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.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29 #include <linux/mm_inline.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned
;
61 /* Number of pages freed so far during a call to shrink_zones() */
62 unsigned long nr_reclaimed
;
64 /* How many pages shrink_list() should reclaim */
65 unsigned long nr_to_reclaim
;
67 unsigned long hibernation_mode
;
69 /* This context's GFP mask */
74 /* Can mapped pages be reclaimed? */
77 /* Can pages be swapped as part of reclaim? */
82 /* Scan (total_size >> priority) pages at once */
86 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
89 struct mem_cgroup
*target_mem_cgroup
;
92 * Nodemask of nodes allowed by the caller. If NULL, all nodes
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
129 * From 0 .. 100. Higher means more swappy.
131 int vm_swappiness
= 60;
132 unsigned long vm_total_pages
; /* The total number of pages which the VM controls */
134 static LIST_HEAD(shrinker_list
);
135 static DECLARE_RWSEM(shrinker_rwsem
);
138 static bool global_reclaim(struct scan_control
*sc
)
140 return !sc
->target_mem_cgroup
;
143 static bool global_reclaim(struct scan_control
*sc
)
149 unsigned long zone_reclaimable_pages(struct zone
*zone
)
153 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
154 zone_page_state(zone
, NR_INACTIVE_FILE
);
156 if (get_nr_swap_pages() > 0)
157 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
158 zone_page_state(zone
, NR_INACTIVE_ANON
);
163 bool zone_reclaimable(struct zone
*zone
)
165 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
168 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
170 if (!mem_cgroup_disabled())
171 return mem_cgroup_get_lru_size(lruvec
, lru
);
173 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
177 * Add a shrinker callback to be called from the vm.
179 int register_shrinker(struct shrinker
*shrinker
)
181 size_t size
= sizeof(*shrinker
->nr_deferred
);
184 * If we only have one possible node in the system anyway, save
185 * ourselves the trouble and disable NUMA aware behavior. This way we
186 * will save memory and some small loop time later.
188 if (nr_node_ids
== 1)
189 shrinker
->flags
&= ~SHRINKER_NUMA_AWARE
;
191 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
194 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
195 if (!shrinker
->nr_deferred
)
198 down_write(&shrinker_rwsem
);
199 list_add_tail(&shrinker
->list
, &shrinker_list
);
200 up_write(&shrinker_rwsem
);
203 EXPORT_SYMBOL(register_shrinker
);
208 void unregister_shrinker(struct shrinker
*shrinker
)
210 down_write(&shrinker_rwsem
);
211 list_del(&shrinker
->list
);
212 up_write(&shrinker_rwsem
);
214 EXPORT_SYMBOL(unregister_shrinker
);
216 #define SHRINK_BATCH 128
219 shrink_slab_node(struct shrink_control
*shrinkctl
, struct shrinker
*shrinker
,
220 unsigned long nr_pages_scanned
, unsigned long lru_pages
)
222 unsigned long freed
= 0;
223 unsigned long long delta
;
228 int nid
= shrinkctl
->nid
;
229 long batch_size
= shrinker
->batch
? shrinker
->batch
232 max_pass
= shrinker
->count_objects(shrinker
, shrinkctl
);
237 * copy the current shrinker scan count into a local variable
238 * and zero it so that other concurrent shrinker invocations
239 * don't also do this scanning work.
241 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
244 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
246 do_div(delta
, lru_pages
+ 1);
248 if (total_scan
< 0) {
250 "shrink_slab: %pF negative objects to delete nr=%ld\n",
251 shrinker
->scan_objects
, total_scan
);
252 total_scan
= max_pass
;
256 * We need to avoid excessive windup on filesystem shrinkers
257 * due to large numbers of GFP_NOFS allocations causing the
258 * shrinkers to return -1 all the time. This results in a large
259 * nr being built up so when a shrink that can do some work
260 * comes along it empties the entire cache due to nr >>>
261 * max_pass. This is bad for sustaining a working set in
264 * Hence only allow the shrinker to scan the entire cache when
265 * a large delta change is calculated directly.
267 if (delta
< max_pass
/ 4)
268 total_scan
= min(total_scan
, max_pass
/ 2);
271 * Avoid risking looping forever due to too large nr value:
272 * never try to free more than twice the estimate number of
275 if (total_scan
> max_pass
* 2)
276 total_scan
= max_pass
* 2;
278 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
279 nr_pages_scanned
, lru_pages
,
280 max_pass
, delta
, total_scan
);
282 while (total_scan
>= batch_size
) {
285 shrinkctl
->nr_to_scan
= batch_size
;
286 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
287 if (ret
== SHRINK_STOP
)
291 count_vm_events(SLABS_SCANNED
, batch_size
);
292 total_scan
-= batch_size
;
298 * move the unused scan count back into the shrinker in a
299 * manner that handles concurrent updates. If we exhausted the
300 * scan, there is no need to do an update.
303 new_nr
= atomic_long_add_return(total_scan
,
304 &shrinker
->nr_deferred
[nid
]);
306 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
308 trace_mm_shrink_slab_end(shrinker
, freed
, nr
, new_nr
);
313 * Call the shrink functions to age shrinkable caches
315 * Here we assume it costs one seek to replace a lru page and that it also
316 * takes a seek to recreate a cache object. With this in mind we age equal
317 * percentages of the lru and ageable caches. This should balance the seeks
318 * generated by these structures.
320 * If the vm encountered mapped pages on the LRU it increase the pressure on
321 * slab to avoid swapping.
323 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
325 * `lru_pages' represents the number of on-LRU pages in all the zones which
326 * are eligible for the caller's allocation attempt. It is used for balancing
327 * slab reclaim versus page reclaim.
329 * Returns the number of slab objects which we shrunk.
331 unsigned long shrink_slab(struct shrink_control
*shrinkctl
,
332 unsigned long nr_pages_scanned
,
333 unsigned long lru_pages
)
335 struct shrinker
*shrinker
;
336 unsigned long freed
= 0;
338 if (nr_pages_scanned
== 0)
339 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
341 if (!down_read_trylock(&shrinker_rwsem
)) {
343 * If we would return 0, our callers would understand that we
344 * have nothing else to shrink and give up trying. By returning
345 * 1 we keep it going and assume we'll be able to shrink next
352 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
353 for_each_node_mask(shrinkctl
->nid
, shrinkctl
->nodes_to_scan
) {
354 if (!node_online(shrinkctl
->nid
))
357 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
) &&
358 (shrinkctl
->nid
!= 0))
361 freed
+= shrink_slab_node(shrinkctl
, shrinker
,
362 nr_pages_scanned
, lru_pages
);
366 up_read(&shrinker_rwsem
);
372 static inline int is_page_cache_freeable(struct page
*page
)
375 * A freeable page cache page is referenced only by the caller
376 * that isolated the page, the page cache radix tree and
377 * optional buffer heads at page->private.
379 return page_count(page
) - page_has_private(page
) == 2;
382 static int may_write_to_queue(struct backing_dev_info
*bdi
,
383 struct scan_control
*sc
)
385 if (current
->flags
& PF_SWAPWRITE
)
387 if (!bdi_write_congested(bdi
))
389 if (bdi
== current
->backing_dev_info
)
395 * We detected a synchronous write error writing a page out. Probably
396 * -ENOSPC. We need to propagate that into the address_space for a subsequent
397 * fsync(), msync() or close().
399 * The tricky part is that after writepage we cannot touch the mapping: nothing
400 * prevents it from being freed up. But we have a ref on the page and once
401 * that page is locked, the mapping is pinned.
403 * We're allowed to run sleeping lock_page() here because we know the caller has
406 static void handle_write_error(struct address_space
*mapping
,
407 struct page
*page
, int error
)
410 if (page_mapping(page
) == mapping
)
411 mapping_set_error(mapping
, error
);
415 /* possible outcome of pageout() */
417 /* failed to write page out, page is locked */
419 /* move page to the active list, page is locked */
421 /* page has been sent to the disk successfully, page is unlocked */
423 /* page is clean and locked */
428 * pageout is called by shrink_page_list() for each dirty page.
429 * Calls ->writepage().
431 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
432 struct scan_control
*sc
)
435 * If the page is dirty, only perform writeback if that write
436 * will be non-blocking. To prevent this allocation from being
437 * stalled by pagecache activity. But note that there may be
438 * stalls if we need to run get_block(). We could test
439 * PagePrivate for that.
441 * If this process is currently in __generic_file_aio_write() against
442 * this page's queue, we can perform writeback even if that
445 * If the page is swapcache, write it back even if that would
446 * block, for some throttling. This happens by accident, because
447 * swap_backing_dev_info is bust: it doesn't reflect the
448 * congestion state of the swapdevs. Easy to fix, if needed.
450 if (!is_page_cache_freeable(page
))
454 * Some data journaling orphaned pages can have
455 * page->mapping == NULL while being dirty with clean buffers.
457 if (page_has_private(page
)) {
458 if (try_to_free_buffers(page
)) {
459 ClearPageDirty(page
);
460 printk("%s: orphaned page\n", __func__
);
466 if (mapping
->a_ops
->writepage
== NULL
)
467 return PAGE_ACTIVATE
;
468 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
471 if (clear_page_dirty_for_io(page
)) {
473 struct writeback_control wbc
= {
474 .sync_mode
= WB_SYNC_NONE
,
475 .nr_to_write
= SWAP_CLUSTER_MAX
,
477 .range_end
= LLONG_MAX
,
481 SetPageReclaim(page
);
482 res
= mapping
->a_ops
->writepage(page
, &wbc
);
484 handle_write_error(mapping
, page
, res
);
485 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
486 ClearPageReclaim(page
);
487 return PAGE_ACTIVATE
;
490 if (!PageWriteback(page
)) {
491 /* synchronous write or broken a_ops? */
492 ClearPageReclaim(page
);
494 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
495 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
503 * Same as remove_mapping, but if the page is removed from the mapping, it
504 * gets returned with a refcount of 0.
506 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
508 BUG_ON(!PageLocked(page
));
509 BUG_ON(mapping
!= page_mapping(page
));
511 spin_lock_irq(&mapping
->tree_lock
);
513 * The non racy check for a busy page.
515 * Must be careful with the order of the tests. When someone has
516 * a ref to the page, it may be possible that they dirty it then
517 * drop the reference. So if PageDirty is tested before page_count
518 * here, then the following race may occur:
520 * get_user_pages(&page);
521 * [user mapping goes away]
523 * !PageDirty(page) [good]
524 * SetPageDirty(page);
526 * !page_count(page) [good, discard it]
528 * [oops, our write_to data is lost]
530 * Reversing the order of the tests ensures such a situation cannot
531 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
532 * load is not satisfied before that of page->_count.
534 * Note that if SetPageDirty is always performed via set_page_dirty,
535 * and thus under tree_lock, then this ordering is not required.
537 if (!page_freeze_refs(page
, 2))
539 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
540 if (unlikely(PageDirty(page
))) {
541 page_unfreeze_refs(page
, 2);
545 if (PageSwapCache(page
)) {
546 swp_entry_t swap
= { .val
= page_private(page
) };
547 __delete_from_swap_cache(page
);
548 spin_unlock_irq(&mapping
->tree_lock
);
549 swapcache_free(swap
, page
);
551 void (*freepage
)(struct page
*);
553 freepage
= mapping
->a_ops
->freepage
;
555 __delete_from_page_cache(page
);
556 spin_unlock_irq(&mapping
->tree_lock
);
557 mem_cgroup_uncharge_cache_page(page
);
559 if (freepage
!= NULL
)
566 spin_unlock_irq(&mapping
->tree_lock
);
571 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
572 * someone else has a ref on the page, abort and return 0. If it was
573 * successfully detached, return 1. Assumes the caller has a single ref on
576 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
578 if (__remove_mapping(mapping
, page
)) {
580 * Unfreezing the refcount with 1 rather than 2 effectively
581 * drops the pagecache ref for us without requiring another
584 page_unfreeze_refs(page
, 1);
591 * putback_lru_page - put previously isolated page onto appropriate LRU list
592 * @page: page to be put back to appropriate lru list
594 * Add previously isolated @page to appropriate LRU list.
595 * Page may still be unevictable for other reasons.
597 * lru_lock must not be held, interrupts must be enabled.
599 void putback_lru_page(struct page
*page
)
602 int was_unevictable
= PageUnevictable(page
);
604 VM_BUG_ON(PageLRU(page
));
607 ClearPageUnevictable(page
);
609 if (page_evictable(page
)) {
611 * For evictable pages, we can use the cache.
612 * In event of a race, worst case is we end up with an
613 * unevictable page on [in]active list.
614 * We know how to handle that.
616 is_unevictable
= false;
620 * Put unevictable pages directly on zone's unevictable
623 is_unevictable
= true;
624 add_page_to_unevictable_list(page
);
626 * When racing with an mlock or AS_UNEVICTABLE clearing
627 * (page is unlocked) make sure that if the other thread
628 * does not observe our setting of PG_lru and fails
629 * isolation/check_move_unevictable_pages,
630 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
631 * the page back to the evictable list.
633 * The other side is TestClearPageMlocked() or shmem_lock().
639 * page's status can change while we move it among lru. If an evictable
640 * page is on unevictable list, it never be freed. To avoid that,
641 * check after we added it to the list, again.
643 if (is_unevictable
&& page_evictable(page
)) {
644 if (!isolate_lru_page(page
)) {
648 /* This means someone else dropped this page from LRU
649 * So, it will be freed or putback to LRU again. There is
650 * nothing to do here.
654 if (was_unevictable
&& !is_unevictable
)
655 count_vm_event(UNEVICTABLE_PGRESCUED
);
656 else if (!was_unevictable
&& is_unevictable
)
657 count_vm_event(UNEVICTABLE_PGCULLED
);
659 put_page(page
); /* drop ref from isolate */
662 enum page_references
{
664 PAGEREF_RECLAIM_CLEAN
,
669 static enum page_references
page_check_references(struct page
*page
,
670 struct scan_control
*sc
)
672 int referenced_ptes
, referenced_page
;
673 unsigned long vm_flags
;
675 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
677 referenced_page
= TestClearPageReferenced(page
);
680 * Mlock lost the isolation race with us. Let try_to_unmap()
681 * move the page to the unevictable list.
683 if (vm_flags
& VM_LOCKED
)
684 return PAGEREF_RECLAIM
;
686 if (referenced_ptes
) {
687 if (PageSwapBacked(page
))
688 return PAGEREF_ACTIVATE
;
690 * All mapped pages start out with page table
691 * references from the instantiating fault, so we need
692 * to look twice if a mapped file page is used more
695 * Mark it and spare it for another trip around the
696 * inactive list. Another page table reference will
697 * lead to its activation.
699 * Note: the mark is set for activated pages as well
700 * so that recently deactivated but used pages are
703 SetPageReferenced(page
);
705 if (referenced_page
|| referenced_ptes
> 1)
706 return PAGEREF_ACTIVATE
;
709 * Activate file-backed executable pages after first usage.
711 if (vm_flags
& VM_EXEC
)
712 return PAGEREF_ACTIVATE
;
717 /* Reclaim if clean, defer dirty pages to writeback */
718 if (referenced_page
&& !PageSwapBacked(page
))
719 return PAGEREF_RECLAIM_CLEAN
;
721 return PAGEREF_RECLAIM
;
724 /* Check if a page is dirty or under writeback */
725 static void page_check_dirty_writeback(struct page
*page
,
726 bool *dirty
, bool *writeback
)
728 struct address_space
*mapping
;
731 * Anonymous pages are not handled by flushers and must be written
732 * from reclaim context. Do not stall reclaim based on them
734 if (!page_is_file_cache(page
)) {
740 /* By default assume that the page flags are accurate */
741 *dirty
= PageDirty(page
);
742 *writeback
= PageWriteback(page
);
744 /* Verify dirty/writeback state if the filesystem supports it */
745 if (!page_has_private(page
))
748 mapping
= page_mapping(page
);
749 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
750 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
754 * shrink_page_list() returns the number of reclaimed pages
756 static unsigned long shrink_page_list(struct list_head
*page_list
,
758 struct scan_control
*sc
,
759 enum ttu_flags ttu_flags
,
760 unsigned long *ret_nr_dirty
,
761 unsigned long *ret_nr_unqueued_dirty
,
762 unsigned long *ret_nr_congested
,
763 unsigned long *ret_nr_writeback
,
764 unsigned long *ret_nr_immediate
,
767 LIST_HEAD(ret_pages
);
768 LIST_HEAD(free_pages
);
770 unsigned long nr_unqueued_dirty
= 0;
771 unsigned long nr_dirty
= 0;
772 unsigned long nr_congested
= 0;
773 unsigned long nr_reclaimed
= 0;
774 unsigned long nr_writeback
= 0;
775 unsigned long nr_immediate
= 0;
779 mem_cgroup_uncharge_start();
780 while (!list_empty(page_list
)) {
781 struct address_space
*mapping
;
784 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
785 bool dirty
, writeback
;
789 page
= lru_to_page(page_list
);
790 list_del(&page
->lru
);
792 if (!trylock_page(page
))
795 VM_BUG_ON(PageActive(page
));
796 VM_BUG_ON(page_zone(page
) != zone
);
800 if (unlikely(!page_evictable(page
)))
803 if (!sc
->may_unmap
&& page_mapped(page
))
806 /* Double the slab pressure for mapped and swapcache pages */
807 if (page_mapped(page
) || PageSwapCache(page
))
810 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
811 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
814 * The number of dirty pages determines if a zone is marked
815 * reclaim_congested which affects wait_iff_congested. kswapd
816 * will stall and start writing pages if the tail of the LRU
817 * is all dirty unqueued pages.
819 page_check_dirty_writeback(page
, &dirty
, &writeback
);
820 if (dirty
|| writeback
)
823 if (dirty
&& !writeback
)
827 * Treat this page as congested if the underlying BDI is or if
828 * pages are cycling through the LRU so quickly that the
829 * pages marked for immediate reclaim are making it to the
830 * end of the LRU a second time.
832 mapping
= page_mapping(page
);
833 if ((mapping
&& bdi_write_congested(mapping
->backing_dev_info
)) ||
834 (writeback
&& PageReclaim(page
)))
838 * If a page at the tail of the LRU is under writeback, there
839 * are three cases to consider.
841 * 1) If reclaim is encountering an excessive number of pages
842 * under writeback and this page is both under writeback and
843 * PageReclaim then it indicates that pages are being queued
844 * for IO but are being recycled through the LRU before the
845 * IO can complete. Waiting on the page itself risks an
846 * indefinite stall if it is impossible to writeback the
847 * page due to IO error or disconnected storage so instead
848 * note that the LRU is being scanned too quickly and the
849 * caller can stall after page list has been processed.
851 * 2) Global reclaim encounters a page, memcg encounters a
852 * page that is not marked for immediate reclaim or
853 * the caller does not have __GFP_IO. In this case mark
854 * the page for immediate reclaim and continue scanning.
856 * __GFP_IO is checked because a loop driver thread might
857 * enter reclaim, and deadlock if it waits on a page for
858 * which it is needed to do the write (loop masks off
859 * __GFP_IO|__GFP_FS for this reason); but more thought
860 * would probably show more reasons.
862 * Don't require __GFP_FS, since we're not going into the
863 * FS, just waiting on its writeback completion. Worryingly,
864 * ext4 gfs2 and xfs allocate pages with
865 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
866 * may_enter_fs here is liable to OOM on them.
868 * 3) memcg encounters a page that is not already marked
869 * PageReclaim. memcg does not have any dirty pages
870 * throttling so we could easily OOM just because too many
871 * pages are in writeback and there is nothing else to
872 * reclaim. Wait for the writeback to complete.
874 if (PageWriteback(page
)) {
876 if (current_is_kswapd() &&
878 zone_is_reclaim_writeback(zone
)) {
883 } else if (global_reclaim(sc
) ||
884 !PageReclaim(page
) || !(sc
->gfp_mask
& __GFP_IO
)) {
886 * This is slightly racy - end_page_writeback()
887 * might have just cleared PageReclaim, then
888 * setting PageReclaim here end up interpreted
889 * as PageReadahead - but that does not matter
890 * enough to care. What we do want is for this
891 * page to have PageReclaim set next time memcg
892 * reclaim reaches the tests above, so it will
893 * then wait_on_page_writeback() to avoid OOM;
894 * and it's also appropriate in global reclaim.
896 SetPageReclaim(page
);
903 wait_on_page_writeback(page
);
908 references
= page_check_references(page
, sc
);
910 switch (references
) {
911 case PAGEREF_ACTIVATE
:
912 goto activate_locked
;
915 case PAGEREF_RECLAIM
:
916 case PAGEREF_RECLAIM_CLEAN
:
917 ; /* try to reclaim the page below */
921 * Anonymous process memory has backing store?
922 * Try to allocate it some swap space here.
924 if (PageAnon(page
) && !PageSwapCache(page
)) {
925 if (!(sc
->gfp_mask
& __GFP_IO
))
927 if (!add_to_swap(page
, page_list
))
928 goto activate_locked
;
931 /* Adding to swap updated mapping */
932 mapping
= page_mapping(page
);
936 * The page is mapped into the page tables of one or more
937 * processes. Try to unmap it here.
939 if (page_mapped(page
) && mapping
) {
940 switch (try_to_unmap(page
, ttu_flags
)) {
942 goto activate_locked
;
948 ; /* try to free the page below */
952 if (PageDirty(page
)) {
954 * Only kswapd can writeback filesystem pages to
955 * avoid risk of stack overflow but only writeback
956 * if many dirty pages have been encountered.
958 if (page_is_file_cache(page
) &&
959 (!current_is_kswapd() ||
960 !zone_is_reclaim_dirty(zone
))) {
962 * Immediately reclaim when written back.
963 * Similar in principal to deactivate_page()
964 * except we already have the page isolated
965 * and know it's dirty
967 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
968 SetPageReclaim(page
);
973 if (references
== PAGEREF_RECLAIM_CLEAN
)
977 if (!sc
->may_writepage
)
980 /* Page is dirty, try to write it out here */
981 switch (pageout(page
, mapping
, sc
)) {
985 goto activate_locked
;
987 if (PageWriteback(page
))
993 * A synchronous write - probably a ramdisk. Go
994 * ahead and try to reclaim the page.
996 if (!trylock_page(page
))
998 if (PageDirty(page
) || PageWriteback(page
))
1000 mapping
= page_mapping(page
);
1002 ; /* try to free the page below */
1007 * If the page has buffers, try to free the buffer mappings
1008 * associated with this page. If we succeed we try to free
1011 * We do this even if the page is PageDirty().
1012 * try_to_release_page() does not perform I/O, but it is
1013 * possible for a page to have PageDirty set, but it is actually
1014 * clean (all its buffers are clean). This happens if the
1015 * buffers were written out directly, with submit_bh(). ext3
1016 * will do this, as well as the blockdev mapping.
1017 * try_to_release_page() will discover that cleanness and will
1018 * drop the buffers and mark the page clean - it can be freed.
1020 * Rarely, pages can have buffers and no ->mapping. These are
1021 * the pages which were not successfully invalidated in
1022 * truncate_complete_page(). We try to drop those buffers here
1023 * and if that worked, and the page is no longer mapped into
1024 * process address space (page_count == 1) it can be freed.
1025 * Otherwise, leave the page on the LRU so it is swappable.
1027 if (page_has_private(page
)) {
1028 if (!try_to_release_page(page
, sc
->gfp_mask
))
1029 goto activate_locked
;
1030 if (!mapping
&& page_count(page
) == 1) {
1032 if (put_page_testzero(page
))
1036 * rare race with speculative reference.
1037 * the speculative reference will free
1038 * this page shortly, so we may
1039 * increment nr_reclaimed here (and
1040 * leave it off the LRU).
1048 if (!mapping
|| !__remove_mapping(mapping
, page
))
1052 * At this point, we have no other references and there is
1053 * no way to pick any more up (removed from LRU, removed
1054 * from pagecache). Can use non-atomic bitops now (and
1055 * we obviously don't have to worry about waking up a process
1056 * waiting on the page lock, because there are no references.
1058 __clear_page_locked(page
);
1063 * Is there need to periodically free_page_list? It would
1064 * appear not as the counts should be low
1066 list_add(&page
->lru
, &free_pages
);
1070 if (PageSwapCache(page
))
1071 try_to_free_swap(page
);
1073 putback_lru_page(page
);
1077 /* Not a candidate for swapping, so reclaim swap space. */
1078 if (PageSwapCache(page
) && vm_swap_full())
1079 try_to_free_swap(page
);
1080 VM_BUG_ON(PageActive(page
));
1081 SetPageActive(page
);
1086 list_add(&page
->lru
, &ret_pages
);
1087 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
1090 free_hot_cold_page_list(&free_pages
, 1);
1092 list_splice(&ret_pages
, page_list
);
1093 count_vm_events(PGACTIVATE
, pgactivate
);
1094 mem_cgroup_uncharge_end();
1095 *ret_nr_dirty
+= nr_dirty
;
1096 *ret_nr_congested
+= nr_congested
;
1097 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1098 *ret_nr_writeback
+= nr_writeback
;
1099 *ret_nr_immediate
+= nr_immediate
;
1100 return nr_reclaimed
;
1103 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1104 struct list_head
*page_list
)
1106 struct scan_control sc
= {
1107 .gfp_mask
= GFP_KERNEL
,
1108 .priority
= DEF_PRIORITY
,
1111 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1112 struct page
*page
, *next
;
1113 LIST_HEAD(clean_pages
);
1115 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1116 if (page_is_file_cache(page
) && !PageDirty(page
)) {
1117 ClearPageActive(page
);
1118 list_move(&page
->lru
, &clean_pages
);
1122 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1123 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1124 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1125 list_splice(&clean_pages
, page_list
);
1126 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1131 * Attempt to remove the specified page from its LRU. Only take this page
1132 * if it is of the appropriate PageActive status. Pages which are being
1133 * freed elsewhere are also ignored.
1135 * page: page to consider
1136 * mode: one of the LRU isolation modes defined above
1138 * returns 0 on success, -ve errno on failure.
1140 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1144 /* Only take pages on the LRU. */
1148 /* Compaction should not handle unevictable pages but CMA can do so */
1149 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1155 * To minimise LRU disruption, the caller can indicate that it only
1156 * wants to isolate pages it will be able to operate on without
1157 * blocking - clean pages for the most part.
1159 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1160 * is used by reclaim when it is cannot write to backing storage
1162 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1163 * that it is possible to migrate without blocking
1165 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1166 /* All the caller can do on PageWriteback is block */
1167 if (PageWriteback(page
))
1170 if (PageDirty(page
)) {
1171 struct address_space
*mapping
;
1173 /* ISOLATE_CLEAN means only clean pages */
1174 if (mode
& ISOLATE_CLEAN
)
1178 * Only pages without mappings or that have a
1179 * ->migratepage callback are possible to migrate
1182 mapping
= page_mapping(page
);
1183 if (mapping
&& !mapping
->a_ops
->migratepage
)
1188 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1191 if (likely(get_page_unless_zero(page
))) {
1193 * Be careful not to clear PageLRU until after we're
1194 * sure the page is not being freed elsewhere -- the
1195 * page release code relies on it.
1205 * zone->lru_lock is heavily contended. Some of the functions that
1206 * shrink the lists perform better by taking out a batch of pages
1207 * and working on them outside the LRU lock.
1209 * For pagecache intensive workloads, this function is the hottest
1210 * spot in the kernel (apart from copy_*_user functions).
1212 * Appropriate locks must be held before calling this function.
1214 * @nr_to_scan: The number of pages to look through on the list.
1215 * @lruvec: The LRU vector to pull pages from.
1216 * @dst: The temp list to put pages on to.
1217 * @nr_scanned: The number of pages that were scanned.
1218 * @sc: The scan_control struct for this reclaim session
1219 * @mode: One of the LRU isolation modes
1220 * @lru: LRU list id for isolating
1222 * returns how many pages were moved onto *@dst.
1224 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1225 struct lruvec
*lruvec
, struct list_head
*dst
,
1226 unsigned long *nr_scanned
, struct scan_control
*sc
,
1227 isolate_mode_t mode
, enum lru_list lru
)
1229 struct list_head
*src
= &lruvec
->lists
[lru
];
1230 unsigned long nr_taken
= 0;
1233 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1237 page
= lru_to_page(src
);
1238 prefetchw_prev_lru_page(page
, src
, flags
);
1240 VM_BUG_ON(!PageLRU(page
));
1242 switch (__isolate_lru_page(page
, mode
)) {
1244 nr_pages
= hpage_nr_pages(page
);
1245 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1246 list_move(&page
->lru
, dst
);
1247 nr_taken
+= nr_pages
;
1251 /* else it is being freed elsewhere */
1252 list_move(&page
->lru
, src
);
1261 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1262 nr_taken
, mode
, is_file_lru(lru
));
1267 * isolate_lru_page - tries to isolate a page from its LRU list
1268 * @page: page to isolate from its LRU list
1270 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1271 * vmstat statistic corresponding to whatever LRU list the page was on.
1273 * Returns 0 if the page was removed from an LRU list.
1274 * Returns -EBUSY if the page was not on an LRU list.
1276 * The returned page will have PageLRU() cleared. If it was found on
1277 * the active list, it will have PageActive set. If it was found on
1278 * the unevictable list, it will have the PageUnevictable bit set. That flag
1279 * may need to be cleared by the caller before letting the page go.
1281 * The vmstat statistic corresponding to the list on which the page was
1282 * found will be decremented.
1285 * (1) Must be called with an elevated refcount on the page. This is a
1286 * fundamentnal difference from isolate_lru_pages (which is called
1287 * without a stable reference).
1288 * (2) the lru_lock must not be held.
1289 * (3) interrupts must be enabled.
1291 int isolate_lru_page(struct page
*page
)
1295 VM_BUG_ON(!page_count(page
));
1297 if (PageLRU(page
)) {
1298 struct zone
*zone
= page_zone(page
);
1299 struct lruvec
*lruvec
;
1301 spin_lock_irq(&zone
->lru_lock
);
1302 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1303 if (PageLRU(page
)) {
1304 int lru
= page_lru(page
);
1307 del_page_from_lru_list(page
, lruvec
, lru
);
1310 spin_unlock_irq(&zone
->lru_lock
);
1316 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1317 * then get resheduled. When there are massive number of tasks doing page
1318 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1319 * the LRU list will go small and be scanned faster than necessary, leading to
1320 * unnecessary swapping, thrashing and OOM.
1322 static int too_many_isolated(struct zone
*zone
, int file
,
1323 struct scan_control
*sc
)
1325 unsigned long inactive
, isolated
;
1327 if (current_is_kswapd())
1330 if (!global_reclaim(sc
))
1334 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1335 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1337 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1338 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1342 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1343 * won't get blocked by normal direct-reclaimers, forming a circular
1346 if ((sc
->gfp_mask
& GFP_IOFS
) == GFP_IOFS
)
1349 return isolated
> inactive
;
1352 static noinline_for_stack
void
1353 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1355 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1356 struct zone
*zone
= lruvec_zone(lruvec
);
1357 LIST_HEAD(pages_to_free
);
1360 * Put back any unfreeable pages.
1362 while (!list_empty(page_list
)) {
1363 struct page
*page
= lru_to_page(page_list
);
1366 VM_BUG_ON(PageLRU(page
));
1367 list_del(&page
->lru
);
1368 if (unlikely(!page_evictable(page
))) {
1369 spin_unlock_irq(&zone
->lru_lock
);
1370 putback_lru_page(page
);
1371 spin_lock_irq(&zone
->lru_lock
);
1375 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1378 lru
= page_lru(page
);
1379 add_page_to_lru_list(page
, lruvec
, lru
);
1381 if (is_active_lru(lru
)) {
1382 int file
= is_file_lru(lru
);
1383 int numpages
= hpage_nr_pages(page
);
1384 reclaim_stat
->recent_rotated
[file
] += numpages
;
1386 if (put_page_testzero(page
)) {
1387 __ClearPageLRU(page
);
1388 __ClearPageActive(page
);
1389 del_page_from_lru_list(page
, lruvec
, lru
);
1391 if (unlikely(PageCompound(page
))) {
1392 spin_unlock_irq(&zone
->lru_lock
);
1393 (*get_compound_page_dtor(page
))(page
);
1394 spin_lock_irq(&zone
->lru_lock
);
1396 list_add(&page
->lru
, &pages_to_free
);
1401 * To save our caller's stack, now use input list for pages to free.
1403 list_splice(&pages_to_free
, page_list
);
1407 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1408 * of reclaimed pages
1410 static noinline_for_stack
unsigned long
1411 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1412 struct scan_control
*sc
, enum lru_list lru
)
1414 LIST_HEAD(page_list
);
1415 unsigned long nr_scanned
;
1416 unsigned long nr_reclaimed
= 0;
1417 unsigned long nr_taken
;
1418 unsigned long nr_dirty
= 0;
1419 unsigned long nr_congested
= 0;
1420 unsigned long nr_unqueued_dirty
= 0;
1421 unsigned long nr_writeback
= 0;
1422 unsigned long nr_immediate
= 0;
1423 isolate_mode_t isolate_mode
= 0;
1424 int file
= is_file_lru(lru
);
1425 struct zone
*zone
= lruvec_zone(lruvec
);
1426 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1428 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1429 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1431 /* We are about to die and free our memory. Return now. */
1432 if (fatal_signal_pending(current
))
1433 return SWAP_CLUSTER_MAX
;
1439 isolate_mode
|= ISOLATE_UNMAPPED
;
1440 if (!sc
->may_writepage
)
1441 isolate_mode
|= ISOLATE_CLEAN
;
1443 spin_lock_irq(&zone
->lru_lock
);
1445 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1446 &nr_scanned
, sc
, isolate_mode
, lru
);
1448 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1449 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1451 if (global_reclaim(sc
)) {
1452 zone
->pages_scanned
+= nr_scanned
;
1453 if (current_is_kswapd())
1454 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1456 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1458 spin_unlock_irq(&zone
->lru_lock
);
1463 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1464 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1465 &nr_writeback
, &nr_immediate
,
1468 spin_lock_irq(&zone
->lru_lock
);
1470 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1472 if (global_reclaim(sc
)) {
1473 if (current_is_kswapd())
1474 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1477 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1481 putback_inactive_pages(lruvec
, &page_list
);
1483 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1485 spin_unlock_irq(&zone
->lru_lock
);
1487 free_hot_cold_page_list(&page_list
, 1);
1490 * If reclaim is isolating dirty pages under writeback, it implies
1491 * that the long-lived page allocation rate is exceeding the page
1492 * laundering rate. Either the global limits are not being effective
1493 * at throttling processes due to the page distribution throughout
1494 * zones or there is heavy usage of a slow backing device. The
1495 * only option is to throttle from reclaim context which is not ideal
1496 * as there is no guarantee the dirtying process is throttled in the
1497 * same way balance_dirty_pages() manages.
1499 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1500 * of pages under pages flagged for immediate reclaim and stall if any
1501 * are encountered in the nr_immediate check below.
1503 if (nr_writeback
&& nr_writeback
== nr_taken
)
1504 zone_set_flag(zone
, ZONE_WRITEBACK
);
1507 * memcg will stall in page writeback so only consider forcibly
1508 * stalling for global reclaim
1510 if (global_reclaim(sc
)) {
1512 * Tag a zone as congested if all the dirty pages scanned were
1513 * backed by a congested BDI and wait_iff_congested will stall.
1515 if (nr_dirty
&& nr_dirty
== nr_congested
)
1516 zone_set_flag(zone
, ZONE_CONGESTED
);
1519 * If dirty pages are scanned that are not queued for IO, it
1520 * implies that flushers are not keeping up. In this case, flag
1521 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1522 * pages from reclaim context. It will forcibly stall in the
1525 if (nr_unqueued_dirty
== nr_taken
)
1526 zone_set_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
1529 * In addition, if kswapd scans pages marked marked for
1530 * immediate reclaim and under writeback (nr_immediate), it
1531 * implies that pages are cycling through the LRU faster than
1532 * they are written so also forcibly stall.
1534 if (nr_unqueued_dirty
== nr_taken
|| nr_immediate
)
1535 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1539 * Stall direct reclaim for IO completions if underlying BDIs or zone
1540 * is congested. Allow kswapd to continue until it starts encountering
1541 * unqueued dirty pages or cycling through the LRU too quickly.
1543 if (!sc
->hibernation_mode
&& !current_is_kswapd())
1544 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1546 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1548 nr_scanned
, nr_reclaimed
,
1550 trace_shrink_flags(file
));
1551 return nr_reclaimed
;
1555 * This moves pages from the active list to the inactive list.
1557 * We move them the other way if the page is referenced by one or more
1558 * processes, from rmap.
1560 * If the pages are mostly unmapped, the processing is fast and it is
1561 * appropriate to hold zone->lru_lock across the whole operation. But if
1562 * the pages are mapped, the processing is slow (page_referenced()) so we
1563 * should drop zone->lru_lock around each page. It's impossible to balance
1564 * this, so instead we remove the pages from the LRU while processing them.
1565 * It is safe to rely on PG_active against the non-LRU pages in here because
1566 * nobody will play with that bit on a non-LRU page.
1568 * The downside is that we have to touch page->_count against each page.
1569 * But we had to alter page->flags anyway.
1572 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1573 struct list_head
*list
,
1574 struct list_head
*pages_to_free
,
1577 struct zone
*zone
= lruvec_zone(lruvec
);
1578 unsigned long pgmoved
= 0;
1582 while (!list_empty(list
)) {
1583 page
= lru_to_page(list
);
1584 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1586 VM_BUG_ON(PageLRU(page
));
1589 nr_pages
= hpage_nr_pages(page
);
1590 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1591 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1592 pgmoved
+= nr_pages
;
1594 if (put_page_testzero(page
)) {
1595 __ClearPageLRU(page
);
1596 __ClearPageActive(page
);
1597 del_page_from_lru_list(page
, lruvec
, lru
);
1599 if (unlikely(PageCompound(page
))) {
1600 spin_unlock_irq(&zone
->lru_lock
);
1601 (*get_compound_page_dtor(page
))(page
);
1602 spin_lock_irq(&zone
->lru_lock
);
1604 list_add(&page
->lru
, pages_to_free
);
1607 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1608 if (!is_active_lru(lru
))
1609 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1612 static void shrink_active_list(unsigned long nr_to_scan
,
1613 struct lruvec
*lruvec
,
1614 struct scan_control
*sc
,
1617 unsigned long nr_taken
;
1618 unsigned long nr_scanned
;
1619 unsigned long vm_flags
;
1620 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1621 LIST_HEAD(l_active
);
1622 LIST_HEAD(l_inactive
);
1624 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1625 unsigned long nr_rotated
= 0;
1626 isolate_mode_t isolate_mode
= 0;
1627 int file
= is_file_lru(lru
);
1628 struct zone
*zone
= lruvec_zone(lruvec
);
1633 isolate_mode
|= ISOLATE_UNMAPPED
;
1634 if (!sc
->may_writepage
)
1635 isolate_mode
|= ISOLATE_CLEAN
;
1637 spin_lock_irq(&zone
->lru_lock
);
1639 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1640 &nr_scanned
, sc
, isolate_mode
, lru
);
1641 if (global_reclaim(sc
))
1642 zone
->pages_scanned
+= nr_scanned
;
1644 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1646 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1647 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1648 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1649 spin_unlock_irq(&zone
->lru_lock
);
1651 while (!list_empty(&l_hold
)) {
1653 page
= lru_to_page(&l_hold
);
1654 list_del(&page
->lru
);
1656 if (unlikely(!page_evictable(page
))) {
1657 putback_lru_page(page
);
1661 if (unlikely(buffer_heads_over_limit
)) {
1662 if (page_has_private(page
) && trylock_page(page
)) {
1663 if (page_has_private(page
))
1664 try_to_release_page(page
, 0);
1669 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1671 nr_rotated
+= hpage_nr_pages(page
);
1673 * Identify referenced, file-backed active pages and
1674 * give them one more trip around the active list. So
1675 * that executable code get better chances to stay in
1676 * memory under moderate memory pressure. Anon pages
1677 * are not likely to be evicted by use-once streaming
1678 * IO, plus JVM can create lots of anon VM_EXEC pages,
1679 * so we ignore them here.
1681 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1682 list_add(&page
->lru
, &l_active
);
1687 ClearPageActive(page
); /* we are de-activating */
1688 list_add(&page
->lru
, &l_inactive
);
1692 * Move pages back to the lru list.
1694 spin_lock_irq(&zone
->lru_lock
);
1696 * Count referenced pages from currently used mappings as rotated,
1697 * even though only some of them are actually re-activated. This
1698 * helps balance scan pressure between file and anonymous pages in
1701 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1703 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1704 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1705 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1706 spin_unlock_irq(&zone
->lru_lock
);
1708 free_hot_cold_page_list(&l_hold
, 1);
1712 static int inactive_anon_is_low_global(struct zone
*zone
)
1714 unsigned long active
, inactive
;
1716 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1717 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1719 if (inactive
* zone
->inactive_ratio
< active
)
1726 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1727 * @lruvec: LRU vector to check
1729 * Returns true if the zone does not have enough inactive anon pages,
1730 * meaning some active anon pages need to be deactivated.
1732 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1735 * If we don't have swap space, anonymous page deactivation
1738 if (!total_swap_pages
)
1741 if (!mem_cgroup_disabled())
1742 return mem_cgroup_inactive_anon_is_low(lruvec
);
1744 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1747 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1754 * inactive_file_is_low - check if file pages need to be deactivated
1755 * @lruvec: LRU vector to check
1757 * When the system is doing streaming IO, memory pressure here
1758 * ensures that active file pages get deactivated, until more
1759 * than half of the file pages are on the inactive list.
1761 * Once we get to that situation, protect the system's working
1762 * set from being evicted by disabling active file page aging.
1764 * This uses a different ratio than the anonymous pages, because
1765 * the page cache uses a use-once replacement algorithm.
1767 static int inactive_file_is_low(struct lruvec
*lruvec
)
1769 unsigned long inactive
;
1770 unsigned long active
;
1772 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1773 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1775 return active
> inactive
;
1778 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1780 if (is_file_lru(lru
))
1781 return inactive_file_is_low(lruvec
);
1783 return inactive_anon_is_low(lruvec
);
1786 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1787 struct lruvec
*lruvec
, struct scan_control
*sc
)
1789 if (is_active_lru(lru
)) {
1790 if (inactive_list_is_low(lruvec
, lru
))
1791 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1795 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1798 static int vmscan_swappiness(struct scan_control
*sc
)
1800 if (global_reclaim(sc
))
1801 return vm_swappiness
;
1802 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1813 * Determine how aggressively the anon and file LRU lists should be
1814 * scanned. The relative value of each set of LRU lists is determined
1815 * by looking at the fraction of the pages scanned we did rotate back
1816 * onto the active list instead of evict.
1818 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1819 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1821 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
1824 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1826 u64 denominator
= 0; /* gcc */
1827 struct zone
*zone
= lruvec_zone(lruvec
);
1828 unsigned long anon_prio
, file_prio
;
1829 enum scan_balance scan_balance
;
1830 unsigned long anon
, file
, free
;
1831 bool force_scan
= false;
1832 unsigned long ap
, fp
;
1836 * If the zone or memcg is small, nr[l] can be 0. This
1837 * results in no scanning on this priority and a potential
1838 * priority drop. Global direct reclaim can go to the next
1839 * zone and tends to have no problems. Global kswapd is for
1840 * zone balancing and it needs to scan a minimum amount. When
1841 * reclaiming for a memcg, a priority drop can cause high
1842 * latencies, so it's better to scan a minimum amount there as
1845 if (current_is_kswapd() && !zone_reclaimable(zone
))
1847 if (!global_reclaim(sc
))
1850 /* If we have no swap space, do not bother scanning anon pages. */
1851 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
1852 scan_balance
= SCAN_FILE
;
1857 * Global reclaim will swap to prevent OOM even with no
1858 * swappiness, but memcg users want to use this knob to
1859 * disable swapping for individual groups completely when
1860 * using the memory controller's swap limit feature would be
1863 if (!global_reclaim(sc
) && !vmscan_swappiness(sc
)) {
1864 scan_balance
= SCAN_FILE
;
1869 * Do not apply any pressure balancing cleverness when the
1870 * system is close to OOM, scan both anon and file equally
1871 * (unless the swappiness setting disagrees with swapping).
1873 if (!sc
->priority
&& vmscan_swappiness(sc
)) {
1874 scan_balance
= SCAN_EQUAL
;
1878 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1879 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1880 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1881 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1884 * If it's foreseeable that reclaiming the file cache won't be
1885 * enough to get the zone back into a desirable shape, we have
1886 * to swap. Better start now and leave the - probably heavily
1887 * thrashing - remaining file pages alone.
1889 if (global_reclaim(sc
)) {
1890 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1891 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1892 scan_balance
= SCAN_ANON
;
1898 * There is enough inactive page cache, do not reclaim
1899 * anything from the anonymous working set right now.
1901 if (!inactive_file_is_low(lruvec
)) {
1902 scan_balance
= SCAN_FILE
;
1906 scan_balance
= SCAN_FRACT
;
1909 * With swappiness at 100, anonymous and file have the same priority.
1910 * This scanning priority is essentially the inverse of IO cost.
1912 anon_prio
= vmscan_swappiness(sc
);
1913 file_prio
= 200 - anon_prio
;
1916 * OK, so we have swap space and a fair amount of page cache
1917 * pages. We use the recently rotated / recently scanned
1918 * ratios to determine how valuable each cache is.
1920 * Because workloads change over time (and to avoid overflow)
1921 * we keep these statistics as a floating average, which ends
1922 * up weighing recent references more than old ones.
1924 * anon in [0], file in [1]
1926 spin_lock_irq(&zone
->lru_lock
);
1927 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1928 reclaim_stat
->recent_scanned
[0] /= 2;
1929 reclaim_stat
->recent_rotated
[0] /= 2;
1932 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1933 reclaim_stat
->recent_scanned
[1] /= 2;
1934 reclaim_stat
->recent_rotated
[1] /= 2;
1938 * The amount of pressure on anon vs file pages is inversely
1939 * proportional to the fraction of recently scanned pages on
1940 * each list that were recently referenced and in active use.
1942 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1943 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1945 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1946 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1947 spin_unlock_irq(&zone
->lru_lock
);
1951 denominator
= ap
+ fp
+ 1;
1953 for_each_evictable_lru(lru
) {
1954 int file
= is_file_lru(lru
);
1958 size
= get_lru_size(lruvec
, lru
);
1959 scan
= size
>> sc
->priority
;
1961 if (!scan
&& force_scan
)
1962 scan
= min(size
, SWAP_CLUSTER_MAX
);
1964 switch (scan_balance
) {
1966 /* Scan lists relative to size */
1970 * Scan types proportional to swappiness and
1971 * their relative recent reclaim efficiency.
1973 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1977 /* Scan one type exclusively */
1978 if ((scan_balance
== SCAN_FILE
) != file
)
1982 /* Look ma, no brain */
1990 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1992 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
1994 unsigned long nr
[NR_LRU_LISTS
];
1995 unsigned long targets
[NR_LRU_LISTS
];
1996 unsigned long nr_to_scan
;
1998 unsigned long nr_reclaimed
= 0;
1999 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2000 struct blk_plug plug
;
2001 bool scan_adjusted
= false;
2003 get_scan_count(lruvec
, sc
, nr
);
2005 /* Record the original scan target for proportional adjustments later */
2006 memcpy(targets
, nr
, sizeof(nr
));
2008 blk_start_plug(&plug
);
2009 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2010 nr
[LRU_INACTIVE_FILE
]) {
2011 unsigned long nr_anon
, nr_file
, percentage
;
2012 unsigned long nr_scanned
;
2014 for_each_evictable_lru(lru
) {
2016 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2017 nr
[lru
] -= nr_to_scan
;
2019 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2024 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2028 * For global direct reclaim, reclaim only the number of pages
2029 * requested. Less care is taken to scan proportionally as it
2030 * is more important to minimise direct reclaim stall latency
2031 * than it is to properly age the LRU lists.
2033 if (global_reclaim(sc
) && !current_is_kswapd())
2037 * For kswapd and memcg, reclaim at least the number of pages
2038 * requested. Ensure that the anon and file LRUs shrink
2039 * proportionally what was requested by get_scan_count(). We
2040 * stop reclaiming one LRU and reduce the amount scanning
2041 * proportional to the original scan target.
2043 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2044 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2046 if (nr_file
> nr_anon
) {
2047 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2048 targets
[LRU_ACTIVE_ANON
] + 1;
2050 percentage
= nr_anon
* 100 / scan_target
;
2052 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2053 targets
[LRU_ACTIVE_FILE
] + 1;
2055 percentage
= nr_file
* 100 / scan_target
;
2058 /* Stop scanning the smaller of the LRU */
2060 nr
[lru
+ LRU_ACTIVE
] = 0;
2063 * Recalculate the other LRU scan count based on its original
2064 * scan target and the percentage scanning already complete
2066 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2067 nr_scanned
= targets
[lru
] - nr
[lru
];
2068 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2069 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2072 nr_scanned
= targets
[lru
] - nr
[lru
];
2073 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2074 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2076 scan_adjusted
= true;
2078 blk_finish_plug(&plug
);
2079 sc
->nr_reclaimed
+= nr_reclaimed
;
2082 * Even if we did not try to evict anon pages at all, we want to
2083 * rebalance the anon lru active/inactive ratio.
2085 if (inactive_anon_is_low(lruvec
))
2086 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2087 sc
, LRU_ACTIVE_ANON
);
2089 throttle_vm_writeout(sc
->gfp_mask
);
2092 /* Use reclaim/compaction for costly allocs or under memory pressure */
2093 static bool in_reclaim_compaction(struct scan_control
*sc
)
2095 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2096 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2097 sc
->priority
< DEF_PRIORITY
- 2))
2104 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2105 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2106 * true if more pages should be reclaimed such that when the page allocator
2107 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2108 * It will give up earlier than that if there is difficulty reclaiming pages.
2110 static inline bool should_continue_reclaim(struct zone
*zone
,
2111 unsigned long nr_reclaimed
,
2112 unsigned long nr_scanned
,
2113 struct scan_control
*sc
)
2115 unsigned long pages_for_compaction
;
2116 unsigned long inactive_lru_pages
;
2118 /* If not in reclaim/compaction mode, stop */
2119 if (!in_reclaim_compaction(sc
))
2122 /* Consider stopping depending on scan and reclaim activity */
2123 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2125 * For __GFP_REPEAT allocations, stop reclaiming if the
2126 * full LRU list has been scanned and we are still failing
2127 * to reclaim pages. This full LRU scan is potentially
2128 * expensive but a __GFP_REPEAT caller really wants to succeed
2130 if (!nr_reclaimed
&& !nr_scanned
)
2134 * For non-__GFP_REPEAT allocations which can presumably
2135 * fail without consequence, stop if we failed to reclaim
2136 * any pages from the last SWAP_CLUSTER_MAX number of
2137 * pages that were scanned. This will return to the
2138 * caller faster at the risk reclaim/compaction and
2139 * the resulting allocation attempt fails
2146 * If we have not reclaimed enough pages for compaction and the
2147 * inactive lists are large enough, continue reclaiming
2149 pages_for_compaction
= (2UL << sc
->order
);
2150 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2151 if (get_nr_swap_pages() > 0)
2152 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2153 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2154 inactive_lru_pages
> pages_for_compaction
)
2157 /* If compaction would go ahead or the allocation would succeed, stop */
2158 switch (compaction_suitable(zone
, sc
->order
)) {
2159 case COMPACT_PARTIAL
:
2160 case COMPACT_CONTINUE
:
2167 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
2169 unsigned long nr_reclaimed
, nr_scanned
;
2172 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2173 struct mem_cgroup_reclaim_cookie reclaim
= {
2175 .priority
= sc
->priority
,
2177 struct mem_cgroup
*memcg
;
2179 nr_reclaimed
= sc
->nr_reclaimed
;
2180 nr_scanned
= sc
->nr_scanned
;
2182 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2184 struct lruvec
*lruvec
;
2186 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2188 shrink_lruvec(lruvec
, sc
);
2191 * Direct reclaim and kswapd have to scan all memory
2192 * cgroups to fulfill the overall scan target for the
2195 * Limit reclaim, on the other hand, only cares about
2196 * nr_to_reclaim pages to be reclaimed and it will
2197 * retry with decreasing priority if one round over the
2198 * whole hierarchy is not sufficient.
2200 if (!global_reclaim(sc
) &&
2201 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2202 mem_cgroup_iter_break(root
, memcg
);
2205 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
2208 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2209 sc
->nr_scanned
- nr_scanned
,
2210 sc
->nr_reclaimed
- nr_reclaimed
);
2212 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2213 sc
->nr_scanned
- nr_scanned
, sc
));
2216 /* Returns true if compaction should go ahead for a high-order request */
2217 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2219 unsigned long balance_gap
, watermark
;
2222 /* Do not consider compaction for orders reclaim is meant to satisfy */
2223 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
2227 * Compaction takes time to run and there are potentially other
2228 * callers using the pages just freed. Continue reclaiming until
2229 * there is a buffer of free pages available to give compaction
2230 * a reasonable chance of completing and allocating the page
2232 balance_gap
= min(low_wmark_pages(zone
),
2233 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2234 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2235 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
2236 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2239 * If compaction is deferred, reclaim up to a point where
2240 * compaction will have a chance of success when re-enabled
2242 if (compaction_deferred(zone
, sc
->order
))
2243 return watermark_ok
;
2245 /* If compaction is not ready to start, keep reclaiming */
2246 if (!compaction_suitable(zone
, sc
->order
))
2249 return watermark_ok
;
2253 * This is the direct reclaim path, for page-allocating processes. We only
2254 * try to reclaim pages from zones which will satisfy the caller's allocation
2257 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2259 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2261 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2262 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2263 * zone defense algorithm.
2265 * If a zone is deemed to be full of pinned pages then just give it a light
2266 * scan then give up on it.
2268 * This function returns true if a zone is being reclaimed for a costly
2269 * high-order allocation and compaction is ready to begin. This indicates to
2270 * the caller that it should consider retrying the allocation instead of
2273 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2277 unsigned long nr_soft_reclaimed
;
2278 unsigned long nr_soft_scanned
;
2279 bool aborted_reclaim
= false;
2282 * If the number of buffer_heads in the machine exceeds the maximum
2283 * allowed level, force direct reclaim to scan the highmem zone as
2284 * highmem pages could be pinning lowmem pages storing buffer_heads
2286 if (buffer_heads_over_limit
)
2287 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2289 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2290 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2291 if (!populated_zone(zone
))
2294 * Take care memory controller reclaiming has small influence
2297 if (global_reclaim(sc
)) {
2298 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2300 if (sc
->priority
!= DEF_PRIORITY
&&
2301 !zone_reclaimable(zone
))
2302 continue; /* Let kswapd poll it */
2303 if (IS_ENABLED(CONFIG_COMPACTION
)) {
2305 * If we already have plenty of memory free for
2306 * compaction in this zone, don't free any more.
2307 * Even though compaction is invoked for any
2308 * non-zero order, only frequent costly order
2309 * reclamation is disruptive enough to become a
2310 * noticeable problem, like transparent huge
2313 if (compaction_ready(zone
, sc
)) {
2314 aborted_reclaim
= true;
2319 * This steals pages from memory cgroups over softlimit
2320 * and returns the number of reclaimed pages and
2321 * scanned pages. This works for global memory pressure
2322 * and balancing, not for a memcg's limit.
2324 nr_soft_scanned
= 0;
2325 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2326 sc
->order
, sc
->gfp_mask
,
2328 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2329 sc
->nr_scanned
+= nr_soft_scanned
;
2330 /* need some check for avoid more shrink_zone() */
2333 shrink_zone(zone
, sc
);
2336 return aborted_reclaim
;
2339 /* All zones in zonelist are unreclaimable? */
2340 static bool all_unreclaimable(struct zonelist
*zonelist
,
2341 struct scan_control
*sc
)
2346 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2347 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2348 if (!populated_zone(zone
))
2350 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2352 if (zone_reclaimable(zone
))
2360 * This is the main entry point to direct page reclaim.
2362 * If a full scan of the inactive list fails to free enough memory then we
2363 * are "out of memory" and something needs to be killed.
2365 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2366 * high - the zone may be full of dirty or under-writeback pages, which this
2367 * caller can't do much about. We kick the writeback threads and take explicit
2368 * naps in the hope that some of these pages can be written. But if the
2369 * allocating task holds filesystem locks which prevent writeout this might not
2370 * work, and the allocation attempt will fail.
2372 * returns: 0, if no pages reclaimed
2373 * else, the number of pages reclaimed
2375 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2376 struct scan_control
*sc
,
2377 struct shrink_control
*shrink
)
2379 unsigned long total_scanned
= 0;
2380 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2383 unsigned long writeback_threshold
;
2384 bool aborted_reclaim
;
2386 delayacct_freepages_start();
2388 if (global_reclaim(sc
))
2389 count_vm_event(ALLOCSTALL
);
2392 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2395 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2398 * Don't shrink slabs when reclaiming memory from over limit
2399 * cgroups but do shrink slab at least once when aborting
2400 * reclaim for compaction to avoid unevenly scanning file/anon
2401 * LRU pages over slab pages.
2403 if (global_reclaim(sc
)) {
2404 unsigned long lru_pages
= 0;
2406 nodes_clear(shrink
->nodes_to_scan
);
2407 for_each_zone_zonelist(zone
, z
, zonelist
,
2408 gfp_zone(sc
->gfp_mask
)) {
2409 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2412 lru_pages
+= zone_reclaimable_pages(zone
);
2413 node_set(zone_to_nid(zone
),
2414 shrink
->nodes_to_scan
);
2417 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2418 if (reclaim_state
) {
2419 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2420 reclaim_state
->reclaimed_slab
= 0;
2423 total_scanned
+= sc
->nr_scanned
;
2424 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2428 * If we're getting trouble reclaiming, start doing
2429 * writepage even in laptop mode.
2431 if (sc
->priority
< DEF_PRIORITY
- 2)
2432 sc
->may_writepage
= 1;
2435 * Try to write back as many pages as we just scanned. This
2436 * tends to cause slow streaming writers to write data to the
2437 * disk smoothly, at the dirtying rate, which is nice. But
2438 * that's undesirable in laptop mode, where we *want* lumpy
2439 * writeout. So in laptop mode, write out the whole world.
2441 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2442 if (total_scanned
> writeback_threshold
) {
2443 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2444 WB_REASON_TRY_TO_FREE_PAGES
);
2445 sc
->may_writepage
= 1;
2447 } while (--sc
->priority
>= 0 && !aborted_reclaim
);
2450 delayacct_freepages_end();
2452 if (sc
->nr_reclaimed
)
2453 return sc
->nr_reclaimed
;
2456 * As hibernation is going on, kswapd is freezed so that it can't mark
2457 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2460 if (oom_killer_disabled
)
2463 /* Aborted reclaim to try compaction? don't OOM, then */
2464 if (aborted_reclaim
)
2467 /* top priority shrink_zones still had more to do? don't OOM, then */
2468 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2474 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2477 unsigned long pfmemalloc_reserve
= 0;
2478 unsigned long free_pages
= 0;
2482 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2483 zone
= &pgdat
->node_zones
[i
];
2484 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2485 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2488 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2490 /* kswapd must be awake if processes are being throttled */
2491 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2492 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2493 (enum zone_type
)ZONE_NORMAL
);
2494 wake_up_interruptible(&pgdat
->kswapd_wait
);
2501 * Throttle direct reclaimers if backing storage is backed by the network
2502 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2503 * depleted. kswapd will continue to make progress and wake the processes
2504 * when the low watermark is reached.
2506 * Returns true if a fatal signal was delivered during throttling. If this
2507 * happens, the page allocator should not consider triggering the OOM killer.
2509 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2510 nodemask_t
*nodemask
)
2513 int high_zoneidx
= gfp_zone(gfp_mask
);
2517 * Kernel threads should not be throttled as they may be indirectly
2518 * responsible for cleaning pages necessary for reclaim to make forward
2519 * progress. kjournald for example may enter direct reclaim while
2520 * committing a transaction where throttling it could forcing other
2521 * processes to block on log_wait_commit().
2523 if (current
->flags
& PF_KTHREAD
)
2527 * If a fatal signal is pending, this process should not throttle.
2528 * It should return quickly so it can exit and free its memory
2530 if (fatal_signal_pending(current
))
2533 /* Check if the pfmemalloc reserves are ok */
2534 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
, &zone
);
2535 pgdat
= zone
->zone_pgdat
;
2536 if (pfmemalloc_watermark_ok(pgdat
))
2539 /* Account for the throttling */
2540 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2543 * If the caller cannot enter the filesystem, it's possible that it
2544 * is due to the caller holding an FS lock or performing a journal
2545 * transaction in the case of a filesystem like ext[3|4]. In this case,
2546 * it is not safe to block on pfmemalloc_wait as kswapd could be
2547 * blocked waiting on the same lock. Instead, throttle for up to a
2548 * second before continuing.
2550 if (!(gfp_mask
& __GFP_FS
)) {
2551 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2552 pfmemalloc_watermark_ok(pgdat
), HZ
);
2557 /* Throttle until kswapd wakes the process */
2558 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2559 pfmemalloc_watermark_ok(pgdat
));
2562 if (fatal_signal_pending(current
))
2569 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2570 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2572 unsigned long nr_reclaimed
;
2573 struct scan_control sc
= {
2574 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2575 .may_writepage
= !laptop_mode
,
2576 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2580 .priority
= DEF_PRIORITY
,
2581 .target_mem_cgroup
= NULL
,
2582 .nodemask
= nodemask
,
2584 struct shrink_control shrink
= {
2585 .gfp_mask
= sc
.gfp_mask
,
2589 * Do not enter reclaim if fatal signal was delivered while throttled.
2590 * 1 is returned so that the page allocator does not OOM kill at this
2593 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2596 trace_mm_vmscan_direct_reclaim_begin(order
,
2600 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2602 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2604 return nr_reclaimed
;
2609 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2610 gfp_t gfp_mask
, bool noswap
,
2612 unsigned long *nr_scanned
)
2614 struct scan_control sc
= {
2616 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2617 .may_writepage
= !laptop_mode
,
2619 .may_swap
= !noswap
,
2622 .target_mem_cgroup
= memcg
,
2624 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2626 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2627 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2629 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2634 * NOTE: Although we can get the priority field, using it
2635 * here is not a good idea, since it limits the pages we can scan.
2636 * if we don't reclaim here, the shrink_zone from balance_pgdat
2637 * will pick up pages from other mem cgroup's as well. We hack
2638 * the priority and make it zero.
2640 shrink_lruvec(lruvec
, &sc
);
2642 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2644 *nr_scanned
= sc
.nr_scanned
;
2645 return sc
.nr_reclaimed
;
2648 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2652 struct zonelist
*zonelist
;
2653 unsigned long nr_reclaimed
;
2655 struct scan_control sc
= {
2656 .may_writepage
= !laptop_mode
,
2658 .may_swap
= !noswap
,
2659 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2661 .priority
= DEF_PRIORITY
,
2662 .target_mem_cgroup
= memcg
,
2663 .nodemask
= NULL
, /* we don't care the placement */
2664 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2665 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2667 struct shrink_control shrink
= {
2668 .gfp_mask
= sc
.gfp_mask
,
2672 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2673 * take care of from where we get pages. So the node where we start the
2674 * scan does not need to be the current node.
2676 nid
= mem_cgroup_select_victim_node(memcg
);
2678 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2680 trace_mm_vmscan_memcg_reclaim_begin(0,
2684 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2686 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2688 return nr_reclaimed
;
2692 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2694 struct mem_cgroup
*memcg
;
2696 if (!total_swap_pages
)
2699 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2701 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2703 if (inactive_anon_is_low(lruvec
))
2704 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2705 sc
, LRU_ACTIVE_ANON
);
2707 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2711 static bool zone_balanced(struct zone
*zone
, int order
,
2712 unsigned long balance_gap
, int classzone_idx
)
2714 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2715 balance_gap
, classzone_idx
, 0))
2718 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&&
2719 !compaction_suitable(zone
, order
))
2726 * pgdat_balanced() is used when checking if a node is balanced.
2728 * For order-0, all zones must be balanced!
2730 * For high-order allocations only zones that meet watermarks and are in a
2731 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2732 * total of balanced pages must be at least 25% of the zones allowed by
2733 * classzone_idx for the node to be considered balanced. Forcing all zones to
2734 * be balanced for high orders can cause excessive reclaim when there are
2736 * The choice of 25% is due to
2737 * o a 16M DMA zone that is balanced will not balance a zone on any
2738 * reasonable sized machine
2739 * o On all other machines, the top zone must be at least a reasonable
2740 * percentage of the middle zones. For example, on 32-bit x86, highmem
2741 * would need to be at least 256M for it to be balance a whole node.
2742 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2743 * to balance a node on its own. These seemed like reasonable ratios.
2745 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2747 unsigned long managed_pages
= 0;
2748 unsigned long balanced_pages
= 0;
2751 /* Check the watermark levels */
2752 for (i
= 0; i
<= classzone_idx
; i
++) {
2753 struct zone
*zone
= pgdat
->node_zones
+ i
;
2755 if (!populated_zone(zone
))
2758 managed_pages
+= zone
->managed_pages
;
2761 * A special case here:
2763 * balance_pgdat() skips over all_unreclaimable after
2764 * DEF_PRIORITY. Effectively, it considers them balanced so
2765 * they must be considered balanced here as well!
2767 if (!zone_reclaimable(zone
)) {
2768 balanced_pages
+= zone
->managed_pages
;
2772 if (zone_balanced(zone
, order
, 0, i
))
2773 balanced_pages
+= zone
->managed_pages
;
2779 return balanced_pages
>= (managed_pages
>> 2);
2785 * Prepare kswapd for sleeping. This verifies that there are no processes
2786 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2788 * Returns true if kswapd is ready to sleep
2790 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
2793 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2798 * There is a potential race between when kswapd checks its watermarks
2799 * and a process gets throttled. There is also a potential race if
2800 * processes get throttled, kswapd wakes, a large process exits therby
2801 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2802 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2803 * so wake them now if necessary. If necessary, processes will wake
2804 * kswapd and get throttled again
2806 if (waitqueue_active(&pgdat
->pfmemalloc_wait
)) {
2807 wake_up(&pgdat
->pfmemalloc_wait
);
2811 return pgdat_balanced(pgdat
, order
, classzone_idx
);
2815 * kswapd shrinks the zone by the number of pages required to reach
2816 * the high watermark.
2818 * Returns true if kswapd scanned at least the requested number of pages to
2819 * reclaim or if the lack of progress was due to pages under writeback.
2820 * This is used to determine if the scanning priority needs to be raised.
2822 static bool kswapd_shrink_zone(struct zone
*zone
,
2824 struct scan_control
*sc
,
2825 unsigned long lru_pages
,
2826 unsigned long *nr_attempted
)
2828 int testorder
= sc
->order
;
2829 unsigned long balance_gap
;
2830 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2831 struct shrink_control shrink
= {
2832 .gfp_mask
= sc
->gfp_mask
,
2834 bool lowmem_pressure
;
2836 /* Reclaim above the high watermark. */
2837 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
2840 * Kswapd reclaims only single pages with compaction enabled. Trying
2841 * too hard to reclaim until contiguous free pages have become
2842 * available can hurt performance by evicting too much useful data
2843 * from memory. Do not reclaim more than needed for compaction.
2845 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2846 compaction_suitable(zone
, sc
->order
) !=
2851 * We put equal pressure on every zone, unless one zone has way too
2852 * many pages free already. The "too many pages" is defined as the
2853 * high wmark plus a "gap" where the gap is either the low
2854 * watermark or 1% of the zone, whichever is smaller.
2856 balance_gap
= min(low_wmark_pages(zone
),
2857 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2858 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2861 * If there is no low memory pressure or the zone is balanced then no
2862 * reclaim is necessary
2864 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
2865 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
2866 balance_gap
, classzone_idx
))
2869 shrink_zone(zone
, sc
);
2870 nodes_clear(shrink
.nodes_to_scan
);
2871 node_set(zone_to_nid(zone
), shrink
.nodes_to_scan
);
2873 reclaim_state
->reclaimed_slab
= 0;
2874 shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2875 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2877 /* Account for the number of pages attempted to reclaim */
2878 *nr_attempted
+= sc
->nr_to_reclaim
;
2880 zone_clear_flag(zone
, ZONE_WRITEBACK
);
2883 * If a zone reaches its high watermark, consider it to be no longer
2884 * congested. It's possible there are dirty pages backed by congested
2885 * BDIs but as pressure is relieved, speculatively avoid congestion
2888 if (zone_reclaimable(zone
) &&
2889 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
2890 zone_clear_flag(zone
, ZONE_CONGESTED
);
2891 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2894 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
2898 * For kswapd, balance_pgdat() will work across all this node's zones until
2899 * they are all at high_wmark_pages(zone).
2901 * Returns the final order kswapd was reclaiming at
2903 * There is special handling here for zones which are full of pinned pages.
2904 * This can happen if the pages are all mlocked, or if they are all used by
2905 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2906 * What we do is to detect the case where all pages in the zone have been
2907 * scanned twice and there has been zero successful reclaim. Mark the zone as
2908 * dead and from now on, only perform a short scan. Basically we're polling
2909 * the zone for when the problem goes away.
2911 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2912 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2913 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2914 * lower zones regardless of the number of free pages in the lower zones. This
2915 * interoperates with the page allocator fallback scheme to ensure that aging
2916 * of pages is balanced across the zones.
2918 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2922 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2923 unsigned long nr_soft_reclaimed
;
2924 unsigned long nr_soft_scanned
;
2925 struct scan_control sc
= {
2926 .gfp_mask
= GFP_KERNEL
,
2927 .priority
= DEF_PRIORITY
,
2930 .may_writepage
= !laptop_mode
,
2932 .target_mem_cgroup
= NULL
,
2934 count_vm_event(PAGEOUTRUN
);
2937 unsigned long lru_pages
= 0;
2938 unsigned long nr_attempted
= 0;
2939 bool raise_priority
= true;
2940 bool pgdat_needs_compaction
= (order
> 0);
2942 sc
.nr_reclaimed
= 0;
2945 * Scan in the highmem->dma direction for the highest
2946 * zone which needs scanning
2948 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2949 struct zone
*zone
= pgdat
->node_zones
+ i
;
2951 if (!populated_zone(zone
))
2954 if (sc
.priority
!= DEF_PRIORITY
&&
2955 !zone_reclaimable(zone
))
2959 * Do some background aging of the anon list, to give
2960 * pages a chance to be referenced before reclaiming.
2962 age_active_anon(zone
, &sc
);
2965 * If the number of buffer_heads in the machine
2966 * exceeds the maximum allowed level and this node
2967 * has a highmem zone, force kswapd to reclaim from
2968 * it to relieve lowmem pressure.
2970 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2975 if (!zone_balanced(zone
, order
, 0, 0)) {
2980 * If balanced, clear the dirty and congested
2983 zone_clear_flag(zone
, ZONE_CONGESTED
);
2984 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2991 for (i
= 0; i
<= end_zone
; i
++) {
2992 struct zone
*zone
= pgdat
->node_zones
+ i
;
2994 if (!populated_zone(zone
))
2997 lru_pages
+= zone_reclaimable_pages(zone
);
3000 * If any zone is currently balanced then kswapd will
3001 * not call compaction as it is expected that the
3002 * necessary pages are already available.
3004 if (pgdat_needs_compaction
&&
3005 zone_watermark_ok(zone
, order
,
3006 low_wmark_pages(zone
),
3008 pgdat_needs_compaction
= false;
3012 * If we're getting trouble reclaiming, start doing writepage
3013 * even in laptop mode.
3015 if (sc
.priority
< DEF_PRIORITY
- 2)
3016 sc
.may_writepage
= 1;
3019 * Now scan the zone in the dma->highmem direction, stopping
3020 * at the last zone which needs scanning.
3022 * We do this because the page allocator works in the opposite
3023 * direction. This prevents the page allocator from allocating
3024 * pages behind kswapd's direction of progress, which would
3025 * cause too much scanning of the lower zones.
3027 for (i
= 0; i
<= end_zone
; i
++) {
3028 struct zone
*zone
= pgdat
->node_zones
+ i
;
3030 if (!populated_zone(zone
))
3033 if (sc
.priority
!= DEF_PRIORITY
&&
3034 !zone_reclaimable(zone
))
3039 nr_soft_scanned
= 0;
3041 * Call soft limit reclaim before calling shrink_zone.
3043 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
3046 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3049 * There should be no need to raise the scanning
3050 * priority if enough pages are already being scanned
3051 * that that high watermark would be met at 100%
3054 if (kswapd_shrink_zone(zone
, end_zone
, &sc
,
3055 lru_pages
, &nr_attempted
))
3056 raise_priority
= false;
3060 * If the low watermark is met there is no need for processes
3061 * to be throttled on pfmemalloc_wait as they should not be
3062 * able to safely make forward progress. Wake them
3064 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3065 pfmemalloc_watermark_ok(pgdat
))
3066 wake_up(&pgdat
->pfmemalloc_wait
);
3069 * Fragmentation may mean that the system cannot be rebalanced
3070 * for high-order allocations in all zones. If twice the
3071 * allocation size has been reclaimed and the zones are still
3072 * not balanced then recheck the watermarks at order-0 to
3073 * prevent kswapd reclaiming excessively. Assume that a
3074 * process requested a high-order can direct reclaim/compact.
3076 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3077 order
= sc
.order
= 0;
3079 /* Check if kswapd should be suspending */
3080 if (try_to_freeze() || kthread_should_stop())
3084 * Compact if necessary and kswapd is reclaiming at least the
3085 * high watermark number of pages as requsted
3087 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3088 compact_pgdat(pgdat
, order
);
3091 * Raise priority if scanning rate is too low or there was no
3092 * progress in reclaiming pages
3094 if (raise_priority
|| !sc
.nr_reclaimed
)
3096 } while (sc
.priority
>= 1 &&
3097 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3101 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3102 * makes a decision on the order we were last reclaiming at. However,
3103 * if another caller entered the allocator slow path while kswapd
3104 * was awake, order will remain at the higher level
3106 *classzone_idx
= end_zone
;
3110 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3115 if (freezing(current
) || kthread_should_stop())
3118 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3120 /* Try to sleep for a short interval */
3121 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3122 remaining
= schedule_timeout(HZ
/10);
3123 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3124 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3128 * After a short sleep, check if it was a premature sleep. If not, then
3129 * go fully to sleep until explicitly woken up.
3131 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3132 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3135 * vmstat counters are not perfectly accurate and the estimated
3136 * value for counters such as NR_FREE_PAGES can deviate from the
3137 * true value by nr_online_cpus * threshold. To avoid the zone
3138 * watermarks being breached while under pressure, we reduce the
3139 * per-cpu vmstat threshold while kswapd is awake and restore
3140 * them before going back to sleep.
3142 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3145 * Compaction records what page blocks it recently failed to
3146 * isolate pages from and skips them in the future scanning.
3147 * When kswapd is going to sleep, it is reasonable to assume
3148 * that pages and compaction may succeed so reset the cache.
3150 reset_isolation_suitable(pgdat
);
3152 if (!kthread_should_stop())
3155 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3158 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3160 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3162 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3166 * The background pageout daemon, started as a kernel thread
3167 * from the init process.
3169 * This basically trickles out pages so that we have _some_
3170 * free memory available even if there is no other activity
3171 * that frees anything up. This is needed for things like routing
3172 * etc, where we otherwise might have all activity going on in
3173 * asynchronous contexts that cannot page things out.
3175 * If there are applications that are active memory-allocators
3176 * (most normal use), this basically shouldn't matter.
3178 static int kswapd(void *p
)
3180 unsigned long order
, new_order
;
3181 unsigned balanced_order
;
3182 int classzone_idx
, new_classzone_idx
;
3183 int balanced_classzone_idx
;
3184 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3185 struct task_struct
*tsk
= current
;
3187 struct reclaim_state reclaim_state
= {
3188 .reclaimed_slab
= 0,
3190 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3192 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3194 if (!cpumask_empty(cpumask
))
3195 set_cpus_allowed_ptr(tsk
, cpumask
);
3196 current
->reclaim_state
= &reclaim_state
;
3199 * Tell the memory management that we're a "memory allocator",
3200 * and that if we need more memory we should get access to it
3201 * regardless (see "__alloc_pages()"). "kswapd" should
3202 * never get caught in the normal page freeing logic.
3204 * (Kswapd normally doesn't need memory anyway, but sometimes
3205 * you need a small amount of memory in order to be able to
3206 * page out something else, and this flag essentially protects
3207 * us from recursively trying to free more memory as we're
3208 * trying to free the first piece of memory in the first place).
3210 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3213 order
= new_order
= 0;
3215 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3216 balanced_classzone_idx
= classzone_idx
;
3221 * If the last balance_pgdat was unsuccessful it's unlikely a
3222 * new request of a similar or harder type will succeed soon
3223 * so consider going to sleep on the basis we reclaimed at
3225 if (balanced_classzone_idx
>= new_classzone_idx
&&
3226 balanced_order
== new_order
) {
3227 new_order
= pgdat
->kswapd_max_order
;
3228 new_classzone_idx
= pgdat
->classzone_idx
;
3229 pgdat
->kswapd_max_order
= 0;
3230 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3233 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3235 * Don't sleep if someone wants a larger 'order'
3236 * allocation or has tigher zone constraints
3239 classzone_idx
= new_classzone_idx
;
3241 kswapd_try_to_sleep(pgdat
, balanced_order
,
3242 balanced_classzone_idx
);
3243 order
= pgdat
->kswapd_max_order
;
3244 classzone_idx
= pgdat
->classzone_idx
;
3246 new_classzone_idx
= classzone_idx
;
3247 pgdat
->kswapd_max_order
= 0;
3248 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3251 ret
= try_to_freeze();
3252 if (kthread_should_stop())
3256 * We can speed up thawing tasks if we don't call balance_pgdat
3257 * after returning from the refrigerator
3260 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3261 balanced_classzone_idx
= classzone_idx
;
3262 balanced_order
= balance_pgdat(pgdat
, order
,
3263 &balanced_classzone_idx
);
3267 current
->reclaim_state
= NULL
;
3272 * A zone is low on free memory, so wake its kswapd task to service it.
3274 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3278 if (!populated_zone(zone
))
3281 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
3283 pgdat
= zone
->zone_pgdat
;
3284 if (pgdat
->kswapd_max_order
< order
) {
3285 pgdat
->kswapd_max_order
= order
;
3286 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3288 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3290 if (zone_balanced(zone
, order
, 0, 0))
3293 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3294 wake_up_interruptible(&pgdat
->kswapd_wait
);
3298 * The reclaimable count would be mostly accurate.
3299 * The less reclaimable pages may be
3300 * - mlocked pages, which will be moved to unevictable list when encountered
3301 * - mapped pages, which may require several travels to be reclaimed
3302 * - dirty pages, which is not "instantly" reclaimable
3304 unsigned long global_reclaimable_pages(void)
3308 nr
= global_page_state(NR_ACTIVE_FILE
) +
3309 global_page_state(NR_INACTIVE_FILE
);
3311 if (get_nr_swap_pages() > 0)
3312 nr
+= global_page_state(NR_ACTIVE_ANON
) +
3313 global_page_state(NR_INACTIVE_ANON
);
3318 #ifdef CONFIG_HIBERNATION
3320 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3323 * Rather than trying to age LRUs the aim is to preserve the overall
3324 * LRU order by reclaiming preferentially
3325 * inactive > active > active referenced > active mapped
3327 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3329 struct reclaim_state reclaim_state
;
3330 struct scan_control sc
= {
3331 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3335 .nr_to_reclaim
= nr_to_reclaim
,
3336 .hibernation_mode
= 1,
3338 .priority
= DEF_PRIORITY
,
3340 struct shrink_control shrink
= {
3341 .gfp_mask
= sc
.gfp_mask
,
3343 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3344 struct task_struct
*p
= current
;
3345 unsigned long nr_reclaimed
;
3347 p
->flags
|= PF_MEMALLOC
;
3348 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3349 reclaim_state
.reclaimed_slab
= 0;
3350 p
->reclaim_state
= &reclaim_state
;
3352 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3354 p
->reclaim_state
= NULL
;
3355 lockdep_clear_current_reclaim_state();
3356 p
->flags
&= ~PF_MEMALLOC
;
3358 return nr_reclaimed
;
3360 #endif /* CONFIG_HIBERNATION */
3362 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3363 not required for correctness. So if the last cpu in a node goes
3364 away, we get changed to run anywhere: as the first one comes back,
3365 restore their cpu bindings. */
3366 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3371 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3372 for_each_node_state(nid
, N_MEMORY
) {
3373 pg_data_t
*pgdat
= NODE_DATA(nid
);
3374 const struct cpumask
*mask
;
3376 mask
= cpumask_of_node(pgdat
->node_id
);
3378 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3379 /* One of our CPUs online: restore mask */
3380 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3387 * This kswapd start function will be called by init and node-hot-add.
3388 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3390 int kswapd_run(int nid
)
3392 pg_data_t
*pgdat
= NODE_DATA(nid
);
3398 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3399 if (IS_ERR(pgdat
->kswapd
)) {
3400 /* failure at boot is fatal */
3401 BUG_ON(system_state
== SYSTEM_BOOTING
);
3402 pr_err("Failed to start kswapd on node %d\n", nid
);
3403 ret
= PTR_ERR(pgdat
->kswapd
);
3404 pgdat
->kswapd
= NULL
;
3410 * Called by memory hotplug when all memory in a node is offlined. Caller must
3411 * hold lock_memory_hotplug().
3413 void kswapd_stop(int nid
)
3415 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3418 kthread_stop(kswapd
);
3419 NODE_DATA(nid
)->kswapd
= NULL
;
3423 static int __init
kswapd_init(void)
3428 for_each_node_state(nid
, N_MEMORY
)
3430 hotcpu_notifier(cpu_callback
, 0);
3434 module_init(kswapd_init
)
3440 * If non-zero call zone_reclaim when the number of free pages falls below
3443 int zone_reclaim_mode __read_mostly
;
3445 #define RECLAIM_OFF 0
3446 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3447 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3448 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3451 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3452 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3455 #define ZONE_RECLAIM_PRIORITY 4
3458 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3461 int sysctl_min_unmapped_ratio
= 1;
3464 * If the number of slab pages in a zone grows beyond this percentage then
3465 * slab reclaim needs to occur.
3467 int sysctl_min_slab_ratio
= 5;
3469 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3471 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3472 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3473 zone_page_state(zone
, NR_ACTIVE_FILE
);
3476 * It's possible for there to be more file mapped pages than
3477 * accounted for by the pages on the file LRU lists because
3478 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3480 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3483 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3484 static long zone_pagecache_reclaimable(struct zone
*zone
)
3486 long nr_pagecache_reclaimable
;
3490 * If RECLAIM_SWAP is set, then all file pages are considered
3491 * potentially reclaimable. Otherwise, we have to worry about
3492 * pages like swapcache and zone_unmapped_file_pages() provides
3495 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3496 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3498 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3500 /* If we can't clean pages, remove dirty pages from consideration */
3501 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3502 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3504 /* Watch for any possible underflows due to delta */
3505 if (unlikely(delta
> nr_pagecache_reclaimable
))
3506 delta
= nr_pagecache_reclaimable
;
3508 return nr_pagecache_reclaimable
- delta
;
3512 * Try to free up some pages from this zone through reclaim.
3514 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3516 /* Minimum pages needed in order to stay on node */
3517 const unsigned long nr_pages
= 1 << order
;
3518 struct task_struct
*p
= current
;
3519 struct reclaim_state reclaim_state
;
3520 struct scan_control sc
= {
3521 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3522 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3524 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3525 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3527 .priority
= ZONE_RECLAIM_PRIORITY
,
3529 struct shrink_control shrink
= {
3530 .gfp_mask
= sc
.gfp_mask
,
3532 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3536 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3537 * and we also need to be able to write out pages for RECLAIM_WRITE
3540 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3541 lockdep_set_current_reclaim_state(gfp_mask
);
3542 reclaim_state
.reclaimed_slab
= 0;
3543 p
->reclaim_state
= &reclaim_state
;
3545 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3547 * Free memory by calling shrink zone with increasing
3548 * priorities until we have enough memory freed.
3551 shrink_zone(zone
, &sc
);
3552 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3555 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3556 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3558 * shrink_slab() does not currently allow us to determine how
3559 * many pages were freed in this zone. So we take the current
3560 * number of slab pages and shake the slab until it is reduced
3561 * by the same nr_pages that we used for reclaiming unmapped
3564 nodes_clear(shrink
.nodes_to_scan
);
3565 node_set(zone_to_nid(zone
), shrink
.nodes_to_scan
);
3567 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3569 /* No reclaimable slab or very low memory pressure */
3570 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3573 /* Freed enough memory */
3574 nr_slab_pages1
= zone_page_state(zone
,
3575 NR_SLAB_RECLAIMABLE
);
3576 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3581 * Update nr_reclaimed by the number of slab pages we
3582 * reclaimed from this zone.
3584 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3585 if (nr_slab_pages1
< nr_slab_pages0
)
3586 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3589 p
->reclaim_state
= NULL
;
3590 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3591 lockdep_clear_current_reclaim_state();
3592 return sc
.nr_reclaimed
>= nr_pages
;
3595 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3601 * Zone reclaim reclaims unmapped file backed pages and
3602 * slab pages if we are over the defined limits.
3604 * A small portion of unmapped file backed pages is needed for
3605 * file I/O otherwise pages read by file I/O will be immediately
3606 * thrown out if the zone is overallocated. So we do not reclaim
3607 * if less than a specified percentage of the zone is used by
3608 * unmapped file backed pages.
3610 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3611 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3612 return ZONE_RECLAIM_FULL
;
3614 if (!zone_reclaimable(zone
))
3615 return ZONE_RECLAIM_FULL
;
3618 * Do not scan if the allocation should not be delayed.
3620 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3621 return ZONE_RECLAIM_NOSCAN
;
3624 * Only run zone reclaim on the local zone or on zones that do not
3625 * have associated processors. This will favor the local processor
3626 * over remote processors and spread off node memory allocations
3627 * as wide as possible.
3629 node_id
= zone_to_nid(zone
);
3630 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3631 return ZONE_RECLAIM_NOSCAN
;
3633 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3634 return ZONE_RECLAIM_NOSCAN
;
3636 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3637 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3640 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3647 * page_evictable - test whether a page is evictable
3648 * @page: the page to test
3650 * Test whether page is evictable--i.e., should be placed on active/inactive
3651 * lists vs unevictable list.
3653 * Reasons page might not be evictable:
3654 * (1) page's mapping marked unevictable
3655 * (2) page is part of an mlocked VMA
3658 int page_evictable(struct page
*page
)
3660 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3665 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3666 * @pages: array of pages to check
3667 * @nr_pages: number of pages to check
3669 * Checks pages for evictability and moves them to the appropriate lru list.
3671 * This function is only used for SysV IPC SHM_UNLOCK.
3673 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3675 struct lruvec
*lruvec
;
3676 struct zone
*zone
= NULL
;
3681 for (i
= 0; i
< nr_pages
; i
++) {
3682 struct page
*page
= pages
[i
];
3683 struct zone
*pagezone
;
3686 pagezone
= page_zone(page
);
3687 if (pagezone
!= zone
) {
3689 spin_unlock_irq(&zone
->lru_lock
);
3691 spin_lock_irq(&zone
->lru_lock
);
3693 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3695 if (!PageLRU(page
) || !PageUnevictable(page
))
3698 if (page_evictable(page
)) {
3699 enum lru_list lru
= page_lru_base_type(page
);
3701 VM_BUG_ON(PageActive(page
));
3702 ClearPageUnevictable(page
);
3703 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3704 add_page_to_lru_list(page
, lruvec
, lru
);
3710 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3711 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3712 spin_unlock_irq(&zone
->lru_lock
);
3715 #endif /* CONFIG_SHMEM */
3717 static void warn_scan_unevictable_pages(void)
3719 printk_once(KERN_WARNING
3720 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3721 "disabled for lack of a legitimate use case. If you have "
3722 "one, please send an email to linux-mm@kvack.org.\n",
3727 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3728 * all nodes' unevictable lists for evictable pages
3730 unsigned long scan_unevictable_pages
;
3732 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3733 void __user
*buffer
,
3734 size_t *length
, loff_t
*ppos
)
3736 warn_scan_unevictable_pages();
3737 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3738 scan_unevictable_pages
= 0;
3744 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3745 * a specified node's per zone unevictable lists for evictable pages.
3748 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3749 struct device_attribute
*attr
,
3752 warn_scan_unevictable_pages();
3753 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3756 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3757 struct device_attribute
*attr
,
3758 const char *buf
, size_t count
)
3760 warn_scan_unevictable_pages();
3765 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3766 read_scan_unevictable_node
,
3767 write_scan_unevictable_node
);
3769 int scan_unevictable_register_node(struct node
*node
)
3771 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3774 void scan_unevictable_unregister_node(struct node
*node
)
3776 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);