIM_NOSCAN
;
3774 if (test_and_set_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
))
3775 return ZONE_RECLAIM_NOSCAN
;
3777 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3778 clear_bit(ZONE_RECLAIM_LOCKED
, &zone
->flags
);
3781 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3788 * page_evictable - test whether a page is evictable
3789 * @page: the page to test
3791 * Test whether page is evictable--i.e., should be placed on active/inactive
3792 * lists vs unevictable list.
3794 * Reasons page might not be evictable:
3795 * (1) page's mapping marked unevictable
3796 * (2) page is part of an mlocked VMA
3799 int page_evictable(struct page
*page
)
3801 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3806 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3807 * @pages: array of pages to check
3808 * @nr_pages: number of pages to check
3810 * Checks pages for evictability and moves them to the appropriate lru list.
3812 * This function is only used for SysV IPC SHM_UNLOCK.
3814 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3816 struct lruvec
*lruvec
;
3817 struct zone
*zone
= NULL
;
3822 for (i
= 0; i
< nr_pages
; i
++) {
3823 struct page
*page
= pages
[i
];
3824 struct zone
*pagezone
;
3827 pagezone
= page_zone(page
);
3828 if (pagezone
!= zone
) {
3830 spin_unlock_irq(zone_lru_lock(zone
));
3832 spin_lock_irq(zone_lru_lock(zone
));
3834 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
3836 if (!PageLRU(page
) || !PageUnevictable(page
))
3839 if (page_evictable(page
)) {
3840 enum lru_list lru
= page_lru_base_type(page
);
3842 VM_BUG_ON_PAGE(PageActive(page
), page
);
3843 ClearPageUnevictable(page
);
3844 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3845 add_page_to_lru_list(page
, lruvec
, lru
);
3851 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3852 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3853 spin_unlock_irq(zone_lru_lock(zone
));
3856 #endif /* CONFIG_SHMEM */